INFO 04-12 15:55:59 [__init__.py:216] Automatically detected platform cuda.
2026-04-12 15:56:00,391 | INFO | train_grpo_train | mode=train

Filter:   0%|          | 0/109070 [00:00<?, ? examples/s]
Filter:   1%|          | 1000/109070 [00:00<00:23, 4582.83 examples/s]
Filter:   2%|▏         | 2000/109070 [00:00<00:22, 4760.85 examples/s]
Filter:   3%|▎         | 3000/109070 [00:00<00:22, 4717.41 examples/s]
Filter:   4%|▎         | 4000/109070 [00:00<00:22, 4613.40 examples/s]
Filter:   5%|▍         | 5000/109070 [00:01<00:22, 4620.28 examples/s]
Filter:   6%|▌         | 6000/109070 [00:01<00:22, 4606.96 examples/s]
Filter:   6%|▋         | 7000/109070 [00:01<00:21, 4697.66 examples/s]
Filter:   7%|▋         | 8000/109070 [00:01<00:21, 4787.07 examples/s]
Filter:   8%|▊         | 9000/109070 [00:01<00:20, 4916.39 examples/s]
Filter:   9%|▉         | 10000/109070 [00:02<00:19, 4994.75 examples/s]
Filter:  10%|█         | 11000/109070 [00:02<00:20, 4864.76 examples/s]
Filter:  11%|█         | 12000/109070 [00:02<00:20, 4758.41 examples/s]
Filter:  12%|█▏        | 13000/109070 [00:02<00:20, 4722.20 examples/s]
Filter:  13%|█▎        | 14000/109070 [00:02<00:20, 4719.04 examples/s]
Filter:  14%|█▍        | 15000/109070 [00:03<00:19, 4736.86 examples/s]
Filter:  15%|█▍        | 16000/109070 [00:03<00:19, 4869.97 examples/s]
Filter:  16%|█▌        | 17000/109070 [00:03<00:18, 4884.70 examples/s]
Filter:  17%|█▋        | 18000/109070 [00:03<00:18, 4867.83 examples/s]
Filter:  17%|█▋        | 19000/109070 [00:03<00:18, 4807.59 examples/s]
Filter:  18%|█▊        | 20000/109070 [00:04<00:18, 4761.57 examples/s]
Filter:  19%|█▉        | 21000/109070 [00:04<00:18, 4740.17 examples/s]
Filter:  20%|██        | 22000/109070 [00:04<00:18, 4775.75 examples/s]
Filter:  21%|██        | 23000/109070 [00:04<00:17, 4787.95 examples/s]
Filter:  22%|██▏       | 24000/109070 [00:05<00:17, 4901.97 examples/s]
Filter:  23%|██▎       | 25000/109070 [00:05<00:16, 4987.39 examples/s]
Filter:  24%|██▍       | 26000/109070 [00:05<00:16, 4917.25 examples/s]
Filter:  25%|██▍       | 27000/109070 [00:05<00:17, 4822.57 examples/s]
Filter:  26%|██▌       | 28000/109070 [00:05<00:17, 4683.01 examples/s]
Filter:  27%|██▋       | 29000/109070 [00:06<00:17, 4662.51 examples/s]
Filter:  28%|██▊       | 30000/109070 [00:06<00:16, 4734.17 examples/s]
Filter:  28%|██▊       | 31000/109070 [00:06<00:16, 4817.68 examples/s]
Filter:  29%|██▉       | 32000/109070 [00:06<00:15, 4907.13 examples/s]
Filter:  30%|███       | 33000/109070 [00:06<00:15, 4892.10 examples/s]
Filter:  31%|███       | 34000/109070 [00:07<00:15, 4745.50 examples/s]
Filter:  32%|███▏      | 35000/109070 [00:07<00:15, 4688.67 examples/s]
Filter:  33%|███▎      | 36000/109070 [00:07<00:15, 4650.63 examples/s]
Filter:  34%|███▍      | 37000/109070 [00:07<00:15, 4670.21 examples/s]
Filter:  35%|███▍      | 38000/109070 [00:07<00:14, 4773.01 examples/s]
Filter:  36%|███▌      | 39000/109070 [00:08<00:14, 4837.89 examples/s]
Filter:  37%|███▋      | 40000/109070 [00:08<00:14, 4821.93 examples/s]
Filter:  38%|███▊      | 41000/109070 [00:08<00:14, 4803.80 examples/s]
Filter:  39%|███▊      | 42000/109070 [00:08<00:14, 4726.11 examples/s]
Filter:  39%|███▉      | 43000/109070 [00:09<00:14, 4701.03 examples/s]
Filter:  40%|████      | 44000/109070 [00:09<00:13, 4695.50 examples/s]
Filter:  41%|████▏     | 45000/109070 [00:09<00:13, 4690.01 examples/s]
Filter:  42%|████▏     | 46000/109070 [00:09<00:13, 4639.54 examples/s]
Filter:  43%|████▎     | 47000/109070 [00:09<00:13, 4685.62 examples/s]
Filter:  44%|████▍     | 48000/109070 [00:10<00:12, 4724.90 examples/s]
Filter:  45%|████▍     | 49000/109070 [00:10<00:12, 4639.64 examples/s]
Filter:  46%|████▌     | 50000/109070 [00:10<00:12, 4589.83 examples/s]
Filter:  47%|████▋     | 51000/109070 [00:10<00:12, 4599.19 examples/s]
Filter:  48%|████▊     | 52000/109070 [00:10<00:12, 4629.70 examples/s]
Filter:  49%|████▊     | 53000/109070 [00:11<00:11, 4733.52 examples/s]
Filter:  50%|████▉     | 54000/109070 [00:11<00:11, 4691.66 examples/s]
Filter:  50%|█████     | 55000/109070 [00:11<00:11, 4709.40 examples/s]
Filter:  51%|█████▏    | 56000/109070 [00:11<00:11, 4695.29 examples/s]
Filter:  52%|█████▏    | 57000/109070 [00:12<00:11, 4698.11 examples/s]
Filter:  53%|█████▎    | 58000/109070 [00:12<00:10, 4684.28 examples/s]
Filter:  54%|█████▍    | 59000/109070 [00:12<00:10, 4694.89 examples/s]
Filter:  55%|█████▌    | 60000/109070 [00:12<00:10, 4744.23 examples/s]
Filter:  56%|█████▌    | 61000/109070 [00:12<00:09, 4865.72 examples/s]
Filter:  57%|█████▋    | 62000/109070 [00:13<00:09, 4949.94 examples/s]
Filter:  58%|█████▊    | 63000/109070 [00:13<00:09, 4973.73 examples/s]
Filter:  59%|█████▊    | 64000/109070 [00:13<00:09, 4869.34 examples/s]
Filter:  60%|█████▉    | 65000/109070 [00:13<00:09, 4803.71 examples/s]
Filter:  61%|██████    | 66000/109070 [00:13<00:09, 4749.12 examples/s]
Filter:  61%|██████▏   | 67000/109070 [00:14<00:08, 4806.44 examples/s]
Filter:  62%|██████▏   | 68000/109070 [00:14<00:08, 4891.21 examples/s]
Filter:  63%|██████▎   | 69000/109070 [00:14<00:08, 4974.66 examples/s]
Filter:  64%|██████▍   | 70000/109070 [00:14<00:07, 5006.18 examples/s]
Filter:  65%|██████▌   | 71000/109070 [00:14<00:07, 4882.64 examples/s]
Filter:  66%|██████▌   | 72000/109070 [00:15<00:07, 4778.39 examples/s]
Filter:  67%|██████▋   | 73000/109070 [00:15<00:07, 4736.82 examples/s]
Filter:  68%|██████▊   | 74000/109070 [00:15<00:07, 4722.19 examples/s]
Filter:  69%|██████▉   | 75000/109070 [00:15<00:07, 4715.64 examples/s]
Filter:  70%|██████▉   | 76000/109070 [00:15<00:06, 4764.80 examples/s]
Filter:  71%|███████   | 77000/109070 [00:16<00:06, 4824.62 examples/s]
Filter:  72%|███████▏  | 78000/109070 [00:16<00:06, 4826.07 examples/s]
Filter:  72%|███████▏  | 79000/109070 [00:16<00:09, 3288.51 examples/s]
Filter:  73%|███████▎  | 80000/109070 [00:17<00:08, 3549.02 examples/s]
Filter:  74%|███████▍  | 81000/109070 [00:17<00:07, 3782.97 examples/s]
Filter:  75%|███████▌  | 82000/109070 [00:17<00:07, 3478.82 examples/s]
Filter:  76%|███████▌  | 83000/109070 [00:17<00:07, 3472.19 examples/s]
Filter:  77%|███████▋  | 84000/109070 [00:18<00:06, 3689.27 examples/s]
Filter:  78%|███████▊  | 85000/109070 [00:18<00:06, 3939.70 examples/s]
Filter:  79%|███████▉  | 86000/109070 [00:18<00:05, 3934.36 examples/s]
Filter:  80%|███████▉  | 87000/109070 [00:18<00:05, 4114.85 examples/s]
Filter:  81%|████████  | 88000/109070 [00:19<00:04, 4336.26 examples/s]
Filter:  82%|████████▏ | 89000/109070 [00:19<00:04, 4474.74 examples/s]
Filter:  83%|████████▎ | 90000/109070 [00:19<00:04, 4599.06 examples/s]
Filter:  83%|████████▎ | 91000/109070 [00:19<00:03, 4546.62 examples/s]
Filter:  84%|████████▍ | 92000/109070 [00:19<00:03, 4575.00 examples/s]
Filter:  85%|████████▌ | 93000/109070 [00:20<00:03, 4589.85 examples/s]
Filter:  86%|████████▌ | 94000/109070 [00:20<00:03, 4168.23 examples/s]
Filter:  87%|████████▋ | 95000/109070 [00:20<00:03, 3602.62 examples/s]
Filter:  88%|████████▊ | 96000/109070 [00:21<00:03, 3679.33 examples/s]
Filter:  89%|████████▉ | 97000/109070 [00:21<00:03, 3916.44 examples/s]
Filter:  90%|████████▉ | 98000/109070 [00:21<00:02, 4074.30 examples/s]
Filter:  91%|█████████ | 99000/109070 [00:21<00:02, 4176.35 examples/s]
Filter:  92%|█████████▏| 100000/109070 [00:21<00:02, 4295.61 examples/s]
Filter:  93%|█████████▎| 101000/109070 [00:22<00:01, 4446.12 examples/s]
Filter:  94%|█████████▎| 102000/109070 [00:22<00:01, 4573.66 examples/s]
Filter:  94%|█████████▍| 103000/109070 [00:22<00:01, 4649.67 examples/s]
Filter:  95%|█████████▌| 104000/109070 [00:22<00:01, 4712.50 examples/s]
Filter:  96%|█████████▋| 105000/109070 [00:22<00:00, 4693.81 examples/s]
Filter:  97%|█████████▋| 106000/109070 [00:23<00:00, 4671.79 examples/s]
Filter:  98%|█████████▊| 107000/109070 [00:23<00:00, 4661.43 examples/s]
Filter:  99%|█████████▉| 108000/109070 [00:23<00:00, 4694.42 examples/s]
Filter: 100%|█████████▉| 109000/109070 [00:23<00:00, 4769.58 examples/s]
Filter: 100%|██████████| 109070/109070 [00:23<00:00, 4574.27 examples/s]

Fetching 128 files:   0%|          | 0/128 [00:00<?, ?it/s]
Fetching 128 files: 100%|██████████| 128/128 [00:00<00:00, 9956.80it/s]
`torch_dtype` is deprecated! Use `dtype` instead!

Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]
Loading checkpoint shards:  33%|███▎      | 1/3 [00:00<00:01,  1.19it/s]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:01<00:00,  1.20it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.25it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.23it/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-12 15:58:06 [utils.py:233] non-default args: {'seed': 0, 'max_model_len': 1536, 'distributed_executor_backend': 'external_launcher', 'gpu_memory_utilization': 0.5, 'max_num_batched_tokens': 4096, 'max_num_seqs': 8, 'disable_log_stats': True, 'model_impl': 'vllm', 'model': 'Navid-AI/Yehia-7B-preview'}
INFO 04-12 15:58:06 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-12 15:58:06 [model.py:1510] Using max model len 1536
INFO 04-12 15:58:07 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-12 15:58:07 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-12 15:58:07 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-12 15:58:07 [core.py:77] Initializing a V1 LLM engine (v0.11.0) with config: model='Navid-AI/Yehia-7B-preview', speculative_config=None, tokenizer='Navid-AI/Yehia-7B-preview', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=1536, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Navid-AI/Yehia-7B-preview, enable_prefix_caching=True, chunked_prefill_enabled=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2","vllm.mamba_mixer","vllm.short_conv","vllm.linear_attention","vllm.plamo2_mamba_mixer","vllm.gdn_attention","vllm.sparse_attn_indexer"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"cudagraph_mode":[2,1],"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"use_inductor_graph_partition":false,"pass_config":{},"max_capture_size":16,"local_cache_dir":null}
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
INFO 04-12 15:58:08 [parallel_state.py:1208] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
WARNING 04-12 15:58:08 [topk_topp_sampler.py:66] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
INFO 04-12 15:58:08 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-12 15:58:08 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-12 15:58:09 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-12 15:58:09 [weight_utils.py:392] Using model weights format ['*.safetensors']

Loading safetensors checkpoint shards:   0% Completed | 0/3 [00:00<?, ?it/s]

Loading safetensors checkpoint shards:  33% Completed | 1/3 [00:00<00:01,  1.95it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.59it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.48it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.53it/s]

INFO 04-12 15:58:11 [default_loader.py:267] Loading weights took 2.15 seconds
INFO 04-12 15:58:12 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.609677 seconds
INFO 04-12 15:58:17 [backends.py:548] Using cache directory: /root/.cache/vllm/torch_compile_cache/8b012be8b8/rank_0_0/backbone for vLLM's torch.compile
INFO 04-12 15:58:17 [backends.py:559] Dynamo bytecode transform time: 5.47 s
INFO 04-12 15:58:21 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.634 s
INFO 04-12 15:58:22 [monitor.py:34] torch.compile takes 5.47 s in total
INFO 04-12 15:58:23 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-12 15:58:24 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-12 15:58:24 [kv_cache_utils.py:1091] Maximum concurrency for 1,536 tokens per request: 34.80x

Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):   0%|          | 0/5 [00:00<?, ?it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  40%|████      | 2/5 [00:00<00:00, 19.36it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 17.35it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 17.55it/s]

Capturing CUDA graphs (decode, FULL):   0%|          | 0/4 [00:00<?, ?it/s]
Capturing CUDA graphs (decode, FULL):  75%|███████▌  | 3/4 [00:00<00:00, 20.88it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 21.82it/s]
INFO 04-12 15:58:25 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-12 15:58:25 [core.py:210] init engine (profile, create kv cache, warmup model) took 13.18 seconds
INFO 04-12 15:58:25 [llm.py:306] Supported_tasks: ('generate',)
2026-04-12 15:58:25,741 | INFO | train_grpo_train | dataset_id=Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1 train_size=19001 eval_size=104
2026-04-12 15:58:25,741 | INFO | train_grpo_train | source_dataset_id=Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits test_size=204 hard_diagnostic_size=3126
2026-04-12 15:58:25,742 | INFO | train_grpo_train | active_rewards=['meter', 'count_adherence', 'arabic_clean', 'hard_gate', 'arabic_floor', 'count_floor', 'lexical_plausibility', 'judge_quality', 'repeat_penalty', 'repeat_floor', 'near_duplicate_penalty', 'opening_diversity', 'distinct_2', 'total_composite']
2026-04-12 15:58:25,742 | INFO | train_grpo_train | reward_weights={'meter': 0.0, 'count_adherence': 0.0, 'arabic_clean': 0.0, 'hard_gate': 0.0, 'arabic_floor': 0.0, 'count_floor': 0.0, 'lexical_plausibility': 0.0, 'judge_quality': 0.0, 'repeat_penalty': 0.0, 'repeat_floor': 0.0, 'near_duplicate_penalty': 0.0, 'opening_diversity': 0.0, 'distinct_2': 0.0, 'total_composite': 1.0}
2026-04-12 15:58:25,742 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo-friend-v1-easyfirst
2026-04-12 15:58:25,742 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_drop_trio/cap_3000/selected_manifest.csv
2026-04-12 15:58:25,742 | INFO | train_grpo_train | hard_diagnostic_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_drop_trio/hard_diagnostic_cap_256/selected_manifest.csv
2026-04-12 15:58:25,742 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-12 15:58:25,742 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-12T15:58:06Z", "resume_mode": "fresh", "requested_spec": "", "result": "fresh", "reason": "explicit_fresh", "local_resume_path": null, "remote_repo": null, "remote_prefix": null, "compatibility": "unchecked", "config_fingerprint": "429b334e6c80dfa9b81f2594e6165bdb27e9289404e07f5b1493fe2fe6227203"}
2026-04-12 15:58:25,742 | INFO | train_grpo_train | starting trainer.train()
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'pad_token_id': 2}.

  0%|          | 0/3300 [00:00<?, ?it/s]INFO 04-12 15:58:26 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  0%|          | 1/3300 [00:09<8:32:09,  9.31s/it]2026-04-12 15:58:35,463 | INFO | train_grpo_train | metrics_logged mode=train step=1

                                                  
{'loss': 0.1988, 'grad_norm': 20.09337043762207, 'learning_rate': 1e-05, 'num_tokens': 1556.0, 'completions/mean_length': 32.5, 'completions/min_length': 20.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.653188943862915, 'rewards/meter/std': 0.44728460907936096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9054144620895386, 'rewards/lexical_plausibility/std': 0.1331765353679657, 'rewards/judge_quality/mean': 0.27125000953674316, 'rewards/judge_quality/std': 0.2713162302970886, 'rewards/repeat_penalty/mean': 0.7492251992225647, 'rewards/repeat_penalty/std': 0.002191446488723159, 'rewards/repeat_floor/mean': 0.9848450422286987, 'rewards/repeat_floor/std': 0.0004382850893307477, 'rewards/near_duplicate_penalty/mean': 0.9989669322967529, 'rewards/near_duplicate_penalty/std': 0.002921935636550188, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2107067108154297, 'rewards/total_composite/std': 0.28980186581611633, 'reward': 0.2107067108154297, 'reward_std': 0.28980186581611633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22281700372695923, 'sampling/sampling_logp_difference/max': 2.2011396884918213, 'sampling/importance_sampling_ratio/min': 0.11067695170640945, 'sampling/importance_sampling_ratio/mean': 0.9999842643737793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5754356235265732, 'clip_ratio/low_mean': 0.1252038492821157, 'clip_ratio/low_min': 0.1252038492821157, 'clip_ratio/high_mean': 0.08266128972172737, 'clip_ratio/high_max': 0.08266128972172737, 'clip_ratio/region_mean': 0.20786513900384307, 'reward_total_mean': 0.2107067108154297, 'reward_meter_mean': 0.653188943862915, 'reward_meter_std': 0.44728460907936096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9054144620895386, 'reward_lexical_plausibility_std': 0.1331765353679657, 'reward_repeat_penalty_mean': 0.7492251992225647, 'reward_repeat_penalty_std': 0.002191446488723159, 'reward_repeat_floor_mean': 0.9848450422286987, 'reward_repeat_floor_std': 0.0004382850893307477, 'reward_near_duplicate_penalty_mean': 0.9989669322967529, 'reward_near_duplicate_penalty_std': 0.002921935636550188, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.27125000953674316, 'reward_judge_quality_std': 0.2713162302970886, 'reward_total_composite_mean': 0.2107067108154297, 'reward_total_composite_std': 0.28980186581611633, 'epoch': 0.0}

  0%|          | 1/3300 [00:09<8:32:09,  9.31s/it]INFO 04-12 15:58:35 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 2/3300 [00:20<9:40:53, 10.57s/it]2026-04-12 15:58:46,910 | INFO | train_grpo_train | metrics_logged mode=train step=2

                                                  
{'loss': 0.0541, 'grad_norm': 9.90222454071045, 'learning_rate': 9.996969696969698e-06, 'num_tokens': 4696.0, 'completions/mean_length': 202.5, 'completions/min_length': 195.0, 'completions/max_length': 212.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 202.5, 'completions/min_terminated_length': 195.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.40368834137916565, 'rewards/meter/std': 0.27416813373565674, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.017251653596758842, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9722908735275269, 'rewards/repeat_penalty/std': 0.023187976330518723, 'rewards/repeat_floor/mean': 0.996211051940918, 'rewards/repeat_floor/std': 0.0025169842410832644, 'rewards/near_duplicate_penalty/mean': 0.9787192344665527, 'rewards/near_duplicate_penalty/std': 0.008377094753086567, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.14162755012512207, 'rewards/total_composite/std': 0.09327300637960434, 'reward': 0.14162755012512207, 'reward_std': 0.09327300637960434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18262526392936707, 'sampling/sampling_logp_difference/max': 3.432194709777832, 'sampling/importance_sampling_ratio/min': 0.0323159396648407, 'sampling/importance_sampling_ratio/mean': 0.9920792579650879, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.697647899389267, 'clip_ratio/low_mean': 0.0823319898918271, 'clip_ratio/low_min': 0.0823319898918271, 'clip_ratio/high_mean': 0.06837169267237186, 'clip_ratio/high_max': 0.06837169267237186, 'clip_ratio/region_mean': 0.15070368256419897, 'reward_total_mean': 0.14162755012512207, 'reward_meter_mean': 0.40368834137916565, 'reward_meter_std': 0.27416813373565674, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.017251653596758842, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9722908735275269, 'reward_repeat_penalty_std': 0.023187976330518723, 'reward_repeat_floor_mean': 0.996211051940918, 'reward_repeat_floor_std': 0.0025169842410832644, 'reward_near_duplicate_penalty_mean': 0.9787192344665527, 'reward_near_duplicate_penalty_std': 0.008377094753086567, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.14162755012512207, 'reward_total_composite_std': 0.09327300637960434, 'epoch': 0.0}

  0%|          | 2/3300 [00:20<9:40:53, 10.57s/it]INFO 04-12 15:58:47 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 3/3300 [00:31<9:34:50, 10.46s/it]2026-04-12 15:58:57,241 | INFO | train_grpo_train | metrics_logged mode=train step=3

                                                  
{'loss': 0.0513, 'grad_norm': 12.573843002319336, 'learning_rate': 9.993939393939395e-06, 'num_tokens': 6843.0, 'completions/mean_length': 84.375, 'completions/min_length': 64.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.6097227334976196, 'rewards/meter/std': 0.37561020255088806, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9436274766921997, 'rewards/lexical_plausibility/std': 0.10242299735546112, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9936990737915039, 'rewards/repeat_penalty/std': 0.003470573341473937, 'rewards/repeat_floor/mean': 0.9990548491477966, 'rewards/repeat_floor/std': 0.0005205817287787795, 'rewards/near_duplicate_penalty/mean': 0.9936990737915039, 'rewards/near_duplicate_penalty/std': 0.003470573341473937, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1688469648361206, 'rewards/total_composite/std': 0.12597337365150452, 'reward': 0.1688469648361206, 'reward_std': 0.12597337365150452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2326127141714096, 'sampling/sampling_logp_difference/max': 2.259091377258301, 'sampling/importance_sampling_ratio/min': 0.10444534569978714, 'sampling/importance_sampling_ratio/mean': 1.0231131315231323, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9617455154657364, 'clip_ratio/low_mean': 0.126793647184968, 'clip_ratio/low_min': 0.126793647184968, 'clip_ratio/high_mean': 0.0766318216919899, 'clip_ratio/high_max': 0.0766318216919899, 'clip_ratio/region_mean': 0.2034254688769579, 'reward_total_mean': 0.1688469648361206, 'reward_meter_mean': 0.6097227334976196, 'reward_meter_std': 0.37561020255088806, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9436274766921997, 'reward_lexical_plausibility_std': 0.10242299735546112, 'reward_repeat_penalty_mean': 0.9936990737915039, 'reward_repeat_penalty_std': 0.003470573341473937, 'reward_repeat_floor_mean': 0.9990548491477966, 'reward_repeat_floor_std': 0.0005205817287787795, 'reward_near_duplicate_penalty_mean': 0.9936990737915039, 'reward_near_duplicate_penalty_std': 0.003470573341473937, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1688469648361206, 'reward_total_composite_std': 0.12597337365150452, 'epoch': 0.0}

  0%|          | 3/3300 [00:31<9:34:50, 10.46s/it]INFO 04-12 15:58:57 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 4/3300 [00:40<9:21:15, 10.22s/it]2026-04-12 15:59:07,085 | INFO | train_grpo_train | metrics_logged mode=train step=4

                                                  
{'loss': 0.0515, 'grad_norm': 14.311335563659668, 'learning_rate': 9.990909090909093e-06, 'num_tokens': 8345.0, 'completions/mean_length': 46.75, 'completions/min_length': 29.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6345551013946533, 'rewards/meter/std': 0.4597727954387665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9836230278015137, 'rewards/lexical_plausibility/std': 0.04632115736603737, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.9836134910583496, 'rewards/repeat_penalty/std': 0.02421470545232296, 'rewards/repeat_floor/mean': 0.9975420236587524, 'rewards/repeat_floor/std': 0.0036322178784757853, 'rewards/near_duplicate_penalty/mean': 0.9836134910583496, 'rewards/near_duplicate_penalty/std': 0.02421470545232296, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31212854385375977, 'rewards/total_composite/std': 0.28672724962234497, 'reward': 0.31212854385375977, 'reward_std': 0.28672724962234497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23379011452198029, 'sampling/sampling_logp_difference/max': 1.7445611953735352, 'sampling/importance_sampling_ratio/min': 0.1747216433286667, 'sampling/importance_sampling_ratio/mean': 1.0515234470367432, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7860644608736038, 'clip_ratio/low_mean': 0.10407150536775589, 'clip_ratio/low_min': 0.10407150536775589, 'clip_ratio/high_mean': 0.12573344260454178, 'clip_ratio/high_max': 0.12573344260454178, 'clip_ratio/region_mean': 0.22980494797229767, 'reward_total_mean': 0.31212854385375977, 'reward_meter_mean': 0.6345551013946533, 'reward_meter_std': 0.4597727954387665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9836230278015137, 'reward_lexical_plausibility_std': 0.04632115736603737, 'reward_repeat_penalty_mean': 0.9836134910583496, 'reward_repeat_penalty_std': 0.02421470545232296, 'reward_repeat_floor_mean': 0.9975420236587524, 'reward_repeat_floor_std': 0.0036322178784757853, 'reward_near_duplicate_penalty_mean': 0.9836134910583496, 'reward_near_duplicate_penalty_std': 0.02421470545232296, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.31212854385375977, 'reward_total_composite_std': 0.28672724962234497, 'epoch': 0.0}

  0%|          | 4/3300 [00:40<9:21:15, 10.22s/it]INFO 04-12 15:59:07 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 5/3300 [00:50<8:58:14,  9.80s/it]2026-04-12 15:59:16,150 | INFO | train_grpo_train | metrics_logged mode=train step=5

                                                  
{'loss': 0.0456, 'grad_norm': 11.538812637329102, 'learning_rate': 9.987878787878788e-06, 'num_tokens': 10140.0, 'completions/mean_length': 64.375, 'completions/min_length': 47.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.7967674732208252, 'rewards/meter/std': 0.34278610348701477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9899553656578064, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.36625000834465027, 'rewards/judge_quality/std': 0.22639645636081696, 'rewards/repeat_penalty/mean': 0.999139130115509, 'rewards/repeat_penalty/std': 0.002434949856251478, 'rewards/repeat_floor/mean': 0.9998708963394165, 'rewards/repeat_floor/std': 0.0003652445739135146, 'rewards/near_duplicate_penalty/mean': 0.999139130115509, 'rewards/near_duplicate_penalty/std': 0.002434949856251478, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3292407989501953, 'rewards/total_composite/std': 0.25814321637153625, 'reward': 0.3292407989501953, 'reward_std': 0.25814321637153625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18061591684818268, 'sampling/sampling_logp_difference/max': 1.6605291366577148, 'sampling/importance_sampling_ratio/min': 0.19003839790821075, 'sampling/importance_sampling_ratio/mean': 1.0090373754501343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8136060386896133, 'clip_ratio/low_mean': 0.08147293515503407, 'clip_ratio/low_min': 0.08147293515503407, 'clip_ratio/high_mean': 0.1212398149073124, 'clip_ratio/high_max': 0.1212398149073124, 'clip_ratio/region_mean': 0.20271275006234646, 'reward_total_mean': 0.3292407989501953, 'reward_meter_mean': 0.7967674732208252, 'reward_meter_std': 0.34278610348701477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9899553656578064, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.999139130115509, 'reward_repeat_penalty_std': 0.002434949856251478, 'reward_repeat_floor_mean': 0.9998708963394165, 'reward_repeat_floor_std': 0.0003652445739135146, 'reward_near_duplicate_penalty_mean': 0.999139130115509, 'reward_near_duplicate_penalty_std': 0.002434949856251478, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36625000834465027, 'reward_judge_quality_std': 0.22639645636081696, 'reward_total_composite_mean': 0.3292407989501953, 'reward_total_composite_std': 0.25814321637153625, 'epoch': 0.0}

  0%|          | 5/3300 [00:50<8:58:14,  9.80s/it]INFO 04-12 15:59:16 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 6/3300 [00:59<8:55:01,  9.75s/it]2026-04-12 15:59:25,788 | INFO | train_grpo_train | metrics_logged mode=train step=6

                                                  
{'loss': 0.0275, 'grad_norm': 12.958381652832031, 'learning_rate': 9.984848484848485e-06, 'num_tokens': 11899.0, 'completions/mean_length': 58.875, 'completions/min_length': 40.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5878450274467468, 'rewards/meter/std': 0.46913498640060425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9604833126068115, 'rewards/lexical_plausibility/std': 0.07437513023614883, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2639264166355133, 'rewards/repeat_penalty/mean': 0.9688584804534912, 'rewards/repeat_penalty/std': 0.0629393607378006, 'rewards/repeat_floor/mean': 0.9958750009536743, 'rewards/repeat_floor/std': 0.00793202593922615, 'rewards/near_duplicate_penalty/mean': 0.9797113537788391, 'rewards/near_duplicate_penalty/std': 0.033393166959285736, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9879807829856873, 'rewards/distinct_2/std': 0.03399552404880524, 'rewards/total_composite/mean': 0.3447781205177307, 'rewards/total_composite/std': 0.37308818101882935, 'reward': 0.3447781205177307, 'reward_std': 0.37308821082115173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21205370128154755, 'sampling/sampling_logp_difference/max': 2.3066372871398926, 'sampling/importance_sampling_ratio/min': 0.09959560632705688, 'sampling/importance_sampling_ratio/mean': 1.027489423751831, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.516320526599884, 'clip_ratio/low_mean': 0.12161067128181458, 'clip_ratio/low_min': 0.12161067128181458, 'clip_ratio/high_mean': 0.07525585032999516, 'clip_ratio/high_max': 0.07525585032999516, 'clip_ratio/region_mean': 0.19686652161180973, 'reward_total_mean': 0.3447781205177307, 'reward_meter_mean': 0.5878450274467468, 'reward_meter_std': 0.46913498640060425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9604833126068115, 'reward_lexical_plausibility_std': 0.07437513023614883, 'reward_repeat_penalty_mean': 0.9688584804534912, 'reward_repeat_penalty_std': 0.0629393607378006, 'reward_repeat_floor_mean': 0.9958750009536743, 'reward_repeat_floor_std': 0.00793202593922615, 'reward_near_duplicate_penalty_mean': 0.9797113537788391, 'reward_near_duplicate_penalty_std': 0.033393166959285736, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9879807829856873, 'reward_distinct_2_std': 0.03399552404880524, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2639264166355133, 'reward_total_composite_mean': 0.3447781205177307, 'reward_total_composite_std': 0.37308818101882935, 'epoch': 0.0}

  0%|          | 6/3300 [00:59<8:55:01,  9.75s/it]INFO 04-12 15:59:26 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 7/3300 [01:08<8:40:12,  9.48s/it]2026-04-12 15:59:34,715 | INFO | train_grpo_train | metrics_logged mode=train step=7

                                                  
{'loss': 0.2477, 'grad_norm': 14.628622055053711, 'learning_rate': 9.981818181818183e-06, 'num_tokens': 13575.0, 'completions/mean_length': 50.5, 'completions/min_length': 32.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6770910024642944, 'rewards/meter/std': 0.374254435300827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.990183413028717, 'rewards/repeat_penalty/std': 0.009771591052412987, 'rewards/repeat_floor/mean': 0.9985275268554688, 'rewards/repeat_floor/std': 0.0014657394494861364, 'rewards/near_duplicate_penalty/mean': 0.990183413028717, 'rewards/near_duplicate_penalty/std': 0.009771591052412987, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32056713104248047, 'rewards/total_composite/std': 0.2870679199695587, 'reward': 0.32056713104248047, 'reward_std': 0.2870679199695587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1809655725955963, 'sampling/sampling_logp_difference/max': 1.6068577766418457, 'sampling/importance_sampling_ratio/min': 0.2005166858434677, 'sampling/importance_sampling_ratio/mean': 1.040963888168335, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2859416902065277, 'clip_ratio/low_mean': 0.08865558542311192, 'clip_ratio/low_min': 0.08865558542311192, 'clip_ratio/high_mean': 0.04387405142188072, 'clip_ratio/high_max': 0.04387405142188072, 'clip_ratio/region_mean': 0.13252963684499264, 'reward_total_mean': 0.32056713104248047, 'reward_meter_mean': 0.6770910024642944, 'reward_meter_std': 0.374254435300827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.990183413028717, 'reward_repeat_penalty_std': 0.009771591052412987, 'reward_repeat_floor_mean': 0.9985275268554688, 'reward_repeat_floor_std': 0.0014657394494861364, 'reward_near_duplicate_penalty_mean': 0.990183413028717, 'reward_near_duplicate_penalty_std': 0.009771591052412987, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.32056713104248047, 'reward_total_composite_std': 0.2870679199695587, 'epoch': 0.0}

  0%|          | 7/3300 [01:08<8:40:12,  9.48s/it]INFO 04-12 15:59:34 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 8/3300 [01:17<8:33:04,  9.35s/it]2026-04-12 15:59:43,794 | INFO | train_grpo_train | metrics_logged mode=train step=8

                                                  
{'loss': -0.0288, 'grad_norm': 19.75635528564453, 'learning_rate': 9.97878787878788e-06, 'num_tokens': 15216.0, 'completions/mean_length': 41.125, 'completions/min_length': 29.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7961204648017883, 'rewards/meter/std': 0.20362447202205658, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9633767008781433, 'rewards/lexical_plausibility/std': 0.1035863608121872, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9952322244644165, 'rewards/repeat_penalty/std': 0.0048536453396081924, 'rewards/repeat_floor/mean': 0.9992848634719849, 'rewards/repeat_floor/std': 0.0007280543795786798, 'rewards/near_duplicate_penalty/mean': 0.9952322244644165, 'rewards/near_duplicate_penalty/std': 0.0048536453396081924, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.574169933795929, 'rewards/total_composite/std': 0.3322398364543915, 'reward': 0.574169933795929, 'reward_std': 0.3322398364543915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23442627489566803, 'sampling/sampling_logp_difference/max': 1.3025455474853516, 'sampling/importance_sampling_ratio/min': 0.2718389332294464, 'sampling/importance_sampling_ratio/mean': 1.0258030891418457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0661709010601044, 'clip_ratio/low_mean': 0.12354860082268715, 'clip_ratio/low_min': 0.12354860082268715, 'clip_ratio/high_mean': 0.1104191280901432, 'clip_ratio/high_max': 0.1104191280901432, 'clip_ratio/region_mean': 0.23396772891283035, 'reward_total_mean': 0.574169933795929, 'reward_meter_mean': 0.7961204648017883, 'reward_meter_std': 0.20362447202205658, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9633767008781433, 'reward_lexical_plausibility_std': 0.1035863608121872, 'reward_repeat_penalty_mean': 0.9952322244644165, 'reward_repeat_penalty_std': 0.0048536453396081924, 'reward_repeat_floor_mean': 0.9992848634719849, 'reward_repeat_floor_std': 0.0007280543795786798, 'reward_near_duplicate_penalty_mean': 0.9952322244644165, 'reward_near_duplicate_penalty_std': 0.0048536453396081924, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.574169933795929, 'reward_total_composite_std': 0.3322398364543915, 'epoch': 0.0}

  0%|          | 8/3300 [01:17<8:33:04,  9.35s/it]INFO 04-12 15:59:44 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 9/3300 [01:27<8:38:49,  9.46s/it]2026-04-12 15:59:53,490 | INFO | train_grpo_train | metrics_logged mode=train step=9

                                                  
{'loss': 0.1489, 'grad_norm': 13.455046653747559, 'learning_rate': 9.975757575757577e-06, 'num_tokens': 16927.0, 'completions/mean_length': 52.875, 'completions/min_length': 31.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9079039096832275, 'rewards/meter/std': 0.11305537074804306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4587499797344208, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9866733551025391, 'rewards/repeat_penalty/std': 0.014050498604774475, 'rewards/repeat_floor/mean': 0.998000979423523, 'rewards/repeat_floor/std': 0.0021075811237096786, 'rewards/near_duplicate_penalty/mean': 0.9866733551025391, 'rewards/near_duplicate_penalty/std': 0.014050498604774475, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.410701185464859, 'rewards/total_composite/std': 0.15337826311588287, 'reward': 0.410701185464859, 'reward_std': 0.15337826311588287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20689186453819275, 'sampling/sampling_logp_difference/max': 1.2887544631958008, 'sampling/importance_sampling_ratio/min': 0.27561384439468384, 'sampling/importance_sampling_ratio/mean': 1.0080140829086304, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.917743593454361, 'clip_ratio/low_mean': 0.1186765506863594, 'clip_ratio/low_min': 0.1186765506863594, 'clip_ratio/high_mean': 0.09586916491389275, 'clip_ratio/high_max': 0.09586916491389275, 'clip_ratio/region_mean': 0.21454571560025215, 'reward_total_mean': 0.410701185464859, 'reward_meter_mean': 0.9079039096832275, 'reward_meter_std': 0.11305537074804306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9866733551025391, 'reward_repeat_penalty_std': 0.014050498604774475, 'reward_repeat_floor_mean': 0.998000979423523, 'reward_repeat_floor_std': 0.0021075811237096786, 'reward_near_duplicate_penalty_mean': 0.9866733551025391, 'reward_near_duplicate_penalty_std': 0.014050498604774475, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4587499797344208, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.410701185464859, 'reward_total_composite_std': 0.15337826311588287, 'epoch': 0.0}

  0%|          | 9/3300 [01:27<8:38:49,  9.46s/it]INFO 04-12 15:59:53 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 10/3300 [01:37<8:55:01,  9.76s/it]2026-04-12 16:00:03,916 | INFO | train_grpo_train | metrics_logged mode=train step=10

                                                   
{'loss': -0.0781, 'grad_norm': 15.111937522888184, 'learning_rate': 9.972727272727274e-06, 'num_tokens': 18909.0, 'completions/mean_length': 67.75, 'completions/min_length': 49.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.5851961970329285, 'rewards/meter/std': 0.31481993198394775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9904689788818359, 'rewards/lexical_plausibility/std': 0.02695779874920845, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9864856004714966, 'rewards/repeat_penalty/std': 0.01194615289568901, 'rewards/repeat_floor/mean': 0.997972846031189, 'rewards/repeat_floor/std': 0.0017919203964993358, 'rewards/near_duplicate_penalty/mean': 0.9864856004714966, 'rewards/near_duplicate_penalty/std': 0.01194615289568901, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24945786595344543, 'rewards/total_composite/std': 0.14811070263385773, 'reward': 0.24945786595344543, 'reward_std': 0.14811071753501892, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22759608924388885, 'sampling/sampling_logp_difference/max': 2.27805233001709, 'sampling/importance_sampling_ratio/min': 0.10248362272977829, 'sampling/importance_sampling_ratio/mean': 1.0295257568359375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0101440995931625, 'clip_ratio/low_mean': 0.13846822828054428, 'clip_ratio/low_min': 0.13846822828054428, 'clip_ratio/high_mean': 0.08745039813220501, 'clip_ratio/high_max': 0.08745039813220501, 'clip_ratio/region_mean': 0.2259186264127493, 'reward_total_mean': 0.24945786595344543, 'reward_meter_mean': 0.5851961970329285, 'reward_meter_std': 0.31481993198394775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9904689788818359, 'reward_lexical_plausibility_std': 0.02695779874920845, 'reward_repeat_penalty_mean': 0.9864856004714966, 'reward_repeat_penalty_std': 0.01194615289568901, 'reward_repeat_floor_mean': 0.997972846031189, 'reward_repeat_floor_std': 0.0017919203964993358, 'reward_near_duplicate_penalty_mean': 0.9864856004714966, 'reward_near_duplicate_penalty_std': 0.01194615289568901, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.24945786595344543, 'reward_total_composite_std': 0.14811070263385773, 'epoch': 0.0}

  0%|          | 10/3300 [01:37<8:55:01,  9.76s/it]INFO 04-12 16:00:04 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 11/3300 [01:47<8:52:48,  9.72s/it]2026-04-12 16:00:13,554 | INFO | train_grpo_train | metrics_logged mode=train step=11

                                                   
{'loss': 0.1132, 'grad_norm': 20.49909210205078, 'learning_rate': 9.96969696969697e-06, 'num_tokens': 20513.0, 'completions/mean_length': 42.5, 'completions/min_length': 22.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.3109018802642822, 'rewards/meter/std': 0.37683284282684326, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.8933573961257935, 'rewards/lexical_plausibility/std': 0.15487238764762878, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136148393154144, 'rewards/repeat_penalty/mean': 0.9677474498748779, 'rewards/repeat_penalty/std': 0.026140382513403893, 'rewards/repeat_floor/mean': 0.9951621294021606, 'rewards/repeat_floor/std': 0.003921049647033215, 'rewards/near_duplicate_penalty/mean': 0.9677474498748779, 'rewards/near_duplicate_penalty/std': 0.026140382513403893, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10126905143260956, 'rewards/total_composite/std': 0.11812613904476166, 'reward': 0.10126905143260956, 'reward_std': 0.11812613904476166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21175774931907654, 'sampling/sampling_logp_difference/max': 1.951096534729004, 'sampling/importance_sampling_ratio/min': 0.1421181559562683, 'sampling/importance_sampling_ratio/mean': 1.0338985919952393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5531779527664185, 'clip_ratio/low_mean': 0.13862179778516293, 'clip_ratio/low_min': 0.13862179778516293, 'clip_ratio/high_mean': 0.08591545931994915, 'clip_ratio/high_max': 0.08591545931994915, 'clip_ratio/region_mean': 0.22453725710511208, 'reward_total_mean': 0.10126905143260956, 'reward_meter_mean': 0.3109018802642822, 'reward_meter_std': 0.37683284282684326, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.8933573961257935, 'reward_lexical_plausibility_std': 0.15487238764762878, 'reward_repeat_penalty_mean': 0.9677474498748779, 'reward_repeat_penalty_std': 0.026140382513403893, 'reward_repeat_floor_mean': 0.9951621294021606, 'reward_repeat_floor_std': 0.003921049647033215, 'reward_near_duplicate_penalty_mean': 0.9677474498748779, 'reward_near_duplicate_penalty_std': 0.026140382513403893, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136148393154144, 'reward_total_composite_mean': 0.10126905143260956, 'reward_total_composite_std': 0.11812613904476166, 'epoch': 0.0}

  0%|          | 11/3300 [01:47<8:52:48,  9.72s/it]INFO 04-12 16:00:13 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 12/3300 [01:54<8:16:30,  9.06s/it]2026-04-12 16:00:21,102 | INFO | train_grpo_train | metrics_logged mode=train step=12

                                                   
{'loss': 0.1082, 'grad_norm': 28.53273582458496, 'learning_rate': 9.966666666666667e-06, 'num_tokens': 22129.0, 'completions/mean_length': 42.0, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6041836738586426, 'rewards/meter/std': 0.42832210659980774, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8725000023841858, 'rewards/judge_quality/std': 0.1713601052761078, 'rewards/repeat_penalty/mean': 0.999159038066864, 'rewards/repeat_penalty/std': 0.0021877859253436327, 'rewards/repeat_floor/mean': 0.9998738765716553, 'rewards/repeat_floor/std': 0.0003281752287875861, 'rewards/near_duplicate_penalty/mean': 0.999159038066864, 'rewards/near_duplicate_penalty/std': 0.0021877859253436327, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5607736110687256, 'rewards/total_composite/std': 0.41370680928230286, 'reward': 0.5607736110687256, 'reward_std': 0.41370677947998047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16803176701068878, 'sampling/sampling_logp_difference/max': 2.5533061027526855, 'sampling/importance_sampling_ratio/min': 0.0778239443898201, 'sampling/importance_sampling_ratio/mean': 1.0045020580291748, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5017555598169565, 'clip_ratio/low_mean': 0.10567478649318218, 'clip_ratio/low_min': 0.10567478649318218, 'clip_ratio/high_mean': 0.03419225220568478, 'clip_ratio/high_max': 0.03419225220568478, 'clip_ratio/region_mean': 0.13986703869886696, 'reward_total_mean': 0.5607736110687256, 'reward_meter_mean': 0.6041836738586426, 'reward_meter_std': 0.42832210659980774, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.999159038066864, 'reward_repeat_penalty_std': 0.0021877859253436327, 'reward_repeat_floor_mean': 0.9998738765716553, 'reward_repeat_floor_std': 0.0003281752287875861, 'reward_near_duplicate_penalty_mean': 0.999159038066864, 'reward_near_duplicate_penalty_std': 0.0021877859253436327, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8725000023841858, 'reward_judge_quality_std': 0.1713601052761078, 'reward_total_composite_mean': 0.5607736110687256, 'reward_total_composite_std': 0.41370680928230286, 'epoch': 0.0}

  0%|          | 12/3300 [01:54<8:16:30,  9.06s/it]INFO 04-12 16:00:21 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 13/3300 [02:04<8:18:17,  9.10s/it]2026-04-12 16:00:30,280 | INFO | train_grpo_train | metrics_logged mode=train step=13

                                                   
{'loss': 0.1447, 'grad_norm': 19.016687393188477, 'learning_rate': 9.963636363636364e-06, 'num_tokens': 23838.0, 'completions/mean_length': 49.625, 'completions/min_length': 33.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5367907285690308, 'rewards/meter/std': 0.3970416188240051, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.9804203510284424, 'rewards/repeat_penalty/std': 0.053552690893411636, 'rewards/repeat_floor/mean': 0.9976269006729126, 'rewards/repeat_floor/std': 0.006438884884119034, 'rewards/near_duplicate_penalty/mean': 0.9909669160842896, 'rewards/near_duplicate_penalty/std': 0.023752670735120773, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.23786820471286774, 'rewards/total_composite/std': 0.1558588743209839, 'reward': 0.23786820471286774, 'reward_std': 0.15585888922214508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20183372497558594, 'sampling/sampling_logp_difference/max': 1.2698707580566406, 'sampling/importance_sampling_ratio/min': 0.3338880240917206, 'sampling/importance_sampling_ratio/mean': 1.017527461051941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8827667981386185, 'clip_ratio/low_mean': 0.09016565605998039, 'clip_ratio/low_min': 0.09016565605998039, 'clip_ratio/high_mean': 0.0960752535611391, 'clip_ratio/high_max': 0.0960752535611391, 'clip_ratio/region_mean': 0.1862409096211195, 'reward_total_mean': 0.23786820471286774, 'reward_meter_mean': 0.5367907285690308, 'reward_meter_std': 0.3970416188240051, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9804203510284424, 'reward_repeat_penalty_std': 0.053552690893411636, 'reward_repeat_floor_mean': 0.9976269006729126, 'reward_repeat_floor_std': 0.006438884884119034, 'reward_near_duplicate_penalty_mean': 0.9909669160842896, 'reward_near_duplicate_penalty_std': 0.023752670735120773, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.23786820471286774, 'reward_total_composite_std': 0.1558588743209839, 'epoch': 0.0}

  0%|          | 13/3300 [02:04<8:18:17,  9.10s/it]INFO 04-12 16:00:30 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 14/3300 [02:13<8:16:33,  9.07s/it]2026-04-12 16:00:39,281 | INFO | train_grpo_train | metrics_logged mode=train step=14

                                                   
{'loss': 0.0354, 'grad_norm': 29.63947105407715, 'learning_rate': 9.960606060606062e-06, 'num_tokens': 25210.0, 'completions/mean_length': 24.5, 'completions/min_length': 17.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7364341616630554, 'rewards/meter/std': 0.24760419130325317, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7262499928474426, 'rewards/judge_quality/std': 0.2859039604663849, 'rewards/repeat_penalty/mean': 0.7446338534355164, 'rewards/repeat_penalty/std': 0.015177796594798565, 'rewards/repeat_floor/mean': 0.9839267730712891, 'rewards/repeat_floor/std': 0.0030355635099112988, 'rewards/near_duplicate_penalty/mean': 0.9928451180458069, 'rewards/near_duplicate_penalty/std': 0.020237062126398087, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5186588168144226, 'rewards/total_composite/std': 0.29185327887535095, 'reward': 0.5186588168144226, 'reward_std': 0.29185324907302856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2531023621559143, 'sampling/sampling_logp_difference/max': 1.7571344375610352, 'sampling/importance_sampling_ratio/min': 0.17253857851028442, 'sampling/importance_sampling_ratio/mean': 1.0228863954544067, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8715475499629974, 'clip_ratio/low_mean': 0.07210934348404408, 'clip_ratio/low_min': 0.07210934348404408, 'clip_ratio/high_mean': 0.13814338576048613, 'clip_ratio/high_max': 0.13814338576048613, 'clip_ratio/region_mean': 0.2102527292445302, 'reward_total_mean': 0.5186588168144226, 'reward_meter_mean': 0.7364341616630554, 'reward_meter_std': 0.24760419130325317, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7446338534355164, 'reward_repeat_penalty_std': 0.015177796594798565, 'reward_repeat_floor_mean': 0.9839267730712891, 'reward_repeat_floor_std': 0.0030355635099112988, 'reward_near_duplicate_penalty_mean': 0.9928451180458069, 'reward_near_duplicate_penalty_std': 0.020237062126398087, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7262499928474426, 'reward_judge_quality_std': 0.2859039604663849, 'reward_total_composite_mean': 0.5186588168144226, 'reward_total_composite_std': 0.29185327887535095, 'epoch': 0.0}

  0%|          | 14/3300 [02:13<8:16:33,  9.07s/it]INFO 04-12 16:00:39 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 15/3300 [02:21<8:07:57,  8.91s/it]2026-04-12 16:00:47,835 | INFO | train_grpo_train | metrics_logged mode=train step=15

                                                   
{'loss': 0.0117, 'grad_norm': 12.585134506225586, 'learning_rate': 9.957575757575757e-06, 'num_tokens': 26721.0, 'completions/mean_length': 33.875, 'completions/min_length': 30.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.3934001326560974, 'rewards/meter/std': 0.44469642639160156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9656643867492676, 'rewards/lexical_plausibility/std': 0.07511628419160843, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.3123185336589813, 'rewards/repeat_penalty/mean': 0.7492251992225647, 'rewards/repeat_penalty/std': 0.002191446488723159, 'rewards/repeat_floor/mean': 0.9848450422286987, 'rewards/repeat_floor/std': 0.0004382850893307477, 'rewards/near_duplicate_penalty/mean': 0.9989669322967529, 'rewards/near_duplicate_penalty/std': 0.002921935636550188, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21219727396965027, 'rewards/total_composite/std': 0.29582908749580383, 'reward': 0.21219727396965027, 'reward_std': 0.2958291172981262, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19646772742271423, 'sampling/sampling_logp_difference/max': 1.3255577087402344, 'sampling/importance_sampling_ratio/min': 0.2656547725200653, 'sampling/importance_sampling_ratio/mean': 1.0391937494277954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.652011975646019, 'clip_ratio/low_mean': 0.11778115015476942, 'clip_ratio/low_min': 0.11778115015476942, 'clip_ratio/high_mean': 0.09087481535971165, 'clip_ratio/high_max': 0.09087481535971165, 'clip_ratio/region_mean': 0.20865596551448107, 'reward_total_mean': 0.21219727396965027, 'reward_meter_mean': 0.3934001326560974, 'reward_meter_std': 0.44469642639160156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9656643867492676, 'reward_lexical_plausibility_std': 0.07511628419160843, 'reward_repeat_penalty_mean': 0.7492251992225647, 'reward_repeat_penalty_std': 0.002191446488723159, 'reward_repeat_floor_mean': 0.9848450422286987, 'reward_repeat_floor_std': 0.0004382850893307477, 'reward_near_duplicate_penalty_mean': 0.9989669322967529, 'reward_near_duplicate_penalty_std': 0.002921935636550188, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.3123185336589813, 'reward_total_composite_mean': 0.21219727396965027, 'reward_total_composite_std': 0.29582908749580383, 'epoch': 0.0}

  0%|          | 15/3300 [02:21<8:07:57,  8.91s/it]INFO 04-12 16:00:48 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 16/3300 [02:30<7:58:13,  8.74s/it]2026-04-12 16:00:56,166 | INFO | train_grpo_train | metrics_logged mode=train step=16

                                                   
{'loss': -0.063, 'grad_norm': 23.451108932495117, 'learning_rate': 9.954545454545456e-06, 'num_tokens': 28371.0, 'completions/mean_length': 36.25, 'completions/min_length': 31.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6007693409919739, 'rewards/meter/std': 0.3601871728897095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9852941036224365, 'rewards/lexical_plausibility/std': 0.04159451276063919, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9692580699920654, 'rewards/repeat_penalty/std': 0.04013483226299286, 'rewards/repeat_floor/mean': 0.9959830045700073, 'rewards/repeat_floor/std': 0.0047417571768164635, 'rewards/near_duplicate_penalty/mean': 0.9796303510665894, 'rewards/near_duplicate_penalty/std': 0.023277563974261284, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.30315864086151123, 'rewards/total_composite/std': 0.25658613443374634, 'reward': 0.30315864086151123, 'reward_std': 0.25658613443374634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24032101035118103, 'sampling/sampling_logp_difference/max': 1.9858918190002441, 'sampling/importance_sampling_ratio/min': 0.1372581422328949, 'sampling/importance_sampling_ratio/mean': 1.0078502893447876, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7265100181102753, 'clip_ratio/low_mean': 0.11193907726556063, 'clip_ratio/low_min': 0.11193907726556063, 'clip_ratio/high_mean': 0.06650641094893217, 'clip_ratio/high_max': 0.06650641094893217, 'clip_ratio/region_mean': 0.1784454882144928, 'reward_total_mean': 0.30315864086151123, 'reward_meter_mean': 0.6007693409919739, 'reward_meter_std': 0.3601871728897095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9852941036224365, 'reward_lexical_plausibility_std': 0.04159451276063919, 'reward_repeat_penalty_mean': 0.9692580699920654, 'reward_repeat_penalty_std': 0.04013483226299286, 'reward_repeat_floor_mean': 0.9959830045700073, 'reward_repeat_floor_std': 0.0047417571768164635, 'reward_near_duplicate_penalty_mean': 0.9796303510665894, 'reward_near_duplicate_penalty_std': 0.023277563974261284, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.30315864086151123, 'reward_total_composite_std': 0.25658613443374634, 'epoch': 0.0}

  0%|          | 16/3300 [02:30<7:58:13,  8.74s/it]INFO 04-12 16:00:56 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 17/3300 [02:38<7:51:51,  8.62s/it]2026-04-12 16:01:04,525 | INFO | train_grpo_train | metrics_logged mode=train step=17

                                                   
{'loss': 0.119, 'grad_norm': 24.433513641357422, 'learning_rate': 9.951515151515152e-06, 'num_tokens': 29750.0, 'completions/mean_length': 26.375, 'completions/min_length': 16.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9241923689842224, 'rewards/meter/std': 0.1924278736114502, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7562500238418579, 'rewards/judge_quality/std': 0.31717222929000854, 'rewards/repeat_penalty/mean': 0.778178334236145, 'rewards/repeat_penalty/std': 0.08223339915275574, 'rewards/repeat_floor/mean': 0.9863755106925964, 'rewards/repeat_floor/std': 0.004412782844156027, 'rewards/near_duplicate_penalty/mean': 0.9966700077056885, 'rewards/near_duplicate_penalty/std': 0.006732024718075991, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.709308385848999, 'rewards/total_composite/std': 0.3587949275970459, 'reward': 0.709308385848999, 'reward_std': 0.3587948977947235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18496766686439514, 'sampling/sampling_logp_difference/max': 1.2623233795166016, 'sampling/importance_sampling_ratio/min': 0.2829957604408264, 'sampling/importance_sampling_ratio/mean': 1.0299561023712158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4059355482459068, 'clip_ratio/low_mean': 0.052264267578721046, 'clip_ratio/low_min': 0.052264267578721046, 'clip_ratio/high_mean': 0.12383278645575047, 'clip_ratio/high_max': 0.12383278645575047, 'clip_ratio/region_mean': 0.1760970540344715, 'reward_total_mean': 0.709308385848999, 'reward_meter_mean': 0.9241923689842224, 'reward_meter_std': 0.1924278736114502, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.778178334236145, 'reward_repeat_penalty_std': 0.08223339915275574, 'reward_repeat_floor_mean': 0.9863755106925964, 'reward_repeat_floor_std': 0.004412782844156027, 'reward_near_duplicate_penalty_mean': 0.9966700077056885, 'reward_near_duplicate_penalty_std': 0.006732024718075991, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7562500238418579, 'reward_judge_quality_std': 0.31717222929000854, 'reward_total_composite_mean': 0.709308385848999, 'reward_total_composite_std': 0.3587949275970459, 'epoch': 0.0}

  1%|          | 17/3300 [02:38<7:51:51,  8.62s/it]INFO 04-12 16:01:04 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 18/3300 [02:49<8:30:14,  9.33s/it]2026-04-12 16:01:15,493 | INFO | train_grpo_train | metrics_logged mode=train step=18

                                                   
{'loss': 0.0438, 'grad_norm': 14.398731231689453, 'learning_rate': 9.948484848484849e-06, 'num_tokens': 32077.0, 'completions/mean_length': 113.875, 'completions/min_length': 78.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.875, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.48476648330688477, 'rewards/meter/std': 0.35743826627731323, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9700000286102295, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.23868316411972046, 'rewards/repeat_penalty/mean': 0.9923823475837708, 'rewards/repeat_penalty/std': 0.003096821717917919, 'rewards/repeat_floor/mean': 0.998857319355011, 'rewards/repeat_floor/std': 0.0004645251319743693, 'rewards/near_duplicate_penalty/mean': 0.9923823475837708, 'rewards/near_duplicate_penalty/std': 0.003096821717917919, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2512439489364624, 'rewards/total_composite/std': 0.25274989008903503, 'reward': 0.2512439489364624, 'reward_std': 0.25274989008903503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.250434547662735, 'sampling/sampling_logp_difference/max': 3.4623918533325195, 'sampling/importance_sampling_ratio/min': 0.03135467693209648, 'sampling/importance_sampling_ratio/mean': 1.0127856731414795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.262481302022934, 'clip_ratio/low_mean': 0.16493276320397854, 'clip_ratio/low_min': 0.16493276320397854, 'clip_ratio/high_mean': 0.06476971507072449, 'clip_ratio/high_max': 0.06476971507072449, 'clip_ratio/region_mean': 0.22970247827470303, 'reward_total_mean': 0.2512439489364624, 'reward_meter_mean': 0.48476648330688477, 'reward_meter_std': 0.35743826627731323, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9700000286102295, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9923823475837708, 'reward_repeat_penalty_std': 0.003096821717917919, 'reward_repeat_floor_mean': 0.998857319355011, 'reward_repeat_floor_std': 0.0004645251319743693, 'reward_near_duplicate_penalty_mean': 0.9923823475837708, 'reward_near_duplicate_penalty_std': 0.003096821717917919, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.23868316411972046, 'reward_total_composite_mean': 0.2512439489364624, 'reward_total_composite_std': 0.25274989008903503, 'epoch': 0.0}

  1%|          | 18/3300 [02:49<8:30:14,  9.33s/it]INFO 04-12 16:01:15 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 19/3300 [02:59<8:40:06,  9.51s/it]2026-04-12 16:01:25,431 | INFO | train_grpo_train | metrics_logged mode=train step=19

                                                   
{'loss': 0.0743, 'grad_norm': 20.236949920654297, 'learning_rate': 9.945454545454546e-06, 'num_tokens': 33593.0, 'completions/mean_length': 39.5, 'completions/min_length': 32.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.5016039609909058, 'rewards/meter/std': 0.406399667263031, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9754565358161926, 'rewards/repeat_penalty/std': 0.04103349521756172, 'rewards/repeat_floor/mean': 0.996688723564148, 'rewards/repeat_floor/std': 0.005164620000869036, 'rewards/near_duplicate_penalty/mean': 0.9823631048202515, 'rewards/near_duplicate_penalty/std': 0.023129036650061607, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.19481655955314636, 'rewards/total_composite/std': 0.18020491302013397, 'reward': 0.19481655955314636, 'reward_std': 0.18020489811897278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24541553854942322, 'sampling/sampling_logp_difference/max': 1.647770881652832, 'sampling/importance_sampling_ratio/min': 0.1924784928560257, 'sampling/importance_sampling_ratio/mean': 1.0177640914916992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0277532786130905, 'clip_ratio/low_mean': 0.12244594004005194, 'clip_ratio/low_min': 0.12244594004005194, 'clip_ratio/high_mean': 0.0850139930844307, 'clip_ratio/high_max': 0.0850139930844307, 'clip_ratio/region_mean': 0.20745993312448263, 'reward_total_mean': 0.19481655955314636, 'reward_meter_mean': 0.5016039609909058, 'reward_meter_std': 0.406399667263031, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9754565358161926, 'reward_repeat_penalty_std': 0.04103349521756172, 'reward_repeat_floor_mean': 0.996688723564148, 'reward_repeat_floor_std': 0.005164620000869036, 'reward_near_duplicate_penalty_mean': 0.9823631048202515, 'reward_near_duplicate_penalty_std': 0.023129036650061607, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.19481655955314636, 'reward_total_composite_std': 0.18020491302013397, 'epoch': 0.0}

  1%|          | 19/3300 [02:59<8:40:06,  9.51s/it]INFO 04-12 16:01:25 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 20/3300 [03:08<8:33:36,  9.40s/it]2026-04-12 16:01:34,560 | INFO | train_grpo_train | metrics_logged mode=train step=20

                                                   
{'loss': 0.0393, 'grad_norm': 19.540916442871094, 'learning_rate': 9.942424242424244e-06, 'num_tokens': 35120.0, 'completions/mean_length': 40.875, 'completions/min_length': 31.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.4095902442932129, 'rewards/meter/std': 0.3453020453453064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.1642080694437027, 'rewards/repeat_penalty/mean': 0.9901515245437622, 'rewards/repeat_penalty/std': 0.02066391333937645, 'rewards/repeat_floor/mean': 0.9985227584838867, 'rewards/repeat_floor/std': 0.003099583787843585, 'rewards/near_duplicate_penalty/mean': 0.9901515245437622, 'rewards/near_duplicate_penalty/std': 0.02066391333937645, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1880517303943634, 'rewards/total_composite/std': 0.15503939986228943, 'reward': 0.1880517303943634, 'reward_std': 0.15503938496112823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24018830060958862, 'sampling/sampling_logp_difference/max': 1.510390281677246, 'sampling/importance_sampling_ratio/min': 0.22082377970218658, 'sampling/importance_sampling_ratio/mean': 1.040675163269043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2183572202920914, 'clip_ratio/low_mean': 0.15957853570580482, 'clip_ratio/low_min': 0.15957853570580482, 'clip_ratio/high_mean': 0.05311078391969204, 'clip_ratio/high_max': 0.05311078391969204, 'clip_ratio/region_mean': 0.21268931962549686, 'reward_total_mean': 0.1880517303943634, 'reward_meter_mean': 0.4095902442932129, 'reward_meter_std': 0.3453020453453064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9901515245437622, 'reward_repeat_penalty_std': 0.02066391333937645, 'reward_repeat_floor_mean': 0.9985227584838867, 'reward_repeat_floor_std': 0.003099583787843585, 'reward_near_duplicate_penalty_mean': 0.9901515245437622, 'reward_near_duplicate_penalty_std': 0.02066391333937645, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.1642080694437027, 'reward_total_composite_mean': 0.1880517303943634, 'reward_total_composite_std': 0.15503939986228943, 'epoch': 0.0}

  1%|          | 20/3300 [03:08<8:33:36,  9.40s/it]INFO 04-12 16:01:34 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 21/3300 [03:19<9:01:16,  9.90s/it]2026-04-12 16:01:45,647 | INFO | train_grpo_train | metrics_logged mode=train step=21

                                                   
{'loss': -0.0459, 'grad_norm': 11.305293083190918, 'learning_rate': 9.939393939393939e-06, 'num_tokens': 37183.0, 'completions/mean_length': 87.875, 'completions/min_length': 60.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.5517024993896484, 'rewards/meter/std': 0.45141252875328064, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9992185235023499, 'rewards/repeat_penalty/std': 0.0016037474852055311, 'rewards/repeat_floor/mean': 0.9998828172683716, 'rewards/repeat_floor/std': 0.0002405582636129111, 'rewards/near_duplicate_penalty/mean': 0.9992185235023499, 'rewards/near_duplicate_penalty/std': 0.0016037474852055311, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24817641079425812, 'rewards/total_composite/std': 0.18709984421730042, 'reward': 0.24817641079425812, 'reward_std': 0.18709984421730042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1770317554473877, 'sampling/sampling_logp_difference/max': 2.3745388984680176, 'sampling/importance_sampling_ratio/min': 0.09305738657712936, 'sampling/importance_sampling_ratio/mean': 1.0006312131881714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0936824977397919, 'clip_ratio/low_mean': 0.09472911804914474, 'clip_ratio/low_min': 0.09472911804914474, 'clip_ratio/high_mean': 0.07718820963054895, 'clip_ratio/high_max': 0.07718820963054895, 'clip_ratio/region_mean': 0.1719173276796937, 'reward_total_mean': 0.24817641079425812, 'reward_meter_mean': 0.5517024993896484, 'reward_meter_std': 0.45141252875328064, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9992185235023499, 'reward_repeat_penalty_std': 0.0016037474852055311, 'reward_repeat_floor_mean': 0.9998828172683716, 'reward_repeat_floor_std': 0.0002405582636129111, 'reward_near_duplicate_penalty_mean': 0.9992185235023499, 'reward_near_duplicate_penalty_std': 0.0016037474852055311, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.24817641079425812, 'reward_total_composite_std': 0.18709984421730042, 'epoch': 0.0}

  1%|          | 21/3300 [03:19<9:01:16,  9.90s/it]INFO 04-12 16:01:45 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 22/3300 [03:27<8:29:37,  9.33s/it]2026-04-12 16:01:53,633 | INFO | train_grpo_train | metrics_logged mode=train step=22

                                                   
{'loss': 0.0045, 'grad_norm': 20.91036605834961, 'learning_rate': 9.936363636363638e-06, 'num_tokens': 38762.0, 'completions/mean_length': 31.375, 'completions/min_length': 30.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9042378067970276, 'rewards/meter/std': 0.16538074612617493, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.831250011920929, 'rewards/judge_quality/std': 0.27559998631477356, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 1.0, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7410659790039062, 'rewards/total_composite/std': 0.2911670207977295, 'reward': 0.7410659790039062, 'reward_std': 0.2911670207977295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10895714163780212, 'sampling/sampling_logp_difference/max': 1.3790559768676758, 'sampling/importance_sampling_ratio/min': 0.25181618332862854, 'sampling/importance_sampling_ratio/mean': 0.9773324131965637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3513520695269108, 'clip_ratio/low_mean': 0.04734848625957966, 'clip_ratio/low_min': 0.04734848625957966, 'clip_ratio/high_mean': 0.07209188491106033, 'clip_ratio/high_max': 0.07209188491106033, 'clip_ratio/region_mean': 0.11944037117063999, 'reward_total_mean': 0.7410659790039062, 'reward_meter_mean': 0.9042378067970276, 'reward_meter_std': 0.16538074612617493, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 1.0, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.831250011920929, 'reward_judge_quality_std': 0.27559998631477356, 'reward_total_composite_mean': 0.7410659790039062, 'reward_total_composite_std': 0.2911670207977295, 'epoch': 0.0}

  1%|          | 22/3300 [03:27<8:29:37,  9.33s/it]INFO 04-12 16:01:53 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 23/3300 [03:37<8:38:30,  9.49s/it]2026-04-12 16:02:03,511 | INFO | train_grpo_train | metrics_logged mode=train step=23

                                                   
{'loss': 0.0812, 'grad_norm': 10.112000465393066, 'learning_rate': 9.933333333333334e-06, 'num_tokens': 41325.0, 'completions/mean_length': 133.375, 'completions/min_length': 99.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.375, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.8597559332847595, 'rewards/meter/std': 0.13285112380981445, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9840116500854492, 'rewards/lexical_plausibility/std': 0.04522194713354111, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9802836775779724, 'rewards/repeat_penalty/std': 0.023504260927438736, 'rewards/repeat_floor/mean': 0.9976856708526611, 'rewards/repeat_floor/std': 0.002427757252007723, 'rewards/near_duplicate_penalty/mean': 0.991152286529541, 'rewards/near_duplicate_penalty/std': 0.009577685035765171, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9890313148498535, 'rewards/distinct_2/std': 0.02148415334522724, 'rewards/total_composite/mean': 0.35202527046203613, 'rewards/total_composite/std': 0.07803290337324142, 'reward': 0.35202527046203613, 'reward_std': 0.07803290337324142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20107419788837433, 'sampling/sampling_logp_difference/max': 1.7424888610839844, 'sampling/importance_sampling_ratio/min': 0.17508408427238464, 'sampling/importance_sampling_ratio/mean': 1.0296818017959595, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1603490710258484, 'clip_ratio/low_mean': 0.10855937749147415, 'clip_ratio/low_min': 0.10855937749147415, 'clip_ratio/high_mean': 0.06738803535699844, 'clip_ratio/high_max': 0.06738803535699844, 'clip_ratio/region_mean': 0.1759474128484726, 'reward_total_mean': 0.35202527046203613, 'reward_meter_mean': 0.8597559332847595, 'reward_meter_std': 0.13285112380981445, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9840116500854492, 'reward_lexical_plausibility_std': 0.04522194713354111, 'reward_repeat_penalty_mean': 0.9802836775779724, 'reward_repeat_penalty_std': 0.023504260927438736, 'reward_repeat_floor_mean': 0.9976856708526611, 'reward_repeat_floor_std': 0.002427757252007723, 'reward_near_duplicate_penalty_mean': 0.991152286529541, 'reward_near_duplicate_penalty_std': 0.009577685035765171, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9890313148498535, 'reward_distinct_2_std': 0.02148415334522724, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.35202527046203613, 'reward_total_composite_std': 0.07803290337324142, 'epoch': 0.0}

  1%|          | 23/3300 [03:37<8:38:30,  9.49s/it]INFO 04-12 16:02:03 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 24/3300 [03:45<8:12:36,  9.02s/it]2026-04-12 16:02:11,433 | INFO | train_grpo_train | metrics_logged mode=train step=24

                                                   
{'loss': 0.046, 'grad_norm': 20.005218505859375, 'learning_rate': 9.930303030303031e-06, 'num_tokens': 43068.0, 'completions/mean_length': 54.875, 'completions/min_length': 54.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7139286994934082, 'rewards/meter/std': 0.35483241081237793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9424999952316284, 'rewards/judge_quality/std': 0.013887288980185986, 'rewards/repeat_penalty/mean': 0.9425437450408936, 'rewards/repeat_penalty/std': 0.023631354793906212, 'rewards/repeat_floor/mean': 0.9919066429138184, 'rewards/repeat_floor/std': 0.0029942041728645563, 'rewards/near_duplicate_penalty/mean': 0.9515388011932373, 'rewards/near_duplicate_penalty/std': 0.02329661138355732, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.6711771488189697, 'rewards/total_composite/std': 0.3355337381362915, 'reward': 0.6711771488189697, 'reward_std': 0.3355337381362915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06163253262639046, 'sampling/sampling_logp_difference/max': 1.5534675121307373, 'sampling/importance_sampling_ratio/min': 0.23674805462360382, 'sampling/importance_sampling_ratio/mean': 1.0056684017181396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.20610892679542303, 'clip_ratio/low_mean': 0.028719397261738777, 'clip_ratio/low_min': 0.028719397261738777, 'clip_ratio/high_mean': 0.02050264598801732, 'clip_ratio/high_max': 0.02050264598801732, 'clip_ratio/region_mean': 0.0492220432497561, 'reward_total_mean': 0.6711771488189697, 'reward_meter_mean': 0.7139286994934082, 'reward_meter_std': 0.35483241081237793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9425437450408936, 'reward_repeat_penalty_std': 0.023631354793906212, 'reward_repeat_floor_mean': 0.9919066429138184, 'reward_repeat_floor_std': 0.0029942041728645563, 'reward_near_duplicate_penalty_mean': 0.9515388011932373, 'reward_near_duplicate_penalty_std': 0.02329661138355732, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.9424999952316284, 'reward_judge_quality_std': 0.013887288980185986, 'reward_total_composite_mean': 0.6711771488189697, 'reward_total_composite_std': 0.3355337381362915, 'epoch': 0.0}

  1%|          | 24/3300 [03:45<8:12:36,  9.02s/it]INFO 04-12 16:02:11 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 25/3300 [03:54<8:18:48,  9.14s/it]2026-04-12 16:02:20,846 | INFO | train_grpo_train | metrics_logged mode=train step=25

                                                   
{'loss': 0.156, 'grad_norm': 16.10081672668457, 'learning_rate': 9.927272727272728e-06, 'num_tokens': 44673.0, 'completions/mean_length': 45.625, 'completions/min_length': 35.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.5014105439186096, 'rewards/meter/std': 0.308937668800354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.3378080725669861, 'rewards/repeat_penalty/mean': 0.9912397861480713, 'rewards/repeat_penalty/std': 0.013325787149369717, 'rewards/repeat_floor/mean': 0.9986859560012817, 'rewards/repeat_floor/std': 0.0019988620188087225, 'rewards/near_duplicate_penalty/mean': 0.9912397861480713, 'rewards/near_duplicate_penalty/std': 0.013325787149369717, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23511448502540588, 'rewards/total_composite/std': 0.16499361395835876, 'reward': 0.23511448502540588, 'reward_std': 0.16499361395835876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.250480979681015, 'sampling/sampling_logp_difference/max': 1.9168643951416016, 'sampling/importance_sampling_ratio/min': 0.1470673829317093, 'sampling/importance_sampling_ratio/mean': 1.0359736680984497, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0624225437641144, 'clip_ratio/low_mean': 0.1553692352026701, 'clip_ratio/low_min': 0.1553692352026701, 'clip_ratio/high_mean': 0.0832142885774374, 'clip_ratio/high_max': 0.0832142885774374, 'clip_ratio/region_mean': 0.2385835237801075, 'reward_total_mean': 0.23511448502540588, 'reward_meter_mean': 0.5014105439186096, 'reward_meter_std': 0.308937668800354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9912397861480713, 'reward_repeat_penalty_std': 0.013325787149369717, 'reward_repeat_floor_mean': 0.9986859560012817, 'reward_repeat_floor_std': 0.0019988620188087225, 'reward_near_duplicate_penalty_mean': 0.9912397861480713, 'reward_near_duplicate_penalty_std': 0.013325787149369717, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.3378080725669861, 'reward_total_composite_mean': 0.23511448502540588, 'reward_total_composite_std': 0.16499361395835876, 'epoch': 0.0}

  1%|          | 25/3300 [03:54<8:18:48,  9.14s/it]INFO 04-12 16:02:21 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 26/3300 [04:04<8:32:08,  9.39s/it]2026-04-12 16:02:30,805 | INFO | train_grpo_train | metrics_logged mode=train step=26

                                                   
{'loss': 0.0318, 'grad_norm': 13.510664939880371, 'learning_rate': 9.924242424242425e-06, 'num_tokens': 46531.0, 'completions/mean_length': 61.25, 'completions/min_length': 36.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.31675535440444946, 'rewards/meter/std': 0.33889898657798767, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9831216335296631, 'rewards/lexical_plausibility/std': 0.03366175666451454, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9987928867340088, 'rewards/repeat_penalty/std': 0.0027606007643043995, 'rewards/repeat_floor/mean': 0.9998189210891724, 'rewards/repeat_floor/std': 0.0004140916280448437, 'rewards/near_duplicate_penalty/mean': 0.9987928867340088, 'rewards/near_duplicate_penalty/std': 0.0027606007643043995, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14853471517562866, 'rewards/total_composite/std': 0.17436335980892181, 'reward': 0.14853471517562866, 'reward_std': 0.17436335980892181, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.228131502866745, 'sampling/sampling_logp_difference/max': 1.9820013046264648, 'sampling/importance_sampling_ratio/min': 0.13779319822788239, 'sampling/importance_sampling_ratio/mean': 1.0108184814453125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.186939314007759, 'clip_ratio/low_mean': 0.11863413453102112, 'clip_ratio/low_min': 0.11863413453102112, 'clip_ratio/high_mean': 0.09134547226130962, 'clip_ratio/high_max': 0.09134547226130962, 'clip_ratio/region_mean': 0.20997960679233074, 'reward_total_mean': 0.14853471517562866, 'reward_meter_mean': 0.31675535440444946, 'reward_meter_std': 0.33889898657798767, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9831216335296631, 'reward_lexical_plausibility_std': 0.03366175666451454, 'reward_repeat_penalty_mean': 0.9987928867340088, 'reward_repeat_penalty_std': 0.0027606007643043995, 'reward_repeat_floor_mean': 0.9998189210891724, 'reward_repeat_floor_std': 0.0004140916280448437, 'reward_near_duplicate_penalty_mean': 0.9987928867340088, 'reward_near_duplicate_penalty_std': 0.0027606007643043995, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.14853471517562866, 'reward_total_composite_std': 0.17436335980892181, 'epoch': 0.0}

  1%|          | 26/3300 [04:04<8:32:08,  9.39s/it]INFO 04-12 16:02:31 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 27/3300 [04:16<9:16:09, 10.20s/it]2026-04-12 16:02:42,890 | INFO | train_grpo_train | metrics_logged mode=train step=27

                                                   
{'loss': 0.4783, 'grad_norm': 11.73549747467041, 'learning_rate': 9.921212121212121e-06, 'num_tokens': 48667.0, 'completions/mean_length': 90.0, 'completions/min_length': 44.0, 'completions/max_length': 306.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.0, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 306.0, 'rewards/meter/mean': 0.36504948139190674, 'rewards/meter/std': 0.26965948939323425, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9226248264312744, 'rewards/lexical_plausibility/std': 0.16758441925048828, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.15059404075145721, 'rewards/repeat_penalty/mean': 0.9928141832351685, 'rewards/repeat_penalty/std': 0.010346044786274433, 'rewards/repeat_floor/mean': 0.9990363717079163, 'rewards/repeat_floor/std': 0.0013709188206121325, 'rewards/near_duplicate_penalty/mean': 0.9947412610054016, 'rewards/near_duplicate_penalty/std': 0.008094299584627151, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.99805748462677, 'rewards/distinct_2/std': 0.005494221113622189, 'rewards/total_composite/mean': 0.11421902477741241, 'rewards/total_composite/std': 0.09893438965082169, 'reward': 0.11421902477741241, 'reward_std': 0.09893438965082169, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2172243744134903, 'sampling/sampling_logp_difference/max': 1.404275894165039, 'sampling/importance_sampling_ratio/min': 0.24554479122161865, 'sampling/importance_sampling_ratio/mean': 1.0379116535186768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.275842383503914, 'clip_ratio/low_mean': 0.10809156112372875, 'clip_ratio/low_min': 0.10809156112372875, 'clip_ratio/high_mean': 0.12825565226376057, 'clip_ratio/high_max': 0.12825565226376057, 'clip_ratio/region_mean': 0.23634721338748932, 'reward_total_mean': 0.11421902477741241, 'reward_meter_mean': 0.36504948139190674, 'reward_meter_std': 0.26965948939323425, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9226248264312744, 'reward_lexical_plausibility_std': 0.16758441925048828, 'reward_repeat_penalty_mean': 0.9928141832351685, 'reward_repeat_penalty_std': 0.010346044786274433, 'reward_repeat_floor_mean': 0.9990363717079163, 'reward_repeat_floor_std': 0.0013709188206121325, 'reward_near_duplicate_penalty_mean': 0.9947412610054016, 'reward_near_duplicate_penalty_std': 0.008094299584627151, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.99805748462677, 'reward_distinct_2_std': 0.005494221113622189, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.15059404075145721, 'reward_total_composite_mean': 0.11421902477741241, 'reward_total_composite_std': 0.09893438965082169, 'epoch': 0.0}

  1%|          | 27/3300 [04:16<9:16:09, 10.20s/it]INFO 04-12 16:02:43 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 28/3300 [04:26<9:14:23, 10.17s/it]2026-04-12 16:02:52,988 | INFO | train_grpo_train | metrics_logged mode=train step=28

                                                   
{'loss': -0.1519, 'grad_norm': 12.113202095031738, 'learning_rate': 9.918181818181818e-06, 'num_tokens': 50777.0, 'completions/mean_length': 96.75, 'completions/min_length': 66.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.4711955189704895, 'rewards/meter/std': 0.31360843777656555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9659777283668518, 'rewards/lexical_plausibility/std': 0.044811539351940155, 'rewards/judge_quality/mean': 0.23749998211860657, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9891072511672974, 'rewards/repeat_penalty/std': 0.019332988187670708, 'rewards/repeat_floor/mean': 0.9987300038337708, 'rewards/repeat_floor/std': 0.0018919405993074179, 'rewards/near_duplicate_penalty/mean': 0.9952554106712341, 'rewards/near_duplicate_penalty/std': 0.0035783271305263042, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9937965273857117, 'rewards/distinct_2/std': 0.017546070739626884, 'rewards/total_composite/mean': 0.11048850417137146, 'rewards/total_composite/std': 0.10270975530147552, 'reward': 0.11048850417137146, 'reward_std': 0.10270974785089493, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.28622543811798096, 'sampling/sampling_logp_difference/max': 3.824678897857666, 'sampling/importance_sampling_ratio/min': 0.02182544209063053, 'sampling/importance_sampling_ratio/mean': 1.0070923566818237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5282831639051437, 'clip_ratio/low_mean': 0.15044707618653774, 'clip_ratio/low_min': 0.15044707618653774, 'clip_ratio/high_mean': 0.08182669058442116, 'clip_ratio/high_max': 0.08182669058442116, 'clip_ratio/region_mean': 0.2322737667709589, 'reward_total_mean': 0.11048850417137146, 'reward_meter_mean': 0.4711955189704895, 'reward_meter_std': 0.31360843777656555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9659777283668518, 'reward_lexical_plausibility_std': 0.044811539351940155, 'reward_repeat_penalty_mean': 0.9891072511672974, 'reward_repeat_penalty_std': 0.019332988187670708, 'reward_repeat_floor_mean': 0.9987300038337708, 'reward_repeat_floor_std': 0.0018919405993074179, 'reward_near_duplicate_penalty_mean': 0.9952554106712341, 'reward_near_duplicate_penalty_std': 0.0035783271305263042, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9937965273857117, 'reward_distinct_2_std': 0.017546070739626884, 'reward_judge_quality_mean': 0.23749998211860657, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.11048850417137146, 'reward_total_composite_std': 0.10270975530147552, 'epoch': 0.0}

  1%|          | 28/3300 [04:26<9:14:23, 10.17s/it]INFO 04-12 16:02:53 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 29/3300 [04:35<8:50:52,  9.74s/it]2026-04-12 16:03:01,728 | INFO | train_grpo_train | metrics_logged mode=train step=29

                                                   
{'loss': 0.0238, 'grad_norm': 16.65239715576172, 'learning_rate': 9.915151515151515e-06, 'num_tokens': 52511.0, 'completions/mean_length': 53.75, 'completions/min_length': 37.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6935027241706848, 'rewards/meter/std': 0.417043000459671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9473055601119995, 'rewards/lexical_plausibility/std': 0.08244762569665909, 'rewards/judge_quality/mean': 0.5424999594688416, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.9909010529518127, 'rewards/repeat_penalty/std': 0.013155777007341385, 'rewards/repeat_floor/mean': 0.9986351728439331, 'rewards/repeat_floor/std': 0.001973371487110853, 'rewards/near_duplicate_penalty/mean': 0.9909010529518127, 'rewards/near_duplicate_penalty/std': 0.013155777007341385, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3540693521499634, 'rewards/total_composite/std': 0.272655189037323, 'reward': 0.3540693521499634, 'reward_std': 0.2726552188396454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20613504946231842, 'sampling/sampling_logp_difference/max': 1.5167832374572754, 'sampling/importance_sampling_ratio/min': 0.277983695268631, 'sampling/importance_sampling_ratio/mean': 1.0124094486236572, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.05456180870533, 'clip_ratio/low_mean': 0.09870942868292332, 'clip_ratio/low_min': 0.09870942868292332, 'clip_ratio/high_mean': 0.10838798806071281, 'clip_ratio/high_max': 0.10838798806071281, 'clip_ratio/region_mean': 0.20709741674363613, 'reward_total_mean': 0.3540693521499634, 'reward_meter_mean': 0.6935027241706848, 'reward_meter_std': 0.417043000459671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9473055601119995, 'reward_lexical_plausibility_std': 0.08244762569665909, 'reward_repeat_penalty_mean': 0.9909010529518127, 'reward_repeat_penalty_std': 0.013155777007341385, 'reward_repeat_floor_mean': 0.9986351728439331, 'reward_repeat_floor_std': 0.001973371487110853, 'reward_near_duplicate_penalty_mean': 0.9909010529518127, 'reward_near_duplicate_penalty_std': 0.013155777007341385, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5424999594688416, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.3540693521499634, 'reward_total_composite_std': 0.272655189037323, 'epoch': 0.0}

  1%|          | 29/3300 [04:35<8:50:52,  9.74s/it]INFO 04-12 16:03:01 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 30/3300 [04:44<8:38:07,  9.51s/it]2026-04-12 16:03:10,694 | INFO | train_grpo_train | metrics_logged mode=train step=30

                                                   
{'loss': -0.0115, 'grad_norm': 19.788253784179688, 'learning_rate': 9.912121212121213e-06, 'num_tokens': 54588.0, 'completions/mean_length': 75.625, 'completions/min_length': 66.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.6688518524169922, 'rewards/meter/std': 0.295454204082489, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.45500001311302185, 'rewards/judge_quality/std': 0.3097464442253113, 'rewards/repeat_penalty/mean': 0.9855921268463135, 'rewards/repeat_penalty/std': 0.01247401349246502, 'rewards/repeat_floor/mean': 0.9978388547897339, 'rewards/repeat_floor/std': 0.0018710998119786382, 'rewards/near_duplicate_penalty/mean': 0.9855921268463135, 'rewards/near_duplicate_penalty/std': 0.01247401349246502, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22063979506492615, 'rewards/total_composite/std': 0.07142075896263123, 'reward': 0.22063979506492615, 'reward_std': 0.07142075896263123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24945597350597382, 'sampling/sampling_logp_difference/max': 3.091531753540039, 'sampling/importance_sampling_ratio/min': 0.04543231055140495, 'sampling/importance_sampling_ratio/mean': 0.9944108128547668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8818572908639908, 'clip_ratio/low_mean': 0.14803308434784412, 'clip_ratio/low_min': 0.14803308434784412, 'clip_ratio/high_mean': 0.06854380294680595, 'clip_ratio/high_max': 0.06854380294680595, 'clip_ratio/region_mean': 0.21657688729465008, 'reward_total_mean': 0.22063979506492615, 'reward_meter_mean': 0.6688518524169922, 'reward_meter_std': 0.295454204082489, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9855921268463135, 'reward_repeat_penalty_std': 0.01247401349246502, 'reward_repeat_floor_mean': 0.9978388547897339, 'reward_repeat_floor_std': 0.0018710998119786382, 'reward_near_duplicate_penalty_mean': 0.9855921268463135, 'reward_near_duplicate_penalty_std': 0.01247401349246502, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45500001311302185, 'reward_judge_quality_std': 0.3097464442253113, 'reward_total_composite_mean': 0.22063979506492615, 'reward_total_composite_std': 0.07142075896263123, 'epoch': 0.0}

  1%|          | 30/3300 [04:44<8:38:07,  9.51s/it]INFO 04-12 16:03:10 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 31/3300 [04:54<8:41:53,  9.58s/it]2026-04-12 16:03:20,441 | INFO | train_grpo_train | metrics_logged mode=train step=31

                                                   
{'loss': 0.1996, 'grad_norm': 20.691179275512695, 'learning_rate': 9.90909090909091e-06, 'num_tokens': 56130.0, 'completions/mean_length': 28.75, 'completions/min_length': 22.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8624352812767029, 'rewards/meter/std': 0.34842345118522644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9824293255805969, 'rewards/lexical_plausibility/std': 0.049697354435920715, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.3651198744773865, 'rewards/repeat_penalty/mean': 0.7369969487190247, 'rewards/repeat_penalty/std': 0.03434223309159279, 'rewards/repeat_floor/mean': 0.9823994040489197, 'rewards/repeat_floor/std': 0.006868451833724976, 'rewards/near_duplicate_penalty/mean': 0.982662558555603, 'rewards/near_duplicate_penalty/std': 0.04578967019915581, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5399028658866882, 'rewards/total_composite/std': 0.39929813146591187, 'reward': 0.5399028658866882, 'reward_std': 0.3992981016635895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19227752089500427, 'sampling/sampling_logp_difference/max': 1.1030511856079102, 'sampling/importance_sampling_ratio/min': 0.33185696601867676, 'sampling/importance_sampling_ratio/mean': 1.0560588836669922, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8176650553941727, 'clip_ratio/low_mean': 0.11976448819041252, 'clip_ratio/low_min': 0.11976448819041252, 'clip_ratio/high_mean': 0.13372859358787537, 'clip_ratio/high_max': 0.13372859358787537, 'clip_ratio/region_mean': 0.2534930817782879, 'reward_total_mean': 0.5399028658866882, 'reward_meter_mean': 0.8624352812767029, 'reward_meter_std': 0.34842345118522644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9824293255805969, 'reward_lexical_plausibility_std': 0.049697354435920715, 'reward_repeat_penalty_mean': 0.7369969487190247, 'reward_repeat_penalty_std': 0.03434223309159279, 'reward_repeat_floor_mean': 0.9823994040489197, 'reward_repeat_floor_std': 0.006868451833724976, 'reward_near_duplicate_penalty_mean': 0.982662558555603, 'reward_near_duplicate_penalty_std': 0.04578967019915581, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.3651198744773865, 'reward_total_composite_mean': 0.5399028658866882, 'reward_total_composite_std': 0.39929813146591187, 'epoch': 0.0}

  1%|          | 31/3300 [04:54<8:41:53,  9.58s/it]INFO 04-12 16:03:20 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 32/3300 [05:04<8:44:52,  9.64s/it]2026-04-12 16:03:30,214 | INFO | train_grpo_train | metrics_logged mode=train step=32

                                                   
{'loss': -0.0146, 'grad_norm': 12.498920440673828, 'learning_rate': 9.906060606060607e-06, 'num_tokens': 58343.0, 'completions/mean_length': 98.625, 'completions/min_length': 76.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.625, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.3425254821777344, 'rewards/meter/std': 0.3815731108188629, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.04242641106247902, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9930620193481445, 'rewards/repeat_penalty/std': 0.005063355900347233, 'rewards/repeat_floor/mean': 0.9989593029022217, 'rewards/repeat_floor/std': 0.0007595062488690019, 'rewards/near_duplicate_penalty/mean': 0.9930620193481445, 'rewards/near_duplicate_penalty/std': 0.005063355900347233, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15067604184150696, 'rewards/total_composite/std': 0.17259570956230164, 'reward': 0.15067604184150696, 'reward_std': 0.17259569466114044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2743317186832428, 'sampling/sampling_logp_difference/max': 1.9858273267745972, 'sampling/importance_sampling_ratio/min': 0.13726700842380524, 'sampling/importance_sampling_ratio/mean': 1.0336440801620483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.797087222337723, 'clip_ratio/low_mean': 0.1716446615755558, 'clip_ratio/low_min': 0.1716446615755558, 'clip_ratio/high_mean': 0.05666547268629074, 'clip_ratio/high_max': 0.05666547268629074, 'clip_ratio/region_mean': 0.22831013426184654, 'reward_total_mean': 0.15067604184150696, 'reward_meter_mean': 0.3425254821777344, 'reward_meter_std': 0.3815731108188629, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.04242641106247902, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9930620193481445, 'reward_repeat_penalty_std': 0.005063355900347233, 'reward_repeat_floor_mean': 0.9989593029022217, 'reward_repeat_floor_std': 0.0007595062488690019, 'reward_near_duplicate_penalty_mean': 0.9930620193481445, 'reward_near_duplicate_penalty_std': 0.005063355900347233, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.15067604184150696, 'reward_total_composite_std': 0.17259570956230164, 'epoch': 0.0}

  1%|          | 32/3300 [05:04<8:44:52,  9.64s/it]INFO 04-12 16:03:30 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 33/3300 [05:17<9:51:14, 10.86s/it]2026-04-12 16:03:43,922 | INFO | train_grpo_train | metrics_logged mode=train step=33

                                                   
{'loss': 0.0628, 'grad_norm': 5.068249225616455, 'learning_rate': 9.903030303030305e-06, 'num_tokens': 62874.0, 'completions/mean_length': 357.375, 'completions/min_length': 248.0, 'completions/max_length': 476.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 357.375, 'completions/min_terminated_length': 248.0, 'completions/max_terminated_length': 476.0, 'rewards/meter/mean': 0.5966706275939941, 'rewards/meter/std': 0.4446839988231659, 'rewards/count_adherence/mean': 0.9659091234207153, 'rewards/count_adherence/std': 0.047049909830093384, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9897727370262146, 'rewards/count_floor/std': 0.014114963822066784, 'rewards/lexical_plausibility/mean': 0.9950504899024963, 'rewards/lexical_plausibility/std': 0.009223292581737041, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9939025640487671, 'rewards/repeat_penalty/std': 0.005231776740401983, 'rewards/repeat_floor/mean': 0.9991873502731323, 'rewards/repeat_floor/std': 0.0005345784593373537, 'rewards/near_duplicate_penalty/mean': 0.9956123232841492, 'rewards/near_duplicate_penalty/std': 0.002133031142875552, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9982829689979553, 'rewards/distinct_2/std': 0.004856497049331665, 'rewards/total_composite/mean': 0.18759429454803467, 'rewards/total_composite/std': 0.1536462903022766, 'reward': 0.18759429454803467, 'reward_std': 0.15364627540111542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19716061651706696, 'sampling/sampling_logp_difference/max': 1.774923324584961, 'sampling/importance_sampling_ratio/min': 0.16949644684791565, 'sampling/importance_sampling_ratio/mean': 1.0434120893478394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7610519528388977, 'clip_ratio/low_mean': 0.0867075864225626, 'clip_ratio/low_min': 0.0867075864225626, 'clip_ratio/high_mean': 0.095170047134161, 'clip_ratio/high_max': 0.095170047134161, 'clip_ratio/region_mean': 0.1818776335567236, 'reward_total_mean': 0.18759429454803467, 'reward_meter_mean': 0.5966706275939941, 'reward_meter_std': 0.4446839988231659, 'reward_count_adherence_mean': 0.9659091234207153, 'reward_count_adherence_std': 0.047049909830093384, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9897727370262146, 'reward_count_floor_std': 0.014114963822066784, 'reward_lexical_plausibility_mean': 0.9950504899024963, 'reward_lexical_plausibility_std': 0.009223292581737041, 'reward_repeat_penalty_mean': 0.9939025640487671, 'reward_repeat_penalty_std': 0.005231776740401983, 'reward_repeat_floor_mean': 0.9991873502731323, 'reward_repeat_floor_std': 0.0005345784593373537, 'reward_near_duplicate_penalty_mean': 0.9956123232841492, 'reward_near_duplicate_penalty_std': 0.002133031142875552, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9982829689979553, 'reward_distinct_2_std': 0.004856497049331665, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.18759429454803467, 'reward_total_composite_std': 0.1536462903022766, 'epoch': 0.0}

  1%|          | 33/3300 [05:17<9:51:14, 10.86s/it]INFO 04-12 16:03:44 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 34/3300 [05:26<9:22:21, 10.33s/it]2026-04-12 16:03:53,024 | INFO | train_grpo_train | metrics_logged mode=train step=34

                                                   
{'loss': 0.0031, 'grad_norm': 9.778609275817871, 'learning_rate': 9.9e-06, 'num_tokens': 64898.0, 'completions/mean_length': 98.0, 'completions/min_length': 64.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.6410538554191589, 'rewards/meter/std': 0.39640089869499207, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.2314550131559372, 'rewards/repeat_penalty/mean': 0.9879550933837891, 'rewards/repeat_penalty/std': 0.020515216514468193, 'rewards/repeat_floor/mean': 0.9985188245773315, 'rewards/repeat_floor/std': 0.0022078033071011305, 'rewards/near_duplicate_penalty/mean': 0.9935190081596375, 'rewards/near_duplicate_penalty/std': 0.00695214094594121, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9943438768386841, 'rewards/distinct_2/std': 0.015997890383005142, 'rewards/total_composite/mean': 0.2914750277996063, 'rewards/total_composite/std': 0.29660576581954956, 'reward': 0.2914750277996063, 'reward_std': 0.2966057360172272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2059471756219864, 'sampling/sampling_logp_difference/max': 2.3541133403778076, 'sampling/importance_sampling_ratio/min': 0.09497768431901932, 'sampling/importance_sampling_ratio/mean': 1.0350326299667358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3347521275281906, 'clip_ratio/low_mean': 0.11685441620647907, 'clip_ratio/low_min': 0.11685441620647907, 'clip_ratio/high_mean': 0.07928240671753883, 'clip_ratio/high_max': 0.07928240671753883, 'clip_ratio/region_mean': 0.1961368229240179, 'reward_total_mean': 0.2914750277996063, 'reward_meter_mean': 0.6410538554191589, 'reward_meter_std': 0.39640089869499207, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9879550933837891, 'reward_repeat_penalty_std': 0.020515216514468193, 'reward_repeat_floor_mean': 0.9985188245773315, 'reward_repeat_floor_std': 0.0022078033071011305, 'reward_near_duplicate_penalty_mean': 0.9935190081596375, 'reward_near_duplicate_penalty_std': 0.00695214094594121, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9943438768386841, 'reward_distinct_2_std': 0.015997890383005142, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.2314550131559372, 'reward_total_composite_mean': 0.2914750277996063, 'reward_total_composite_std': 0.29660576581954956, 'epoch': 0.0}

  1%|          | 34/3300 [05:26<9:22:21, 10.33s/it]INFO 04-12 16:03:53 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 35/3300 [05:46<11:49:48, 13.04s/it]2026-04-12 16:04:12,397 | INFO | train_grpo_train | metrics_logged mode=train step=35

                                                    
{'loss': 0.5092, 'grad_norm': 16.70918846130371, 'learning_rate': 9.896969696969699e-06, 'num_tokens': 66565.0, 'completions/mean_length': 61.375, 'completions/min_length': 38.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.692173957824707, 'rewards/meter/std': 0.4213980436325073, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.9863764643669128, 'rewards/lexical_plausibility/std': 0.038533199578523636, 'rewards/judge_quality/mean': 0.4374999701976776, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9922046065330505, 'rewards/repeat_penalty/std': 0.010986856184899807, 'rewards/repeat_floor/mean': 0.9988306760787964, 'rewards/repeat_floor/std': 0.0016480127815157175, 'rewards/near_duplicate_penalty/mean': 0.9922046065330505, 'rewards/near_duplicate_penalty/std': 0.010986856184899807, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31467336416244507, 'rewards/total_composite/std': 0.21308283507823944, 'reward': 0.31467336416244507, 'reward_std': 0.21308282017707825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2405560463666916, 'sampling/sampling_logp_difference/max': 2.1018619537353516, 'sampling/importance_sampling_ratio/min': 0.12222862988710403, 'sampling/importance_sampling_ratio/mean': 1.037939429283142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.62863227725029, 'clip_ratio/low_mean': 0.05899621360003948, 'clip_ratio/low_min': 0.05899621360003948, 'clip_ratio/high_mean': 0.1803486943244934, 'clip_ratio/high_max': 0.1803486943244934, 'clip_ratio/region_mean': 0.2393449079245329, 'reward_total_mean': 0.31467336416244507, 'reward_meter_mean': 0.692173957824707, 'reward_meter_std': 0.4213980436325073, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.9863764643669128, 'reward_lexical_plausibility_std': 0.038533199578523636, 'reward_repeat_penalty_mean': 0.9922046065330505, 'reward_repeat_penalty_std': 0.010986856184899807, 'reward_repeat_floor_mean': 0.9988306760787964, 'reward_repeat_floor_std': 0.0016480127815157175, 'reward_near_duplicate_penalty_mean': 0.9922046065330505, 'reward_near_duplicate_penalty_std': 0.010986856184899807, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4374999701976776, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.31467336416244507, 'reward_total_composite_std': 0.21308283507823944, 'epoch': 0.0}

  1%|          | 35/3300 [05:46<11:49:48, 13.04s/it]INFO 04-12 16:04:12 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 36/3300 [05:55<10:55:18, 12.05s/it]2026-04-12 16:04:22,116 | INFO | train_grpo_train | metrics_logged mode=train step=36

                                                    
{'loss': 0.0688, 'grad_norm': 13.143230438232422, 'learning_rate': 9.893939393939395e-06, 'num_tokens': 68772.0, 'completions/mean_length': 90.875, 'completions/min_length': 71.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.875, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.722322404384613, 'rewards/meter/std': 0.25282078981399536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9908853769302368, 'rewards/lexical_plausibility/std': 0.021876772865653038, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9927700757980347, 'rewards/repeat_penalty/std': 0.004573230631649494, 'rewards/repeat_floor/mean': 0.9989154934883118, 'rewards/repeat_floor/std': 0.0006859906134195626, 'rewards/near_duplicate_penalty/mean': 0.9927700757980347, 'rewards/near_duplicate_penalty/std': 0.004573230631649494, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27753379940986633, 'rewards/total_composite/std': 0.10023543238639832, 'reward': 0.27753379940986633, 'reward_std': 0.10023541003465652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2324392944574356, 'sampling/sampling_logp_difference/max': 3.417316198348999, 'sampling/importance_sampling_ratio/min': 0.0328003466129303, 'sampling/importance_sampling_ratio/mean': 1.0347892045974731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4474022686481476, 'clip_ratio/low_mean': 0.0823497623205185, 'clip_ratio/low_min': 0.0823497623205185, 'clip_ratio/high_mean': 0.13351451978087425, 'clip_ratio/high_max': 0.13351451978087425, 'clip_ratio/region_mean': 0.21586428210139275, 'reward_total_mean': 0.27753379940986633, 'reward_meter_mean': 0.722322404384613, 'reward_meter_std': 0.25282078981399536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9908853769302368, 'reward_lexical_plausibility_std': 0.021876772865653038, 'reward_repeat_penalty_mean': 0.9927700757980347, 'reward_repeat_penalty_std': 0.004573230631649494, 'reward_repeat_floor_mean': 0.9989154934883118, 'reward_repeat_floor_std': 0.0006859906134195626, 'reward_near_duplicate_penalty_mean': 0.9927700757980347, 'reward_near_duplicate_penalty_std': 0.004573230631649494, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.27753379940986633, 'reward_total_composite_std': 0.10023543238639832, 'epoch': 0.0}

  1%|          | 36/3300 [05:55<10:55:18, 12.05s/it]INFO 04-12 16:04:22 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 37/3300 [06:04<10:04:10, 11.11s/it]2026-04-12 16:04:31,040 | INFO | train_grpo_train | metrics_logged mode=train step=37

                                                    
{'loss': 0.3362, 'grad_norm': 22.158899307250977, 'learning_rate': 9.890909090909092e-06, 'num_tokens': 70327.0, 'completions/mean_length': 30.375, 'completions/min_length': 17.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.45229798555374146, 'rewards/meter/std': 0.41445446014404297, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.0289318785071373, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.2924496531486511, 'rewards/repeat_penalty/mean': 0.8070650696754456, 'rewards/repeat_penalty/std': 0.10630100965499878, 'rewards/repeat_floor/mean': 0.9879347681999207, 'rewards/repeat_floor/std': 0.005706820636987686, 'rewards/near_duplicate_penalty/mean': 0.9945650696754456, 'rewards/near_duplicate_penalty/std': 0.015372347086668015, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24394118785858154, 'rewards/total_composite/std': 0.25945141911506653, 'reward': 0.24394118785858154, 'reward_std': 0.25945141911506653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25791147351264954, 'sampling/sampling_logp_difference/max': 2.059619903564453, 'sampling/importance_sampling_ratio/min': 0.1275024265050888, 'sampling/importance_sampling_ratio/mean': 1.031569242477417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6196595281362534, 'clip_ratio/low_mean': 0.13278313353657722, 'clip_ratio/low_min': 0.13278313353657722, 'clip_ratio/high_mean': 0.05720588192343712, 'clip_ratio/high_max': 0.05720588192343712, 'clip_ratio/region_mean': 0.18998901546001434, 'reward_total_mean': 0.24394118785858154, 'reward_meter_mean': 0.45229798555374146, 'reward_meter_std': 0.41445446014404297, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.0289318785071373, 'reward_repeat_penalty_mean': 0.8070650696754456, 'reward_repeat_penalty_std': 0.10630100965499878, 'reward_repeat_floor_mean': 0.9879347681999207, 'reward_repeat_floor_std': 0.005706820636987686, 'reward_near_duplicate_penalty_mean': 0.9945650696754456, 'reward_near_duplicate_penalty_std': 0.015372347086668015, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.2924496531486511, 'reward_total_composite_mean': 0.24394118785858154, 'reward_total_composite_std': 0.25945141911506653, 'epoch': 0.0}

  1%|          | 37/3300 [06:04<10:04:10, 11.11s/it]INFO 04-12 16:04:31 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 38/3300 [06:14<9:33:35, 10.55s/it] 2026-04-12 16:04:40,285 | INFO | train_grpo_train | metrics_logged mode=train step=38

                                                   
{'loss': -0.019, 'grad_norm': 17.26567268371582, 'learning_rate': 9.887878787878789e-06, 'num_tokens': 71817.0, 'completions/mean_length': 43.25, 'completions/min_length': 32.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5460525751113892, 'rewards/meter/std': 0.4408157765865326, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.9705172181129456, 'rewards/repeat_penalty/std': 0.04875211417675018, 'rewards/repeat_floor/mean': 0.996078372001648, 'rewards/repeat_floor/std': 0.005964566953480244, 'rewards/near_duplicate_penalty/mean': 0.9799288511276245, 'rewards/near_duplicate_penalty/std': 0.024315785616636276, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.3269811272621155, 'rewards/total_composite/std': 0.3435862362384796, 'reward': 0.3269811272621155, 'reward_std': 0.3435862064361572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2100449949502945, 'sampling/sampling_logp_difference/max': 1.4643468856811523, 'sampling/importance_sampling_ratio/min': 0.23122896254062653, 'sampling/importance_sampling_ratio/mean': 1.0345838069915771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8484523594379425, 'clip_ratio/low_mean': 0.13420672714710236, 'clip_ratio/low_min': 0.13420672714710236, 'clip_ratio/high_mean': 0.08962204866111279, 'clip_ratio/high_max': 0.08962204866111279, 'clip_ratio/region_mean': 0.22382877580821514, 'reward_total_mean': 0.3269811272621155, 'reward_meter_mean': 0.5460525751113892, 'reward_meter_std': 0.4408157765865326, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9705172181129456, 'reward_repeat_penalty_std': 0.04875211417675018, 'reward_repeat_floor_mean': 0.996078372001648, 'reward_repeat_floor_std': 0.005964566953480244, 'reward_near_duplicate_penalty_mean': 0.9799288511276245, 'reward_near_duplicate_penalty_std': 0.024315785616636276, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.3269811272621155, 'reward_total_composite_std': 0.3435862362384796, 'epoch': 0.0}

  1%|          | 38/3300 [06:14<9:33:35, 10.55s/it]INFO 04-12 16:04:40 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 39/3300 [06:25<9:41:21, 10.70s/it]2026-04-12 16:04:51,324 | INFO | train_grpo_train | metrics_logged mode=train step=39

                                                   
{'loss': -0.0047, 'grad_norm': 12.93870735168457, 'learning_rate': 9.884848484848486e-06, 'num_tokens': 74018.0, 'completions/mean_length': 104.125, 'completions/min_length': 61.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.125, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.4005318582057953, 'rewards/meter/std': 0.3796837627887726, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9792049527168274, 'rewards/lexical_plausibility/std': 0.04006021469831467, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.97029709815979, 'rewards/repeat_penalty/std': 0.07485032081604004, 'rewards/repeat_floor/mean': 0.9974926710128784, 'rewards/repeat_floor/std': 0.00572331715375185, 'rewards/near_duplicate_penalty/mean': 0.9961960911750793, 'rewards/near_duplicate_penalty/std': 0.0028786661569029093, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.1474762260913849, 'rewards/total_composite/std': 0.1448993980884552, 'reward': 0.1474762260913849, 'reward_std': 0.1448993980884552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22583043575286865, 'sampling/sampling_logp_difference/max': 2.5910415649414062, 'sampling/importance_sampling_ratio/min': 0.07494194060564041, 'sampling/importance_sampling_ratio/mean': 1.029626727104187, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.604298412799835, 'clip_ratio/low_mean': 0.14363864623010159, 'clip_ratio/low_min': 0.14363864623010159, 'clip_ratio/high_mean': 0.04995958134531975, 'clip_ratio/high_max': 0.04995958134531975, 'clip_ratio/region_mean': 0.19359822757542133, 'reward_total_mean': 0.1474762260913849, 'reward_meter_mean': 0.4005318582057953, 'reward_meter_std': 0.3796837627887726, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9792049527168274, 'reward_lexical_plausibility_std': 0.04006021469831467, 'reward_repeat_penalty_mean': 0.97029709815979, 'reward_repeat_penalty_std': 0.07485032081604004, 'reward_repeat_floor_mean': 0.9974926710128784, 'reward_repeat_floor_std': 0.00572331715375185, 'reward_near_duplicate_penalty_mean': 0.9961960911750793, 'reward_near_duplicate_penalty_std': 0.0028786661569029093, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.1474762260913849, 'reward_total_composite_std': 0.1448993980884552, 'epoch': 0.0}

  1%|          | 39/3300 [06:25<9:41:21, 10.70s/it]INFO 04-12 16:04:51 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 40/3300 [06:34<9:23:51, 10.38s/it]2026-04-12 16:05:00,957 | INFO | train_grpo_train | metrics_logged mode=train step=40

                                                   
{'loss': 0.1301, 'grad_norm': 12.409218788146973, 'learning_rate': 9.881818181818182e-06, 'num_tokens': 75996.0, 'completions/mean_length': 70.25, 'completions/min_length': 50.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.36916738748550415, 'rewards/meter/std': 0.3919422924518585, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4662500023841858, 'rewards/judge_quality/std': 0.1743508279323578, 'rewards/repeat_penalty/mean': 0.9277639389038086, 'rewards/repeat_penalty/std': 0.08746004849672318, 'rewards/repeat_floor/mean': 0.9911202192306519, 'rewards/repeat_floor/std': 0.010052579455077648, 'rewards/near_duplicate_penalty/mean': 0.9647765755653381, 'rewards/near_duplicate_penalty/std': 0.029485993087291718, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9600412845611572, 'rewards/distinct_2/std': 0.06480267643928528, 'rewards/total_composite/mean': 0.19634950160980225, 'rewards/total_composite/std': 0.2600710988044739, 'reward': 0.19634950160980225, 'reward_std': 0.2600710690021515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19952291250228882, 'sampling/sampling_logp_difference/max': 1.2535696029663086, 'sampling/importance_sampling_ratio/min': 0.2854839265346527, 'sampling/importance_sampling_ratio/mean': 1.0462639331817627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7411714494228363, 'clip_ratio/low_mean': 0.11788088269531727, 'clip_ratio/low_min': 0.11788088269531727, 'clip_ratio/high_mean': 0.05319135822355747, 'clip_ratio/high_max': 0.05319135822355747, 'clip_ratio/region_mean': 0.17107224091887474, 'reward_total_mean': 0.19634950160980225, 'reward_meter_mean': 0.36916738748550415, 'reward_meter_std': 0.3919422924518585, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9277639389038086, 'reward_repeat_penalty_std': 0.08746004849672318, 'reward_repeat_floor_mean': 0.9911202192306519, 'reward_repeat_floor_std': 0.010052579455077648, 'reward_near_duplicate_penalty_mean': 0.9647765755653381, 'reward_near_duplicate_penalty_std': 0.029485993087291718, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9600412845611572, 'reward_distinct_2_std': 0.06480267643928528, 'reward_judge_quality_mean': 0.4662500023841858, 'reward_judge_quality_std': 0.1743508279323578, 'reward_total_composite_mean': 0.19634950160980225, 'reward_total_composite_std': 0.2600710988044739, 'epoch': 0.0}

  1%|          | 40/3300 [06:34<9:23:51, 10.38s/it]INFO 04-12 16:05:01 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 41/3300 [06:43<8:51:55,  9.79s/it]2026-04-12 16:05:09,387 | INFO | train_grpo_train | metrics_logged mode=train step=41

                                                   
{'loss': 0.0938, 'grad_norm': 22.473737716674805, 'learning_rate': 9.87878787878788e-06, 'num_tokens': 77486.0, 'completions/mean_length': 43.25, 'completions/min_length': 37.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.39646434783935547, 'rewards/meter/std': 0.290080189704895, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.9904521107673645, 'rewards/repeat_penalty/std': 0.014704611152410507, 'rewards/repeat_floor/mean': 0.9985678195953369, 'rewards/repeat_floor/std': 0.0022056829184293747, 'rewards/near_duplicate_penalty/mean': 0.9904521107673645, 'rewards/near_duplicate_penalty/std': 0.014704611152410507, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1848469376564026, 'rewards/total_composite/std': 0.1707412749528885, 'reward': 0.1848469376564026, 'reward_std': 0.1707412749528885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2598678469657898, 'sampling/sampling_logp_difference/max': 1.7882328033447266, 'sampling/importance_sampling_ratio/min': 0.1672554761171341, 'sampling/importance_sampling_ratio/mean': 1.0375196933746338, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.179448515176773, 'clip_ratio/low_mean': 0.10744178667664528, 'clip_ratio/low_min': 0.10744178667664528, 'clip_ratio/high_mean': 0.1179304700344801, 'clip_ratio/high_max': 0.1179304700344801, 'clip_ratio/region_mean': 0.22537225671112537, 'reward_total_mean': 0.1848469376564026, 'reward_meter_mean': 0.39646434783935547, 'reward_meter_std': 0.290080189704895, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9904521107673645, 'reward_repeat_penalty_std': 0.014704611152410507, 'reward_repeat_floor_mean': 0.9985678195953369, 'reward_repeat_floor_std': 0.0022056829184293747, 'reward_near_duplicate_penalty_mean': 0.9904521107673645, 'reward_near_duplicate_penalty_std': 0.014704611152410507, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.1848469376564026, 'reward_total_composite_std': 0.1707412749528885, 'epoch': 0.0}

  1%|          | 41/3300 [06:43<8:51:55,  9.79s/it]INFO 04-12 16:05:09 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 42/3300 [06:52<8:39:26,  9.57s/it]2026-04-12 16:05:18,423 | INFO | train_grpo_train | metrics_logged mode=train step=42

                                                   
{'loss': 0.0573, 'grad_norm': 14.14794635772705, 'learning_rate': 9.875757575757576e-06, 'num_tokens': 79246.0, 'completions/mean_length': 54.0, 'completions/min_length': 35.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.4714590013027191, 'rewards/meter/std': 0.4398857653141022, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.943163275718689, 'rewards/lexical_plausibility/std': 0.1121167242527008, 'rewards/judge_quality/mean': 0.5462499856948853, 'rewards/judge_quality/std': 0.33312106132507324, 'rewards/repeat_penalty/mean': 0.9951434135437012, 'rewards/repeat_penalty/std': 0.010327104479074478, 'rewards/repeat_floor/mean': 0.9992715120315552, 'rewards/repeat_floor/std': 0.0015490736113861203, 'rewards/near_duplicate_penalty/mean': 0.9951434135437012, 'rewards/near_duplicate_penalty/std': 0.010327104479074478, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19690018892288208, 'rewards/total_composite/std': 0.22086982429027557, 'reward': 0.19690018892288208, 'reward_std': 0.22086982429027557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20788058638572693, 'sampling/sampling_logp_difference/max': 1.6384973526000977, 'sampling/importance_sampling_ratio/min': 0.1942717581987381, 'sampling/importance_sampling_ratio/mean': 1.048647165298462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9128017276525497, 'clip_ratio/low_mean': 0.1061682146973908, 'clip_ratio/low_min': 0.1061682146973908, 'clip_ratio/high_mean': 0.07374090049415827, 'clip_ratio/high_max': 0.07374090049415827, 'clip_ratio/region_mean': 0.17990911519154906, 'reward_total_mean': 0.19690018892288208, 'reward_meter_mean': 0.4714590013027191, 'reward_meter_std': 0.4398857653141022, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.943163275718689, 'reward_lexical_plausibility_std': 0.1121167242527008, 'reward_repeat_penalty_mean': 0.9951434135437012, 'reward_repeat_penalty_std': 0.010327104479074478, 'reward_repeat_floor_mean': 0.9992715120315552, 'reward_repeat_floor_std': 0.0015490736113861203, 'reward_near_duplicate_penalty_mean': 0.9951434135437012, 'reward_near_duplicate_penalty_std': 0.010327104479074478, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5462499856948853, 'reward_judge_quality_std': 0.33312106132507324, 'reward_total_composite_mean': 0.19690018892288208, 'reward_total_composite_std': 0.22086982429027557, 'epoch': 0.0}

  1%|▏         | 42/3300 [06:52<8:39:26,  9.57s/it]INFO 04-12 16:05:18 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 43/3300 [07:01<8:39:35,  9.57s/it]2026-04-12 16:05:28,009 | INFO | train_grpo_train | metrics_logged mode=train step=43

                                                   
{'loss': 0.0937, 'grad_norm': 13.095500946044922, 'learning_rate': 9.872727272727274e-06, 'num_tokens': 80854.0, 'completions/mean_length': 51.0, 'completions/min_length': 37.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5143081545829773, 'rewards/meter/std': 0.4423011243343353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5149999856948853, 'rewards/judge_quality/std': 0.3022770881652832, 'rewards/repeat_penalty/mean': 0.9940285682678223, 'rewards/repeat_penalty/std': 0.013220340944826603, 'rewards/repeat_floor/mean': 0.9991042613983154, 'rewards/repeat_floor/std': 0.001983057940378785, 'rewards/near_duplicate_penalty/mean': 0.9940285682678223, 'rewards/near_duplicate_penalty/std': 0.013220340944826603, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2419220358133316, 'rewards/total_composite/std': 0.3246799111366272, 'reward': 0.2419220358133316, 'reward_std': 0.3246799111366272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20194099843502045, 'sampling/sampling_logp_difference/max': 2.0696909427642822, 'sampling/importance_sampling_ratio/min': 0.12622478604316711, 'sampling/importance_sampling_ratio/mean': 1.0088589191436768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7163867801427841, 'clip_ratio/low_mean': 0.10294000897556543, 'clip_ratio/low_min': 0.10294000897556543, 'clip_ratio/high_mean': 0.06398474518209696, 'clip_ratio/high_max': 0.06398474518209696, 'clip_ratio/region_mean': 0.1669247541576624, 'reward_total_mean': 0.2419220358133316, 'reward_meter_mean': 0.5143081545829773, 'reward_meter_std': 0.4423011243343353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9940285682678223, 'reward_repeat_penalty_std': 0.013220340944826603, 'reward_repeat_floor_mean': 0.9991042613983154, 'reward_repeat_floor_std': 0.001983057940378785, 'reward_near_duplicate_penalty_mean': 0.9940285682678223, 'reward_near_duplicate_penalty_std': 0.013220340944826603, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5149999856948853, 'reward_judge_quality_std': 0.3022770881652832, 'reward_total_composite_mean': 0.2419220358133316, 'reward_total_composite_std': 0.3246799111366272, 'epoch': 0.0}

  1%|▏         | 43/3300 [07:01<8:39:35,  9.57s/it]INFO 04-12 16:05:28 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 44/3300 [07:12<8:56:57,  9.89s/it]2026-04-12 16:05:38,658 | INFO | train_grpo_train | metrics_logged mode=train step=44

                                                   
{'loss': 0.0317, 'grad_norm': 11.550451278686523, 'learning_rate': 9.869696969696971e-06, 'num_tokens': 83197.0, 'completions/mean_length': 107.875, 'completions/min_length': 90.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.875, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.544021725654602, 'rewards/meter/std': 0.39922788739204407, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9986978769302368, 'rewards/lexical_plausibility/std': 0.0036828548181802034, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9485529661178589, 'rewards/repeat_penalty/std': 0.12392938137054443, 'rewards/repeat_floor/mean': 0.9940599203109741, 'rewards/repeat_floor/std': 0.013579227030277252, 'rewards/near_duplicate_penalty/mean': 0.9828356504440308, 'rewards/near_duplicate_penalty/std': 0.027978375554084778, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9626068472862244, 'rewards/distinct_2/std': 0.10576383024454117, 'rewards/total_composite/mean': 0.1766035556793213, 'rewards/total_composite/std': 0.13081301748752594, 'reward': 0.1766035556793213, 'reward_std': 0.13081301748752594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22205466032028198, 'sampling/sampling_logp_difference/max': 1.7351970672607422, 'sampling/importance_sampling_ratio/min': 0.1763654351234436, 'sampling/importance_sampling_ratio/mean': 1.0308425426483154, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0413359254598618, 'clip_ratio/low_mean': 0.10913269873708487, 'clip_ratio/low_min': 0.10913269873708487, 'clip_ratio/high_mean': 0.08570929616689682, 'clip_ratio/high_max': 0.08570929616689682, 'clip_ratio/region_mean': 0.19484199490398169, 'reward_total_mean': 0.1766035556793213, 'reward_meter_mean': 0.544021725654602, 'reward_meter_std': 0.39922788739204407, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9986978769302368, 'reward_lexical_plausibility_std': 0.0036828548181802034, 'reward_repeat_penalty_mean': 0.9485529661178589, 'reward_repeat_penalty_std': 0.12392938137054443, 'reward_repeat_floor_mean': 0.9940599203109741, 'reward_repeat_floor_std': 0.013579227030277252, 'reward_near_duplicate_penalty_mean': 0.9828356504440308, 'reward_near_duplicate_penalty_std': 0.027978375554084778, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9626068472862244, 'reward_distinct_2_std': 0.10576383024454117, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1766035556793213, 'reward_total_composite_std': 0.13081301748752594, 'epoch': 0.0}

  1%|▏         | 44/3300 [07:12<8:56:57,  9.89s/it]INFO 04-12 16:05:38 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 45/3300 [07:22<9:00:02,  9.95s/it]2026-04-12 16:05:48,751 | INFO | train_grpo_train | metrics_logged mode=train step=45

                                                   
{'loss': 0.1632, 'grad_norm': 17.604976654052734, 'learning_rate': 9.866666666666668e-06, 'num_tokens': 85068.0, 'completions/mean_length': 51.875, 'completions/min_length': 28.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.2960636019706726, 'rewards/meter/std': 0.32714810967445374, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9284577965736389, 'rewards/lexical_plausibility/std': 0.13728460669517517, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.9990079402923584, 'rewards/repeat_penalty/std': 0.0028059897013008595, 'rewards/repeat_floor/mean': 0.9998512268066406, 'rewards/repeat_floor/std': 0.00042089950875379145, 'rewards/near_duplicate_penalty/mean': 0.9990079402923584, 'rewards/near_duplicate_penalty/std': 0.0028059897013008595, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08450189232826233, 'rewards/total_composite/std': 0.08134256303310394, 'reward': 0.08450189232826233, 'reward_std': 0.08134256303310394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2405218482017517, 'sampling/sampling_logp_difference/max': 1.4257278442382812, 'sampling/importance_sampling_ratio/min': 0.24033348262310028, 'sampling/importance_sampling_ratio/mean': 1.0540187358856201, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0648261457681656, 'clip_ratio/low_mean': 0.13810977526009083, 'clip_ratio/low_min': 0.13810977526009083, 'clip_ratio/high_mean': 0.06894841324537992, 'clip_ratio/high_max': 0.06894841324537992, 'clip_ratio/region_mean': 0.20705818850547075, 'reward_total_mean': 0.08450189232826233, 'reward_meter_mean': 0.2960636019706726, 'reward_meter_std': 0.32714810967445374, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9284577965736389, 'reward_lexical_plausibility_std': 0.13728460669517517, 'reward_repeat_penalty_mean': 0.9990079402923584, 'reward_repeat_penalty_std': 0.0028059897013008595, 'reward_repeat_floor_mean': 0.9998512268066406, 'reward_repeat_floor_std': 0.00042089950875379145, 'reward_near_duplicate_penalty_mean': 0.9990079402923584, 'reward_near_duplicate_penalty_std': 0.0028059897013008595, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.08450189232826233, 'reward_total_composite_std': 0.08134256303310394, 'epoch': 0.0}

  1%|▏         | 45/3300 [07:22<9:00:02,  9.95s/it]INFO 04-12 16:05:48 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 46/3300 [07:32<9:01:05,  9.98s/it]2026-04-12 16:05:58,779 | INFO | train_grpo_train | metrics_logged mode=train step=46

                                                   
{'loss': 0.0327, 'grad_norm': 10.066476821899414, 'learning_rate': 9.863636363636364e-06, 'num_tokens': 87623.0, 'completions/mean_length': 123.375, 'completions/min_length': 92.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.375, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.47935622930526733, 'rewards/meter/std': 0.39885953068733215, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9589699506759644, 'rewards/lexical_plausibility/std': 0.10753607749938965, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9956408739089966, 'rewards/repeat_penalty/std': 0.004476652015000582, 'rewards/repeat_floor/mean': 0.999346137046814, 'rewards/repeat_floor/std': 0.0006714908522553742, 'rewards/near_duplicate_penalty/mean': 0.9956408739089966, 'rewards/near_duplicate_penalty/std': 0.004476652015000582, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1525011956691742, 'rewards/total_composite/std': 0.1323809027671814, 'reward': 0.1525011956691742, 'reward_std': 0.1323809027671814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21697646379470825, 'sampling/sampling_logp_difference/max': 1.9224927425384521, 'sampling/importance_sampling_ratio/min': 0.1975652426481247, 'sampling/importance_sampling_ratio/mean': 1.0245466232299805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.017793223261833, 'clip_ratio/low_mean': 0.1330940742045641, 'clip_ratio/low_min': 0.1330940742045641, 'clip_ratio/high_mean': 0.08683452941477299, 'clip_ratio/high_max': 0.08683452941477299, 'clip_ratio/region_mean': 0.21992860361933708, 'reward_total_mean': 0.1525011956691742, 'reward_meter_mean': 0.47935622930526733, 'reward_meter_std': 0.39885953068733215, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9589699506759644, 'reward_lexical_plausibility_std': 0.10753607749938965, 'reward_repeat_penalty_mean': 0.9956408739089966, 'reward_repeat_penalty_std': 0.004476652015000582, 'reward_repeat_floor_mean': 0.999346137046814, 'reward_repeat_floor_std': 0.0006714908522553742, 'reward_near_duplicate_penalty_mean': 0.9956408739089966, 'reward_near_duplicate_penalty_std': 0.004476652015000582, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.1525011956691742, 'reward_total_composite_std': 0.1323809027671814, 'epoch': 0.0}

  1%|▏         | 46/3300 [07:32<9:01:05,  9.98s/it]INFO 04-12 16:05:59 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 47/3300 [07:40<8:30:35,  9.42s/it]2026-04-12 16:06:06,893 | INFO | train_grpo_train | metrics_logged mode=train step=47

                                                   
{'loss': 0.0981, 'grad_norm': 19.66780662536621, 'learning_rate': 9.860606060606061e-06, 'num_tokens': 89279.0, 'completions/mean_length': 47.0, 'completions/min_length': 34.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6978145241737366, 'rewards/meter/std': 0.3271360695362091, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9867098331451416, 'rewards/repeat_penalty/std': 0.02383718639612198, 'rewards/repeat_floor/mean': 0.9985309839248657, 'rewards/repeat_floor/std': 0.0021909738425165415, 'rewards/near_duplicate_penalty/mean': 0.9954510927200317, 'rewards/near_duplicate_penalty/std': 0.006915302015841007, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.2271214872598648, 'rewards/total_composite/std': 0.12039513885974884, 'reward': 0.2271214872598648, 'reward_std': 0.12039513140916824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24266086518764496, 'sampling/sampling_logp_difference/max': 1.4173450469970703, 'sampling/importance_sampling_ratio/min': 0.24235661327838898, 'sampling/importance_sampling_ratio/mean': 1.0233699083328247, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.640699565410614, 'clip_ratio/low_mean': 0.08001583628356457, 'clip_ratio/low_min': 0.08001583628356457, 'clip_ratio/high_mean': 0.16471003741025925, 'clip_ratio/high_max': 0.16471003741025925, 'clip_ratio/region_mean': 0.24472587369382381, 'reward_total_mean': 0.2271214872598648, 'reward_meter_mean': 0.6978145241737366, 'reward_meter_std': 0.3271360695362091, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.9867098331451416, 'reward_repeat_penalty_std': 0.02383718639612198, 'reward_repeat_floor_mean': 0.9985309839248657, 'reward_repeat_floor_std': 0.0021909738425165415, 'reward_near_duplicate_penalty_mean': 0.9954510927200317, 'reward_near_duplicate_penalty_std': 0.006915302015841007, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.2271214872598648, 'reward_total_composite_std': 0.12039513885974884, 'epoch': 0.0}

  1%|▏         | 47/3300 [07:40<8:30:35,  9.42s/it]INFO 04-12 16:06:07 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 48/3300 [07:52<9:06:06, 10.08s/it]2026-04-12 16:06:18,504 | INFO | train_grpo_train | metrics_logged mode=train step=48

                                                   
{'loss': 0.0175, 'grad_norm': 18.717687606811523, 'learning_rate': 9.857575757575758e-06, 'num_tokens': 91807.0, 'completions/mean_length': 110.0, 'completions/min_length': 77.0, 'completions/max_length': 174.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 174.0, 'rewards/meter/mean': 0.5925639271736145, 'rewards/meter/std': 0.2528626024723053, 'rewards/count_adherence/mean': 0.9821428656578064, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.9946428537368774, 'rewards/count_floor/std': 0.015152297914028168, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9914956092834473, 'rewards/repeat_penalty/std': 0.002611305331811309, 'rewards/repeat_floor/mean': 0.9987243413925171, 'rewards/repeat_floor/std': 0.0003916955611202866, 'rewards/near_duplicate_penalty/mean': 0.9914956092834473, 'rewards/near_duplicate_penalty/std': 0.002611305331811309, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23373478651046753, 'rewards/total_composite/std': 0.10138335824012756, 'reward': 0.23373478651046753, 'reward_std': 0.10138335078954697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.3012349009513855, 'sampling/sampling_logp_difference/max': 2.999422073364258, 'sampling/importance_sampling_ratio/min': 0.04981584846973419, 'sampling/importance_sampling_ratio/mean': 0.9902380704879761, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4447583556175232, 'clip_ratio/low_mean': 0.15489736385643482, 'clip_ratio/low_min': 0.15489736385643482, 'clip_ratio/high_mean': 0.09630028530955315, 'clip_ratio/high_max': 0.09630028530955315, 'clip_ratio/region_mean': 0.25119764916598797, 'reward_total_mean': 0.23373478651046753, 'reward_meter_mean': 0.5925639271736145, 'reward_meter_std': 0.2528626024723053, 'reward_count_adherence_mean': 0.9821428656578064, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.9946428537368774, 'reward_count_floor_std': 0.015152297914028168, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9914956092834473, 'reward_repeat_penalty_std': 0.002611305331811309, 'reward_repeat_floor_mean': 0.9987243413925171, 'reward_repeat_floor_std': 0.0003916955611202866, 'reward_near_duplicate_penalty_mean': 0.9914956092834473, 'reward_near_duplicate_penalty_std': 0.002611305331811309, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.23373478651046753, 'reward_total_composite_std': 0.10138335824012756, 'epoch': 0.0}

  1%|▏         | 48/3300 [07:52<9:06:06, 10.08s/it]INFO 04-12 16:06:18 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 49/3300 [08:01<8:52:40,  9.83s/it]2026-04-12 16:06:27,763 | INFO | train_grpo_train | metrics_logged mode=train step=49

                                                   
{'loss': 0.3139, 'grad_norm': 10.536018371582031, 'learning_rate': 9.854545454545456e-06, 'num_tokens': 93540.0, 'completions/mean_length': 73.625, 'completions/min_length': 35.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.5893219709396362, 'rewards/meter/std': 0.42576706409454346, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 0.9814241528511047, 'rewards/lexical_plausibility/std': 0.052540455013513565, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.9992559552192688, 'rewards/repeat_penalty/std': 0.002104476559907198, 'rewards/repeat_floor/mean': 0.9998884201049805, 'rewards/repeat_floor/std': 0.00031567990663461387, 'rewards/near_duplicate_penalty/mean': 0.9992559552192688, 'rewards/near_duplicate_penalty/std': 0.002104476559907198, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20694305002689362, 'rewards/total_composite/std': 0.2326822578907013, 'reward': 0.20694305002689362, 'reward_std': 0.2326822429895401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2126588523387909, 'sampling/sampling_logp_difference/max': 1.9608492851257324, 'sampling/importance_sampling_ratio/min': 0.140738844871521, 'sampling/importance_sampling_ratio/mean': 1.036598563194275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3167301416397095, 'clip_ratio/low_mean': 0.1291656717658043, 'clip_ratio/low_min': 0.1291656717658043, 'clip_ratio/high_mean': 0.07552650291472673, 'clip_ratio/high_max': 0.07552650291472673, 'clip_ratio/region_mean': 0.20469217468053102, 'reward_total_mean': 0.20694305002689362, 'reward_meter_mean': 0.5893219709396362, 'reward_meter_std': 0.42576706409454346, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 0.9814241528511047, 'reward_lexical_plausibility_std': 0.052540455013513565, 'reward_repeat_penalty_mean': 0.9992559552192688, 'reward_repeat_penalty_std': 0.002104476559907198, 'reward_repeat_floor_mean': 0.9998884201049805, 'reward_repeat_floor_std': 0.00031567990663461387, 'reward_near_duplicate_penalty_mean': 0.9992559552192688, 'reward_near_duplicate_penalty_std': 0.002104476559907198, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.20694305002689362, 'reward_total_composite_std': 0.2326822578907013, 'epoch': 0.0}

  1%|▏         | 49/3300 [08:01<8:52:40,  9.83s/it]INFO 04-12 16:06:27 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 50/3300 [08:10<8:40:58,  9.62s/it]2026-04-12 16:06:36,886 | INFO | train_grpo_train | metrics_logged mode=train step=50

                                                   
{'loss': 0.0626, 'grad_norm': 14.482155799865723, 'learning_rate': 9.851515151515151e-06, 'num_tokens': 95434.0, 'completions/mean_length': 54.75, 'completions/min_length': 40.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.8728687167167664, 'rewards/meter/std': 0.22255796194076538, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9908762574195862, 'rewards/lexical_plausibility/std': 0.02509334124624729, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9907007813453674, 'rewards/repeat_penalty/std': 0.013375191949307919, 'rewards/repeat_floor/mean': 0.9986051321029663, 'rewards/repeat_floor/std': 0.0020062841940671206, 'rewards/near_duplicate_penalty/mean': 0.9907007813453674, 'rewards/near_duplicate_penalty/std': 0.013375191949307919, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4573856294155121, 'rewards/total_composite/std': 0.303283154964447, 'reward': 0.4573856294155121, 'reward_std': 0.303283154964447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20837847888469696, 'sampling/sampling_logp_difference/max': 1.2507715225219727, 'sampling/importance_sampling_ratio/min': 0.28628382086753845, 'sampling/importance_sampling_ratio/mean': 1.0178985595703125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.950836792588234, 'clip_ratio/low_mean': 0.1635638177394867, 'clip_ratio/low_min': 0.1635638177394867, 'clip_ratio/high_mean': 0.05538522079586983, 'clip_ratio/high_max': 0.05538522079586983, 'clip_ratio/region_mean': 0.21894903853535652, 'reward_total_mean': 0.4573856294155121, 'reward_meter_mean': 0.8728687167167664, 'reward_meter_std': 0.22255796194076538, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9908762574195862, 'reward_lexical_plausibility_std': 0.02509334124624729, 'reward_repeat_penalty_mean': 0.9907007813453674, 'reward_repeat_penalty_std': 0.013375191949307919, 'reward_repeat_floor_mean': 0.9986051321029663, 'reward_repeat_floor_std': 0.0020062841940671206, 'reward_near_duplicate_penalty_mean': 0.9907007813453674, 'reward_near_duplicate_penalty_std': 0.013375191949307919, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.4573856294155121, 'reward_total_composite_std': 0.303283154964447, 'epoch': 0.0}

  2%|▏         | 50/3300 [08:10<8:40:58,  9.62s/it]INFO 04-12 16:06:37 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:10<00:56,  5.10s/it]

 23%|██▎       | 3/13 [00:17<01:00,  6.03s/it]

 31%|███       | 4/13 [00:27<01:08,  7.60s/it]

 38%|███▊      | 5/13 [00:38<01:08,  8.57s/it]

 46%|████▌     | 6/13 [00:47<01:01,  8.76s/it]

 54%|█████▍    | 7/13 [00:55<00:52,  8.73s/it]

 62%|██████▏   | 8/13 [01:07<00:48,  9.76s/it]

 69%|██████▉   | 9/13 [01:19<00:40, 10.19s/it]

 77%|███████▋  | 10/13 [01:31<00:32, 10.72s/it]

 85%|████████▍ | 11/13 [01:41<00:21, 10.64s/it]

 92%|█████████▏| 12/13 [01:52<00:10, 10.82s/it]

100%|██████████| 13/13 [02:03<00:00, 10.88s/it]2026-04-12 16:08:51,138 | INFO | train_grpo_train | metrics_logged mode=eval step=50

                                                   

                                               
{'eval_loss': nan, 'eval_runtime': 134.2424, 'eval_samples_per_second': 0.775, 'eval_steps_per_second': 0.097, 'eval_num_tokens': 95434.0, 'eval_completions/mean_length': 177.29807692307693, 'eval_completions/min_length': 44.38461538461539, 'eval_completions/max_length': 405.7692307692308, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 170.3804955115685, 'eval_completions/min_terminated_length': 44.38461538461539, 'eval_completions/max_terminated_length': 369.61538461538464, 'eval_rewards/meter/mean': 0.445071117236064, 'eval_rewards/meter/std': 0.34558682029063886, 'eval_rewards/count_adherence/mean': 0.9169511153147771, 'eval_rewards/count_adherence/std': 0.0925693979056982, 'eval_rewards/arabic_clean/mean': 0.8942307692307693, 'eval_rewards/arabic_clean/std': 0.2573827917759235, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9750853272584769, 'eval_rewards/count_floor/std': 0.027770821793148152, 'eval_rewards/lexical_plausibility/mean': 0.9905767853443439, 'eval_rewards/lexical_plausibility/std': 0.02665287284896924, 'eval_rewards/judge_quality/mean': 0.40894230283223665, 'eval_rewards/judge_quality/std': 0.1612841243354174, 'eval_rewards/repeat_penalty/mean': 0.9728505198772137, 'eval_rewards/repeat_penalty/std': 0.04411904696518412, 'eval_rewards/repeat_floor/mean': 0.9967490618045514, 'eval_rewards/repeat_floor/std': 0.00508836470096587, 'eval_rewards/near_duplicate_penalty/mean': 0.9845332365769607, 'eval_rewards/near_duplicate_penalty/std': 0.025225410219998315, 'eval_rewards/opening_diversity/mean': 0.9927884615384616, 'eval_rewards/opening_diversity/std': 0.020397310646680687, 'eval_rewards/distinct_2/mean': 0.9944640352175786, 'eval_rewards/distinct_2/std': 0.010853729497354764, 'eval_rewards/total_composite/mean': 0.1737233320107827, 'eval_rewards/total_composite/std': 0.1642189031610122, 'eval_reward': 0.1737233320107827, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.13728128946744478, 'eval_sampling/sampling_logp_difference/max': 1.2141184440025916, 'eval_sampling/importance_sampling_ratio/min': 0.3012317717075348, 'eval_sampling/importance_sampling_ratio/mean': 1.0408150782951942, 'eval_sampling/importance_sampling_ratio/max': 1.6635784644346971, 'eval_entropy': 2.2389675012001624, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1737233320107827, 'eval_reward_meter_mean': 0.445071117236064, 'eval_reward_meter_std': 0.34558682029063886, 'eval_reward_count_adherence_mean': 0.9169511153147771, 'eval_reward_count_adherence_std': 0.0925693979056982, 'eval_reward_arabic_clean_mean': 0.8942307692307693, 'eval_reward_arabic_clean_std': 0.2573827917759235, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9750853272584769, 'eval_reward_count_floor_std': 0.027770821793148152, 'eval_reward_lexical_plausibility_mean': 0.9905767853443439, 'eval_reward_lexical_plausibility_std': 0.02665287284896924, 'eval_reward_repeat_penalty_mean': 0.9728505198772137, 'eval_reward_repeat_penalty_std': 0.04411904696518412, 'eval_reward_repeat_floor_mean': 0.9967490618045514, 'eval_reward_repeat_floor_std': 0.00508836470096587, 'eval_reward_near_duplicate_penalty_mean': 0.9845332365769607, 'eval_reward_near_duplicate_penalty_std': 0.025225410219998315, 'eval_reward_opening_diversity_mean': 0.9927884615384616, 'eval_reward_opening_diversity_std': 0.020397310646680687, 'eval_reward_distinct_2_mean': 0.9944640352175786, 'eval_reward_distinct_2_std': 0.010853729497354764, 'eval_reward_judge_quality_mean': 0.40894230283223665, 'eval_reward_judge_quality_std': 0.1612841243354174, 'eval_reward_total_composite_mean': 0.1737233320107827, 'eval_reward_total_composite_std': 0.1642189031610122, 'epoch': 0.0}

  2%|▏         | 50/3300 [10:24<8:40:58,  9.62s/it]

100%|██████████| 13/13 [02:03<00:00, 10.88s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  2%|▏         | 51/3300 [10:37<45:43:16, 50.66s/it]2026-04-12 16:09:03,313 | INFO | train_grpo_train | metrics_logged mode=train step=51

                                                    
{'loss': -0.0176, 'grad_norm': 10.391440391540527, 'learning_rate': 9.84848484848485e-06, 'num_tokens': 97567.0, 'completions/mean_length': 86.625, 'completions/min_length': 59.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.625, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.7797682285308838, 'rewards/meter/std': 0.2274528443813324, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.932199239730835, 'rewards/lexical_plausibility/std': 0.11454705148935318, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9820003509521484, 'rewards/repeat_penalty/std': 0.025386661291122437, 'rewards/repeat_floor/mean': 0.9980157613754272, 'rewards/repeat_floor/std': 0.002557407831773162, 'rewards/near_duplicate_penalty/mean': 0.9940996170043945, 'rewards/near_duplicate_penalty/std': 0.007026705890893936, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9877868890762329, 'rewards/distinct_2/std': 0.022618044167757034, 'rewards/total_composite/mean': 0.24208964407444, 'rewards/total_composite/std': 0.1124430000782013, 'reward': 0.24208964407444, 'reward_std': 0.1124429851770401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21936972439289093, 'sampling/sampling_logp_difference/max': 1.5279293060302734, 'sampling/importance_sampling_ratio/min': 0.21698451042175293, 'sampling/importance_sampling_ratio/mean': 1.0266475677490234, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.54919296503067, 'clip_ratio/low_mean': 0.10702678561210632, 'clip_ratio/low_min': 0.10702678561210632, 'clip_ratio/high_mean': 0.1289989035576582, 'clip_ratio/high_max': 0.1289989035576582, 'clip_ratio/region_mean': 0.23602568916976452, 'reward_total_mean': 0.24208964407444, 'reward_meter_mean': 0.7797682285308838, 'reward_meter_std': 0.2274528443813324, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.932199239730835, 'reward_lexical_plausibility_std': 0.11454705148935318, 'reward_repeat_penalty_mean': 0.9820003509521484, 'reward_repeat_penalty_std': 0.025386661291122437, 'reward_repeat_floor_mean': 0.9980157613754272, 'reward_repeat_floor_std': 0.002557407831773162, 'reward_near_duplicate_penalty_mean': 0.9940996170043945, 'reward_near_duplicate_penalty_std': 0.007026705890893936, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9877868890762329, 'reward_distinct_2_std': 0.022618044167757034, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.24208964407444, 'reward_total_composite_std': 0.1124430000782013, 'epoch': 0.0}

  2%|▏         | 51/3300 [10:37<45:43:16, 50.66s/it]INFO 04-12 16:09:03 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 52/3300 [10:46<34:36:23, 38.36s/it]2026-04-12 16:09:12,961 | INFO | train_grpo_train | metrics_logged mode=train step=52

                                                    
{'loss': 0.1254, 'grad_norm': 10.570046424865723, 'learning_rate': 9.845454545454546e-06, 'num_tokens': 99438.0, 'completions/mean_length': 75.875, 'completions/min_length': 53.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9224824905395508, 'rewards/meter/std': 0.14314211905002594, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9236583113670349, 'rewards/lexical_plausibility/std': 0.10479297488927841, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.2773052752017975, 'rewards/repeat_penalty/mean': 0.9874131083488464, 'rewards/repeat_penalty/std': 0.02840070240199566, 'rewards/repeat_floor/mean': 0.998478889465332, 'rewards/repeat_floor/std': 0.003229753579944372, 'rewards/near_duplicate_penalty/mean': 0.9938383102416992, 'rewards/near_duplicate_penalty/std': 0.010502070188522339, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.2647750675678253, 'rewards/total_composite/std': 0.18174003064632416, 'reward': 0.2647750675678253, 'reward_std': 0.18174003064632416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.197578564286232, 'sampling/sampling_logp_difference/max': 1.3047370910644531, 'sampling/importance_sampling_ratio/min': 0.2712438404560089, 'sampling/importance_sampling_ratio/mean': 1.0037709474563599, 'sampling/importance_sampling_ratio/max': 1.9497319459915161, 'entropy': 2.3698359429836273, 'clip_ratio/low_mean': 0.11267345305532217, 'clip_ratio/low_min': 0.11267345305532217, 'clip_ratio/high_mean': 0.08525996468961239, 'clip_ratio/high_max': 0.08525996468961239, 'clip_ratio/region_mean': 0.19793341774493456, 'reward_total_mean': 0.2647750675678253, 'reward_meter_mean': 0.9224824905395508, 'reward_meter_std': 0.14314211905002594, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9236583113670349, 'reward_lexical_plausibility_std': 0.10479297488927841, 'reward_repeat_penalty_mean': 0.9874131083488464, 'reward_repeat_penalty_std': 0.02840070240199566, 'reward_repeat_floor_mean': 0.998478889465332, 'reward_repeat_floor_std': 0.003229753579944372, 'reward_near_duplicate_penalty_mean': 0.9938383102416992, 'reward_near_duplicate_penalty_std': 0.010502070188522339, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.2773052752017975, 'reward_total_composite_mean': 0.2647750675678253, 'reward_total_composite_std': 0.18174003064632416, 'epoch': 0.0}

  2%|▏         | 52/3300 [10:46<34:36:23, 38.36s/it]INFO 04-12 16:09:13 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 53/3300 [10:56<26:47:57, 29.71s/it]2026-04-12 16:09:22,504 | INFO | train_grpo_train | metrics_logged mode=train step=53

                                                    
{'loss': 0.103, 'grad_norm': 10.738615036010742, 'learning_rate': 9.842424242424243e-06, 'num_tokens': 101564.0, 'completions/mean_length': 82.75, 'completions/min_length': 55.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.6636631488800049, 'rewards/meter/std': 0.4369257092475891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9588133096694946, 'rewards/lexical_plausibility/std': 0.07631121575832367, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9947271347045898, 'rewards/repeat_penalty/std': 0.004523387644439936, 'rewards/repeat_floor/mean': 0.9992091059684753, 'rewards/repeat_floor/std': 0.0006784949800930917, 'rewards/near_duplicate_penalty/mean': 0.9947271347045898, 'rewards/near_duplicate_penalty/std': 0.004523387644439936, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29524222016334534, 'rewards/total_composite/std': 0.292190819978714, 'reward': 0.29524222016334534, 'reward_std': 0.292190819978714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2270900309085846, 'sampling/sampling_logp_difference/max': 3.872091293334961, 'sampling/importance_sampling_ratio/min': 0.020814793184399605, 'sampling/importance_sampling_ratio/mean': 1.0320154428482056, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.151434540748596, 'clip_ratio/low_mean': 0.1154552698135376, 'clip_ratio/low_min': 0.1154552698135376, 'clip_ratio/high_mean': 0.10241159051656723, 'clip_ratio/high_max': 0.10241159051656723, 'clip_ratio/region_mean': 0.21786686033010483, 'reward_total_mean': 0.29524222016334534, 'reward_meter_mean': 0.6636631488800049, 'reward_meter_std': 0.4369257092475891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9588133096694946, 'reward_lexical_plausibility_std': 0.07631121575832367, 'reward_repeat_penalty_mean': 0.9947271347045898, 'reward_repeat_penalty_std': 0.004523387644439936, 'reward_repeat_floor_mean': 0.9992091059684753, 'reward_repeat_floor_std': 0.0006784949800930917, 'reward_near_duplicate_penalty_mean': 0.9947271347045898, 'reward_near_duplicate_penalty_std': 0.004523387644439936, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.29524222016334534, 'reward_total_composite_std': 0.292190819978714, 'epoch': 0.0}

  2%|▏         | 53/3300 [10:56<26:47:57, 29.71s/it]INFO 04-12 16:09:22 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 54/3300 [11:04<20:56:07, 23.22s/it]2026-04-12 16:09:30,570 | INFO | train_grpo_train | metrics_logged mode=train step=54

                                                    
{'loss': -0.0546, 'grad_norm': 31.716785430908203, 'learning_rate': 9.83939393939394e-06, 'num_tokens': 102989.0, 'completions/mean_length': 23.125, 'completions/min_length': 18.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8562110662460327, 'rewards/meter/std': 0.2180831879377365, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7562500238418579, 'rewards/judge_quality/std': 0.31717222929000854, 'rewards/repeat_penalty/mean': 0.7398061752319336, 'rewards/repeat_penalty/std': 0.028832554817199707, 'rewards/repeat_floor/mean': 0.9829612374305725, 'rewards/repeat_floor/std': 0.005766510963439941, 'rewards/near_duplicate_penalty/mean': 0.9864082336425781, 'rewards/near_duplicate_penalty/std': 0.0384434312582016, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6877180337905884, 'rewards/total_composite/std': 0.3359270691871643, 'reward': 0.6877180337905884, 'reward_std': 0.3359270691871643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15803538262844086, 'sampling/sampling_logp_difference/max': 1.237335205078125, 'sampling/importance_sampling_ratio/min': 0.29015639424324036, 'sampling/importance_sampling_ratio/mean': 1.0263099670410156, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5656054466962814, 'clip_ratio/low_mean': 0.04255952499806881, 'clip_ratio/low_min': 0.04255952499806881, 'clip_ratio/high_mean': 0.13260012725368142, 'clip_ratio/high_max': 0.13260012725368142, 'clip_ratio/region_mean': 0.17515965225175023, 'reward_total_mean': 0.6877180337905884, 'reward_meter_mean': 0.8562110662460327, 'reward_meter_std': 0.2180831879377365, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7398061752319336, 'reward_repeat_penalty_std': 0.028832554817199707, 'reward_repeat_floor_mean': 0.9829612374305725, 'reward_repeat_floor_std': 0.005766510963439941, 'reward_near_duplicate_penalty_mean': 0.9864082336425781, 'reward_near_duplicate_penalty_std': 0.0384434312582016, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7562500238418579, 'reward_judge_quality_std': 0.31717222929000854, 'reward_total_composite_mean': 0.6877180337905884, 'reward_total_composite_std': 0.3359270691871643, 'epoch': 0.0}

  2%|▏         | 54/3300 [11:04<20:56:07, 23.22s/it]INFO 04-12 16:09:30 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 55/3300 [11:12<16:49:46, 18.67s/it]2026-04-12 16:09:38,628 | INFO | train_grpo_train | metrics_logged mode=train step=55

                                                    
{'loss': 0.0867, 'grad_norm': 23.254289627075195, 'learning_rate': 9.836363636363637e-06, 'num_tokens': 104420.0, 'completions/mean_length': 33.875, 'completions/min_length': 26.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.27538812160491943, 'rewards/meter/std': 0.2869819700717926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9988390207290649, 'rewards/lexical_plausibility/std': 0.0032837872859090567, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9970964193344116, 'rewards/repeat_penalty/std': 0.0052824467420578, 'rewards/repeat_floor/mean': 0.9995644688606262, 'rewards/repeat_floor/std': 0.0007923676166683435, 'rewards/near_duplicate_penalty/mean': 0.9970964193344116, 'rewards/near_duplicate_penalty/std': 0.0052824467420578, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10117396712303162, 'rewards/total_composite/std': 0.09911399334669113, 'reward': 0.10117396712303162, 'reward_std': 0.09911399334669113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24441500008106232, 'sampling/sampling_logp_difference/max': 1.9817352294921875, 'sampling/importance_sampling_ratio/min': 0.13782986998558044, 'sampling/importance_sampling_ratio/mean': 1.0106151103973389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6633601784706116, 'clip_ratio/low_mean': 0.13752583041787148, 'clip_ratio/low_min': 0.13752583041787148, 'clip_ratio/high_mean': 0.07858373317867517, 'clip_ratio/high_max': 0.07858373317867517, 'clip_ratio/region_mean': 0.21610956359654665, 'reward_total_mean': 0.10117396712303162, 'reward_meter_mean': 0.27538812160491943, 'reward_meter_std': 0.2869819700717926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9988390207290649, 'reward_lexical_plausibility_std': 0.0032837872859090567, 'reward_repeat_penalty_mean': 0.9970964193344116, 'reward_repeat_penalty_std': 0.0052824467420578, 'reward_repeat_floor_mean': 0.9995644688606262, 'reward_repeat_floor_std': 0.0007923676166683435, 'reward_near_duplicate_penalty_mean': 0.9970964193344116, 'reward_near_duplicate_penalty_std': 0.0052824467420578, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.10117396712303162, 'reward_total_composite_std': 0.09911399334669113, 'epoch': 0.0}

  2%|▏         | 55/3300 [11:12<16:49:46, 18.67s/it]INFO 04-12 16:09:38 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 56/3300 [11:21<14:15:11, 15.82s/it]2026-04-12 16:09:47,788 | INFO | train_grpo_train | metrics_logged mode=train step=56

                                                    
{'loss': -0.1011, 'grad_norm': 16.747817993164062, 'learning_rate': 9.833333333333333e-06, 'num_tokens': 106204.0, 'completions/mean_length': 51.0, 'completions/min_length': 25.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.541499137878418, 'rewards/meter/std': 0.4344416856765747, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.964999258518219, 'rewards/repeat_penalty/std': 0.08750505745410919, 'rewards/repeat_floor/mean': 0.9975624084472656, 'rewards/repeat_floor/std': 0.005314696114510298, 'rewards/near_duplicate_penalty/mean': 0.996249258518219, 'rewards/near_duplicate_penalty/std': 0.010608699172735214, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31828606128692627, 'rewards/total_composite/std': 0.33471304178237915, 'reward': 0.31828606128692627, 'reward_std': 0.33471301198005676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20916934311389923, 'sampling/sampling_logp_difference/max': 1.670210838317871, 'sampling/importance_sampling_ratio/min': 0.18820738792419434, 'sampling/importance_sampling_ratio/mean': 1.0528364181518555, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.318334147334099, 'clip_ratio/low_mean': 0.09745830856263638, 'clip_ratio/low_min': 0.09745830856263638, 'clip_ratio/high_mean': 0.06373097561299801, 'clip_ratio/high_max': 0.06373097561299801, 'clip_ratio/region_mean': 0.16118928417563438, 'reward_total_mean': 0.31828606128692627, 'reward_meter_mean': 0.541499137878418, 'reward_meter_std': 0.4344416856765747, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.964999258518219, 'reward_repeat_penalty_std': 0.08750505745410919, 'reward_repeat_floor_mean': 0.9975624084472656, 'reward_repeat_floor_std': 0.005314696114510298, 'reward_near_duplicate_penalty_mean': 0.996249258518219, 'reward_near_duplicate_penalty_std': 0.010608699172735214, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.31828606128692627, 'reward_total_composite_std': 0.33471304178237915, 'epoch': 0.0}

  2%|▏         | 56/3300 [11:21<14:15:11, 15.82s/it]INFO 04-12 16:09:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 57/3300 [11:34<13:26:58, 14.93s/it]2026-04-12 16:10:00,649 | INFO | train_grpo_train | metrics_logged mode=train step=57

                                                    
{'loss': 0.0134, 'grad_norm': 4.642459869384766, 'learning_rate': 9.830303030303032e-06, 'num_tokens': 111258.0, 'completions/mean_length': 400.75, 'completions/min_length': 307.0, 'completions/max_length': 474.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 400.75, 'completions/min_terminated_length': 307.0, 'completions/max_terminated_length': 474.0, 'rewards/meter/mean': 0.9013170003890991, 'rewards/meter/std': 0.08074897527694702, 'rewards/count_adherence/mean': 0.8636363744735718, 'rewards/count_adherence/std': 0.04859296977519989, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9590909481048584, 'rewards/count_floor/std': 0.014577903784811497, 'rewards/lexical_plausibility/mean': 0.983328104019165, 'rewards/lexical_plausibility/std': 0.0458899661898613, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.09425459802150726, 'rewards/repeat_penalty/mean': 0.9893382787704468, 'rewards/repeat_penalty/std': 0.009818924590945244, 'rewards/repeat_floor/mean': 0.9988276958465576, 'rewards/repeat_floor/std': 0.0009705271222628653, 'rewards/near_duplicate_penalty/mean': 0.9965053200721741, 'rewards/near_duplicate_penalty/std': 0.0016852014232426882, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9927985668182373, 'rewards/distinct_2/std': 0.00856027565896511, 'rewards/total_composite/mean': 0.13366711139678955, 'rewards/total_composite/std': 0.08221367001533508, 'reward': 0.13366711139678955, 'reward_std': 0.08221367001533508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20859874784946442, 'sampling/sampling_logp_difference/max': 2.716672897338867, 'sampling/importance_sampling_ratio/min': 0.0660942941904068, 'sampling/importance_sampling_ratio/mean': 1.037036657333374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8429089784622192, 'clip_ratio/low_mean': 0.11098390258848667, 'clip_ratio/low_min': 0.11098390258848667, 'clip_ratio/high_mean': 0.07725444622337818, 'clip_ratio/high_max': 0.07725444622337818, 'clip_ratio/region_mean': 0.18823834881186485, 'reward_total_mean': 0.13366711139678955, 'reward_meter_mean': 0.9013170003890991, 'reward_meter_std': 0.08074897527694702, 'reward_count_adherence_mean': 0.8636363744735718, 'reward_count_adherence_std': 0.04859296977519989, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9590909481048584, 'reward_count_floor_std': 0.014577903784811497, 'reward_lexical_plausibility_mean': 0.983328104019165, 'reward_lexical_plausibility_std': 0.0458899661898613, 'reward_repeat_penalty_mean': 0.9893382787704468, 'reward_repeat_penalty_std': 0.009818924590945244, 'reward_repeat_floor_mean': 0.9988276958465576, 'reward_repeat_floor_std': 0.0009705271222628653, 'reward_near_duplicate_penalty_mean': 0.9965053200721741, 'reward_near_duplicate_penalty_std': 0.0016852014232426882, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9927985668182373, 'reward_distinct_2_std': 0.00856027565896511, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.09425459802150726, 'reward_total_composite_mean': 0.13366711139678955, 'reward_total_composite_std': 0.08221367001533508, 'epoch': 0.0}

  2%|▏         | 57/3300 [11:34<13:26:58, 14.93s/it]INFO 04-12 16:10:00 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 58/3300 [11:43<11:53:37, 13.21s/it]2026-04-12 16:10:09,834 | INFO | train_grpo_train | metrics_logged mode=train step=58

                                                    
{'loss': -0.062, 'grad_norm': 10.932044982910156, 'learning_rate': 9.827272727272729e-06, 'num_tokens': 113545.0, 'completions/mean_length': 96.875, 'completions/min_length': 73.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.875, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.8158788681030273, 'rewards/meter/std': 0.2542024850845337, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9748784303665161, 'rewards/repeat_penalty/std': 0.05859052762389183, 'rewards/repeat_floor/mean': 0.9971906542778015, 'rewards/repeat_floor/std': 0.00609612138941884, 'rewards/near_duplicate_penalty/mean': 0.9917606115341187, 'rewards/near_duplicate_penalty/std': 0.012072098441421986, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.2996920943260193, 'rewards/total_composite/std': 0.08764141798019409, 'reward': 0.2996920943260193, 'reward_std': 0.08764141798019409, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.222623810172081, 'sampling/sampling_logp_difference/max': 1.4790148735046387, 'sampling/importance_sampling_ratio/min': 0.22786207497119904, 'sampling/importance_sampling_ratio/mean': 1.0376818180084229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.836358606815338, 'clip_ratio/low_mean': 0.0490132924169302, 'clip_ratio/low_min': 0.0490132924169302, 'clip_ratio/high_mean': 0.16877817921340466, 'clip_ratio/high_max': 0.16877817921340466, 'clip_ratio/region_mean': 0.21779147163033485, 'reward_total_mean': 0.2996920943260193, 'reward_meter_mean': 0.8158788681030273, 'reward_meter_std': 0.2542024850845337, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9748784303665161, 'reward_repeat_penalty_std': 0.05859052762389183, 'reward_repeat_floor_mean': 0.9971906542778015, 'reward_repeat_floor_std': 0.00609612138941884, 'reward_near_duplicate_penalty_mean': 0.9917606115341187, 'reward_near_duplicate_penalty_std': 0.012072098441421986, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2996920943260193, 'reward_total_composite_std': 0.08764141798019409, 'epoch': 0.0}

  2%|▏         | 58/3300 [11:43<11:53:37, 13.21s/it]INFO 04-12 16:10:10 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 59/3300 [11:52<10:45:49, 11.96s/it]2026-04-12 16:10:18,876 | INFO | train_grpo_train | metrics_logged mode=train step=59

                                                    
{'loss': 0.1433, 'grad_norm': 17.224931716918945, 'learning_rate': 9.824242424242425e-06, 'num_tokens': 115263.0, 'completions/mean_length': 43.75, 'completions/min_length': 28.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8781561851501465, 'rewards/meter/std': 0.13444408774375916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9930555820465088, 'rewards/lexical_plausibility/std': 0.01285861898213625, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2745385766029358, 'rewards/repeat_penalty/mean': 0.9897637367248535, 'rewards/repeat_penalty/std': 0.02021852321922779, 'rewards/repeat_floor/mean': 0.9984645843505859, 'rewards/repeat_floor/std': 0.003032783279195428, 'rewards/near_duplicate_penalty/mean': 0.9897637367248535, 'rewards/near_duplicate_penalty/std': 0.02021852321922779, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.43237680196762085, 'rewards/total_composite/std': 0.22993886470794678, 'reward': 0.43237680196762085, 'reward_std': 0.22993884980678558, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2185475379228592, 'sampling/sampling_logp_difference/max': 2.5989527702331543, 'sampling/importance_sampling_ratio/min': 0.07435140013694763, 'sampling/importance_sampling_ratio/mean': 1.0197927951812744, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.218057304620743, 'clip_ratio/low_mean': 0.13719338364899158, 'clip_ratio/low_min': 0.13719338364899158, 'clip_ratio/high_mean': 0.07783209532499313, 'clip_ratio/high_max': 0.07783209532499313, 'clip_ratio/region_mean': 0.21502547897398472, 'reward_total_mean': 0.43237680196762085, 'reward_meter_mean': 0.8781561851501465, 'reward_meter_std': 0.13444408774375916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9930555820465088, 'reward_lexical_plausibility_std': 0.01285861898213625, 'reward_repeat_penalty_mean': 0.9897637367248535, 'reward_repeat_penalty_std': 0.02021852321922779, 'reward_repeat_floor_mean': 0.9984645843505859, 'reward_repeat_floor_std': 0.003032783279195428, 'reward_near_duplicate_penalty_mean': 0.9897637367248535, 'reward_near_duplicate_penalty_std': 0.02021852321922779, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2745385766029358, 'reward_total_composite_mean': 0.43237680196762085, 'reward_total_composite_std': 0.22993886470794678, 'epoch': 0.0}

  2%|▏         | 59/3300 [11:52<10:45:49, 11.96s/it]INFO 04-12 16:10:19 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 60/3300 [12:03<10:25:44, 11.59s/it]2026-04-12 16:10:29,601 | INFO | train_grpo_train | metrics_logged mode=train step=60

                                                    
{'loss': 0.0639, 'grad_norm': 9.0322265625, 'learning_rate': 9.821212121212122e-06, 'num_tokens': 117897.0, 'completions/mean_length': 155.25, 'completions/min_length': 115.0, 'completions/max_length': 187.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 155.25, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.9913378953933716, 'rewards/meter/std': 0.010462315753102303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9424874782562256, 'rewards/lexical_plausibility/std': 0.110664963722229, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.98736971616745, 'rewards/repeat_penalty/std': 0.025898857042193413, 'rewards/repeat_floor/mean': 0.9985493421554565, 'rewards/repeat_floor/std': 0.002653527772054076, 'rewards/near_duplicate_penalty/mean': 0.9949444532394409, 'rewards/near_duplicate_penalty/std': 0.005881176795810461, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.39576077461242676, 'rewards/total_composite/std': 0.10535090416669846, 'reward': 0.39576077461242676, 'reward_std': 0.10535089671611786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1954849362373352, 'sampling/sampling_logp_difference/max': 1.7156519889831543, 'sampling/importance_sampling_ratio/min': 0.17984643578529358, 'sampling/importance_sampling_ratio/mean': 1.0383236408233643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2503845244646072, 'clip_ratio/low_mean': 0.04047378897666931, 'clip_ratio/low_min': 0.04047378897666931, 'clip_ratio/high_mean': 0.1617762576788664, 'clip_ratio/high_max': 0.1617762576788664, 'clip_ratio/region_mean': 0.2022500466555357, 'reward_total_mean': 0.39576077461242676, 'reward_meter_mean': 0.9913378953933716, 'reward_meter_std': 0.010462315753102303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9424874782562256, 'reward_lexical_plausibility_std': 0.110664963722229, 'reward_repeat_penalty_mean': 0.98736971616745, 'reward_repeat_penalty_std': 0.025898857042193413, 'reward_repeat_floor_mean': 0.9985493421554565, 'reward_repeat_floor_std': 0.002653527772054076, 'reward_near_duplicate_penalty_mean': 0.9949444532394409, 'reward_near_duplicate_penalty_std': 0.005881176795810461, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.39576077461242676, 'reward_total_composite_std': 0.10535090416669846, 'epoch': 0.0}

  2%|▏         | 60/3300 [12:03<10:25:44, 11.59s/it]INFO 04-12 16:10:29 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 61/3300 [12:12<9:42:43, 10.79s/it] 2026-04-12 16:10:38,544 | INFO | train_grpo_train | metrics_logged mode=train step=61

                                                   
{'loss': 0.0872, 'grad_norm': 27.523418426513672, 'learning_rate': 9.81818181818182e-06, 'num_tokens': 119366.0, 'completions/mean_length': 35.625, 'completions/min_length': 25.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.42723244428634644, 'rewards/meter/std': 0.40925148129463196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5250000357627869, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9827656149864197, 'rewards/repeat_penalty/std': 0.019679687917232513, 'rewards/repeat_floor/mean': 0.9974148273468018, 'rewards/repeat_floor/std': 0.00295194867067039, 'rewards/near_duplicate_penalty/mean': 0.9827656149864197, 'rewards/near_duplicate_penalty/std': 0.019679687917232513, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25965416431427, 'rewards/total_composite/std': 0.27791792154312134, 'reward': 0.25965416431427, 'reward_std': 0.27791789174079895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2621625065803528, 'sampling/sampling_logp_difference/max': 2.3092856407165527, 'sampling/importance_sampling_ratio/min': 0.09933219105005264, 'sampling/importance_sampling_ratio/mean': 0.9767956137657166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.357659012079239, 'clip_ratio/low_mean': 0.10717914719134569, 'clip_ratio/low_min': 0.10717914719134569, 'clip_ratio/high_mean': 0.13505913875997066, 'clip_ratio/high_max': 0.13505913875997066, 'clip_ratio/region_mean': 0.24223828595131636, 'reward_total_mean': 0.25965416431427, 'reward_meter_mean': 0.42723244428634644, 'reward_meter_std': 0.40925148129463196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9827656149864197, 'reward_repeat_penalty_std': 0.019679687917232513, 'reward_repeat_floor_mean': 0.9974148273468018, 'reward_repeat_floor_std': 0.00295194867067039, 'reward_near_duplicate_penalty_mean': 0.9827656149864197, 'reward_near_duplicate_penalty_std': 0.019679687917232513, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5250000357627869, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.25965416431427, 'reward_total_composite_std': 0.27791792154312134, 'epoch': 0.0}

  2%|▏         | 61/3300 [12:12<9:42:43, 10.79s/it]INFO 04-12 16:10:38 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 62/3300 [12:20<8:56:07,  9.93s/it]2026-04-12 16:10:46,472 | INFO | train_grpo_train | metrics_logged mode=train step=62

                                                   
{'loss': 0.1218, 'grad_norm': 21.601524353027344, 'learning_rate': 9.815151515151516e-06, 'num_tokens': 120932.0, 'completions/mean_length': 34.75, 'completions/min_length': 28.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.3043733835220337, 'rewards/meter/std': 0.30216702818870544, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9836230278015137, 'rewards/lexical_plausibility/std': 0.04632115736603737, 'rewards/judge_quality/mean': 0.7012499570846558, 'rewards/judge_quality/std': 0.3290001451969147, 'rewards/repeat_penalty/mean': 0.9963341355323792, 'rewards/repeat_penalty/std': 0.005656497087329626, 'rewards/repeat_floor/mean': 0.9994500875473022, 'rewards/repeat_floor/std': 0.0008484762511216104, 'rewards/near_duplicate_penalty/mean': 0.9963341355323792, 'rewards/near_duplicate_penalty/std': 0.005656497087329626, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18301305174827576, 'rewards/total_composite/std': 0.18151044845581055, 'reward': 0.18301305174827576, 'reward_std': 0.18151043355464935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2644853889942169, 'sampling/sampling_logp_difference/max': 2.5253634452819824, 'sampling/importance_sampling_ratio/min': 0.08002921938896179, 'sampling/importance_sampling_ratio/mean': 0.9787096381187439, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7581471502780914, 'clip_ratio/low_mean': 0.11003666650503874, 'clip_ratio/low_min': 0.11003666650503874, 'clip_ratio/high_mean': 0.08930934779345989, 'clip_ratio/high_max': 0.08930934779345989, 'clip_ratio/region_mean': 0.19934601429849863, 'reward_total_mean': 0.18301305174827576, 'reward_meter_mean': 0.3043733835220337, 'reward_meter_std': 0.30216702818870544, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9836230278015137, 'reward_lexical_plausibility_std': 0.04632115736603737, 'reward_repeat_penalty_mean': 0.9963341355323792, 'reward_repeat_penalty_std': 0.005656497087329626, 'reward_repeat_floor_mean': 0.9994500875473022, 'reward_repeat_floor_std': 0.0008484762511216104, 'reward_near_duplicate_penalty_mean': 0.9963341355323792, 'reward_near_duplicate_penalty_std': 0.005656497087329626, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7012499570846558, 'reward_judge_quality_std': 0.3290001451969147, 'reward_total_composite_mean': 0.18301305174827576, 'reward_total_composite_std': 0.18151044845581055, 'epoch': 0.0}

  2%|▏         | 62/3300 [12:20<8:56:07,  9.93s/it]INFO 04-12 16:10:46 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 63/3300 [12:28<8:30:00,  9.45s/it]2026-04-12 16:10:54,808 | INFO | train_grpo_train | metrics_logged mode=train step=63

                                                   
{'loss': 0.2275, 'grad_norm': 26.84090232849121, 'learning_rate': 9.812121212121212e-06, 'num_tokens': 122192.0, 'completions/mean_length': 16.5, 'completions/min_length': 12.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.5, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.4654954671859741, 'rewards/meter/std': 0.4026661217212677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.21104755997657776, 'rewards/repeat_penalty/mean': 0.7432166337966919, 'rewards/repeat_penalty/std': 0.019186342135071754, 'rewards/repeat_floor/mean': 0.9836432933807373, 'rewards/repeat_floor/std': 0.0038372809067368507, 'rewards/near_duplicate_penalty/mean': 0.9909554719924927, 'rewards/near_duplicate_penalty/std': 0.025581790134310722, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20217357575893402, 'rewards/total_composite/std': 0.1694164276123047, 'reward': 0.20217357575893402, 'reward_std': 0.1694164276123047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21298496425151825, 'sampling/sampling_logp_difference/max': 1.5629849433898926, 'sampling/importance_sampling_ratio/min': 0.20950976014137268, 'sampling/importance_sampling_ratio/mean': 0.9909963011741638, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2390024811029434, 'clip_ratio/low_mean': 0.06486344616860151, 'clip_ratio/low_min': 0.06486344616860151, 'clip_ratio/high_mean': 0.11815476510673761, 'clip_ratio/high_max': 0.11815476510673761, 'clip_ratio/region_mean': 0.18301821127533913, 'reward_total_mean': 0.20217357575893402, 'reward_meter_mean': 0.4654954671859741, 'reward_meter_std': 0.4026661217212677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7432166337966919, 'reward_repeat_penalty_std': 0.019186342135071754, 'reward_repeat_floor_mean': 0.9836432933807373, 'reward_repeat_floor_std': 0.0038372809067368507, 'reward_near_duplicate_penalty_mean': 0.9909554719924927, 'reward_near_duplicate_penalty_std': 0.025581790134310722, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.21104755997657776, 'reward_total_composite_mean': 0.20217357575893402, 'reward_total_composite_std': 0.1694164276123047, 'epoch': 0.0}

  2%|▏         | 63/3300 [12:28<8:30:00,  9.45s/it]INFO 04-12 16:10:55 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 64/3300 [12:38<8:30:13,  9.46s/it]2026-04-12 16:11:04,280 | INFO | train_grpo_train | metrics_logged mode=train step=64

                                                   
{'loss': 0.1121, 'grad_norm': 14.572439193725586, 'learning_rate': 9.809090909090911e-06, 'num_tokens': 124051.0, 'completions/mean_length': 62.375, 'completions/min_length': 48.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.8821189403533936, 'rewards/meter/std': 0.26891547441482544, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9881824254989624, 'rewards/repeat_penalty/std': 0.014187062159180641, 'rewards/repeat_floor/mean': 0.9982273578643799, 'rewards/repeat_floor/std': 0.002128050895407796, 'rewards/near_duplicate_penalty/mean': 0.9881824254989624, 'rewards/near_duplicate_penalty/std': 0.014187062159180641, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3443075120449066, 'rewards/total_composite/std': 0.1484030932188034, 'reward': 0.3443075120449066, 'reward_std': 0.1484030932188034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22710561752319336, 'sampling/sampling_logp_difference/max': 1.7819933891296387, 'sampling/importance_sampling_ratio/min': 0.16830232739448547, 'sampling/importance_sampling_ratio/mean': 1.0271949768066406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0756715685129166, 'clip_ratio/low_mean': 0.05536968074738979, 'clip_ratio/low_min': 0.05536968074738979, 'clip_ratio/high_mean': 0.14837482012808323, 'clip_ratio/high_max': 0.14837482012808323, 'clip_ratio/region_mean': 0.20374450087547302, 'reward_total_mean': 0.3443075120449066, 'reward_meter_mean': 0.8821189403533936, 'reward_meter_std': 0.26891547441482544, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9881824254989624, 'reward_repeat_penalty_std': 0.014187062159180641, 'reward_repeat_floor_mean': 0.9982273578643799, 'reward_repeat_floor_std': 0.002128050895407796, 'reward_near_duplicate_penalty_mean': 0.9881824254989624, 'reward_near_duplicate_penalty_std': 0.014187062159180641, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.3443075120449066, 'reward_total_composite_std': 0.1484030932188034, 'epoch': 0.0}

  2%|▏         | 64/3300 [12:38<8:30:13,  9.46s/it]INFO 04-12 16:11:04 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 65/3300 [12:47<8:23:44,  9.34s/it]2026-04-12 16:11:13,348 | INFO | train_grpo_train | metrics_logged mode=train step=65

                                                   
{'loss': -0.1235, 'grad_norm': 25.44385528564453, 'learning_rate': 9.806060606060607e-06, 'num_tokens': 125739.0, 'completions/mean_length': 47.0, 'completions/min_length': 34.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7405123710632324, 'rewards/meter/std': 0.3356536626815796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4587499797344208, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9792635440826416, 'rewards/repeat_penalty/std': 0.012044358998537064, 'rewards/repeat_floor/mean': 0.9968895316123962, 'rewards/repeat_floor/std': 0.0018066534539684653, 'rewards/near_duplicate_penalty/mean': 0.9792635440826416, 'rewards/near_duplicate_penalty/std': 0.012044358998537064, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3144078254699707, 'rewards/total_composite/std': 0.13536033034324646, 'reward': 0.3144078254699707, 'reward_std': 0.13536033034324646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24648673832416534, 'sampling/sampling_logp_difference/max': 1.7683677673339844, 'sampling/importance_sampling_ratio/min': 0.17061124742031097, 'sampling/importance_sampling_ratio/mean': 1.0266461372375488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4860416650772095, 'clip_ratio/low_mean': 0.046078432351350784, 'clip_ratio/low_min': 0.046078432351350784, 'clip_ratio/high_mean': 0.15715246833860874, 'clip_ratio/high_max': 0.15715246833860874, 'clip_ratio/region_mean': 0.20323090068995953, 'reward_total_mean': 0.3144078254699707, 'reward_meter_mean': 0.7405123710632324, 'reward_meter_std': 0.3356536626815796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9792635440826416, 'reward_repeat_penalty_std': 0.012044358998537064, 'reward_repeat_floor_mean': 0.9968895316123962, 'reward_repeat_floor_std': 0.0018066534539684653, 'reward_near_duplicate_penalty_mean': 0.9792635440826416, 'reward_near_duplicate_penalty_std': 0.012044358998537064, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4587499797344208, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.3144078254699707, 'reward_total_composite_std': 0.13536033034324646, 'epoch': 0.0}

  2%|▏         | 65/3300 [12:47<8:23:44,  9.34s/it]INFO 04-12 16:11:13 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 66/3300 [12:55<8:02:30,  8.95s/it]2026-04-12 16:11:21,387 | INFO | train_grpo_train | metrics_logged mode=train step=66

                                                   
{'loss': -0.0324, 'grad_norm': 15.109299659729004, 'learning_rate': 9.803030303030304e-06, 'num_tokens': 127343.0, 'completions/mean_length': 40.5, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.857477605342865, 'rewards/meter/std': 0.13007476925849915, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9572981595993042, 'rewards/lexical_plausibility/std': 0.08722753822803497, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9931166172027588, 'rewards/repeat_penalty/std': 0.015105213969945908, 'rewards/repeat_floor/mean': 0.9989675283432007, 'rewards/repeat_floor/std': 0.002265774644911289, 'rewards/near_duplicate_penalty/mean': 0.9931166172027588, 'rewards/near_duplicate_penalty/std': 0.015105213969945908, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4047844409942627, 'rewards/total_composite/std': 0.12278708815574646, 'reward': 0.4047844409942627, 'reward_std': 0.12278708815574646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17524486780166626, 'sampling/sampling_logp_difference/max': 1.338212013244629, 'sampling/importance_sampling_ratio/min': 0.2623142600059509, 'sampling/importance_sampling_ratio/mean': 1.0493476390838623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.170139789581299, 'clip_ratio/low_mean': 0.1144349854439497, 'clip_ratio/low_min': 0.1144349854439497, 'clip_ratio/high_mean': 0.06920644082129002, 'clip_ratio/high_max': 0.06920644082129002, 'clip_ratio/region_mean': 0.18364142626523972, 'reward_total_mean': 0.4047844409942627, 'reward_meter_mean': 0.857477605342865, 'reward_meter_std': 0.13007476925849915, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9572981595993042, 'reward_lexical_plausibility_std': 0.08722753822803497, 'reward_repeat_penalty_mean': 0.9931166172027588, 'reward_repeat_penalty_std': 0.015105213969945908, 'reward_repeat_floor_mean': 0.9989675283432007, 'reward_repeat_floor_std': 0.002265774644911289, 'reward_near_duplicate_penalty_mean': 0.9931166172027588, 'reward_near_duplicate_penalty_std': 0.015105213969945908, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.4047844409942627, 'reward_total_composite_std': 0.12278708815574646, 'epoch': 0.0}

  2%|▏         | 66/3300 [12:55<8:02:30,  8.95s/it]INFO 04-12 16:11:21 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 67/3300 [13:06<8:36:29,  9.59s/it]2026-04-12 16:11:32,451 | INFO | train_grpo_train | metrics_logged mode=train step=67

                                                   
{'loss': 0.0426, 'grad_norm': 14.470446586608887, 'learning_rate': 9.800000000000001e-06, 'num_tokens': 129036.0, 'completions/mean_length': 57.625, 'completions/min_length': 36.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4118611812591553, 'rewards/meter/std': 0.4184275269508362, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9930555820465088, 'rewards/lexical_plausibility/std': 0.01285861898213625, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121347069740295, 'rewards/repeat_penalty/mean': 0.9856113791465759, 'rewards/repeat_penalty/std': 0.018535489216446877, 'rewards/repeat_floor/mean': 0.9978417158126831, 'rewards/repeat_floor/std': 0.002780318958684802, 'rewards/near_duplicate_penalty/mean': 0.9856113791465759, 'rewards/near_duplicate_penalty/std': 0.018535489216446877, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17825430631637573, 'rewards/total_composite/std': 0.26260918378829956, 'reward': 0.17825430631637573, 'reward_std': 0.26260918378829956, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2115630805492401, 'sampling/sampling_logp_difference/max': 1.5753626823425293, 'sampling/importance_sampling_ratio/min': 0.20693248510360718, 'sampling/importance_sampling_ratio/mean': 1.0093439817428589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9403171986341476, 'clip_ratio/low_mean': 0.13015037402510643, 'clip_ratio/low_min': 0.13015037402510643, 'clip_ratio/high_mean': 0.06842435523867607, 'clip_ratio/high_max': 0.06842435523867607, 'clip_ratio/region_mean': 0.1985747292637825, 'reward_total_mean': 0.17825430631637573, 'reward_meter_mean': 0.4118611812591553, 'reward_meter_std': 0.4184275269508362, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9930555820465088, 'reward_lexical_plausibility_std': 0.01285861898213625, 'reward_repeat_penalty_mean': 0.9856113791465759, 'reward_repeat_penalty_std': 0.018535489216446877, 'reward_repeat_floor_mean': 0.9978417158126831, 'reward_repeat_floor_std': 0.002780318958684802, 'reward_near_duplicate_penalty_mean': 0.9856113791465759, 'reward_near_duplicate_penalty_std': 0.018535489216446877, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121347069740295, 'reward_total_composite_mean': 0.17825430631637573, 'reward_total_composite_std': 0.26260918378829956, 'epoch': 0.0}

  2%|▏         | 67/3300 [13:06<8:36:29,  9.59s/it]INFO 04-12 16:11:32 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 68/3300 [13:17<9:03:05, 10.08s/it]2026-04-12 16:11:43,691 | INFO | train_grpo_train | metrics_logged mode=train step=68

                                                   
{'loss': 0.1064, 'grad_norm': 24.879629135131836, 'learning_rate': 9.796969696969698e-06, 'num_tokens': 130351.0, 'completions/mean_length': 20.375, 'completions/min_length': 16.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.5425297021865845, 'rewards/meter/std': 0.4662072956562042, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.2950030267238617, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3311706781387329, 'rewards/total_composite/std': 0.35036519169807434, 'reward': 0.3311706781387329, 'reward_std': 0.35036519169807434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18561644852161407, 'sampling/sampling_logp_difference/max': 1.0416240692138672, 'sampling/importance_sampling_ratio/min': 0.3528811037540436, 'sampling/importance_sampling_ratio/mean': 1.0162744522094727, 'sampling/importance_sampling_ratio/max': 1.9564259052276611, 'entropy': 1.7219845056533813, 'clip_ratio/low_mean': 0.05209703976288438, 'clip_ratio/low_min': 0.05209703976288438, 'clip_ratio/high_mean': 0.10329861473292112, 'clip_ratio/high_max': 0.10329861473292112, 'clip_ratio/region_mean': 0.1553956544958055, 'reward_total_mean': 0.3311706781387329, 'reward_meter_mean': 0.5425297021865845, 'reward_meter_std': 0.4662072956562042, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.2950030267238617, 'reward_total_composite_mean': 0.3311706781387329, 'reward_total_composite_std': 0.35036519169807434, 'epoch': 0.0}

  2%|▏         | 68/3300 [13:17<9:03:05, 10.08s/it]INFO 04-12 16:11:43 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 69/3300 [13:26<8:37:28,  9.61s/it]2026-04-12 16:11:52,200 | INFO | train_grpo_train | metrics_logged mode=train step=69

                                                   
{'loss': 0.0433, 'grad_norm': 19.02853012084961, 'learning_rate': 9.793939393939394e-06, 'num_tokens': 132038.0, 'completions/mean_length': 49.875, 'completions/min_length': 35.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.546741247177124, 'rewards/meter/std': 0.43295955657958984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9810117483139038, 'rewards/repeat_penalty/std': 0.022440442815423012, 'rewards/repeat_floor/mean': 0.9971517324447632, 'rewards/repeat_floor/std': 0.0033660747576504946, 'rewards/near_duplicate_penalty/mean': 0.9810117483139038, 'rewards/near_duplicate_penalty/std': 0.022440442815423012, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2481343001127243, 'rewards/total_composite/std': 0.1825486123561859, 'reward': 0.2481343001127243, 'reward_std': 0.1825486123561859, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21970173716545105, 'sampling/sampling_logp_difference/max': 2.9506092071533203, 'sampling/importance_sampling_ratio/min': 0.05230783298611641, 'sampling/importance_sampling_ratio/mean': 1.0293940305709839, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1979509592056274, 'clip_ratio/low_mean': 0.05561224650591612, 'clip_ratio/low_min': 0.05561224650591612, 'clip_ratio/high_mean': 0.12462344113737345, 'clip_ratio/high_max': 0.12462344113737345, 'clip_ratio/region_mean': 0.18023568764328957, 'reward_total_mean': 0.2481343001127243, 'reward_meter_mean': 0.546741247177124, 'reward_meter_std': 0.43295955657958984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9810117483139038, 'reward_repeat_penalty_std': 0.022440442815423012, 'reward_repeat_floor_mean': 0.9971517324447632, 'reward_repeat_floor_std': 0.0033660747576504946, 'reward_near_duplicate_penalty_mean': 0.9810117483139038, 'reward_near_duplicate_penalty_std': 0.022440442815423012, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.2481343001127243, 'reward_total_composite_std': 0.1825486123561859, 'epoch': 0.0}

  2%|▏         | 69/3300 [13:26<8:37:28,  9.61s/it]INFO 04-12 16:11:52 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 70/3300 [13:34<8:21:24,  9.31s/it]2026-04-12 16:12:00,825 | INFO | train_grpo_train | metrics_logged mode=train step=70

                                                   
{'loss': 0.0476, 'grad_norm': 14.128938674926758, 'learning_rate': 9.790909090909093e-06, 'num_tokens': 134009.0, 'completions/mean_length': 80.375, 'completions/min_length': 68.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.5790213346481323, 'rewards/meter/std': 0.36007389426231384, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9857064485549927, 'rewards/lexical_plausibility/std': 0.02878989279270172, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.10350983589887619, 'rewards/repeat_penalty/mean': 0.9832158088684082, 'rewards/repeat_penalty/std': 0.02878197841346264, 'rewards/repeat_floor/mean': 0.9978755116462708, 'rewards/repeat_floor/std': 0.0032227092888206244, 'rewards/near_duplicate_penalty/mean': 0.9901101589202881, 'rewards/near_duplicate_penalty/std': 0.009958839043974876, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9928774833679199, 'rewards/distinct_2/std': 0.02014549821615219, 'rewards/total_composite/mean': 0.22911961376667023, 'rewards/total_composite/std': 0.14408215880393982, 'reward': 0.22911961376667023, 'reward_std': 0.14408215880393982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22844818234443665, 'sampling/sampling_logp_difference/max': 1.4417133331298828, 'sampling/importance_sampling_ratio/min': 0.23652216792106628, 'sampling/importance_sampling_ratio/mean': 1.029520869255066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.37846240401268, 'clip_ratio/low_mean': 0.08246947452425957, 'clip_ratio/low_min': 0.08246947452425957, 'clip_ratio/high_mean': 0.13380922749638557, 'clip_ratio/high_max': 0.13380922749638557, 'clip_ratio/region_mean': 0.21627870202064514, 'reward_total_mean': 0.22911961376667023, 'reward_meter_mean': 0.5790213346481323, 'reward_meter_std': 0.36007389426231384, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9857064485549927, 'reward_lexical_plausibility_std': 0.02878989279270172, 'reward_repeat_penalty_mean': 0.9832158088684082, 'reward_repeat_penalty_std': 0.02878197841346264, 'reward_repeat_floor_mean': 0.9978755116462708, 'reward_repeat_floor_std': 0.0032227092888206244, 'reward_near_duplicate_penalty_mean': 0.9901101589202881, 'reward_near_duplicate_penalty_std': 0.009958839043974876, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9928774833679199, 'reward_distinct_2_std': 0.02014549821615219, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.10350983589887619, 'reward_total_composite_mean': 0.22911961376667023, 'reward_total_composite_std': 0.14408215880393982, 'epoch': 0.0}

  2%|▏         | 70/3300 [13:34<8:21:24,  9.31s/it]INFO 04-12 16:12:01 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 71/3300 [13:43<8:07:29,  9.06s/it]2026-04-12 16:12:09,287 | INFO | train_grpo_train | metrics_logged mode=train step=71

                                                   
{'loss': 0.0156, 'grad_norm': 20.836307525634766, 'learning_rate': 9.787878787878788e-06, 'num_tokens': 135523.0, 'completions/mean_length': 27.25, 'completions/min_length': 19.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.3806454539299011, 'rewards/meter/std': 0.4722149670124054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9046452045440674, 'rewards/lexical_plausibility/std': 0.2059667706489563, 'rewards/judge_quality/mean': 0.476250022649765, 'rewards/judge_quality/std': 0.38119879364967346, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1801045835018158, 'rewards/total_composite/std': 0.303378701210022, 'reward': 0.1801045835018158, 'reward_std': 0.303378701210022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20752005279064178, 'sampling/sampling_logp_difference/max': 1.965214729309082, 'sampling/importance_sampling_ratio/min': 0.1401257961988449, 'sampling/importance_sampling_ratio/mean': 1.0337620973587036, 'sampling/importance_sampling_ratio/max': 1.8718876838684082, 'entropy': 2.0731808096170425, 'clip_ratio/low_mean': 0.1582103967666626, 'clip_ratio/low_min': 0.1582103967666626, 'clip_ratio/high_mean': 0.0467391312122345, 'clip_ratio/high_max': 0.0467391312122345, 'clip_ratio/region_mean': 0.2049495279788971, 'reward_total_mean': 0.1801045835018158, 'reward_meter_mean': 0.3806454539299011, 'reward_meter_std': 0.4722149670124054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9046452045440674, 'reward_lexical_plausibility_std': 0.2059667706489563, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.476250022649765, 'reward_judge_quality_std': 0.38119879364967346, 'reward_total_composite_mean': 0.1801045835018158, 'reward_total_composite_std': 0.303378701210022, 'epoch': 0.0}

  2%|▏         | 71/3300 [13:43<8:07:29,  9.06s/it]INFO 04-12 16:12:09 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 72/3300 [13:59<9:59:04, 11.14s/it]2026-04-12 16:12:25,266 | INFO | train_grpo_train | metrics_logged mode=train step=72

                                                   
{'loss': 0.0315, 'grad_norm': 20.252107620239258, 'learning_rate': 9.784848484848486e-06, 'num_tokens': 137084.0, 'completions/mean_length': 43.125, 'completions/min_length': 31.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5920624732971191, 'rewards/meter/std': 0.44187507033348083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.25616681575775146, 'rewards/repeat_penalty/mean': 0.9964206218719482, 'rewards/repeat_penalty/std': 0.009782219305634499, 'rewards/repeat_floor/mean': 0.9994630813598633, 'rewards/repeat_floor/std': 0.0014673316618427634, 'rewards/near_duplicate_penalty/mean': 0.9964206218719482, 'rewards/near_duplicate_penalty/std': 0.009782219305634499, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4168849587440491, 'rewards/total_composite/std': 0.3681540787220001, 'reward': 0.4168849587440491, 'reward_std': 0.3681540787220001, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2308512032032013, 'sampling/sampling_logp_difference/max': 2.053664207458496, 'sampling/importance_sampling_ratio/min': 0.12826405465602875, 'sampling/importance_sampling_ratio/mean': 1.0262924432754517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.046980783343315, 'clip_ratio/low_mean': 0.09719309955835342, 'clip_ratio/low_min': 0.09719309955835342, 'clip_ratio/high_mean': 0.07912491727620363, 'clip_ratio/high_max': 0.07912491727620363, 'clip_ratio/region_mean': 0.17631801683455706, 'reward_total_mean': 0.4168849587440491, 'reward_meter_mean': 0.5920624732971191, 'reward_meter_std': 0.44187507033348083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9964206218719482, 'reward_repeat_penalty_std': 0.009782219305634499, 'reward_repeat_floor_mean': 0.9994630813598633, 'reward_repeat_floor_std': 0.0014673316618427634, 'reward_near_duplicate_penalty_mean': 0.9964206218719482, 'reward_near_duplicate_penalty_std': 0.009782219305634499, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.25616681575775146, 'reward_total_composite_mean': 0.4168849587440491, 'reward_total_composite_std': 0.3681540787220001, 'epoch': 0.0}

  2%|▏         | 72/3300 [13:59<9:59:04, 11.14s/it]INFO 04-12 16:12:25 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 73/3300 [14:07<9:19:22, 10.40s/it]2026-04-12 16:12:33,953 | INFO | train_grpo_train | metrics_logged mode=train step=73

                                                   
{'loss': 0.031, 'grad_norm': 12.774842262268066, 'learning_rate': 9.781818181818183e-06, 'num_tokens': 138820.0, 'completions/mean_length': 49.0, 'completions/min_length': 36.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5381691455841064, 'rewards/meter/std': 0.34407538175582886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9801508188247681, 'rewards/lexical_plausibility/std': 0.04595785215497017, 'rewards/judge_quality/mean': 0.6462500095367432, 'rewards/judge_quality/std': 0.2263333648443222, 'rewards/repeat_penalty/mean': 0.9840354919433594, 'rewards/repeat_penalty/std': 0.029167495667934418, 'rewards/repeat_floor/mean': 0.9976053237915039, 'rewards/repeat_floor/std': 0.004375113174319267, 'rewards/near_duplicate_penalty/mean': 0.9840354919433594, 'rewards/near_duplicate_penalty/std': 0.029167495667934418, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.35036981105804443, 'rewards/total_composite/std': 0.2823674976825714, 'reward': 0.35036981105804443, 'reward_std': 0.282367467880249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1923993080854416, 'sampling/sampling_logp_difference/max': 1.5187273025512695, 'sampling/importance_sampling_ratio/min': 0.21899043023586273, 'sampling/importance_sampling_ratio/mean': 1.018026351928711, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7377461791038513, 'clip_ratio/low_mean': 0.13482744060456753, 'clip_ratio/low_min': 0.13482744060456753, 'clip_ratio/high_mean': 0.07108242809772491, 'clip_ratio/high_max': 0.07108242809772491, 'clip_ratio/region_mean': 0.20590986870229244, 'reward_total_mean': 0.35036981105804443, 'reward_meter_mean': 0.5381691455841064, 'reward_meter_std': 0.34407538175582886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9801508188247681, 'reward_lexical_plausibility_std': 0.04595785215497017, 'reward_repeat_penalty_mean': 0.9840354919433594, 'reward_repeat_penalty_std': 0.029167495667934418, 'reward_repeat_floor_mean': 0.9976053237915039, 'reward_repeat_floor_std': 0.004375113174319267, 'reward_near_duplicate_penalty_mean': 0.9840354919433594, 'reward_near_duplicate_penalty_std': 0.029167495667934418, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6462500095367432, 'reward_judge_quality_std': 0.2263333648443222, 'reward_total_composite_mean': 0.35036981105804443, 'reward_total_composite_std': 0.2823674976825714, 'epoch': 0.0}

  2%|▏         | 73/3300 [14:07<9:19:22, 10.40s/it]INFO 04-12 16:12:34 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 74/3300 [14:17<9:03:22, 10.11s/it]2026-04-12 16:12:43,374 | INFO | train_grpo_train | metrics_logged mode=train step=74

                                                   
{'loss': 0.1424, 'grad_norm': 21.010709762573242, 'learning_rate': 9.77878787878788e-06, 'num_tokens': 140446.0, 'completions/mean_length': 48.25, 'completions/min_length': 32.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8615536093711853, 'rewards/meter/std': 0.25530722737312317, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8962432742118835, 'rewards/lexical_plausibility/std': 0.15018872916698456, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9534638524055481, 'rewards/repeat_penalty/std': 0.11381767690181732, 'rewards/repeat_floor/mean': 0.9959642887115479, 'rewards/repeat_floor/std': 0.008830887265503407, 'rewards/near_duplicate_penalty/mean': 0.9920052289962769, 'rewards/near_duplicate_penalty/std': 0.012510445900261402, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.30540943145751953, 'rewards/total_composite/std': 0.1472395956516266, 'reward': 0.30540943145751953, 'reward_std': 0.1472395807504654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24463339149951935, 'sampling/sampling_logp_difference/max': 1.5558185577392578, 'sampling/importance_sampling_ratio/min': 0.21101658046245575, 'sampling/importance_sampling_ratio/mean': 1.0142935514450073, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.618894636631012, 'clip_ratio/low_mean': 0.09531590715050697, 'clip_ratio/low_min': 0.09531590715050697, 'clip_ratio/high_mean': 0.15906223095953465, 'clip_ratio/high_max': 0.15906223095953465, 'clip_ratio/region_mean': 0.2543781381100416, 'reward_total_mean': 0.30540943145751953, 'reward_meter_mean': 0.8615536093711853, 'reward_meter_std': 0.25530722737312317, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8962432742118835, 'reward_lexical_plausibility_std': 0.15018872916698456, 'reward_repeat_penalty_mean': 0.9534638524055481, 'reward_repeat_penalty_std': 0.11381767690181732, 'reward_repeat_floor_mean': 0.9959642887115479, 'reward_repeat_floor_std': 0.008830887265503407, 'reward_near_duplicate_penalty_mean': 0.9920052289962769, 'reward_near_duplicate_penalty_std': 0.012510445900261402, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.30540943145751953, 'reward_total_composite_std': 0.1472395956516266, 'epoch': 0.0}

  2%|▏         | 74/3300 [14:17<9:03:22, 10.11s/it]INFO 04-12 16:12:43 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 75/3300 [14:26<8:42:28,  9.72s/it]2026-04-12 16:12:52,193 | INFO | train_grpo_train | metrics_logged mode=train step=75

                                                   
{'loss': -0.0507, 'grad_norm': 12.79999828338623, 'learning_rate': 9.775757575757576e-06, 'num_tokens': 142160.0, 'completions/mean_length': 57.25, 'completions/min_length': 39.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.7840166091918945, 'rewards/meter/std': 0.372598797082901, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9724822044372559, 'rewards/lexical_plausibility/std': 0.0778321623802185, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.9930688738822937, 'rewards/repeat_penalty/std': 0.012655957601964474, 'rewards/repeat_floor/mean': 0.9989603161811829, 'rewards/repeat_floor/std': 0.0018983840709552169, 'rewards/near_duplicate_penalty/mean': 0.9930688738822937, 'rewards/near_duplicate_penalty/std': 0.012655957601964474, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32773613929748535, 'rewards/total_composite/std': 0.13446436822414398, 'reward': 0.32773613929748535, 'reward_std': 0.13446436822414398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19810864329338074, 'sampling/sampling_logp_difference/max': 1.2086148262023926, 'sampling/importance_sampling_ratio/min': 0.2986105978488922, 'sampling/importance_sampling_ratio/mean': 1.0505841970443726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4203799813985825, 'clip_ratio/low_mean': 0.037847790867090225, 'clip_ratio/low_min': 0.037847790867090225, 'clip_ratio/high_mean': 0.1543244794011116, 'clip_ratio/high_max': 0.1543244794011116, 'clip_ratio/region_mean': 0.19217227026820183, 'reward_total_mean': 0.32773613929748535, 'reward_meter_mean': 0.7840166091918945, 'reward_meter_std': 0.372598797082901, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9724822044372559, 'reward_lexical_plausibility_std': 0.0778321623802185, 'reward_repeat_penalty_mean': 0.9930688738822937, 'reward_repeat_penalty_std': 0.012655957601964474, 'reward_repeat_floor_mean': 0.9989603161811829, 'reward_repeat_floor_std': 0.0018983840709552169, 'reward_near_duplicate_penalty_mean': 0.9930688738822937, 'reward_near_duplicate_penalty_std': 0.012655957601964474, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.32773613929748535, 'reward_total_composite_std': 0.13446436822414398, 'epoch': 0.0}

  2%|▏         | 75/3300 [14:26<8:42:28,  9.72s/it]INFO 04-12 16:12:52 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 76/3300 [14:37<9:03:48, 10.12s/it]2026-04-12 16:13:03,247 | INFO | train_grpo_train | metrics_logged mode=train step=76

                                                   
{'loss': 0.0274, 'grad_norm': 7.083038330078125, 'learning_rate': 9.772727272727273e-06, 'num_tokens': 145483.0, 'completions/mean_length': 231.375, 'completions/min_length': 161.0, 'completions/max_length': 280.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 231.375, 'completions/min_terminated_length': 161.0, 'completions/max_terminated_length': 280.0, 'rewards/meter/mean': 0.34000200033187866, 'rewards/meter/std': 0.24787180125713348, 'rewards/count_adherence/mean': 0.953125, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9859374761581421, 'rewards/count_floor/std': 0.019408106803894043, 'rewards/lexical_plausibility/mean': 0.9986978769302368, 'rewards/lexical_plausibility/std': 0.0036828548181802034, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9847592115402222, 'rewards/repeat_penalty/std': 0.01706889271736145, 'rewards/repeat_floor/mean': 0.9981634616851807, 'rewards/repeat_floor/std': 0.0017077922821044922, 'rewards/near_duplicate_penalty/mean': 0.9923731684684753, 'rewards/near_duplicate_penalty/std': 0.0040103597566485405, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923049211502075, 'rewards/distinct_2/std': 0.015163577161729336, 'rewards/total_composite/mean': 0.08127839863300323, 'rewards/total_composite/std': 0.06733237206935883, 'reward': 0.08127839863300323, 'reward_std': 0.06733237206935883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21238939464092255, 'sampling/sampling_logp_difference/max': 1.9760198593139648, 'sampling/importance_sampling_ratio/min': 0.13861986994743347, 'sampling/importance_sampling_ratio/mean': 1.0330545902252197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4870296269655228, 'clip_ratio/low_mean': 0.13166139274835587, 'clip_ratio/low_min': 0.13166139274835587, 'clip_ratio/high_mean': 0.08023050799965858, 'clip_ratio/high_max': 0.08023050799965858, 'clip_ratio/region_mean': 0.21189190074801445, 'reward_total_mean': 0.08127839863300323, 'reward_meter_mean': 0.34000200033187866, 'reward_meter_std': 0.24787180125713348, 'reward_count_adherence_mean': 0.953125, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9859374761581421, 'reward_count_floor_std': 0.019408106803894043, 'reward_lexical_plausibility_mean': 0.9986978769302368, 'reward_lexical_plausibility_std': 0.0036828548181802034, 'reward_repeat_penalty_mean': 0.9847592115402222, 'reward_repeat_penalty_std': 0.01706889271736145, 'reward_repeat_floor_mean': 0.9981634616851807, 'reward_repeat_floor_std': 0.0017077922821044922, 'reward_near_duplicate_penalty_mean': 0.9923731684684753, 'reward_near_duplicate_penalty_std': 0.0040103597566485405, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923049211502075, 'reward_distinct_2_std': 0.015163577161729336, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.08127839863300323, 'reward_total_composite_std': 0.06733237206935883, 'epoch': 0.0}

  2%|▏         | 76/3300 [14:37<9:03:48, 10.12s/it]INFO 04-12 16:13:03 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 77/3300 [14:46<8:44:35,  9.77s/it]2026-04-12 16:13:12,191 | INFO | train_grpo_train | metrics_logged mode=train step=77

                                                   
{'loss': 0.1016, 'grad_norm': 17.043352127075195, 'learning_rate': 9.76969696969697e-06, 'num_tokens': 147045.0, 'completions/mean_length': 47.25, 'completions/min_length': 34.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.44776201248168945, 'rewards/meter/std': 0.3801688253879547, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9914492964744568, 'rewards/lexical_plausibility/std': 0.024185042828321457, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9849174618721008, 'rewards/repeat_penalty/std': 0.03660213202238083, 'rewards/repeat_floor/mean': 0.998343825340271, 'rewards/repeat_floor/std': 0.0037844826001673937, 'rewards/near_duplicate_penalty/mean': 0.9953689575195312, 'rewards/near_duplicate_penalty/std': 0.0075595383532345295, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.16858382523059845, 'rewards/total_composite/std': 0.1656939685344696, 'reward': 0.16858382523059845, 'reward_std': 0.1656939685344696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.219539076089859, 'sampling/sampling_logp_difference/max': 1.158310890197754, 'sampling/importance_sampling_ratio/min': 0.3140161335468292, 'sampling/importance_sampling_ratio/mean': 1.0317708253860474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.245242640376091, 'clip_ratio/low_mean': 0.1158177200704813, 'clip_ratio/low_min': 0.1158177200704813, 'clip_ratio/high_mean': 0.09649939276278019, 'clip_ratio/high_max': 0.09649939276278019, 'clip_ratio/region_mean': 0.2123171128332615, 'reward_total_mean': 0.16858382523059845, 'reward_meter_mean': 0.44776201248168945, 'reward_meter_std': 0.3801688253879547, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9914492964744568, 'reward_lexical_plausibility_std': 0.024185042828321457, 'reward_repeat_penalty_mean': 0.9849174618721008, 'reward_repeat_penalty_std': 0.03660213202238083, 'reward_repeat_floor_mean': 0.998343825340271, 'reward_repeat_floor_std': 0.0037844826001673937, 'reward_near_duplicate_penalty_mean': 0.9953689575195312, 'reward_near_duplicate_penalty_std': 0.0075595383532345295, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.16858382523059845, 'reward_total_composite_std': 0.1656939685344696, 'epoch': 0.0}

  2%|▏         | 77/3300 [14:46<8:44:35,  9.77s/it]INFO 04-12 16:13:12 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 78/3300 [14:54<8:23:27,  9.38s/it]2026-04-12 16:13:20,650 | INFO | train_grpo_train | metrics_logged mode=train step=78

                                                   
{'loss': 0.0846, 'grad_norm': 15.409075736999512, 'learning_rate': 9.766666666666667e-06, 'num_tokens': 148889.0, 'completions/mean_length': 51.5, 'completions/min_length': 36.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7258580923080444, 'rewards/meter/std': 0.44764217734336853, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9095508456230164, 'rewards/lexical_plausibility/std': 0.13177736103534698, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3464488387107849, 'rewards/repeat_penalty/mean': 0.9927314519882202, 'rewards/repeat_penalty/std': 0.01101653277873993, 'rewards/repeat_floor/mean': 0.9989097118377686, 'rewards/repeat_floor/std': 0.0016524747479707003, 'rewards/near_duplicate_penalty/mean': 0.9927314519882202, 'rewards/near_duplicate_penalty/std': 0.01101653277873993, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.46346306800842285, 'rewards/total_composite/std': 0.3897399604320526, 'reward': 0.46346306800842285, 'reward_std': 0.3897399306297302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21735115349292755, 'sampling/sampling_logp_difference/max': 1.364924430847168, 'sampling/importance_sampling_ratio/min': 0.2553999722003937, 'sampling/importance_sampling_ratio/mean': 1.0262371301651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2506941705942154, 'clip_ratio/low_mean': 0.14349130354821682, 'clip_ratio/low_min': 0.14349130354821682, 'clip_ratio/high_mean': 0.08258798345923424, 'clip_ratio/high_max': 0.08258798345923424, 'clip_ratio/region_mean': 0.22607928700745106, 'reward_total_mean': 0.46346306800842285, 'reward_meter_mean': 0.7258580923080444, 'reward_meter_std': 0.44764217734336853, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9095508456230164, 'reward_lexical_plausibility_std': 0.13177736103534698, 'reward_repeat_penalty_mean': 0.9927314519882202, 'reward_repeat_penalty_std': 0.01101653277873993, 'reward_repeat_floor_mean': 0.9989097118377686, 'reward_repeat_floor_std': 0.0016524747479707003, 'reward_near_duplicate_penalty_mean': 0.9927314519882202, 'reward_near_duplicate_penalty_std': 0.01101653277873993, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3464488387107849, 'reward_total_composite_mean': 0.46346306800842285, 'reward_total_composite_std': 0.3897399604320526, 'epoch': 0.0}

  2%|▏         | 78/3300 [14:54<8:23:27,  9.38s/it]INFO 04-12 16:13:20 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 79/3300 [15:03<8:15:42,  9.23s/it]2026-04-12 16:13:29,556 | INFO | train_grpo_train | metrics_logged mode=train step=79

                                                   
{'loss': 0.1152, 'grad_norm': 18.246423721313477, 'learning_rate': 9.763636363636365e-06, 'num_tokens': 150407.0, 'completions/mean_length': 43.75, 'completions/min_length': 36.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5016579627990723, 'rewards/meter/std': 0.39960169792175293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9975490570068359, 'rewards/lexical_plausibility/std': 0.006932419259101152, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9798001050949097, 'rewards/repeat_penalty/std': 0.02479006163775921, 'rewards/repeat_floor/mean': 0.9974507689476013, 'rewards/repeat_floor/std': 0.002905119676142931, 'rewards/near_duplicate_penalty/mean': 0.9878129363059998, 'rewards/near_duplicate_penalty/std': 0.018003087490797043, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.24432101845741272, 'rewards/total_composite/std': 0.18906839191913605, 'reward': 0.24432101845741272, 'reward_std': 0.18906840682029724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22122344374656677, 'sampling/sampling_logp_difference/max': 1.349376916885376, 'sampling/importance_sampling_ratio/min': 0.2594018280506134, 'sampling/importance_sampling_ratio/mean': 1.0465655326843262, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.168430730700493, 'clip_ratio/low_mean': 0.10910411970689893, 'clip_ratio/low_min': 0.10910411970689893, 'clip_ratio/high_mean': 0.11801265180110931, 'clip_ratio/high_max': 0.11801265180110931, 'clip_ratio/region_mean': 0.22711677150800824, 'reward_total_mean': 0.24432101845741272, 'reward_meter_mean': 0.5016579627990723, 'reward_meter_std': 0.39960169792175293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9975490570068359, 'reward_lexical_plausibility_std': 0.006932419259101152, 'reward_repeat_penalty_mean': 0.9798001050949097, 'reward_repeat_penalty_std': 0.02479006163775921, 'reward_repeat_floor_mean': 0.9974507689476013, 'reward_repeat_floor_std': 0.002905119676142931, 'reward_near_duplicate_penalty_mean': 0.9878129363059998, 'reward_near_duplicate_penalty_std': 0.018003087490797043, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.24432101845741272, 'reward_total_composite_std': 0.18906839191913605, 'epoch': 0.0}

  2%|▏         | 79/3300 [15:03<8:15:42,  9.23s/it]INFO 04-12 16:13:29 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 80/3300 [15:12<8:10:59,  9.15s/it]2026-04-12 16:13:38,506 | INFO | train_grpo_train | metrics_logged mode=train step=80

                                                   
{'loss': 0.0199, 'grad_norm': 10.943520545959473, 'learning_rate': 9.760606060606062e-06, 'num_tokens': 152386.0, 'completions/mean_length': 85.375, 'completions/min_length': 54.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.7308220863342285, 'rewards/meter/std': 0.29062309861183167, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9831444621086121, 'rewards/repeat_penalty/std': 0.024900658056139946, 'rewards/repeat_floor/mean': 0.9981099367141724, 'rewards/repeat_floor/std': 0.0025812743697315454, 'rewards/near_duplicate_penalty/mean': 0.993998110294342, 'rewards/near_duplicate_penalty/std': 0.006261636968702078, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.989001989364624, 'rewards/distinct_2/std': 0.02037539705634117, 'rewards/total_composite/mean': 0.2483464777469635, 'rewards/total_composite/std': 0.12048420310020447, 'reward': 0.2483464777469635, 'reward_std': 0.12048420310020447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21638713777065277, 'sampling/sampling_logp_difference/max': 1.9478286504745483, 'sampling/importance_sampling_ratio/min': 0.14258332550525665, 'sampling/importance_sampling_ratio/mean': 1.0444848537445068, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8526629209518433, 'clip_ratio/low_mean': 0.10550958663225174, 'clip_ratio/low_min': 0.10550958663225174, 'clip_ratio/high_mean': 0.10242243856191635, 'clip_ratio/high_max': 0.10242243856191635, 'clip_ratio/region_mean': 0.2079320251941681, 'reward_total_mean': 0.2483464777469635, 'reward_meter_mean': 0.7308220863342285, 'reward_meter_std': 0.29062309861183167, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9831444621086121, 'reward_repeat_penalty_std': 0.024900658056139946, 'reward_repeat_floor_mean': 0.9981099367141724, 'reward_repeat_floor_std': 0.0025812743697315454, 'reward_near_duplicate_penalty_mean': 0.993998110294342, 'reward_near_duplicate_penalty_std': 0.006261636968702078, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.989001989364624, 'reward_distinct_2_std': 0.02037539705634117, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.2483464777469635, 'reward_total_composite_std': 0.12048420310020447, 'epoch': 0.0}

  2%|▏         | 80/3300 [15:12<8:10:59,  9.15s/it]INFO 04-12 16:13:38 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 81/3300 [15:21<8:11:39,  9.16s/it]2026-04-12 16:13:47,705 | INFO | train_grpo_train | metrics_logged mode=train step=81

                                                   
{'loss': 0.126, 'grad_norm': 14.844765663146973, 'learning_rate': 9.757575757575758e-06, 'num_tokens': 154020.0, 'completions/mean_length': 49.25, 'completions/min_length': 35.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.610368549823761, 'rewards/meter/std': 0.32375213503837585, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9947047233581543, 'rewards/repeat_penalty/std': 0.009650482796132565, 'rewards/repeat_floor/mean': 0.9992057085037231, 'rewards/repeat_floor/std': 0.001447572954930365, 'rewards/near_duplicate_penalty/mean': 0.9947047233581543, 'rewards/near_duplicate_penalty/std': 0.009650482796132565, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25115853548049927, 'rewards/total_composite/std': 0.14330866932868958, 'reward': 0.25115853548049927, 'reward_std': 0.14330866932868958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21455712616443634, 'sampling/sampling_logp_difference/max': 1.2710590362548828, 'sampling/importance_sampling_ratio/min': 0.28053438663482666, 'sampling/importance_sampling_ratio/mean': 1.0214149951934814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4596636295318604, 'clip_ratio/low_mean': 0.09379999339580536, 'clip_ratio/low_min': 0.09379999339580536, 'clip_ratio/high_mean': 0.11031364649534225, 'clip_ratio/high_max': 0.11031364649534225, 'clip_ratio/region_mean': 0.2041136398911476, 'reward_total_mean': 0.25115853548049927, 'reward_meter_mean': 0.610368549823761, 'reward_meter_std': 0.32375213503837585, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9947047233581543, 'reward_repeat_penalty_std': 0.009650482796132565, 'reward_repeat_floor_mean': 0.9992057085037231, 'reward_repeat_floor_std': 0.001447572954930365, 'reward_near_duplicate_penalty_mean': 0.9947047233581543, 'reward_near_duplicate_penalty_std': 0.009650482796132565, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.25115853548049927, 'reward_total_composite_std': 0.14330866932868958, 'epoch': 0.0}

  2%|▏         | 81/3300 [15:21<8:11:39,  9.16s/it]INFO 04-12 16:13:47 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 82/3300 [15:35<9:31:47, 10.66s/it]2026-04-12 16:14:01,861 | INFO | train_grpo_train | metrics_logged mode=train step=82

                                                   
{'loss': 0.1333, 'grad_norm': 6.31158447265625, 'learning_rate': 9.754545454545455e-06, 'num_tokens': 157714.0, 'completions/mean_length': 279.75, 'completions/min_length': 215.0, 'completions/max_length': 392.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 279.75, 'completions/min_terminated_length': 215.0, 'completions/max_terminated_length': 392.0, 'rewards/meter/mean': 0.3813129663467407, 'rewards/meter/std': 0.27879804372787476, 'rewards/count_adherence/mean': 0.886363685131073, 'rewards/count_adherence/std': 0.06428244709968567, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9659091234207153, 'rewards/count_floor/std': 0.019284740090370178, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.14500615000724792, 'rewards/repeat_penalty/mean': 0.9832416772842407, 'rewards/repeat_penalty/std': 0.012843562290072441, 'rewards/repeat_floor/mean': 0.9980295300483704, 'rewards/repeat_floor/std': 0.0013376965653151274, 'rewards/near_duplicate_penalty/mean': 0.9924200773239136, 'rewards/near_duplicate_penalty/std': 0.004430524539202452, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9907390475273132, 'rewards/distinct_2/std': 0.010929741896688938, 'rewards/total_composite/mean': 0.11373062431812286, 'rewards/total_composite/std': 0.09236416220664978, 'reward': 0.11373062431812286, 'reward_std': 0.09236415475606918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2386922985315323, 'sampling/sampling_logp_difference/max': 2.0198888778686523, 'sampling/importance_sampling_ratio/min': 0.13267020881175995, 'sampling/importance_sampling_ratio/mean': 1.0501742362976074, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.024958550930023, 'clip_ratio/low_mean': 0.12591862678527832, 'clip_ratio/low_min': 0.12591862678527832, 'clip_ratio/high_mean': 0.09771881811320782, 'clip_ratio/high_max': 0.09771881811320782, 'clip_ratio/region_mean': 0.22363744489848614, 'reward_total_mean': 0.11373062431812286, 'reward_meter_mean': 0.3813129663467407, 'reward_meter_std': 0.27879804372787476, 'reward_count_adherence_mean': 0.886363685131073, 'reward_count_adherence_std': 0.06428244709968567, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9659091234207153, 'reward_count_floor_std': 0.019284740090370178, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9832416772842407, 'reward_repeat_penalty_std': 0.012843562290072441, 'reward_repeat_floor_mean': 0.9980295300483704, 'reward_repeat_floor_std': 0.0013376965653151274, 'reward_near_duplicate_penalty_mean': 0.9924200773239136, 'reward_near_duplicate_penalty_std': 0.004430524539202452, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9907390475273132, 'reward_distinct_2_std': 0.010929741896688938, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.14500615000724792, 'reward_total_composite_mean': 0.11373062431812286, 'reward_total_composite_std': 0.09236416220664978, 'epoch': 0.0}

  2%|▏         | 82/3300 [15:35<9:31:47, 10.66s/it]INFO 04-12 16:14:02 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 83/3300 [15:43<8:50:27,  9.89s/it]2026-04-12 16:14:09,962 | INFO | train_grpo_train | metrics_logged mode=train step=83

                                                   
{'loss': -0.0137, 'grad_norm': 18.539670944213867, 'learning_rate': 9.751515151515152e-06, 'num_tokens': 159231.0, 'completions/mean_length': 36.625, 'completions/min_length': 28.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6339971423149109, 'rewards/meter/std': 0.38485148549079895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.2503390908241272, 'rewards/repeat_penalty/mean': 0.9952200651168823, 'rewards/repeat_penalty/std': 0.010680274106562138, 'rewards/repeat_floor/mean': 0.9992830157279968, 'rewards/repeat_floor/std': 0.0016020411858335137, 'rewards/near_duplicate_penalty/mean': 0.9952200651168823, 'rewards/near_duplicate_penalty/std': 0.010680274106562138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4572411775588989, 'rewards/total_composite/std': 0.3553715944290161, 'reward': 0.4572411775588989, 'reward_std': 0.3553715646266937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20119155943393707, 'sampling/sampling_logp_difference/max': 1.817361831665039, 'sampling/importance_sampling_ratio/min': 0.32714101672172546, 'sampling/importance_sampling_ratio/mean': 1.04038667678833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0735863745212555, 'clip_ratio/low_mean': 0.062157207168638706, 'clip_ratio/low_min': 0.062157207168638706, 'clip_ratio/high_mean': 0.1074504666030407, 'clip_ratio/high_max': 0.1074504666030407, 'clip_ratio/region_mean': 0.1696076737716794, 'reward_total_mean': 0.4572411775588989, 'reward_meter_mean': 0.6339971423149109, 'reward_meter_std': 0.38485148549079895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9952200651168823, 'reward_repeat_penalty_std': 0.010680274106562138, 'reward_repeat_floor_mean': 0.9992830157279968, 'reward_repeat_floor_std': 0.0016020411858335137, 'reward_near_duplicate_penalty_mean': 0.9952200651168823, 'reward_near_duplicate_penalty_std': 0.010680274106562138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.2503390908241272, 'reward_total_composite_mean': 0.4572411775588989, 'reward_total_composite_std': 0.3553715944290161, 'epoch': 0.0}

  3%|▎         | 83/3300 [15:43<8:50:27,  9.89s/it]INFO 04-12 16:14:10 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 84/3300 [15:52<8:29:49,  9.51s/it]2026-04-12 16:14:18,582 | INFO | train_grpo_train | metrics_logged mode=train step=84

                                                   
{'loss': 0.1023, 'grad_norm': 20.06077766418457, 'learning_rate': 9.74848484848485e-06, 'num_tokens': 161106.0, 'completions/mean_length': 45.375, 'completions/min_length': 37.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6297622919082642, 'rewards/meter/std': 0.38552671670913696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9930555820465088, 'rewards/lexical_plausibility/std': 0.01285861898213625, 'rewards/judge_quality/mean': 0.492499977350235, 'rewards/judge_quality/std': 0.2796298861503601, 'rewards/repeat_penalty/mean': 0.9873254299163818, 'rewards/repeat_penalty/std': 0.01144490484148264, 'rewards/repeat_floor/mean': 0.9980988502502441, 'rewards/repeat_floor/std': 0.0017167360056191683, 'rewards/near_duplicate_penalty/mean': 0.9873254299163818, 'rewards/near_duplicate_penalty/std': 0.01144490484148264, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.36389321088790894, 'rewards/total_composite/std': 0.32425564527511597, 'reward': 0.36389321088790894, 'reward_std': 0.32425564527511597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20729809999465942, 'sampling/sampling_logp_difference/max': 1.8689308166503906, 'sampling/importance_sampling_ratio/min': 0.15428853034973145, 'sampling/importance_sampling_ratio/mean': 1.006610631942749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.140326365828514, 'clip_ratio/low_mean': 0.12810066901147366, 'clip_ratio/low_min': 0.12810066901147366, 'clip_ratio/high_mean': 0.07445101346820593, 'clip_ratio/high_max': 0.07445101346820593, 'clip_ratio/region_mean': 0.20255168247967958, 'reward_total_mean': 0.36389321088790894, 'reward_meter_mean': 0.6297622919082642, 'reward_meter_std': 0.38552671670913696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9930555820465088, 'reward_lexical_plausibility_std': 0.01285861898213625, 'reward_repeat_penalty_mean': 0.9873254299163818, 'reward_repeat_penalty_std': 0.01144490484148264, 'reward_repeat_floor_mean': 0.9980988502502441, 'reward_repeat_floor_std': 0.0017167360056191683, 'reward_near_duplicate_penalty_mean': 0.9873254299163818, 'reward_near_duplicate_penalty_std': 0.01144490484148264, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.492499977350235, 'reward_judge_quality_std': 0.2796298861503601, 'reward_total_composite_mean': 0.36389321088790894, 'reward_total_composite_std': 0.32425564527511597, 'epoch': 0.0}

  3%|▎         | 84/3300 [15:52<8:29:49,  9.51s/it]INFO 04-12 16:14:18 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 85/3300 [16:02<8:45:58,  9.82s/it]2026-04-12 16:14:29,108 | INFO | train_grpo_train | metrics_logged mode=train step=85

                                                   
{'loss': -0.0317, 'grad_norm': 19.909093856811523, 'learning_rate': 9.745454545454547e-06, 'num_tokens': 162453.0, 'completions/mean_length': 24.375, 'completions/min_length': 19.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.8942579030990601, 'rewards/meter/std': 0.18176580965518951, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8967865109443665, 'rewards/lexical_plausibility/std': 0.2128681242465973, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.343386709690094, 'rewards/total_composite/std': 0.2046266794204712, 'reward': 0.343386709690094, 'reward_std': 0.2046266794204712, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2179541289806366, 'sampling/sampling_logp_difference/max': 2.4439144134521484, 'sampling/importance_sampling_ratio/min': 0.0868203341960907, 'sampling/importance_sampling_ratio/mean': 1.0402379035949707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2473599910736084, 'clip_ratio/low_mean': 0.16301407292485237, 'clip_ratio/low_min': 0.16301407292485237, 'clip_ratio/high_mean': 0.046408251859247684, 'clip_ratio/high_max': 0.046408251859247684, 'clip_ratio/region_mean': 0.20942232478410006, 'reward_total_mean': 0.343386709690094, 'reward_meter_mean': 0.8942579030990601, 'reward_meter_std': 0.18176580965518951, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8967865109443665, 'reward_lexical_plausibility_std': 0.2128681242465973, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.343386709690094, 'reward_total_composite_std': 0.2046266794204712, 'epoch': 0.0}

  3%|▎         | 85/3300 [16:02<8:45:58,  9.82s/it]INFO 04-12 16:14:29 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 86/3300 [16:10<8:13:10,  9.21s/it]2026-04-12 16:14:36,893 | INFO | train_grpo_train | metrics_logged mode=train step=86

                                                   
{'loss': 0.0513, 'grad_norm': 24.302440643310547, 'learning_rate': 9.742424242424244e-06, 'num_tokens': 164038.0, 'completions/mean_length': 34.125, 'completions/min_length': 29.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.596794843673706, 'rewards/meter/std': 0.4008859395980835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 1.0, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.37639322876930237, 'rewards/total_composite/std': 0.2644297480583191, 'reward': 0.37639322876930237, 'reward_std': 0.2644297480583191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24157695472240448, 'sampling/sampling_logp_difference/max': 2.1380763053894043, 'sampling/importance_sampling_ratio/min': 0.1178813949227333, 'sampling/importance_sampling_ratio/mean': 0.9839524626731873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9065810143947601, 'clip_ratio/low_mean': 0.070782121270895, 'clip_ratio/low_min': 0.070782121270895, 'clip_ratio/high_mean': 0.10850926674902439, 'clip_ratio/high_max': 0.10850926674902439, 'clip_ratio/region_mean': 0.1792913880199194, 'reward_total_mean': 0.37639322876930237, 'reward_meter_mean': 0.596794843673706, 'reward_meter_std': 0.4008859395980835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 1.0, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.37639322876930237, 'reward_total_composite_std': 0.2644297480583191, 'epoch': 0.0}

  3%|▎         | 86/3300 [16:10<8:13:10,  9.21s/it]INFO 04-12 16:14:37 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 87/3300 [16:19<8:10:15,  9.16s/it]2026-04-12 16:14:45,928 | INFO | train_grpo_train | metrics_logged mode=train step=87

                                                   
{'loss': 0.1047, 'grad_norm': 17.62166404724121, 'learning_rate': 9.739393939393941e-06, 'num_tokens': 165884.0, 'completions/mean_length': 58.75, 'completions/min_length': 46.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.42708075046539307, 'rewards/meter/std': 0.293794184923172, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9741393327713013, 'rewards/lexical_plausibility/std': 0.05264858156442642, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9964795112609863, 'rewards/repeat_penalty/std': 0.005044713616371155, 'rewards/repeat_floor/mean': 0.99947190284729, 'rewards/repeat_floor/std': 0.0007567146676592529, 'rewards/near_duplicate_penalty/mean': 0.9964795112609863, 'rewards/near_duplicate_penalty/std': 0.005044713616371155, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18397322297096252, 'rewards/total_composite/std': 0.13547933101654053, 'reward': 0.18397322297096252, 'reward_std': 0.13547933101654053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2125231772661209, 'sampling/sampling_logp_difference/max': 2.169562816619873, 'sampling/importance_sampling_ratio/min': 0.1142275482416153, 'sampling/importance_sampling_ratio/mean': 1.0063490867614746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.717762142419815, 'clip_ratio/low_mean': 0.11891158856451511, 'clip_ratio/low_min': 0.11891158856451511, 'clip_ratio/high_mean': 0.052052177488803864, 'clip_ratio/high_max': 0.052052177488803864, 'clip_ratio/region_mean': 0.17096376605331898, 'reward_total_mean': 0.18397322297096252, 'reward_meter_mean': 0.42708075046539307, 'reward_meter_std': 0.293794184923172, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9741393327713013, 'reward_lexical_plausibility_std': 0.05264858156442642, 'reward_repeat_penalty_mean': 0.9964795112609863, 'reward_repeat_penalty_std': 0.005044713616371155, 'reward_repeat_floor_mean': 0.99947190284729, 'reward_repeat_floor_std': 0.0007567146676592529, 'reward_near_duplicate_penalty_mean': 0.9964795112609863, 'reward_near_duplicate_penalty_std': 0.005044713616371155, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.18397322297096252, 'reward_total_composite_std': 0.13547933101654053, 'epoch': 0.0}

  3%|▎         | 87/3300 [16:19<8:10:15,  9.16s/it]INFO 04-12 16:14:46 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 88/3300 [16:32<9:04:20, 10.17s/it]2026-04-12 16:14:58,462 | INFO | train_grpo_train | metrics_logged mode=train step=88

                                                   
{'loss': 0.0324, 'grad_norm': 6.4770917892456055, 'learning_rate': 9.736363636363637e-06, 'num_tokens': 169572.0, 'completions/mean_length': 235.0, 'completions/min_length': 162.0, 'completions/max_length': 361.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 235.0, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 361.0, 'rewards/meter/mean': 0.9215198755264282, 'rewards/meter/std': 0.09319359064102173, 'rewards/count_adherence/mean': 0.953125, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9859374761581421, 'rewards/count_floor/std': 0.019408106803894043, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9822944402694702, 'rewards/repeat_penalty/std': 0.016634171828627586, 'rewards/repeat_floor/mean': 0.9981192946434021, 'rewards/repeat_floor/std': 0.0015382225392386317, 'rewards/near_duplicate_penalty/mean': 0.995309591293335, 'rewards/near_duplicate_penalty/std': 0.002218200359493494, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9869205951690674, 'rewards/distinct_2/std': 0.016350824385881424, 'rewards/total_composite/mean': 0.2676911950111389, 'rewards/total_composite/std': 0.07761240750551224, 'reward': 0.2676911950111389, 'reward_std': 0.07761240750551224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20899732410907745, 'sampling/sampling_logp_difference/max': 1.692124366760254, 'sampling/importance_sampling_ratio/min': 0.18412795662879944, 'sampling/importance_sampling_ratio/mean': 1.0465455055236816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8870257139205933, 'clip_ratio/low_mean': 0.0739701446145773, 'clip_ratio/low_min': 0.0739701446145773, 'clip_ratio/high_mean': 0.13612034544348717, 'clip_ratio/high_max': 0.13612034544348717, 'clip_ratio/region_mean': 0.21009049005806446, 'reward_total_mean': 0.2676911950111389, 'reward_meter_mean': 0.9215198755264282, 'reward_meter_std': 0.09319359064102173, 'reward_count_adherence_mean': 0.953125, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9859374761581421, 'reward_count_floor_std': 0.019408106803894043, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9822944402694702, 'reward_repeat_penalty_std': 0.016634171828627586, 'reward_repeat_floor_mean': 0.9981192946434021, 'reward_repeat_floor_std': 0.0015382225392386317, 'reward_near_duplicate_penalty_mean': 0.995309591293335, 'reward_near_duplicate_penalty_std': 0.002218200359493494, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9869205951690674, 'reward_distinct_2_std': 0.016350824385881424, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.2676911950111389, 'reward_total_composite_std': 0.07761240750551224, 'epoch': 0.0}

  3%|▎         | 88/3300 [16:32<9:04:20, 10.17s/it]INFO 04-12 16:14:58 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 89/3300 [16:40<8:34:30,  9.61s/it]2026-04-12 16:15:06,781 | INFO | train_grpo_train | metrics_logged mode=train step=89

                                                   
{'loss': -0.0413, 'grad_norm': 15.227660179138184, 'learning_rate': 9.733333333333334e-06, 'num_tokens': 171157.0, 'completions/mean_length': 46.125, 'completions/min_length': 39.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8218026757240295, 'rewards/meter/std': 0.30785462260246277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9754703640937805, 'rewards/lexical_plausibility/std': 0.06938032805919647, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.28630343079566956, 'rewards/repeat_penalty/mean': 0.9967880845069885, 'rewards/repeat_penalty/std': 0.006415999960154295, 'rewards/repeat_floor/mean': 0.9995182156562805, 'rewards/repeat_floor/std': 0.0009623938240110874, 'rewards/near_duplicate_penalty/mean': 0.9967880845069885, 'rewards/near_duplicate_penalty/std': 0.006415999960154295, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.44428551197052, 'rewards/total_composite/std': 0.32935091853141785, 'reward': 0.44428551197052, 'reward_std': 0.32935088872909546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2216944545507431, 'sampling/sampling_logp_difference/max': 1.4072818756103516, 'sampling/importance_sampling_ratio/min': 0.24480780959129333, 'sampling/importance_sampling_ratio/mean': 1.0472164154052734, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6710088551044464, 'clip_ratio/low_mean': 0.14506504870951176, 'clip_ratio/low_min': 0.14506504870951176, 'clip_ratio/high_mean': 0.05743970721960068, 'clip_ratio/high_max': 0.05743970721960068, 'clip_ratio/region_mean': 0.20250475592911243, 'reward_total_mean': 0.44428551197052, 'reward_meter_mean': 0.8218026757240295, 'reward_meter_std': 0.30785462260246277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9754703640937805, 'reward_lexical_plausibility_std': 0.06938032805919647, 'reward_repeat_penalty_mean': 0.9967880845069885, 'reward_repeat_penalty_std': 0.006415999960154295, 'reward_repeat_floor_mean': 0.9995182156562805, 'reward_repeat_floor_std': 0.0009623938240110874, 'reward_near_duplicate_penalty_mean': 0.9967880845069885, 'reward_near_duplicate_penalty_std': 0.006415999960154295, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.28630343079566956, 'reward_total_composite_mean': 0.44428551197052, 'reward_total_composite_std': 0.32935091853141785, 'epoch': 0.0}

  3%|▎         | 89/3300 [16:40<8:34:30,  9.61s/it]INFO 04-12 16:15:07 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 90/3300 [16:48<8:12:06,  9.20s/it]2026-04-12 16:15:15,010 | INFO | train_grpo_train | metrics_logged mode=train step=90

                                                   
{'loss': 0.1647, 'grad_norm': 21.623760223388672, 'learning_rate': 9.730303030303031e-06, 'num_tokens': 172686.0, 'completions/mean_length': 41.125, 'completions/min_length': 30.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.32222986221313477, 'rewards/meter/std': 0.24403588473796844, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9518775939941406, 'rewards/lexical_plausibility/std': 0.10010452568531036, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.3620477616786957, 'rewards/repeat_penalty/mean': 0.9765366315841675, 'rewards/repeat_penalty/std': 0.03547777608036995, 'rewards/repeat_floor/mean': 0.9969081878662109, 'rewards/repeat_floor/std': 0.004214660730212927, 'rewards/near_duplicate_penalty/mean': 0.9841955900192261, 'rewards/near_duplicate_penalty/std': 0.017114752903580666, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.23835045099258423, 'rewards/total_composite/std': 0.25593531131744385, 'reward': 0.23835045099258423, 'reward_std': 0.25593531131744385, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2359987199306488, 'sampling/sampling_logp_difference/max': 1.8980355262756348, 'sampling/importance_sampling_ratio/min': 0.14986273646354675, 'sampling/importance_sampling_ratio/mean': 0.9931842684745789, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0835602059960365, 'clip_ratio/low_mean': 0.13202047906816006, 'clip_ratio/low_min': 0.13202047906816006, 'clip_ratio/high_mean': 0.0401960788294673, 'clip_ratio/high_max': 0.0401960788294673, 'clip_ratio/region_mean': 0.17221655789762735, 'reward_total_mean': 0.23835045099258423, 'reward_meter_mean': 0.32222986221313477, 'reward_meter_std': 0.24403588473796844, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9518775939941406, 'reward_lexical_plausibility_std': 0.10010452568531036, 'reward_repeat_penalty_mean': 0.9765366315841675, 'reward_repeat_penalty_std': 0.03547777608036995, 'reward_repeat_floor_mean': 0.9969081878662109, 'reward_repeat_floor_std': 0.004214660730212927, 'reward_near_duplicate_penalty_mean': 0.9841955900192261, 'reward_near_duplicate_penalty_std': 0.017114752903580666, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.3620477616786957, 'reward_total_composite_mean': 0.23835045099258423, 'reward_total_composite_std': 0.25593531131744385, 'epoch': 0.0}

  3%|▎         | 90/3300 [16:48<8:12:06,  9.20s/it]INFO 04-12 16:15:15 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 91/3300 [16:59<8:40:58,  9.74s/it]2026-04-12 16:15:26,018 | INFO | train_grpo_train | metrics_logged mode=train step=91

                                                   
{'loss': 0.0241, 'grad_norm': 6.1096696853637695, 'learning_rate': 9.727272727272728e-06, 'num_tokens': 175756.0, 'completions/mean_length': 174.75, 'completions/min_length': 159.0, 'completions/max_length': 188.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 174.75, 'completions/min_terminated_length': 159.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.878669023513794, 'rewards/meter/std': 0.24484211206436157, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9925000071525574, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9234609603881836, 'rewards/lexical_plausibility/std': 0.1143912598490715, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.9974187016487122, 'rewards/repeat_penalty/std': 0.0032706044148653746, 'rewards/repeat_floor/mean': 0.9996128082275391, 'rewards/repeat_floor/std': 0.0004905856912955642, 'rewards/near_duplicate_penalty/mean': 0.9974187016487122, 'rewards/near_duplicate_penalty/std': 0.0032706044148653746, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33547407388687134, 'rewards/total_composite/std': 0.24134448170661926, 'reward': 0.33547407388687134, 'reward_std': 0.24134448170661926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19223442673683167, 'sampling/sampling_logp_difference/max': 1.791189193725586, 'sampling/importance_sampling_ratio/min': 0.16676174104213715, 'sampling/importance_sampling_ratio/mean': 1.023423194885254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.378643646836281, 'clip_ratio/low_mean': 0.08139126002788544, 'clip_ratio/low_min': 0.08139126002788544, 'clip_ratio/high_mean': 0.09280398488044739, 'clip_ratio/high_max': 0.09280398488044739, 'clip_ratio/region_mean': 0.17419524490833282, 'reward_total_mean': 0.33547407388687134, 'reward_meter_mean': 0.878669023513794, 'reward_meter_std': 0.24484211206436157, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9925000071525574, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9234609603881836, 'reward_lexical_plausibility_std': 0.1143912598490715, 'reward_repeat_penalty_mean': 0.9974187016487122, 'reward_repeat_penalty_std': 0.0032706044148653746, 'reward_repeat_floor_mean': 0.9996128082275391, 'reward_repeat_floor_std': 0.0004905856912955642, 'reward_near_duplicate_penalty_mean': 0.9974187016487122, 'reward_near_duplicate_penalty_std': 0.0032706044148653746, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.33547407388687134, 'reward_total_composite_std': 0.24134448170661926, 'epoch': 0.0}

  3%|▎         | 91/3300 [16:59<8:40:58,  9.74s/it]INFO 04-12 16:15:26 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 92/3300 [17:13<9:47:39, 10.99s/it]2026-04-12 16:15:39,925 | INFO | train_grpo_train | metrics_logged mode=train step=92

                                                   
{'loss': 0.1676, 'grad_norm': 5.645833969116211, 'learning_rate': 9.724242424242426e-06, 'num_tokens': 179653.0, 'completions/mean_length': 269.125, 'completions/min_length': 180.0, 'completions/max_length': 370.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 269.125, 'completions/min_terminated_length': 180.0, 'completions/max_terminated_length': 370.0, 'rewards/meter/mean': 0.2120237946510315, 'rewards/meter/std': 0.1717546433210373, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.05345224589109421, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9700000286102295, 'rewards/count_floor/std': 0.01603567600250244, 'rewards/lexical_plausibility/mean': 0.9818199872970581, 'rewards/lexical_plausibility/std': 0.03753276914358139, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.988950252532959, 'rewards/repeat_penalty/std': 0.009890228509902954, 'rewards/repeat_floor/mean': 0.9986698627471924, 'rewards/repeat_floor/std': 0.001043530646711588, 'rewards/near_duplicate_penalty/mean': 0.9944734573364258, 'rewards/near_duplicate_penalty/std': 0.0028089345432817936, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9944319725036621, 'rewards/distinct_2/std': 0.0076876552775502205, 'rewards/total_composite/mean': 0.05339657515287399, 'rewards/total_composite/std': 0.05815418064594269, 'reward': 0.05339657515287399, 'reward_std': 0.05815417692065239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2234933078289032, 'sampling/sampling_logp_difference/max': 2.155268669128418, 'sampling/importance_sampling_ratio/min': 0.11587205529212952, 'sampling/importance_sampling_ratio/mean': 1.033098816871643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0099581480026245, 'clip_ratio/low_mean': 0.13186335749924183, 'clip_ratio/low_min': 0.13186335749924183, 'clip_ratio/high_mean': 0.048218170180916786, 'clip_ratio/high_max': 0.048218170180916786, 'clip_ratio/region_mean': 0.18008152768015862, 'reward_total_mean': 0.05339657515287399, 'reward_meter_mean': 0.2120237946510315, 'reward_meter_std': 0.1717546433210373, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.05345224589109421, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9700000286102295, 'reward_count_floor_std': 0.01603567600250244, 'reward_lexical_plausibility_mean': 0.9818199872970581, 'reward_lexical_plausibility_std': 0.03753276914358139, 'reward_repeat_penalty_mean': 0.988950252532959, 'reward_repeat_penalty_std': 0.009890228509902954, 'reward_repeat_floor_mean': 0.9986698627471924, 'reward_repeat_floor_std': 0.001043530646711588, 'reward_near_duplicate_penalty_mean': 0.9944734573364258, 'reward_near_duplicate_penalty_std': 0.0028089345432817936, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9944319725036621, 'reward_distinct_2_std': 0.0076876552775502205, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.05339657515287399, 'reward_total_composite_std': 0.05815418064594269, 'epoch': 0.0}

  3%|▎         | 92/3300 [17:13<9:47:39, 10.99s/it]INFO 04-12 16:15:40 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 93/3300 [17:22<9:04:38, 10.19s/it]2026-04-12 16:15:48,245 | INFO | train_grpo_train | metrics_logged mode=train step=93

                                                   
{'loss': 0.0822, 'grad_norm': 20.925451278686523, 'learning_rate': 9.721212121212123e-06, 'num_tokens': 181102.0, 'completions/mean_length': 24.125, 'completions/min_length': 18.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7652628421783447, 'rewards/meter/std': 0.4062955975532532, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8676470518112183, 'rewards/lexical_plausibility/std': 0.17168965935707092, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.3660210967063904, 'rewards/repeat_penalty/mean': 0.7455357313156128, 'rewards/repeat_penalty/std': 0.012626901268959045, 'rewards/repeat_floor/mean': 0.9841071367263794, 'rewards/repeat_floor/std': 0.002525375923141837, 'rewards/near_duplicate_penalty/mean': 0.9940476417541504, 'rewards/near_duplicate_penalty/std': 0.016835875809192657, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5161031484603882, 'rewards/total_composite/std': 0.4155314564704895, 'reward': 0.5161031484603882, 'reward_std': 0.4155314266681671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23831449449062347, 'sampling/sampling_logp_difference/max': 1.624668836593628, 'sampling/importance_sampling_ratio/min': 0.196976900100708, 'sampling/importance_sampling_ratio/mean': 1.0388902425765991, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.498924896121025, 'clip_ratio/low_mean': 0.0980448741465807, 'clip_ratio/low_min': 0.0980448741465807, 'clip_ratio/high_mean': 0.1560349315404892, 'clip_ratio/high_max': 0.1560349315404892, 'clip_ratio/region_mean': 0.2540798056870699, 'reward_total_mean': 0.5161031484603882, 'reward_meter_mean': 0.7652628421783447, 'reward_meter_std': 0.4062955975532532, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8676470518112183, 'reward_lexical_plausibility_std': 0.17168965935707092, 'reward_repeat_penalty_mean': 0.7455357313156128, 'reward_repeat_penalty_std': 0.012626901268959045, 'reward_repeat_floor_mean': 0.9841071367263794, 'reward_repeat_floor_std': 0.002525375923141837, 'reward_near_duplicate_penalty_mean': 0.9940476417541504, 'reward_near_duplicate_penalty_std': 0.016835875809192657, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.3660210967063904, 'reward_total_composite_mean': 0.5161031484603882, 'reward_total_composite_std': 0.4155314564704895, 'epoch': 0.0}

  3%|▎         | 93/3300 [17:22<9:04:38, 10.19s/it]INFO 04-12 16:15:48 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 94/3300 [17:31<8:50:18,  9.92s/it]2026-04-12 16:15:57,552 | INFO | train_grpo_train | metrics_logged mode=train step=94

                                                   
{'loss': 0.0026, 'grad_norm': 16.18235969543457, 'learning_rate': 9.718181818181818e-06, 'num_tokens': 182594.0, 'completions/mean_length': 42.5, 'completions/min_length': 34.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.4126415252685547, 'rewards/meter/std': 0.463853657245636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.9979338645935059, 'rewards/repeat_penalty/std': 0.005309219006448984, 'rewards/repeat_floor/mean': 0.9996901154518127, 'rewards/repeat_floor/std': 0.0007963832467794418, 'rewards/near_duplicate_penalty/mean': 0.9979338645935059, 'rewards/near_duplicate_penalty/std': 0.005309219006448984, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.144917830824852, 'rewards/total_composite/std': 0.18419717252254486, 'reward': 0.144917830824852, 'reward_std': 0.18419715762138367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22092963755130768, 'sampling/sampling_logp_difference/max': 1.6048192977905273, 'sampling/importance_sampling_ratio/min': 0.20092585682868958, 'sampling/importance_sampling_ratio/mean': 1.0474412441253662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.75510573387146, 'clip_ratio/low_mean': 0.0901009663939476, 'clip_ratio/low_min': 0.0901009663939476, 'clip_ratio/high_mean': 0.07163957506418228, 'clip_ratio/high_max': 0.07163957506418228, 'clip_ratio/region_mean': 0.16174054145812988, 'reward_total_mean': 0.144917830824852, 'reward_meter_mean': 0.4126415252685547, 'reward_meter_std': 0.463853657245636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9979338645935059, 'reward_repeat_penalty_std': 0.005309219006448984, 'reward_repeat_floor_mean': 0.9996901154518127, 'reward_repeat_floor_std': 0.0007963832467794418, 'reward_near_duplicate_penalty_mean': 0.9979338645935059, 'reward_near_duplicate_penalty_std': 0.005309219006448984, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.144917830824852, 'reward_total_composite_std': 0.18419717252254486, 'epoch': 0.0}

  3%|▎         | 94/3300 [17:31<8:50:18,  9.92s/it]INFO 04-12 16:15:57 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 95/3300 [17:39<8:26:27,  9.48s/it]2026-04-12 16:16:06,000 | INFO | train_grpo_train | metrics_logged mode=train step=95

                                                   
{'loss': 0.1012, 'grad_norm': 16.56658935546875, 'learning_rate': 9.715151515151516e-06, 'num_tokens': 184188.0, 'completions/mean_length': 42.25, 'completions/min_length': 34.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7237671613693237, 'rewards/meter/std': 0.43647003173828125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9895833730697632, 'rewards/lexical_plausibility/std': 0.014376373030245304, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.21836650371551514, 'rewards/repeat_penalty/mean': 0.9969465136528015, 'rewards/repeat_penalty/std': 0.00863658543676138, 'rewards/repeat_floor/mean': 0.9995419979095459, 'rewards/repeat_floor/std': 0.0012954887934029102, 'rewards/near_duplicate_penalty/mean': 0.9969465136528015, 'rewards/near_duplicate_penalty/std': 0.00863658543676138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.36743414402008057, 'rewards/total_composite/std': 0.3021846413612366, 'reward': 0.36743414402008057, 'reward_std': 0.3021846115589142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20720338821411133, 'sampling/sampling_logp_difference/max': 1.6059870719909668, 'sampling/importance_sampling_ratio/min': 0.2006913721561432, 'sampling/importance_sampling_ratio/mean': 1.0434907674789429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.336642026901245, 'clip_ratio/low_mean': 0.06482128519564867, 'clip_ratio/low_min': 0.06482128519564867, 'clip_ratio/high_mean': 0.12250485830008984, 'clip_ratio/high_max': 0.12250485830008984, 'clip_ratio/region_mean': 0.1873261434957385, 'reward_total_mean': 0.36743414402008057, 'reward_meter_mean': 0.7237671613693237, 'reward_meter_std': 0.43647003173828125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9895833730697632, 'reward_lexical_plausibility_std': 0.014376373030245304, 'reward_repeat_penalty_mean': 0.9969465136528015, 'reward_repeat_penalty_std': 0.00863658543676138, 'reward_repeat_floor_mean': 0.9995419979095459, 'reward_repeat_floor_std': 0.0012954887934029102, 'reward_near_duplicate_penalty_mean': 0.9969465136528015, 'reward_near_duplicate_penalty_std': 0.00863658543676138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.21836650371551514, 'reward_total_composite_mean': 0.36743414402008057, 'reward_total_composite_std': 0.3021846413612366, 'epoch': 0.0}

  3%|▎         | 95/3300 [17:39<8:26:27,  9.48s/it]INFO 04-12 16:16:06 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 96/3300 [17:49<8:21:32,  9.39s/it]2026-04-12 16:16:15,182 | INFO | train_grpo_train | metrics_logged mode=train step=96

                                                   
{'loss': 0.0625, 'grad_norm': 12.924627304077148, 'learning_rate': 9.712121212121213e-06, 'num_tokens': 185838.0, 'completions/mean_length': 45.25, 'completions/min_length': 30.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9345506429672241, 'rewards/meter/std': 0.05481686070561409, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9885380268096924, 'rewards/lexical_plausibility/std': 0.032419249415397644, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9998278617858887, 'rewards/repeat_penalty/std': 0.0004869857511948794, 'rewards/repeat_floor/mean': 0.9999741911888123, 'rewards/repeat_floor/std': 7.304049358936027e-05, 'rewards/near_duplicate_penalty/mean': 0.9998278617858887, 'rewards/near_duplicate_penalty/std': 0.0004869857511948794, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38547563552856445, 'rewards/total_composite/std': 0.05341857299208641, 'reward': 0.38547563552856445, 'reward_std': 0.05341857671737671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2202989012002945, 'sampling/sampling_logp_difference/max': 1.2113046646118164, 'sampling/importance_sampling_ratio/min': 0.2978084683418274, 'sampling/importance_sampling_ratio/mean': 1.045016884803772, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.342909559607506, 'clip_ratio/low_mean': 0.08094632439315319, 'clip_ratio/low_min': 0.08094632439315319, 'clip_ratio/high_mean': 0.11059252172708511, 'clip_ratio/high_max': 0.11059252172708511, 'clip_ratio/region_mean': 0.1915388461202383, 'reward_total_mean': 0.38547563552856445, 'reward_meter_mean': 0.9345506429672241, 'reward_meter_std': 0.05481686070561409, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9885380268096924, 'reward_lexical_plausibility_std': 0.032419249415397644, 'reward_repeat_penalty_mean': 0.9998278617858887, 'reward_repeat_penalty_std': 0.0004869857511948794, 'reward_repeat_floor_mean': 0.9999741911888123, 'reward_repeat_floor_std': 7.304049358936027e-05, 'reward_near_duplicate_penalty_mean': 0.9998278617858887, 'reward_near_duplicate_penalty_std': 0.0004869857511948794, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.38547563552856445, 'reward_total_composite_std': 0.05341857299208641, 'epoch': 0.01}

  3%|▎         | 96/3300 [17:49<8:21:32,  9.39s/it]INFO 04-12 16:16:15 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 97/3300 [17:59<8:42:38,  9.79s/it]2026-04-12 16:16:25,902 | INFO | train_grpo_train | metrics_logged mode=train step=97

                                                   
{'loss': 0.2, 'grad_norm': 14.041024208068848, 'learning_rate': 9.70909090909091e-06, 'num_tokens': 187970.0, 'completions/mean_length': 91.5, 'completions/min_length': 14.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.5290300846099854, 'rewards/meter/std': 0.30417460203170776, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.08017838001251221, 'rewards/lexical_plausibility/mean': 0.9743640422821045, 'rewards/lexical_plausibility/std': 0.07250935584306717, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9626697301864624, 'rewards/repeat_penalty/std': 0.0860561728477478, 'rewards/repeat_floor/mean': 0.9972129464149475, 'rewards/repeat_floor/std': 0.00498339906334877, 'rewards/near_duplicate_penalty/mean': 0.9939197301864624, 'rewards/near_duplicate_penalty/std': 0.005241386592388153, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20124690234661102, 'rewards/total_composite/std': 0.11275546252727509, 'reward': 0.20124690234661102, 'reward_std': 0.11275545507669449, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2275000810623169, 'sampling/sampling_logp_difference/max': 1.7260465621948242, 'sampling/importance_sampling_ratio/min': 0.17798668146133423, 'sampling/importance_sampling_ratio/mean': 1.0286104679107666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.283098965883255, 'clip_ratio/low_mean': 0.11474193446338177, 'clip_ratio/low_min': 0.11474193446338177, 'clip_ratio/high_mean': 0.06822820007801056, 'clip_ratio/high_max': 0.06822820007801056, 'clip_ratio/region_mean': 0.18297013454139233, 'reward_total_mean': 0.20124690234661102, 'reward_meter_mean': 0.5290300846099854, 'reward_meter_std': 0.30417460203170776, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.08017838001251221, 'reward_lexical_plausibility_mean': 0.9743640422821045, 'reward_lexical_plausibility_std': 0.07250935584306717, 'reward_repeat_penalty_mean': 0.9626697301864624, 'reward_repeat_penalty_std': 0.0860561728477478, 'reward_repeat_floor_mean': 0.9972129464149475, 'reward_repeat_floor_std': 0.00498339906334877, 'reward_near_duplicate_penalty_mean': 0.9939197301864624, 'reward_near_duplicate_penalty_std': 0.005241386592388153, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.20124690234661102, 'reward_total_composite_std': 0.11275546252727509, 'epoch': 0.01}

  3%|▎         | 97/3300 [17:59<8:42:38,  9.79s/it]INFO 04-12 16:16:26 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 98/3300 [18:09<8:43:34,  9.81s/it]2026-04-12 16:16:35,761 | INFO | train_grpo_train | metrics_logged mode=train step=98

                                                   
{'loss': -0.0143, 'grad_norm': 19.608434677124023, 'learning_rate': 9.706060606060606e-06, 'num_tokens': 189687.0, 'completions/mean_length': 38.625, 'completions/min_length': 28.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.18261590600013733, 'rewards/meter/std': 0.28733447194099426, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9499838352203369, 'rewards/lexical_plausibility/std': 0.10509025305509567, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.2945577800273895, 'rewards/repeat_penalty/mean': 0.9624144434928894, 'rewards/repeat_penalty/std': 0.04290363937616348, 'rewards/repeat_floor/mean': 0.9951818585395813, 'rewards/repeat_floor/std': 0.005092092324048281, 'rewards/near_duplicate_penalty/mean': 0.9761209487915039, 'rewards/near_duplicate_penalty/std': 0.03190719336271286, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.04863385111093521, 'rewards/total_composite/std': 0.04955475777387619, 'reward': 0.04863385111093521, 'reward_std': 0.04955475777387619, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22066283226013184, 'sampling/sampling_logp_difference/max': 1.746760368347168, 'sampling/importance_sampling_ratio/min': 0.17433781921863556, 'sampling/importance_sampling_ratio/mean': 1.032572627067566, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3004300892353058, 'clip_ratio/low_mean': 0.14087301678955555, 'clip_ratio/low_min': 0.14087301678955555, 'clip_ratio/high_mean': 0.10493478737771511, 'clip_ratio/high_max': 0.10493478737771511, 'clip_ratio/region_mean': 0.24580780416727066, 'reward_total_mean': 0.04863385111093521, 'reward_meter_mean': 0.18261590600013733, 'reward_meter_std': 0.28733447194099426, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9499838352203369, 'reward_lexical_plausibility_std': 0.10509025305509567, 'reward_repeat_penalty_mean': 0.9624144434928894, 'reward_repeat_penalty_std': 0.04290363937616348, 'reward_repeat_floor_mean': 0.9951818585395813, 'reward_repeat_floor_std': 0.005092092324048281, 'reward_near_duplicate_penalty_mean': 0.9761209487915039, 'reward_near_duplicate_penalty_std': 0.03190719336271286, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.2945577800273895, 'reward_total_composite_mean': 0.04863385111093521, 'reward_total_composite_std': 0.04955475777387619, 'epoch': 0.01}

  3%|▎         | 98/3300 [18:09<8:43:34,  9.81s/it]INFO 04-12 16:16:36 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 99/3300 [18:19<8:43:53,  9.82s/it]2026-04-12 16:16:45,601 | INFO | train_grpo_train | metrics_logged mode=train step=99

                                                   
{'loss': -0.0025, 'grad_norm': 10.482253074645996, 'learning_rate': 9.703030303030305e-06, 'num_tokens': 191617.0, 'completions/mean_length': 83.25, 'completions/min_length': 18.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.5146980285644531, 'rewards/meter/std': 0.41274702548980713, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.35197150707244873, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.10559145361185074, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9783871173858643, 'rewards/repeat_penalty/std': 0.033628061413764954, 'rewards/repeat_floor/mean': 0.997295081615448, 'rewards/repeat_floor/std': 0.0035862107761204243, 'rewards/near_duplicate_penalty/mean': 0.9877362251281738, 'rewards/near_duplicate_penalty/std': 0.010089682415127754, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.18637636303901672, 'rewards/total_composite/std': 0.14430412650108337, 'reward': 0.18637636303901672, 'reward_std': 0.14430411159992218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20706245303153992, 'sampling/sampling_logp_difference/max': 1.6014900207519531, 'sampling/importance_sampling_ratio/min': 0.25874075293540955, 'sampling/importance_sampling_ratio/mean': 1.048789143562317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.520952507853508, 'clip_ratio/low_mean': 0.07201317511498928, 'clip_ratio/low_min': 0.07201317511498928, 'clip_ratio/high_mean': 0.14008009247481823, 'clip_ratio/high_max': 0.14008009247481823, 'clip_ratio/region_mean': 0.2120932675898075, 'reward_total_mean': 0.18637636303901672, 'reward_meter_mean': 0.5146980285644531, 'reward_meter_std': 0.41274702548980713, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.35197150707244873, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.10559145361185074, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9783871173858643, 'reward_repeat_penalty_std': 0.033628061413764954, 'reward_repeat_floor_mean': 0.997295081615448, 'reward_repeat_floor_std': 0.0035862107761204243, 'reward_near_duplicate_penalty_mean': 0.9877362251281738, 'reward_near_duplicate_penalty_std': 0.010089682415127754, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.18637636303901672, 'reward_total_composite_std': 0.14430412650108337, 'epoch': 0.01}

  3%|▎         | 99/3300 [18:19<8:43:53,  9.82s/it]INFO 04-12 16:16:45 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 100/3300 [18:30<9:03:25, 10.19s/it]2026-04-12 16:16:56,652 | INFO | train_grpo_train | metrics_logged mode=train step=100

                                                    
{'loss': -0.0098, 'grad_norm': 8.801105499267578, 'learning_rate': 9.7e-06, 'num_tokens': 194164.0, 'completions/mean_length': 124.375, 'completions/min_length': 84.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.375, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.6300970315933228, 'rewards/meter/std': 0.3762970566749573, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9907407760620117, 'rewards/lexical_plausibility/std': 0.02618914656341076, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.986890435218811, 'rewards/repeat_penalty/std': 0.019385388121008873, 'rewards/repeat_floor/mean': 0.9985101819038391, 'rewards/repeat_floor/std': 0.0020619009155780077, 'rewards/near_duplicate_penalty/mean': 0.9949702024459839, 'rewards/near_duplicate_penalty/std': 0.006119845900684595, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.991829514503479, 'rewards/distinct_2/std': 0.0152190662920475, 'rewards/total_composite/mean': 0.21604344248771667, 'rewards/total_composite/std': 0.12924325466156006, 'reward': 0.21604344248771667, 'reward_std': 0.12924323976039886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21135327219963074, 'sampling/sampling_logp_difference/max': 1.5679607391357422, 'sampling/importance_sampling_ratio/min': 0.20846986770629883, 'sampling/importance_sampling_ratio/mean': 1.0488430261611938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6173838078975677, 'clip_ratio/low_mean': 0.07352012768387794, 'clip_ratio/low_min': 0.07352012768387794, 'clip_ratio/high_mean': 0.11888312362134457, 'clip_ratio/high_max': 0.11888312362134457, 'clip_ratio/region_mean': 0.1924032513052225, 'reward_total_mean': 0.21604344248771667, 'reward_meter_mean': 0.6300970315933228, 'reward_meter_std': 0.3762970566749573, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9907407760620117, 'reward_lexical_plausibility_std': 0.02618914656341076, 'reward_repeat_penalty_mean': 0.986890435218811, 'reward_repeat_penalty_std': 0.019385388121008873, 'reward_repeat_floor_mean': 0.9985101819038391, 'reward_repeat_floor_std': 0.0020619009155780077, 'reward_near_duplicate_penalty_mean': 0.9949702024459839, 'reward_near_duplicate_penalty_std': 0.006119845900684595, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.991829514503479, 'reward_distinct_2_std': 0.0152190662920475, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.21604344248771667, 'reward_total_composite_std': 0.12924325466156006, 'epoch': 0.01}

  3%|▎         | 100/3300 [18:30<9:03:25, 10.19s/it]INFO 04-12 16:16:56 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:05,  5.91s/it]

 23%|██▎       | 3/13 [00:19<01:08,  6.80s/it]

 31%|███       | 4/13 [00:29<01:09,  7.75s/it]

 38%|███▊      | 5/13 [00:40<01:12,  9.08s/it]

 46%|████▌     | 6/13 [00:51<01:07,  9.67s/it]

 54%|█████▍    | 7/13 [01:02<00:59,  9.90s/it]

 62%|██████▏   | 8/13 [01:13<00:51, 10.35s/it]

 69%|██████▉   | 9/13 [01:26<00:45, 11.31s/it]

 77%|███████▋  | 10/13 [01:36<00:32, 10.84s/it]

 85%|████████▍ | 11/13 [01:51<00:24, 12.16s/it]

 92%|█████████▏| 12/13 [02:09<00:13, 13.78s/it]

100%|██████████| 13/13 [02:21<00:00, 13.37s/it]2026-04-12 16:19:30,443 | INFO | train_grpo_train | metrics_logged mode=eval step=100

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 153.78, 'eval_samples_per_second': 0.676, 'eval_steps_per_second': 0.085, 'eval_num_tokens': 194164.0, 'eval_completions/mean_length': 193.84615384615384, 'eval_completions/min_length': 42.92307692307692, 'eval_completions/max_length': 423.3076923076923, 'eval_completions/clipped_ratio': 0.0673076923076923, 'eval_completions/mean_terminated_length': 171.4107149564303, 'eval_completions/min_terminated_length': 42.92307692307692, 'eval_completions/max_terminated_length': 364.61538461538464, 'eval_rewards/meter/mean': 0.32941878071198094, 'eval_rewards/meter/std': 0.36498406300177944, 'eval_rewards/count_adherence/mean': 0.9112801781067481, 'eval_rewards/count_adherence/std': 0.12825587879006678, 'eval_rewards/arabic_clean/mean': 0.9038461538461539, 'eval_rewards/arabic_clean/std': 0.21140201504413897, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.973384059392489, 'eval_rewards/count_floor/std': 0.03847676479759125, 'eval_rewards/lexical_plausibility/mean': 0.9768699315878061, 'eval_rewards/lexical_plausibility/std': 0.05473032858795845, 'eval_rewards/judge_quality/mean': 0.32932692078443676, 'eval_rewards/judge_quality/std': 0.16876431153370783, 'eval_rewards/repeat_penalty/mean': 0.9646368301831759, 'eval_rewards/repeat_penalty/std': 0.05871333593789201, 'eval_rewards/repeat_floor/mean': 0.9964348903069129, 'eval_rewards/repeat_floor/std': 0.005291556106665387, 'eval_rewards/near_duplicate_penalty/mean': 0.9886650855724628, 'eval_rewards/near_duplicate_penalty/std': 0.014827047564232579, 'eval_rewards/opening_diversity/mean': 0.9879807692307693, 'eval_rewards/opening_diversity/std': 0.03399551774446781, 'eval_rewards/distinct_2/mean': 0.9872922255442693, 'eval_rewards/distinct_2/std': 0.025999667015499793, 'eval_rewards/total_composite/mean': 0.11386545432301667, 'eval_rewards/total_composite/std': 0.1390101150251352, 'eval_reward': 0.11386545432301667, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15946110624533433, 'eval_sampling/sampling_logp_difference/max': 1.2660664778489332, 'eval_sampling/importance_sampling_ratio/min': 0.2837856881893598, 'eval_sampling/importance_sampling_ratio/mean': 1.0492045054068933, 'eval_sampling/importance_sampling_ratio/max': 1.6505950414217436, 'eval_entropy': 2.7171743649702806, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.11386545432301667, 'eval_reward_meter_mean': 0.32941878071198094, 'eval_reward_meter_std': 0.36498406300177944, 'eval_reward_count_adherence_mean': 0.9112801781067481, 'eval_reward_count_adherence_std': 0.12825587879006678, 'eval_reward_arabic_clean_mean': 0.9038461538461539, 'eval_reward_arabic_clean_std': 0.21140201504413897, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.973384059392489, 'eval_reward_count_floor_std': 0.03847676479759125, 'eval_reward_lexical_plausibility_mean': 0.9768699315878061, 'eval_reward_lexical_plausibility_std': 0.05473032858795845, 'eval_reward_repeat_penalty_mean': 0.9646368301831759, 'eval_reward_repeat_penalty_std': 0.05871333593789201, 'eval_reward_repeat_floor_mean': 0.9964348903069129, 'eval_reward_repeat_floor_std': 0.005291556106665387, 'eval_reward_near_duplicate_penalty_mean': 0.9886650855724628, 'eval_reward_near_duplicate_penalty_std': 0.014827047564232579, 'eval_reward_opening_diversity_mean': 0.9879807692307693, 'eval_reward_opening_diversity_std': 0.03399551774446781, 'eval_reward_distinct_2_mean': 0.9872922255442693, 'eval_reward_distinct_2_std': 0.025999667015499793, 'eval_reward_judge_quality_mean': 0.32932692078443676, 'eval_reward_judge_quality_std': 0.16876431153370783, 'eval_reward_total_composite_mean': 0.11386545432301667, 'eval_reward_total_composite_std': 0.1390101150251352, 'epoch': 0.01}

  3%|▎         | 100/3300 [21:04<9:03:25, 10.19s/it]

100%|██████████| 13/13 [02:21<00:00, 13.37s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  3%|▎         | 101/3300 [21:15<50:19:27, 56.63s/it]2026-04-12 16:19:41,653 | INFO | train_grpo_train | metrics_logged mode=train step=101

                                                     
{'loss': 0.1291, 'grad_norm': 24.660491943359375, 'learning_rate': 9.696969696969698e-06, 'num_tokens': 195639.0, 'completions/mean_length': 20.375, 'completions/min_length': 14.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.375, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.48119914531707764, 'rewards/meter/std': 0.508114218711853, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.350265234708786, 'rewards/repeat_penalty/mean': 0.7407185435295105, 'rewards/repeat_penalty/std': 0.01720251515507698, 'rewards/repeat_floor/mean': 0.98314368724823, 'rewards/repeat_floor/std': 0.003440506523475051, 'rewards/near_duplicate_penalty/mean': 0.9876246452331543, 'rewards/near_duplicate_penalty/std': 0.022936683148145676, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3458004295825958, 'rewards/total_composite/std': 0.43915578722953796, 'reward': 0.3458004295825958, 'reward_std': 0.43915584683418274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2215399593114853, 'sampling/sampling_logp_difference/max': 2.1100997924804688, 'sampling/importance_sampling_ratio/min': 0.12122586369514465, 'sampling/importance_sampling_ratio/mean': 0.9900545477867126, 'sampling/importance_sampling_ratio/max': 1.6320496797561646, 'entropy': 1.846018135547638, 'clip_ratio/low_mean': 0.06352419592440128, 'clip_ratio/low_min': 0.06352419592440128, 'clip_ratio/high_mean': 0.1071428582072258, 'clip_ratio/high_max': 0.1071428582072258, 'clip_ratio/region_mean': 0.17066705413162708, 'reward_total_mean': 0.3458004295825958, 'reward_meter_mean': 0.48119914531707764, 'reward_meter_std': 0.508114218711853, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7407185435295105, 'reward_repeat_penalty_std': 0.01720251515507698, 'reward_repeat_floor_mean': 0.98314368724823, 'reward_repeat_floor_std': 0.003440506523475051, 'reward_near_duplicate_penalty_mean': 0.9876246452331543, 'reward_near_duplicate_penalty_std': 0.022936683148145676, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.350265234708786, 'reward_total_composite_mean': 0.3458004295825958, 'reward_total_composite_std': 0.43915578722953796, 'epoch': 0.01}

  3%|▎         | 101/3300 [21:15<50:19:27, 56.63s/it]INFO 04-12 16:19:41 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 102/3300 [21:25<37:47:19, 42.54s/it]2026-04-12 16:19:51,397 | INFO | train_grpo_train | metrics_logged mode=train step=102

                                                     
{'loss': 0.1026, 'grad_norm': 13.553353309631348, 'learning_rate': 9.693939393939395e-06, 'num_tokens': 197312.0, 'completions/mean_length': 52.125, 'completions/min_length': 36.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.838103175163269, 'rewards/meter/std': 0.3429650664329529, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.9850345849990845, 'rewards/repeat_penalty/std': 0.038038067519664764, 'rewards/repeat_floor/mean': 0.9982517957687378, 'rewards/repeat_floor/std': 0.004305039066821337, 'rewards/near_duplicate_penalty/mean': 0.9938398599624634, 'rewards/near_duplicate_penalty/std': 0.01328597404062748, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.37283021211624146, 'rewards/total_composite/std': 0.2614906132221222, 'reward': 0.37283021211624146, 'reward_std': 0.2614906430244446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22789521515369415, 'sampling/sampling_logp_difference/max': 1.8274550437927246, 'sampling/importance_sampling_ratio/min': 0.160822331905365, 'sampling/importance_sampling_ratio/mean': 1.0536524057388306, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5214386582374573, 'clip_ratio/low_mean': 0.10784188099205494, 'clip_ratio/low_min': 0.10784188099205494, 'clip_ratio/high_mean': 0.07347913924604654, 'clip_ratio/high_max': 0.07347913924604654, 'clip_ratio/region_mean': 0.18132102023810148, 'reward_total_mean': 0.37283021211624146, 'reward_meter_mean': 0.838103175163269, 'reward_meter_std': 0.3429650664329529, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9850345849990845, 'reward_repeat_penalty_std': 0.038038067519664764, 'reward_repeat_floor_mean': 0.9982517957687378, 'reward_repeat_floor_std': 0.004305039066821337, 'reward_near_duplicate_penalty_mean': 0.9938398599624634, 'reward_near_duplicate_penalty_std': 0.01328597404062748, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.37283021211624146, 'reward_total_composite_std': 0.2614906132221222, 'epoch': 0.01}

  3%|▎         | 102/3300 [21:25<37:47:19, 42.54s/it]INFO 04-12 16:19:51 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 103/3300 [21:33<28:45:37, 32.39s/it]2026-04-12 16:20:00,004 | INFO | train_grpo_train | metrics_logged mode=train step=103

                                                     
{'loss': 0.0504, 'grad_norm': 25.127071380615234, 'learning_rate': 9.690909090909092e-06, 'num_tokens': 198839.0, 'completions/mean_length': 43.875, 'completions/min_length': 41.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.5449619293212891, 'rewards/meter/std': 0.30401355028152466, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8400000333786011, 'rewards/judge_quality/std': 0.16843822598457336, 'rewards/repeat_penalty/mean': 0.9917404651641846, 'rewards/repeat_penalty/std': 0.01016718428581953, 'rewards/repeat_floor/mean': 0.9987610578536987, 'rewards/repeat_floor/std': 0.0015250721480697393, 'rewards/near_duplicate_penalty/mean': 0.9917404651641846, 'rewards/near_duplicate_penalty/std': 0.01016718428581953, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.47210636734962463, 'rewards/total_composite/std': 0.28807532787323, 'reward': 0.47210636734962463, 'reward_std': 0.2880752980709076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10407081246376038, 'sampling/sampling_logp_difference/max': 1.7363204956054688, 'sampling/importance_sampling_ratio/min': 0.17616741359233856, 'sampling/importance_sampling_ratio/mean': 1.0125762224197388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5438592098653316, 'clip_ratio/low_mean': 0.05730749759823084, 'clip_ratio/low_min': 0.05730749759823084, 'clip_ratio/high_mean': 0.048939047614112496, 'clip_ratio/high_max': 0.048939047614112496, 'clip_ratio/region_mean': 0.10624654521234334, 'reward_total_mean': 0.47210636734962463, 'reward_meter_mean': 0.5449619293212891, 'reward_meter_std': 0.30401355028152466, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9917404651641846, 'reward_repeat_penalty_std': 0.01016718428581953, 'reward_repeat_floor_mean': 0.9987610578536987, 'reward_repeat_floor_std': 0.0015250721480697393, 'reward_near_duplicate_penalty_mean': 0.9917404651641846, 'reward_near_duplicate_penalty_std': 0.01016718428581953, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8400000333786011, 'reward_judge_quality_std': 0.16843822598457336, 'reward_total_composite_mean': 0.47210636734962463, 'reward_total_composite_std': 0.28807532787323, 'epoch': 0.01}

  3%|▎         | 103/3300 [21:33<28:45:37, 32.39s/it]INFO 04-12 16:20:00 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 104/3300 [21:41<22:15:05, 25.06s/it]2026-04-12 16:20:07,984 | INFO | train_grpo_train | metrics_logged mode=train step=104

                                                     
{'loss': 0.0611, 'grad_norm': 21.231082916259766, 'learning_rate': 9.687878787878788e-06, 'num_tokens': 200299.0, 'completions/mean_length': 25.5, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9572399854660034, 'rewards/meter/std': 0.051584795117378235, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9752674102783203, 'rewards/lexical_plausibility/std': 0.06995441019535065, 'rewards/judge_quality/mean': 0.5724999904632568, 'rewards/judge_quality/std': 0.38425251841545105, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5437753200531006, 'rewards/total_composite/std': 0.36814752221107483, 'reward': 0.5437753200531006, 'reward_std': 0.36814752221107483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23689311742782593, 'sampling/sampling_logp_difference/max': 1.550304889678955, 'sampling/importance_sampling_ratio/min': 0.21218328177928925, 'sampling/importance_sampling_ratio/mean': 1.055704951286316, 'sampling/importance_sampling_ratio/max': 1.796012043952942, 'entropy': 2.577810436487198, 'clip_ratio/low_mean': 0.08415292482823133, 'clip_ratio/low_min': 0.08415292482823133, 'clip_ratio/high_mean': 0.1052370686084032, 'clip_ratio/high_max': 0.1052370686084032, 'clip_ratio/region_mean': 0.18938999343663454, 'reward_total_mean': 0.5437753200531006, 'reward_meter_mean': 0.9572399854660034, 'reward_meter_std': 0.051584795117378235, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9752674102783203, 'reward_lexical_plausibility_std': 0.06995441019535065, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5724999904632568, 'reward_judge_quality_std': 0.38425251841545105, 'reward_total_composite_mean': 0.5437753200531006, 'reward_total_composite_std': 0.36814752221107483, 'epoch': 0.01}

  3%|▎         | 104/3300 [21:41<22:15:05, 25.06s/it]INFO 04-12 16:20:08 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 105/3300 [21:50<17:51:55, 20.13s/it]2026-04-12 16:20:16,601 | INFO | train_grpo_train | metrics_logged mode=train step=105

                                                     
{'loss': -0.0739, 'grad_norm': 14.191681861877441, 'learning_rate': 9.684848484848487e-06, 'num_tokens': 202184.0, 'completions/mean_length': 50.625, 'completions/min_length': 41.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5926839113235474, 'rewards/meter/std': 0.3989444971084595, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8915603160858154, 'rewards/lexical_plausibility/std': 0.11506184935569763, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9921516180038452, 'rewards/repeat_penalty/std': 0.01117697823792696, 'rewards/repeat_floor/mean': 0.9988227486610413, 'rewards/repeat_floor/std': 0.0016765488544479012, 'rewards/near_duplicate_penalty/mean': 0.9921516180038452, 'rewards/near_duplicate_penalty/std': 0.01117697823792696, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1712530255317688, 'rewards/total_composite/std': 0.11353282630443573, 'reward': 0.1712530255317688, 'reward_std': 0.11353281885385513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25293460488319397, 'sampling/sampling_logp_difference/max': 1.8832950592041016, 'sampling/importance_sampling_ratio/min': 0.15208813548088074, 'sampling/importance_sampling_ratio/mean': 1.0083962678909302, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.698224186897278, 'clip_ratio/low_mean': 0.12342285364866257, 'clip_ratio/low_min': 0.12342285364866257, 'clip_ratio/high_mean': 0.10810508020222187, 'clip_ratio/high_max': 0.10810508020222187, 'clip_ratio/region_mean': 0.23152793385088444, 'reward_total_mean': 0.1712530255317688, 'reward_meter_mean': 0.5926839113235474, 'reward_meter_std': 0.3989444971084595, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8915603160858154, 'reward_lexical_plausibility_std': 0.11506184935569763, 'reward_repeat_penalty_mean': 0.9921516180038452, 'reward_repeat_penalty_std': 0.01117697823792696, 'reward_repeat_floor_mean': 0.9988227486610413, 'reward_repeat_floor_std': 0.0016765488544479012, 'reward_near_duplicate_penalty_mean': 0.9921516180038452, 'reward_near_duplicate_penalty_std': 0.01117697823792696, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.1712530255317688, 'reward_total_composite_std': 0.11353282630443573, 'epoch': 0.01}

  3%|▎         | 105/3300 [21:50<17:51:55, 20.13s/it]INFO 04-12 16:20:16 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 106/3300 [21:58<14:35:35, 16.45s/it]2026-04-12 16:20:24,456 | INFO | train_grpo_train | metrics_logged mode=train step=106

                                                     
{'loss': 0.1553, 'grad_norm': 27.88457489013672, 'learning_rate': 9.681818181818182e-06, 'num_tokens': 203817.0, 'completions/mean_length': 42.125, 'completions/min_length': 35.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6516113877296448, 'rewards/meter/std': 0.44471949338912964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8050000071525574, 'rewards/judge_quality/std': 0.2517368495464325, 'rewards/repeat_penalty/mean': 0.9780062437057495, 'rewards/repeat_penalty/std': 0.02115113101899624, 'rewards/repeat_floor/mean': 0.9967009425163269, 'rewards/repeat_floor/std': 0.003172681899741292, 'rewards/near_duplicate_penalty/mean': 0.9780062437057495, 'rewards/near_duplicate_penalty/std': 0.02115113101899624, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5909601449966431, 'rewards/total_composite/std': 0.44367197155952454, 'reward': 0.5909601449966431, 'reward_std': 0.44367194175720215, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20126840472221375, 'sampling/sampling_logp_difference/max': 2.3270578384399414, 'sampling/importance_sampling_ratio/min': 0.09758242964744568, 'sampling/importance_sampling_ratio/mean': 0.9800254702568054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9679724723100662, 'clip_ratio/low_mean': 0.06499474868178368, 'clip_ratio/low_min': 0.06499474868178368, 'clip_ratio/high_mean': 0.11691626626998186, 'clip_ratio/high_max': 0.11691626626998186, 'clip_ratio/region_mean': 0.18191101495176554, 'reward_total_mean': 0.5909601449966431, 'reward_meter_mean': 0.6516113877296448, 'reward_meter_std': 0.44471949338912964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9780062437057495, 'reward_repeat_penalty_std': 0.02115113101899624, 'reward_repeat_floor_mean': 0.9967009425163269, 'reward_repeat_floor_std': 0.003172681899741292, 'reward_near_duplicate_penalty_mean': 0.9780062437057495, 'reward_near_duplicate_penalty_std': 0.02115113101899624, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8050000071525574, 'reward_judge_quality_std': 0.2517368495464325, 'reward_total_composite_mean': 0.5909601449966431, 'reward_total_composite_std': 0.44367197155952454, 'epoch': 0.01}

  3%|▎         | 106/3300 [21:58<14:35:35, 16.45s/it]INFO 04-12 16:20:24 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 107/3300 [22:10<13:28:20, 15.19s/it]2026-04-12 16:20:36,711 | INFO | train_grpo_train | metrics_logged mode=train step=107

                                                     
{'loss': 0.0966, 'grad_norm': 4.674691677093506, 'learning_rate': 9.67878787878788e-06, 'num_tokens': 207778.0, 'completions/mean_length': 307.125, 'completions/min_length': 258.0, 'completions/max_length': 390.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 307.125, 'completions/min_terminated_length': 258.0, 'completions/max_terminated_length': 390.0, 'rewards/meter/mean': 0.6583197116851807, 'rewards/meter/std': 0.3852737545967102, 'rewards/count_adherence/mean': 0.9305555820465088, 'rewards/count_adherence/std': 0.08266931027173996, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9791666865348816, 'rewards/count_floor/std': 0.024800799787044525, 'rewards/lexical_plausibility/mean': 0.9650596380233765, 'rewards/lexical_plausibility/std': 0.08201706409454346, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9906517267227173, 'rewards/repeat_penalty/std': 0.0076460340060293674, 'rewards/repeat_floor/mean': 0.9989129900932312, 'rewards/repeat_floor/std': 0.0007341052987612784, 'rewards/near_duplicate_penalty/mean': 0.9959393739700317, 'rewards/near_duplicate_penalty/std': 0.0019260674016550183, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9946901798248291, 'rewards/distinct_2/std': 0.007329006679356098, 'rewards/total_composite/mean': 0.13285447657108307, 'rewards/total_composite/std': 0.12690594792366028, 'reward': 0.13285447657108307, 'reward_std': 0.12690596282482147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20803570747375488, 'sampling/sampling_logp_difference/max': 1.7416906356811523, 'sampling/importance_sampling_ratio/min': 0.1752239167690277, 'sampling/importance_sampling_ratio/mean': 1.0514438152313232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.267181158065796, 'clip_ratio/low_mean': 0.09589782729744911, 'clip_ratio/low_min': 0.09589782729744911, 'clip_ratio/high_mean': 0.07560446858406067, 'clip_ratio/high_max': 0.07560446858406067, 'clip_ratio/region_mean': 0.17150229588150978, 'reward_total_mean': 0.13285447657108307, 'reward_meter_mean': 0.6583197116851807, 'reward_meter_std': 0.3852737545967102, 'reward_count_adherence_mean': 0.9305555820465088, 'reward_count_adherence_std': 0.08266931027173996, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9791666865348816, 'reward_count_floor_std': 0.024800799787044525, 'reward_lexical_plausibility_mean': 0.9650596380233765, 'reward_lexical_plausibility_std': 0.08201706409454346, 'reward_repeat_penalty_mean': 0.9906517267227173, 'reward_repeat_penalty_std': 0.0076460340060293674, 'reward_repeat_floor_mean': 0.9989129900932312, 'reward_repeat_floor_std': 0.0007341052987612784, 'reward_near_duplicate_penalty_mean': 0.9959393739700317, 'reward_near_duplicate_penalty_std': 0.0019260674016550183, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9946901798248291, 'reward_distinct_2_std': 0.007329006679356098, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.13285447657108307, 'reward_total_composite_std': 0.12690594792366028, 'epoch': 0.01}

  3%|▎         | 107/3300 [22:10<13:28:20, 15.19s/it]INFO 04-12 16:20:36 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 108/3300 [22:19<11:50:40, 13.36s/it]2026-04-12 16:20:45,809 | INFO | train_grpo_train | metrics_logged mode=train step=108

                                                     
{'loss': -0.0681, 'grad_norm': 18.262678146362305, 'learning_rate': 9.675757575757577e-06, 'num_tokens': 209438.0, 'completions/mean_length': 49.5, 'completions/min_length': 33.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4589425325393677, 'rewards/meter/std': 0.31086745858192444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.2707628309726715, 'rewards/repeat_penalty/mean': 0.9637178182601929, 'rewards/repeat_penalty/std': 0.04730316251516342, 'rewards/repeat_floor/mean': 0.9956634044647217, 'rewards/repeat_floor/std': 0.0054037608206272125, 'rewards/near_duplicate_penalty/mean': 0.9829314947128296, 'rewards/near_duplicate_penalty/std': 0.023687105625867844, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9802631139755249, 'rewards/distinct_2/std': 0.036561544984579086, 'rewards/total_composite/mean': 0.3096131682395935, 'rewards/total_composite/std': 0.2605850100517273, 'reward': 0.3096131682395935, 'reward_std': 0.2605850100517273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20179329812526703, 'sampling/sampling_logp_difference/max': 1.2871370315551758, 'sampling/importance_sampling_ratio/min': 0.27606001496315, 'sampling/importance_sampling_ratio/mean': 1.0199841260910034, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8905853629112244, 'clip_ratio/low_mean': 0.10233294777572155, 'clip_ratio/low_min': 0.10233294777572155, 'clip_ratio/high_mean': 0.06369313038885593, 'clip_ratio/high_max': 0.06369313038885593, 'clip_ratio/region_mean': 0.16602607816457748, 'reward_total_mean': 0.3096131682395935, 'reward_meter_mean': 0.4589425325393677, 'reward_meter_std': 0.31086745858192444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9637178182601929, 'reward_repeat_penalty_std': 0.04730316251516342, 'reward_repeat_floor_mean': 0.9956634044647217, 'reward_repeat_floor_std': 0.0054037608206272125, 'reward_near_duplicate_penalty_mean': 0.9829314947128296, 'reward_near_duplicate_penalty_std': 0.023687105625867844, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9802631139755249, 'reward_distinct_2_std': 0.036561544984579086, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.2707628309726715, 'reward_total_composite_mean': 0.3096131682395935, 'reward_total_composite_std': 0.2605850100517273, 'epoch': 0.01}

  3%|▎         | 108/3300 [22:19<11:50:40, 13.36s/it]INFO 04-12 16:20:46 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 109/3300 [22:28<10:46:35, 12.16s/it]2026-04-12 16:20:55,154 | INFO | train_grpo_train | metrics_logged mode=train step=109

                                                     
{'loss': 0.019, 'grad_norm': 16.318899154663086, 'learning_rate': 9.672727272727274e-06, 'num_tokens': 211082.0, 'completions/mean_length': 52.5, 'completions/min_length': 35.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.39329585433006287, 'rewards/meter/std': 0.3182377219200134, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9126887321472168, 'rewards/lexical_plausibility/std': 0.1678655445575714, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.13562028110027313, 'rewards/repeat_penalty/mean': 0.9663114547729492, 'rewards/repeat_penalty/std': 0.0535384863615036, 'rewards/repeat_floor/mean': 0.9959038496017456, 'rewards/repeat_floor/std': 0.006516526453197002, 'rewards/near_duplicate_penalty/mean': 0.9833259582519531, 'rewards/near_duplicate_penalty/std': 0.029908593744039536, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.982277512550354, 'rewards/distinct_2/std': 0.03444872051477432, 'rewards/total_composite/mean': 0.17002442479133606, 'rewards/total_composite/std': 0.13673201203346252, 'reward': 0.17002442479133606, 'reward_std': 0.13673201203346252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22045652568340302, 'sampling/sampling_logp_difference/max': 2.1182918548583984, 'sampling/importance_sampling_ratio/min': 0.120236836373806, 'sampling/importance_sampling_ratio/mean': 1.0091578960418701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5120889097452164, 'clip_ratio/low_mean': 0.10537504777312279, 'clip_ratio/low_min': 0.10537504777312279, 'clip_ratio/high_mean': 0.07010242925025523, 'clip_ratio/high_max': 0.07010242925025523, 'clip_ratio/region_mean': 0.17547747702337801, 'reward_total_mean': 0.17002442479133606, 'reward_meter_mean': 0.39329585433006287, 'reward_meter_std': 0.3182377219200134, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9126887321472168, 'reward_lexical_plausibility_std': 0.1678655445575714, 'reward_repeat_penalty_mean': 0.9663114547729492, 'reward_repeat_penalty_std': 0.0535384863615036, 'reward_repeat_floor_mean': 0.9959038496017456, 'reward_repeat_floor_std': 0.006516526453197002, 'reward_near_duplicate_penalty_mean': 0.9833259582519531, 'reward_near_duplicate_penalty_std': 0.029908593744039536, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.982277512550354, 'reward_distinct_2_std': 0.03444872051477432, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.13562028110027313, 'reward_total_composite_mean': 0.17002442479133606, 'reward_total_composite_std': 0.13673201203346252, 'epoch': 0.01}

  3%|▎         | 109/3300 [22:29<10:46:35, 12.16s/it]INFO 04-12 16:20:55 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 110/3300 [22:37<9:41:39, 10.94s/it] 2026-04-12 16:21:03,253 | INFO | train_grpo_train | metrics_logged mode=train step=110

                                                    
{'loss': 0.0067, 'grad_norm': 25.61552619934082, 'learning_rate': 9.66969696969697e-06, 'num_tokens': 212636.0, 'completions/mean_length': 34.25, 'completions/min_length': 27.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8282620310783386, 'rewards/meter/std': 0.2704026401042938, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9994343519210815, 'rewards/lexical_plausibility/std': 0.0015997890150174499, 'rewards/judge_quality/mean': 0.8799999952316284, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/repeat_penalty/mean': 0.9926075339317322, 'rewards/repeat_penalty/std': 0.02090907283127308, 'rewards/repeat_floor/mean': 0.9988911151885986, 'rewards/repeat_floor/std': 0.003136357991024852, 'rewards/near_duplicate_penalty/mean': 0.9926075339317322, 'rewards/near_duplicate_penalty/std': 0.02090907283127308, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7320461273193359, 'rewards/total_composite/std': 0.2981451749801636, 'reward': 0.7320461273193359, 'reward_std': 0.2981451749801636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1422923356294632, 'sampling/sampling_logp_difference/max': 2.8497915267944336, 'sampling/importance_sampling_ratio/min': 0.05785638466477394, 'sampling/importance_sampling_ratio/mean': 0.9579386115074158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.566255196928978, 'clip_ratio/low_mean': 0.02813390363007784, 'clip_ratio/low_min': 0.02813390363007784, 'clip_ratio/high_mean': 0.1177370531950146, 'clip_ratio/high_max': 0.1177370531950146, 'clip_ratio/region_mean': 0.14587095682509243, 'reward_total_mean': 0.7320461273193359, 'reward_meter_mean': 0.8282620310783386, 'reward_meter_std': 0.2704026401042938, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9994343519210815, 'reward_lexical_plausibility_std': 0.0015997890150174499, 'reward_repeat_penalty_mean': 0.9926075339317322, 'reward_repeat_penalty_std': 0.02090907283127308, 'reward_repeat_floor_mean': 0.9988911151885986, 'reward_repeat_floor_std': 0.003136357991024852, 'reward_near_duplicate_penalty_mean': 0.9926075339317322, 'reward_near_duplicate_penalty_std': 0.02090907283127308, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8799999952316284, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.7320461273193359, 'reward_total_composite_std': 0.2981451749801636, 'epoch': 0.01}

  3%|▎         | 110/3300 [22:37<9:41:39, 10.94s/it]INFO 04-12 16:21:03 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 111/3300 [22:45<9:02:14, 10.20s/it]2026-04-12 16:21:11,741 | INFO | train_grpo_train | metrics_logged mode=train step=111

                                                    
{'loss': 0.1901, 'grad_norm': 20.622541427612305, 'learning_rate': 9.666666666666667e-06, 'num_tokens': 214072.0, 'completions/mean_length': 25.5, 'completions/min_length': 16.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7410401105880737, 'rewards/meter/std': 0.4375188648700714, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9715073704719543, 'rewards/lexical_plausibility/std': 0.08058936893939972, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.36246922612190247, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45010489225387573, 'rewards/total_composite/std': 0.3942350745201111, 'reward': 0.45010489225387573, 'reward_std': 0.3942350447177887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20283541083335876, 'sampling/sampling_logp_difference/max': 1.5572848320007324, 'sampling/importance_sampling_ratio/min': 0.2107073962688446, 'sampling/importance_sampling_ratio/mean': 1.0446093082427979, 'sampling/importance_sampling_ratio/max': 1.9071470499038696, 'entropy': 2.180199012160301, 'clip_ratio/low_mean': 0.13912497088313103, 'clip_ratio/low_min': 0.13912497088313103, 'clip_ratio/high_mean': 0.08583960123360157, 'clip_ratio/high_max': 0.08583960123360157, 'clip_ratio/region_mean': 0.2249645721167326, 'reward_total_mean': 0.45010489225387573, 'reward_meter_mean': 0.7410401105880737, 'reward_meter_std': 0.4375188648700714, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9715073704719543, 'reward_lexical_plausibility_std': 0.08058936893939972, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.36246922612190247, 'reward_total_composite_mean': 0.45010489225387573, 'reward_total_composite_std': 0.3942350745201111, 'epoch': 0.01}

  3%|▎         | 111/3300 [22:45<9:02:14, 10.20s/it]INFO 04-12 16:21:12 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 112/3300 [22:54<8:45:52,  9.90s/it]2026-04-12 16:21:20,926 | INFO | train_grpo_train | metrics_logged mode=train step=112

                                                    
{'loss': -0.0453, 'grad_norm': 19.2539005279541, 'learning_rate': 9.663636363636364e-06, 'num_tokens': 215478.0, 'completions/mean_length': 28.75, 'completions/min_length': 16.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5128393173217773, 'rewards/meter/std': 0.4806661009788513, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.7517600655555725, 'rewards/lexical_plausibility/std': 0.3265290856361389, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.3340979516506195, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2956763505935669, 'rewards/total_composite/std': 0.3917238712310791, 'reward': 0.2956763505935669, 'reward_std': 0.3917238712310791, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22882838547229767, 'sampling/sampling_logp_difference/max': 1.7279586791992188, 'sampling/importance_sampling_ratio/min': 0.1776466816663742, 'sampling/importance_sampling_ratio/mean': 1.031433343887329, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9021585285663605, 'clip_ratio/low_mean': 0.09303139289841056, 'clip_ratio/low_min': 0.09303139289841056, 'clip_ratio/high_mean': 0.07412941940128803, 'clip_ratio/high_max': 0.07412941940128803, 'clip_ratio/region_mean': 0.1671608122996986, 'reward_total_mean': 0.2956763505935669, 'reward_meter_mean': 0.5128393173217773, 'reward_meter_std': 0.4806661009788513, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.7517600655555725, 'reward_lexical_plausibility_std': 0.3265290856361389, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.3340979516506195, 'reward_total_composite_mean': 0.2956763505935669, 'reward_total_composite_std': 0.3917238712310791, 'epoch': 0.01}

  3%|▎         | 112/3300 [22:54<8:45:52,  9.90s/it]INFO 04-12 16:21:21 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 113/3300 [23:03<8:29:21,  9.59s/it]2026-04-12 16:21:29,796 | INFO | train_grpo_train | metrics_logged mode=train step=113

                                                    
{'loss': 0.0237, 'grad_norm': 16.712873458862305, 'learning_rate': 9.660606060606061e-06, 'num_tokens': 217193.0, 'completions/mean_length': 43.375, 'completions/min_length': 31.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.3598339557647705, 'rewards/meter/std': 0.2768198847770691, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9061805009841919, 'rewards/lexical_plausibility/std': 0.1414359211921692, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9927659630775452, 'rewards/repeat_penalty/std': 0.018347030505537987, 'rewards/repeat_floor/mean': 0.998914897441864, 'rewards/repeat_floor/std': 0.0027520437724888325, 'rewards/near_duplicate_penalty/mean': 0.9927659630775452, 'rewards/near_duplicate_penalty/std': 0.018347030505537987, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12368791550397873, 'rewards/total_composite/std': 0.1086086556315422, 'reward': 0.12368791550397873, 'reward_std': 0.1086086556315422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2594793140888214, 'sampling/sampling_logp_difference/max': 1.387596607208252, 'sampling/importance_sampling_ratio/min': 0.24967466294765472, 'sampling/importance_sampling_ratio/mean': 1.034879207611084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.802516222000122, 'clip_ratio/low_mean': 0.14920755475759506, 'clip_ratio/low_min': 0.14920755475759506, 'clip_ratio/high_mean': 0.09501327015459538, 'clip_ratio/high_max': 0.09501327015459538, 'clip_ratio/region_mean': 0.24422082491219044, 'reward_total_mean': 0.12368791550397873, 'reward_meter_mean': 0.3598339557647705, 'reward_meter_std': 0.2768198847770691, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9061805009841919, 'reward_lexical_plausibility_std': 0.1414359211921692, 'reward_repeat_penalty_mean': 0.9927659630775452, 'reward_repeat_penalty_std': 0.018347030505537987, 'reward_repeat_floor_mean': 0.998914897441864, 'reward_repeat_floor_std': 0.0027520437724888325, 'reward_near_duplicate_penalty_mean': 0.9927659630775452, 'reward_near_duplicate_penalty_std': 0.018347030505537987, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.12368791550397873, 'reward_total_composite_std': 0.1086086556315422, 'epoch': 0.01}

  3%|▎         | 113/3300 [23:03<8:29:21,  9.59s/it]INFO 04-12 16:21:30 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 114/3300 [23:13<8:30:01,  9.61s/it]2026-04-12 16:21:39,438 | INFO | train_grpo_train | metrics_logged mode=train step=114

                                                    
{'loss': 0.0008, 'grad_norm': 9.978446006774902, 'learning_rate': 9.657575757575758e-06, 'num_tokens': 219248.0, 'completions/mean_length': 92.875, 'completions/min_length': 70.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.875, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.8164951801300049, 'rewards/meter/std': 0.320004403591156, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.961208701133728, 'rewards/lexical_plausibility/std': 0.10971830040216446, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9959613084793091, 'rewards/repeat_penalty/std': 0.004583663772791624, 'rewards/repeat_floor/mean': 0.9993941783905029, 'rewards/repeat_floor/std': 0.0006875431863591075, 'rewards/near_duplicate_penalty/mean': 0.9959613084793091, 'rewards/near_duplicate_penalty/std': 0.004583663772791624, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25260108709335327, 'rewards/total_composite/std': 0.1469898819923401, 'reward': 0.25260108709335327, 'reward_std': 0.1469898819923401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20012064278125763, 'sampling/sampling_logp_difference/max': 1.081636905670166, 'sampling/importance_sampling_ratio/min': 0.3390401005744934, 'sampling/importance_sampling_ratio/mean': 1.0457664728164673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8036739826202393, 'clip_ratio/low_mean': 0.08223419077694416, 'clip_ratio/low_min': 0.08223419077694416, 'clip_ratio/high_mean': 0.12496751919388771, 'clip_ratio/high_max': 0.12496751919388771, 'clip_ratio/region_mean': 0.20720170997083187, 'reward_total_mean': 0.25260108709335327, 'reward_meter_mean': 0.8164951801300049, 'reward_meter_std': 0.320004403591156, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.961208701133728, 'reward_lexical_plausibility_std': 0.10971830040216446, 'reward_repeat_penalty_mean': 0.9959613084793091, 'reward_repeat_penalty_std': 0.004583663772791624, 'reward_repeat_floor_mean': 0.9993941783905029, 'reward_repeat_floor_std': 0.0006875431863591075, 'reward_near_duplicate_penalty_mean': 0.9959613084793091, 'reward_near_duplicate_penalty_std': 0.004583663772791624, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.25260108709335327, 'reward_total_composite_std': 0.1469898819923401, 'epoch': 0.01}

  3%|▎         | 114/3300 [23:13<8:30:01,  9.61s/it]INFO 04-12 16:21:39 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 115/3300 [23:22<8:28:56,  9.59s/it]2026-04-12 16:21:48,982 | INFO | train_grpo_train | metrics_logged mode=train step=115

                                                    
{'loss': 0.1745, 'grad_norm': 13.326382637023926, 'learning_rate': 9.654545454545456e-06, 'num_tokens': 221079.0, 'completions/mean_length': 54.875, 'completions/min_length': 41.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.7106096744537354, 'rewards/meter/std': 0.3569812476634979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9928355813026428, 'rewards/lexical_plausibility/std': 0.020263994112610817, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.11259915679693222, 'rewards/repeat_penalty/mean': 0.9856451749801636, 'rewards/repeat_penalty/std': 0.012042542919516563, 'rewards/repeat_floor/mean': 0.997846782207489, 'rewards/repeat_floor/std': 0.0018063782481476665, 'rewards/near_duplicate_penalty/mean': 0.9856451749801636, 'rewards/near_duplicate_penalty/std': 0.012042542919516563, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2591401934623718, 'rewards/total_composite/std': 0.16632743179798126, 'reward': 0.2591401934623718, 'reward_std': 0.16632743179798126, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19514282047748566, 'sampling/sampling_logp_difference/max': 1.364858627319336, 'sampling/importance_sampling_ratio/min': 0.25541678071022034, 'sampling/importance_sampling_ratio/mean': 1.0378247499465942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.176184728741646, 'clip_ratio/low_mean': 0.10042064264416695, 'clip_ratio/low_min': 0.10042064264416695, 'clip_ratio/high_mean': 0.09279343206435442, 'clip_ratio/high_max': 0.09279343206435442, 'clip_ratio/region_mean': 0.19321407470852137, 'reward_total_mean': 0.2591401934623718, 'reward_meter_mean': 0.7106096744537354, 'reward_meter_std': 0.3569812476634979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9928355813026428, 'reward_lexical_plausibility_std': 0.020263994112610817, 'reward_repeat_penalty_mean': 0.9856451749801636, 'reward_repeat_penalty_std': 0.012042542919516563, 'reward_repeat_floor_mean': 0.997846782207489, 'reward_repeat_floor_std': 0.0018063782481476665, 'reward_near_duplicate_penalty_mean': 0.9856451749801636, 'reward_near_duplicate_penalty_std': 0.012042542919516563, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.11259915679693222, 'reward_total_composite_mean': 0.2591401934623718, 'reward_total_composite_std': 0.16632743179798126, 'epoch': 0.01}

  3%|▎         | 115/3300 [23:22<8:28:56,  9.59s/it]INFO 04-12 16:21:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 116/3300 [23:33<8:47:51,  9.95s/it]2026-04-12 16:21:59,765 | INFO | train_grpo_train | metrics_logged mode=train step=116

                                                    
{'loss': 0.1931, 'grad_norm': 9.80229377746582, 'learning_rate': 9.651515151515153e-06, 'num_tokens': 223223.0, 'completions/mean_length': 96.0, 'completions/min_length': 77.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.7641827464103699, 'rewards/meter/std': 0.32104983925819397, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9899832010269165, 'rewards/repeat_penalty/std': 0.011829826049506664, 'rewards/repeat_floor/mean': 0.998497486114502, 'rewards/repeat_floor/std': 0.0017744723008945584, 'rewards/near_duplicate_penalty/mean': 0.9899832010269165, 'rewards/near_duplicate_penalty/std': 0.011829826049506664, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31298011541366577, 'rewards/total_composite/std': 0.1595117598772049, 'reward': 0.31298011541366577, 'reward_std': 0.1595117449760437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20624369382858276, 'sampling/sampling_logp_difference/max': 1.57826566696167, 'sampling/importance_sampling_ratio/min': 0.20633263885974884, 'sampling/importance_sampling_ratio/mean': 1.0410480499267578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.560793563723564, 'clip_ratio/low_mean': 0.06803693901747465, 'clip_ratio/low_min': 0.06803693901747465, 'clip_ratio/high_mean': 0.13042096979916096, 'clip_ratio/high_max': 0.13042096979916096, 'clip_ratio/region_mean': 0.1984579088166356, 'reward_total_mean': 0.31298011541366577, 'reward_meter_mean': 0.7641827464103699, 'reward_meter_std': 0.32104983925819397, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9899832010269165, 'reward_repeat_penalty_std': 0.011829826049506664, 'reward_repeat_floor_mean': 0.998497486114502, 'reward_repeat_floor_std': 0.0017744723008945584, 'reward_near_duplicate_penalty_mean': 0.9899832010269165, 'reward_near_duplicate_penalty_std': 0.011829826049506664, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.31298011541366577, 'reward_total_composite_std': 0.1595117598772049, 'epoch': 0.01}

  4%|▎         | 116/3300 [23:33<8:47:51,  9.95s/it]INFO 04-12 16:22:00 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 117/3300 [23:43<8:50:28, 10.00s/it]2026-04-12 16:22:09,885 | INFO | train_grpo_train | metrics_logged mode=train step=117

                                                    
{'loss': 0.0745, 'grad_norm': 13.605415344238281, 'learning_rate': 9.648484848484849e-06, 'num_tokens': 224921.0, 'completions/mean_length': 51.25, 'completions/min_length': 35.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5380079746246338, 'rewards/meter/std': 0.4626144468784332, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9967199563980103, 'rewards/repeat_penalty/std': 0.006169808097183704, 'rewards/repeat_floor/mean': 0.9995079636573792, 'rewards/repeat_floor/std': 0.0009254738688468933, 'rewards/near_duplicate_penalty/mean': 0.9967199563980103, 'rewards/near_duplicate_penalty/std': 0.006169808097183704, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20348688960075378, 'rewards/total_composite/std': 0.18260757625102997, 'reward': 0.20348688960075378, 'reward_std': 0.18260756134986877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2219952493906021, 'sampling/sampling_logp_difference/max': 1.4091901779174805, 'sampling/importance_sampling_ratio/min': 0.24434107542037964, 'sampling/importance_sampling_ratio/mean': 1.0295130014419556, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8360769152641296, 'clip_ratio/low_mean': 0.09715486317873001, 'clip_ratio/low_min': 0.09715486317873001, 'clip_ratio/high_mean': 0.09219090826809406, 'clip_ratio/high_max': 0.09219090826809406, 'clip_ratio/region_mean': 0.18934577144682407, 'reward_total_mean': 0.20348688960075378, 'reward_meter_mean': 0.5380079746246338, 'reward_meter_std': 0.4626144468784332, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9967199563980103, 'reward_repeat_penalty_std': 0.006169808097183704, 'reward_repeat_floor_mean': 0.9995079636573792, 'reward_repeat_floor_std': 0.0009254738688468933, 'reward_near_duplicate_penalty_mean': 0.9967199563980103, 'reward_near_duplicate_penalty_std': 0.006169808097183704, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.20348688960075378, 'reward_total_composite_std': 0.18260757625102997, 'epoch': 0.01}

  4%|▎         | 117/3300 [23:43<8:50:28, 10.00s/it]INFO 04-12 16:22:10 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 118/3300 [23:52<8:24:59,  9.52s/it]2026-04-12 16:22:18,300 | INFO | train_grpo_train | metrics_logged mode=train step=118

                                                    
{'loss': 0.0664, 'grad_norm': 15.919730186462402, 'learning_rate': 9.645454545454548e-06, 'num_tokens': 226333.0, 'completions/mean_length': 24.5, 'completions/min_length': 21.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6992982625961304, 'rewards/meter/std': 0.4218030273914337, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9170168042182922, 'rewards/lexical_plausibility/std': 0.2347119152545929, 'rewards/judge_quality/mean': 0.6875, 'rewards/judge_quality/std': 0.3739270567893982, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5495831966400146, 'rewards/total_composite/std': 0.42945122718811035, 'reward': 0.5495831966400146, 'reward_std': 0.42945122718811035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1402398943901062, 'sampling/sampling_logp_difference/max': 1.2016280889511108, 'sampling/importance_sampling_ratio/min': 0.3007042407989502, 'sampling/importance_sampling_ratio/mean': 1.0070768594741821, 'sampling/importance_sampling_ratio/max': 1.9456191062927246, 'entropy': 1.0227503515779972, 'clip_ratio/low_mean': 0.06757575832307339, 'clip_ratio/low_min': 0.06757575832307339, 'clip_ratio/high_mean': 0.060659651178866625, 'clip_ratio/high_max': 0.060659651178866625, 'clip_ratio/region_mean': 0.12823540950194, 'reward_total_mean': 0.5495831966400146, 'reward_meter_mean': 0.6992982625961304, 'reward_meter_std': 0.4218030273914337, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9170168042182922, 'reward_lexical_plausibility_std': 0.2347119152545929, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6875, 'reward_judge_quality_std': 0.3739270567893982, 'reward_total_composite_mean': 0.5495831966400146, 'reward_total_composite_std': 0.42945122718811035, 'epoch': 0.01}

  4%|▎         | 118/3300 [23:52<8:24:59,  9.52s/it]INFO 04-12 16:22:18 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 119/3300 [24:00<7:58:29,  9.03s/it]2026-04-12 16:22:26,160 | INFO | train_grpo_train | metrics_logged mode=train step=119

                                                    
{'loss': 0.2095, 'grad_norm': 27.368627548217773, 'learning_rate': 9.642424242424243e-06, 'num_tokens': 227751.0, 'completions/mean_length': 17.25, 'completions/min_length': 10.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.25, 'completions/min_terminated_length': 10.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.1176612600684166, 'rewards/meter/std': 0.1472979336977005, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.970588207244873, 'rewards/lexical_plausibility/std': 0.05446000397205353, 'rewards/judge_quality/mean': 0.7762500047683716, 'rewards/judge_quality/std': 0.30500292778015137, 'rewards/repeat_penalty/mean': 0.7320075631141663, 'rewards/repeat_penalty/std': 0.050890251994132996, 'rewards/repeat_floor/mean': 0.9814015626907349, 'rewards/repeat_floor/std': 0.010178063064813614, 'rewards/near_duplicate_penalty/mean': 0.9760100841522217, 'rewards/near_duplicate_penalty/std': 0.06785368919372559, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.09496261179447174, 'rewards/total_composite/std': 0.13875754177570343, 'reward': 0.09496261179447174, 'reward_std': 0.13875754177570343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2491590529680252, 'sampling/sampling_logp_difference/max': 1.505394458770752, 'sampling/importance_sampling_ratio/min': 0.22192974388599396, 'sampling/importance_sampling_ratio/mean': 1.0080736875534058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4507152736186981, 'clip_ratio/low_mean': 0.189669081941247, 'clip_ratio/low_min': 0.189669081941247, 'clip_ratio/high_mean': 0.07500000018626451, 'clip_ratio/high_max': 0.07500000018626451, 'clip_ratio/region_mean': 0.2646690821275115, 'reward_total_mean': 0.09496261179447174, 'reward_meter_mean': 0.1176612600684166, 'reward_meter_std': 0.1472979336977005, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.970588207244873, 'reward_lexical_plausibility_std': 0.05446000397205353, 'reward_repeat_penalty_mean': 0.7320075631141663, 'reward_repeat_penalty_std': 0.050890251994132996, 'reward_repeat_floor_mean': 0.9814015626907349, 'reward_repeat_floor_std': 0.010178063064813614, 'reward_near_duplicate_penalty_mean': 0.9760100841522217, 'reward_near_duplicate_penalty_std': 0.06785368919372559, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7762500047683716, 'reward_judge_quality_std': 0.30500292778015137, 'reward_total_composite_mean': 0.09496261179447174, 'reward_total_composite_std': 0.13875754177570343, 'epoch': 0.01}

  4%|▎         | 119/3300 [24:00<7:58:29,  9.03s/it]INFO 04-12 16:22:26 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 120/3300 [24:09<8:10:20,  9.25s/it]2026-04-12 16:22:35,945 | INFO | train_grpo_train | metrics_logged mode=train step=120

                                                    
{'loss': 0.0361, 'grad_norm': 11.578890800476074, 'learning_rate': 9.63939393939394e-06, 'num_tokens': 229648.0, 'completions/mean_length': 55.125, 'completions/min_length': 44.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5034680962562561, 'rewards/meter/std': 0.40532049536705017, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.908367395401001, 'rewards/lexical_plausibility/std': 0.16053621470928192, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9956713914871216, 'rewards/repeat_penalty/std': 0.008281382732093334, 'rewards/repeat_floor/mean': 0.9993507266044617, 'rewards/repeat_floor/std': 0.0012422117870301008, 'rewards/near_duplicate_penalty/mean': 0.9956713914871216, 'rewards/near_duplicate_penalty/std': 0.008281382732093334, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15261298418045044, 'rewards/total_composite/std': 0.1726185828447342, 'reward': 0.15261298418045044, 'reward_std': 0.17261859774589539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22696611285209656, 'sampling/sampling_logp_difference/max': 1.343881607055664, 'sampling/importance_sampling_ratio/min': 0.2608312666416168, 'sampling/importance_sampling_ratio/mean': 1.028020977973938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.654230237007141, 'clip_ratio/low_mean': 0.13845311105251312, 'clip_ratio/low_min': 0.13845311105251312, 'clip_ratio/high_mean': 0.05308493785560131, 'clip_ratio/high_max': 0.05308493785560131, 'clip_ratio/region_mean': 0.19153804890811443, 'reward_total_mean': 0.15261298418045044, 'reward_meter_mean': 0.5034680962562561, 'reward_meter_std': 0.40532049536705017, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.908367395401001, 'reward_lexical_plausibility_std': 0.16053621470928192, 'reward_repeat_penalty_mean': 0.9956713914871216, 'reward_repeat_penalty_std': 0.008281382732093334, 'reward_repeat_floor_mean': 0.9993507266044617, 'reward_repeat_floor_std': 0.0012422117870301008, 'reward_near_duplicate_penalty_mean': 0.9956713914871216, 'reward_near_duplicate_penalty_std': 0.008281382732093334, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.15261298418045044, 'reward_total_composite_std': 0.1726185828447342, 'epoch': 0.01}

  4%|▎         | 120/3300 [24:09<8:10:20,  9.25s/it]INFO 04-12 16:22:36 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 121/3300 [24:22<9:10:19, 10.39s/it]2026-04-12 16:22:48,975 | INFO | train_grpo_train | metrics_logged mode=train step=121

                                                    
{'loss': 0.0606, 'grad_norm': 5.397149085998535, 'learning_rate': 9.636363636363638e-06, 'num_tokens': 233918.0, 'completions/mean_length': 301.75, 'completions/min_length': 158.0, 'completions/max_length': 381.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 301.75, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 381.0, 'rewards/meter/mean': 0.22918570041656494, 'rewards/meter/std': 0.12608183920383453, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.13849148154258728, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9374999403953552, 'rewards/count_floor/std': 0.041547439992427826, 'rewards/lexical_plausibility/mean': 0.8940296173095703, 'rewards/lexical_plausibility/std': 0.11361392587423325, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.053452249616384506, 'rewards/repeat_penalty/mean': 0.9928607940673828, 'rewards/repeat_penalty/std': 0.011244257912039757, 'rewards/repeat_floor/mean': 0.9992445707321167, 'rewards/repeat_floor/std': 0.0010756528936326504, 'rewards/near_duplicate_penalty/mean': 0.9981446266174316, 'rewards/near_duplicate_penalty/std': 0.0015453349333256483, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9946988821029663, 'rewards/distinct_2/std': 0.010371833108365536, 'rewards/total_composite/mean': 0.033423878252506256, 'rewards/total_composite/std': 0.026064569130539894, 'reward': 0.033423878252506256, 'reward_std': 0.026064569130539894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22285933792591095, 'sampling/sampling_logp_difference/max': 1.6598758697509766, 'sampling/importance_sampling_ratio/min': 0.19016258418560028, 'sampling/importance_sampling_ratio/mean': 1.0520243644714355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.5553068220615387, 'clip_ratio/low_mean': 0.07398652471601963, 'clip_ratio/low_min': 0.07398652471601963, 'clip_ratio/high_mean': 0.09162890538573265, 'clip_ratio/high_max': 0.09162890538573265, 'clip_ratio/region_mean': 0.16561543010175228, 'reward_total_mean': 0.033423878252506256, 'reward_meter_mean': 0.22918570041656494, 'reward_meter_std': 0.12608183920383453, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.13849148154258728, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9374999403953552, 'reward_count_floor_std': 0.041547439992427826, 'reward_lexical_plausibility_mean': 0.8940296173095703, 'reward_lexical_plausibility_std': 0.11361392587423325, 'reward_repeat_penalty_mean': 0.9928607940673828, 'reward_repeat_penalty_std': 0.011244257912039757, 'reward_repeat_floor_mean': 0.9992445707321167, 'reward_repeat_floor_std': 0.0010756528936326504, 'reward_near_duplicate_penalty_mean': 0.9981446266174316, 'reward_near_duplicate_penalty_std': 0.0015453349333256483, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9946988821029663, 'reward_distinct_2_std': 0.010371833108365536, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.053452249616384506, 'reward_total_composite_mean': 0.033423878252506256, 'reward_total_composite_std': 0.026064569130539894, 'epoch': 0.01}

  4%|▎         | 121/3300 [24:22<9:10:19, 10.39s/it]INFO 04-12 16:22:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 122/3300 [24:34<9:35:41, 10.87s/it]2026-04-12 16:23:01,030 | INFO | train_grpo_train | metrics_logged mode=train step=122

                                                    
{'loss': 0.1563, 'grad_norm': 7.213613510131836, 'learning_rate': 9.633333333333335e-06, 'num_tokens': 236969.0, 'completions/mean_length': 197.375, 'completions/min_length': 146.0, 'completions/max_length': 332.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.375, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 332.0, 'rewards/meter/mean': 0.33072274923324585, 'rewards/meter/std': 0.33584359288215637, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9678571224212646, 'rewards/count_floor/std': 0.01983901485800743, 'rewards/lexical_plausibility/mean': 0.9770627021789551, 'rewards/lexical_plausibility/std': 0.0626121461391449, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9937963485717773, 'rewards/repeat_penalty/std': 0.005309552885591984, 'rewards/repeat_floor/mean': 0.9991875886917114, 'rewards/repeat_floor/std': 0.0005017775110900402, 'rewards/near_duplicate_penalty/mean': 0.9957737922668457, 'rewards/near_duplicate_penalty/std': 0.00208354857750237, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.998017430305481, 'rewards/distinct_2/std': 0.0056075118482112885, 'rewards/total_composite/mean': 0.12869678437709808, 'rewards/total_composite/std': 0.13592872023582458, 'reward': 0.12869678437709808, 'reward_std': 0.13592872023582458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24395523965358734, 'sampling/sampling_logp_difference/max': 2.389272451400757, 'sampling/importance_sampling_ratio/min': 0.0916963741183281, 'sampling/importance_sampling_ratio/mean': 1.0517562627792358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2598190009593964, 'clip_ratio/low_mean': 0.11706330999732018, 'clip_ratio/low_min': 0.11706330999732018, 'clip_ratio/high_mean': 0.08133545145392418, 'clip_ratio/high_max': 0.08133545145392418, 'clip_ratio/region_mean': 0.19839876145124435, 'reward_total_mean': 0.12869678437709808, 'reward_meter_mean': 0.33072274923324585, 'reward_meter_std': 0.33584359288215637, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9678571224212646, 'reward_count_floor_std': 0.01983901485800743, 'reward_lexical_plausibility_mean': 0.9770627021789551, 'reward_lexical_plausibility_std': 0.0626121461391449, 'reward_repeat_penalty_mean': 0.9937963485717773, 'reward_repeat_penalty_std': 0.005309552885591984, 'reward_repeat_floor_mean': 0.9991875886917114, 'reward_repeat_floor_std': 0.0005017775110900402, 'reward_near_duplicate_penalty_mean': 0.9957737922668457, 'reward_near_duplicate_penalty_std': 0.00208354857750237, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.998017430305481, 'reward_distinct_2_std': 0.0056075118482112885, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.12869678437709808, 'reward_total_composite_std': 0.13592872023582458, 'epoch': 0.01}

  4%|▎         | 122/3300 [24:34<9:35:41, 10.87s/it]INFO 04-12 16:23:01 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 123/3300 [24:45<9:36:45, 10.89s/it]2026-04-12 16:23:11,917 | INFO | train_grpo_train | metrics_logged mode=train step=123

                                                    
{'loss': -0.006, 'grad_norm': 7.073686599731445, 'learning_rate': 9.63030303030303e-06, 'num_tokens': 239390.0, 'completions/mean_length': 132.625, 'completions/min_length': 88.0, 'completions/max_length': 246.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.625, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 246.0, 'rewards/meter/mean': 0.28302252292633057, 'rewards/meter/std': 0.3588270843029022, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.29124119877815247, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.08737236261367798, 'rewards/lexical_plausibility/mean': 0.9583333134651184, 'rewards/lexical_plausibility/std': 0.08147962391376495, 'rewards/judge_quality/mean': 0.3412500023841858, 'rewards/judge_quality/std': 0.2391316145658493, 'rewards/repeat_penalty/mean': 0.9990323185920715, 'rewards/repeat_penalty/std': 0.0012910085497424006, 'rewards/repeat_floor/mean': 0.9998548626899719, 'rewards/repeat_floor/std': 0.00019365194020792842, 'rewards/near_duplicate_penalty/mean': 0.9990323185920715, 'rewards/near_duplicate_penalty/std': 0.0012910085497424006, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.0783679261803627, 'rewards/total_composite/std': 0.11503598839044571, 'reward': 0.0783679261803627, 'reward_std': 0.11503598093986511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21352066099643707, 'sampling/sampling_logp_difference/max': 1.3450571298599243, 'sampling/importance_sampling_ratio/min': 0.26052483916282654, 'sampling/importance_sampling_ratio/mean': 1.0447856187820435, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9213546067476273, 'clip_ratio/low_mean': 0.14327157847583294, 'clip_ratio/low_min': 0.14327157847583294, 'clip_ratio/high_mean': 0.051379503682255745, 'clip_ratio/high_max': 0.051379503682255745, 'clip_ratio/region_mean': 0.19465108215808868, 'reward_total_mean': 0.0783679261803627, 'reward_meter_mean': 0.28302252292633057, 'reward_meter_std': 0.3588270843029022, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.29124119877815247, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.08737236261367798, 'reward_lexical_plausibility_mean': 0.9583333134651184, 'reward_lexical_plausibility_std': 0.08147962391376495, 'reward_repeat_penalty_mean': 0.9990323185920715, 'reward_repeat_penalty_std': 0.0012910085497424006, 'reward_repeat_floor_mean': 0.9998548626899719, 'reward_repeat_floor_std': 0.00019365194020792842, 'reward_near_duplicate_penalty_mean': 0.9990323185920715, 'reward_near_duplicate_penalty_std': 0.0012910085497424006, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3412500023841858, 'reward_judge_quality_std': 0.2391316145658493, 'reward_total_composite_mean': 0.0783679261803627, 'reward_total_composite_std': 0.11503598839044571, 'epoch': 0.01}

  4%|▎         | 123/3300 [24:45<9:36:45, 10.89s/it]INFO 04-12 16:23:12 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 124/3300 [24:54<9:06:36, 10.33s/it]2026-04-12 16:23:20,924 | INFO | train_grpo_train | metrics_logged mode=train step=124

                                                    
{'loss': 0.1116, 'grad_norm': 18.304222106933594, 'learning_rate': 9.627272727272728e-06, 'num_tokens': 240881.0, 'completions/mean_length': 48.375, 'completions/min_length': 42.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.25827205181121826, 'rewards/meter/std': 0.3599354922771454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9861021041870117, 'rewards/lexical_plausibility/std': 0.03930910304188728, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136146903038025, 'rewards/repeat_penalty/mean': 0.9810866117477417, 'rewards/repeat_penalty/std': 0.02608991600573063, 'rewards/repeat_floor/mean': 0.9971629977226257, 'rewards/repeat_floor/std': 0.003913478925824165, 'rewards/near_duplicate_penalty/mean': 0.9810866117477417, 'rewards/near_duplicate_penalty/std': 0.02608991600573063, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1445094496011734, 'rewards/total_composite/std': 0.2417820245027542, 'reward': 0.1445094496011734, 'reward_std': 0.24178199470043182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22063003480434418, 'sampling/sampling_logp_difference/max': 1.8402786254882812, 'sampling/importance_sampling_ratio/min': 0.15877318382263184, 'sampling/importance_sampling_ratio/mean': 1.0315083265304565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4374092072248459, 'clip_ratio/low_mean': 0.17583125084638596, 'clip_ratio/low_min': 0.17583125084638596, 'clip_ratio/high_mean': 0.06101190485060215, 'clip_ratio/high_max': 0.06101190485060215, 'clip_ratio/region_mean': 0.2368431556969881, 'reward_total_mean': 0.1445094496011734, 'reward_meter_mean': 0.25827205181121826, 'reward_meter_std': 0.3599354922771454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9861021041870117, 'reward_lexical_plausibility_std': 0.03930910304188728, 'reward_repeat_penalty_mean': 0.9810866117477417, 'reward_repeat_penalty_std': 0.02608991600573063, 'reward_repeat_floor_mean': 0.9971629977226257, 'reward_repeat_floor_std': 0.003913478925824165, 'reward_near_duplicate_penalty_mean': 0.9810866117477417, 'reward_near_duplicate_penalty_std': 0.02608991600573063, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136146903038025, 'reward_total_composite_mean': 0.1445094496011734, 'reward_total_composite_std': 0.2417820245027542, 'epoch': 0.01}

  4%|▍         | 124/3300 [24:54<9:06:36, 10.33s/it]INFO 04-12 16:23:21 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 125/3300 [25:03<8:35:49,  9.75s/it]2026-04-12 16:23:29,322 | INFO | train_grpo_train | metrics_logged mode=train step=125

                                                    
{'loss': -0.1071, 'grad_norm': 17.003002166748047, 'learning_rate': 9.624242424242425e-06, 'num_tokens': 242305.0, 'completions/mean_length': 25.0, 'completions/min_length': 16.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.551611065864563, 'rewards/meter/std': 0.42970213294029236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.801948070526123, 'rewards/lexical_plausibility/std': 0.14781975746154785, 'rewards/judge_quality/mean': 0.5712499618530273, 'rewards/judge_quality/std': 0.31921499967575073, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38685962557792664, 'rewards/total_composite/std': 0.40095359086990356, 'reward': 0.38685962557792664, 'reward_std': 0.40095359086990356, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21530112624168396, 'sampling/sampling_logp_difference/max': 1.3130431175231934, 'sampling/importance_sampling_ratio/min': 0.26900020241737366, 'sampling/importance_sampling_ratio/mean': 1.0635402202606201, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5896531641483307, 'clip_ratio/low_mean': 0.09175990149378777, 'clip_ratio/low_min': 0.09175990149378777, 'clip_ratio/high_mean': 0.04286616249009967, 'clip_ratio/high_max': 0.04286616249009967, 'clip_ratio/region_mean': 0.13462606398388743, 'reward_total_mean': 0.38685962557792664, 'reward_meter_mean': 0.551611065864563, 'reward_meter_std': 0.42970213294029236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.801948070526123, 'reward_lexical_plausibility_std': 0.14781975746154785, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5712499618530273, 'reward_judge_quality_std': 0.31921499967575073, 'reward_total_composite_mean': 0.38685962557792664, 'reward_total_composite_std': 0.40095359086990356, 'epoch': 0.01}

  4%|▍         | 125/3300 [25:03<8:35:49,  9.75s/it]INFO 04-12 16:23:29 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 126/3300 [25:15<9:23:00, 10.64s/it]2026-04-12 16:23:42,052 | INFO | train_grpo_train | metrics_logged mode=train step=126

                                                    
{'loss': 0.0092, 'grad_norm': 12.127283096313477, 'learning_rate': 9.621212121212122e-06, 'num_tokens': 244523.0, 'completions/mean_length': 99.25, 'completions/min_length': 66.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.28973257541656494, 'rewards/meter/std': 0.2882465124130249, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9866727590560913, 'rewards/lexical_plausibility/std': 0.02516147866845131, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.977553129196167, 'rewards/repeat_penalty/std': 0.02701600268483162, 'rewards/repeat_floor/mean': 0.9974371194839478, 'rewards/repeat_floor/std': 0.0029925296548753977, 'rewards/near_duplicate_penalty/mean': 0.991321325302124, 'rewards/near_duplicate_penalty/std': 0.009695005603134632, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9859883785247803, 'rewards/distinct_2/std': 0.01951143704354763, 'rewards/total_composite/mean': 0.10840226709842682, 'rewards/total_composite/std': 0.13197503983974457, 'reward': 0.10840226709842682, 'reward_std': 0.13197503983974457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24986715614795685, 'sampling/sampling_logp_difference/max': 1.5561456680297852, 'sampling/importance_sampling_ratio/min': 0.21094755828380585, 'sampling/importance_sampling_ratio/mean': 1.066483974456787, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9682925045490265, 'clip_ratio/low_mean': 0.1341007500886917, 'clip_ratio/low_min': 0.1341007500886917, 'clip_ratio/high_mean': 0.07857861742377281, 'clip_ratio/high_max': 0.07857861742377281, 'clip_ratio/region_mean': 0.21267936751246452, 'reward_total_mean': 0.10840226709842682, 'reward_meter_mean': 0.28973257541656494, 'reward_meter_std': 0.2882465124130249, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9866727590560913, 'reward_lexical_plausibility_std': 0.02516147866845131, 'reward_repeat_penalty_mean': 0.977553129196167, 'reward_repeat_penalty_std': 0.02701600268483162, 'reward_repeat_floor_mean': 0.9974371194839478, 'reward_repeat_floor_std': 0.0029925296548753977, 'reward_near_duplicate_penalty_mean': 0.991321325302124, 'reward_near_duplicate_penalty_std': 0.009695005603134632, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9859883785247803, 'reward_distinct_2_std': 0.01951143704354763, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.10840226709842682, 'reward_total_composite_std': 0.13197503983974457, 'epoch': 0.01}

  4%|▍         | 126/3300 [25:15<9:23:00, 10.64s/it]INFO 04-12 16:23:42 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 127/3300 [25:24<8:44:54,  9.93s/it]2026-04-12 16:23:50,305 | INFO | train_grpo_train | metrics_logged mode=train step=127

                                                    
{'loss': -0.1594, 'grad_norm': 22.551822662353516, 'learning_rate': 9.61818181818182e-06, 'num_tokens': 246024.0, 'completions/mean_length': 18.625, 'completions/min_length': 14.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.35911935567855835, 'rewards/meter/std': 0.4125699996948242, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9397977590560913, 'rewards/lexical_plausibility/std': 0.09944810718297958, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2728912830352783, 'rewards/repeat_penalty/mean': 0.7033942937850952, 'rewards/repeat_penalty/std': 0.08695917576551437, 'rewards/repeat_floor/mean': 0.9779176115989685, 'rewards/repeat_floor/std': 0.013899269513785839, 'rewards/near_duplicate_penalty/mean': 0.9800275564193726, 'rewards/near_duplicate_penalty/std': 0.053229887038469315, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9545940160751343, 'rewards/distinct_2/std': 0.08426927030086517, 'rewards/total_composite/mean': 0.07746857404708862, 'rewards/total_composite/std': 0.11463268101215363, 'reward': 0.07746857404708862, 'reward_std': 0.11463267356157303, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19300872087478638, 'sampling/sampling_logp_difference/max': 1.3846263885498047, 'sampling/importance_sampling_ratio/min': 0.2504173517227173, 'sampling/importance_sampling_ratio/mean': 1.0627801418304443, 'sampling/importance_sampling_ratio/max': 1.9003657102584839, 'entropy': 2.346334755420685, 'clip_ratio/low_mean': 0.12997037917375565, 'clip_ratio/low_min': 0.12997037917375565, 'clip_ratio/high_mean': 0.08439814671874046, 'clip_ratio/high_max': 0.08439814671874046, 'clip_ratio/region_mean': 0.2143685258924961, 'reward_total_mean': 0.07746857404708862, 'reward_meter_mean': 0.35911935567855835, 'reward_meter_std': 0.4125699996948242, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9397977590560913, 'reward_lexical_plausibility_std': 0.09944810718297958, 'reward_repeat_penalty_mean': 0.7033942937850952, 'reward_repeat_penalty_std': 0.08695917576551437, 'reward_repeat_floor_mean': 0.9779176115989685, 'reward_repeat_floor_std': 0.013899269513785839, 'reward_near_duplicate_penalty_mean': 0.9800275564193726, 'reward_near_duplicate_penalty_std': 0.053229887038469315, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9545940160751343, 'reward_distinct_2_std': 0.08426927030086517, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2728912830352783, 'reward_total_composite_mean': 0.07746857404708862, 'reward_total_composite_std': 0.11463268101215363, 'epoch': 0.01}

  4%|▍         | 127/3300 [25:24<8:44:54,  9.93s/it]INFO 04-12 16:23:50 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 128/3300 [25:33<8:28:09,  9.61s/it]2026-04-12 16:23:59,185 | INFO | train_grpo_train | metrics_logged mode=train step=128

                                                    
{'loss': 0.1323, 'grad_norm': 16.545969009399414, 'learning_rate': 9.615151515151517e-06, 'num_tokens': 247659.0, 'completions/mean_length': 45.375, 'completions/min_length': 33.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.4964056611061096, 'rewards/meter/std': 0.4586038589477539, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.512499988079071, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.9968196153640747, 'rewards/repeat_penalty/std': 0.005409465171396732, 'rewards/repeat_floor/mean': 0.9995229244232178, 'rewards/repeat_floor/std': 0.0008114197407849133, 'rewards/near_duplicate_penalty/mean': 0.9968196153640747, 'rewards/near_duplicate_penalty/std': 0.005409465171396732, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2368769347667694, 'rewards/total_composite/std': 0.23428308963775635, 'reward': 0.2368769347667694, 'reward_std': 0.23428308963775635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23393887281417847, 'sampling/sampling_logp_difference/max': 1.572005271911621, 'sampling/importance_sampling_ratio/min': 0.20762839913368225, 'sampling/importance_sampling_ratio/mean': 1.0549741983413696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4049852788448334, 'clip_ratio/low_mean': 0.09577205963432789, 'clip_ratio/low_min': 0.09577205963432789, 'clip_ratio/high_mean': 0.0936755957081914, 'clip_ratio/high_max': 0.0936755957081914, 'clip_ratio/region_mean': 0.18944765534251928, 'reward_total_mean': 0.2368769347667694, 'reward_meter_mean': 0.4964056611061096, 'reward_meter_std': 0.4586038589477539, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9968196153640747, 'reward_repeat_penalty_std': 0.005409465171396732, 'reward_repeat_floor_mean': 0.9995229244232178, 'reward_repeat_floor_std': 0.0008114197407849133, 'reward_near_duplicate_penalty_mean': 0.9968196153640747, 'reward_near_duplicate_penalty_std': 0.005409465171396732, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.512499988079071, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.2368769347667694, 'reward_total_composite_std': 0.23428308963775635, 'epoch': 0.01}

  4%|▍         | 128/3300 [25:33<8:28:09,  9.61s/it]INFO 04-12 16:23:59 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 129/3300 [25:42<8:22:24,  9.51s/it]2026-04-12 16:24:08,446 | INFO | train_grpo_train | metrics_logged mode=train step=129

                                                    
{'loss': 0.206, 'grad_norm': 17.825443267822266, 'learning_rate': 9.612121212121212e-06, 'num_tokens': 249076.0, 'completions/mean_length': 33.125, 'completions/min_length': 19.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5484859943389893, 'rewards/meter/std': 0.36270710825920105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2639264166355133, 'rewards/repeat_penalty/mean': 0.9728013277053833, 'rewards/repeat_penalty/std': 0.04434370622038841, 'rewards/repeat_floor/mean': 0.996694803237915, 'rewards/repeat_floor/std': 0.004625349305570126, 'rewards/near_duplicate_penalty/mean': 0.9862070083618164, 'rewards/near_duplicate_penalty/std': 0.014626423828303814, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.26354318857192993, 'rewards/total_composite/std': 0.2759113609790802, 'reward': 0.26354318857192993, 'reward_std': 0.2759113311767578, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20456117391586304, 'sampling/sampling_logp_difference/max': 1.6085231304168701, 'sampling/importance_sampling_ratio/min': 0.20018304884433746, 'sampling/importance_sampling_ratio/mean': 0.9972600936889648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3621015101671219, 'clip_ratio/low_mean': 0.12675178702920675, 'clip_ratio/low_min': 0.12675178702920675, 'clip_ratio/high_mean': 0.08021616749465466, 'clip_ratio/high_max': 0.08021616749465466, 'clip_ratio/region_mean': 0.2069679545238614, 'reward_total_mean': 0.26354318857192993, 'reward_meter_mean': 0.5484859943389893, 'reward_meter_std': 0.36270710825920105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9728013277053833, 'reward_repeat_penalty_std': 0.04434370622038841, 'reward_repeat_floor_mean': 0.996694803237915, 'reward_repeat_floor_std': 0.004625349305570126, 'reward_near_duplicate_penalty_mean': 0.9862070083618164, 'reward_near_duplicate_penalty_std': 0.014626423828303814, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2639264166355133, 'reward_total_composite_mean': 0.26354318857192993, 'reward_total_composite_std': 0.2759113609790802, 'epoch': 0.01}

  4%|▍         | 129/3300 [25:42<8:22:24,  9.51s/it]INFO 04-12 16:24:08 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 130/3300 [25:52<8:29:08,  9.64s/it]2026-04-12 16:24:18,388 | INFO | train_grpo_train | metrics_logged mode=train step=130

                                                    
{'loss': 0.0948, 'grad_norm': 13.174181938171387, 'learning_rate': 9.60909090909091e-06, 'num_tokens': 250797.0, 'completions/mean_length': 61.125, 'completions/min_length': 50.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.3797729015350342, 'rewards/meter/std': 0.27309468388557434, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.993243932723999, 'rewards/repeat_penalty/std': 0.0058365557342767715, 'rewards/repeat_floor/mean': 0.9989866018295288, 'rewards/repeat_floor/std': 0.0008754919399507344, 'rewards/near_duplicate_penalty/mean': 0.993243932723999, 'rewards/near_duplicate_penalty/std': 0.0058365557342767715, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14400315284729004, 'rewards/total_composite/std': 0.10686323046684265, 'reward': 0.14400315284729004, 'reward_std': 0.10686323791742325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21904681622982025, 'sampling/sampling_logp_difference/max': 2.2688913345336914, 'sampling/importance_sampling_ratio/min': 0.10342678427696228, 'sampling/importance_sampling_ratio/mean': 1.0242271423339844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2305078506469727, 'clip_ratio/low_mean': 0.09416982345283031, 'clip_ratio/low_min': 0.09416982345283031, 'clip_ratio/high_mean': 0.08246147818863392, 'clip_ratio/high_max': 0.08246147818863392, 'clip_ratio/region_mean': 0.17663130164146423, 'reward_total_mean': 0.14400315284729004, 'reward_meter_mean': 0.3797729015350342, 'reward_meter_std': 0.27309468388557434, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.993243932723999, 'reward_repeat_penalty_std': 0.0058365557342767715, 'reward_repeat_floor_mean': 0.9989866018295288, 'reward_repeat_floor_std': 0.0008754919399507344, 'reward_near_duplicate_penalty_mean': 0.993243932723999, 'reward_near_duplicate_penalty_std': 0.0058365557342767715, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.14400315284729004, 'reward_total_composite_std': 0.10686323046684265, 'epoch': 0.01}

  4%|▍         | 130/3300 [25:52<8:29:08,  9.64s/it]INFO 04-12 16:24:18 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 131/3300 [26:02<8:44:07,  9.92s/it]2026-04-12 16:24:29,020 | INFO | train_grpo_train | metrics_logged mode=train step=131

                                                    
{'loss': -0.0146, 'grad_norm': 8.176576614379883, 'learning_rate': 9.606060606060607e-06, 'num_tokens': 253173.0, 'completions/mean_length': 112.0, 'completions/min_length': 63.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.5862672328948975, 'rewards/meter/std': 0.4476952850818634, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.07763238251209259, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.04419417306780815, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9982891082763672, 'rewards/repeat_penalty/std': 0.0020285688806325197, 'rewards/repeat_floor/mean': 0.9997433423995972, 'rewards/repeat_floor/std': 0.00030428983154706657, 'rewards/near_duplicate_penalty/mean': 0.9982891082763672, 'rewards/near_duplicate_penalty/std': 0.0020285688806325197, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21008649468421936, 'rewards/total_composite/std': 0.1683793067932129, 'reward': 0.21008649468421936, 'reward_std': 0.1683792918920517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20571686327457428, 'sampling/sampling_logp_difference/max': 1.4152746200561523, 'sampling/importance_sampling_ratio/min': 0.2428589165210724, 'sampling/importance_sampling_ratio/mean': 1.0346027612686157, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.915090709924698, 'clip_ratio/low_mean': 0.0838995985686779, 'clip_ratio/low_min': 0.0838995985686779, 'clip_ratio/high_mean': 0.10998196713626385, 'clip_ratio/high_max': 0.10998196713626385, 'clip_ratio/region_mean': 0.19388156570494175, 'reward_total_mean': 0.21008649468421936, 'reward_meter_mean': 0.5862672328948975, 'reward_meter_std': 0.4476952850818634, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.07763238251209259, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.04419417306780815, 'reward_repeat_penalty_mean': 0.9982891082763672, 'reward_repeat_penalty_std': 0.0020285688806325197, 'reward_repeat_floor_mean': 0.9997433423995972, 'reward_repeat_floor_std': 0.00030428983154706657, 'reward_near_duplicate_penalty_mean': 0.9982891082763672, 'reward_near_duplicate_penalty_std': 0.0020285688806325197, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.21008649468421936, 'reward_total_composite_std': 0.1683793067932129, 'epoch': 0.01}

  4%|▍         | 131/3300 [26:02<8:44:07,  9.92s/it]INFO 04-12 16:24:29 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 132/3300 [26:11<8:22:06,  9.51s/it]2026-04-12 16:24:37,631 | INFO | train_grpo_train | metrics_logged mode=train step=132

                                                    
{'loss': 0.0517, 'grad_norm': 14.145821571350098, 'learning_rate': 9.603030303030304e-06, 'num_tokens': 254954.0, 'completions/mean_length': 47.625, 'completions/min_length': 36.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7534605264663696, 'rewards/meter/std': 0.4120592474937439, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9898897409439087, 'rewards/lexical_plausibility/std': 0.028596237301826477, 'rewards/judge_quality/mean': 0.5799999833106995, 'rewards/judge_quality/std': 0.24395550787448883, 'rewards/repeat_penalty/mean': 0.9879139065742493, 'rewards/repeat_penalty/std': 0.0105670765042305, 'rewards/repeat_floor/mean': 0.9981870651245117, 'rewards/repeat_floor/std': 0.0015850637573748827, 'rewards/near_duplicate_penalty/mean': 0.9879139065742493, 'rewards/near_duplicate_penalty/std': 0.0105670765042305, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4782547950744629, 'rewards/total_composite/std': 0.34520453214645386, 'reward': 0.4782547950744629, 'reward_std': 0.34520453214645386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20471051335334778, 'sampling/sampling_logp_difference/max': 1.7622137069702148, 'sampling/importance_sampling_ratio/min': 0.17166441679000854, 'sampling/importance_sampling_ratio/mean': 1.0301520824432373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.475692719221115, 'clip_ratio/low_mean': 0.09087053965777159, 'clip_ratio/low_min': 0.09087053965777159, 'clip_ratio/high_mean': 0.07378472201526165, 'clip_ratio/high_max': 0.07378472201526165, 'clip_ratio/region_mean': 0.16465526167303324, 'reward_total_mean': 0.4782547950744629, 'reward_meter_mean': 0.7534605264663696, 'reward_meter_std': 0.4120592474937439, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9898897409439087, 'reward_lexical_plausibility_std': 0.028596237301826477, 'reward_repeat_penalty_mean': 0.9879139065742493, 'reward_repeat_penalty_std': 0.0105670765042305, 'reward_repeat_floor_mean': 0.9981870651245117, 'reward_repeat_floor_std': 0.0015850637573748827, 'reward_near_duplicate_penalty_mean': 0.9879139065742493, 'reward_near_duplicate_penalty_std': 0.0105670765042305, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5799999833106995, 'reward_judge_quality_std': 0.24395550787448883, 'reward_total_composite_mean': 0.4782547950744629, 'reward_total_composite_std': 0.34520453214645386, 'epoch': 0.01}

  4%|▍         | 132/3300 [26:11<8:22:06,  9.51s/it]INFO 04-12 16:24:37 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 133/3300 [26:19<8:07:56,  9.24s/it]2026-04-12 16:24:46,148 | INFO | train_grpo_train | metrics_logged mode=train step=133

                                                    
{'loss': 0.0629, 'grad_norm': 19.718576431274414, 'learning_rate': 9.600000000000001e-06, 'num_tokens': 256310.0, 'completions/mean_length': 18.5, 'completions/min_length': 13.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.5, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.4098302721977234, 'rewards/meter/std': 0.44320231676101685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9459689855575562, 'rewards/lexical_plausibility/std': 0.12943005561828613, 'rewards/judge_quality/mean': 0.29625001549720764, 'rewards/judge_quality/std': 0.26773855090141296, 'rewards/repeat_penalty/mean': 0.7464906573295593, 'rewards/repeat_penalty/std': 0.009925962425768375, 'rewards/repeat_floor/mean': 0.9842981100082397, 'rewards/repeat_floor/std': 0.0019852009136229753, 'rewards/near_duplicate_penalty/mean': 0.9953208565711975, 'rewards/near_duplicate_penalty/std': 0.013234616257250309, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14574924111366272, 'rewards/total_composite/std': 0.272583931684494, 'reward': 0.14574924111366272, 'reward_std': 0.272583931684494, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24437402188777924, 'sampling/sampling_logp_difference/max': 0.9934263229370117, 'sampling/importance_sampling_ratio/min': 0.37030574679374695, 'sampling/importance_sampling_ratio/mean': 1.0548073053359985, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.498630940914154, 'clip_ratio/low_mean': 0.15323186246678233, 'clip_ratio/low_min': 0.15323186246678233, 'clip_ratio/high_mean': 0.02777777798473835, 'clip_ratio/high_max': 0.02777777798473835, 'clip_ratio/region_mean': 0.18100964045152068, 'reward_total_mean': 0.14574924111366272, 'reward_meter_mean': 0.4098302721977234, 'reward_meter_std': 0.44320231676101685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9459689855575562, 'reward_lexical_plausibility_std': 0.12943005561828613, 'reward_repeat_penalty_mean': 0.7464906573295593, 'reward_repeat_penalty_std': 0.009925962425768375, 'reward_repeat_floor_mean': 0.9842981100082397, 'reward_repeat_floor_std': 0.0019852009136229753, 'reward_near_duplicate_penalty_mean': 0.9953208565711975, 'reward_near_duplicate_penalty_std': 0.013234616257250309, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.29625001549720764, 'reward_judge_quality_std': 0.26773855090141296, 'reward_total_composite_mean': 0.14574924111366272, 'reward_total_composite_std': 0.272583931684494, 'epoch': 0.01}

  4%|▍         | 133/3300 [26:20<8:07:56,  9.24s/it]INFO 04-12 16:24:46 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 134/3300 [26:29<8:07:32,  9.24s/it]2026-04-12 16:24:55,379 | INFO | train_grpo_train | metrics_logged mode=train step=134

                                                    
{'loss': 0.1667, 'grad_norm': 12.369037628173828, 'learning_rate': 9.596969696969699e-06, 'num_tokens': 257969.0, 'completions/mean_length': 45.375, 'completions/min_length': 30.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.37220871448516846, 'rewards/meter/std': 0.42236220836639404, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9910990595817566, 'rewards/lexical_plausibility/std': 0.02517561987042427, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9431490302085876, 'rewards/repeat_penalty/std': 0.08836318552494049, 'rewards/repeat_floor/mean': 0.9942848682403564, 'rewards/repeat_floor/std': 0.007260276470333338, 'rewards/near_duplicate_penalty/mean': 0.9743990302085876, 'rewards/near_duplicate_penalty/std': 0.04271058738231659, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13243892788887024, 'rewards/total_composite/std': 0.1523837298154831, 'reward': 0.13243892788887024, 'reward_std': 0.1523837149143219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21118704974651337, 'sampling/sampling_logp_difference/max': 1.2088298797607422, 'sampling/importance_sampling_ratio/min': 0.2985464036464691, 'sampling/importance_sampling_ratio/mean': 1.0220602750778198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.166370913386345, 'clip_ratio/low_mean': 0.12005884945392609, 'clip_ratio/low_min': 0.12005884945392609, 'clip_ratio/high_mean': 0.07563676871359348, 'clip_ratio/high_max': 0.07563676871359348, 'clip_ratio/region_mean': 0.19569561816751957, 'reward_total_mean': 0.13243892788887024, 'reward_meter_mean': 0.37220871448516846, 'reward_meter_std': 0.42236220836639404, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9910990595817566, 'reward_lexical_plausibility_std': 0.02517561987042427, 'reward_repeat_penalty_mean': 0.9431490302085876, 'reward_repeat_penalty_std': 0.08836318552494049, 'reward_repeat_floor_mean': 0.9942848682403564, 'reward_repeat_floor_std': 0.007260276470333338, 'reward_near_duplicate_penalty_mean': 0.9743990302085876, 'reward_near_duplicate_penalty_std': 0.04271058738231659, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.13243892788887024, 'reward_total_composite_std': 0.1523837298154831, 'epoch': 0.01}

  4%|▍         | 134/3300 [26:29<8:07:32,  9.24s/it]INFO 04-12 16:24:55 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 135/3300 [26:38<8:08:34,  9.26s/it]2026-04-12 16:25:04,693 | INFO | train_grpo_train | metrics_logged mode=train step=135

                                                    
{'loss': 0.0681, 'grad_norm': 11.468616485595703, 'learning_rate': 9.593939393939394e-06, 'num_tokens': 259663.0, 'completions/mean_length': 48.75, 'completions/min_length': 40.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.6985104084014893, 'rewards/meter/std': 0.3451581299304962, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6387500166893005, 'rewards/judge_quality/std': 0.2994489371776581, 'rewards/repeat_penalty/mean': 0.9809602499008179, 'rewards/repeat_penalty/std': 0.017124870792031288, 'rewards/repeat_floor/mean': 0.9971440434455872, 'rewards/repeat_floor/std': 0.00256872596219182, 'rewards/near_duplicate_penalty/mean': 0.9809602499008179, 'rewards/near_duplicate_penalty/std': 0.017124870792031288, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.510533332824707, 'rewards/total_composite/std': 0.35021281242370605, 'reward': 0.510533332824707, 'reward_std': 0.35021281242370605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17869427800178528, 'sampling/sampling_logp_difference/max': 1.879021167755127, 'sampling/importance_sampling_ratio/min': 0.1527395397424698, 'sampling/importance_sampling_ratio/mean': 1.0203851461410522, 'sampling/importance_sampling_ratio/max': 1.786812663078308, 'entropy': 1.7431702762842178, 'clip_ratio/low_mean': 0.0785891329869628, 'clip_ratio/low_min': 0.0785891329869628, 'clip_ratio/high_mean': 0.061967531219124794, 'clip_ratio/high_max': 0.061967531219124794, 'clip_ratio/region_mean': 0.1405566642060876, 'reward_total_mean': 0.510533332824707, 'reward_meter_mean': 0.6985104084014893, 'reward_meter_std': 0.3451581299304962, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9809602499008179, 'reward_repeat_penalty_std': 0.017124870792031288, 'reward_repeat_floor_mean': 0.9971440434455872, 'reward_repeat_floor_std': 0.00256872596219182, 'reward_near_duplicate_penalty_mean': 0.9809602499008179, 'reward_near_duplicate_penalty_std': 0.017124870792031288, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6387500166893005, 'reward_judge_quality_std': 0.2994489371776581, 'reward_total_composite_mean': 0.510533332824707, 'reward_total_composite_std': 0.35021281242370605, 'epoch': 0.01}

  4%|▍         | 135/3300 [26:38<8:08:34,  9.26s/it]INFO 04-12 16:25:04 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 136/3300 [26:52<9:22:36, 10.67s/it]2026-04-12 16:25:18,651 | INFO | train_grpo_train | metrics_logged mode=train step=136

                                                    
{'loss': 0.1192, 'grad_norm': 5.362037658691406, 'learning_rate': 9.590909090909091e-06, 'num_tokens': 262090.0, 'completions/mean_length': 199.375, 'completions/min_length': 88.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 154.71429443359375, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 403.0, 'rewards/meter/mean': 0.526153564453125, 'rewards/meter/std': 0.3998994529247284, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.38816189765930176, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.11644857376813889, 'rewards/lexical_plausibility/mean': 0.9769138693809509, 'rewards/lexical_plausibility/std': 0.041646022349596024, 'rewards/judge_quality/mean': 0.19374999403953552, 'rewards/judge_quality/std': 0.13999362289905548, 'rewards/repeat_penalty/mean': 0.9327306747436523, 'rewards/repeat_penalty/std': 0.09829148650169373, 'rewards/repeat_floor/mean': 0.9942677021026611, 'rewards/repeat_floor/std': 0.007390270940959454, 'rewards/near_duplicate_penalty/mean': 0.9916704893112183, 'rewards/near_duplicate_penalty/std': 0.01151853334158659, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9710232019424438, 'rewards/distinct_2/std': 0.04885513335466385, 'rewards/total_composite/mean': 0.11389347165822983, 'rewards/total_composite/std': 0.13071997463703156, 'reward': 0.11389347165822983, 'reward_std': 0.13071998953819275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23406687378883362, 'sampling/sampling_logp_difference/max': 1.371506690979004, 'sampling/importance_sampling_ratio/min': 0.2537243664264679, 'sampling/importance_sampling_ratio/mean': 1.0817837715148926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9461817145347595, 'clip_ratio/low_mean': 0.11688897851854563, 'clip_ratio/low_min': 0.11688897851854563, 'clip_ratio/high_mean': 0.046653544530272484, 'clip_ratio/high_max': 0.046653544530272484, 'clip_ratio/region_mean': 0.1635425230488181, 'reward_total_mean': 0.11389347165822983, 'reward_meter_mean': 0.526153564453125, 'reward_meter_std': 0.3998994529247284, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.38816189765930176, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.11644857376813889, 'reward_lexical_plausibility_mean': 0.9769138693809509, 'reward_lexical_plausibility_std': 0.041646022349596024, 'reward_repeat_penalty_mean': 0.9327306747436523, 'reward_repeat_penalty_std': 0.09829148650169373, 'reward_repeat_floor_mean': 0.9942677021026611, 'reward_repeat_floor_std': 0.007390270940959454, 'reward_near_duplicate_penalty_mean': 0.9916704893112183, 'reward_near_duplicate_penalty_std': 0.01151853334158659, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9710232019424438, 'reward_distinct_2_std': 0.04885513335466385, 'reward_judge_quality_mean': 0.19374999403953552, 'reward_judge_quality_std': 0.13999362289905548, 'reward_total_composite_mean': 0.11389347165822983, 'reward_total_composite_std': 0.13071997463703156, 'epoch': 0.01}

  4%|▍         | 136/3300 [26:52<9:22:36, 10.67s/it]INFO 04-12 16:25:18 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 137/3300 [27:06<10:16:47, 11.70s/it]2026-04-12 16:25:32,753 | INFO | train_grpo_train | metrics_logged mode=train step=137

                                                     
{'loss': 0.0377, 'grad_norm': 4.935665130615234, 'learning_rate': 9.587878787878789e-06, 'num_tokens': 265220.0, 'completions/mean_length': 260.25, 'completions/min_length': 164.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 224.2857208251953, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 306.0, 'rewards/meter/mean': 0.1846785694360733, 'rewards/meter/std': 0.26168984174728394, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.11080066114664078, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.033240195363759995, 'rewards/lexical_plausibility/mean': 0.9290204048156738, 'rewards/lexical_plausibility/std': 0.18416956067085266, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.09161254018545151, 'rewards/repeat_penalty/mean': 0.98967444896698, 'rewards/repeat_penalty/std': 0.008269112557172775, 'rewards/repeat_floor/mean': 0.9988254308700562, 'rewards/repeat_floor/std': 0.0007386037614196539, 'rewards/near_duplicate_penalty/mean': 0.9959423542022705, 'rewards/near_duplicate_penalty/std': 0.0018977454164996743, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9937118291854858, 'rewards/distinct_2/std': 0.008787008002400398, 'rewards/total_composite/mean': 0.036731235682964325, 'rewards/total_composite/std': 0.06452368944883347, 'reward': 0.036731235682964325, 'reward_std': 0.06452368944883347, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22951792180538177, 'sampling/sampling_logp_difference/max': 1.558380126953125, 'sampling/importance_sampling_ratio/min': 0.21047674119472504, 'sampling/importance_sampling_ratio/mean': 1.0506075620651245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0635197460651398, 'clip_ratio/low_mean': 0.10556838102638721, 'clip_ratio/low_min': 0.10556838102638721, 'clip_ratio/high_mean': 0.0488334558904171, 'clip_ratio/high_max': 0.0488334558904171, 'clip_ratio/region_mean': 0.1544018369168043, 'reward_total_mean': 0.036731235682964325, 'reward_meter_mean': 0.1846785694360733, 'reward_meter_std': 0.26168984174728394, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.11080066114664078, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.033240195363759995, 'reward_lexical_plausibility_mean': 0.9290204048156738, 'reward_lexical_plausibility_std': 0.18416956067085266, 'reward_repeat_penalty_mean': 0.98967444896698, 'reward_repeat_penalty_std': 0.008269112557172775, 'reward_repeat_floor_mean': 0.9988254308700562, 'reward_repeat_floor_std': 0.0007386037614196539, 'reward_near_duplicate_penalty_mean': 0.9959423542022705, 'reward_near_duplicate_penalty_std': 0.0018977454164996743, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9937118291854858, 'reward_distinct_2_std': 0.008787008002400398, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.09161254018545151, 'reward_total_composite_mean': 0.036731235682964325, 'reward_total_composite_std': 0.06452368944883347, 'epoch': 0.01}

  4%|▍         | 137/3300 [27:06<10:16:47, 11.70s/it]INFO 04-12 16:25:33 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 138/3300 [27:15<9:33:17, 10.88s/it] 2026-04-12 16:25:41,710 | INFO | train_grpo_train | metrics_logged mode=train step=138

                                                    
{'loss': 0.0286, 'grad_norm': 15.345233917236328, 'learning_rate': 9.584848484848486e-06, 'num_tokens': 266887.0, 'completions/mean_length': 35.375, 'completions/min_length': 32.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8309788703918457, 'rewards/meter/std': 0.29639732837677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9928355813026428, 'rewards/lexical_plausibility/std': 0.020263994112610817, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.995775580406189, 'rewards/repeat_penalty/std': 0.010963790118694305, 'rewards/repeat_floor/mean': 0.999366283416748, 'rewards/repeat_floor/std': 0.0016445706132799387, 'rewards/near_duplicate_penalty/mean': 0.995775580406189, 'rewards/near_duplicate_penalty/std': 0.010963790118694305, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.35061946511268616, 'rewards/total_composite/std': 0.13179932534694672, 'reward': 0.35061946511268616, 'reward_std': 0.13179932534694672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2372082620859146, 'sampling/sampling_logp_difference/max': 1.6083354949951172, 'sampling/importance_sampling_ratio/min': 0.20022061467170715, 'sampling/importance_sampling_ratio/mean': 1.0378944873809814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.649111896753311, 'clip_ratio/low_mean': 0.10363711230456829, 'clip_ratio/low_min': 0.10363711230456829, 'clip_ratio/high_mean': 0.12226535938680172, 'clip_ratio/high_max': 0.12226535938680172, 'clip_ratio/region_mean': 0.22590247169137, 'reward_total_mean': 0.35061946511268616, 'reward_meter_mean': 0.8309788703918457, 'reward_meter_std': 0.29639732837677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9928355813026428, 'reward_lexical_plausibility_std': 0.020263994112610817, 'reward_repeat_penalty_mean': 0.995775580406189, 'reward_repeat_penalty_std': 0.010963790118694305, 'reward_repeat_floor_mean': 0.999366283416748, 'reward_repeat_floor_std': 0.0016445706132799387, 'reward_near_duplicate_penalty_mean': 0.995775580406189, 'reward_near_duplicate_penalty_std': 0.010963790118694305, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.35061946511268616, 'reward_total_composite_std': 0.13179932534694672, 'epoch': 0.01}

  4%|▍         | 138/3300 [27:15<9:33:17, 10.88s/it]INFO 04-12 16:25:41 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 139/3300 [27:23<8:53:37, 10.13s/it]2026-04-12 16:25:50,090 | INFO | train_grpo_train | metrics_logged mode=train step=139

                                                    
{'loss': -0.0332, 'grad_norm': 23.95005226135254, 'learning_rate': 9.581818181818181e-06, 'num_tokens': 268312.0, 'completions/mean_length': 24.125, 'completions/min_length': 17.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.3752234876155853, 'rewards/meter/std': 0.4822108745574951, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 0.8075591325759888, 'rewards/lexical_plausibility/std': 0.22869481146335602, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3414648175239563, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.1157275140285492, 'rewards/repeat_floor/mean': 0.9887499809265137, 'rewards/repeat_floor/std': 0.0069436440244317055, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3305409252643585, 'rewards/total_composite/std': 0.4436834752559662, 'reward': 0.3305409252643585, 'reward_std': 0.4436834752559662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2235724776983261, 'sampling/sampling_logp_difference/max': 1.3785057067871094, 'sampling/importance_sampling_ratio/min': 0.3296750783920288, 'sampling/importance_sampling_ratio/mean': 1.0513975620269775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4279952496290207, 'clip_ratio/low_mean': 0.15241304971277714, 'clip_ratio/low_min': 0.15241304971277714, 'clip_ratio/high_mean': 0.08189449831843376, 'clip_ratio/high_max': 0.08189449831843376, 'clip_ratio/region_mean': 0.2343075480312109, 'reward_total_mean': 0.3305409252643585, 'reward_meter_mean': 0.3752234876155853, 'reward_meter_std': 0.4822108745574951, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 0.8075591325759888, 'reward_lexical_plausibility_std': 0.22869481146335602, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.1157275140285492, 'reward_repeat_floor_mean': 0.9887499809265137, 'reward_repeat_floor_std': 0.0069436440244317055, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3414648175239563, 'reward_total_composite_mean': 0.3305409252643585, 'reward_total_composite_std': 0.4436834752559662, 'epoch': 0.01}

  4%|▍         | 139/3300 [27:23<8:53:37, 10.13s/it]INFO 04-12 16:25:50 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 140/3300 [27:32<8:25:35,  9.60s/it]2026-04-12 16:25:58,455 | INFO | train_grpo_train | metrics_logged mode=train step=140

                                                    
{'loss': 0.0396, 'grad_norm': 14.158329010009766, 'learning_rate': 9.57878787878788e-06, 'num_tokens': 270035.0, 'completions/mean_length': 46.375, 'completions/min_length': 41.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6709508299827576, 'rewards/meter/std': 0.42303919792175293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9716345071792603, 'rewards/lexical_plausibility/std': 0.0802297294139862, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.268537700176239, 'rewards/repeat_penalty/mean': 0.9590060710906982, 'rewards/repeat_penalty/std': 0.10046610981225967, 'rewards/repeat_floor/mean': 0.9948601722717285, 'rewards/repeat_floor/std': 0.012219909578561783, 'rewards/near_duplicate_penalty/mean': 0.980157732963562, 'rewards/near_duplicate_penalty/std': 0.04083683714270592, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9759615659713745, 'rewards/distinct_2/std': 0.06799104809761047, 'rewards/total_composite/mean': 0.23097580671310425, 'rewards/total_composite/std': 0.2973477840423584, 'reward': 0.23097580671310425, 'reward_std': 0.2973477840423584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.240529403090477, 'sampling/sampling_logp_difference/max': 1.2370967864990234, 'sampling/importance_sampling_ratio/min': 0.2902255952358246, 'sampling/importance_sampling_ratio/mean': 1.0806950330734253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8457294404506683, 'clip_ratio/low_mean': 0.14311196468770504, 'clip_ratio/low_min': 0.14311196468770504, 'clip_ratio/high_mean': 0.034863551147282124, 'clip_ratio/high_max': 0.034863551147282124, 'clip_ratio/region_mean': 0.17797551583498716, 'reward_total_mean': 0.23097580671310425, 'reward_meter_mean': 0.6709508299827576, 'reward_meter_std': 0.42303919792175293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9716345071792603, 'reward_lexical_plausibility_std': 0.0802297294139862, 'reward_repeat_penalty_mean': 0.9590060710906982, 'reward_repeat_penalty_std': 0.10046610981225967, 'reward_repeat_floor_mean': 0.9948601722717285, 'reward_repeat_floor_std': 0.012219909578561783, 'reward_near_duplicate_penalty_mean': 0.980157732963562, 'reward_near_duplicate_penalty_std': 0.04083683714270592, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9759615659713745, 'reward_distinct_2_std': 0.06799104809761047, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.268537700176239, 'reward_total_composite_mean': 0.23097580671310425, 'reward_total_composite_std': 0.2973477840423584, 'epoch': 0.01}

  4%|▍         | 140/3300 [27:32<8:25:35,  9.60s/it]INFO 04-12 16:25:58 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 141/3300 [27:40<8:01:24,  9.14s/it]2026-04-12 16:26:06,535 | INFO | train_grpo_train | metrics_logged mode=train step=141

                                                    
{'loss': 0.1518, 'grad_norm': 13.101373672485352, 'learning_rate': 9.575757575757576e-06, 'num_tokens': 271662.0, 'completions/mean_length': 46.375, 'completions/min_length': 33.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.5712951421737671, 'rewards/meter/std': 0.41170933842658997, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9915022850036621, 'rewards/repeat_penalty/std': 0.014074904844164848, 'rewards/repeat_floor/mean': 0.9987253546714783, 'rewards/repeat_floor/std': 0.0021112330723553896, 'rewards/near_duplicate_penalty/mean': 0.9915022850036621, 'rewards/near_duplicate_penalty/std': 0.014074904844164848, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18343859910964966, 'rewards/total_composite/std': 0.14860112965106964, 'reward': 0.18343859910964966, 'reward_std': 0.14860112965106964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23140616714954376, 'sampling/sampling_logp_difference/max': 1.531193733215332, 'sampling/importance_sampling_ratio/min': 0.2162773311138153, 'sampling/importance_sampling_ratio/mean': 1.0546079874038696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.774710536003113, 'clip_ratio/low_mean': 0.09014561213552952, 'clip_ratio/low_min': 0.09014561213552952, 'clip_ratio/high_mean': 0.10207689367234707, 'clip_ratio/high_max': 0.10207689367234707, 'clip_ratio/region_mean': 0.1922225058078766, 'reward_total_mean': 0.18343859910964966, 'reward_meter_mean': 0.5712951421737671, 'reward_meter_std': 0.41170933842658997, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9915022850036621, 'reward_repeat_penalty_std': 0.014074904844164848, 'reward_repeat_floor_mean': 0.9987253546714783, 'reward_repeat_floor_std': 0.0021112330723553896, 'reward_near_duplicate_penalty_mean': 0.9915022850036621, 'reward_near_duplicate_penalty_std': 0.014074904844164848, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.18343859910964966, 'reward_total_composite_std': 0.14860112965106964, 'epoch': 0.01}

  4%|▍         | 141/3300 [27:40<8:01:24,  9.14s/it]INFO 04-12 16:26:06 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 142/3300 [27:49<7:58:01,  9.08s/it]2026-04-12 16:26:15,475 | INFO | train_grpo_train | metrics_logged mode=train step=142

                                                    
{'loss': -0.0063, 'grad_norm': 15.231253623962402, 'learning_rate': 9.572727272727273e-06, 'num_tokens': 273390.0, 'completions/mean_length': 44.0, 'completions/min_length': 36.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5280349254608154, 'rewards/meter/std': 0.41948628425598145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5587500333786011, 'rewards/judge_quality/std': 0.2158992886543274, 'rewards/repeat_penalty/mean': 0.9944033026695251, 'rewards/repeat_penalty/std': 0.0121182044968009, 'rewards/repeat_floor/mean': 0.9991604685783386, 'rewards/repeat_floor/std': 0.001817729091271758, 'rewards/near_duplicate_penalty/mean': 0.9944033026695251, 'rewards/near_duplicate_penalty/std': 0.0121182044968009, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2706297039985657, 'rewards/total_composite/std': 0.2414269894361496, 'reward': 0.2706297039985657, 'reward_std': 0.2414269894361496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21337202191352844, 'sampling/sampling_logp_difference/max': 1.4742069244384766, 'sampling/importance_sampling_ratio/min': 0.22896023094654083, 'sampling/importance_sampling_ratio/mean': 1.0421977043151855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.118324249982834, 'clip_ratio/low_mean': 0.10393263585865498, 'clip_ratio/low_min': 0.10393263585865498, 'clip_ratio/high_mean': 0.07075795065611601, 'clip_ratio/high_max': 0.07075795065611601, 'clip_ratio/region_mean': 0.17469058651477098, 'reward_total_mean': 0.2706297039985657, 'reward_meter_mean': 0.5280349254608154, 'reward_meter_std': 0.41948628425598145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9944033026695251, 'reward_repeat_penalty_std': 0.0121182044968009, 'reward_repeat_floor_mean': 0.9991604685783386, 'reward_repeat_floor_std': 0.001817729091271758, 'reward_near_duplicate_penalty_mean': 0.9944033026695251, 'reward_near_duplicate_penalty_std': 0.0121182044968009, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5587500333786011, 'reward_judge_quality_std': 0.2158992886543274, 'reward_total_composite_mean': 0.2706297039985657, 'reward_total_composite_std': 0.2414269894361496, 'epoch': 0.01}

  4%|▍         | 142/3300 [27:49<7:58:01,  9.08s/it]INFO 04-12 16:26:15 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 143/3300 [28:03<9:16:44, 10.58s/it]2026-04-12 16:26:29,552 | INFO | train_grpo_train | metrics_logged mode=train step=143

                                                    
{'loss': 0.3349, 'grad_norm': 7.965107440948486, 'learning_rate': 9.56969696969697e-06, 'num_tokens': 275990.0, 'completions/mean_length': 145.0, 'completions/min_length': 81.0, 'completions/max_length': 475.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 145.0, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 475.0, 'rewards/meter/mean': 0.12463633716106415, 'rewards/meter/std': 0.11814981698989868, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.05580177903175354, 'rewards/lexical_plausibility/mean': 0.9836630821228027, 'rewards/lexical_plausibility/std': 0.026990195736289024, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9874579906463623, 'rewards/repeat_penalty/std': 0.020340239629149437, 'rewards/repeat_floor/mean': 0.9984275698661804, 'rewards/repeat_floor/std': 0.00232161907479167, 'rewards/near_duplicate_penalty/mean': 0.992782711982727, 'rewards/near_duplicate_penalty/std': 0.00817783921957016, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9945576190948486, 'rewards/distinct_2/std': 0.013459553010761738, 'rewards/total_composite/mean': 0.052760906517505646, 'rewards/total_composite/std': 0.0543060302734375, 'reward': 0.052760906517505646, 'reward_std': 0.0543060265481472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24614021182060242, 'sampling/sampling_logp_difference/max': 1.5540504455566406, 'sampling/importance_sampling_ratio/min': 0.21139000356197357, 'sampling/importance_sampling_ratio/mean': 1.0575592517852783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.552743285894394, 'clip_ratio/low_mean': 0.11233380064368248, 'clip_ratio/low_min': 0.11233380064368248, 'clip_ratio/high_mean': 0.07718076556921005, 'clip_ratio/high_max': 0.07718076556921005, 'clip_ratio/region_mean': 0.18951456621289253, 'reward_total_mean': 0.052760906517505646, 'reward_meter_mean': 0.12463633716106415, 'reward_meter_std': 0.11814981698989868, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.05580177903175354, 'reward_lexical_plausibility_mean': 0.9836630821228027, 'reward_lexical_plausibility_std': 0.026990195736289024, 'reward_repeat_penalty_mean': 0.9874579906463623, 'reward_repeat_penalty_std': 0.020340239629149437, 'reward_repeat_floor_mean': 0.9984275698661804, 'reward_repeat_floor_std': 0.00232161907479167, 'reward_near_duplicate_penalty_mean': 0.992782711982727, 'reward_near_duplicate_penalty_std': 0.00817783921957016, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9945576190948486, 'reward_distinct_2_std': 0.013459553010761738, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.052760906517505646, 'reward_total_composite_std': 0.0543060302734375, 'epoch': 0.01}

  4%|▍         | 143/3300 [28:03<9:16:44, 10.58s/it]INFO 04-12 16:26:29 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 144/3300 [28:12<8:57:10, 10.21s/it]2026-04-12 16:26:38,904 | INFO | train_grpo_train | metrics_logged mode=train step=144

                                                    
{'loss': 0.1612, 'grad_norm': 12.269087791442871, 'learning_rate': 9.566666666666668e-06, 'num_tokens': 277709.0, 'completions/mean_length': 49.875, 'completions/min_length': 32.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.4163627624511719, 'rewards/meter/std': 0.40220409631729126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8012499809265137, 'rewards/judge_quality/std': 0.24948161840438843, 'rewards/repeat_penalty/mean': 0.9986408352851868, 'rewards/repeat_penalty/std': 0.0038443405646830797, 'rewards/repeat_floor/mean': 0.9997961521148682, 'rewards/repeat_floor/std': 0.0005766532267443836, 'rewards/near_duplicate_penalty/mean': 0.9986408352851868, 'rewards/near_duplicate_penalty/std': 0.0038443405646830797, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32693901658058167, 'rewards/total_composite/std': 0.36526110768318176, 'reward': 0.32693901658058167, 'reward_std': 0.36526110768318176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16840322315692902, 'sampling/sampling_logp_difference/max': 1.1883630752563477, 'sampling/importance_sampling_ratio/min': 0.3047196567058563, 'sampling/importance_sampling_ratio/mean': 1.043142557144165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7192247211933136, 'clip_ratio/low_mean': 0.1265910305082798, 'clip_ratio/low_min': 0.1265910305082798, 'clip_ratio/high_mean': 0.017189984442666173, 'clip_ratio/high_max': 0.017189984442666173, 'clip_ratio/region_mean': 0.14378101495094597, 'reward_total_mean': 0.32693901658058167, 'reward_meter_mean': 0.4163627624511719, 'reward_meter_std': 0.40220409631729126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9986408352851868, 'reward_repeat_penalty_std': 0.0038443405646830797, 'reward_repeat_floor_mean': 0.9997961521148682, 'reward_repeat_floor_std': 0.0005766532267443836, 'reward_near_duplicate_penalty_mean': 0.9986408352851868, 'reward_near_duplicate_penalty_std': 0.0038443405646830797, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8012499809265137, 'reward_judge_quality_std': 0.24948161840438843, 'reward_total_composite_mean': 0.32693901658058167, 'reward_total_composite_std': 0.36526110768318176, 'epoch': 0.01}

  4%|▍         | 144/3300 [28:12<8:57:10, 10.21s/it]INFO 04-12 16:26:39 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 145/3300 [28:22<8:52:38, 10.13s/it]2026-04-12 16:26:48,841 | INFO | train_grpo_train | metrics_logged mode=train step=145

                                                    
{'loss': 0.1939, 'grad_norm': 8.608485221862793, 'learning_rate': 9.563636363636365e-06, 'num_tokens': 279792.0, 'completions/mean_length': 97.375, 'completions/min_length': 56.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.08041495084762573, 'rewards/meter/std': 0.09178543090820312, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 0.8635621070861816, 'rewards/lexical_plausibility/std': 0.13445277512073517, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9962135553359985, 'rewards/repeat_penalty/std': 0.006941632833331823, 'rewards/repeat_floor/mean': 0.9994320273399353, 'rewards/repeat_floor/std': 0.0010412477422505617, 'rewards/near_duplicate_penalty/mean': 0.9962135553359985, 'rewards/near_duplicate_penalty/std': 0.006941632833331823, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.022487379610538483, 'rewards/total_composite/std': 0.03471127897500992, 'reward': 0.022487379610538483, 'reward_std': 0.03471127897500992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2191181629896164, 'sampling/sampling_logp_difference/max': 1.4294586181640625, 'sampling/importance_sampling_ratio/min': 0.2394385188817978, 'sampling/importance_sampling_ratio/mean': 1.040704369544983, 'sampling/importance_sampling_ratio/max': 1.9703952074050903, 'entropy': 2.8126401007175446, 'clip_ratio/low_mean': 0.1424967236816883, 'clip_ratio/low_min': 0.1424967236816883, 'clip_ratio/high_mean': 0.04397551529109478, 'clip_ratio/high_max': 0.04397551529109478, 'clip_ratio/region_mean': 0.1864722389727831, 'reward_total_mean': 0.022487379610538483, 'reward_meter_mean': 0.08041495084762573, 'reward_meter_std': 0.09178543090820312, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 0.8635621070861816, 'reward_lexical_plausibility_std': 0.13445277512073517, 'reward_repeat_penalty_mean': 0.9962135553359985, 'reward_repeat_penalty_std': 0.006941632833331823, 'reward_repeat_floor_mean': 0.9994320273399353, 'reward_repeat_floor_std': 0.0010412477422505617, 'reward_near_duplicate_penalty_mean': 0.9962135553359985, 'reward_near_duplicate_penalty_std': 0.006941632833331823, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.022487379610538483, 'reward_total_composite_std': 0.03471127897500992, 'epoch': 0.01}

  4%|▍         | 145/3300 [28:22<8:52:38, 10.13s/it]INFO 04-12 16:26:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 146/3300 [28:40<10:55:28, 12.47s/it]2026-04-12 16:27:06,770 | INFO | train_grpo_train | metrics_logged mode=train step=146

                                                     
{'loss': 0.1073, 'grad_norm': 3.361783266067505, 'learning_rate': 9.56060606060606e-06, 'num_tokens': 281931.0, 'completions/mean_length': 223.375, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 127.16667175292969, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 449.0, 'rewards/meter/mean': 0.7488503456115723, 'rewards/meter/std': 0.3192792236804962, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.8579171299934387, 'rewards/lexical_plausibility/std': 0.2528241276741028, 'rewards/judge_quality/mean': 0.18125000596046448, 'rewards/judge_quality/std': 0.14865468442440033, 'rewards/repeat_penalty/mean': 0.9988870620727539, 'rewards/repeat_penalty/std': 0.0026286349166184664, 'rewards/repeat_floor/mean': 0.9998893737792969, 'rewards/repeat_floor/std': 0.0002397261851001531, 'rewards/near_duplicate_penalty/mean': 0.9998251795768738, 'rewards/near_duplicate_penalty/std': 0.000494487932883203, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9990619421005249, 'rewards/distinct_2/std': 0.0026533128693699837, 'rewards/total_composite/mean': 0.13748405873775482, 'rewards/total_composite/std': 0.1326868236064911, 'reward': 0.13748405873775482, 'reward_std': 0.1326868087053299, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2458665817975998, 'sampling/sampling_logp_difference/max': 1.2333660125732422, 'sampling/importance_sampling_ratio/min': 0.2913103699684143, 'sampling/importance_sampling_ratio/mean': 1.0718419551849365, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9780437648296356, 'clip_ratio/low_mean': 0.0679066926240921, 'clip_ratio/low_min': 0.0679066926240921, 'clip_ratio/high_mean': 0.0773675087839365, 'clip_ratio/high_max': 0.0773675087839365, 'clip_ratio/region_mean': 0.1452742014080286, 'reward_total_mean': 0.13748405873775482, 'reward_meter_mean': 0.7488503456115723, 'reward_meter_std': 0.3192792236804962, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.8579171299934387, 'reward_lexical_plausibility_std': 0.2528241276741028, 'reward_repeat_penalty_mean': 0.9988870620727539, 'reward_repeat_penalty_std': 0.0026286349166184664, 'reward_repeat_floor_mean': 0.9998893737792969, 'reward_repeat_floor_std': 0.0002397261851001531, 'reward_near_duplicate_penalty_mean': 0.9998251795768738, 'reward_near_duplicate_penalty_std': 0.000494487932883203, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9990619421005249, 'reward_distinct_2_std': 0.0026533128693699837, 'reward_judge_quality_mean': 0.18125000596046448, 'reward_judge_quality_std': 0.14865468442440033, 'reward_total_composite_mean': 0.13748405873775482, 'reward_total_composite_std': 0.1326868236064911, 'epoch': 0.01}

  4%|▍         | 146/3300 [28:40<10:55:28, 12.47s/it]INFO 04-12 16:27:07 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 147/3300 [28:55<11:27:58, 13.09s/it]2026-04-12 16:27:21,314 | INFO | train_grpo_train | metrics_logged mode=train step=147

                                                     
{'loss': 0.0819, 'grad_norm': 10.808486938476562, 'learning_rate': 9.55757575757576e-06, 'num_tokens': 284258.0, 'completions/mean_length': 120.875, 'completions/min_length': 104.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.875, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.3757374882698059, 'rewards/meter/std': 0.25155457854270935, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9784359931945801, 'rewards/repeat_penalty/std': 0.021663790568709373, 'rewards/repeat_floor/mean': 0.9973344802856445, 'rewards/repeat_floor/std': 0.002300889464095235, 'rewards/near_duplicate_penalty/mean': 0.9881511926651001, 'rewards/near_duplicate_penalty/std': 0.008078720420598984, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9901315569877625, 'rewards/distinct_2/std': 0.018280785530805588, 'rewards/total_composite/mean': 0.1059810072183609, 'rewards/total_composite/std': 0.07456544041633606, 'reward': 0.1059810072183609, 'reward_std': 0.07456544041633606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22779150307178497, 'sampling/sampling_logp_difference/max': 2.154332160949707, 'sampling/importance_sampling_ratio/min': 0.1159806177020073, 'sampling/importance_sampling_ratio/mean': 1.0102511644363403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5819136500358582, 'clip_ratio/low_mean': 0.10269808769226074, 'clip_ratio/low_min': 0.10269808769226074, 'clip_ratio/high_mean': 0.10698944143950939, 'clip_ratio/high_max': 0.10698944143950939, 'clip_ratio/region_mean': 0.20968752913177013, 'reward_total_mean': 0.1059810072183609, 'reward_meter_mean': 0.3757374882698059, 'reward_meter_std': 0.25155457854270935, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.9784359931945801, 'reward_repeat_penalty_std': 0.021663790568709373, 'reward_repeat_floor_mean': 0.9973344802856445, 'reward_repeat_floor_std': 0.002300889464095235, 'reward_near_duplicate_penalty_mean': 0.9881511926651001, 'reward_near_duplicate_penalty_std': 0.008078720420598984, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9901315569877625, 'reward_distinct_2_std': 0.018280785530805588, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.1059810072183609, 'reward_total_composite_std': 0.07456544041633606, 'epoch': 0.01}

  4%|▍         | 147/3300 [28:55<11:27:58, 13.09s/it]INFO 04-12 16:27:21 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 148/3300 [29:08<11:38:21, 13.29s/it]2026-04-12 16:27:35,080 | INFO | train_grpo_train | metrics_logged mode=train step=148

                                                     
{'loss': -0.1492, 'grad_norm': 3.1610093116760254, 'learning_rate': 9.554545454545455e-06, 'num_tokens': 287190.0, 'completions/mean_length': 390.5, 'completions/min_length': 173.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 269.0, 'completions/min_terminated_length': 173.0, 'completions/max_terminated_length': 348.0, 'rewards/meter/mean': 0.4640042781829834, 'rewards/meter/std': 0.4690435528755188, 'rewards/count_adherence/mean': 0.4027777910232544, 'rewards/count_adherence/std': 0.33564019203186035, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8208333253860474, 'rewards/count_floor/std': 0.10069204121828079, 'rewards/lexical_plausibility/mean': 0.5631389617919922, 'rewards/lexical_plausibility/std': 0.2559228241443634, 'rewards/judge_quality/mean': 0.07500000298023224, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.9368826150894165, 'rewards/repeat_penalty/std': 0.11535022407770157, 'rewards/repeat_floor/mean': 0.9961574077606201, 'rewards/repeat_floor/std': 0.006887897849082947, 'rewards/near_duplicate_penalty/mean': 0.9993826150894165, 'rewards/near_duplicate_penalty/std': 0.0010040045017376542, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.04006575420498848, 'rewards/total_composite/std': 0.05277764052152634, 'reward': 0.04006575420498848, 'reward_std': 0.05277763679623604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21735039353370667, 'sampling/sampling_logp_difference/max': 1.3076105117797852, 'sampling/importance_sampling_ratio/min': 0.27046555280685425, 'sampling/importance_sampling_ratio/mean': 1.0642019510269165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2763786613941193, 'clip_ratio/low_mean': 0.03180444426834583, 'clip_ratio/low_min': 0.03180444426834583, 'clip_ratio/high_mean': 0.046840742230415344, 'clip_ratio/high_max': 0.046840742230415344, 'clip_ratio/region_mean': 0.07864518649876118, 'reward_total_mean': 0.04006575420498848, 'reward_meter_mean': 0.4640042781829834, 'reward_meter_std': 0.4690435528755188, 'reward_count_adherence_mean': 0.4027777910232544, 'reward_count_adherence_std': 0.33564019203186035, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8208333253860474, 'reward_count_floor_std': 0.10069204121828079, 'reward_lexical_plausibility_mean': 0.5631389617919922, 'reward_lexical_plausibility_std': 0.2559228241443634, 'reward_repeat_penalty_mean': 0.9368826150894165, 'reward_repeat_penalty_std': 0.11535022407770157, 'reward_repeat_floor_mean': 0.9961574077606201, 'reward_repeat_floor_std': 0.006887897849082947, 'reward_near_duplicate_penalty_mean': 0.9993826150894165, 'reward_near_duplicate_penalty_std': 0.0010040045017376542, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.07500000298023224, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.04006575420498848, 'reward_total_composite_std': 0.05277764052152634, 'epoch': 0.01}

  4%|▍         | 148/3300 [29:08<11:38:21, 13.29s/it]INFO 04-12 16:27:35 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 149/3300 [29:24<12:09:02, 13.88s/it]2026-04-12 16:27:50,334 | INFO | train_grpo_train | metrics_logged mode=train step=149

                                                     
{'loss': -0.0507, 'grad_norm': 2.1130027770996094, 'learning_rate': 9.551515151515152e-06, 'num_tokens': 290328.0, 'completions/mean_length': 439.25, 'completions/min_length': 171.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 366.5, 'completions/min_terminated_length': 171.0, 'completions/max_terminated_length': 494.0, 'rewards/meter/mean': 0.4967094957828522, 'rewards/meter/std': 0.3946586847305298, 'rewards/count_adherence/mean': 0.4583333730697632, 'rewards/count_adherence/std': 0.19187648594379425, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8374999761581421, 'rewards/count_floor/std': 0.05756295099854469, 'rewards/lexical_plausibility/mean': 0.6219012141227722, 'rewards/lexical_plausibility/std': 0.23513641953468323, 'rewards/judge_quality/mean': 0.0625, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9974699020385742, 'rewards/repeat_penalty/std': 0.003077274886891246, 'rewards/repeat_floor/mean': 0.9996856451034546, 'rewards/repeat_floor/std': 0.00032589014153927565, 'rewards/near_duplicate_penalty/mean': 0.9985562562942505, 'rewards/near_duplicate_penalty/std': 0.0018052110681310296, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9989135265350342, 'rewards/distinct_2/std': 0.003073032246902585, 'rewards/total_composite/mean': 0.03064427338540554, 'rewards/total_composite/std': 0.03845704346895218, 'reward': 0.03064427338540554, 'reward_std': 0.03845703974366188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23426313698291779, 'sampling/sampling_logp_difference/max': 1.581960678100586, 'sampling/importance_sampling_ratio/min': 0.20557163655757904, 'sampling/importance_sampling_ratio/mean': 1.0527914762496948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4563476741313934, 'clip_ratio/low_mean': 0.028711112216114998, 'clip_ratio/low_min': 0.028711112216114998, 'clip_ratio/high_mean': 0.04554655961692333, 'clip_ratio/high_max': 0.04554655961692333, 'clip_ratio/region_mean': 0.07425767183303833, 'reward_total_mean': 0.03064427338540554, 'reward_meter_mean': 0.4967094957828522, 'reward_meter_std': 0.3946586847305298, 'reward_count_adherence_mean': 0.4583333730697632, 'reward_count_adherence_std': 0.19187648594379425, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8374999761581421, 'reward_count_floor_std': 0.05756295099854469, 'reward_lexical_plausibility_mean': 0.6219012141227722, 'reward_lexical_plausibility_std': 0.23513641953468323, 'reward_repeat_penalty_mean': 0.9974699020385742, 'reward_repeat_penalty_std': 0.003077274886891246, 'reward_repeat_floor_mean': 0.9996856451034546, 'reward_repeat_floor_std': 0.00032589014153927565, 'reward_near_duplicate_penalty_mean': 0.9985562562942505, 'reward_near_duplicate_penalty_std': 0.0018052110681310296, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9989135265350342, 'reward_distinct_2_std': 0.003073032246902585, 'reward_judge_quality_mean': 0.0625, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.03064427338540554, 'reward_total_composite_std': 0.03845704346895218, 'epoch': 0.01}

  5%|▍         | 149/3300 [29:24<12:09:02, 13.88s/it]INFO 04-12 16:27:50 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 150/3300 [29:32<10:37:54, 12.15s/it]2026-04-12 16:27:58,445 | INFO | train_grpo_train | metrics_logged mode=train step=150

                                                     
{'loss': 0.0257, 'grad_norm': 13.331134796142578, 'learning_rate': 9.54848484848485e-06, 'num_tokens': 291954.0, 'completions/mean_length': 39.25, 'completions/min_length': 26.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.14966915547847748, 'rewards/meter/std': 0.2939712703227997, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.3624691963195801, 'rewards/repeat_penalty/mean': 0.996006190776825, 'rewards/repeat_penalty/std': 0.007941397838294506, 'rewards/repeat_floor/mean': 0.9994009137153625, 'rewards/repeat_floor/std': 0.001191212679259479, 'rewards/near_duplicate_penalty/mean': 0.996006190776825, 'rewards/near_duplicate_penalty/std': 0.007941397838294506, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08787254989147186, 'rewards/total_composite/std': 0.14439496397972107, 'reward': 0.08787254989147186, 'reward_std': 0.14439494907855988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18074624240398407, 'sampling/sampling_logp_difference/max': 1.9926366806030273, 'sampling/importance_sampling_ratio/min': 0.13633549213409424, 'sampling/importance_sampling_ratio/mean': 1.0147531032562256, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.20120470225811, 'clip_ratio/low_mean': 0.135904380120337, 'clip_ratio/low_min': 0.135904380120337, 'clip_ratio/high_mean': 0.041025642305612564, 'clip_ratio/high_max': 0.041025642305612564, 'clip_ratio/region_mean': 0.17693002242594957, 'reward_total_mean': 0.08787254989147186, 'reward_meter_mean': 0.14966915547847748, 'reward_meter_std': 0.2939712703227997, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.996006190776825, 'reward_repeat_penalty_std': 0.007941397838294506, 'reward_repeat_floor_mean': 0.9994009137153625, 'reward_repeat_floor_std': 0.001191212679259479, 'reward_near_duplicate_penalty_mean': 0.996006190776825, 'reward_near_duplicate_penalty_std': 0.007941397838294506, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.3624691963195801, 'reward_total_composite_mean': 0.08787254989147186, 'reward_total_composite_std': 0.14439496397972107, 'epoch': 0.01}

  5%|▍         | 150/3300 [29:32<10:37:54, 12.15s/it]INFO 04-12 16:27:58 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:02,  5.68s/it]

 23%|██▎       | 3/13 [00:22<01:19,  7.94s/it]

 31%|███       | 4/13 [00:34<01:25,  9.55s/it]

 38%|███▊      | 5/13 [00:46<01:22, 10.34s/it]

 46%|████▌     | 6/13 [00:57<01:14, 10.62s/it]

 54%|█████▍    | 7/13 [01:08<01:04, 10.71s/it]

 62%|██████▏   | 8/13 [01:20<00:55, 11.08s/it]

 69%|██████▉   | 9/13 [01:31<00:44, 11.06s/it]

 77%|███████▋  | 10/13 [01:42<00:33, 11.01s/it]

 85%|████████▍ | 11/13 [01:53<00:22, 11.18s/it]

 92%|█████████▏| 12/13 [02:04<00:11, 11.11s/it]

100%|██████████| 13/13 [02:16<00:00, 11.38s/it]2026-04-12 16:30:27,318 | INFO | train_grpo_train | metrics_logged mode=eval step=150

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 148.8576, 'eval_samples_per_second': 0.699, 'eval_steps_per_second': 0.087, 'eval_num_tokens': 291954.0, 'eval_completions/mean_length': 231.90384615384616, 'eval_completions/min_length': 47.84615384615385, 'eval_completions/max_length': 511.9230769230769, 'eval_completions/clipped_ratio': 0.19230769230769232, 'eval_completions/mean_terminated_length': 163.97528017484225, 'eval_completions/min_terminated_length': 47.84615384615385, 'eval_completions/max_terminated_length': 358.3076923076923, 'eval_rewards/meter/mean': 0.2257346401994045, 'eval_rewards/meter/std': 0.3091386625399956, 'eval_rewards/count_adherence/mean': 0.7586572468280792, 'eval_rewards/count_adherence/std': 0.27057631944234556, 'eval_rewards/arabic_clean/mean': 0.7115384615384616, 'eval_rewards/arabic_clean/std': 0.4383068313965431, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9927884569534888, 'eval_rewards/arabic_floor/std': 0.020397308354194347, 'eval_rewards/count_floor/mean': 0.9275971605227544, 'eval_rewards/count_floor/std': 0.08117289626254486, 'eval_rewards/lexical_plausibility/mean': 0.8448186241663419, 'eval_rewards/lexical_plausibility/std': 0.20366024054013765, 'eval_rewards/judge_quality/mean': 0.23086538853553626, 'eval_rewards/judge_quality/std': 0.171858167132506, 'eval_rewards/repeat_penalty/mean': 0.9710512023705703, 'eval_rewards/repeat_penalty/std': 0.05339848382685047, 'eval_rewards/repeat_floor/mean': 0.9975112263972943, 'eval_rewards/repeat_floor/std': 0.0038020770656518065, 'eval_rewards/near_duplicate_penalty/mean': 0.9936564977352436, 'eval_rewards/near_duplicate_penalty/std': 0.009320403108946405, 'eval_rewards/opening_diversity/mean': 0.9831730769230769, 'eval_rewards/opening_diversity/std': 0.04289763420820236, 'eval_rewards/distinct_2/mean': 0.9941373284046466, 'eval_rewards/distinct_2/std': 0.012226779443713335, 'eval_rewards/total_composite/mean': 0.05975959184937752, 'eval_rewards/total_composite/std': 0.10096489437497579, 'eval_reward': 0.05975959184937752, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.18483873972525963, 'eval_sampling/sampling_logp_difference/max': 1.2555984350351186, 'eval_sampling/importance_sampling_ratio/min': 0.2868260466135465, 'eval_sampling/importance_sampling_ratio/mean': 1.0590465068817139, 'eval_sampling/importance_sampling_ratio/max': 1.7238773290927594, 'eval_entropy': 3.7728551167708178, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.05975959184937752, 'eval_reward_meter_mean': 0.2257346401994045, 'eval_reward_meter_std': 0.3091386625399956, 'eval_reward_count_adherence_mean': 0.7586572468280792, 'eval_reward_count_adherence_std': 0.27057631944234556, 'eval_reward_arabic_clean_mean': 0.7115384615384616, 'eval_reward_arabic_clean_std': 0.4383068313965431, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9927884569534888, 'eval_reward_arabic_floor_std': 0.020397308354194347, 'eval_reward_count_floor_mean': 0.9275971605227544, 'eval_reward_count_floor_std': 0.08117289626254486, 'eval_reward_lexical_plausibility_mean': 0.8448186241663419, 'eval_reward_lexical_plausibility_std': 0.20366024054013765, 'eval_reward_repeat_penalty_mean': 0.9710512023705703, 'eval_reward_repeat_penalty_std': 0.05339848382685047, 'eval_reward_repeat_floor_mean': 0.9975112263972943, 'eval_reward_repeat_floor_std': 0.0038020770656518065, 'eval_reward_near_duplicate_penalty_mean': 0.9936564977352436, 'eval_reward_near_duplicate_penalty_std': 0.009320403108946405, 'eval_reward_opening_diversity_mean': 0.9831730769230769, 'eval_reward_opening_diversity_std': 0.04289763420820236, 'eval_reward_distinct_2_mean': 0.9941373284046466, 'eval_reward_distinct_2_std': 0.012226779443713335, 'eval_reward_judge_quality_mean': 0.23086538853553626, 'eval_reward_judge_quality_std': 0.171858167132506, 'eval_reward_total_composite_mean': 0.05975959184937752, 'eval_reward_total_composite_std': 0.10096489437497579, 'epoch': 0.01}

  5%|▍         | 150/3300 [32:01<10:37:54, 12.15s/it]

100%|██████████| 13/13 [02:16<00:00, 11.38s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  5%|▍         | 151/3300 [32:11<49:09:32, 56.20s/it]2026-04-12 16:30:37,425 | INFO | train_grpo_train | metrics_logged mode=train step=151

                                                     
{'loss': 0.1278, 'grad_norm': 15.257999420166016, 'learning_rate': 9.545454545454547e-06, 'num_tokens': 293270.0, 'completions/mean_length': 27.5, 'completions/min_length': 21.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.15804384648799896, 'rewards/meter/std': 0.2840515971183777, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.3620477616786957, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1102190762758255, 'rewards/total_composite/std': 0.2489222139120102, 'reward': 0.1102190762758255, 'reward_std': 0.2489222139120102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2210453897714615, 'sampling/sampling_logp_difference/max': 1.2390928268432617, 'sampling/importance_sampling_ratio/min': 0.28964686393737793, 'sampling/importance_sampling_ratio/mean': 1.0759471654891968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3047438114881516, 'clip_ratio/low_mean': 0.14588989783078432, 'clip_ratio/low_min': 0.14588989783078432, 'clip_ratio/high_mean': 0.02976190485060215, 'clip_ratio/high_max': 0.02976190485060215, 'clip_ratio/region_mean': 0.17565180268138647, 'reward_total_mean': 0.1102190762758255, 'reward_meter_mean': 0.15804384648799896, 'reward_meter_std': 0.2840515971183777, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.3620477616786957, 'reward_total_composite_mean': 0.1102190762758255, 'reward_total_composite_std': 0.2489222139120102, 'epoch': 0.01}

  5%|▍         | 151/3300 [32:11<49:09:32, 56.20s/it]INFO 04-12 16:30:37 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 152/3300 [32:20<36:49:44, 42.12s/it]2026-04-12 16:30:46,684 | INFO | train_grpo_train | metrics_logged mode=train step=152

                                                     
{'loss': 0.0611, 'grad_norm': 14.929104804992676, 'learning_rate': 9.542424242424242e-06, 'num_tokens': 295050.0, 'completions/mean_length': 61.5, 'completions/min_length': 44.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.38935643434524536, 'rewards/meter/std': 0.29865187406539917, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9754464626312256, 'rewards/lexical_plausibility/std': 0.06944799423217773, 'rewards/judge_quality/mean': 0.5712500214576721, 'rewards/judge_quality/std': 0.2811424434185028, 'rewards/repeat_penalty/mean': 0.9804328680038452, 'rewards/repeat_penalty/std': 0.0327908992767334, 'rewards/repeat_floor/mean': 0.9973775744438171, 'rewards/repeat_floor/std': 0.004061189945787191, 'rewards/near_duplicate_penalty/mean': 0.9861229658126831, 'rewards/near_duplicate_penalty/std': 0.01746893674135208, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.993990421295166, 'rewards/distinct_2/std': 0.01699776202440262, 'rewards/total_composite/mean': 0.1978502869606018, 'rewards/total_composite/std': 0.1373213827610016, 'reward': 0.1978502869606018, 'reward_std': 0.1373213678598404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2262590527534485, 'sampling/sampling_logp_difference/max': 1.6137351989746094, 'sampling/importance_sampling_ratio/min': 0.19914238154888153, 'sampling/importance_sampling_ratio/mean': 1.0549683570861816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5685294568538666, 'clip_ratio/low_mean': 0.12222131341695786, 'clip_ratio/low_min': 0.12222131341695786, 'clip_ratio/high_mean': 0.09842611104249954, 'clip_ratio/high_max': 0.09842611104249954, 'clip_ratio/region_mean': 0.2206474244594574, 'reward_total_mean': 0.1978502869606018, 'reward_meter_mean': 0.38935643434524536, 'reward_meter_std': 0.29865187406539917, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9754464626312256, 'reward_lexical_plausibility_std': 0.06944799423217773, 'reward_repeat_penalty_mean': 0.9804328680038452, 'reward_repeat_penalty_std': 0.0327908992767334, 'reward_repeat_floor_mean': 0.9973775744438171, 'reward_repeat_floor_std': 0.004061189945787191, 'reward_near_duplicate_penalty_mean': 0.9861229658126831, 'reward_near_duplicate_penalty_std': 0.01746893674135208, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.993990421295166, 'reward_distinct_2_std': 0.01699776202440262, 'reward_judge_quality_mean': 0.5712500214576721, 'reward_judge_quality_std': 0.2811424434185028, 'reward_total_composite_mean': 0.1978502869606018, 'reward_total_composite_std': 0.1373213827610016, 'epoch': 0.01}

  5%|▍         | 152/3300 [32:20<36:49:44, 42.12s/it]INFO 04-12 16:30:46 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 153/3300 [32:29<28:09:50, 32.22s/it]2026-04-12 16:30:55,806 | INFO | train_grpo_train | metrics_logged mode=train step=153

                                                     
{'loss': 0.283, 'grad_norm': 16.348724365234375, 'learning_rate': 9.539393939393941e-06, 'num_tokens': 297032.0, 'completions/mean_length': 75.75, 'completions/min_length': 43.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.29824861884117126, 'rewards/meter/std': 0.29439404606819153, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9386405944824219, 'rewards/lexical_plausibility/std': 0.09032656252384186, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.32936087250709534, 'rewards/repeat_penalty/mean': 0.9406213760375977, 'rewards/repeat_penalty/std': 0.1538749784231186, 'rewards/repeat_floor/mean': 0.9931595325469971, 'rewards/repeat_floor/std': 0.017247755080461502, 'rewards/near_duplicate_penalty/mean': 0.982021689414978, 'rewards/near_duplicate_penalty/std': 0.03746762499213219, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9622926115989685, 'rewards/distinct_2/std': 0.10665260255336761, 'rewards/total_composite/mean': 0.1481320708990097, 'rewards/total_composite/std': 0.187057763338089, 'reward': 0.1481320708990097, 'reward_std': 0.187057763338089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2677787244319916, 'sampling/sampling_logp_difference/max': 2.189413070678711, 'sampling/importance_sampling_ratio/min': 0.11198245733976364, 'sampling/importance_sampling_ratio/mean': 1.0200051069259644, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.132720023393631, 'clip_ratio/low_mean': 0.1375386230647564, 'clip_ratio/low_min': 0.1375386230647564, 'clip_ratio/high_mean': 0.09088260866701603, 'clip_ratio/high_max': 0.09088260866701603, 'clip_ratio/region_mean': 0.22842123173177242, 'reward_total_mean': 0.1481320708990097, 'reward_meter_mean': 0.29824861884117126, 'reward_meter_std': 0.29439404606819153, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9386405944824219, 'reward_lexical_plausibility_std': 0.09032656252384186, 'reward_repeat_penalty_mean': 0.9406213760375977, 'reward_repeat_penalty_std': 0.1538749784231186, 'reward_repeat_floor_mean': 0.9931595325469971, 'reward_repeat_floor_std': 0.017247755080461502, 'reward_near_duplicate_penalty_mean': 0.982021689414978, 'reward_near_duplicate_penalty_std': 0.03746762499213219, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9622926115989685, 'reward_distinct_2_std': 0.10665260255336761, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.32936087250709534, 'reward_total_composite_mean': 0.1481320708990097, 'reward_total_composite_std': 0.187057763338089, 'epoch': 0.01}

  5%|▍         | 153/3300 [32:29<28:09:50, 32.22s/it]INFO 04-12 16:30:56 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 154/3300 [32:38<22:08:07, 25.33s/it]2026-04-12 16:31:05,060 | INFO | train_grpo_train | metrics_logged mode=train step=154

                                                     
{'loss': 0.1705, 'grad_norm': 14.417436599731445, 'learning_rate': 9.536363636363637e-06, 'num_tokens': 298848.0, 'completions/mean_length': 44.0, 'completions/min_length': 32.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.3378413915634155, 'rewards/meter/std': 0.40106305480003357, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390502154827118, 'rewards/repeat_penalty/mean': 0.9758144617080688, 'rewards/repeat_penalty/std': 0.03136686980724335, 'rewards/repeat_floor/mean': 0.9969303607940674, 'rewards/repeat_floor/std': 0.003498316276818514, 'rewards/near_duplicate_penalty/mean': 0.9853047132492065, 'rewards/near_duplicate_penalty/std': 0.017314862459897995, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.20318013429641724, 'rewards/total_composite/std': 0.3137761652469635, 'reward': 0.20318013429641724, 'reward_std': 0.3137761652469635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22353847324848175, 'sampling/sampling_logp_difference/max': 1.8340167999267578, 'sampling/importance_sampling_ratio/min': 0.15977051854133606, 'sampling/importance_sampling_ratio/mean': 1.045520544052124, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.528806522488594, 'clip_ratio/low_mean': 0.1583808809518814, 'clip_ratio/low_min': 0.1583808809518814, 'clip_ratio/high_mean': 0.060891544446349144, 'clip_ratio/high_max': 0.060891544446349144, 'clip_ratio/region_mean': 0.21927242539823055, 'reward_total_mean': 0.20318013429641724, 'reward_meter_mean': 0.3378413915634155, 'reward_meter_std': 0.40106305480003357, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9758144617080688, 'reward_repeat_penalty_std': 0.03136686980724335, 'reward_repeat_floor_mean': 0.9969303607940674, 'reward_repeat_floor_std': 0.003498316276818514, 'reward_near_duplicate_penalty_mean': 0.9853047132492065, 'reward_near_duplicate_penalty_std': 0.017314862459897995, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390502154827118, 'reward_total_composite_mean': 0.20318013429641724, 'reward_total_composite_std': 0.3137761652469635, 'epoch': 0.01}

  5%|▍         | 154/3300 [32:38<22:08:07, 25.33s/it]INFO 04-12 16:31:05 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 155/3300 [32:46<17:35:04, 20.13s/it]2026-04-12 16:31:13,054 | INFO | train_grpo_train | metrics_logged mode=train step=155

                                                     
{'loss': 0.111, 'grad_norm': 24.618492126464844, 'learning_rate': 9.533333333333334e-06, 'num_tokens': 300247.0, 'completions/mean_length': 18.875, 'completions/min_length': 15.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9403250813484192, 'rewards/meter/std': 0.06395131349563599, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.945521354675293, 'rewards/lexical_plausibility/std': 0.11963294446468353, 'rewards/judge_quality/mean': 0.7387499809265137, 'rewards/judge_quality/std': 0.3636496663093567, 'rewards/repeat_penalty/mean': 0.5773205757141113, 'rewards/repeat_penalty/std': 0.248104527592659, 'rewards/repeat_floor/mean': 0.9476659297943115, 'rewards/repeat_floor/std': 0.05473656207323074, 'rewards/near_duplicate_penalty/mean': 0.8023881316184998, 'rewards/near_duplicate_penalty/std': 0.27356192469596863, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9145299196243286, 'rewards/distinct_2/std': 0.15825983881950378, 'rewards/total_composite/mean': 0.6672157049179077, 'rewards/total_composite/std': 0.33616819977760315, 'reward': 0.6672157049179077, 'reward_std': 0.33616816997528076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17333175241947174, 'sampling/sampling_logp_difference/max': 0.9076957702636719, 'sampling/importance_sampling_ratio/min': 0.4034528136253357, 'sampling/importance_sampling_ratio/mean': 1.0622689723968506, 'sampling/importance_sampling_ratio/max': 1.9556593894958496, 'entropy': 1.8294667154550552, 'clip_ratio/low_mean': 0.07727272808551788, 'clip_ratio/low_min': 0.07727272808551788, 'clip_ratio/high_mean': 0.153308829292655, 'clip_ratio/high_max': 0.153308829292655, 'clip_ratio/region_mean': 0.23058155737817287, 'reward_total_mean': 0.6672157049179077, 'reward_meter_mean': 0.9403250813484192, 'reward_meter_std': 0.06395131349563599, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.945521354675293, 'reward_lexical_plausibility_std': 0.11963294446468353, 'reward_repeat_penalty_mean': 0.5773205757141113, 'reward_repeat_penalty_std': 0.248104527592659, 'reward_repeat_floor_mean': 0.9476659297943115, 'reward_repeat_floor_std': 0.05473656207323074, 'reward_near_duplicate_penalty_mean': 0.8023881316184998, 'reward_near_duplicate_penalty_std': 0.27356192469596863, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9145299196243286, 'reward_distinct_2_std': 0.15825983881950378, 'reward_judge_quality_mean': 0.7387499809265137, 'reward_judge_quality_std': 0.3636496663093567, 'reward_total_composite_mean': 0.6672157049179077, 'reward_total_composite_std': 0.33616819977760315, 'epoch': 0.01}

  5%|▍         | 155/3300 [32:46<17:35:04, 20.13s/it]INFO 04-12 16:31:13 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 156/3300 [33:00<15:57:09, 18.27s/it]2026-04-12 16:31:26,975 | INFO | train_grpo_train | metrics_logged mode=train step=156

                                                     
{'loss': 0.0807, 'grad_norm': 6.091550350189209, 'learning_rate': 9.530303030303031e-06, 'num_tokens': 302442.0, 'completions/mean_length': 174.375, 'completions/min_length': 83.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 126.14286041259766, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 259.0, 'rewards/meter/mean': 0.27547502517700195, 'rewards/meter/std': 0.33542144298553467, 'rewards/count_adherence/mean': 0.7999999523162842, 'rewards/count_adherence/std': 0.2618614733219147, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.07855844497680664, 'rewards/lexical_plausibility/mean': 0.9448052644729614, 'rewards/lexical_plausibility/std': 0.12211354076862335, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9574728012084961, 'rewards/repeat_penalty/std': 0.03196243569254875, 'rewards/repeat_floor/mean': 0.9954248666763306, 'rewards/repeat_floor/std': 0.0035908089485019445, 'rewards/near_duplicate_penalty/mean': 0.9883695840835571, 'rewards/near_duplicate_penalty/std': 0.011273902840912342, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9685494899749756, 'rewards/distinct_2/std': 0.02286594547331333, 'rewards/total_composite/mean': 0.10009932518005371, 'rewards/total_composite/std': 0.1299850195646286, 'reward': 0.10009932518005371, 'reward_std': 0.1299850195646286, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2429916262626648, 'sampling/sampling_logp_difference/max': 1.702545166015625, 'sampling/importance_sampling_ratio/min': 0.18221914768218994, 'sampling/importance_sampling_ratio/mean': 1.0676641464233398, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.298069953918457, 'clip_ratio/low_mean': 0.12119883298873901, 'clip_ratio/low_min': 0.12119883298873901, 'clip_ratio/high_mean': 0.053592994809150696, 'clip_ratio/high_max': 0.053592994809150696, 'clip_ratio/region_mean': 0.1747918277978897, 'reward_total_mean': 0.10009932518005371, 'reward_meter_mean': 0.27547502517700195, 'reward_meter_std': 0.33542144298553467, 'reward_count_adherence_mean': 0.7999999523162842, 'reward_count_adherence_std': 0.2618614733219147, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.07855844497680664, 'reward_lexical_plausibility_mean': 0.9448052644729614, 'reward_lexical_plausibility_std': 0.12211354076862335, 'reward_repeat_penalty_mean': 0.9574728012084961, 'reward_repeat_penalty_std': 0.03196243569254875, 'reward_repeat_floor_mean': 0.9954248666763306, 'reward_repeat_floor_std': 0.0035908089485019445, 'reward_near_duplicate_penalty_mean': 0.9883695840835571, 'reward_near_duplicate_penalty_std': 0.011273902840912342, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9685494899749756, 'reward_distinct_2_std': 0.02286594547331333, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.10009932518005371, 'reward_total_composite_std': 0.1299850195646286, 'epoch': 0.01}

  5%|▍         | 156/3300 [33:00<15:57:09, 18.27s/it]INFO 04-12 16:31:27 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 157/3300 [33:08<13:08:29, 15.05s/it]2026-04-12 16:31:34,530 | INFO | train_grpo_train | metrics_logged mode=train step=157

                                                     
{'loss': 0.0154, 'grad_norm': 17.993427276611328, 'learning_rate': 9.527272727272729e-06, 'num_tokens': 303802.0, 'completions/mean_length': 22.0, 'completions/min_length': 17.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9873130321502686, 'rewards/meter/std': 0.017658350989222527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7425000071525574, 'rewards/judge_quality/std': 0.27185869216918945, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7237325310707092, 'rewards/total_composite/std': 0.26927655935287476, 'reward': 0.7237325310707092, 'reward_std': 0.26927658915519714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19480332732200623, 'sampling/sampling_logp_difference/max': 1.1242094039916992, 'sampling/importance_sampling_ratio/min': 0.3249092400074005, 'sampling/importance_sampling_ratio/mean': 1.061350703239441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8669251203536987, 'clip_ratio/low_mean': 0.09261363931000233, 'clip_ratio/low_min': 0.09261363931000233, 'clip_ratio/high_mean': 0.09373304760083556, 'clip_ratio/high_max': 0.09373304760083556, 'clip_ratio/region_mean': 0.1863466869108379, 'reward_total_mean': 0.7237325310707092, 'reward_meter_mean': 0.9873130321502686, 'reward_meter_std': 0.017658350989222527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7425000071525574, 'reward_judge_quality_std': 0.27185869216918945, 'reward_total_composite_mean': 0.7237325310707092, 'reward_total_composite_std': 0.26927655935287476, 'epoch': 0.01}

  5%|▍         | 157/3300 [33:08<13:08:29, 15.05s/it]INFO 04-12 16:31:34 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 158/3300 [33:21<12:45:17, 14.61s/it]2026-04-12 16:31:48,122 | INFO | train_grpo_train | metrics_logged mode=train step=158

                                                     
{'loss': 0.0181, 'grad_norm': 5.933853626251221, 'learning_rate': 9.524242424242424e-06, 'num_tokens': 305477.0, 'completions/mean_length': 108.375, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.71428680419922, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.38190749287605286, 'rewards/meter/std': 0.4354288876056671, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9033068418502808, 'rewards/lexical_plausibility/std': 0.21514086425304413, 'rewards/judge_quality/mean': 0.20624998211860657, 'rewards/judge_quality/std': 0.1590990275144577, 'rewards/repeat_penalty/mean': 1.0, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 1.0, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11781704425811768, 'rewards/total_composite/std': 0.14058606326580048, 'reward': 0.11781704425811768, 'reward_std': 0.14058606326580048, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23430916666984558, 'sampling/sampling_logp_difference/max': 1.268911361694336, 'sampling/importance_sampling_ratio/min': 0.28113749623298645, 'sampling/importance_sampling_ratio/mean': 1.0673094987869263, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0684922635555267, 'clip_ratio/low_mean': 0.07401142083108425, 'clip_ratio/low_min': 0.07401142083108425, 'clip_ratio/high_mean': 0.14546815305948257, 'clip_ratio/high_max': 0.14546815305948257, 'clip_ratio/region_mean': 0.21947957389056683, 'reward_total_mean': 0.11781704425811768, 'reward_meter_mean': 0.38190749287605286, 'reward_meter_std': 0.4354288876056671, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9033068418502808, 'reward_lexical_plausibility_std': 0.21514086425304413, 'reward_repeat_penalty_mean': 1.0, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 1.0, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.20624998211860657, 'reward_judge_quality_std': 0.1590990275144577, 'reward_total_composite_mean': 0.11781704425811768, 'reward_total_composite_std': 0.14058606326580048, 'epoch': 0.01}

  5%|▍         | 158/3300 [33:21<12:45:17, 14.61s/it]INFO 04-12 16:31:48 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 159/3300 [33:35<12:32:28, 14.37s/it]2026-04-12 16:32:01,943 | INFO | train_grpo_train | metrics_logged mode=train step=159

                                                     
{'loss': -0.2268, 'grad_norm': 2.426312208175659, 'learning_rate': 9.521212121212121e-06, 'num_tokens': 308165.0, 'completions/mean_length': 333.0, 'completions/min_length': 141.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 225.60000610351562, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 360.0, 'rewards/meter/mean': 0.353273868560791, 'rewards/meter/std': 0.2707987129688263, 'rewards/count_adherence/mean': 0.6785714626312256, 'rewards/count_adherence/std': 0.26175010204315186, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9035714268684387, 'rewards/count_floor/std': 0.07852502912282944, 'rewards/lexical_plausibility/mean': 0.6502748727798462, 'rewards/lexical_plausibility/std': 0.21541759371757507, 'rewards/judge_quality/mean': 0.10000000149011612, 'rewards/judge_quality/std': 0.053452253341674805, 'rewards/repeat_penalty/mean': 0.9957379102706909, 'rewards/repeat_penalty/std': 0.0068611567839980125, 'rewards/repeat_floor/mean': 0.999508261680603, 'rewards/repeat_floor/std': 0.0006383705767802894, 'rewards/near_duplicate_penalty/mean': 0.9982011914253235, 'rewards/near_duplicate_penalty/std': 0.0019292989745736122, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9975345134735107, 'rewards/distinct_2/std': 0.006973449140787125, 'rewards/total_composite/mean': 0.03163018077611923, 'rewards/total_composite/std': 0.023396549746394157, 'reward': 0.03163018077611923, 'reward_std': 0.023396549746394157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24087728559970856, 'sampling/sampling_logp_difference/max': 1.4648590087890625, 'sampling/importance_sampling_ratio/min': 0.2311105877161026, 'sampling/importance_sampling_ratio/mean': 1.0651675462722778, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.174529790878296, 'clip_ratio/low_mean': 0.02570921927690506, 'clip_ratio/low_min': 0.02570921927690506, 'clip_ratio/high_mean': 0.10626691207289696, 'clip_ratio/high_max': 0.10626691207289696, 'clip_ratio/region_mean': 0.13197613134980202, 'reward_total_mean': 0.03163018077611923, 'reward_meter_mean': 0.353273868560791, 'reward_meter_std': 0.2707987129688263, 'reward_count_adherence_mean': 0.6785714626312256, 'reward_count_adherence_std': 0.26175010204315186, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9035714268684387, 'reward_count_floor_std': 0.07852502912282944, 'reward_lexical_plausibility_mean': 0.6502748727798462, 'reward_lexical_plausibility_std': 0.21541759371757507, 'reward_repeat_penalty_mean': 0.9957379102706909, 'reward_repeat_penalty_std': 0.0068611567839980125, 'reward_repeat_floor_mean': 0.999508261680603, 'reward_repeat_floor_std': 0.0006383705767802894, 'reward_near_duplicate_penalty_mean': 0.9982011914253235, 'reward_near_duplicate_penalty_std': 0.0019292989745736122, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9975345134735107, 'reward_distinct_2_std': 0.006973449140787125, 'reward_judge_quality_mean': 0.10000000149011612, 'reward_judge_quality_std': 0.053452253341674805, 'reward_total_composite_mean': 0.03163018077611923, 'reward_total_composite_std': 0.023396549746394157, 'epoch': 0.01}

  5%|▍         | 159/3300 [33:35<12:32:28, 14.37s/it]INFO 04-12 16:32:02 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 160/3300 [33:50<12:41:16, 14.55s/it]2026-04-12 16:32:16,892 | INFO | train_grpo_train | metrics_logged mode=train step=160

                                                     
{'loss': -0.0537, 'grad_norm': 7.381004333496094, 'learning_rate': 9.518181818181819e-06, 'num_tokens': 310465.0, 'completions/mean_length': 154.5, 'completions/min_length': 74.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 103.42857360839844, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.28334954380989075, 'rewards/meter/std': 0.33286550641059875, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9006118178367615, 'rewards/lexical_plausibility/std': 0.20599594712257385, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9929472208023071, 'rewards/repeat_penalty/std': 0.006425357423722744, 'rewards/repeat_floor/mean': 0.9989420771598816, 'rewards/repeat_floor/std': 0.0009637992479838431, 'rewards/near_duplicate_penalty/mean': 0.9929472208023071, 'rewards/near_duplicate_penalty/std': 0.006425357423722744, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.039745062589645386, 'rewards/total_composite/std': 0.040800224989652634, 'reward': 0.039745062589645386, 'reward_std': 0.040800221264362335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22394897043704987, 'sampling/sampling_logp_difference/max': 1.5978527069091797, 'sampling/importance_sampling_ratio/min': 0.20233049988746643, 'sampling/importance_sampling_ratio/mean': 1.0648385286331177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8074288964271545, 'clip_ratio/low_mean': 0.09691987931728363, 'clip_ratio/low_min': 0.09691987931728363, 'clip_ratio/high_mean': 0.07455702871084213, 'clip_ratio/high_max': 0.07455702871084213, 'clip_ratio/region_mean': 0.17147690802812576, 'reward_total_mean': 0.039745062589645386, 'reward_meter_mean': 0.28334954380989075, 'reward_meter_std': 0.33286550641059875, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9006118178367615, 'reward_lexical_plausibility_std': 0.20599594712257385, 'reward_repeat_penalty_mean': 0.9929472208023071, 'reward_repeat_penalty_std': 0.006425357423722744, 'reward_repeat_floor_mean': 0.9989420771598816, 'reward_repeat_floor_std': 0.0009637992479838431, 'reward_near_duplicate_penalty_mean': 0.9929472208023071, 'reward_near_duplicate_penalty_std': 0.006425357423722744, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.039745062589645386, 'reward_total_composite_std': 0.040800224989652634, 'epoch': 0.01}

  5%|▍         | 160/3300 [33:50<12:41:16, 14.55s/it]INFO 04-12 16:32:17 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 161/3300 [34:05<12:51:03, 14.74s/it]2026-04-12 16:32:32,078 | INFO | train_grpo_train | metrics_logged mode=train step=161

                                                     
{'loss': 0.0495, 'grad_norm': 1.618666648864746, 'learning_rate': 9.515151515151516e-06, 'num_tokens': 313485.0, 'completions/mean_length': 435.5, 'completions/min_length': 307.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 359.0, 'completions/min_terminated_length': 307.0, 'completions/max_terminated_length': 392.0, 'rewards/meter/mean': 0.026769960299134254, 'rewards/meter/std': 0.0258164219558239, 'rewards/count_adherence/mean': 0.6911765336990356, 'rewards/count_adherence/std': 0.23894231021404266, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9073529243469238, 'rewards/count_floor/std': 0.07168269902467728, 'rewards/lexical_plausibility/mean': 0.7938939332962036, 'rewards/lexical_plausibility/std': 0.22051064670085907, 'rewards/judge_quality/mean': 0.10000000149011612, 'rewards/judge_quality/std': 0.053452253341674805, 'rewards/repeat_penalty/mean': 0.9705691337585449, 'rewards/repeat_penalty/std': 0.04052077978849411, 'rewards/repeat_floor/mean': 0.997039794921875, 'rewards/repeat_floor/std': 0.003988719079643488, 'rewards/near_duplicate_penalty/mean': 0.9952114224433899, 'rewards/near_duplicate_penalty/std': 0.004913213662803173, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9750899076461792, 'rewards/distinct_2/std': 0.03648589551448822, 'rewards/total_composite/mean': 0.0027222498320043087, 'rewards/total_composite/std': 0.003158347448334098, 'reward': 0.0027222498320043087, 'reward_std': 0.003158347215503454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2468816339969635, 'sampling/sampling_logp_difference/max': 1.328948974609375, 'sampling/importance_sampling_ratio/min': 0.26475536823272705, 'sampling/importance_sampling_ratio/mean': 1.0749328136444092, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4229578971862793, 'clip_ratio/low_mean': 0.06084258295595646, 'clip_ratio/low_min': 0.06084258295595646, 'clip_ratio/high_mean': 0.02207792177796364, 'clip_ratio/high_max': 0.02207792177796364, 'clip_ratio/region_mean': 0.0829205047339201, 'reward_total_mean': 0.0027222498320043087, 'reward_meter_mean': 0.026769960299134254, 'reward_meter_std': 0.0258164219558239, 'reward_count_adherence_mean': 0.6911765336990356, 'reward_count_adherence_std': 0.23894231021404266, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9073529243469238, 'reward_count_floor_std': 0.07168269902467728, 'reward_lexical_plausibility_mean': 0.7938939332962036, 'reward_lexical_plausibility_std': 0.22051064670085907, 'reward_repeat_penalty_mean': 0.9705691337585449, 'reward_repeat_penalty_std': 0.04052077978849411, 'reward_repeat_floor_mean': 0.997039794921875, 'reward_repeat_floor_std': 0.003988719079643488, 'reward_near_duplicate_penalty_mean': 0.9952114224433899, 'reward_near_duplicate_penalty_std': 0.004913213662803173, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9750899076461792, 'reward_distinct_2_std': 0.03648589551448822, 'reward_judge_quality_mean': 0.10000000149011612, 'reward_judge_quality_std': 0.053452253341674805, 'reward_total_composite_mean': 0.0027222498320043087, 'reward_total_composite_std': 0.003158347448334098, 'epoch': 0.01}

  5%|▍         | 161/3300 [34:05<12:51:03, 14.74s/it]INFO 04-12 16:32:32 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 162/3300 [34:15<11:28:24, 13.16s/it]2026-04-12 16:32:41,567 | INFO | train_grpo_train | metrics_logged mode=train step=162

                                                     
{'loss': 0.0407, 'grad_norm': 23.070953369140625, 'learning_rate': 9.512121212121213e-06, 'num_tokens': 315070.0, 'completions/mean_length': 28.125, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.125, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.6077741384506226, 'rewards/meter/std': 0.3664523959159851, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6974999904632568, 'rewards/judge_quality/std': 0.25172266364097595, 'rewards/repeat_penalty/mean': 0.9586921334266663, 'rewards/repeat_penalty/std': 0.03407951444387436, 'rewards/repeat_floor/mean': 0.9945210814476013, 'rewards/repeat_floor/std': 0.004034711048007011, 'rewards/near_duplicate_penalty/mean': 0.9706853628158569, 'rewards/near_duplicate_penalty/std': 0.027476971969008446, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9879807829856873, 'rewards/distinct_2/std': 0.03399552404880524, 'rewards/total_composite/mean': 0.4812217652797699, 'rewards/total_composite/std': 0.35587385296821594, 'reward': 0.4812217652797699, 'reward_std': 0.35587382316589355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1815616935491562, 'sampling/sampling_logp_difference/max': 1.334096908569336, 'sampling/importance_sampling_ratio/min': 0.26339593529701233, 'sampling/importance_sampling_ratio/mean': 1.0680123567581177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3512811288237572, 'clip_ratio/low_mean': 0.10988849867135286, 'clip_ratio/low_min': 0.10988849867135286, 'clip_ratio/high_mean': 0.08059663884341717, 'clip_ratio/high_max': 0.08059663884341717, 'clip_ratio/region_mean': 0.19048513751477003, 'reward_total_mean': 0.4812217652797699, 'reward_meter_mean': 0.6077741384506226, 'reward_meter_std': 0.3664523959159851, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9586921334266663, 'reward_repeat_penalty_std': 0.03407951444387436, 'reward_repeat_floor_mean': 0.9945210814476013, 'reward_repeat_floor_std': 0.004034711048007011, 'reward_near_duplicate_penalty_mean': 0.9706853628158569, 'reward_near_duplicate_penalty_std': 0.027476971969008446, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9879807829856873, 'reward_distinct_2_std': 0.03399552404880524, 'reward_judge_quality_mean': 0.6974999904632568, 'reward_judge_quality_std': 0.25172266364097595, 'reward_total_composite_mean': 0.4812217652797699, 'reward_total_composite_std': 0.35587385296821594, 'epoch': 0.01}

  5%|▍         | 162/3300 [34:15<11:28:24, 13.16s/it]INFO 04-12 16:32:42 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 163/3300 [34:23<10:14:16, 11.75s/it]2026-04-12 16:32:50,006 | INFO | train_grpo_train | metrics_logged mode=train step=163

                                                     
{'loss': 0.0529, 'grad_norm': 17.180492401123047, 'learning_rate': 9.50909090909091e-06, 'num_tokens': 316574.0, 'completions/mean_length': 38.0, 'completions/min_length': 27.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.510293185710907, 'rewards/meter/std': 0.4052880108356476, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.2655990719795227, 'rewards/repeat_penalty/mean': 0.9861856698989868, 'rewards/repeat_penalty/std': 0.01695219799876213, 'rewards/repeat_floor/mean': 0.9979279041290283, 'rewards/repeat_floor/std': 0.002542835893109441, 'rewards/near_duplicate_penalty/mean': 0.9861856698989868, 'rewards/near_duplicate_penalty/std': 0.01695219799876213, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3274025619029999, 'rewards/total_composite/std': 0.27338168025016785, 'reward': 0.3274025619029999, 'reward_std': 0.27338168025016785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16291232407093048, 'sampling/sampling_logp_difference/max': 2.105924606323242, 'sampling/importance_sampling_ratio/min': 0.12173306196928024, 'sampling/importance_sampling_ratio/mean': 1.0006645917892456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7504936307668686, 'clip_ratio/low_mean': 0.08019230980426073, 'clip_ratio/low_min': 0.08019230980426073, 'clip_ratio/high_mean': 0.07817300595343113, 'clip_ratio/high_max': 0.07817300595343113, 'clip_ratio/region_mean': 0.15836531575769186, 'reward_total_mean': 0.3274025619029999, 'reward_meter_mean': 0.510293185710907, 'reward_meter_std': 0.4052880108356476, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9861856698989868, 'reward_repeat_penalty_std': 0.01695219799876213, 'reward_repeat_floor_mean': 0.9979279041290283, 'reward_repeat_floor_std': 0.002542835893109441, 'reward_near_duplicate_penalty_mean': 0.9861856698989868, 'reward_near_duplicate_penalty_std': 0.01695219799876213, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.2655990719795227, 'reward_total_composite_mean': 0.3274025619029999, 'reward_total_composite_std': 0.27338168025016785, 'epoch': 0.01}

  5%|▍         | 163/3300 [34:23<10:14:16, 11.75s/it]INFO 04-12 16:32:50 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 164/3300 [34:33<9:45:21, 11.20s/it] 2026-04-12 16:32:59,922 | INFO | train_grpo_train | metrics_logged mode=train step=164

                                                    
{'loss': 0.0941, 'grad_norm': 8.719386100769043, 'learning_rate': 9.506060606060606e-06, 'num_tokens': 319141.0, 'completions/mean_length': 119.875, 'completions/min_length': 92.0, 'completions/max_length': 168.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.875, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 168.0, 'rewards/meter/mean': 0.25987565517425537, 'rewards/meter/std': 0.3236730992794037, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.9806985259056091, 'rewards/lexical_plausibility/std': 0.054592814296483994, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9858438968658447, 'rewards/repeat_penalty/std': 0.020412007346749306, 'rewards/repeat_floor/mean': 0.9982364177703857, 'rewards/repeat_floor/std': 0.0021149625536054373, 'rewards/near_duplicate_penalty/mean': 0.9919647574424744, 'rewards/near_duplicate_penalty/std': 0.0065601542592048645, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9937965273857117, 'rewards/distinct_2/std': 0.017546070739626884, 'rewards/total_composite/mean': 0.09222100675106049, 'rewards/total_composite/std': 0.11238480359315872, 'reward': 0.09222100675106049, 'reward_std': 0.11238480359315872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22538737952709198, 'sampling/sampling_logp_difference/max': 1.495957374572754, 'sampling/importance_sampling_ratio/min': 0.22403402626514435, 'sampling/importance_sampling_ratio/mean': 1.0768688917160034, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.218059927225113, 'clip_ratio/low_mean': 0.10372049733996391, 'clip_ratio/low_min': 0.10372049733996391, 'clip_ratio/high_mean': 0.06451714038848877, 'clip_ratio/high_max': 0.06451714038848877, 'clip_ratio/region_mean': 0.16823763772845268, 'reward_total_mean': 0.09222100675106049, 'reward_meter_mean': 0.25987565517425537, 'reward_meter_std': 0.3236730992794037, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.9806985259056091, 'reward_lexical_plausibility_std': 0.054592814296483994, 'reward_repeat_penalty_mean': 0.9858438968658447, 'reward_repeat_penalty_std': 0.020412007346749306, 'reward_repeat_floor_mean': 0.9982364177703857, 'reward_repeat_floor_std': 0.0021149625536054373, 'reward_near_duplicate_penalty_mean': 0.9919647574424744, 'reward_near_duplicate_penalty_std': 0.0065601542592048645, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9937965273857117, 'reward_distinct_2_std': 0.017546070739626884, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.09222100675106049, 'reward_total_composite_std': 0.11238480359315872, 'epoch': 0.01}

  5%|▍         | 164/3300 [34:33<9:45:21, 11.20s/it]INFO 04-12 16:33:00 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 165/3300 [34:42<9:11:47, 10.56s/it]2026-04-12 16:33:08,995 | INFO | train_grpo_train | metrics_logged mode=train step=165

                                                    
{'loss': 0.2982, 'grad_norm': 14.512274742126465, 'learning_rate': 9.503030303030303e-06, 'num_tokens': 320704.0, 'completions/mean_length': 45.375, 'completions/min_length': 21.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.5098780989646912, 'rewards/meter/std': 0.5024057030677795, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.11338934302330017, 'rewards/lexical_plausibility/mean': 0.9523521661758423, 'rewards/lexical_plausibility/std': 0.10907252132892609, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.23754698038101196, 'rewards/repeat_penalty/mean': 0.9246719479560852, 'rewards/repeat_penalty/std': 0.11142340302467346, 'rewards/repeat_floor/mean': 0.9947507381439209, 'rewards/repeat_floor/std': 0.006766745820641518, 'rewards/near_duplicate_penalty/mean': 0.9945298433303833, 'rewards/near_duplicate_penalty/std': 0.01021417137235403, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9924585223197937, 'rewards/distinct_2/std': 0.021330520510673523, 'rewards/total_composite/mean': 0.23539981245994568, 'rewards/total_composite/std': 0.28147608041763306, 'reward': 0.23539981245994568, 'reward_std': 0.28147611021995544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25541940331459045, 'sampling/sampling_logp_difference/max': 1.4899091720581055, 'sampling/importance_sampling_ratio/min': 0.2253931164741516, 'sampling/importance_sampling_ratio/mean': 1.0630767345428467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7429435700178146, 'clip_ratio/low_mean': 0.09702238254249096, 'clip_ratio/low_min': 0.09702238254249096, 'clip_ratio/high_mean': 0.08932022284716368, 'clip_ratio/high_max': 0.08932022284716368, 'clip_ratio/region_mean': 0.18634260538965464, 'reward_total_mean': 0.23539981245994568, 'reward_meter_mean': 0.5098780989646912, 'reward_meter_std': 0.5024057030677795, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.11338934302330017, 'reward_lexical_plausibility_mean': 0.9523521661758423, 'reward_lexical_plausibility_std': 0.10907252132892609, 'reward_repeat_penalty_mean': 0.9246719479560852, 'reward_repeat_penalty_std': 0.11142340302467346, 'reward_repeat_floor_mean': 0.9947507381439209, 'reward_repeat_floor_std': 0.006766745820641518, 'reward_near_duplicate_penalty_mean': 0.9945298433303833, 'reward_near_duplicate_penalty_std': 0.01021417137235403, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9924585223197937, 'reward_distinct_2_std': 0.021330520510673523, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.23754698038101196, 'reward_total_composite_mean': 0.23539981245994568, 'reward_total_composite_std': 0.28147608041763306, 'epoch': 0.01}

  5%|▌         | 165/3300 [34:42<9:11:47, 10.56s/it]INFO 04-12 16:33:09 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 166/3300 [34:56<10:03:27, 11.55s/it]2026-04-12 16:33:22,863 | INFO | train_grpo_train | metrics_logged mode=train step=166

                                                     
{'loss': 0.0185, 'grad_norm': 5.241777420043945, 'learning_rate': 9.5e-06, 'num_tokens': 322245.0, 'completions/mean_length': 109.625, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.142860412597656, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.38890665769577026, 'rewards/meter/std': 0.36653921008110046, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9018011093139648, 'rewards/lexical_plausibility/std': 0.2394527643918991, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.29001232981681824, 'rewards/repeat_penalty/mean': 0.9259210824966431, 'rewards/repeat_penalty/std': 0.13395890593528748, 'rewards/repeat_floor/mean': 0.9946435689926147, 'rewards/repeat_floor/std': 0.009626256301999092, 'rewards/near_duplicate_penalty/mean': 0.9957007169723511, 'rewards/near_duplicate_penalty/std': 0.0070299566723406315, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.19523629546165466, 'rewards/total_composite/std': 0.2460174411535263, 'reward': 0.19523629546165466, 'reward_std': 0.2460174411535263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.218777135014534, 'sampling/sampling_logp_difference/max': 1.8099956512451172, 'sampling/importance_sampling_ratio/min': 0.1636548489332199, 'sampling/importance_sampling_ratio/mean': 1.0517387390136719, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.281535819172859, 'clip_ratio/low_mean': 0.09778554923832417, 'clip_ratio/low_min': 0.09778554923832417, 'clip_ratio/high_mean': 0.054242050275206566, 'clip_ratio/high_max': 0.054242050275206566, 'clip_ratio/region_mean': 0.15202759951353073, 'reward_total_mean': 0.19523629546165466, 'reward_meter_mean': 0.38890665769577026, 'reward_meter_std': 0.36653921008110046, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9018011093139648, 'reward_lexical_plausibility_std': 0.2394527643918991, 'reward_repeat_penalty_mean': 0.9259210824966431, 'reward_repeat_penalty_std': 0.13395890593528748, 'reward_repeat_floor_mean': 0.9946435689926147, 'reward_repeat_floor_std': 0.009626256301999092, 'reward_near_duplicate_penalty_mean': 0.9957007169723511, 'reward_near_duplicate_penalty_std': 0.0070299566723406315, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.29001232981681824, 'reward_total_composite_mean': 0.19523629546165466, 'reward_total_composite_std': 0.2460174411535263, 'epoch': 0.01}

  5%|▌         | 166/3300 [34:56<10:03:27, 11.55s/it]INFO 04-12 16:33:23 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 167/3300 [35:05<9:25:51, 10.84s/it] 2026-04-12 16:33:32,028 | INFO | train_grpo_train | metrics_logged mode=train step=167

                                                    
{'loss': 0.0312, 'grad_norm': 12.906083106994629, 'learning_rate': 9.496969696969698e-06, 'num_tokens': 324231.0, 'completions/mean_length': 79.25, 'completions/min_length': 64.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.1399468183517456, 'rewards/meter/std': 0.11115916073322296, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9864832758903503, 'rewards/lexical_plausibility/std': 0.03823106735944748, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9820793271064758, 'rewards/repeat_penalty/std': 0.015891199931502342, 'rewards/repeat_floor/mean': 0.997779130935669, 'rewards/repeat_floor/std': 0.0016564702382311225, 'rewards/near_duplicate_penalty/mean': 0.9899773001670837, 'rewards/near_duplicate_penalty/std': 0.00716781709343195, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9920282363891602, 'rewards/distinct_2/std': 0.014811928384006023, 'rewards/total_composite/mean': 0.0491066500544548, 'rewards/total_composite/std': 0.04577239602804184, 'reward': 0.0491066500544548, 'reward_std': 0.04577239602804184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24595601856708527, 'sampling/sampling_logp_difference/max': 1.7702323198318481, 'sampling/importance_sampling_ratio/min': 0.1702934205532074, 'sampling/importance_sampling_ratio/mean': 1.049609661102295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7938202917575836, 'clip_ratio/low_mean': 0.14112098701298237, 'clip_ratio/low_min': 0.14112098701298237, 'clip_ratio/high_mean': 0.057857686653733253, 'clip_ratio/high_max': 0.057857686653733253, 'clip_ratio/region_mean': 0.19897867366671562, 'reward_total_mean': 0.0491066500544548, 'reward_meter_mean': 0.1399468183517456, 'reward_meter_std': 0.11115916073322296, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9864832758903503, 'reward_lexical_plausibility_std': 0.03823106735944748, 'reward_repeat_penalty_mean': 0.9820793271064758, 'reward_repeat_penalty_std': 0.015891199931502342, 'reward_repeat_floor_mean': 0.997779130935669, 'reward_repeat_floor_std': 0.0016564702382311225, 'reward_near_duplicate_penalty_mean': 0.9899773001670837, 'reward_near_duplicate_penalty_std': 0.00716781709343195, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9920282363891602, 'reward_distinct_2_std': 0.014811928384006023, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.0491066500544548, 'reward_total_composite_std': 0.04577239602804184, 'epoch': 0.01}

  5%|▌         | 167/3300 [35:05<9:25:51, 10.84s/it]INFO 04-12 16:33:32 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 168/3300 [35:19<10:09:15, 11.67s/it]2026-04-12 16:33:45,648 | INFO | train_grpo_train | metrics_logged mode=train step=168

                                                     
{'loss': 0.0082, 'grad_norm': 3.6465823650360107, 'learning_rate': 9.493939393939395e-06, 'num_tokens': 327106.0, 'completions/mean_length': 298.375, 'completions/min_length': 176.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 227.1666717529297, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 456.0, 'rewards/meter/mean': 0.46919679641723633, 'rewards/meter/std': 0.3492467403411865, 'rewards/count_adherence/mean': 0.828125, 'rewards/count_adherence/std': 0.11451567709445953, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9484374523162842, 'rewards/count_floor/std': 0.03435469791293144, 'rewards/lexical_plausibility/mean': 0.8188076019287109, 'rewards/lexical_plausibility/std': 0.19052761793136597, 'rewards/judge_quality/mean': 0.11250000447034836, 'rewards/judge_quality/std': 0.051754921674728394, 'rewards/repeat_penalty/mean': 0.9575713872909546, 'rewards/repeat_penalty/std': 0.07485056668519974, 'rewards/repeat_floor/mean': 0.9963113069534302, 'rewards/repeat_floor/std': 0.0055867028422653675, 'rewards/near_duplicate_penalty/mean': 0.9936656951904297, 'rewards/near_duplicate_penalty/std': 0.006218735594302416, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9695711731910706, 'rewards/distinct_2/std': 0.0545896552503109, 'rewards/total_composite/mean': 0.05623955279588699, 'rewards/total_composite/std': 0.05702324956655502, 'reward': 0.05623955279588699, 'reward_std': 0.057023245841264725, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22682708501815796, 'sampling/sampling_logp_difference/max': 1.3170084953308105, 'sampling/importance_sampling_ratio/min': 0.2679356336593628, 'sampling/importance_sampling_ratio/mean': 1.0585166215896606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.532101571559906, 'clip_ratio/low_mean': 0.04777298215776682, 'clip_ratio/low_min': 0.04777298215776682, 'clip_ratio/high_mean': 0.06866891868412495, 'clip_ratio/high_max': 0.06866891868412495, 'clip_ratio/region_mean': 0.11644190084189177, 'reward_total_mean': 0.05623955279588699, 'reward_meter_mean': 0.46919679641723633, 'reward_meter_std': 0.3492467403411865, 'reward_count_adherence_mean': 0.828125, 'reward_count_adherence_std': 0.11451567709445953, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9484374523162842, 'reward_count_floor_std': 0.03435469791293144, 'reward_lexical_plausibility_mean': 0.8188076019287109, 'reward_lexical_plausibility_std': 0.19052761793136597, 'reward_repeat_penalty_mean': 0.9575713872909546, 'reward_repeat_penalty_std': 0.07485056668519974, 'reward_repeat_floor_mean': 0.9963113069534302, 'reward_repeat_floor_std': 0.0055867028422653675, 'reward_near_duplicate_penalty_mean': 0.9936656951904297, 'reward_near_duplicate_penalty_std': 0.006218735594302416, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9695711731910706, 'reward_distinct_2_std': 0.0545896552503109, 'reward_judge_quality_mean': 0.11250000447034836, 'reward_judge_quality_std': 0.051754921674728394, 'reward_total_composite_mean': 0.05623955279588699, 'reward_total_composite_std': 0.05702324956655502, 'epoch': 0.01}

  5%|▌         | 168/3300 [35:19<10:09:15, 11.67s/it]INFO 04-12 16:33:45 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 169/3300 [35:30<10:02:00, 11.54s/it]2026-04-12 16:33:56,869 | INFO | train_grpo_train | metrics_logged mode=train step=169

                                                     
{'loss': 0.1387, 'grad_norm': 7.401317596435547, 'learning_rate': 9.490909090909092e-06, 'num_tokens': 329621.0, 'completions/mean_length': 129.375, 'completions/min_length': 73.0, 'completions/max_length': 272.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.375, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 272.0, 'rewards/meter/mean': 0.23914751410484314, 'rewards/meter/std': 0.31655964255332947, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9106929302215576, 'rewards/lexical_plausibility/std': 0.2039894312620163, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9846621155738831, 'rewards/repeat_penalty/std': 0.01714184693992138, 'rewards/repeat_floor/mean': 0.9981511831283569, 'rewards/repeat_floor/std': 0.0017984056612476707, 'rewards/near_duplicate_penalty/mean': 0.9923198223114014, 'rewards/near_duplicate_penalty/std': 0.005941803567111492, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9922581315040588, 'rewards/distinct_2/std': 0.014396161772310734, 'rewards/total_composite/mean': 0.04120022431015968, 'rewards/total_composite/std': 0.04593072086572647, 'reward': 0.04120022431015968, 'reward_std': 0.045930713415145874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24399013817310333, 'sampling/sampling_logp_difference/max': 1.256594181060791, 'sampling/importance_sampling_ratio/min': 0.2846217453479767, 'sampling/importance_sampling_ratio/mean': 1.0619888305664062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7711445093154907, 'clip_ratio/low_mean': 0.12760332319885492, 'clip_ratio/low_min': 0.12760332319885492, 'clip_ratio/high_mean': 0.07233880087733269, 'clip_ratio/high_max': 0.07233880087733269, 'clip_ratio/region_mean': 0.1999421240761876, 'reward_total_mean': 0.04120022431015968, 'reward_meter_mean': 0.23914751410484314, 'reward_meter_std': 0.31655964255332947, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9106929302215576, 'reward_lexical_plausibility_std': 0.2039894312620163, 'reward_repeat_penalty_mean': 0.9846621155738831, 'reward_repeat_penalty_std': 0.01714184693992138, 'reward_repeat_floor_mean': 0.9981511831283569, 'reward_repeat_floor_std': 0.0017984056612476707, 'reward_near_duplicate_penalty_mean': 0.9923198223114014, 'reward_near_duplicate_penalty_std': 0.005941803567111492, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9922581315040588, 'reward_distinct_2_std': 0.014396161772310734, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.04120022431015968, 'reward_total_composite_std': 0.04593072086572647, 'epoch': 0.01}

  5%|▌         | 169/3300 [35:30<10:02:00, 11.54s/it]INFO 04-12 16:33:57 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 170/3300 [35:38<9:06:47, 10.48s/it] 2026-04-12 16:34:04,890 | INFO | train_grpo_train | metrics_logged mode=train step=170

                                                    
{'loss': 0.1843, 'grad_norm': 14.099448204040527, 'learning_rate': 9.487878787878788e-06, 'num_tokens': 331351.0, 'completions/mean_length': 41.25, 'completions/min_length': 28.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.3891289234161377, 'rewards/meter/std': 0.49643561244010925, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9196504354476929, 'rewards/lexical_plausibility/std': 0.15237295627593994, 'rewards/judge_quality/mean': 0.5, 'rewards/judge_quality/std': 0.26613637804985046, 'rewards/repeat_penalty/mean': 0.9899561405181885, 'rewards/repeat_penalty/std': 0.012933881022036076, 'rewards/repeat_floor/mean': 0.9984934329986572, 'rewards/repeat_floor/std': 0.0019400803139433265, 'rewards/near_duplicate_penalty/mean': 0.9899561405181885, 'rewards/near_duplicate_penalty/std': 0.012933881022036076, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2277049720287323, 'rewards/total_composite/std': 0.32318681478500366, 'reward': 0.2277049720287323, 'reward_std': 0.3231867849826813, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23590445518493652, 'sampling/sampling_logp_difference/max': 1.0857219696044922, 'sampling/importance_sampling_ratio/min': 0.3376579284667969, 'sampling/importance_sampling_ratio/mean': 1.066724181175232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8991546034812927, 'clip_ratio/low_mean': 0.11641414184123278, 'clip_ratio/low_min': 0.11641414184123278, 'clip_ratio/high_mean': 0.05895737465471029, 'clip_ratio/high_max': 0.05895737465471029, 'clip_ratio/region_mean': 0.17537151649594307, 'reward_total_mean': 0.2277049720287323, 'reward_meter_mean': 0.3891289234161377, 'reward_meter_std': 0.49643561244010925, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9196504354476929, 'reward_lexical_plausibility_std': 0.15237295627593994, 'reward_repeat_penalty_mean': 0.9899561405181885, 'reward_repeat_penalty_std': 0.012933881022036076, 'reward_repeat_floor_mean': 0.9984934329986572, 'reward_repeat_floor_std': 0.0019400803139433265, 'reward_near_duplicate_penalty_mean': 0.9899561405181885, 'reward_near_duplicate_penalty_std': 0.012933881022036076, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5, 'reward_judge_quality_std': 0.26613637804985046, 'reward_total_composite_mean': 0.2277049720287323, 'reward_total_composite_std': 0.32318681478500366, 'epoch': 0.01}

  5%|▌         | 170/3300 [35:38<9:06:47, 10.48s/it]INFO 04-12 16:34:05 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 171/3300 [35:48<8:55:55, 10.28s/it]2026-04-12 16:34:14,687 | INFO | train_grpo_train | metrics_logged mode=train step=171

                                                    
{'loss': 0.2631, 'grad_norm': 12.088873863220215, 'learning_rate': 9.484848484848485e-06, 'num_tokens': 332990.0, 'completions/mean_length': 53.875, 'completions/min_length': 33.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.5055831074714661, 'rewards/meter/std': 0.45644906163215637, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9804325103759766, 'rewards/lexical_plausibility/std': 0.03285950794816017, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.25035688281059265, 'rewards/repeat_penalty/mean': 0.9683129191398621, 'rewards/repeat_penalty/std': 0.08822023123502731, 'rewards/repeat_floor/mean': 0.9980594515800476, 'rewards/repeat_floor/std': 0.005279997363686562, 'rewards/near_duplicate_penalty/mean': 0.9995629191398621, 'rewards/near_duplicate_penalty/std': 0.0012362091802060604, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2332489937543869, 'rewards/total_composite/std': 0.28420260548591614, 'reward': 0.2332489937543869, 'reward_std': 0.28420257568359375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23966380953788757, 'sampling/sampling_logp_difference/max': 1.5754125118255615, 'sampling/importance_sampling_ratio/min': 0.2928496301174164, 'sampling/importance_sampling_ratio/mean': 1.0594810247421265, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.178064525127411, 'clip_ratio/low_mean': 0.10319321043789387, 'clip_ratio/low_min': 0.10319321043789387, 'clip_ratio/high_mean': 0.07592468895018101, 'clip_ratio/high_max': 0.07592468895018101, 'clip_ratio/region_mean': 0.17911789938807487, 'reward_total_mean': 0.2332489937543869, 'reward_meter_mean': 0.5055831074714661, 'reward_meter_std': 0.45644906163215637, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9804325103759766, 'reward_lexical_plausibility_std': 0.03285950794816017, 'reward_repeat_penalty_mean': 0.9683129191398621, 'reward_repeat_penalty_std': 0.08822023123502731, 'reward_repeat_floor_mean': 0.9980594515800476, 'reward_repeat_floor_std': 0.005279997363686562, 'reward_near_duplicate_penalty_mean': 0.9995629191398621, 'reward_near_duplicate_penalty_std': 0.0012362091802060604, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.25035688281059265, 'reward_total_composite_mean': 0.2332489937543869, 'reward_total_composite_std': 0.28420260548591614, 'epoch': 0.01}

  5%|▌         | 171/3300 [35:48<8:55:55, 10.28s/it]INFO 04-12 16:34:14 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 172/3300 [36:01<9:45:16, 11.23s/it]2026-04-12 16:34:28,130 | INFO | train_grpo_train | metrics_logged mode=train step=172

                                                    
{'loss': 0.0568, 'grad_norm': 6.325699806213379, 'learning_rate': 9.481818181818182e-06, 'num_tokens': 334935.0, 'completions/mean_length': 136.125, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 82.42857360839844, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.5953339338302612, 'rewards/meter/std': 0.3515208959579468, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.07954952120780945, 'rewards/lexical_plausibility/mean': 0.8813091516494751, 'rewards/lexical_plausibility/std': 0.25190088152885437, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9240643382072449, 'rewards/repeat_penalty/std': 0.15501122176647186, 'rewards/repeat_floor/mean': 0.9928107857704163, 'rewards/repeat_floor/std': 0.013184435665607452, 'rewards/near_duplicate_penalty/mean': 0.9815835952758789, 'rewards/near_duplicate_penalty/std': 0.023980464786291122, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9508135914802551, 'rewards/distinct_2/std': 0.10906938463449478, 'rewards/total_composite/mean': 0.21437200903892517, 'rewards/total_composite/std': 0.154502272605896, 'reward': 0.21437200903892517, 'reward_std': 0.154502272605896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2241189181804657, 'sampling/sampling_logp_difference/max': 1.5689411163330078, 'sampling/importance_sampling_ratio/min': 0.20826558768749237, 'sampling/importance_sampling_ratio/mean': 1.0423859357833862, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6517246663570404, 'clip_ratio/low_mean': 0.04110402427613735, 'clip_ratio/low_min': 0.04110402427613735, 'clip_ratio/high_mean': 0.1448630504310131, 'clip_ratio/high_max': 0.1448630504310131, 'clip_ratio/region_mean': 0.18596707470715046, 'reward_total_mean': 0.21437200903892517, 'reward_meter_mean': 0.5953339338302612, 'reward_meter_std': 0.3515208959579468, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.07954952120780945, 'reward_lexical_plausibility_mean': 0.8813091516494751, 'reward_lexical_plausibility_std': 0.25190088152885437, 'reward_repeat_penalty_mean': 0.9240643382072449, 'reward_repeat_penalty_std': 0.15501122176647186, 'reward_repeat_floor_mean': 0.9928107857704163, 'reward_repeat_floor_std': 0.013184435665607452, 'reward_near_duplicate_penalty_mean': 0.9815835952758789, 'reward_near_duplicate_penalty_std': 0.023980464786291122, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9508135914802551, 'reward_distinct_2_std': 0.10906938463449478, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.21437200903892517, 'reward_total_composite_std': 0.154502272605896, 'epoch': 0.01}

  5%|▌         | 172/3300 [36:01<9:45:16, 11.23s/it]INFO 04-12 16:34:28 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 173/3300 [36:10<9:05:04, 10.46s/it]2026-04-12 16:34:36,800 | INFO | train_grpo_train | metrics_logged mode=train step=173

                                                    
{'loss': 0.1524, 'grad_norm': 15.391106605529785, 'learning_rate': 9.47878787878788e-06, 'num_tokens': 336529.0, 'completions/mean_length': 39.25, 'completions/min_length': 28.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6029083728790283, 'rewards/meter/std': 0.4984326660633087, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9895833730697632, 'rewards/lexical_plausibility/std': 0.029462775215506554, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9780707359313965, 'rewards/repeat_penalty/std': 0.04785861819982529, 'rewards/repeat_floor/mean': 0.9973043203353882, 'rewards/repeat_floor/std': 0.005578472279012203, 'rewards/near_duplicate_penalty/mean': 0.9888160228729248, 'rewards/near_duplicate_penalty/std': 0.02026197873055935, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.26817238330841064, 'rewards/total_composite/std': 0.3063661754131317, 'reward': 0.26817238330841064, 'reward_std': 0.3063661754131317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24825410544872284, 'sampling/sampling_logp_difference/max': 1.511880874633789, 'sampling/importance_sampling_ratio/min': 0.2204948514699936, 'sampling/importance_sampling_ratio/mean': 1.0907857418060303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.249059647321701, 'clip_ratio/low_mean': 0.11488499492406845, 'clip_ratio/low_min': 0.11488499492406845, 'clip_ratio/high_mean': 0.09170249849557877, 'clip_ratio/high_max': 0.09170249849557877, 'clip_ratio/region_mean': 0.20658749341964722, 'reward_total_mean': 0.26817238330841064, 'reward_meter_mean': 0.6029083728790283, 'reward_meter_std': 0.4984326660633087, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9895833730697632, 'reward_lexical_plausibility_std': 0.029462775215506554, 'reward_repeat_penalty_mean': 0.9780707359313965, 'reward_repeat_penalty_std': 0.04785861819982529, 'reward_repeat_floor_mean': 0.9973043203353882, 'reward_repeat_floor_std': 0.005578472279012203, 'reward_near_duplicate_penalty_mean': 0.9888160228729248, 'reward_near_duplicate_penalty_std': 0.02026197873055935, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.26817238330841064, 'reward_total_composite_std': 0.3063661754131317, 'epoch': 0.01}

  5%|▌         | 173/3300 [36:10<9:05:04, 10.46s/it]INFO 04-12 16:34:37 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 174/3300 [36:18<8:30:43,  9.80s/it]2026-04-12 16:34:45,072 | INFO | train_grpo_train | metrics_logged mode=train step=174

                                                    
{'loss': 0.013, 'grad_norm': 14.048095703125, 'learning_rate': 9.475757575757577e-06, 'num_tokens': 338062.0, 'completions/mean_length': 39.625, 'completions/min_length': 35.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.36875519156455994, 'rewards/meter/std': 0.3963353931903839, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9579613208770752, 'rewards/lexical_plausibility/std': 0.06439344584941864, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.26029860973358154, 'rewards/repeat_penalty/mean': 0.9898826479911804, 'rewards/repeat_penalty/std': 0.011895010247826576, 'rewards/repeat_floor/mean': 0.9984824061393738, 'rewards/repeat_floor/std': 0.0017842425731942058, 'rewards/near_duplicate_penalty/mean': 0.9898826479911804, 'rewards/near_duplicate_penalty/std': 0.011895010247826576, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20255500078201294, 'rewards/total_composite/std': 0.30077502131462097, 'reward': 0.20255500078201294, 'reward_std': 0.30077502131462097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20269235968589783, 'sampling/sampling_logp_difference/max': 1.1932525634765625, 'sampling/importance_sampling_ratio/min': 0.30323338508605957, 'sampling/importance_sampling_ratio/mean': 1.0353174209594727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.374999389052391, 'clip_ratio/low_mean': 0.130195040255785, 'clip_ratio/low_min': 0.130195040255785, 'clip_ratio/high_mean': 0.05330598400905728, 'clip_ratio/high_max': 0.05330598400905728, 'clip_ratio/region_mean': 0.18350102426484227, 'reward_total_mean': 0.20255500078201294, 'reward_meter_mean': 0.36875519156455994, 'reward_meter_std': 0.3963353931903839, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9579613208770752, 'reward_lexical_plausibility_std': 0.06439344584941864, 'reward_repeat_penalty_mean': 0.9898826479911804, 'reward_repeat_penalty_std': 0.011895010247826576, 'reward_repeat_floor_mean': 0.9984824061393738, 'reward_repeat_floor_std': 0.0017842425731942058, 'reward_near_duplicate_penalty_mean': 0.9898826479911804, 'reward_near_duplicate_penalty_std': 0.011895010247826576, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.26029860973358154, 'reward_total_composite_mean': 0.20255500078201294, 'reward_total_composite_std': 0.30077502131462097, 'epoch': 0.01}

  5%|▌         | 174/3300 [36:18<8:30:43,  9.80s/it]INFO 04-12 16:34:45 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 175/3300 [36:26<8:03:48,  9.29s/it]2026-04-12 16:34:53,163 | INFO | train_grpo_train | metrics_logged mode=train step=175

                                                    
{'loss': 0.1704, 'grad_norm': 16.34452247619629, 'learning_rate': 9.472727272727274e-06, 'num_tokens': 339664.0, 'completions/mean_length': 40.25, 'completions/min_length': 31.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.47447168827056885, 'rewards/meter/std': 0.4072783589363098, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9674941897392273, 'rewards/lexical_plausibility/std': 0.04732557013630867, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9900354146957397, 'rewards/repeat_penalty/std': 0.014336712658405304, 'rewards/repeat_floor/mean': 0.9985052943229675, 'rewards/repeat_floor/std': 0.0021505055483430624, 'rewards/near_duplicate_penalty/mean': 0.9900354146957397, 'rewards/near_duplicate_penalty/std': 0.014336712658405304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13026830554008484, 'rewards/total_composite/std': 0.16197744011878967, 'reward': 0.13026830554008484, 'reward_std': 0.16197744011878967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2528935372829437, 'sampling/sampling_logp_difference/max': 1.213531494140625, 'sampling/importance_sampling_ratio/min': 0.2971460521221161, 'sampling/importance_sampling_ratio/mean': 1.0664143562316895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0034386217594147, 'clip_ratio/low_mean': 0.1279802629724145, 'clip_ratio/low_min': 0.1279802629724145, 'clip_ratio/high_mean': 0.05409564450383186, 'clip_ratio/high_max': 0.05409564450383186, 'clip_ratio/region_mean': 0.18207590747624636, 'reward_total_mean': 0.13026830554008484, 'reward_meter_mean': 0.47447168827056885, 'reward_meter_std': 0.4072783589363098, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9674941897392273, 'reward_lexical_plausibility_std': 0.04732557013630867, 'reward_repeat_penalty_mean': 0.9900354146957397, 'reward_repeat_penalty_std': 0.014336712658405304, 'reward_repeat_floor_mean': 0.9985052943229675, 'reward_repeat_floor_std': 0.0021505055483430624, 'reward_near_duplicate_penalty_mean': 0.9900354146957397, 'reward_near_duplicate_penalty_std': 0.014336712658405304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.13026830554008484, 'reward_total_composite_std': 0.16197744011878967, 'epoch': 0.01}

  5%|▌         | 175/3300 [36:27<8:03:48,  9.29s/it]INFO 04-12 16:34:53 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 176/3300 [36:40<9:09:30, 10.55s/it]2026-04-12 16:35:06,666 | INFO | train_grpo_train | metrics_logged mode=train step=176

                                                    
{'loss': 0.0399, 'grad_norm': 2.9782471656799316, 'learning_rate': 9.469696969696971e-06, 'num_tokens': 342295.0, 'completions/mean_length': 280.875, 'completions/min_length': 12.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 203.83334350585938, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 476.0, 'rewards/meter/mean': 0.4379279613494873, 'rewards/meter/std': 0.3314565420150757, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.3204349875450134, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.09613049775362015, 'rewards/lexical_plausibility/mean': 0.5992525815963745, 'rewards/lexical_plausibility/std': 0.26424071192741394, 'rewards/judge_quality/mean': 0.0625, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9964256882667542, 'rewards/repeat_penalty/std': 0.003698437474668026, 'rewards/repeat_floor/mean': 0.9995236396789551, 'rewards/repeat_floor/std': 0.0004962141974829137, 'rewards/near_duplicate_penalty/mean': 0.9974220991134644, 'rewards/near_duplicate_penalty/std': 0.0034069011453539133, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9990035891532898, 'rewards/distinct_2/std': 0.002818275475874543, 'rewards/total_composite/mean': 0.028571154922246933, 'rewards/total_composite/std': 0.030546991154551506, 'reward': 0.028571154922246933, 'reward_std': 0.030546991154551506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21626479923725128, 'sampling/sampling_logp_difference/max': 1.0515127182006836, 'sampling/importance_sampling_ratio/min': 0.3494087755680084, 'sampling/importance_sampling_ratio/mean': 1.0579288005828857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.235089272260666, 'clip_ratio/low_mean': 0.04965019319206476, 'clip_ratio/low_min': 0.04965019319206476, 'clip_ratio/high_mean': 0.06005795765668154, 'clip_ratio/high_max': 0.06005795765668154, 'clip_ratio/region_mean': 0.1097081508487463, 'reward_total_mean': 0.028571154922246933, 'reward_meter_mean': 0.4379279613494873, 'reward_meter_std': 0.3314565420150757, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.3204349875450134, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.09613049775362015, 'reward_lexical_plausibility_mean': 0.5992525815963745, 'reward_lexical_plausibility_std': 0.26424071192741394, 'reward_repeat_penalty_mean': 0.9964256882667542, 'reward_repeat_penalty_std': 0.003698437474668026, 'reward_repeat_floor_mean': 0.9995236396789551, 'reward_repeat_floor_std': 0.0004962141974829137, 'reward_near_duplicate_penalty_mean': 0.9974220991134644, 'reward_near_duplicate_penalty_std': 0.0034069011453539133, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9990035891532898, 'reward_distinct_2_std': 0.002818275475874543, 'reward_judge_quality_mean': 0.0625, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.028571154922246933, 'reward_total_composite_std': 0.030546991154551506, 'epoch': 0.01}

  5%|▌         | 176/3300 [36:40<9:09:30, 10.55s/it]INFO 04-12 16:35:06 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 177/3300 [36:53<9:46:00, 11.26s/it]2026-04-12 16:35:19,569 | INFO | train_grpo_train | metrics_logged mode=train step=177

                                                    
{'loss': 0.2103, 'grad_norm': 4.683366775512695, 'learning_rate': 9.466666666666667e-06, 'num_tokens': 343915.0, 'completions/mean_length': 114.5, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.71428680419922, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 280.0, 'rewards/meter/mean': 0.48603808879852295, 'rewards/meter/std': 0.46289294958114624, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.6432515382766724, 'rewards/lexical_plausibility/std': 0.31876441836357117, 'rewards/judge_quality/mean': 0.3425000309944153, 'rewards/judge_quality/std': 0.3682681620121002, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.1157275140285492, 'rewards/repeat_floor/mean': 0.9887499809265137, 'rewards/repeat_floor/std': 0.0069436440244317055, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.181174635887146, 'rewards/total_composite/std': 0.21942174434661865, 'reward': 0.181174635887146, 'reward_std': 0.21942174434661865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20456665754318237, 'sampling/sampling_logp_difference/max': 1.0288105010986328, 'sampling/importance_sampling_ratio/min': 0.35743188858032227, 'sampling/importance_sampling_ratio/mean': 1.0502562522888184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.5841771960258484, 'clip_ratio/low_mean': 0.10271092038601637, 'clip_ratio/low_min': 0.10271092038601637, 'clip_ratio/high_mean': 0.09828431438654661, 'clip_ratio/high_max': 0.09828431438654661, 'clip_ratio/region_mean': 0.20099523477256298, 'reward_total_mean': 0.181174635887146, 'reward_meter_mean': 0.48603808879852295, 'reward_meter_std': 0.46289294958114624, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.6432515382766724, 'reward_lexical_plausibility_std': 0.31876441836357117, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.1157275140285492, 'reward_repeat_floor_mean': 0.9887499809265137, 'reward_repeat_floor_std': 0.0069436440244317055, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3425000309944153, 'reward_judge_quality_std': 0.3682681620121002, 'reward_total_composite_mean': 0.181174635887146, 'reward_total_composite_std': 0.21942174434661865, 'epoch': 0.01}

  5%|▌         | 177/3300 [36:53<9:46:00, 11.26s/it]INFO 04-12 16:35:19 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 178/3300 [37:07<10:24:11, 12.00s/it]2026-04-12 16:35:33,294 | INFO | train_grpo_train | metrics_logged mode=train step=178

                                                     
{'loss': -0.0233, 'grad_norm': 6.025400638580322, 'learning_rate': 9.463636363636364e-06, 'num_tokens': 345819.0, 'completions/mean_length': 126.0, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.85714721679688, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.4014953672885895, 'rewards/meter/std': 0.3172938823699951, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.05580177903175354, 'rewards/lexical_plausibility/mean': 0.8826035261154175, 'rewards/lexical_plausibility/std': 0.3079337477684021, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9888430833816528, 'rewards/repeat_penalty/std': 0.010254358872771263, 'rewards/repeat_floor/mean': 0.9983264207839966, 'rewards/repeat_floor/std': 0.001538152457214892, 'rewards/near_duplicate_penalty/mean': 0.9888430833816528, 'rewards/near_duplicate_penalty/std': 0.010254358872771263, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13863034546375275, 'rewards/total_composite/std': 0.1386062204837799, 'reward': 0.13863034546375275, 'reward_std': 0.1386062055826187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21773995459079742, 'sampling/sampling_logp_difference/max': 1.9745492935180664, 'sampling/importance_sampling_ratio/min': 0.13882386684417725, 'sampling/importance_sampling_ratio/mean': 1.04308021068573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5849375128746033, 'clip_ratio/low_mean': 0.09830117598176003, 'clip_ratio/low_min': 0.09830117598176003, 'clip_ratio/high_mean': 0.06714709103107452, 'clip_ratio/high_max': 0.06714709103107452, 'clip_ratio/region_mean': 0.16544826701283455, 'reward_total_mean': 0.13863034546375275, 'reward_meter_mean': 0.4014953672885895, 'reward_meter_std': 0.3172938823699951, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.05580177903175354, 'reward_lexical_plausibility_mean': 0.8826035261154175, 'reward_lexical_plausibility_std': 0.3079337477684021, 'reward_repeat_penalty_mean': 0.9888430833816528, 'reward_repeat_penalty_std': 0.010254358872771263, 'reward_repeat_floor_mean': 0.9983264207839966, 'reward_repeat_floor_std': 0.001538152457214892, 'reward_near_duplicate_penalty_mean': 0.9888430833816528, 'reward_near_duplicate_penalty_std': 0.010254358872771263, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.13863034546375275, 'reward_total_composite_std': 0.1386062204837799, 'epoch': 0.01}

  5%|▌         | 178/3300 [37:07<10:24:11, 12.00s/it]INFO 04-12 16:35:33 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 179/3300 [37:20<10:51:56, 12.53s/it]2026-04-12 16:35:47,075 | INFO | train_grpo_train | metrics_logged mode=train step=179

                                                     
{'loss': 0.0374, 'grad_norm': 6.673909664154053, 'learning_rate': 9.460606060606061e-06, 'num_tokens': 347320.0, 'completions/mean_length': 106.625, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.2345580756664276, 'rewards/meter/std': 0.3012939989566803, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.7145659923553467, 'rewards/lexical_plausibility/std': 0.3297322690486908, 'rewards/judge_quality/mean': 0.21249999105930328, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.982830286026001, 'rewards/repeat_penalty/std': 0.027915718033909798, 'rewards/repeat_floor/mean': 0.997907280921936, 'rewards/repeat_floor/std': 0.002971539506688714, 'rewards/near_duplicate_penalty/mean': 0.9910648465156555, 'rewards/near_duplicate_penalty/std': 0.010946473106741905, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.0730305165052414, 'rewards/total_composite/std': 0.13297533988952637, 'reward': 0.0730305165052414, 'reward_std': 0.13297533988952637, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2518250644207001, 'sampling/sampling_logp_difference/max': 1.1850132942199707, 'sampling/importance_sampling_ratio/min': 0.3057421147823334, 'sampling/importance_sampling_ratio/mean': 1.0697705745697021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3620694875717163, 'clip_ratio/low_mean': 0.12779998034238815, 'clip_ratio/low_min': 0.12779998034238815, 'clip_ratio/high_mean': 0.014285714365541935, 'clip_ratio/high_max': 0.014285714365541935, 'clip_ratio/region_mean': 0.1420856947079301, 'reward_total_mean': 0.0730305165052414, 'reward_meter_mean': 0.2345580756664276, 'reward_meter_std': 0.3012939989566803, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.7145659923553467, 'reward_lexical_plausibility_std': 0.3297322690486908, 'reward_repeat_penalty_mean': 0.982830286026001, 'reward_repeat_penalty_std': 0.027915718033909798, 'reward_repeat_floor_mean': 0.997907280921936, 'reward_repeat_floor_std': 0.002971539506688714, 'reward_near_duplicate_penalty_mean': 0.9910648465156555, 'reward_near_duplicate_penalty_std': 0.010946473106741905, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.0730305165052414, 'reward_total_composite_std': 0.13297533988952637, 'epoch': 0.01}

  5%|▌         | 179/3300 [37:20<10:51:56, 12.53s/it]INFO 04-12 16:35:47 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 180/3300 [37:34<11:12:36, 12.93s/it]2026-04-12 16:36:00,944 | INFO | train_grpo_train | metrics_logged mode=train step=180

                                                     
{'loss': -0.1523, 'grad_norm': 2.562324285507202, 'learning_rate': 9.457575757575759e-06, 'num_tokens': 350334.0, 'completions/mean_length': 384.75, 'completions/min_length': 193.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 308.3999938964844, 'completions/min_terminated_length': 193.0, 'completions/max_terminated_length': 449.0, 'rewards/meter/mean': 0.499294251203537, 'rewards/meter/std': 0.41147667169570923, 'rewards/count_adherence/mean': 0.640625, 'rewards/count_adherence/std': 0.2538762092590332, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8921874761581421, 'rewards/count_floor/std': 0.07616285979747772, 'rewards/lexical_plausibility/mean': 0.5289751291275024, 'rewards/lexical_plausibility/std': 0.2826981246471405, 'rewards/judge_quality/mean': 0.06875000149011612, 'rewards/judge_quality/std': 0.0530330128967762, 'rewards/repeat_penalty/mean': 0.9891723990440369, 'rewards/repeat_penalty/std': 0.015472057275474072, 'rewards/repeat_floor/mean': 0.9987034797668457, 'rewards/repeat_floor/std': 0.0017474140040576458, 'rewards/near_duplicate_penalty/mean': 0.9947054982185364, 'rewards/near_duplicate_penalty/std': 0.007706289645284414, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9944184422492981, 'rewards/distinct_2/std': 0.011827335692942142, 'rewards/total_composite/mean': 0.023752262815833092, 'rewards/total_composite/std': 0.02465471811592579, 'reward': 0.023752262815833092, 'reward_std': 0.02465471811592579, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21988987922668457, 'sampling/sampling_logp_difference/max': 1.5030994415283203, 'sampling/importance_sampling_ratio/min': 0.22243966162204742, 'sampling/importance_sampling_ratio/mean': 1.0672351121902466, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.988249957561493, 'clip_ratio/low_mean': 0.03225304465740919, 'clip_ratio/low_min': 0.03225304465740919, 'clip_ratio/high_mean': 0.04911600053310394, 'clip_ratio/high_max': 0.04911600053310394, 'clip_ratio/region_mean': 0.08136904519051313, 'reward_total_mean': 0.023752262815833092, 'reward_meter_mean': 0.499294251203537, 'reward_meter_std': 0.41147667169570923, 'reward_count_adherence_mean': 0.640625, 'reward_count_adherence_std': 0.2538762092590332, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8921874761581421, 'reward_count_floor_std': 0.07616285979747772, 'reward_lexical_plausibility_mean': 0.5289751291275024, 'reward_lexical_plausibility_std': 0.2826981246471405, 'reward_repeat_penalty_mean': 0.9891723990440369, 'reward_repeat_penalty_std': 0.015472057275474072, 'reward_repeat_floor_mean': 0.9987034797668457, 'reward_repeat_floor_std': 0.0017474140040576458, 'reward_near_duplicate_penalty_mean': 0.9947054982185364, 'reward_near_duplicate_penalty_std': 0.007706289645284414, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9944184422492981, 'reward_distinct_2_std': 0.011827335692942142, 'reward_judge_quality_mean': 0.06875000149011612, 'reward_judge_quality_std': 0.0530330128967762, 'reward_total_composite_mean': 0.023752262815833092, 'reward_total_composite_std': 0.02465471811592579, 'epoch': 0.01}

  5%|▌         | 180/3300 [37:34<11:12:36, 12.93s/it]INFO 04-12 16:36:01 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 181/3300 [37:45<10:40:33, 12.32s/it]2026-04-12 16:36:11,838 | INFO | train_grpo_train | metrics_logged mode=train step=181

                                                     
{'loss': 0.3118, 'grad_norm': 6.628081321716309, 'learning_rate': 9.454545454545456e-06, 'num_tokens': 352540.0, 'completions/mean_length': 122.75, 'completions/min_length': 49.0, 'completions/max_length': 289.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 289.0, 'rewards/meter/mean': 0.09586291015148163, 'rewards/meter/std': 0.11569578945636749, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 0.8156722784042358, 'rewards/lexical_plausibility/std': 0.15160100162029266, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9882614612579346, 'rewards/repeat_penalty/std': 0.013631528243422508, 'rewards/repeat_floor/mean': 0.9982392191886902, 'rewards/repeat_floor/std': 0.0020447387360036373, 'rewards/near_duplicate_penalty/mean': 0.9882614612579346, 'rewards/near_duplicate_penalty/std': 0.013631528243422508, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.01775287464261055, 'rewards/total_composite/std': 0.0199081189930439, 'reward': 0.01775287464261055, 'reward_std': 0.0199081189930439, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2173110842704773, 'sampling/sampling_logp_difference/max': 1.7365942001342773, 'sampling/importance_sampling_ratio/min': 0.17611920833587646, 'sampling/importance_sampling_ratio/mean': 1.0354623794555664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.3025761395692825, 'clip_ratio/low_mean': 0.09172592684626579, 'clip_ratio/low_min': 0.09172592684626579, 'clip_ratio/high_mean': 0.07841318473219872, 'clip_ratio/high_max': 0.07841318473219872, 'clip_ratio/region_mean': 0.1701391115784645, 'reward_total_mean': 0.01775287464261055, 'reward_meter_mean': 0.09586291015148163, 'reward_meter_std': 0.11569578945636749, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 0.8156722784042358, 'reward_lexical_plausibility_std': 0.15160100162029266, 'reward_repeat_penalty_mean': 0.9882614612579346, 'reward_repeat_penalty_std': 0.013631528243422508, 'reward_repeat_floor_mean': 0.9982392191886902, 'reward_repeat_floor_std': 0.0020447387360036373, 'reward_near_duplicate_penalty_mean': 0.9882614612579346, 'reward_near_duplicate_penalty_std': 0.013631528243422508, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.01775287464261055, 'reward_total_composite_std': 0.0199081189930439, 'epoch': 0.01}

  5%|▌         | 181/3300 [37:45<10:40:33, 12.32s/it]INFO 04-12 16:36:12 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 182/3300 [37:54<9:44:33, 11.25s/it] 2026-04-12 16:36:20,580 | INFO | train_grpo_train | metrics_logged mode=train step=182

                                                    
{'loss': 0.0071, 'grad_norm': 17.36798095703125, 'learning_rate': 9.451515151515153e-06, 'num_tokens': 354531.0, 'completions/mean_length': 73.875, 'completions/min_length': 59.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.317579984664917, 'rewards/meter/std': 0.15974120795726776, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9326201677322388, 'rewards/repeat_penalty/std': 0.08470673114061356, 'rewards/repeat_floor/mean': 0.9924519062042236, 'rewards/repeat_floor/std': 0.009214295074343681, 'rewards/near_duplicate_penalty/mean': 0.9785886406898499, 'rewards/near_duplicate_penalty/std': 0.02176840789616108, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9518173933029175, 'rewards/distinct_2/std': 0.06827271729707718, 'rewards/total_composite/mean': 0.11482786387205124, 'rewards/total_composite/std': 0.06405065953731537, 'reward': 0.11482786387205124, 'reward_std': 0.06405065953731537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2649470269680023, 'sampling/sampling_logp_difference/max': 3.0607335567474365, 'sampling/importance_sampling_ratio/min': 0.046853311359882355, 'sampling/importance_sampling_ratio/mean': 0.9959365725517273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7446686774492264, 'clip_ratio/low_mean': 0.1078990250825882, 'clip_ratio/low_min': 0.1078990250825882, 'clip_ratio/high_mean': 0.1264716386795044, 'clip_ratio/high_max': 0.1264716386795044, 'clip_ratio/region_mean': 0.2343706637620926, 'reward_total_mean': 0.11482786387205124, 'reward_meter_mean': 0.317579984664917, 'reward_meter_std': 0.15974120795726776, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9326201677322388, 'reward_repeat_penalty_std': 0.08470673114061356, 'reward_repeat_floor_mean': 0.9924519062042236, 'reward_repeat_floor_std': 0.009214295074343681, 'reward_near_duplicate_penalty_mean': 0.9785886406898499, 'reward_near_duplicate_penalty_std': 0.02176840789616108, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9518173933029175, 'reward_distinct_2_std': 0.06827271729707718, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.11482786387205124, 'reward_total_composite_std': 0.06405065953731537, 'epoch': 0.01}

  6%|▌         | 182/3300 [37:54<9:44:33, 11.25s/it]INFO 04-12 16:36:20 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 183/3300 [38:01<8:42:27, 10.06s/it]2026-04-12 16:36:27,858 | INFO | train_grpo_train | metrics_logged mode=train step=183

                                                    
{'loss': 0.0507, 'grad_norm': 26.580228805541992, 'learning_rate': 9.448484848484849e-06, 'num_tokens': 355870.0, 'completions/mean_length': 27.375, 'completions/min_length': 24.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.5567831993103027, 'rewards/meter/std': 0.4256998598575592, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8212499618530273, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/repeat_penalty/mean': 0.9842255115509033, 'rewards/repeat_penalty/std': 0.02475259266793728, 'rewards/repeat_floor/mean': 0.9976338148117065, 'rewards/repeat_floor/std': 0.003712882287800312, 'rewards/near_duplicate_penalty/mean': 0.9842255115509033, 'rewards/near_duplicate_penalty/std': 0.02475259266793728, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5049673914909363, 'rewards/total_composite/std': 0.42627090215682983, 'reward': 0.5049673914909363, 'reward_std': 0.42627090215682983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1320503056049347, 'sampling/sampling_logp_difference/max': 1.4269943237304688, 'sampling/importance_sampling_ratio/min': 0.24002930521965027, 'sampling/importance_sampling_ratio/mean': 1.0096726417541504, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.871697224676609, 'clip_ratio/low_mean': 0.09298029448837042, 'clip_ratio/low_min': 0.09298029448837042, 'clip_ratio/high_mean': 0.03928571566939354, 'clip_ratio/high_max': 0.03928571566939354, 'clip_ratio/region_mean': 0.13226601015776396, 'reward_total_mean': 0.5049673914909363, 'reward_meter_mean': 0.5567831993103027, 'reward_meter_std': 0.4256998598575592, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9842255115509033, 'reward_repeat_penalty_std': 0.02475259266793728, 'reward_repeat_floor_mean': 0.9976338148117065, 'reward_repeat_floor_std': 0.003712882287800312, 'reward_near_duplicate_penalty_mean': 0.9842255115509033, 'reward_near_duplicate_penalty_std': 0.02475259266793728, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8212499618530273, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.5049673914909363, 'reward_total_composite_std': 0.42627090215682983, 'epoch': 0.01}

  6%|▌         | 183/3300 [38:01<8:42:27, 10.06s/it]INFO 04-12 16:36:28 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 184/3300 [38:14<9:31:50, 11.01s/it]2026-04-12 16:36:41,094 | INFO | train_grpo_train | metrics_logged mode=train step=184

                                                    
{'loss': -0.0154, 'grad_norm': 4.983717918395996, 'learning_rate': 9.445454545454546e-06, 'num_tokens': 357258.0, 'completions/mean_length': 81.5, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.6330931186676025, 'rewards/meter/std': 0.48437026143074036, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8900034427642822, 'rewards/lexical_plausibility/std': 0.26902082562446594, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.18015369772911072, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2401992529630661, 'rewards/total_composite/std': 0.2160726934671402, 'reward': 0.2401992529630661, 'reward_std': 0.216072678565979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24982981383800507, 'sampling/sampling_logp_difference/max': 1.1990976333618164, 'sampling/importance_sampling_ratio/min': 0.30146610736846924, 'sampling/importance_sampling_ratio/mean': 1.0588208436965942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7792413532733917, 'clip_ratio/low_mean': 0.06853590346872807, 'clip_ratio/low_min': 0.06853590346872807, 'clip_ratio/high_mean': 0.10964912362396717, 'clip_ratio/high_max': 0.10964912362396717, 'clip_ratio/region_mean': 0.17818502709269524, 'reward_total_mean': 0.2401992529630661, 'reward_meter_mean': 0.6330931186676025, 'reward_meter_std': 0.48437026143074036, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8900034427642822, 'reward_lexical_plausibility_std': 0.26902082562446594, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.18015369772911072, 'reward_total_composite_mean': 0.2401992529630661, 'reward_total_composite_std': 0.2160726934671402, 'epoch': 0.01}

  6%|▌         | 184/3300 [38:14<9:31:50, 11.01s/it]INFO 04-12 16:36:41 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 185/3300 [38:25<9:25:46, 10.90s/it]2026-04-12 16:36:51,728 | INFO | train_grpo_train | metrics_logged mode=train step=185

                                                    
{'loss': 0.4115, 'grad_norm': 7.150838375091553, 'learning_rate': 9.442424242424243e-06, 'num_tokens': 359283.0, 'completions/mean_length': 100.125, 'completions/min_length': 33.0, 'completions/max_length': 279.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 279.0, 'rewards/meter/mean': 0.3759284019470215, 'rewards/meter/std': 0.4123052656650543, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.7504698038101196, 'rewards/lexical_plausibility/std': 0.3222554326057434, 'rewards/judge_quality/mean': 0.21249999105930328, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.9969080686569214, 'rewards/repeat_penalty/std': 0.006644115783274174, 'rewards/repeat_floor/mean': 0.9995362162590027, 'rewards/repeat_floor/std': 0.0009966063080355525, 'rewards/near_duplicate_penalty/mean': 0.9969080686569214, 'rewards/near_duplicate_penalty/std': 0.006644115783274174, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08624979853630066, 'rewards/total_composite/std': 0.11350648105144501, 'reward': 0.08624979853630066, 'reward_std': 0.11350647360086441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2148657739162445, 'sampling/sampling_logp_difference/max': 1.4891471862792969, 'sampling/importance_sampling_ratio/min': 0.22556494176387787, 'sampling/importance_sampling_ratio/mean': 1.044112205505371, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.610387206077576, 'clip_ratio/low_mean': 0.10242792032659054, 'clip_ratio/low_min': 0.10242792032659054, 'clip_ratio/high_mean': 0.07306698057800531, 'clip_ratio/high_max': 0.07306698057800531, 'clip_ratio/region_mean': 0.17549490090459585, 'reward_total_mean': 0.08624979853630066, 'reward_meter_mean': 0.3759284019470215, 'reward_meter_std': 0.4123052656650543, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.7504698038101196, 'reward_lexical_plausibility_std': 0.3222554326057434, 'reward_repeat_penalty_mean': 0.9969080686569214, 'reward_repeat_penalty_std': 0.006644115783274174, 'reward_repeat_floor_mean': 0.9995362162590027, 'reward_repeat_floor_std': 0.0009966063080355525, 'reward_near_duplicate_penalty_mean': 0.9969080686569214, 'reward_near_duplicate_penalty_std': 0.006644115783274174, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.08624979853630066, 'reward_total_composite_std': 0.11350648105144501, 'epoch': 0.01}

  6%|▌         | 185/3300 [38:25<9:25:46, 10.90s/it]INFO 04-12 16:36:51 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 186/3300 [38:39<10:10:48, 11.77s/it]2026-04-12 16:37:05,531 | INFO | train_grpo_train | metrics_logged mode=train step=186

                                                     
{'loss': 0.0533, 'grad_norm': 2.1375834941864014, 'learning_rate': 9.43939393939394e-06, 'num_tokens': 360950.0, 'completions/mean_length': 243.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 82.20000457763672, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 266.0, 'rewards/meter/mean': 0.558614194393158, 'rewards/meter/std': 0.36556172370910645, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.727321982383728, 'rewards/lexical_plausibility/std': 0.30975380539894104, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.9993149638175964, 'rewards/repeat_penalty/std': 0.0019375960109755397, 'rewards/repeat_floor/mean': 0.9998972415924072, 'rewards/repeat_floor/std': 0.0002906446752604097, 'rewards/near_duplicate_penalty/mean': 0.9993149638175964, 'rewards/near_duplicate_penalty/std': 0.0019375960109755397, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13475856184959412, 'rewards/total_composite/std': 0.16865652799606323, 'reward': 0.13475856184959412, 'reward_std': 0.16865651309490204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21962442994117737, 'sampling/sampling_logp_difference/max': 1.4570684432983398, 'sampling/importance_sampling_ratio/min': 0.23291809856891632, 'sampling/importance_sampling_ratio/mean': 1.0197837352752686, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.900775671005249, 'clip_ratio/low_mean': 0.018703008070588112, 'clip_ratio/low_min': 0.018703008070588112, 'clip_ratio/high_mean': 0.07720588054507971, 'clip_ratio/high_max': 0.07720588054507971, 'clip_ratio/region_mean': 0.09590888861566782, 'reward_total_mean': 0.13475856184959412, 'reward_meter_mean': 0.558614194393158, 'reward_meter_std': 0.36556172370910645, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.727321982383728, 'reward_lexical_plausibility_std': 0.30975380539894104, 'reward_repeat_penalty_mean': 0.9993149638175964, 'reward_repeat_penalty_std': 0.0019375960109755397, 'reward_repeat_floor_mean': 0.9998972415924072, 'reward_repeat_floor_std': 0.0002906446752604097, 'reward_near_duplicate_penalty_mean': 0.9993149638175964, 'reward_near_duplicate_penalty_std': 0.0019375960109755397, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.13475856184959412, 'reward_total_composite_std': 0.16865652799606323, 'epoch': 0.01}

  6%|▌         | 186/3300 [38:39<10:10:48, 11.77s/it]INFO 04-12 16:37:05 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 187/3300 [38:53<10:43:22, 12.40s/it]2026-04-12 16:37:19,403 | INFO | train_grpo_train | metrics_logged mode=train step=187

                                                     
{'loss': -0.0061, 'grad_norm': 5.491358280181885, 'learning_rate': 9.436363636363636e-06, 'num_tokens': 362522.0, 'completions/mean_length': 96.5, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.142860412597656, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5083314180374146, 'rewards/meter/std': 0.4606216251850128, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8983074426651001, 'rewards/lexical_plausibility/std': 0.27657681703567505, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.19955307245254517, 'rewards/repeat_penalty/mean': 0.9808831214904785, 'rewards/repeat_penalty/std': 0.043287038803100586, 'rewards/repeat_floor/mean': 0.9976341128349304, 'rewards/repeat_floor/std': 0.00509362667798996, 'rewards/near_duplicate_penalty/mean': 0.9899966716766357, 'rewards/near_duplicate_penalty/std': 0.018181046470999718, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.16275985538959503, 'rewards/total_composite/std': 0.1781010776758194, 'reward': 0.16275985538959503, 'reward_std': 0.1781010925769806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22770176827907562, 'sampling/sampling_logp_difference/max': 1.481790542602539, 'sampling/importance_sampling_ratio/min': 0.2272304743528366, 'sampling/importance_sampling_ratio/mean': 1.0485010147094727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9776471555233002, 'clip_ratio/low_mean': 0.09363758750259876, 'clip_ratio/low_min': 0.09363758750259876, 'clip_ratio/high_mean': 0.08203125, 'clip_ratio/high_max': 0.08203125, 'clip_ratio/region_mean': 0.17566883750259876, 'reward_total_mean': 0.16275985538959503, 'reward_meter_mean': 0.5083314180374146, 'reward_meter_std': 0.4606216251850128, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8983074426651001, 'reward_lexical_plausibility_std': 0.27657681703567505, 'reward_repeat_penalty_mean': 0.9808831214904785, 'reward_repeat_penalty_std': 0.043287038803100586, 'reward_repeat_floor_mean': 0.9976341128349304, 'reward_repeat_floor_std': 0.00509362667798996, 'reward_near_duplicate_penalty_mean': 0.9899966716766357, 'reward_near_duplicate_penalty_std': 0.018181046470999718, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.19955307245254517, 'reward_total_composite_mean': 0.16275985538959503, 'reward_total_composite_std': 0.1781010776758194, 'epoch': 0.01}

  6%|▌         | 187/3300 [38:53<10:43:22, 12.40s/it]INFO 04-12 16:37:19 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 188/3300 [39:01<9:36:54, 11.12s/it] 2026-04-12 16:37:27,546 | INFO | train_grpo_train | metrics_logged mode=train step=188

                                                    
{'loss': 0.0468, 'grad_norm': 15.839167594909668, 'learning_rate': 9.433333333333335e-06, 'num_tokens': 364111.0, 'completions/mean_length': 37.625, 'completions/min_length': 29.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.4937368631362915, 'rewards/meter/std': 0.36918631196022034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9329289197921753, 'rewards/lexical_plausibility/std': 0.1064477264881134, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9874900579452515, 'rewards/repeat_penalty/std': 0.016355164349079132, 'rewards/repeat_floor/mean': 0.9981235265731812, 'rewards/repeat_floor/std': 0.002453265246003866, 'rewards/near_duplicate_penalty/mean': 0.9874900579452515, 'rewards/near_duplicate_penalty/std': 0.016355164349079132, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.188018798828125, 'rewards/total_composite/std': 0.15379884839057922, 'reward': 0.188018798828125, 'reward_std': 0.15379883348941803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24071529507637024, 'sampling/sampling_logp_difference/max': 1.2050809860229492, 'sampling/importance_sampling_ratio/min': 0.29966774582862854, 'sampling/importance_sampling_ratio/mean': 1.0537582635879517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.142243593931198, 'clip_ratio/low_mean': 0.10029411502182484, 'clip_ratio/low_min': 0.10029411502182484, 'clip_ratio/high_mean': 0.10462926793843508, 'clip_ratio/high_max': 0.10462926793843508, 'clip_ratio/region_mean': 0.20492338296025991, 'reward_total_mean': 0.188018798828125, 'reward_meter_mean': 0.4937368631362915, 'reward_meter_std': 0.36918631196022034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9329289197921753, 'reward_lexical_plausibility_std': 0.1064477264881134, 'reward_repeat_penalty_mean': 0.9874900579452515, 'reward_repeat_penalty_std': 0.016355164349079132, 'reward_repeat_floor_mean': 0.9981235265731812, 'reward_repeat_floor_std': 0.002453265246003866, 'reward_near_duplicate_penalty_mean': 0.9874900579452515, 'reward_near_duplicate_penalty_std': 0.016355164349079132, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.188018798828125, 'reward_total_composite_std': 0.15379884839057922, 'epoch': 0.01}

  6%|▌         | 188/3300 [39:01<9:36:54, 11.12s/it]INFO 04-12 16:37:27 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 189/3300 [39:14<10:11:50, 11.80s/it]2026-04-12 16:37:40,930 | INFO | train_grpo_train | metrics_logged mode=train step=189

                                                     
{'loss': 0.3546, 'grad_norm': 5.5799479484558105, 'learning_rate': 9.43030303030303e-06, 'num_tokens': 366066.0, 'completions/mean_length': 159.375, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 109.00000762939453, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 458.0, 'rewards/meter/mean': 0.6374885439872742, 'rewards/meter/std': 0.3750973641872406, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.8109548091888428, 'rewards/lexical_plausibility/std': 0.31919920444488525, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9798671007156372, 'rewards/repeat_penalty/std': 0.04848518967628479, 'rewards/repeat_floor/mean': 0.9976696372032166, 'rewards/repeat_floor/std': 0.005331065040081739, 'rewards/near_duplicate_penalty/mean': 0.9921566247940063, 'rewards/near_duplicate_penalty/std': 0.014121531508862972, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.22261124849319458, 'rewards/total_composite/std': 0.18416748940944672, 'reward': 0.22261124849319458, 'reward_std': 0.18416748940944672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2021288424730301, 'sampling/sampling_logp_difference/max': 1.7195062637329102, 'sampling/importance_sampling_ratio/min': 0.17915457487106323, 'sampling/importance_sampling_ratio/mean': 1.0501278638839722, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.273718625307083, 'clip_ratio/low_mean': 0.04993602354079485, 'clip_ratio/low_min': 0.04993602354079485, 'clip_ratio/high_mean': 0.08595150709152222, 'clip_ratio/high_max': 0.08595150709152222, 'clip_ratio/region_mean': 0.13588753063231707, 'reward_total_mean': 0.22261124849319458, 'reward_meter_mean': 0.6374885439872742, 'reward_meter_std': 0.3750973641872406, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.8109548091888428, 'reward_lexical_plausibility_std': 0.31919920444488525, 'reward_repeat_penalty_mean': 0.9798671007156372, 'reward_repeat_penalty_std': 0.04848518967628479, 'reward_repeat_floor_mean': 0.9976696372032166, 'reward_repeat_floor_std': 0.005331065040081739, 'reward_near_duplicate_penalty_mean': 0.9921566247940063, 'reward_near_duplicate_penalty_std': 0.014121531508862972, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.22261124849319458, 'reward_total_composite_std': 0.18416748940944672, 'epoch': 0.01}

  6%|▌         | 189/3300 [39:14<10:11:50, 11.80s/it]INFO 04-12 16:37:41 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 190/3300 [39:25<9:55:41, 11.49s/it] 2026-04-12 16:37:51,703 | INFO | train_grpo_train | metrics_logged mode=train step=190

                                                    
{'loss': 0.1348, 'grad_norm': 11.683950424194336, 'learning_rate': 9.427272727272728e-06, 'num_tokens': 368161.0, 'completions/mean_length': 75.875, 'completions/min_length': 47.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.5286248326301575, 'rewards/meter/std': 0.42756661772727966, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9482710957527161, 'rewards/lexical_plausibility/std': 0.09784351289272308, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9863206148147583, 'rewards/repeat_penalty/std': 0.025610435754060745, 'rewards/repeat_floor/mean': 0.9983108639717102, 'rewards/repeat_floor/std': 0.00290672411210835, 'rewards/near_duplicate_penalty/mean': 0.9925850629806519, 'rewards/near_duplicate_penalty/std': 0.011060180142521858, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9935897588729858, 'rewards/distinct_2/std': 0.018130943179130554, 'rewards/total_composite/mean': 0.17522484064102173, 'rewards/total_composite/std': 0.14982382953166962, 'reward': 0.17522484064102173, 'reward_std': 0.14982382953166962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18190455436706543, 'sampling/sampling_logp_difference/max': 1.6534295082092285, 'sampling/importance_sampling_ratio/min': 0.19139240682125092, 'sampling/importance_sampling_ratio/mean': 1.0633751153945923, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.705846443772316, 'clip_ratio/low_mean': 0.06471885740756989, 'clip_ratio/low_min': 0.06471885740756989, 'clip_ratio/high_mean': 0.06607464654371142, 'clip_ratio/high_max': 0.06607464654371142, 'clip_ratio/region_mean': 0.1307935039512813, 'reward_total_mean': 0.17522484064102173, 'reward_meter_mean': 0.5286248326301575, 'reward_meter_std': 0.42756661772727966, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9482710957527161, 'reward_lexical_plausibility_std': 0.09784351289272308, 'reward_repeat_penalty_mean': 0.9863206148147583, 'reward_repeat_penalty_std': 0.025610435754060745, 'reward_repeat_floor_mean': 0.9983108639717102, 'reward_repeat_floor_std': 0.00290672411210835, 'reward_near_duplicate_penalty_mean': 0.9925850629806519, 'reward_near_duplicate_penalty_std': 0.011060180142521858, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9935897588729858, 'reward_distinct_2_std': 0.018130943179130554, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.17522484064102173, 'reward_total_composite_std': 0.14982382953166962, 'epoch': 0.01}

  6%|▌         | 190/3300 [39:25<9:55:41, 11.49s/it]INFO 04-12 16:37:51 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 191/3300 [39:33<9:04:52, 10.52s/it]2026-04-12 16:37:59,943 | INFO | train_grpo_train | metrics_logged mode=train step=191

                                                    
{'loss': 0.0416, 'grad_norm': 28.455045700073242, 'learning_rate': 9.424242424242425e-06, 'num_tokens': 369609.0, 'completions/mean_length': 20.0, 'completions/min_length': 16.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5563498139381409, 'rewards/meter/std': 0.4520806670188904, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.781009316444397, 'rewards/lexical_plausibility/std': 0.1717691421508789, 'rewards/judge_quality/mean': 0.4050000309944153, 'rewards/judge_quality/std': 0.33713075518608093, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14113321900367737, 'rewards/total_composite/std': 0.10498090833425522, 'reward': 0.14113321900367737, 'reward_std': 0.10498090088367462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.28368139266967773, 'sampling/sampling_logp_difference/max': 1.2664527893066406, 'sampling/importance_sampling_ratio/min': 0.2818295657634735, 'sampling/importance_sampling_ratio/mean': 1.0669454336166382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.475811317563057, 'clip_ratio/low_mean': 0.04173913039267063, 'clip_ratio/low_min': 0.04173913039267063, 'clip_ratio/high_mean': 0.12946121580898762, 'clip_ratio/high_max': 0.12946121580898762, 'clip_ratio/region_mean': 0.17120034620165825, 'reward_total_mean': 0.14113321900367737, 'reward_meter_mean': 0.5563498139381409, 'reward_meter_std': 0.4520806670188904, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.781009316444397, 'reward_lexical_plausibility_std': 0.1717691421508789, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4050000309944153, 'reward_judge_quality_std': 0.33713075518608093, 'reward_total_composite_mean': 0.14113321900367737, 'reward_total_composite_std': 0.10498090833425522, 'epoch': 0.01}

  6%|▌         | 191/3300 [39:33<9:04:52, 10.52s/it]INFO 04-12 16:38:00 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 192/3300 [39:47<9:51:58, 11.43s/it]2026-04-12 16:38:13,495 | INFO | train_grpo_train | metrics_logged mode=train step=192

                                                    
{'loss': 0.0328, 'grad_norm': 2.036179780960083, 'learning_rate': 9.421212121212122e-06, 'num_tokens': 371545.0, 'completions/mean_length': 390.0, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 186.6666717529297, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 447.0, 'rewards/meter/mean': 0.44524112343788147, 'rewards/meter/std': 0.3842077851295471, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 0.5221306085586548, 'rewards/lexical_plausibility/std': 0.314175009727478, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.9935352802276611, 'rewards/repeat_penalty/std': 0.012585675343871117, 'rewards/repeat_floor/mean': 0.9990302920341492, 'rewards/repeat_floor/std': 0.0018878512782976031, 'rewards/near_duplicate_penalty/mean': 0.9935352802276611, 'rewards/near_duplicate_penalty/std': 0.012585675343871117, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10575045645236969, 'rewards/total_composite/std': 0.1727677285671234, 'reward': 0.10575045645236969, 'reward_std': 0.17276771366596222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17472466826438904, 'sampling/sampling_logp_difference/max': 1.1510963439941406, 'sampling/importance_sampling_ratio/min': 0.3162898123264313, 'sampling/importance_sampling_ratio/mean': 1.015725016593933, 'sampling/importance_sampling_ratio/max': 1.9650384187698364, 'entropy': 1.4540944546461105, 'clip_ratio/low_mean': 0.007550335489213467, 'clip_ratio/low_min': 0.007550335489213467, 'clip_ratio/high_mean': 0.04625237174332142, 'clip_ratio/high_max': 0.04625237174332142, 'clip_ratio/region_mean': 0.053802707232534885, 'reward_total_mean': 0.10575045645236969, 'reward_meter_mean': 0.44524112343788147, 'reward_meter_std': 0.3842077851295471, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 0.5221306085586548, 'reward_lexical_plausibility_std': 0.314175009727478, 'reward_repeat_penalty_mean': 0.9935352802276611, 'reward_repeat_penalty_std': 0.012585675343871117, 'reward_repeat_floor_mean': 0.9990302920341492, 'reward_repeat_floor_std': 0.0018878512782976031, 'reward_near_duplicate_penalty_mean': 0.9935352802276611, 'reward_near_duplicate_penalty_std': 0.012585675343871117, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.10575045645236969, 'reward_total_composite_std': 0.1727677285671234, 'epoch': 0.01}

  6%|▌         | 192/3300 [39:47<9:51:58, 11.43s/it]INFO 04-12 16:38:13 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 193/3300 [40:00<10:22:42, 12.03s/it]2026-04-12 16:38:26,913 | INFO | train_grpo_train | metrics_logged mode=train step=193

                                                     
{'loss': 0.0069, 'grad_norm': 2.518564224243164, 'learning_rate': 9.418181818181818e-06, 'num_tokens': 373504.0, 'completions/mean_length': 327.875, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 143.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 362.0, 'rewards/meter/mean': 0.21626828610897064, 'rewards/meter/std': 0.16922105848789215, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.33905068039894104, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.10171521455049515, 'rewards/lexical_plausibility/mean': 0.6645095348358154, 'rewards/lexical_plausibility/std': 0.28426334261894226, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9442843794822693, 'rewards/repeat_penalty/std': 0.12373539805412292, 'rewards/repeat_floor/mean': 0.9955518245697021, 'rewards/repeat_floor/std': 0.009132664650678635, 'rewards/near_duplicate_penalty/mean': 0.9913046360015869, 'rewards/near_duplicate_penalty/std': 0.017040900886058807, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9650679230690002, 'rewards/distinct_2/std': 0.07423978298902512, 'rewards/total_composite/mean': 0.03105197474360466, 'rewards/total_composite/std': 0.03759218379855156, 'reward': 0.03105197474360466, 'reward_std': 0.03759218752384186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19447876513004303, 'sampling/sampling_logp_difference/max': 1.335679292678833, 'sampling/importance_sampling_ratio/min': 0.2629794776439667, 'sampling/importance_sampling_ratio/mean': 1.0487233400344849, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.333795487880707, 'clip_ratio/low_mean': 0.006560773588716984, 'clip_ratio/low_min': 0.006560773588716984, 'clip_ratio/high_mean': 0.0716051533818245, 'clip_ratio/high_max': 0.0716051533818245, 'clip_ratio/region_mean': 0.07816592697054148, 'reward_total_mean': 0.03105197474360466, 'reward_meter_mean': 0.21626828610897064, 'reward_meter_std': 0.16922105848789215, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.33905068039894104, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.10171521455049515, 'reward_lexical_plausibility_mean': 0.6645095348358154, 'reward_lexical_plausibility_std': 0.28426334261894226, 'reward_repeat_penalty_mean': 0.9442843794822693, 'reward_repeat_penalty_std': 0.12373539805412292, 'reward_repeat_floor_mean': 0.9955518245697021, 'reward_repeat_floor_std': 0.009132664650678635, 'reward_near_duplicate_penalty_mean': 0.9913046360015869, 'reward_near_duplicate_penalty_std': 0.017040900886058807, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9650679230690002, 'reward_distinct_2_std': 0.07423978298902512, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.03105197474360466, 'reward_total_composite_std': 0.03759218379855156, 'epoch': 0.01}

  6%|▌         | 193/3300 [40:00<10:22:42, 12.03s/it]INFO 04-12 16:38:27 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 194/3300 [40:13<10:36:24, 12.29s/it]2026-04-12 16:38:39,833 | INFO | train_grpo_train | metrics_logged mode=train step=194

                                                     
{'loss': -0.0852, 'grad_norm': 3.6086113452911377, 'learning_rate': 9.415151515151515e-06, 'num_tokens': 375089.0, 'completions/mean_length': 169.125, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 54.833335876464844, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5615483522415161, 'rewards/meter/std': 0.36815136671066284, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.8159192204475403, 'rewards/lexical_plausibility/std': 0.22040477395057678, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.2840491831302643, 'rewards/repeat_penalty/mean': 0.9941784143447876, 'rewards/repeat_penalty/std': 0.008796388283371925, 'rewards/repeat_floor/mean': 0.9991267323493958, 'rewards/repeat_floor/std': 0.0013194619677960873, 'rewards/near_duplicate_penalty/mean': 0.9941784143447876, 'rewards/near_duplicate_penalty/std': 0.008796388283371925, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17255161702632904, 'rewards/total_composite/std': 0.14791880548000336, 'reward': 0.17255161702632904, 'reward_std': 0.14791880548000336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22499358654022217, 'sampling/sampling_logp_difference/max': 1.6198673248291016, 'sampling/importance_sampling_ratio/min': 0.19792495667934418, 'sampling/importance_sampling_ratio/mean': 1.0764914751052856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8485622704029083, 'clip_ratio/low_mean': 0.06984508596360683, 'clip_ratio/low_min': 0.06984508596360683, 'clip_ratio/high_mean': 0.09305718168616295, 'clip_ratio/high_max': 0.09305718168616295, 'clip_ratio/region_mean': 0.16290226764976978, 'reward_total_mean': 0.17255161702632904, 'reward_meter_mean': 0.5615483522415161, 'reward_meter_std': 0.36815136671066284, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.8159192204475403, 'reward_lexical_plausibility_std': 0.22040477395057678, 'reward_repeat_penalty_mean': 0.9941784143447876, 'reward_repeat_penalty_std': 0.008796388283371925, 'reward_repeat_floor_mean': 0.9991267323493958, 'reward_repeat_floor_std': 0.0013194619677960873, 'reward_near_duplicate_penalty_mean': 0.9941784143447876, 'reward_near_duplicate_penalty_std': 0.008796388283371925, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.2840491831302643, 'reward_total_composite_mean': 0.17255161702632904, 'reward_total_composite_std': 0.14791880548000336, 'epoch': 0.01}

  6%|▌         | 194/3300 [40:13<10:36:24, 12.29s/it]INFO 04-12 16:38:40 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 195/3300 [40:23<9:54:10, 11.48s/it] 2026-04-12 16:38:49,421 | INFO | train_grpo_train | metrics_logged mode=train step=195

                                                    
{'loss': 0.0871, 'grad_norm': 15.280990600585938, 'learning_rate': 9.412121212121212e-06, 'num_tokens': 376659.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.4135478734970093, 'rewards/meter/std': 0.33901268243789673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6387500166893005, 'rewards/judge_quality/std': 0.3224211037158966, 'rewards/repeat_penalty/mean': 0.9999389052391052, 'rewards/repeat_penalty/std': 0.00017280207248404622, 'rewards/repeat_floor/mean': 0.9999908208847046, 'rewards/repeat_floor/std': 2.5920317057170905e-05, 'rewards/near_duplicate_penalty/mean': 0.9999389052391052, 'rewards/near_duplicate_penalty/std': 0.00017280207248404622, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.219503253698349, 'rewards/total_composite/std': 0.13908927142620087, 'reward': 0.219503253698349, 'reward_std': 0.13908925652503967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2435743510723114, 'sampling/sampling_logp_difference/max': 2.39784574508667, 'sampling/importance_sampling_ratio/min': 0.09091359376907349, 'sampling/importance_sampling_ratio/mean': 1.045608401298523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0363870710134506, 'clip_ratio/low_mean': 0.0820359718054533, 'clip_ratio/low_min': 0.0820359718054533, 'clip_ratio/high_mean': 0.16588432900607586, 'clip_ratio/high_max': 0.16588432900607586, 'clip_ratio/region_mean': 0.24792030081152916, 'reward_total_mean': 0.219503253698349, 'reward_meter_mean': 0.4135478734970093, 'reward_meter_std': 0.33901268243789673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9999389052391052, 'reward_repeat_penalty_std': 0.00017280207248404622, 'reward_repeat_floor_mean': 0.9999908208847046, 'reward_repeat_floor_std': 2.5920317057170905e-05, 'reward_near_duplicate_penalty_mean': 0.9999389052391052, 'reward_near_duplicate_penalty_std': 0.00017280207248404622, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6387500166893005, 'reward_judge_quality_std': 0.3224211037158966, 'reward_total_composite_mean': 0.219503253698349, 'reward_total_composite_std': 0.13908927142620087, 'epoch': 0.01}

  6%|▌         | 195/3300 [40:23<9:54:10, 11.48s/it]INFO 04-12 16:38:49 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 196/3300 [40:37<10:30:01, 12.18s/it]2026-04-12 16:39:03,225 | INFO | train_grpo_train | metrics_logged mode=train step=196

                                                     
{'loss': -0.0359, 'grad_norm': 0.9217419028282166, 'learning_rate': 9.40909090909091e-06, 'num_tokens': 378148.0, 'completions/mean_length': 458.125, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.4111250936985016, 'rewards/meter/std': 0.2979764938354492, 'rewards/count_adherence/mean': 0.5250000357627869, 'rewards/count_adherence/std': 0.23754701018333435, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8574999570846558, 'rewards/count_floor/std': 0.07126409560441971, 'rewards/lexical_plausibility/mean': 0.5859890580177307, 'rewards/lexical_plausibility/std': 0.1842985302209854, 'rewards/judge_quality/mean': 0.0625, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9353363513946533, 'rewards/repeat_penalty/std': 0.11443106085062027, 'rewards/repeat_floor/mean': 0.995958685874939, 'rewards/repeat_floor/std': 0.0067739044316112995, 'rewards/near_duplicate_penalty/mean': 0.998394250869751, 'rewards/near_duplicate_penalty/std': 0.002028181916102767, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9994393587112427, 'rewards/distinct_2/std': 0.0015857962425798178, 'rewards/total_composite/mean': 0.021751247346401215, 'rewards/total_composite/std': 0.016394687816500664, 'reward': 0.021751247346401215, 'reward_std': 0.016394687816500664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22108183801174164, 'sampling/sampling_logp_difference/max': 0.7847318649291992, 'sampling/importance_sampling_ratio/min': 0.5413163304328918, 'sampling/importance_sampling_ratio/mean': 1.1084215641021729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6079707741737366, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.023148147389292717, 'clip_ratio/high_max': 0.023148147389292717, 'clip_ratio/region_mean': 0.023148147389292717, 'reward_total_mean': 0.021751247346401215, 'reward_meter_mean': 0.4111250936985016, 'reward_meter_std': 0.2979764938354492, 'reward_count_adherence_mean': 0.5250000357627869, 'reward_count_adherence_std': 0.23754701018333435, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8574999570846558, 'reward_count_floor_std': 0.07126409560441971, 'reward_lexical_plausibility_mean': 0.5859890580177307, 'reward_lexical_plausibility_std': 0.1842985302209854, 'reward_repeat_penalty_mean': 0.9353363513946533, 'reward_repeat_penalty_std': 0.11443106085062027, 'reward_repeat_floor_mean': 0.995958685874939, 'reward_repeat_floor_std': 0.0067739044316112995, 'reward_near_duplicate_penalty_mean': 0.998394250869751, 'reward_near_duplicate_penalty_std': 0.002028181916102767, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9994393587112427, 'reward_distinct_2_std': 0.0015857962425798178, 'reward_judge_quality_mean': 0.0625, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.021751247346401215, 'reward_total_composite_std': 0.016394687816500664, 'epoch': 0.01}

  6%|▌         | 196/3300 [40:37<10:30:01, 12.18s/it]INFO 04-12 16:39:03 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 197/3300 [40:51<11:00:41, 12.78s/it]2026-04-12 16:39:17,394 | INFO | train_grpo_train | metrics_logged mode=train step=197

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.406060606060607e-06, 'num_tokens': 380044.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.2945672869682312, 'rewards/meter/std': 0.2733685374259949, 'rewards/count_adherence/mean': 0.4444444477558136, 'rewards/count_adherence/std': 0.29095718264579773, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8333333134651184, 'rewards/count_floor/std': 0.08728715032339096, 'rewards/lexical_plausibility/mean': 0.5796335339546204, 'rewards/lexical_plausibility/std': 0.05944741517305374, 'rewards/judge_quality/mean': 0.05625000223517418, 'rewards/judge_quality/std': 0.04172614961862564, 'rewards/repeat_penalty/mean': 0.9643219709396362, 'rewards/repeat_penalty/std': 0.08672906458377838, 'rewards/repeat_floor/mean': 0.99753338098526, 'rewards/repeat_floor/std': 0.005097570829093456, 'rewards/near_duplicate_penalty/mean': 0.9967884421348572, 'rewards/near_duplicate_penalty/std': 0.0035242505837231874, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9987789988517761, 'rewards/distinct_2/std': 0.0034535129088908434, 'rewards/total_composite/mean': 0.01453713420778513, 'rewards/total_composite/std': 0.014201870188117027, 'reward': 0.01453713420778513, 'reward_std': 0.014201869256794453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.01453713420778513, 'reward_meter_mean': 0.2945672869682312, 'reward_meter_std': 0.2733685374259949, 'reward_count_adherence_mean': 0.4444444477558136, 'reward_count_adherence_std': 0.29095718264579773, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8333333134651184, 'reward_count_floor_std': 0.08728715032339096, 'reward_lexical_plausibility_mean': 0.5796335339546204, 'reward_lexical_plausibility_std': 0.05944741517305374, 'reward_repeat_penalty_mean': 0.9643219709396362, 'reward_repeat_penalty_std': 0.08672906458377838, 'reward_repeat_floor_mean': 0.99753338098526, 'reward_repeat_floor_std': 0.005097570829093456, 'reward_near_duplicate_penalty_mean': 0.9967884421348572, 'reward_near_duplicate_penalty_std': 0.0035242505837231874, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9987789988517761, 'reward_distinct_2_std': 0.0034535129088908434, 'reward_judge_quality_mean': 0.05625000223517418, 'reward_judge_quality_std': 0.04172614961862564, 'reward_total_composite_mean': 0.01453713420778513, 'reward_total_composite_std': 0.014201870188117027, 'epoch': 0.01}

  6%|▌         | 197/3300 [40:51<11:00:41, 12.78s/it]INFO 04-12 16:39:17 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 198/3300 [41:05<11:21:30, 13.18s/it]2026-04-12 16:39:31,535 | INFO | train_grpo_train | metrics_logged mode=train step=198

                                                     
{'loss': -0.1471, 'grad_norm': 2.7665464878082275, 'learning_rate': 9.403030303030304e-06, 'num_tokens': 382011.0, 'completions/mean_length': 309.875, 'completions/min_length': 83.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 107.75, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.667330265045166, 'rewards/meter/std': 0.3395802676677704, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.29154759645462036, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.08746428042650223, 'rewards/lexical_plausibility/mean': 0.7903605699539185, 'rewards/lexical_plausibility/std': 0.2222806066274643, 'rewards/judge_quality/mean': 0.11875000596046448, 'rewards/judge_quality/std': 0.1099918857216835, 'rewards/repeat_penalty/mean': 0.9035739302635193, 'rewards/repeat_penalty/std': 0.1521126925945282, 'rewards/repeat_floor/mean': 0.9901407957077026, 'rewards/repeat_floor/std': 0.015917491167783737, 'rewards/near_duplicate_penalty/mean': 0.9693999290466309, 'rewards/near_duplicate_penalty/std': 0.03903765231370926, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9497872591018677, 'rewards/distinct_2/std': 0.09576201438903809, 'rewards/total_composite/mean': 0.09143868088722229, 'rewards/total_composite/std': 0.1126495972275734, 'reward': 0.09143868088722229, 'reward_std': 0.1126495897769928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19507019221782684, 'sampling/sampling_logp_difference/max': 1.13051176071167, 'sampling/importance_sampling_ratio/min': 0.3228679895401001, 'sampling/importance_sampling_ratio/mean': 1.0543478727340698, 'sampling/importance_sampling_ratio/max': 1.9720526933670044, 'entropy': 1.7003357410430908, 'clip_ratio/low_mean': 0.017326733097434044, 'clip_ratio/low_min': 0.017326733097434044, 'clip_ratio/high_mean': 0.07335328683257103, 'clip_ratio/high_max': 0.07335328683257103, 'clip_ratio/region_mean': 0.09068001993000507, 'reward_total_mean': 0.09143868088722229, 'reward_meter_mean': 0.667330265045166, 'reward_meter_std': 0.3395802676677704, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.29154759645462036, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.08746428042650223, 'reward_lexical_plausibility_mean': 0.7903605699539185, 'reward_lexical_plausibility_std': 0.2222806066274643, 'reward_repeat_penalty_mean': 0.9035739302635193, 'reward_repeat_penalty_std': 0.1521126925945282, 'reward_repeat_floor_mean': 0.9901407957077026, 'reward_repeat_floor_std': 0.015917491167783737, 'reward_near_duplicate_penalty_mean': 0.9693999290466309, 'reward_near_duplicate_penalty_std': 0.03903765231370926, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9497872591018677, 'reward_distinct_2_std': 0.09576201438903809, 'reward_judge_quality_mean': 0.11875000596046448, 'reward_judge_quality_std': 0.1099918857216835, 'reward_total_composite_mean': 0.09143868088722229, 'reward_total_composite_std': 0.1126495972275734, 'epoch': 0.01}

  6%|▌         | 198/3300 [41:05<11:21:30, 13.18s/it]INFO 04-12 16:39:31 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 199/3300 [41:13<10:03:28, 11.68s/it]2026-04-12 16:39:39,687 | INFO | train_grpo_train | metrics_logged mode=train step=199

                                                     
{'loss': 0.0309, 'grad_norm': 21.764482498168945, 'learning_rate': 9.4e-06, 'num_tokens': 383534.0, 'completions/mean_length': 30.375, 'completions/min_length': 28.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.3349902033805847, 'rewards/meter/std': 0.41731929779052734, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.9688562750816345, 'rewards/repeat_penalty/std': 0.047484010457992554, 'rewards/repeat_floor/mean': 0.9958813190460205, 'rewards/repeat_floor/std': 0.00577416829764843, 'rewards/near_duplicate_penalty/mean': 0.978952169418335, 'rewards/near_duplicate_penalty/std': 0.025598246604204178, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.17050465941429138, 'rewards/total_composite/std': 0.2202007919549942, 'reward': 0.17050465941429138, 'reward_std': 0.2202007919549942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19891545176506042, 'sampling/sampling_logp_difference/max': 1.0123915672302246, 'sampling/importance_sampling_ratio/min': 0.3748926818370819, 'sampling/importance_sampling_ratio/mean': 1.0575103759765625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8103185594081879, 'clip_ratio/low_mean': 0.08342103566974401, 'clip_ratio/low_min': 0.08342103566974401, 'clip_ratio/high_mean': 0.044728269102051854, 'clip_ratio/high_max': 0.044728269102051854, 'clip_ratio/region_mean': 0.12814930477179587, 'reward_total_mean': 0.17050465941429138, 'reward_meter_mean': 0.3349902033805847, 'reward_meter_std': 0.41731929779052734, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9688562750816345, 'reward_repeat_penalty_std': 0.047484010457992554, 'reward_repeat_floor_mean': 0.9958813190460205, 'reward_repeat_floor_std': 0.00577416829764843, 'reward_near_duplicate_penalty_mean': 0.978952169418335, 'reward_near_duplicate_penalty_std': 0.025598246604204178, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.17050465941429138, 'reward_total_composite_std': 0.2202007919549942, 'epoch': 0.01}

  6%|▌         | 199/3300 [41:13<10:03:28, 11.68s/it]INFO 04-12 16:39:39 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 200/3300 [41:21<9:05:59, 10.57s/it] 2026-04-12 16:39:47,669 | INFO | train_grpo_train | metrics_logged mode=train step=200

                                                    
{'loss': 0.0221, 'grad_norm': 16.82535743713379, 'learning_rate': 9.396969696969697e-06, 'num_tokens': 385013.0, 'completions/mean_length': 36.875, 'completions/min_length': 32.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.25020644068717957, 'rewards/meter/std': 0.3476031422615051, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9369837045669556, 'rewards/lexical_plausibility/std': 0.11954433470964432, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9960905313491821, 'rewards/repeat_penalty/std': 0.004692242946475744, 'rewards/repeat_floor/mean': 0.9994135499000549, 'rewards/repeat_floor/std': 0.0007038420881144702, 'rewards/near_duplicate_penalty/mean': 0.9960905313491821, 'rewards/near_duplicate_penalty/std': 0.004692242946475744, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11283959448337555, 'rewards/total_composite/std': 0.1497221738100052, 'reward': 0.11283959448337555, 'reward_std': 0.1497221738100052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23476119339466095, 'sampling/sampling_logp_difference/max': 2.3389034271240234, 'sampling/importance_sampling_ratio/min': 0.09643332660198212, 'sampling/importance_sampling_ratio/mean': 1.0340209007263184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.674425408244133, 'clip_ratio/low_mean': 0.06655457336455584, 'clip_ratio/low_min': 0.06655457336455584, 'clip_ratio/high_mean': 0.05685078911483288, 'clip_ratio/high_max': 0.05685078911483288, 'clip_ratio/region_mean': 0.12340536247938871, 'reward_total_mean': 0.11283959448337555, 'reward_meter_mean': 0.25020644068717957, 'reward_meter_std': 0.3476031422615051, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9369837045669556, 'reward_lexical_plausibility_std': 0.11954433470964432, 'reward_repeat_penalty_mean': 0.9960905313491821, 'reward_repeat_penalty_std': 0.004692242946475744, 'reward_repeat_floor_mean': 0.9994135499000549, 'reward_repeat_floor_std': 0.0007038420881144702, 'reward_near_duplicate_penalty_mean': 0.9960905313491821, 'reward_near_duplicate_penalty_std': 0.004692242946475744, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.11283959448337555, 'reward_total_composite_std': 0.1497221738100052, 'epoch': 0.01}

  6%|▌         | 200/3300 [41:21<9:05:59, 10.57s/it]INFO 04-12 16:39:47 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:04,  5.88s/it]

 23%|██▎       | 3/13 [00:23<01:22,  8.27s/it]

 31%|███       | 4/13 [00:35<01:26,  9.59s/it]

 38%|███▊      | 5/13 [00:46<01:22, 10.29s/it]

 46%|████▌     | 6/13 [00:58<01:14, 10.60s/it]

 54%|█████▍    | 7/13 [01:09<01:05, 10.84s/it]

 62%|██████▏   | 8/13 [01:21<00:56, 11.31s/it]

 69%|██████▉   | 9/13 [01:33<00:45, 11.48s/it]

 77%|███████▋  | 10/13 [01:44<00:34, 11.38s/it]

 85%|████████▍ | 11/13 [01:56<00:23, 11.65s/it]

 92%|█████████▏| 12/13 [02:08<00:11, 11.60s/it]

100%|██████████| 13/13 [02:19<00:00, 11.44s/it]2026-04-12 16:42:18,733 | INFO | train_grpo_train | metrics_logged mode=eval step=200

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 151.0431, 'eval_samples_per_second': 0.689, 'eval_steps_per_second': 0.086, 'eval_num_tokens': 385013.0, 'eval_completions/mean_length': 288.61538461538464, 'eval_completions/min_length': 33.46153846153846, 'eval_completions/max_length': 512.0, 'eval_completions/clipped_ratio': 0.46153846153846156, 'eval_completions/mean_terminated_length': 93.1934080857497, 'eval_completions/min_terminated_length': 33.46153846153846, 'eval_completions/max_terminated_length': 188.46153846153845, 'eval_rewards/meter/mean': 0.27145538880274844, 'eval_rewards/meter/std': 0.3038664013147354, 'eval_rewards/count_adherence/mean': 0.7206993744923518, 'eval_rewards/count_adherence/std': 0.30023181667694676, 'eval_rewards/arabic_clean/mean': 0.5384615384615384, 'eval_rewards/arabic_clean/std': 0.49203900878246015, 'eval_rewards/hard_gate/mean': 0.9807692307692307, 'eval_rewards/hard_gate/std': 0.05439282839114849, 'eval_rewards/arabic_floor/mean': 0.9956730741720933, 'eval_rewards/arabic_floor/std': 0.009420730173587799, 'eval_rewards/count_floor/mean': 0.9162098169326782, 'eval_rewards/count_floor/std': 0.09006954099123295, 'eval_rewards/lexical_plausibility/mean': 0.8058967177684491, 'eval_rewards/lexical_plausibility/std': 0.18502292667443937, 'eval_rewards/judge_quality/mean': 0.21432691869827417, 'eval_rewards/judge_quality/std': 0.19979697867081717, 'eval_rewards/repeat_penalty/mean': 0.9533474032695477, 'eval_rewards/repeat_penalty/std': 0.08043498184102085, 'eval_rewards/repeat_floor/mean': 0.9955630073180566, 'eval_rewards/repeat_floor/std': 0.007148864777543797, 'eval_rewards/near_duplicate_penalty/mean': 0.9879688849815955, 'eval_rewards/near_duplicate_penalty/std': 0.01758513988282245, 'eval_rewards/opening_diversity/mean': 0.975600962455456, 'eval_rewards/opening_diversity/std': 0.061748841061041906, 'eval_rewards/distinct_2/mean': 0.9870180625181931, 'eval_rewards/distinct_2/std': 0.02635298454417632, 'eval_rewards/total_composite/mean': 0.05888123573878637, 'eval_rewards/total_composite/std': 0.0850732745602727, 'eval_reward': 0.05888123573878637, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.16237259484254396, 'eval_sampling/sampling_logp_difference/max': 1.082637016589825, 'eval_sampling/importance_sampling_ratio/min': 0.35368941265803117, 'eval_sampling/importance_sampling_ratio/mean': 1.0550401760981634, 'eval_sampling/importance_sampling_ratio/max': 1.7244445177224965, 'eval_entropy': 3.61222793505742, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.05888123573878637, 'eval_reward_meter_mean': 0.27145538880274844, 'eval_reward_meter_std': 0.3038664013147354, 'eval_reward_count_adherence_mean': 0.7206993744923518, 'eval_reward_count_adherence_std': 0.30023181667694676, 'eval_reward_arabic_clean_mean': 0.5384615384615384, 'eval_reward_arabic_clean_std': 0.49203900878246015, 'eval_reward_hard_gate_mean': 0.9807692307692307, 'eval_reward_hard_gate_std': 0.05439282839114849, 'eval_reward_arabic_floor_mean': 0.9956730741720933, 'eval_reward_arabic_floor_std': 0.009420730173587799, 'eval_reward_count_floor_mean': 0.9162098169326782, 'eval_reward_count_floor_std': 0.09006954099123295, 'eval_reward_lexical_plausibility_mean': 0.8058967177684491, 'eval_reward_lexical_plausibility_std': 0.18502292667443937, 'eval_reward_repeat_penalty_mean': 0.9533474032695477, 'eval_reward_repeat_penalty_std': 0.08043498184102085, 'eval_reward_repeat_floor_mean': 0.9955630073180566, 'eval_reward_repeat_floor_std': 0.007148864777543797, 'eval_reward_near_duplicate_penalty_mean': 0.9879688849815955, 'eval_reward_near_duplicate_penalty_std': 0.01758513988282245, 'eval_reward_opening_diversity_mean': 0.975600962455456, 'eval_reward_opening_diversity_std': 0.061748841061041906, 'eval_reward_distinct_2_mean': 0.9870180625181931, 'eval_reward_distinct_2_std': 0.02635298454417632, 'eval_reward_judge_quality_mean': 0.21432691869827417, 'eval_reward_judge_quality_std': 0.19979697867081717, 'eval_reward_total_composite_mean': 0.05888123573878637, 'eval_reward_total_composite_std': 0.0850732745602727, 'epoch': 0.01}

  6%|▌         | 200/3300 [43:52<9:05:59, 10.57s/it]

100%|██████████| 13/13 [02:19<00:00, 11.44s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  6%|▌         | 201/3300 [44:03<48:13:41, 56.03s/it]2026-04-12 16:42:29,762 | INFO | train_grpo_train | metrics_logged mode=train step=201

                                                     
{'loss': 0.0456, 'grad_norm': 22.318904876708984, 'learning_rate': 9.393939393939396e-06, 'num_tokens': 386583.0, 'completions/mean_length': 34.25, 'completions/min_length': 30.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6802480220794678, 'rewards/meter/std': 0.38063785433769226, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.9978264570236206, 'rewards/repeat_penalty/std': 0.003853553906083107, 'rewards/repeat_floor/mean': 0.9996739625930786, 'rewards/repeat_floor/std': 0.000578021805267781, 'rewards/near_duplicate_penalty/mean': 0.9978264570236206, 'rewards/near_duplicate_penalty/std': 0.003853553906083107, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2922589182853699, 'rewards/total_composite/std': 0.16907870769500732, 'reward': 0.2922589182853699, 'reward_std': 0.16907872259616852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21440432965755463, 'sampling/sampling_logp_difference/max': 1.2445874214172363, 'sampling/importance_sampling_ratio/min': 0.2880597412586212, 'sampling/importance_sampling_ratio/mean': 1.0453404188156128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4124284982681274, 'clip_ratio/low_mean': 0.03623949736356735, 'clip_ratio/low_min': 0.03623949736356735, 'clip_ratio/high_mean': 0.1460014395415783, 'clip_ratio/high_max': 0.1460014395415783, 'clip_ratio/region_mean': 0.18224093690514565, 'reward_total_mean': 0.2922589182853699, 'reward_meter_mean': 0.6802480220794678, 'reward_meter_std': 0.38063785433769226, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9978264570236206, 'reward_repeat_penalty_std': 0.003853553906083107, 'reward_repeat_floor_mean': 0.9996739625930786, 'reward_repeat_floor_std': 0.000578021805267781, 'reward_near_duplicate_penalty_mean': 0.9978264570236206, 'reward_near_duplicate_penalty_std': 0.003853553906083107, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.2922589182853699, 'reward_total_composite_std': 0.16907870769500732, 'epoch': 0.01}

  6%|▌         | 201/3300 [44:03<48:13:41, 56.03s/it]INFO 04-12 16:42:30 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 202/3300 [44:17<37:20:53, 43.40s/it]2026-04-12 16:42:43,702 | INFO | train_grpo_train | metrics_logged mode=train step=202

                                                     
{'loss': -0.0311, 'grad_norm': 3.6717519760131836, 'learning_rate': 9.390909090909092e-06, 'num_tokens': 387989.0, 'completions/mean_length': 152.75, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.156070739030838, 'rewards/meter/std': 0.217662513256073, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9246528148651123, 'rewards/lexical_plausibility/std': 0.14081719517707825, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.946195125579834, 'rewards/repeat_penalty/std': 0.09871681034564972, 'rewards/repeat_floor/mean': 0.9936034679412842, 'rewards/repeat_floor/std': 0.011640418320894241, 'rewards/near_duplicate_penalty/mean': 0.9784188866615295, 'rewards/near_duplicate_penalty/std': 0.03865911811590195, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9648962020874023, 'rewards/distinct_2/std': 0.06701616942882538, 'rewards/total_composite/mean': 0.059857502579689026, 'rewards/total_composite/std': 0.0954614207148552, 'reward': 0.059857502579689026, 'reward_std': 0.0954614207148552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23309345543384552, 'sampling/sampling_logp_difference/max': 0.9677591323852539, 'sampling/importance_sampling_ratio/min': 0.3799334764480591, 'sampling/importance_sampling_ratio/mean': 1.0754752159118652, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1340361535549164, 'clip_ratio/low_mean': 0.11528520844876766, 'clip_ratio/low_min': 0.11528520844876766, 'clip_ratio/high_mean': 0.04817708395421505, 'clip_ratio/high_max': 0.04817708395421505, 'clip_ratio/region_mean': 0.1634622924029827, 'reward_total_mean': 0.059857502579689026, 'reward_meter_mean': 0.156070739030838, 'reward_meter_std': 0.217662513256073, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9246528148651123, 'reward_lexical_plausibility_std': 0.14081719517707825, 'reward_repeat_penalty_mean': 0.946195125579834, 'reward_repeat_penalty_std': 0.09871681034564972, 'reward_repeat_floor_mean': 0.9936034679412842, 'reward_repeat_floor_std': 0.011640418320894241, 'reward_near_duplicate_penalty_mean': 0.9784188866615295, 'reward_near_duplicate_penalty_std': 0.03865911811590195, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9648962020874023, 'reward_distinct_2_std': 0.06701616942882538, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.059857502579689026, 'reward_total_composite_std': 0.0954614207148552, 'epoch': 0.01}

  6%|▌         | 202/3300 [44:17<37:20:53, 43.40s/it]INFO 04-12 16:42:43 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 203/3300 [44:25<28:17:18, 32.88s/it]2026-04-12 16:42:52,052 | INFO | train_grpo_train | metrics_logged mode=train step=203

                                                     
{'loss': 0.1517, 'grad_norm': 22.60735511779785, 'learning_rate': 9.387878787878789e-06, 'num_tokens': 389472.0, 'completions/mean_length': 34.375, 'completions/min_length': 27.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7746254801750183, 'rewards/meter/std': 0.300258994102478, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.993865430355072, 'rewards/repeat_penalty/std': 0.008924965746700764, 'rewards/repeat_floor/mean': 0.9990798234939575, 'rewards/repeat_floor/std': 0.0013387514045462012, 'rewards/near_duplicate_penalty/mean': 0.993865430355072, 'rewards/near_duplicate_penalty/std': 0.008924965746700764, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4030006229877472, 'rewards/total_composite/std': 0.2450082004070282, 'reward': 0.4030006229877472, 'reward_std': 0.2450082153081894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23428988456726074, 'sampling/sampling_logp_difference/max': 1.3909235000610352, 'sampling/importance_sampling_ratio/min': 0.24884538352489471, 'sampling/importance_sampling_ratio/mean': 1.0344312191009521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3992421329021454, 'clip_ratio/low_mean': 0.09450177289545536, 'clip_ratio/low_min': 0.09450177289545536, 'clip_ratio/high_mean': 0.10327311791479588, 'clip_ratio/high_max': 0.10327311791479588, 'clip_ratio/region_mean': 0.19777489081025124, 'reward_total_mean': 0.4030006229877472, 'reward_meter_mean': 0.7746254801750183, 'reward_meter_std': 0.300258994102478, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.993865430355072, 'reward_repeat_penalty_std': 0.008924965746700764, 'reward_repeat_floor_mean': 0.9990798234939575, 'reward_repeat_floor_std': 0.0013387514045462012, 'reward_near_duplicate_penalty_mean': 0.993865430355072, 'reward_near_duplicate_penalty_std': 0.008924965746700764, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.4030006229877472, 'reward_total_composite_std': 0.2450082004070282, 'epoch': 0.01}

  6%|▌         | 203/3300 [44:25<28:17:18, 32.88s/it]INFO 04-12 16:42:52 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 204/3300 [44:40<23:27:13, 27.27s/it]2026-04-12 16:43:06,228 | INFO | train_grpo_train | metrics_logged mode=train step=204

                                                     
{'loss': -0.0813, 'grad_norm': 1.6889697313308716, 'learning_rate': 9.384848484848486e-06, 'num_tokens': 391095.0, 'completions/mean_length': 346.875, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 71.66667175292969, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7293943166732788, 'rewards/meter/std': 0.30243679881095886, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.29880714416503906, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08964214473962784, 'rewards/lexical_plausibility/mean': 0.7510567903518677, 'rewards/lexical_plausibility/std': 0.21321669220924377, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9344176054000854, 'rewards/repeat_penalty/std': 0.09440242499113083, 'rewards/repeat_floor/mean': 0.9948080778121948, 'rewards/repeat_floor/std': 0.006688643712550402, 'rewards/near_duplicate_penalty/mean': 0.9905301928520203, 'rewards/near_duplicate_penalty/std': 0.012749215587973595, 'rewards/opening_diversity/mean': 0.953125, 'rewards/opening_diversity/std': 0.0867956355214119, 'rewards/distinct_2/mean': 0.9893455505371094, 'rewards/distinct_2/std': 0.024057654663920403, 'rewards/total_composite/mean': 0.10282497107982635, 'rewards/total_composite/std': 0.11676634848117828, 'reward': 0.10282497107982635, 'reward_std': 0.11676633358001709, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21512453258037567, 'sampling/sampling_logp_difference/max': 1.3077983856201172, 'sampling/importance_sampling_ratio/min': 0.2704147696495056, 'sampling/importance_sampling_ratio/mean': 1.0816482305526733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9477676451206207, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06274452432990074, 'clip_ratio/high_max': 0.06274452432990074, 'clip_ratio/region_mean': 0.06274452432990074, 'reward_total_mean': 0.10282497107982635, 'reward_meter_mean': 0.7293943166732788, 'reward_meter_std': 0.30243679881095886, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.29880714416503906, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08964214473962784, 'reward_lexical_plausibility_mean': 0.7510567903518677, 'reward_lexical_plausibility_std': 0.21321669220924377, 'reward_repeat_penalty_mean': 0.9344176054000854, 'reward_repeat_penalty_std': 0.09440242499113083, 'reward_repeat_floor_mean': 0.9948080778121948, 'reward_repeat_floor_std': 0.006688643712550402, 'reward_near_duplicate_penalty_mean': 0.9905301928520203, 'reward_near_duplicate_penalty_std': 0.012749215587973595, 'reward_opening_diversity_mean': 0.953125, 'reward_opening_diversity_std': 0.0867956355214119, 'reward_distinct_2_mean': 0.9893455505371094, 'reward_distinct_2_std': 0.024057654663920403, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.10282497107982635, 'reward_total_composite_std': 0.11676634848117828, 'epoch': 0.01}

  6%|▌         | 204/3300 [44:40<23:27:13, 27.27s/it]INFO 04-12 16:43:06 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 205/3300 [44:54<20:07:29, 23.41s/it]2026-04-12 16:43:20,620 | INFO | train_grpo_train | metrics_logged mode=train step=205

                                                     
{'loss': -0.0467, 'grad_norm': 7.729182720184326, 'learning_rate': 9.381818181818183e-06, 'num_tokens': 392838.0, 'completions/mean_length': 117.875, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.57143020629883, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.3980157971382141, 'rewards/meter/std': 0.33278077840805054, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9597687125205994, 'rewards/lexical_plausibility/std': 0.1067589670419693, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.9880841374397278, 'rewards/repeat_penalty/std': 0.008756035007536411, 'rewards/repeat_floor/mean': 0.9982126355171204, 'rewards/repeat_floor/std': 0.0013134105829522014, 'rewards/near_duplicate_penalty/mean': 0.9880841374397278, 'rewards/near_duplicate_penalty/std': 0.008756035007536411, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1487998366355896, 'rewards/total_composite/std': 0.14598247408866882, 'reward': 0.1487998366355896, 'reward_std': 0.14598247408866882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26030126214027405, 'sampling/sampling_logp_difference/max': 2.192049026489258, 'sampling/importance_sampling_ratio/min': 0.11168766766786575, 'sampling/importance_sampling_ratio/mean': 1.0297348499298096, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9395598471164703, 'clip_ratio/low_mean': 0.11876832321286201, 'clip_ratio/low_min': 0.11876832321286201, 'clip_ratio/high_mean': 0.09215893037617207, 'clip_ratio/high_max': 0.09215893037617207, 'clip_ratio/region_mean': 0.21092725358903408, 'reward_total_mean': 0.1487998366355896, 'reward_meter_mean': 0.3980157971382141, 'reward_meter_std': 0.33278077840805054, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9597687125205994, 'reward_lexical_plausibility_std': 0.1067589670419693, 'reward_repeat_penalty_mean': 0.9880841374397278, 'reward_repeat_penalty_std': 0.008756035007536411, 'reward_repeat_floor_mean': 0.9982126355171204, 'reward_repeat_floor_std': 0.0013134105829522014, 'reward_near_duplicate_penalty_mean': 0.9880841374397278, 'reward_near_duplicate_penalty_std': 0.008756035007536411, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.1487998366355896, 'reward_total_composite_std': 0.14598247408866882, 'epoch': 0.01}

  6%|▌         | 205/3300 [44:54<20:07:29, 23.41s/it]INFO 04-12 16:43:20 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 206/3300 [45:08<17:41:47, 20.59s/it]2026-04-12 16:43:34,635 | INFO | train_grpo_train | metrics_logged mode=train step=206

                                                     
{'loss': -0.0657, 'grad_norm': 3.1638717651367188, 'learning_rate': 9.378787878787879e-06, 'num_tokens': 394523.0, 'completions/mean_length': 168.625, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 54.16666793823242, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.622302770614624, 'rewards/meter/std': 0.48749101161956787, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.3535534143447876, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9010752439498901, 'rewards/lexical_plausibility/std': 0.18063318729400635, 'rewards/judge_quality/mean': 0.39750000834465027, 'rewards/judge_quality/std': 0.27395254373550415, 'rewards/repeat_penalty/mean': 0.9985781908035278, 'rewards/repeat_penalty/std': 0.0021619838662445545, 'rewards/repeat_floor/mean': 0.9997867345809937, 'rewards/repeat_floor/std': 0.00032429612474516034, 'rewards/near_duplicate_penalty/mean': 0.9985781908035278, 'rewards/near_duplicate_penalty/std': 0.0021619838662445545, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32643288373947144, 'rewards/total_composite/std': 0.30038806796073914, 'reward': 0.32643288373947144, 'reward_std': 0.30038806796073914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19563747942447662, 'sampling/sampling_logp_difference/max': 1.1974034309387207, 'sampling/importance_sampling_ratio/min': 0.3019773066043854, 'sampling/importance_sampling_ratio/mean': 1.0790159702301025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8306284844875336, 'clip_ratio/low_mean': 0.03214285895228386, 'clip_ratio/low_min': 0.03214285895228386, 'clip_ratio/high_mean': 0.0801125643774867, 'clip_ratio/high_max': 0.0801125643774867, 'clip_ratio/region_mean': 0.11225542332977057, 'reward_total_mean': 0.32643288373947144, 'reward_meter_mean': 0.622302770614624, 'reward_meter_std': 0.48749101161956787, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.3535534143447876, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9010752439498901, 'reward_lexical_plausibility_std': 0.18063318729400635, 'reward_repeat_penalty_mean': 0.9985781908035278, 'reward_repeat_penalty_std': 0.0021619838662445545, 'reward_repeat_floor_mean': 0.9997867345809937, 'reward_repeat_floor_std': 0.00032429612474516034, 'reward_near_duplicate_penalty_mean': 0.9985781908035278, 'reward_near_duplicate_penalty_std': 0.0021619838662445545, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39750000834465027, 'reward_judge_quality_std': 0.27395254373550415, 'reward_total_composite_mean': 0.32643288373947144, 'reward_total_composite_std': 0.30038806796073914, 'epoch': 0.01}

  6%|▌         | 206/3300 [45:08<17:41:47, 20.59s/it]INFO 04-12 16:43:34 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 207/3300 [45:22<16:02:31, 18.67s/it]2026-04-12 16:43:48,830 | INFO | train_grpo_train | metrics_logged mode=train step=207

                                                     
{'loss': -0.0234, 'grad_norm': 5.783321857452393, 'learning_rate': 9.375757575757576e-06, 'num_tokens': 396003.0, 'completions/mean_length': 96.0, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.57143020629883, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6044220328330994, 'rewards/meter/std': 0.38935935497283936, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8877576589584351, 'rewards/lexical_plausibility/std': 0.16585402190685272, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.9972041845321655, 'rewards/repeat_penalty/std': 0.007907760329544544, 'rewards/repeat_floor/mean': 0.9995806217193604, 'rewards/repeat_floor/std': 0.0011861597886309028, 'rewards/near_duplicate_penalty/mean': 0.9972041845321655, 'rewards/near_duplicate_penalty/std': 0.007907760329544544, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15325434505939484, 'rewards/total_composite/std': 0.13274861872196198, 'reward': 0.15325434505939484, 'reward_std': 0.13274861872196198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19770850241184235, 'sampling/sampling_logp_difference/max': 1.696860909461975, 'sampling/importance_sampling_ratio/min': 0.18325787782669067, 'sampling/importance_sampling_ratio/mean': 1.0484166145324707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0697822868824005, 'clip_ratio/low_mean': 0.06460118852555752, 'clip_ratio/low_min': 0.06460118852555752, 'clip_ratio/high_mean': 0.046474359929561615, 'clip_ratio/high_max': 0.046474359929561615, 'clip_ratio/region_mean': 0.11107554845511913, 'reward_total_mean': 0.15325434505939484, 'reward_meter_mean': 0.6044220328330994, 'reward_meter_std': 0.38935935497283936, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8877576589584351, 'reward_lexical_plausibility_std': 0.16585402190685272, 'reward_repeat_penalty_mean': 0.9972041845321655, 'reward_repeat_penalty_std': 0.007907760329544544, 'reward_repeat_floor_mean': 0.9995806217193604, 'reward_repeat_floor_std': 0.0011861597886309028, 'reward_near_duplicate_penalty_mean': 0.9972041845321655, 'reward_near_duplicate_penalty_std': 0.007907760329544544, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.15325434505939484, 'reward_total_composite_std': 0.13274861872196198, 'epoch': 0.01}

  6%|▋         | 207/3300 [45:22<16:02:31, 18.67s/it]INFO 04-12 16:43:49 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 208/3300 [45:31<13:30:06, 15.72s/it]2026-04-12 16:43:57,663 | INFO | train_grpo_train | metrics_logged mode=train step=208

                                                     
{'loss': -0.0052, 'grad_norm': 18.905698776245117, 'learning_rate': 9.372727272727273e-06, 'num_tokens': 397599.0, 'completions/mean_length': 35.5, 'completions/min_length': 30.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6675697565078735, 'rewards/meter/std': 0.31470629572868347, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9751378297805786, 'rewards/repeat_penalty/std': 0.029883703216910362, 'rewards/repeat_floor/mean': 0.9967879056930542, 'rewards/repeat_floor/std': 0.0036519819404929876, 'rewards/near_duplicate_penalty/mean': 0.9838311672210693, 'rewards/near_duplicate_penalty/std': 0.022379130125045776, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.2621288299560547, 'rewards/total_composite/std': 0.13208754360675812, 'reward': 0.2621288299560547, 'reward_std': 0.13208752870559692, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23961612582206726, 'sampling/sampling_logp_difference/max': 1.2466683387756348, 'sampling/importance_sampling_ratio/min': 0.28746092319488525, 'sampling/importance_sampling_ratio/mean': 1.0629518032073975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.228884309530258, 'clip_ratio/low_mean': 0.1445711050182581, 'clip_ratio/low_min': 0.1445711050182581, 'clip_ratio/high_mean': 0.11328125186264515, 'clip_ratio/high_max': 0.11328125186264515, 'clip_ratio/region_mean': 0.25785235688090324, 'reward_total_mean': 0.2621288299560547, 'reward_meter_mean': 0.6675697565078735, 'reward_meter_std': 0.31470629572868347, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9751378297805786, 'reward_repeat_penalty_std': 0.029883703216910362, 'reward_repeat_floor_mean': 0.9967879056930542, 'reward_repeat_floor_std': 0.0036519819404929876, 'reward_near_duplicate_penalty_mean': 0.9838311672210693, 'reward_near_duplicate_penalty_std': 0.022379130125045776, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.2621288299560547, 'reward_total_composite_std': 0.13208754360675812, 'epoch': 0.01}

  6%|▋         | 208/3300 [45:31<13:30:06, 15.72s/it]INFO 04-12 16:43:57 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 209/3300 [45:45<13:08:24, 15.30s/it]2026-04-12 16:44:11,996 | INFO | train_grpo_train | metrics_logged mode=train step=209

                                                     
{'loss': -0.0311, 'grad_norm': 2.745635747909546, 'learning_rate': 9.36969696969697e-06, 'num_tokens': 399100.0, 'completions/mean_length': 213.625, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 34.60000228881836, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.28015777468681335, 'rewards/meter/std': 0.3834420144557953, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8569297790527344, 'rewards/lexical_plausibility/std': 0.17469874024391174, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9683129191398621, 'rewards/repeat_penalty/std': 0.08822023123502731, 'rewards/repeat_floor/mean': 0.9980594515800476, 'rewards/repeat_floor/std': 0.005279997363686562, 'rewards/near_duplicate_penalty/mean': 0.9995629191398621, 'rewards/near_duplicate_penalty/std': 0.0012362091802060604, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11359059810638428, 'rewards/total_composite/std': 0.16369518637657166, 'reward': 0.11359059810638428, 'reward_std': 0.16369518637657166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24144268035888672, 'sampling/sampling_logp_difference/max': 1.072744369506836, 'sampling/importance_sampling_ratio/min': 0.34206846356391907, 'sampling/importance_sampling_ratio/mean': 1.0731685161590576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7406482100486755, 'clip_ratio/low_mean': 0.04836849123239517, 'clip_ratio/low_min': 0.04836849123239517, 'clip_ratio/high_mean': 0.0986643135547638, 'clip_ratio/high_max': 0.0986643135547638, 'clip_ratio/region_mean': 0.14703280478715897, 'reward_total_mean': 0.11359059810638428, 'reward_meter_mean': 0.28015777468681335, 'reward_meter_std': 0.3834420144557953, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8569297790527344, 'reward_lexical_plausibility_std': 0.17469874024391174, 'reward_repeat_penalty_mean': 0.9683129191398621, 'reward_repeat_penalty_std': 0.08822023123502731, 'reward_repeat_floor_mean': 0.9980594515800476, 'reward_repeat_floor_std': 0.005279997363686562, 'reward_near_duplicate_penalty_mean': 0.9995629191398621, 'reward_near_duplicate_penalty_std': 0.0012362091802060604, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.11359059810638428, 'reward_total_composite_std': 0.16369518637657166, 'epoch': 0.01}

  6%|▋         | 209/3300 [45:45<13:08:24, 15.30s/it]INFO 04-12 16:44:12 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 210/3300 [46:00<12:51:15, 14.98s/it]2026-04-12 16:44:26,206 | INFO | train_grpo_train | metrics_logged mode=train step=210

                                                     
{'loss': -0.0112, 'grad_norm': 0.7419185638427734, 'learning_rate': 9.366666666666668e-06, 'num_tokens': 400809.0, 'completions/mean_length': 402.625, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.5791481137275696, 'rewards/meter/std': 0.31973588466644287, 'rewards/count_adherence/mean': 0.4625000059604645, 'rewards/count_adherence/std': 0.2722263038158417, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8387500047683716, 'rewards/count_floor/std': 0.08166788518428802, 'rewards/lexical_plausibility/mean': 0.7092629671096802, 'rewards/lexical_plausibility/std': 0.1857946813106537, 'rewards/judge_quality/mean': 0.07500000298023224, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.9599920511245728, 'rewards/repeat_penalty/std': 0.08602334558963776, 'rewards/repeat_floor/mean': 0.9968570470809937, 'rewards/repeat_floor/std': 0.0050535649061203, 'rewards/near_duplicate_penalty/mean': 0.9921658039093018, 'rewards/near_duplicate_penalty/std': 0.005406332667917013, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9989688396453857, 'rewards/distinct_2/std': 0.002916498575359583, 'rewards/total_composite/mean': 0.037158988416194916, 'rewards/total_composite/std': 0.02932584658265114, 'reward': 0.037158988416194916, 'reward_std': 0.02932584658265114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1843682825565338, 'sampling/sampling_logp_difference/max': 0.7553153038024902, 'sampling/importance_sampling_ratio/min': 0.4817355275154114, 'sampling/importance_sampling_ratio/mean': 1.0175669193267822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4296127259731293, 'clip_ratio/low_mean': 0.011811023578047752, 'clip_ratio/low_min': 0.011811023578047752, 'clip_ratio/high_mean': 0.034090910106897354, 'clip_ratio/high_max': 0.034090910106897354, 'clip_ratio/region_mean': 0.045901933684945107, 'reward_total_mean': 0.037158988416194916, 'reward_meter_mean': 0.5791481137275696, 'reward_meter_std': 0.31973588466644287, 'reward_count_adherence_mean': 0.4625000059604645, 'reward_count_adherence_std': 0.2722263038158417, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8387500047683716, 'reward_count_floor_std': 0.08166788518428802, 'reward_lexical_plausibility_mean': 0.7092629671096802, 'reward_lexical_plausibility_std': 0.1857946813106537, 'reward_repeat_penalty_mean': 0.9599920511245728, 'reward_repeat_penalty_std': 0.08602334558963776, 'reward_repeat_floor_mean': 0.9968570470809937, 'reward_repeat_floor_std': 0.0050535649061203, 'reward_near_duplicate_penalty_mean': 0.9921658039093018, 'reward_near_duplicate_penalty_std': 0.005406332667917013, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9989688396453857, 'reward_distinct_2_std': 0.002916498575359583, 'reward_judge_quality_mean': 0.07500000298023224, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.037158988416194916, 'reward_total_composite_std': 0.02932584658265114, 'epoch': 0.01}

  6%|▋         | 210/3300 [46:00<12:51:15, 14.98s/it]INFO 04-12 16:44:26 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 211/3300 [46:08<11:17:52, 13.17s/it]2026-04-12 16:44:35,151 | INFO | train_grpo_train | metrics_logged mode=train step=211

                                                     
{'loss': 0.0697, 'grad_norm': 19.052175521850586, 'learning_rate': 9.363636363636365e-06, 'num_tokens': 402299.0, 'completions/mean_length': 32.25, 'completions/min_length': 30.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.35857319831848145, 'rewards/meter/std': 0.3390684723854065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.995916485786438, 'rewards/repeat_penalty/std': 0.00799814984202385, 'rewards/repeat_floor/mean': 0.9993874430656433, 'rewards/repeat_floor/std': 0.0011997224064543843, 'rewards/near_duplicate_penalty/mean': 0.995916485786438, 'rewards/near_duplicate_penalty/std': 0.00799814984202385, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16217264533042908, 'rewards/total_composite/std': 0.15186506509780884, 'reward': 0.16217264533042908, 'reward_std': 0.15186506509780884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2352273315191269, 'sampling/sampling_logp_difference/max': 1.5037126541137695, 'sampling/importance_sampling_ratio/min': 0.22230330109596252, 'sampling/importance_sampling_ratio/mean': 1.0422440767288208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.69798144698143, 'clip_ratio/low_mean': 0.08931615576148033, 'clip_ratio/low_min': 0.08931615576148033, 'clip_ratio/high_mean': 0.10882056504487991, 'clip_ratio/high_max': 0.10882056504487991, 'clip_ratio/region_mean': 0.19813672080636024, 'reward_total_mean': 0.16217264533042908, 'reward_meter_mean': 0.35857319831848145, 'reward_meter_std': 0.3390684723854065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.995916485786438, 'reward_repeat_penalty_std': 0.00799814984202385, 'reward_repeat_floor_mean': 0.9993874430656433, 'reward_repeat_floor_std': 0.0011997224064543843, 'reward_near_duplicate_penalty_mean': 0.995916485786438, 'reward_near_duplicate_penalty_std': 0.00799814984202385, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.16217264533042908, 'reward_total_composite_std': 0.15186506509780884, 'epoch': 0.01}

  6%|▋         | 211/3300 [46:09<11:17:52, 13.17s/it]INFO 04-12 16:44:35 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 212/3300 [46:23<11:40:04, 13.60s/it]2026-04-12 16:44:49,771 | INFO | train_grpo_train | metrics_logged mode=train step=212

                                                     
{'loss': -0.0691, 'grad_norm': 1.820805311203003, 'learning_rate': 9.36060606060606e-06, 'num_tokens': 403891.0, 'completions/mean_length': 400.0, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.673837423324585, 'rewards/meter/std': 0.2788621783256531, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687499761581421, 'rewards/count_floor/std': 0.07763238251209259, 'rewards/lexical_plausibility/mean': 0.7303123474121094, 'rewards/lexical_plausibility/std': 0.15362820029258728, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9950567483901978, 'rewards/repeat_penalty/std': 0.008463154546916485, 'rewards/repeat_floor/mean': 0.9992585182189941, 'rewards/repeat_floor/std': 0.001269485685043037, 'rewards/near_duplicate_penalty/mean': 0.9950567483901978, 'rewards/near_duplicate_penalty/std': 0.008463154546916485, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.07967209815979004, 'rewards/total_composite/std': 0.09823612868785858, 'reward': 0.07967209815979004, 'reward_std': 0.09823612868785858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2677590847015381, 'sampling/sampling_logp_difference/max': 0.7996141910552979, 'sampling/importance_sampling_ratio/min': 0.4977577030658722, 'sampling/importance_sampling_ratio/mean': 1.1355384588241577, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0449228286743164, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.04657771252095699, 'clip_ratio/high_max': 0.04657771252095699, 'clip_ratio/region_mean': 0.04657771252095699, 'reward_total_mean': 0.07967209815979004, 'reward_meter_mean': 0.673837423324585, 'reward_meter_std': 0.2788621783256531, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687499761581421, 'reward_count_floor_std': 0.07763238251209259, 'reward_lexical_plausibility_mean': 0.7303123474121094, 'reward_lexical_plausibility_std': 0.15362820029258728, 'reward_repeat_penalty_mean': 0.9950567483901978, 'reward_repeat_penalty_std': 0.008463154546916485, 'reward_repeat_floor_mean': 0.9992585182189941, 'reward_repeat_floor_std': 0.001269485685043037, 'reward_near_duplicate_penalty_mean': 0.9950567483901978, 'reward_near_duplicate_penalty_std': 0.008463154546916485, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.07967209815979004, 'reward_total_composite_std': 0.09823612868785858, 'epoch': 0.01}

  6%|▋         | 212/3300 [46:23<11:40:04, 13.60s/it]INFO 04-12 16:44:50 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 213/3300 [46:30<10:03:28, 11.73s/it]2026-04-12 16:44:57,129 | INFO | train_grpo_train | metrics_logged mode=train step=213

                                                     
{'loss': -0.034, 'grad_norm': 27.234357833862305, 'learning_rate': 9.357575757575758e-06, 'num_tokens': 405291.0, 'completions/mean_length': 22.0, 'completions/min_length': 17.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7079232931137085, 'rewards/meter/std': 0.33456262946128845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.887499988079071, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6232815980911255, 'rewards/total_composite/std': 0.34102189540863037, 'reward': 0.6232815980911255, 'reward_std': 0.34102189540863037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15619590878486633, 'sampling/sampling_logp_difference/max': 1.7358851432800293, 'sampling/importance_sampling_ratio/min': 0.17624413967132568, 'sampling/importance_sampling_ratio/mean': 1.005838394165039, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.035675896331668, 'clip_ratio/low_mean': 0.07611111178994179, 'clip_ratio/low_min': 0.07611111178994179, 'clip_ratio/high_mean': 0.1122333100065589, 'clip_ratio/high_max': 0.1122333100065589, 'clip_ratio/region_mean': 0.18834442179650068, 'reward_total_mean': 0.6232815980911255, 'reward_meter_mean': 0.7079232931137085, 'reward_meter_std': 0.33456262946128845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.887499988079071, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.6232815980911255, 'reward_total_composite_std': 0.34102189540863037, 'epoch': 0.01}

  6%|▋         | 213/3300 [46:30<10:03:28, 11.73s/it]INFO 04-12 16:44:57 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 214/3300 [46:45<10:39:45, 12.44s/it]2026-04-12 16:45:11,224 | INFO | train_grpo_train | metrics_logged mode=train step=214

                                                     
{'loss': -0.0377, 'grad_norm': 1.3515175580978394, 'learning_rate': 9.354545454545455e-06, 'num_tokens': 406633.0, 'completions/mean_length': 454.75, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.38559412956237793, 'rewards/meter/std': 0.37455591559410095, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.05175493285059929, 'rewards/lexical_plausibility/mean': 0.7717058658599854, 'rewards/lexical_plausibility/std': 0.11193247139453888, 'rewards/judge_quality/mean': 0.0625, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9950237274169922, 'rewards/repeat_penalty/std': 0.005315759684890509, 'rewards/repeat_floor/mean': 0.9992535710334778, 'rewards/repeat_floor/std': 0.0007973714964464307, 'rewards/near_duplicate_penalty/mean': 0.9950237274169922, 'rewards/near_duplicate_penalty/std': 0.005315759684890509, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.031039826571941376, 'rewards/total_composite/std': 0.0496603399515152, 'reward': 0.031039826571941376, 'reward_std': 0.0496603362262249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25347810983657837, 'sampling/sampling_logp_difference/max': 0.9176156520843506, 'sampling/importance_sampling_ratio/min': 0.537760853767395, 'sampling/importance_sampling_ratio/mean': 1.139912486076355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48676249384880066, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.025462962687015533, 'clip_ratio/high_max': 0.025462962687015533, 'clip_ratio/region_mean': 0.025462962687015533, 'reward_total_mean': 0.031039826571941376, 'reward_meter_mean': 0.38559412956237793, 'reward_meter_std': 0.37455591559410095, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.05175493285059929, 'reward_lexical_plausibility_mean': 0.7717058658599854, 'reward_lexical_plausibility_std': 0.11193247139453888, 'reward_repeat_penalty_mean': 0.9950237274169922, 'reward_repeat_penalty_std': 0.005315759684890509, 'reward_repeat_floor_mean': 0.9992535710334778, 'reward_repeat_floor_std': 0.0007973714964464307, 'reward_near_duplicate_penalty_mean': 0.9950237274169922, 'reward_near_duplicate_penalty_std': 0.005315759684890509, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.0625, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.031039826571941376, 'reward_total_composite_std': 0.0496603399515152, 'epoch': 0.01}

  6%|▋         | 214/3300 [46:45<10:39:45, 12.44s/it]INFO 04-12 16:45:11 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 215/3300 [46:54<9:47:19, 11.42s/it] 2026-04-12 16:45:20,277 | INFO | train_grpo_train | metrics_logged mode=train step=215

                                                    
{'loss': 0.0507, 'grad_norm': 23.49319839477539, 'learning_rate': 9.351515151515152e-06, 'num_tokens': 408734.0, 'completions/mean_length': 65.625, 'completions/min_length': 51.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.5201544165611267, 'rewards/meter/std': 0.33865422010421753, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.9904114007949829, 'rewards/repeat_penalty/std': 0.0049859690479934216, 'rewards/repeat_floor/mean': 0.9985617399215698, 'rewards/repeat_floor/std': 0.000747889163903892, 'rewards/near_duplicate_penalty/mean': 0.9904114007949829, 'rewards/near_duplicate_penalty/std': 0.0049859690479934216, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2099500298500061, 'rewards/total_composite/std': 0.128269225358963, 'reward': 0.2099500298500061, 'reward_std': 0.12826921045780182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2299450784921646, 'sampling/sampling_logp_difference/max': 2.288926601409912, 'sampling/importance_sampling_ratio/min': 0.10137522220611572, 'sampling/importance_sampling_ratio/mean': 0.9993533492088318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2596945688128471, 'clip_ratio/low_mean': 0.10337468795478344, 'clip_ratio/low_min': 0.10337468795478344, 'clip_ratio/high_mean': 0.08307577297091484, 'clip_ratio/high_max': 0.08307577297091484, 'clip_ratio/region_mean': 0.18645046092569828, 'reward_total_mean': 0.2099500298500061, 'reward_meter_mean': 0.5201544165611267, 'reward_meter_std': 0.33865422010421753, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9904114007949829, 'reward_repeat_penalty_std': 0.0049859690479934216, 'reward_repeat_floor_mean': 0.9985617399215698, 'reward_repeat_floor_std': 0.000747889163903892, 'reward_near_duplicate_penalty_mean': 0.9904114007949829, 'reward_near_duplicate_penalty_std': 0.0049859690479934216, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.2099500298500061, 'reward_total_composite_std': 0.128269225358963, 'epoch': 0.01}

  7%|▋         | 215/3300 [46:54<9:47:19, 11.42s/it]INFO 04-12 16:45:20 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 216/3300 [47:07<10:14:54, 11.96s/it]2026-04-12 16:45:33,499 | INFO | train_grpo_train | metrics_logged mode=train step=216

                                                     
{'loss': -0.0272, 'grad_norm': 3.323742151260376, 'learning_rate': 9.34848484848485e-06, 'num_tokens': 410190.0, 'completions/mean_length': 156.0, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 37.333335876464844, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6257503032684326, 'rewards/meter/std': 0.4384929835796356, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9170446395874023, 'rewards/lexical_plausibility/std': 0.13874001801013947, 'rewards/judge_quality/mean': 0.3712500333786011, 'rewards/judge_quality/std': 0.3466755449771881, 'rewards/repeat_penalty/mean': 0.9333760738372803, 'rewards/repeat_penalty/std': 0.12040144205093384, 'rewards/repeat_floor/mean': 0.9954893589019775, 'rewards/repeat_floor/std': 0.007987740449607372, 'rewards/near_duplicate_penalty/mean': 0.9949291348457336, 'rewards/near_duplicate_penalty/std': 0.01080945786088705, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2878599464893341, 'rewards/total_composite/std': 0.34471791982650757, 'reward': 0.2878599464893341, 'reward_std': 0.3447178900241852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21436162292957306, 'sampling/sampling_logp_difference/max': 0.9169864654541016, 'sampling/importance_sampling_ratio/min': 0.39972180128097534, 'sampling/importance_sampling_ratio/mean': 1.067720890045166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4292386174201965, 'clip_ratio/low_mean': 0.08294450864195824, 'clip_ratio/low_min': 0.08294450864195824, 'clip_ratio/high_mean': 0.03644018666818738, 'clip_ratio/high_max': 0.03644018666818738, 'clip_ratio/region_mean': 0.11938469531014562, 'reward_total_mean': 0.2878599464893341, 'reward_meter_mean': 0.6257503032684326, 'reward_meter_std': 0.4384929835796356, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9170446395874023, 'reward_lexical_plausibility_std': 0.13874001801013947, 'reward_repeat_penalty_mean': 0.9333760738372803, 'reward_repeat_penalty_std': 0.12040144205093384, 'reward_repeat_floor_mean': 0.9954893589019775, 'reward_repeat_floor_std': 0.007987740449607372, 'reward_near_duplicate_penalty_mean': 0.9949291348457336, 'reward_near_duplicate_penalty_std': 0.01080945786088705, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500333786011, 'reward_judge_quality_std': 0.3466755449771881, 'reward_total_composite_mean': 0.2878599464893341, 'reward_total_composite_std': 0.34471791982650757, 'epoch': 0.01}

  7%|▋         | 216/3300 [47:07<10:14:54, 11.96s/it]INFO 04-12 16:45:33 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 217/3300 [47:15<9:20:48, 10.91s/it] 2026-04-12 16:45:41,967 | INFO | train_grpo_train | metrics_logged mode=train step=217

                                                    
{'loss': 0.1321, 'grad_norm': 16.335145950317383, 'learning_rate': 9.345454545454547e-06, 'num_tokens': 411796.0, 'completions/mean_length': 35.75, 'completions/min_length': 30.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6175544261932373, 'rewards/meter/std': 0.4756390154361725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9516937732696533, 'rewards/repeat_penalty/std': 0.02354690432548523, 'rewards/repeat_floor/mean': 0.9927541017532349, 'rewards/repeat_floor/std': 0.0035320345778018236, 'rewards/near_duplicate_penalty/mean': 0.9516937732696533, 'rewards/near_duplicate_penalty/std': 0.02354690432548523, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.43223848938941956, 'rewards/total_composite/std': 0.40692591667175293, 'reward': 0.43223848938941956, 'reward_std': 0.40692591667175293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17292283475399017, 'sampling/sampling_logp_difference/max': 1.6637763977050781, 'sampling/importance_sampling_ratio/min': 0.18942229449748993, 'sampling/importance_sampling_ratio/mean': 1.015834093093872, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5780327171087265, 'clip_ratio/low_mean': 0.10448349639773369, 'clip_ratio/low_min': 0.10448349639773369, 'clip_ratio/high_mean': 0.08014706149697304, 'clip_ratio/high_max': 0.08014706149697304, 'clip_ratio/region_mean': 0.18463055789470673, 'reward_total_mean': 0.43223848938941956, 'reward_meter_mean': 0.6175544261932373, 'reward_meter_std': 0.4756390154361725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9516937732696533, 'reward_repeat_penalty_std': 0.02354690432548523, 'reward_repeat_floor_mean': 0.9927541017532349, 'reward_repeat_floor_std': 0.0035320345778018236, 'reward_near_duplicate_penalty_mean': 0.9516937732696533, 'reward_near_duplicate_penalty_std': 0.02354690432548523, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.43223848938941956, 'reward_total_composite_std': 0.40692591667175293, 'epoch': 0.01}

  7%|▋         | 217/3300 [47:15<9:20:48, 10.91s/it]INFO 04-12 16:45:42 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 218/3300 [47:24<8:48:57, 10.30s/it]2026-04-12 16:45:50,827 | INFO | train_grpo_train | metrics_logged mode=train step=218

                                                    
{'loss': 0.0679, 'grad_norm': 25.940507888793945, 'learning_rate': 9.342424242424243e-06, 'num_tokens': 413121.0, 'completions/mean_length': 17.625, 'completions/min_length': 14.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.625, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7779308557510376, 'rewards/meter/std': 0.3581196069717407, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3586059510707855, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.36226576566696167, 'rewards/total_composite/std': 0.34641146659851074, 'reward': 0.36226576566696167, 'reward_std': 0.34641146659851074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20455262064933777, 'sampling/sampling_logp_difference/max': 1.4882903099060059, 'sampling/importance_sampling_ratio/min': 0.22575829923152924, 'sampling/importance_sampling_ratio/mean': 1.0410339832305908, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.24961219727993, 'clip_ratio/low_mean': 0.09095703344792128, 'clip_ratio/low_min': 0.09095703344792128, 'clip_ratio/high_mean': 0.06939338240772486, 'clip_ratio/high_max': 0.06939338240772486, 'clip_ratio/region_mean': 0.16035041585564613, 'reward_total_mean': 0.36226576566696167, 'reward_meter_mean': 0.7779308557510376, 'reward_meter_std': 0.3581196069717407, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3586059510707855, 'reward_total_composite_mean': 0.36226576566696167, 'reward_total_composite_std': 0.34641146659851074, 'epoch': 0.01}

  7%|▋         | 218/3300 [47:24<8:48:57, 10.30s/it]INFO 04-12 16:45:51 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 219/3300 [47:33<8:18:38,  9.71s/it]2026-04-12 16:45:59,167 | INFO | train_grpo_train | metrics_logged mode=train step=219

                                                    
{'loss': 0.0455, 'grad_norm': 19.09295082092285, 'learning_rate': 9.33939393939394e-06, 'num_tokens': 414570.0, 'completions/mean_length': 34.125, 'completions/min_length': 28.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5822951197624207, 'rewards/meter/std': 0.4585415720939636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9896408915519714, 'rewards/repeat_penalty/std': 0.016224481165409088, 'rewards/repeat_floor/mean': 0.9984461069107056, 'rewards/repeat_floor/std': 0.002433664398267865, 'rewards/near_duplicate_penalty/mean': 0.9896408915519714, 'rewards/near_duplicate_penalty/std': 0.016224481165409088, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2613937556743622, 'rewards/total_composite/std': 0.2057267129421234, 'reward': 0.2613937556743622, 'reward_std': 0.2057267129421234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18743401765823364, 'sampling/sampling_logp_difference/max': 1.6540594100952148, 'sampling/importance_sampling_ratio/min': 0.19127187132835388, 'sampling/importance_sampling_ratio/mean': 1.0398900508880615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9390588402748108, 'clip_ratio/low_mean': 0.05295000830665231, 'clip_ratio/low_min': 0.05295000830665231, 'clip_ratio/high_mean': 0.11012150626629591, 'clip_ratio/high_max': 0.11012150626629591, 'clip_ratio/region_mean': 0.16307151457294822, 'reward_total_mean': 0.2613937556743622, 'reward_meter_mean': 0.5822951197624207, 'reward_meter_std': 0.4585415720939636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9896408915519714, 'reward_repeat_penalty_std': 0.016224481165409088, 'reward_repeat_floor_mean': 0.9984461069107056, 'reward_repeat_floor_std': 0.002433664398267865, 'reward_near_duplicate_penalty_mean': 0.9896408915519714, 'reward_near_duplicate_penalty_std': 0.016224481165409088, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.2613937556743622, 'reward_total_composite_std': 0.2057267129421234, 'epoch': 0.01}

  7%|▋         | 219/3300 [47:33<8:18:38,  9.71s/it]INFO 04-12 16:45:59 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 220/3300 [47:46<9:22:33, 10.96s/it]2026-04-12 16:46:13,040 | INFO | train_grpo_train | metrics_logged mode=train step=220

                                                    
{'loss': -0.0451, 'grad_norm': 1.6660808324813843, 'learning_rate': 9.336363636363637e-06, 'num_tokens': 416198.0, 'completions/mean_length': 280.5, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 49.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.4168480634689331, 'rewards/meter/std': 0.31307777762413025, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.28149792551994324, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.934374988079071, 'rewards/count_floor/std': 0.08444937318563461, 'rewards/lexical_plausibility/mean': 0.8552516102790833, 'rewards/lexical_plausibility/std': 0.16271202266216278, 'rewards/judge_quality/mean': 0.18125000596046448, 'rewards/judge_quality/std': 0.17512750625610352, 'rewards/repeat_penalty/mean': 0.9324818253517151, 'rewards/repeat_penalty/std': 0.0984434187412262, 'rewards/repeat_floor/mean': 0.9956501722335815, 'rewards/repeat_floor/std': 0.005253738723695278, 'rewards/near_duplicate_penalty/mean': 0.9883019328117371, 'rewards/near_duplicate_penalty/std': 0.013734445907175541, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9919986724853516, 'rewards/distinct_2/std': 0.014318671077489853, 'rewards/total_composite/mean': 0.08835402131080627, 'rewards/total_composite/std': 0.10560376197099686, 'reward': 0.08835402131080627, 'reward_std': 0.10560375452041626, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21905475854873657, 'sampling/sampling_logp_difference/max': 0.9840552806854248, 'sampling/importance_sampling_ratio/min': 0.43166446685791016, 'sampling/importance_sampling_ratio/mean': 1.0340116024017334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4396289587020874, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11531941220164299, 'clip_ratio/high_max': 0.11531941220164299, 'clip_ratio/region_mean': 0.11531941220164299, 'reward_total_mean': 0.08835402131080627, 'reward_meter_mean': 0.4168480634689331, 'reward_meter_std': 0.31307777762413025, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.28149792551994324, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.934374988079071, 'reward_count_floor_std': 0.08444937318563461, 'reward_lexical_plausibility_mean': 0.8552516102790833, 'reward_lexical_plausibility_std': 0.16271202266216278, 'reward_repeat_penalty_mean': 0.9324818253517151, 'reward_repeat_penalty_std': 0.0984434187412262, 'reward_repeat_floor_mean': 0.9956501722335815, 'reward_repeat_floor_std': 0.005253738723695278, 'reward_near_duplicate_penalty_mean': 0.9883019328117371, 'reward_near_duplicate_penalty_std': 0.013734445907175541, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9919986724853516, 'reward_distinct_2_std': 0.014318671077489853, 'reward_judge_quality_mean': 0.18125000596046448, 'reward_judge_quality_std': 0.17512750625610352, 'reward_total_composite_mean': 0.08835402131080627, 'reward_total_composite_std': 0.10560376197099686, 'epoch': 0.01}

  7%|▋         | 220/3300 [47:46<9:22:33, 10.96s/it]INFO 04-12 16:46:13 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 221/3300 [48:00<10:01:53, 11.73s/it]2026-04-12 16:46:26,816 | INFO | train_grpo_train | metrics_logged mode=train step=221

                                                     
{'loss': -0.086, 'grad_norm': 4.576457977294922, 'learning_rate': 9.333333333333334e-06, 'num_tokens': 417948.0, 'completions/mean_length': 188.75, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.19065053761005402, 'rewards/meter/std': 0.1801435500383377, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.1511857807636261, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9700000286102295, 'rewards/count_floor/std': 0.04535574093461037, 'rewards/lexical_plausibility/mean': 0.9536890983581543, 'rewards/lexical_plausibility/std': 0.09056153148412704, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9514344930648804, 'rewards/repeat_penalty/std': 0.05912245064973831, 'rewards/repeat_floor/mean': 0.9948757886886597, 'rewards/repeat_floor/std': 0.006274827290326357, 'rewards/near_duplicate_penalty/mean': 0.9889129400253296, 'rewards/near_duplicate_penalty/std': 0.013839642517268658, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.961542546749115, 'rewards/distinct_2/std': 0.047152500599622726, 'rewards/total_composite/mean': 0.07901493459939957, 'rewards/total_composite/std': 0.08410973101854324, 'reward': 0.07901493459939957, 'reward_std': 0.08410973101854324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2276083528995514, 'sampling/sampling_logp_difference/max': 1.8408679962158203, 'sampling/importance_sampling_ratio/min': 0.15867963433265686, 'sampling/importance_sampling_ratio/mean': 1.0345252752304077, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5651446878910065, 'clip_ratio/low_mean': 0.07474303804337978, 'clip_ratio/low_min': 0.07474303804337978, 'clip_ratio/high_mean': 0.08329828828573227, 'clip_ratio/high_max': 0.08329828828573227, 'clip_ratio/region_mean': 0.15804132632911205, 'reward_total_mean': 0.07901493459939957, 'reward_meter_mean': 0.19065053761005402, 'reward_meter_std': 0.1801435500383377, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.1511857807636261, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9700000286102295, 'reward_count_floor_std': 0.04535574093461037, 'reward_lexical_plausibility_mean': 0.9536890983581543, 'reward_lexical_plausibility_std': 0.09056153148412704, 'reward_repeat_penalty_mean': 0.9514344930648804, 'reward_repeat_penalty_std': 0.05912245064973831, 'reward_repeat_floor_mean': 0.9948757886886597, 'reward_repeat_floor_std': 0.006274827290326357, 'reward_near_duplicate_penalty_mean': 0.9889129400253296, 'reward_near_duplicate_penalty_std': 0.013839642517268658, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.961542546749115, 'reward_distinct_2_std': 0.047152500599622726, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.07901493459939957, 'reward_total_composite_std': 0.08410973101854324, 'epoch': 0.01}

  7%|▋         | 221/3300 [48:00<10:01:53, 11.73s/it]INFO 04-12 16:46:27 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 222/3300 [48:15<10:47:07, 12.61s/it]2026-04-12 16:46:41,246 | INFO | train_grpo_train | metrics_logged mode=train step=222

                                                     
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 9.33030303030303e-06, 'num_tokens': 419516.0, 'completions/mean_length': 512.0, 'completions/min_length': 512.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 1.0, 'completions/mean_terminated_length': 0.0, 'completions/min_terminated_length': 0.0, 'completions/max_terminated_length': 0.0, 'rewards/meter/mean': 0.7977553606033325, 'rewards/meter/std': 0.2215038686990738, 'rewards/count_adherence/mean': 0.578125, 'rewards/count_adherence/std': 0.19974872469902039, 'rewards/arabic_clean/mean': 0.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8734374642372131, 'rewards/count_floor/std': 0.059924617409706116, 'rewards/lexical_plausibility/mean': 0.6908619403839111, 'rewards/lexical_plausibility/std': 0.07034913450479507, 'rewards/judge_quality/mean': 0.0625, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9941745400428772, 'rewards/repeat_penalty/std': 0.006332909222692251, 'rewards/repeat_floor/mean': 0.9992550611495972, 'rewards/repeat_floor/std': 0.0008662913460284472, 'rewards/near_duplicate_penalty/mean': 0.9963395595550537, 'rewards/near_duplicate_penalty/std': 0.005102863069623709, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9978231191635132, 'rewards/distinct_2/std': 0.0023297611624002457, 'rewards/total_composite/mean': 0.04376174509525299, 'rewards/total_composite/std': 0.030908260494470596, 'reward': 0.04376174509525299, 'reward_std': 0.030908260494470596, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0, 'sampling/sampling_logp_difference/max': 0.0, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/importance_sampling_ratio/mean': 0.0, 'sampling/importance_sampling_ratio/max': 0.0, 'entropy': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'reward_total_mean': 0.04376174509525299, 'reward_meter_mean': 0.7977553606033325, 'reward_meter_std': 0.2215038686990738, 'reward_count_adherence_mean': 0.578125, 'reward_count_adherence_std': 0.19974872469902039, 'reward_arabic_clean_mean': 0.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8734374642372131, 'reward_count_floor_std': 0.059924617409706116, 'reward_lexical_plausibility_mean': 0.6908619403839111, 'reward_lexical_plausibility_std': 0.07034913450479507, 'reward_repeat_penalty_mean': 0.9941745400428772, 'reward_repeat_penalty_std': 0.006332909222692251, 'reward_repeat_floor_mean': 0.9992550611495972, 'reward_repeat_floor_std': 0.0008662913460284472, 'reward_near_duplicate_penalty_mean': 0.9963395595550537, 'reward_near_duplicate_penalty_std': 0.005102863069623709, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9978231191635132, 'reward_distinct_2_std': 0.0023297611624002457, 'reward_judge_quality_mean': 0.0625, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.04376174509525299, 'reward_total_composite_std': 0.030908260494470596, 'epoch': 0.01}

  7%|▋         | 222/3300 [48:15<10:47:07, 12.61s/it]INFO 04-12 16:46:41 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 223/3300 [48:24<9:51:08, 11.53s/it] 2026-04-12 16:46:50,234 | INFO | train_grpo_train | metrics_logged mode=train step=223

                                                    
{'loss': -0.0243, 'grad_norm': 14.237860679626465, 'learning_rate': 9.327272727272729e-06, 'num_tokens': 421113.0, 'completions/mean_length': 43.625, 'completions/min_length': 32.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7762871384620667, 'rewards/meter/std': 0.323667973279953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.26592159271240234, 'rewards/repeat_penalty/mean': 0.9909542798995972, 'rewards/repeat_penalty/std': 0.00725660240277648, 'rewards/repeat_floor/mean': 0.998643159866333, 'rewards/repeat_floor/std': 0.001088486285880208, 'rewards/near_duplicate_penalty/mean': 0.9909542798995972, 'rewards/near_duplicate_penalty/std': 0.00725660240277648, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30592989921569824, 'rewards/total_composite/std': 0.2920238971710205, 'reward': 0.30592989921569824, 'reward_std': 0.2920239269733429, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21913012862205505, 'sampling/sampling_logp_difference/max': 1.454554557800293, 'sampling/importance_sampling_ratio/min': 0.23350435495376587, 'sampling/importance_sampling_ratio/mean': 1.050878882408142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.473749414086342, 'clip_ratio/low_mean': 0.10079926811158657, 'clip_ratio/low_min': 0.10079926811158657, 'clip_ratio/high_mean': 0.10061218775808811, 'clip_ratio/high_max': 0.10061218775808811, 'clip_ratio/region_mean': 0.20141145586967468, 'reward_total_mean': 0.30592989921569824, 'reward_meter_mean': 0.7762871384620667, 'reward_meter_std': 0.323667973279953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9909542798995972, 'reward_repeat_penalty_std': 0.00725660240277648, 'reward_repeat_floor_mean': 0.998643159866333, 'reward_repeat_floor_std': 0.001088486285880208, 'reward_near_duplicate_penalty_mean': 0.9909542798995972, 'reward_near_duplicate_penalty_std': 0.00725660240277648, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.26592159271240234, 'reward_total_composite_mean': 0.30592989921569824, 'reward_total_composite_std': 0.2920238971710205, 'epoch': 0.01}

  7%|▋         | 223/3300 [48:24<9:51:08, 11.53s/it]INFO 04-12 16:46:50 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 224/3300 [48:37<10:27:33, 12.24s/it]2026-04-12 16:47:04,141 | INFO | train_grpo_train | metrics_logged mode=train step=224

                                                     
{'loss': -0.1116, 'grad_norm': 5.811514854431152, 'learning_rate': 9.324242424242424e-06, 'num_tokens': 422749.0, 'completions/mean_length': 95.5, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9750537872314453, 'rewards/meter/std': 0.035048168152570724, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.9425743222236633, 'rewards/lexical_plausibility/std': 0.11617325246334076, 'rewards/judge_quality/mean': 0.23125000298023224, 'rewards/judge_quality/std': 0.16461534798145294, 'rewards/repeat_penalty/mean': 0.8933764696121216, 'rewards/repeat_penalty/std': 0.12036852538585663, 'rewards/repeat_floor/mean': 0.9929247498512268, 'rewards/repeat_floor/std': 0.006812785752117634, 'rewards/near_duplicate_penalty/mean': 0.9951393008232117, 'rewards/near_duplicate_penalty/std': 0.009027883410453796, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.21982286870479584, 'rewards/total_composite/std': 0.16935597360134125, 'reward': 0.21982286870479584, 'reward_std': 0.16935597360134125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23326432704925537, 'sampling/sampling_logp_difference/max': 1.2762398719787598, 'sampling/importance_sampling_ratio/min': 0.279084712266922, 'sampling/importance_sampling_ratio/mean': 1.0669156312942505, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.077867805957794, 'clip_ratio/low_mean': 0.0736715579405427, 'clip_ratio/low_min': 0.0736715579405427, 'clip_ratio/high_mean': 0.08973146043717861, 'clip_ratio/high_max': 0.08973146043717861, 'clip_ratio/region_mean': 0.1634030183777213, 'reward_total_mean': 0.21982286870479584, 'reward_meter_mean': 0.9750537872314453, 'reward_meter_std': 0.035048168152570724, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.9425743222236633, 'reward_lexical_plausibility_std': 0.11617325246334076, 'reward_repeat_penalty_mean': 0.8933764696121216, 'reward_repeat_penalty_std': 0.12036852538585663, 'reward_repeat_floor_mean': 0.9929247498512268, 'reward_repeat_floor_std': 0.006812785752117634, 'reward_near_duplicate_penalty_mean': 0.9951393008232117, 'reward_near_duplicate_penalty_std': 0.009027883410453796, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.23125000298023224, 'reward_judge_quality_std': 0.16461534798145294, 'reward_total_composite_mean': 0.21982286870479584, 'reward_total_composite_std': 0.16935597360134125, 'epoch': 0.01}

  7%|▋         | 224/3300 [48:37<10:27:33, 12.24s/it]INFO 04-12 16:47:04 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 225/3300 [48:46<9:31:03, 11.14s/it] 2026-04-12 16:47:12,721 | INFO | train_grpo_train | metrics_logged mode=train step=225

                                                    
{'loss': 0.1261, 'grad_norm': 18.091720581054688, 'learning_rate': 9.321212121212122e-06, 'num_tokens': 424350.0, 'completions/mean_length': 36.125, 'completions/min_length': 16.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7392919063568115, 'rewards/meter/std': 0.38587990403175354, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9953668117523193, 'rewards/lexical_plausibility/std': 0.009900491684675217, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.96875, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9981250166893005, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4188479781150818, 'rewards/total_composite/std': 0.34137439727783203, 'reward': 0.4188479781150818, 'reward_std': 0.34137439727783203, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22719928622245789, 'sampling/sampling_logp_difference/max': 1.7320566177368164, 'sampling/importance_sampling_ratio/min': 0.17692016065120697, 'sampling/importance_sampling_ratio/mean': 1.0504812002182007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5301357060670853, 'clip_ratio/low_mean': 0.14875992434099317, 'clip_ratio/low_min': 0.14875992434099317, 'clip_ratio/high_mean': 0.08425324968993664, 'clip_ratio/high_max': 0.08425324968993664, 'clip_ratio/region_mean': 0.2330131740309298, 'reward_total_mean': 0.4188479781150818, 'reward_meter_mean': 0.7392919063568115, 'reward_meter_std': 0.38587990403175354, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9953668117523193, 'reward_lexical_plausibility_std': 0.009900491684675217, 'reward_repeat_penalty_mean': 0.96875, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9981250166893005, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.4188479781150818, 'reward_total_composite_std': 0.34137439727783203, 'epoch': 0.01}

  7%|▋         | 225/3300 [48:46<9:31:03, 11.14s/it]INFO 04-12 16:47:12 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 226/3300 [49:00<10:08:57, 11.89s/it]2026-04-12 16:47:26,341 | INFO | train_grpo_train | metrics_logged mode=train step=226

                                                     
{'loss': -0.0531, 'grad_norm': 4.074066162109375, 'learning_rate': 9.318181818181819e-06, 'num_tokens': 425746.0, 'completions/mean_length': 82.5, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.142858505249023, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.8516990542411804, 'rewards/meter/std': 0.34473299980163574, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9503329992294312, 'rewards/lexical_plausibility/std': 0.1404794305562973, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.3853291869163513, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5613418817520142, 'rewards/total_composite/std': 0.3796902000904083, 'reward': 0.5613418817520142, 'reward_std': 0.3796902000904083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1913546323776245, 'sampling/sampling_logp_difference/max': 1.3582305908203125, 'sampling/importance_sampling_ratio/min': 0.25711530447006226, 'sampling/importance_sampling_ratio/mean': 1.0482648611068726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7384716868400574, 'clip_ratio/low_mean': 0.07791733276098967, 'clip_ratio/low_min': 0.07791733276098967, 'clip_ratio/high_mean': 0.05777899548411369, 'clip_ratio/high_max': 0.05777899548411369, 'clip_ratio/region_mean': 0.13569632824510336, 'reward_total_mean': 0.5613418817520142, 'reward_meter_mean': 0.8516990542411804, 'reward_meter_std': 0.34473299980163574, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9503329992294312, 'reward_lexical_plausibility_std': 0.1404794305562973, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.3853291869163513, 'reward_total_composite_mean': 0.5613418817520142, 'reward_total_composite_std': 0.3796902000904083, 'epoch': 0.01}

  7%|▋         | 226/3300 [49:00<10:08:57, 11.89s/it]INFO 04-12 16:47:26 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 227/3300 [49:09<9:28:08, 11.09s/it] 2026-04-12 16:47:35,584 | INFO | train_grpo_train | metrics_logged mode=train step=227

                                                    
{'loss': 0.062, 'grad_norm': 11.85656452178955, 'learning_rate': 9.315151515151516e-06, 'num_tokens': 427574.0, 'completions/mean_length': 57.5, 'completions/min_length': 48.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9355548620223999, 'rewards/meter/std': 0.10469502210617065, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.971396803855896, 'rewards/repeat_penalty/std': 0.03477616235613823, 'rewards/repeat_floor/mean': 0.995978832244873, 'rewards/repeat_floor/std': 0.004565069917589426, 'rewards/near_duplicate_penalty/mean': 0.9763974547386169, 'rewards/near_duplicate_penalty/std': 0.02343011274933815, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9946581125259399, 'rewards/distinct_2/std': 0.01510911900550127, 'rewards/total_composite/mean': 0.45958274602890015, 'rewards/total_composite/std': 0.18823538720607758, 'reward': 0.45958274602890015, 'reward_std': 0.18823538720607758, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22032207250595093, 'sampling/sampling_logp_difference/max': 1.444075584411621, 'sampling/importance_sampling_ratio/min': 0.2359641194343567, 'sampling/importance_sampling_ratio/mean': 1.0553920269012451, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.563923791050911, 'clip_ratio/low_mean': 0.1450505694374442, 'clip_ratio/low_min': 0.1450505694374442, 'clip_ratio/high_mean': 0.029545454308390617, 'clip_ratio/high_max': 0.029545454308390617, 'clip_ratio/region_mean': 0.17459602374583483, 'reward_total_mean': 0.45958274602890015, 'reward_meter_mean': 0.9355548620223999, 'reward_meter_std': 0.10469502210617065, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.971396803855896, 'reward_repeat_penalty_std': 0.03477616235613823, 'reward_repeat_floor_mean': 0.995978832244873, 'reward_repeat_floor_std': 0.004565069917589426, 'reward_near_duplicate_penalty_mean': 0.9763974547386169, 'reward_near_duplicate_penalty_std': 0.02343011274933815, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9946581125259399, 'reward_distinct_2_std': 0.01510911900550127, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.45958274602890015, 'reward_total_composite_std': 0.18823538720607758, 'epoch': 0.01}

  7%|▋         | 227/3300 [49:09<9:28:08, 11.09s/it]INFO 04-12 16:47:35 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 228/3300 [49:18<8:57:01, 10.49s/it]2026-04-12 16:47:44,664 | INFO | train_grpo_train | metrics_logged mode=train step=228

                                                    
{'loss': -0.009, 'grad_norm': 17.098552703857422, 'learning_rate': 9.312121212121212e-06, 'num_tokens': 429223.0, 'completions/mean_length': 39.125, 'completions/min_length': 32.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.2652822732925415, 'rewards/meter/std': 0.38240987062454224, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.95652174949646, 'rewards/lexical_plausibility/std': 0.12297508865594864, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9891475439071655, 'rewards/repeat_penalty/std': 0.016673490405082703, 'rewards/repeat_floor/mean': 0.9983721375465393, 'rewards/repeat_floor/std': 0.0025010122917592525, 'rewards/near_duplicate_penalty/mean': 0.9891475439071655, 'rewards/near_duplicate_penalty/std': 0.016673490405082703, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10478058457374573, 'rewards/total_composite/std': 0.15999507904052734, 'reward': 0.10478058457374573, 'reward_std': 0.15999507904052734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21070675551891327, 'sampling/sampling_logp_difference/max': 2.224109172821045, 'sampling/importance_sampling_ratio/min': 0.1081637367606163, 'sampling/importance_sampling_ratio/mean': 1.0367555618286133, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9564559906721115, 'clip_ratio/low_mean': 0.0989826051518321, 'clip_ratio/low_min': 0.0989826051518321, 'clip_ratio/high_mean': 0.059379216749221087, 'clip_ratio/high_max': 0.059379216749221087, 'clip_ratio/region_mean': 0.1583618219010532, 'reward_total_mean': 0.10478058457374573, 'reward_meter_mean': 0.2652822732925415, 'reward_meter_std': 0.38240987062454224, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.95652174949646, 'reward_lexical_plausibility_std': 0.12297508865594864, 'reward_repeat_penalty_mean': 0.9891475439071655, 'reward_repeat_penalty_std': 0.016673490405082703, 'reward_repeat_floor_mean': 0.9983721375465393, 'reward_repeat_floor_std': 0.0025010122917592525, 'reward_near_duplicate_penalty_mean': 0.9891475439071655, 'reward_near_duplicate_penalty_std': 0.016673490405082703, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.10478058457374573, 'reward_total_composite_std': 0.15999507904052734, 'epoch': 0.01}

  7%|▋         | 228/3300 [49:18<8:57:01, 10.49s/it]INFO 04-12 16:47:44 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 229/3300 [49:32<9:50:45, 11.54s/it]2026-04-12 16:47:58,663 | INFO | train_grpo_train | metrics_logged mode=train step=229

                                                    
{'loss': -0.0281, 'grad_norm': 3.2970972061157227, 'learning_rate': 9.30909090909091e-06, 'num_tokens': 430888.0, 'completions/mean_length': 238.125, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 73.80000305175781, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.32129067182540894, 'rewards/meter/std': 0.31284385919570923, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.22903135418891907, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8968750238418579, 'rewards/count_floor/std': 0.06870941072702408, 'rewards/lexical_plausibility/mean': 0.8877338171005249, 'rewards/lexical_plausibility/std': 0.1321529895067215, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.9146140813827515, 'rewards/repeat_penalty/std': 0.11323264241218567, 'rewards/repeat_floor/mean': 0.9933590888977051, 'rewards/repeat_floor/std': 0.00719152856618166, 'rewards/near_duplicate_penalty/mean': 0.9878732562065125, 'rewards/near_duplicate_penalty/std': 0.015880271792411804, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9880892634391785, 'rewards/distinct_2/std': 0.02549082227051258, 'rewards/total_composite/mean': 0.087250255048275, 'rewards/total_composite/std': 0.13232587277889252, 'reward': 0.087250255048275, 'reward_std': 0.13232587277889252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2153828740119934, 'sampling/sampling_logp_difference/max': 1.0914678573608398, 'sampling/importance_sampling_ratio/min': 0.335723340511322, 'sampling/importance_sampling_ratio/mean': 1.0501376390457153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.866685301065445, 'clip_ratio/low_mean': 0.061829304322600365, 'clip_ratio/low_min': 0.061829304322600365, 'clip_ratio/high_mean': 0.05949237011373043, 'clip_ratio/high_max': 0.05949237011373043, 'clip_ratio/region_mean': 0.1213216744363308, 'reward_total_mean': 0.087250255048275, 'reward_meter_mean': 0.32129067182540894, 'reward_meter_std': 0.31284385919570923, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.22903135418891907, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8968750238418579, 'reward_count_floor_std': 0.06870941072702408, 'reward_lexical_plausibility_mean': 0.8877338171005249, 'reward_lexical_plausibility_std': 0.1321529895067215, 'reward_repeat_penalty_mean': 0.9146140813827515, 'reward_repeat_penalty_std': 0.11323264241218567, 'reward_repeat_floor_mean': 0.9933590888977051, 'reward_repeat_floor_std': 0.00719152856618166, 'reward_near_duplicate_penalty_mean': 0.9878732562065125, 'reward_near_duplicate_penalty_std': 0.015880271792411804, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9880892634391785, 'reward_distinct_2_std': 0.02549082227051258, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.087250255048275, 'reward_total_composite_std': 0.13232587277889252, 'epoch': 0.01}

  7%|▋         | 229/3300 [49:32<9:50:45, 11.54s/it]INFO 04-12 16:47:58 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 230/3300 [49:39<8:47:20, 10.31s/it]2026-04-12 16:48:06,087 | INFO | train_grpo_train | metrics_logged mode=train step=230

                                                    
{'loss': 0.1667, 'grad_norm': 24.872583389282227, 'learning_rate': 9.306060606060608e-06, 'num_tokens': 432191.0, 'completions/mean_length': 16.875, 'completions/min_length': 14.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7083169221878052, 'rewards/meter/std': 0.42615315318107605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8062499761581421, 'rewards/judge_quality/std': 0.22012577950954437, 'rewards/repeat_penalty/mean': 0.7123380303382874, 'rewards/repeat_penalty/std': 0.03877650946378708, 'rewards/repeat_floor/mean': 0.9774675965309143, 'rewards/repeat_floor/std': 0.007755298633128405, 'rewards/near_duplicate_penalty/mean': 0.9497840404510498, 'rewards/near_duplicate_penalty/std': 0.05170203372836113, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5452650189399719, 'rewards/total_composite/std': 0.3932492136955261, 'reward': 0.5452650189399719, 'reward_std': 0.39324918389320374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15991055965423584, 'sampling/sampling_logp_difference/max': 1.5523312091827393, 'sampling/importance_sampling_ratio/min': 0.2117537558078766, 'sampling/importance_sampling_ratio/mean': 0.9995664358139038, 'sampling/importance_sampling_ratio/max': 1.9226329326629639, 'entropy': 1.2698263004422188, 'clip_ratio/low_mean': 0.059748932253569365, 'clip_ratio/low_min': 0.059748932253569365, 'clip_ratio/high_mean': 0.008928571827709675, 'clip_ratio/high_max': 0.008928571827709675, 'clip_ratio/region_mean': 0.06867750408127904, 'reward_total_mean': 0.5452650189399719, 'reward_meter_mean': 0.7083169221878052, 'reward_meter_std': 0.42615315318107605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7123380303382874, 'reward_repeat_penalty_std': 0.03877650946378708, 'reward_repeat_floor_mean': 0.9774675965309143, 'reward_repeat_floor_std': 0.007755298633128405, 'reward_near_duplicate_penalty_mean': 0.9497840404510498, 'reward_near_duplicate_penalty_std': 0.05170203372836113, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8062499761581421, 'reward_judge_quality_std': 0.22012577950954437, 'reward_total_composite_mean': 0.5452650189399719, 'reward_total_composite_std': 0.3932492136955261, 'epoch': 0.01}

  7%|▋         | 230/3300 [49:39<8:47:20, 10.31s/it]INFO 04-12 16:48:06 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 231/3300 [49:48<8:14:23,  9.67s/it]2026-04-12 16:48:14,257 | INFO | train_grpo_train | metrics_logged mode=train step=231

                                                    
{'loss': 0.1007, 'grad_norm': 15.607954978942871, 'learning_rate': 9.303030303030303e-06, 'num_tokens': 433877.0, 'completions/mean_length': 40.75, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9608279466629028, 'rewards/meter/std': 0.04227001965045929, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.35847893357276917, 'rewards/repeat_penalty/mean': 0.9857639074325562, 'rewards/repeat_penalty/std': 0.024509375914931297, 'rewards/repeat_floor/mean': 0.9978646039962769, 'rewards/repeat_floor/std': 0.0036764012183994055, 'rewards/near_duplicate_penalty/mean': 0.9857639074325562, 'rewards/near_duplicate_penalty/std': 0.024509375914931297, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5745943784713745, 'rewards/total_composite/std': 0.34490159153938293, 'reward': 0.5745943784713745, 'reward_std': 0.34490156173706055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18984870612621307, 'sampling/sampling_logp_difference/max': 1.342778205871582, 'sampling/importance_sampling_ratio/min': 0.26111921668052673, 'sampling/importance_sampling_ratio/mean': 1.0135283470153809, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.680129811167717, 'clip_ratio/low_mean': 0.05300451535731554, 'clip_ratio/low_min': 0.05300451535731554, 'clip_ratio/high_mean': 0.08902871422469616, 'clip_ratio/high_max': 0.08902871422469616, 'clip_ratio/region_mean': 0.1420332295820117, 'reward_total_mean': 0.5745943784713745, 'reward_meter_mean': 0.9608279466629028, 'reward_meter_std': 0.04227001965045929, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9857639074325562, 'reward_repeat_penalty_std': 0.024509375914931297, 'reward_repeat_floor_mean': 0.9978646039962769, 'reward_repeat_floor_std': 0.0036764012183994055, 'reward_near_duplicate_penalty_mean': 0.9857639074325562, 'reward_near_duplicate_penalty_std': 0.024509375914931297, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.35847893357276917, 'reward_total_composite_mean': 0.5745943784713745, 'reward_total_composite_std': 0.34490159153938293, 'epoch': 0.01}

  7%|▋         | 231/3300 [49:48<8:14:23,  9.67s/it]INFO 04-12 16:48:14 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 232/3300 [50:01<9:04:58, 10.66s/it]2026-04-12 16:48:27,232 | INFO | train_grpo_train | metrics_logged mode=train step=232

                                                    
{'loss': -0.0344, 'grad_norm': 3.6098029613494873, 'learning_rate': 9.3e-06, 'num_tokens': 435583.0, 'completions/mean_length': 159.25, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 41.66666793823242, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.4055835008621216, 'rewards/meter/std': 0.39694225788116455, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.8637757301330566, 'rewards/lexical_plausibility/std': 0.17642712593078613, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.24494898319244385, 'rewards/repeat_penalty/mean': 0.9352866411209106, 'rewards/repeat_penalty/std': 0.11439581960439682, 'rewards/repeat_floor/mean': 0.9959179759025574, 'rewards/repeat_floor/std': 0.006751851178705692, 'rewards/near_duplicate_penalty/mean': 0.9977866411209106, 'rewards/near_duplicate_penalty/std': 0.003120151814073324, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17531438171863556, 'rewards/total_composite/std': 0.2520464062690735, 'reward': 0.17531438171863556, 'reward_std': 0.2520463764667511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22814057767391205, 'sampling/sampling_logp_difference/max': 1.343282699584961, 'sampling/importance_sampling_ratio/min': 0.2609875202178955, 'sampling/importance_sampling_ratio/mean': 1.0814465284347534, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4205938577651978, 'clip_ratio/low_mean': 0.060912005603313446, 'clip_ratio/low_min': 0.060912005603313446, 'clip_ratio/high_mean': 0.06919306237250566, 'clip_ratio/high_max': 0.06919306237250566, 'clip_ratio/region_mean': 0.1301050679758191, 'reward_total_mean': 0.17531438171863556, 'reward_meter_mean': 0.4055835008621216, 'reward_meter_std': 0.39694225788116455, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.8637757301330566, 'reward_lexical_plausibility_std': 0.17642712593078613, 'reward_repeat_penalty_mean': 0.9352866411209106, 'reward_repeat_penalty_std': 0.11439581960439682, 'reward_repeat_floor_mean': 0.9959179759025574, 'reward_repeat_floor_std': 0.006751851178705692, 'reward_near_duplicate_penalty_mean': 0.9977866411209106, 'reward_near_duplicate_penalty_std': 0.003120151814073324, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.24494898319244385, 'reward_total_composite_mean': 0.17531438171863556, 'reward_total_composite_std': 0.2520464062690735, 'epoch': 0.01}

  7%|▋         | 232/3300 [50:01<9:04:58, 10.66s/it]INFO 04-12 16:48:27 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 233/3300 [50:09<8:26:28,  9.91s/it]2026-04-12 16:48:35,389 | INFO | train_grpo_train | metrics_logged mode=train step=233

                                                    
{'loss': 0.1102, 'grad_norm': 16.913036346435547, 'learning_rate': 9.296969696969698e-06, 'num_tokens': 437102.0, 'completions/mean_length': 40.875, 'completions/min_length': 33.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7456052899360657, 'rewards/meter/std': 0.41660934686660767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.2924496531486511, 'rewards/repeat_penalty/mean': 0.9727711081504822, 'rewards/repeat_penalty/std': 0.021970974281430244, 'rewards/repeat_floor/mean': 0.9959156513214111, 'rewards/repeat_floor/std': 0.003295645583420992, 'rewards/near_duplicate_penalty/mean': 0.9727711081504822, 'rewards/near_duplicate_penalty/std': 0.021970974281430244, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4961712062358856, 'rewards/total_composite/std': 0.3597232401371002, 'reward': 0.4961712062358856, 'reward_std': 0.35972318053245544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2133798450231552, 'sampling/sampling_logp_difference/max': 1.676072120666504, 'sampling/importance_sampling_ratio/min': 0.18710745871067047, 'sampling/importance_sampling_ratio/mean': 1.0489169359207153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1937512159347534, 'clip_ratio/low_mean': 0.12542528845369816, 'clip_ratio/low_min': 0.12542528845369816, 'clip_ratio/high_mean': 0.07146460469812155, 'clip_ratio/high_max': 0.07146460469812155, 'clip_ratio/region_mean': 0.1968898931518197, 'reward_total_mean': 0.4961712062358856, 'reward_meter_mean': 0.7456052899360657, 'reward_meter_std': 0.41660934686660767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9727711081504822, 'reward_repeat_penalty_std': 0.021970974281430244, 'reward_repeat_floor_mean': 0.9959156513214111, 'reward_repeat_floor_std': 0.003295645583420992, 'reward_near_duplicate_penalty_mean': 0.9727711081504822, 'reward_near_duplicate_penalty_std': 0.021970974281430244, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.2924496531486511, 'reward_total_composite_mean': 0.4961712062358856, 'reward_total_composite_std': 0.3597232401371002, 'epoch': 0.01}

  7%|▋         | 233/3300 [50:09<8:26:28,  9.91s/it]INFO 04-12 16:48:35 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 234/3300 [50:18<8:11:26,  9.62s/it]2026-04-12 16:48:44,331 | INFO | train_grpo_train | metrics_logged mode=train step=234

                                                    
{'loss': -0.0041, 'grad_norm': 13.80274486541748, 'learning_rate': 9.293939393939395e-06, 'num_tokens': 438821.0, 'completions/mean_length': 42.875, 'completions/min_length': 39.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.973192036151886, 'rewards/meter/std': 0.03658362105488777, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9765625, 'rewards/lexical_plausibility/std': 0.04650149121880531, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.9922782182693481, 'rewards/repeat_penalty/std': 0.018763329833745956, 'rewards/repeat_floor/mean': 0.9988417625427246, 'rewards/repeat_floor/std': 0.0028145047836005688, 'rewards/near_duplicate_penalty/mean': 0.9922782182693481, 'rewards/near_duplicate_penalty/std': 0.018763329833745956, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.44965702295303345, 'rewards/total_composite/std': 0.21406090259552002, 'reward': 0.44965702295303345, 'reward_std': 0.21406088769435883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21099594235420227, 'sampling/sampling_logp_difference/max': 1.0102977752685547, 'sampling/importance_sampling_ratio/min': 0.3641105592250824, 'sampling/importance_sampling_ratio/mean': 1.071935772895813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9485606849193573, 'clip_ratio/low_mean': 0.16430298518389463, 'clip_ratio/low_min': 0.16430298518389463, 'clip_ratio/high_mean': 0.02976190485060215, 'clip_ratio/high_max': 0.02976190485060215, 'clip_ratio/region_mean': 0.19406489003449678, 'reward_total_mean': 0.44965702295303345, 'reward_meter_mean': 0.973192036151886, 'reward_meter_std': 0.03658362105488777, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9765625, 'reward_lexical_plausibility_std': 0.04650149121880531, 'reward_repeat_penalty_mean': 0.9922782182693481, 'reward_repeat_penalty_std': 0.018763329833745956, 'reward_repeat_floor_mean': 0.9988417625427246, 'reward_repeat_floor_std': 0.0028145047836005688, 'reward_near_duplicate_penalty_mean': 0.9922782182693481, 'reward_near_duplicate_penalty_std': 0.018763329833745956, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.44965702295303345, 'reward_total_composite_std': 0.21406090259552002, 'epoch': 0.01}

  7%|▋         | 234/3300 [50:18<8:11:26,  9.62s/it]INFO 04-12 16:48:44 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 235/3300 [50:26<7:56:23,  9.33s/it]2026-04-12 16:48:52,974 | INFO | train_grpo_train | metrics_logged mode=train step=235

                                                    
{'loss': 0.0414, 'grad_norm': 23.255542755126953, 'learning_rate': 9.29090909090909e-06, 'num_tokens': 440167.0, 'completions/mean_length': 17.25, 'completions/min_length': 13.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.25, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.3789360225200653, 'rewards/meter/std': 0.45817628502845764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9275209903717041, 'rewards/lexical_plausibility/std': 0.20500154793262482, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.2881684899330139, 'rewards/repeat_penalty/mean': 0.7433712482452393, 'rewards/repeat_penalty/std': 0.018749047070741653, 'rewards/repeat_floor/mean': 0.9836742877960205, 'rewards/repeat_floor/std': 0.0037498052697628736, 'rewards/near_duplicate_penalty/mean': 0.991161584854126, 'rewards/near_duplicate_penalty/std': 0.024998728185892105, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11721844971179962, 'rewards/total_composite/std': 0.1704958826303482, 'reward': 0.11721844971179962, 'reward_std': 0.1704958826303482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25322791934013367, 'sampling/sampling_logp_difference/max': 1.1516380310058594, 'sampling/importance_sampling_ratio/min': 0.3161185383796692, 'sampling/importance_sampling_ratio/mean': 1.0224978923797607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6132565438747406, 'clip_ratio/low_mean': 0.1485345559194684, 'clip_ratio/low_min': 0.1485345559194684, 'clip_ratio/high_mean': 0.1067708358168602, 'clip_ratio/high_max': 0.1067708358168602, 'clip_ratio/region_mean': 0.2553053917363286, 'reward_total_mean': 0.11721844971179962, 'reward_meter_mean': 0.3789360225200653, 'reward_meter_std': 0.45817628502845764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9275209903717041, 'reward_lexical_plausibility_std': 0.20500154793262482, 'reward_repeat_penalty_mean': 0.7433712482452393, 'reward_repeat_penalty_std': 0.018749047070741653, 'reward_repeat_floor_mean': 0.9836742877960205, 'reward_repeat_floor_std': 0.0037498052697628736, 'reward_near_duplicate_penalty_mean': 0.991161584854126, 'reward_near_duplicate_penalty_std': 0.024998728185892105, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.2881684899330139, 'reward_total_composite_mean': 0.11721844971179962, 'reward_total_composite_std': 0.1704958826303482, 'epoch': 0.01}

  7%|▋         | 235/3300 [50:26<7:56:23,  9.33s/it]INFO 04-12 16:48:53 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 236/3300 [50:35<7:46:02,  9.13s/it]2026-04-12 16:49:01,635 | INFO | train_grpo_train | metrics_logged mode=train step=236

                                                    
{'loss': 0.0216, 'grad_norm': 19.179950714111328, 'learning_rate': 9.28787878787879e-06, 'num_tokens': 441812.0, 'completions/mean_length': 33.625, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7460530996322632, 'rewards/meter/std': 0.3232627511024475, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.97265625, 'rewards/lexical_plausibility/std': 0.07733980566263199, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.9724409580230713, 'rewards/repeat_penalty/std': 0.039793893694877625, 'rewards/repeat_floor/mean': 0.9958661198616028, 'rewards/repeat_floor/std': 0.005969088524580002, 'rewards/near_duplicate_penalty/mean': 0.9724409580230713, 'rewards/near_duplicate_penalty/std': 0.039793893694877625, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3454809784889221, 'rewards/total_composite/std': 0.25839027762413025, 'reward': 0.3454809784889221, 'reward_std': 0.25839027762413025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17248158156871796, 'sampling/sampling_logp_difference/max': 1.304800033569336, 'sampling/importance_sampling_ratio/min': 0.2778701186180115, 'sampling/importance_sampling_ratio/mean': 1.0554769039154053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5852476954460144, 'clip_ratio/low_mean': 0.12302587367594242, 'clip_ratio/low_min': 0.12302587367594242, 'clip_ratio/high_mean': 0.041598313488066196, 'clip_ratio/high_max': 0.041598313488066196, 'clip_ratio/region_mean': 0.16462418716400862, 'reward_total_mean': 0.3454809784889221, 'reward_meter_mean': 0.7460530996322632, 'reward_meter_std': 0.3232627511024475, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.97265625, 'reward_lexical_plausibility_std': 0.07733980566263199, 'reward_repeat_penalty_mean': 0.9724409580230713, 'reward_repeat_penalty_std': 0.039793893694877625, 'reward_repeat_floor_mean': 0.9958661198616028, 'reward_repeat_floor_std': 0.005969088524580002, 'reward_near_duplicate_penalty_mean': 0.9724409580230713, 'reward_near_duplicate_penalty_std': 0.039793893694877625, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.3454809784889221, 'reward_total_composite_std': 0.25839027762413025, 'epoch': 0.01}

  7%|▋         | 236/3300 [50:35<7:46:02,  9.13s/it]INFO 04-12 16:49:01 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 237/3300 [50:44<7:43:42,  9.08s/it]2026-04-12 16:49:10,619 | INFO | train_grpo_train | metrics_logged mode=train step=237

                                                    
{'loss': -0.0162, 'grad_norm': 11.643808364868164, 'learning_rate': 9.284848484848485e-06, 'num_tokens': 443893.0, 'completions/mean_length': 65.125, 'completions/min_length': 53.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.8283839225769043, 'rewards/meter/std': 0.2995067834854126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9709752798080444, 'rewards/lexical_plausibility/std': 0.037827182561159134, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9944260120391846, 'rewards/repeat_penalty/std': 0.007627261336892843, 'rewards/repeat_floor/mean': 0.9991638660430908, 'rewards/repeat_floor/std': 0.001144097768701613, 'rewards/near_duplicate_penalty/mean': 0.9944260120391846, 'rewards/near_duplicate_penalty/std': 0.007627261336892843, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27146846055984497, 'rewards/total_composite/std': 0.14824147522449493, 'reward': 0.27146846055984497, 'reward_std': 0.14824147522449493, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21287813782691956, 'sampling/sampling_logp_difference/max': 1.6447405815124512, 'sampling/importance_sampling_ratio/min': 0.19306263327598572, 'sampling/importance_sampling_ratio/mean': 1.0562447309494019, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.782700479030609, 'clip_ratio/low_mean': 0.07221325859427452, 'clip_ratio/low_min': 0.07221325859427452, 'clip_ratio/high_mean': 0.11596426740288734, 'clip_ratio/high_max': 0.11596426740288734, 'clip_ratio/region_mean': 0.18817752599716187, 'reward_total_mean': 0.27146846055984497, 'reward_meter_mean': 0.8283839225769043, 'reward_meter_std': 0.2995067834854126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9709752798080444, 'reward_lexical_plausibility_std': 0.037827182561159134, 'reward_repeat_penalty_mean': 0.9944260120391846, 'reward_repeat_penalty_std': 0.007627261336892843, 'reward_repeat_floor_mean': 0.9991638660430908, 'reward_repeat_floor_std': 0.001144097768701613, 'reward_near_duplicate_penalty_mean': 0.9944260120391846, 'reward_near_duplicate_penalty_std': 0.007627261336892843, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.27146846055984497, 'reward_total_composite_std': 0.14824147522449493, 'epoch': 0.01}

  7%|▋         | 237/3300 [50:44<7:43:42,  9.08s/it]INFO 04-12 16:49:10 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 238/3300 [50:59<9:08:07, 10.74s/it]2026-04-12 16:49:25,225 | INFO | train_grpo_train | metrics_logged mode=train step=238

                                                    
{'loss': -0.1321, 'grad_norm': 5.200933933258057, 'learning_rate': 9.281818181818183e-06, 'num_tokens': 445691.0, 'completions/mean_length': 107.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.000003814697266, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9475964307785034, 'rewards/meter/std': 0.07346004992723465, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.3053750991821289, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.09161254018545151, 'rewards/lexical_plausibility/mean': 0.9088233709335327, 'rewards/lexical_plausibility/std': 0.16669966280460358, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.7485313415527344, 'rewards/repeat_penalty/std': 0.22189751267433167, 'rewards/repeat_floor/mean': 0.974923849105835, 'rewards/repeat_floor/std': 0.028835337609052658, 'rewards/near_duplicate_penalty/mean': 0.9394485354423523, 'rewards/near_duplicate_penalty/std': 0.09532373398542404, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8847955465316772, 'rewards/distinct_2/std': 0.14293035864830017, 'rewards/total_composite/mean': 0.22168511152267456, 'rewards/total_composite/std': 0.13467872142791748, 'reward': 0.22168511152267456, 'reward_std': 0.13467872142791748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19657887518405914, 'sampling/sampling_logp_difference/max': 1.1651115417480469, 'sampling/importance_sampling_ratio/min': 0.31188786029815674, 'sampling/importance_sampling_ratio/mean': 1.0709632635116577, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1937649250030518, 'clip_ratio/low_mean': 0.0767676793038845, 'clip_ratio/low_min': 0.0767676793038845, 'clip_ratio/high_mean': 0.08816965855658054, 'clip_ratio/high_max': 0.08816965855658054, 'clip_ratio/region_mean': 0.16493733786046505, 'reward_total_mean': 0.22168511152267456, 'reward_meter_mean': 0.9475964307785034, 'reward_meter_std': 0.07346004992723465, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.3053750991821289, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.09161254018545151, 'reward_lexical_plausibility_mean': 0.9088233709335327, 'reward_lexical_plausibility_std': 0.16669966280460358, 'reward_repeat_penalty_mean': 0.7485313415527344, 'reward_repeat_penalty_std': 0.22189751267433167, 'reward_repeat_floor_mean': 0.974923849105835, 'reward_repeat_floor_std': 0.028835337609052658, 'reward_near_duplicate_penalty_mean': 0.9394485354423523, 'reward_near_duplicate_penalty_std': 0.09532373398542404, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8847955465316772, 'reward_distinct_2_std': 0.14293035864830017, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.22168511152267456, 'reward_total_composite_std': 0.13467872142791748, 'epoch': 0.01}

  7%|▋         | 238/3300 [50:59<9:08:07, 10.74s/it]INFO 04-12 16:49:25 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 239/3300 [51:08<8:42:13, 10.24s/it]2026-04-12 16:49:34,285 | INFO | train_grpo_train | metrics_logged mode=train step=239

                                                    
{'loss': 0.0353, 'grad_norm': 19.865358352661133, 'learning_rate': 9.27878787878788e-06, 'num_tokens': 447175.0, 'completions/mean_length': 21.5, 'completions/min_length': 14.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7912949919700623, 'rewards/meter/std': 0.3438830077648163, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.7480332255363464, 'rewards/repeat_penalty/std': 0.005562899634242058, 'rewards/repeat_floor/mean': 0.9846066236495972, 'rewards/repeat_floor/std': 0.0011125925229862332, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5536940097808838, 'rewards/total_composite/std': 0.31100213527679443, 'reward': 0.5536940097808838, 'reward_std': 0.31100213527679443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19315433502197266, 'sampling/sampling_logp_difference/max': 1.5123519897460938, 'sampling/importance_sampling_ratio/min': 0.22039102017879486, 'sampling/importance_sampling_ratio/mean': 1.0412137508392334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7123451083898544, 'clip_ratio/low_mean': 0.10180191276594996, 'clip_ratio/low_min': 0.10180191276594996, 'clip_ratio/high_mean': 0.0555108361877501, 'clip_ratio/high_max': 0.0555108361877501, 'clip_ratio/region_mean': 0.15731274895370007, 'reward_total_mean': 0.5536940097808838, 'reward_meter_mean': 0.7912949919700623, 'reward_meter_std': 0.3438830077648163, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7480332255363464, 'reward_repeat_penalty_std': 0.005562899634242058, 'reward_repeat_floor_mean': 0.9846066236495972, 'reward_repeat_floor_std': 0.0011125925229862332, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.5536940097808838, 'reward_total_composite_std': 0.31100213527679443, 'epoch': 0.01}

  7%|▋         | 239/3300 [51:08<8:42:13, 10.24s/it]INFO 04-12 16:49:34 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 240/3300 [51:21<9:37:40, 11.33s/it]2026-04-12 16:49:48,157 | INFO | train_grpo_train | metrics_logged mode=train step=240

                                                    
{'loss': -0.1714, 'grad_norm': 2.913806200027466, 'learning_rate': 9.275757575757577e-06, 'num_tokens': 449294.0, 'completions/mean_length': 376.875, 'completions/min_length': 143.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 151.6666717529297, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.6958625912666321, 'rewards/meter/std': 0.4179382920265198, 'rewards/count_adherence/mean': 0.6607142686843872, 'rewards/count_adherence/std': 0.3234066069126129, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8982142806053162, 'rewards/count_floor/std': 0.09702198207378387, 'rewards/lexical_plausibility/mean': 0.8450936675071716, 'rewards/lexical_plausibility/std': 0.13930977880954742, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9677689075469971, 'rewards/repeat_penalty/std': 0.055305365473032, 'rewards/repeat_floor/mean': 0.9966572523117065, 'rewards/repeat_floor/std': 0.005531198810786009, 'rewards/near_duplicate_penalty/mean': 0.993362307548523, 'rewards/near_duplicate_penalty/std': 0.008659596554934978, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.973921000957489, 'rewards/distinct_2/std': 0.04876678064465523, 'rewards/total_composite/mean': 0.11358870565891266, 'rewards/total_composite/std': 0.13763585686683655, 'reward': 0.11358870565891266, 'reward_std': 0.13763584196567535, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23199371993541718, 'sampling/sampling_logp_difference/max': 1.8094148635864258, 'sampling/importance_sampling_ratio/min': 0.16374991834163666, 'sampling/importance_sampling_ratio/mean': 1.075723648071289, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1606292128562927, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07492086850106716, 'clip_ratio/high_max': 0.07492086850106716, 'clip_ratio/region_mean': 0.07492086850106716, 'reward_total_mean': 0.11358870565891266, 'reward_meter_mean': 0.6958625912666321, 'reward_meter_std': 0.4179382920265198, 'reward_count_adherence_mean': 0.6607142686843872, 'reward_count_adherence_std': 0.3234066069126129, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8982142806053162, 'reward_count_floor_std': 0.09702198207378387, 'reward_lexical_plausibility_mean': 0.8450936675071716, 'reward_lexical_plausibility_std': 0.13930977880954742, 'reward_repeat_penalty_mean': 0.9677689075469971, 'reward_repeat_penalty_std': 0.055305365473032, 'reward_repeat_floor_mean': 0.9966572523117065, 'reward_repeat_floor_std': 0.005531198810786009, 'reward_near_duplicate_penalty_mean': 0.993362307548523, 'reward_near_duplicate_penalty_std': 0.008659596554934978, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.973921000957489, 'reward_distinct_2_std': 0.04876678064465523, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.11358870565891266, 'reward_total_composite_std': 0.13763585686683655, 'epoch': 0.01}

  7%|▋         | 240/3300 [51:22<9:37:40, 11.33s/it]INFO 04-12 16:49:48 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 241/3300 [51:35<10:12:45, 12.02s/it]2026-04-12 16:50:01,789 | INFO | train_grpo_train | metrics_logged mode=train step=241

                                                     
{'loss': -0.0291, 'grad_norm': 6.813591480255127, 'learning_rate': 9.272727272727273e-06, 'num_tokens': 451161.0, 'completions/mean_length': 122.375, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 66.71428680419922, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.3726944029331207, 'rewards/meter/std': 0.37314772605895996, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9808240532875061, 'rewards/lexical_plausibility/std': 0.03568333387374878, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9513874053955078, 'rewards/repeat_penalty/std': 0.053269632160663605, 'rewards/repeat_floor/mean': 0.9943923950195312, 'rewards/repeat_floor/std': 0.005572207272052765, 'rewards/near_duplicate_penalty/mean': 0.9808699488639832, 'rewards/near_duplicate_penalty/std': 0.012911330908536911, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9695764183998108, 'rewards/distinct_2/std': 0.04475013539195061, 'rewards/total_composite/mean': 0.1334598958492279, 'rewards/total_composite/std': 0.14915242791175842, 'reward': 0.1334598958492279, 'reward_std': 0.14915242791175842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21868832409381866, 'sampling/sampling_logp_difference/max': 1.4155559539794922, 'sampling/importance_sampling_ratio/min': 0.2427905946969986, 'sampling/importance_sampling_ratio/mean': 1.0160592794418335, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7901955991983414, 'clip_ratio/low_mean': 0.11398501694202423, 'clip_ratio/low_min': 0.11398501694202423, 'clip_ratio/high_mean': 0.05323639698326588, 'clip_ratio/high_max': 0.05323639698326588, 'clip_ratio/region_mean': 0.1672214139252901, 'reward_total_mean': 0.1334598958492279, 'reward_meter_mean': 0.3726944029331207, 'reward_meter_std': 0.37314772605895996, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9808240532875061, 'reward_lexical_plausibility_std': 0.03568333387374878, 'reward_repeat_penalty_mean': 0.9513874053955078, 'reward_repeat_penalty_std': 0.053269632160663605, 'reward_repeat_floor_mean': 0.9943923950195312, 'reward_repeat_floor_std': 0.005572207272052765, 'reward_near_duplicate_penalty_mean': 0.9808699488639832, 'reward_near_duplicate_penalty_std': 0.012911330908536911, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9695764183998108, 'reward_distinct_2_std': 0.04475013539195061, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.1334598958492279, 'reward_total_composite_std': 0.14915242791175842, 'epoch': 0.01}

  7%|▋         | 241/3300 [51:35<10:12:45, 12.02s/it]INFO 04-12 16:50:02 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 242/3300 [51:43<9:05:17, 10.70s/it] 2026-04-12 16:50:09,431 | INFO | train_grpo_train | metrics_logged mode=train step=242

                                                    
{'loss': 0.0805, 'grad_norm': 31.87870979309082, 'learning_rate': 9.26969696969697e-06, 'num_tokens': 452529.0, 'completions/mean_length': 16.0, 'completions/min_length': 15.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.9850348234176636, 'rewards/meter/std': 0.013144196011126041, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8137500286102295, 'rewards/judge_quality/std': 0.2249404639005661, 'rewards/repeat_penalty/mean': 0.7432166337966919, 'rewards/repeat_penalty/std': 0.019186342135071754, 'rewards/repeat_floor/mean': 0.9836432933807373, 'rewards/repeat_floor/std': 0.0038372809067368507, 'rewards/near_duplicate_penalty/mean': 0.9909554719924927, 'rewards/near_duplicate_penalty/std': 0.025581790134310722, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7880934476852417, 'rewards/total_composite/std': 0.21708205342292786, 'reward': 0.7880934476852417, 'reward_std': 0.21708205342292786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18875885009765625, 'sampling/sampling_logp_difference/max': 1.651460886001587, 'sampling/importance_sampling_ratio/min': 0.191769540309906, 'sampling/importance_sampling_ratio/mean': 1.0643130540847778, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4098713919520378, 'clip_ratio/low_mean': 0.05147058889269829, 'clip_ratio/low_min': 0.05147058889269829, 'clip_ratio/high_mean': 0.09445465821772814, 'clip_ratio/high_max': 0.09445465821772814, 'clip_ratio/region_mean': 0.14592524711042643, 'reward_total_mean': 0.7880934476852417, 'reward_meter_mean': 0.9850348234176636, 'reward_meter_std': 0.013144196011126041, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7432166337966919, 'reward_repeat_penalty_std': 0.019186342135071754, 'reward_repeat_floor_mean': 0.9836432933807373, 'reward_repeat_floor_std': 0.0038372809067368507, 'reward_near_duplicate_penalty_mean': 0.9909554719924927, 'reward_near_duplicate_penalty_std': 0.025581790134310722, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8137500286102295, 'reward_judge_quality_std': 0.2249404639005661, 'reward_total_composite_mean': 0.7880934476852417, 'reward_total_composite_std': 0.21708205342292786, 'epoch': 0.01}

  7%|▋         | 242/3300 [51:43<9:05:17, 10.70s/it]INFO 04-12 16:50:09 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 243/3300 [51:57<9:57:30, 11.73s/it]2026-04-12 16:50:23,535 | INFO | train_grpo_train | metrics_logged mode=train step=243

                                                    
{'loss': -0.1724, 'grad_norm': 2.7820847034454346, 'learning_rate': 9.266666666666667e-06, 'num_tokens': 454720.0, 'completions/mean_length': 331.875, 'completions/min_length': 140.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 151.75, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 163.0, 'rewards/meter/mean': 0.7052256464958191, 'rewards/meter/std': 0.3101499378681183, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.3061862289905548, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.09185586124658585, 'rewards/lexical_plausibility/mean': 0.8455607295036316, 'rewards/lexical_plausibility/std': 0.14618223905563354, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.08864052593708038, 'rewards/repeat_penalty/mean': 0.9469081163406372, 'rewards/repeat_penalty/std': 0.07134432345628738, 'rewards/repeat_floor/mean': 0.9948353171348572, 'rewards/repeat_floor/std': 0.006482340395450592, 'rewards/near_duplicate_penalty/mean': 0.9870542287826538, 'rewards/near_duplicate_penalty/std': 0.01594948209822178, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.9760956764221191, 'rewards/distinct_2/std': 0.03492533788084984, 'rewards/total_composite/mean': 0.09603214263916016, 'rewards/total_composite/std': 0.09496942162513733, 'reward': 0.09603214263916016, 'reward_std': 0.09496942162513733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24829797446727753, 'sampling/sampling_logp_difference/max': 1.7428979873657227, 'sampling/importance_sampling_ratio/min': 0.1750124841928482, 'sampling/importance_sampling_ratio/mean': 1.0601587295532227, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9971315264701843, 'clip_ratio/low_mean': 0.02083333395421505, 'clip_ratio/low_min': 0.02083333395421505, 'clip_ratio/high_mean': 0.06954973749816418, 'clip_ratio/high_max': 0.06954973749816418, 'clip_ratio/region_mean': 0.09038307145237923, 'reward_total_mean': 0.09603214263916016, 'reward_meter_mean': 0.7052256464958191, 'reward_meter_std': 0.3101499378681183, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.3061862289905548, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.09185586124658585, 'reward_lexical_plausibility_mean': 0.8455607295036316, 'reward_lexical_plausibility_std': 0.14618223905563354, 'reward_repeat_penalty_mean': 0.9469081163406372, 'reward_repeat_penalty_std': 0.07134432345628738, 'reward_repeat_floor_mean': 0.9948353171348572, 'reward_repeat_floor_std': 0.006482340395450592, 'reward_near_duplicate_penalty_mean': 0.9870542287826538, 'reward_near_duplicate_penalty_std': 0.01594948209822178, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.9760956764221191, 'reward_distinct_2_std': 0.03492533788084984, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.08864052593708038, 'reward_total_composite_mean': 0.09603214263916016, 'reward_total_composite_std': 0.09496942162513733, 'epoch': 0.01}

  7%|▋         | 243/3300 [51:57<9:57:30, 11.73s/it]INFO 04-12 16:50:23 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 244/3300 [52:05<9:04:03, 10.68s/it]2026-04-12 16:50:31,778 | INFO | train_grpo_train | metrics_logged mode=train step=244

                                                    
{'loss': -0.0167, 'grad_norm': 15.387537002563477, 'learning_rate': 9.263636363636364e-06, 'num_tokens': 456417.0, 'completions/mean_length': 42.125, 'completions/min_length': 35.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9270445704460144, 'rewards/meter/std': 0.09438148140907288, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9555628299713135, 'rewards/lexical_plausibility/std': 0.0722915381193161, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.32262319326400757, 'rewards/repeat_penalty/mean': 0.992059588432312, 'rewards/repeat_penalty/std': 0.010410577058792114, 'rewards/repeat_floor/mean': 0.9988089203834534, 'rewards/repeat_floor/std': 0.0015615939628332853, 'rewards/near_duplicate_penalty/mean': 0.992059588432312, 'rewards/near_duplicate_penalty/std': 0.010410577058792114, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.39052316546440125, 'rewards/total_composite/std': 0.2810784578323364, 'reward': 0.39052316546440125, 'reward_std': 0.2810784578323364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22290082275867462, 'sampling/sampling_logp_difference/max': 1.8054604530334473, 'sampling/importance_sampling_ratio/min': 0.16439872980117798, 'sampling/importance_sampling_ratio/mean': 1.0307132005691528, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.390533074736595, 'clip_ratio/low_mean': 0.09051434975117445, 'clip_ratio/low_min': 0.09051434975117445, 'clip_ratio/high_mean': 0.055435288697481155, 'clip_ratio/high_max': 0.055435288697481155, 'clip_ratio/region_mean': 0.1459496384486556, 'reward_total_mean': 0.39052316546440125, 'reward_meter_mean': 0.9270445704460144, 'reward_meter_std': 0.09438148140907288, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9555628299713135, 'reward_lexical_plausibility_std': 0.0722915381193161, 'reward_repeat_penalty_mean': 0.992059588432312, 'reward_repeat_penalty_std': 0.010410577058792114, 'reward_repeat_floor_mean': 0.9988089203834534, 'reward_repeat_floor_std': 0.0015615939628332853, 'reward_near_duplicate_penalty_mean': 0.992059588432312, 'reward_near_duplicate_penalty_std': 0.010410577058792114, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.32262319326400757, 'reward_total_composite_mean': 0.39052316546440125, 'reward_total_composite_std': 0.2810784578323364, 'epoch': 0.01}

  7%|▋         | 244/3300 [52:05<9:04:03, 10.68s/it]INFO 04-12 16:50:32 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 245/3300 [52:14<8:32:59, 10.08s/it]2026-04-12 16:50:40,437 | INFO | train_grpo_train | metrics_logged mode=train step=245

                                                    
{'loss': 0.0098, 'grad_norm': 20.19318199157715, 'learning_rate': 9.260606060606062e-06, 'num_tokens': 457877.0, 'completions/mean_length': 32.5, 'completions/min_length': 28.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.48916447162628174, 'rewards/meter/std': 0.3811974823474884, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.23868316411972046, 'rewards/repeat_penalty/mean': 0.8754228353500366, 'rewards/repeat_penalty/std': 0.26593077182769775, 'rewards/repeat_floor/mean': 0.9854813814163208, 'rewards/repeat_floor/std': 0.029772454872727394, 'rewards/near_duplicate_penalty/mean': 0.9524838924407959, 'rewards/near_duplicate_penalty/std': 0.07924303412437439, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9387089014053345, 'rewards/distinct_2/std': 0.14200736582279205, 'rewards/total_composite/mean': 0.18750335276126862, 'rewards/total_composite/std': 0.1550569087266922, 'reward': 0.18750335276126862, 'reward_std': 0.1550569236278534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1468329280614853, 'sampling/sampling_logp_difference/max': 1.510289192199707, 'sampling/importance_sampling_ratio/min': 0.22084610164165497, 'sampling/importance_sampling_ratio/mean': 1.0243041515350342, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.174165040254593, 'clip_ratio/low_mean': 0.09172544442117214, 'clip_ratio/low_min': 0.09172544442117214, 'clip_ratio/high_mean': 0.059171365574002266, 'clip_ratio/high_max': 0.059171365574002266, 'clip_ratio/region_mean': 0.1508968099951744, 'reward_total_mean': 0.18750335276126862, 'reward_meter_mean': 0.48916447162628174, 'reward_meter_std': 0.3811974823474884, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8754228353500366, 'reward_repeat_penalty_std': 0.26593077182769775, 'reward_repeat_floor_mean': 0.9854813814163208, 'reward_repeat_floor_std': 0.029772454872727394, 'reward_near_duplicate_penalty_mean': 0.9524838924407959, 'reward_near_duplicate_penalty_std': 0.07924303412437439, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9387089014053345, 'reward_distinct_2_std': 0.14200736582279205, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.23868316411972046, 'reward_total_composite_mean': 0.18750335276126862, 'reward_total_composite_std': 0.1550569087266922, 'epoch': 0.01}

  7%|▋         | 245/3300 [52:14<8:32:59, 10.08s/it]INFO 04-12 16:50:40 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 246/3300 [52:29<9:47:14, 11.54s/it]2026-04-12 16:50:55,386 | INFO | train_grpo_train | metrics_logged mode=train step=246

                                                    
{'loss': -0.1008, 'grad_norm': 6.182137489318848, 'learning_rate': 9.257575757575759e-06, 'num_tokens': 459822.0, 'completions/mean_length': 130.125, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 75.5714340209961, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.6792052984237671, 'rewards/meter/std': 0.428801566362381, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9701545834541321, 'rewards/lexical_plausibility/std': 0.07745271176099777, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.7933319807052612, 'rewards/repeat_penalty/std': 0.2632565200328827, 'rewards/repeat_floor/mean': 0.9757997989654541, 'rewards/repeat_floor/std': 0.036176349967718124, 'rewards/near_duplicate_penalty/mean': 0.9379631876945496, 'rewards/near_duplicate_penalty/std': 0.09891653060913086, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.86054527759552, 'rewards/distinct_2/std': 0.1846321076154709, 'rewards/total_composite/mean': 0.2275792360305786, 'rewards/total_composite/std': 0.14797452092170715, 'reward': 0.2275792360305786, 'reward_std': 0.14797452092170715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21234148740768433, 'sampling/sampling_logp_difference/max': 1.5397872924804688, 'sampling/importance_sampling_ratio/min': 0.21442671120166779, 'sampling/importance_sampling_ratio/mean': 1.0555132627487183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1686637550592422, 'clip_ratio/low_mean': 0.050833333283662796, 'clip_ratio/low_min': 0.050833333283662796, 'clip_ratio/high_mean': 0.09622169937938452, 'clip_ratio/high_max': 0.09622169937938452, 'clip_ratio/region_mean': 0.1470550326630473, 'reward_total_mean': 0.2275792360305786, 'reward_meter_mean': 0.6792052984237671, 'reward_meter_std': 0.428801566362381, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9701545834541321, 'reward_lexical_plausibility_std': 0.07745271176099777, 'reward_repeat_penalty_mean': 0.7933319807052612, 'reward_repeat_penalty_std': 0.2632565200328827, 'reward_repeat_floor_mean': 0.9757997989654541, 'reward_repeat_floor_std': 0.036176349967718124, 'reward_near_duplicate_penalty_mean': 0.9379631876945496, 'reward_near_duplicate_penalty_std': 0.09891653060913086, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.86054527759552, 'reward_distinct_2_std': 0.1846321076154709, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.2275792360305786, 'reward_total_composite_std': 0.14797452092170715, 'epoch': 0.01}

  7%|▋         | 246/3300 [52:29<9:47:14, 11.54s/it]INFO 04-12 16:50:55 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 247/3300 [52:37<8:56:02, 10.53s/it]2026-04-12 16:51:03,583 | INFO | train_grpo_train | metrics_logged mode=train step=247

                                                    
{'loss': 0.0586, 'grad_norm': 18.777450561523438, 'learning_rate': 9.254545454545454e-06, 'num_tokens': 461181.0, 'completions/mean_length': 30.875, 'completions/min_length': 27.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.2120095193386078, 'rewards/meter/std': 0.13663122057914734, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.3358757197856903, 'rewards/repeat_penalty/mean': 0.9969503879547119, 'rewards/repeat_penalty/std': 0.00634540943428874, 'rewards/repeat_floor/mean': 0.9995425343513489, 'rewards/repeat_floor/std': 0.000951816386077553, 'rewards/near_duplicate_penalty/mean': 0.9969503879547119, 'rewards/near_duplicate_penalty/std': 0.00634540943428874, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12416981160640717, 'rewards/total_composite/std': 0.09541624039411545, 'reward': 0.12416981160640717, 'reward_std': 0.09541624039411545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1698114573955536, 'sampling/sampling_logp_difference/max': 1.595409870147705, 'sampling/importance_sampling_ratio/min': 0.2028253823518753, 'sampling/importance_sampling_ratio/mean': 1.03268301486969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.291483037173748, 'clip_ratio/low_mean': 0.037844214122742414, 'clip_ratio/low_min': 0.037844214122742414, 'clip_ratio/high_mean': 0.0858754962682724, 'clip_ratio/high_max': 0.0858754962682724, 'clip_ratio/region_mean': 0.12371971039101481, 'reward_total_mean': 0.12416981160640717, 'reward_meter_mean': 0.2120095193386078, 'reward_meter_std': 0.13663122057914734, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9969503879547119, 'reward_repeat_penalty_std': 0.00634540943428874, 'reward_repeat_floor_mean': 0.9995425343513489, 'reward_repeat_floor_std': 0.000951816386077553, 'reward_near_duplicate_penalty_mean': 0.9969503879547119, 'reward_near_duplicate_penalty_std': 0.00634540943428874, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.3358757197856903, 'reward_total_composite_mean': 0.12416981160640717, 'reward_total_composite_std': 0.09541624039411545, 'epoch': 0.01}

  7%|▋         | 247/3300 [52:37<8:56:02, 10.53s/it]INFO 04-12 16:51:03 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 248/3300 [52:51<9:45:54, 11.52s/it]2026-04-12 16:51:17,396 | INFO | train_grpo_train | metrics_logged mode=train step=248

                                                    
{'loss': -0.061, 'grad_norm': 2.320683479309082, 'learning_rate': 9.251515151515152e-06, 'num_tokens': 462772.0, 'completions/mean_length': 212.875, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 33.400001525878906, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5926539897918701, 'rewards/meter/std': 0.4195498526096344, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9128397703170776, 'rewards/lexical_plausibility/std': 0.12853839993476868, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.25495097041130066, 'rewards/repeat_penalty/mean': 0.8760510683059692, 'rewards/repeat_penalty/std': 0.17846953868865967, 'rewards/repeat_floor/mean': 0.989148736000061, 'rewards/repeat_floor/std': 0.01759549416601658, 'rewards/near_duplicate_penalty/mean': 0.9766964912414551, 'rewards/near_duplicate_penalty/std': 0.047928716987371445, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.24713833630084991, 'rewards/total_composite/std': 0.25866013765335083, 'reward': 0.24713833630084991, 'reward_std': 0.25866013765335083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22261904180049896, 'sampling/sampling_logp_difference/max': 1.378725528717041, 'sampling/importance_sampling_ratio/min': 0.251899391412735, 'sampling/importance_sampling_ratio/mean': 1.0522751808166504, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2500000447034836, 'clip_ratio/low_mean': 0.0173611119389534, 'clip_ratio/low_min': 0.0173611119389534, 'clip_ratio/high_mean': 0.1114067193120718, 'clip_ratio/high_max': 0.1114067193120718, 'clip_ratio/region_mean': 0.1287678312510252, 'reward_total_mean': 0.24713833630084991, 'reward_meter_mean': 0.5926539897918701, 'reward_meter_std': 0.4195498526096344, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9128397703170776, 'reward_lexical_plausibility_std': 0.12853839993476868, 'reward_repeat_penalty_mean': 0.8760510683059692, 'reward_repeat_penalty_std': 0.17846953868865967, 'reward_repeat_floor_mean': 0.989148736000061, 'reward_repeat_floor_std': 0.01759549416601658, 'reward_near_duplicate_penalty_mean': 0.9766964912414551, 'reward_near_duplicate_penalty_std': 0.047928716987371445, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.25495097041130066, 'reward_total_composite_mean': 0.24713833630084991, 'reward_total_composite_std': 0.25866013765335083, 'epoch': 0.01}

  8%|▊         | 248/3300 [52:51<9:45:54, 11.52s/it]INFO 04-12 16:51:17 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 249/3300 [53:01<9:24:51, 11.11s/it]2026-04-12 16:51:27,547 | INFO | train_grpo_train | metrics_logged mode=train step=249

                                                    
{'loss': 0.1052, 'grad_norm': 7.700905799865723, 'learning_rate': 9.248484848484849e-06, 'num_tokens': 465589.0, 'completions/mean_length': 164.125, 'completions/min_length': 122.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 164.125, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.8245628476142883, 'rewards/meter/std': 0.19125983119010925, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.1079898476600647, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9571428298950195, 'rewards/count_floor/std': 0.03239695355296135, 'rewards/lexical_plausibility/mean': 0.9769822359085083, 'rewards/lexical_plausibility/std': 0.043875548988580704, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.5602109432220459, 'rewards/repeat_penalty/std': 0.4092436730861664, 'rewards/repeat_floor/mean': 0.949129045009613, 'rewards/repeat_floor/std': 0.04875320568680763, 'rewards/near_duplicate_penalty/mean': 0.8983430862426758, 'rewards/near_duplicate_penalty/std': 0.10049480199813843, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.6302369832992554, 'rewards/distinct_2/std': 0.3503389358520508, 'rewards/total_composite/mean': 0.18337294459342957, 'rewards/total_composite/std': 0.09522434324026108, 'reward': 0.18337294459342957, 'reward_std': 0.09522434324026108, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18656452000141144, 'sampling/sampling_logp_difference/max': 1.7262134552001953, 'sampling/importance_sampling_ratio/min': 0.1779569834470749, 'sampling/importance_sampling_ratio/mean': 1.05612313747406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4170335233211517, 'clip_ratio/low_mean': 0.07984719797968864, 'clip_ratio/low_min': 0.07984719797968864, 'clip_ratio/high_mean': 0.08934582304209471, 'clip_ratio/high_max': 0.08934582304209471, 'clip_ratio/region_mean': 0.16919302102178335, 'reward_total_mean': 0.18337294459342957, 'reward_meter_mean': 0.8245628476142883, 'reward_meter_std': 0.19125983119010925, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.1079898476600647, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9571428298950195, 'reward_count_floor_std': 0.03239695355296135, 'reward_lexical_plausibility_mean': 0.9769822359085083, 'reward_lexical_plausibility_std': 0.043875548988580704, 'reward_repeat_penalty_mean': 0.5602109432220459, 'reward_repeat_penalty_std': 0.4092436730861664, 'reward_repeat_floor_mean': 0.949129045009613, 'reward_repeat_floor_std': 0.04875320568680763, 'reward_near_duplicate_penalty_mean': 0.8983430862426758, 'reward_near_duplicate_penalty_std': 0.10049480199813843, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.6302369832992554, 'reward_distinct_2_std': 0.3503389358520508, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.18337294459342957, 'reward_total_composite_std': 0.09522434324026108, 'epoch': 0.01}

  8%|▊         | 249/3300 [53:01<9:24:51, 11.11s/it]INFO 04-12 16:51:27 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 250/3300 [53:09<8:46:06, 10.35s/it]2026-04-12 16:51:36,126 | INFO | train_grpo_train | metrics_logged mode=train step=250

                                                    
{'loss': 0.0255, 'grad_norm': 14.192955017089844, 'learning_rate': 9.245454545454546e-06, 'num_tokens': 467371.0, 'completions/mean_length': 58.75, 'completions/min_length': 50.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.383329838514328, 'rewards/meter/std': 0.3598807454109192, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.10350983589887619, 'rewards/repeat_penalty/mean': 0.9382904171943665, 'rewards/repeat_penalty/std': 0.1468064785003662, 'rewards/repeat_floor/mean': 0.9946203231811523, 'rewards/repeat_floor/std': 0.011121366173028946, 'rewards/near_duplicate_penalty/mean': 0.9795204401016235, 'rewards/near_duplicate_penalty/std': 0.03185485675930977, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9743589758872986, 'rewards/distinct_2/std': 0.07252377271652222, 'rewards/total_composite/mean': 0.12034064531326294, 'rewards/total_composite/std': 0.09669128805398941, 'reward': 0.12034064531326294, 'reward_std': 0.09669128805398941, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22268080711364746, 'sampling/sampling_logp_difference/max': 1.697484016418457, 'sampling/importance_sampling_ratio/min': 0.26028621196746826, 'sampling/importance_sampling_ratio/mean': 1.0251330137252808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.087587743997574, 'clip_ratio/low_mean': 0.0946830315515399, 'clip_ratio/low_min': 0.0946830315515399, 'clip_ratio/high_mean': 0.11120689660310745, 'clip_ratio/high_max': 0.11120689660310745, 'clip_ratio/region_mean': 0.20588992815464735, 'reward_total_mean': 0.12034064531326294, 'reward_meter_mean': 0.383329838514328, 'reward_meter_std': 0.3598807454109192, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9382904171943665, 'reward_repeat_penalty_std': 0.1468064785003662, 'reward_repeat_floor_mean': 0.9946203231811523, 'reward_repeat_floor_std': 0.011121366173028946, 'reward_near_duplicate_penalty_mean': 0.9795204401016235, 'reward_near_duplicate_penalty_std': 0.03185485675930977, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9743589758872986, 'reward_distinct_2_std': 0.07252377271652222, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.10350983589887619, 'reward_total_composite_mean': 0.12034064531326294, 'reward_total_composite_std': 0.09669128805398941, 'epoch': 0.01}

  8%|▊         | 250/3300 [53:09<8:46:06, 10.35s/it]INFO 04-12 16:51:36 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:10<00:55,  5.03s/it]

 23%|██▎       | 3/13 [00:16<00:56,  5.63s/it]

 31%|███       | 4/13 [00:30<01:17,  8.57s/it]

 38%|███▊      | 5/13 [00:41<01:15,  9.48s/it]

 46%|████▌     | 6/13 [00:51<01:07,  9.61s/it]

 54%|█████▍    | 7/13 [01:02<01:00, 10.03s/it]

 62%|██████▏   | 8/13 [01:13<00:52, 10.56s/it]

 69%|██████▉   | 9/13 [01:22<00:40, 10.15s/it]

 77%|███████▋  | 10/13 [01:34<00:31, 10.53s/it]

 85%|████████▍ | 11/13 [01:45<00:21, 10.69s/it]

 92%|█████████▏| 12/13 [01:56<00:10, 10.67s/it]

100%|██████████| 13/13 [02:04<00:00,  9.89s/it]2026-04-12 16:53:49,597 | INFO | train_grpo_train | metrics_logged mode=eval step=250

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 133.4458, 'eval_samples_per_second': 0.779, 'eval_steps_per_second': 0.097, 'eval_num_tokens': 467371.0, 'eval_completions/mean_length': 171.3846153846154, 'eval_completions/min_length': 36.30769230769231, 'eval_completions/max_length': 413.0, 'eval_completions/clipped_ratio': 0.10576923076923077, 'eval_completions/mean_terminated_length': 130.67994689941406, 'eval_completions/min_terminated_length': 36.30769230769231, 'eval_completions/max_terminated_length': 267.46153846153845, 'eval_rewards/meter/mean': 0.4022582242122063, 'eval_rewards/meter/std': 0.32037050334306866, 'eval_rewards/count_adherence/mean': 0.8858297421382024, 'eval_rewards/count_adherence/std': 0.15729298671850792, 'eval_rewards/arabic_clean/mean': 0.7884615384615384, 'eval_rewards/arabic_clean/std': 0.3460688109581287, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9956730741720933, 'eval_rewards/arabic_floor/std': 0.012238385012516608, 'eval_rewards/count_floor/mean': 0.9657489107205317, 'eval_rewards/count_floor/std': 0.04718789687523475, 'eval_rewards/lexical_plausibility/mean': 0.9579436320524949, 'eval_rewards/lexical_plausibility/std': 0.07819282206205222, 'eval_rewards/judge_quality/mean': 0.3529807673050807, 'eval_rewards/judge_quality/std': 0.19971058173821524, 'eval_rewards/repeat_penalty/mean': 0.8649896199886615, 'eval_rewards/repeat_penalty/std': 0.1934896529867099, 'eval_rewards/repeat_floor/mean': 0.9865787029266357, 'eval_rewards/repeat_floor/std': 0.019076756524065368, 'eval_rewards/near_duplicate_penalty/mean': 0.972652270243718, 'eval_rewards/near_duplicate_penalty/std': 0.0355529493580644, 'eval_rewards/opening_diversity/mean': 0.9830288474376385, 'eval_rewards/opening_diversity/std': 0.03266540993578159, 'eval_rewards/distinct_2/mean': 0.9077681211324838, 'eval_rewards/distinct_2/std': 0.156504979213843, 'eval_rewards/total_composite/mean': 0.14088497224908608, 'eval_rewards/total_composite/std': 0.14079813802471527, 'eval_reward': 0.14088497224908608, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1258189225426087, 'eval_sampling/sampling_logp_difference/max': 1.347562496478741, 'eval_sampling/importance_sampling_ratio/min': 0.26669255242897916, 'eval_sampling/importance_sampling_ratio/mean': 1.039134511580834, 'eval_sampling/importance_sampling_ratio/max': 1.661945865704463, 'eval_entropy': 1.8623096667803252, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.14088497224908608, 'eval_reward_meter_mean': 0.4022582242122063, 'eval_reward_meter_std': 0.32037050334306866, 'eval_reward_count_adherence_mean': 0.8858297421382024, 'eval_reward_count_adherence_std': 0.15729298671850792, 'eval_reward_arabic_clean_mean': 0.7884615384615384, 'eval_reward_arabic_clean_std': 0.3460688109581287, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9956730741720933, 'eval_reward_arabic_floor_std': 0.012238385012516608, 'eval_reward_count_floor_mean': 0.9657489107205317, 'eval_reward_count_floor_std': 0.04718789687523475, 'eval_reward_lexical_plausibility_mean': 0.9579436320524949, 'eval_reward_lexical_plausibility_std': 0.07819282206205222, 'eval_reward_repeat_penalty_mean': 0.8649896199886615, 'eval_reward_repeat_penalty_std': 0.1934896529867099, 'eval_reward_repeat_floor_mean': 0.9865787029266357, 'eval_reward_repeat_floor_std': 0.019076756524065368, 'eval_reward_near_duplicate_penalty_mean': 0.972652270243718, 'eval_reward_near_duplicate_penalty_std': 0.0355529493580644, 'eval_reward_opening_diversity_mean': 0.9830288474376385, 'eval_reward_opening_diversity_std': 0.03266540993578159, 'eval_reward_distinct_2_mean': 0.9077681211324838, 'eval_reward_distinct_2_std': 0.156504979213843, 'eval_reward_judge_quality_mean': 0.3529807673050807, 'eval_reward_judge_quality_std': 0.19971058173821524, 'eval_reward_total_composite_mean': 0.14088497224908608, 'eval_reward_total_composite_std': 0.14079813802471527, 'epoch': 0.01}

  8%|▊         | 250/3300 [55:23<8:46:06, 10.35s/it]

100%|██████████| 13/13 [02:04<00:00,  9.89s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  8%|▊         | 251/3300 [55:36<43:21:10, 51.19s/it]2026-04-12 16:54:02,603 | INFO | train_grpo_train | metrics_logged mode=train step=251

                                                     
{'loss': 0.0141, 'grad_norm': 10.03214168548584, 'learning_rate': 9.242424242424244e-06, 'num_tokens': 469796.0, 'completions/mean_length': 99.125, 'completions/min_length': 75.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.125, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.781234860420227, 'rewards/meter/std': 0.3135354518890381, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.7317330241203308, 'rewards/repeat_penalty/std': 0.3799841105937958, 'rewards/repeat_floor/mean': 0.9612491130828857, 'rewards/repeat_floor/std': 0.06683868169784546, 'rewards/near_duplicate_penalty/mean': 0.9002248048782349, 'rewards/near_duplicate_penalty/std': 0.1821615844964981, 'rewards/opening_diversity/mean': 0.9453125, 'rewards/opening_diversity/std': 0.15467961132526398, 'rewards/distinct_2/mean': 0.7721851468086243, 'rewards/distinct_2/std': 0.3513650596141815, 'rewards/total_composite/mean': 0.24391454458236694, 'rewards/total_composite/std': 0.1486084908246994, 'reward': 0.24391454458236694, 'reward_std': 0.1486085057258606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18580110371112823, 'sampling/sampling_logp_difference/max': 2.362151622772217, 'sampling/importance_sampling_ratio/min': 0.09421728551387787, 'sampling/importance_sampling_ratio/mean': 1.0311291217803955, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4244952648878098, 'clip_ratio/low_mean': 0.08838836941868067, 'clip_ratio/low_min': 0.08838836941868067, 'clip_ratio/high_mean': 0.08680574130266905, 'clip_ratio/high_max': 0.08680574130266905, 'clip_ratio/region_mean': 0.17519411072134972, 'reward_total_mean': 0.24391454458236694, 'reward_meter_mean': 0.781234860420227, 'reward_meter_std': 0.3135354518890381, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7317330241203308, 'reward_repeat_penalty_std': 0.3799841105937958, 'reward_repeat_floor_mean': 0.9612491130828857, 'reward_repeat_floor_std': 0.06683868169784546, 'reward_near_duplicate_penalty_mean': 0.9002248048782349, 'reward_near_duplicate_penalty_std': 0.1821615844964981, 'reward_opening_diversity_mean': 0.9453125, 'reward_opening_diversity_std': 0.15467961132526398, 'reward_distinct_2_mean': 0.7721851468086243, 'reward_distinct_2_std': 0.3513650596141815, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.24391454458236694, 'reward_total_composite_std': 0.1486084908246994, 'epoch': 0.01}

  8%|▊         | 251/3300 [55:36<43:21:10, 51.19s/it]INFO 04-12 16:54:02 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 252/3300 [55:46<32:53:11, 38.84s/it]2026-04-12 16:54:12,641 | INFO | train_grpo_train | metrics_logged mode=train step=252

                                                     
{'loss': -0.014, 'grad_norm': 22.089492797851562, 'learning_rate': 9.23939393939394e-06, 'num_tokens': 471283.0, 'completions/mean_length': 29.875, 'completions/min_length': 25.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.8945860862731934, 'rewards/meter/std': 0.20781312882900238, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8712499737739563, 'rewards/judge_quality/std': 0.21087828278541565, 'rewards/repeat_penalty/mean': 0.982690155506134, 'rewards/repeat_penalty/std': 0.013326878659427166, 'rewards/repeat_floor/mean': 0.9974035024642944, 'rewards/repeat_floor/std': 0.0019990308210253716, 'rewards/near_duplicate_penalty/mean': 0.982690155506134, 'rewards/near_duplicate_penalty/std': 0.013326878659427166, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7784930467605591, 'rewards/total_composite/std': 0.27445438504219055, 'reward': 0.7784930467605591, 'reward_std': 0.27445438504219055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12639819085597992, 'sampling/sampling_logp_difference/max': 2.0030102729797363, 'sampling/importance_sampling_ratio/min': 0.13492849469184875, 'sampling/importance_sampling_ratio/mean': 1.0100497007369995, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45426201447844505, 'clip_ratio/low_mean': 0.05419354699552059, 'clip_ratio/low_min': 0.05419354699552059, 'clip_ratio/high_mean': 0.03564588166773319, 'clip_ratio/high_max': 0.03564588166773319, 'clip_ratio/region_mean': 0.08983942866325378, 'reward_total_mean': 0.7784930467605591, 'reward_meter_mean': 0.8945860862731934, 'reward_meter_std': 0.20781312882900238, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.982690155506134, 'reward_repeat_penalty_std': 0.013326878659427166, 'reward_repeat_floor_mean': 0.9974035024642944, 'reward_repeat_floor_std': 0.0019990308210253716, 'reward_near_duplicate_penalty_mean': 0.982690155506134, 'reward_near_duplicate_penalty_std': 0.013326878659427166, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8712499737739563, 'reward_judge_quality_std': 0.21087828278541565, 'reward_total_composite_mean': 0.7784930467605591, 'reward_total_composite_std': 0.27445438504219055, 'epoch': 0.01}

  8%|▊         | 252/3300 [55:46<32:53:11, 38.84s/it]INFO 04-12 16:54:12 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 253/3300 [55:59<26:21:23, 31.14s/it]2026-04-12 16:54:25,808 | INFO | train_grpo_train | metrics_logged mode=train step=253

                                                     
{'loss': -0.0213, 'grad_norm': 6.694579124450684, 'learning_rate': 9.236363636363636e-06, 'num_tokens': 472702.0, 'completions/mean_length': 96.375, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.40104883909225464, 'rewards/meter/std': 0.4383257031440735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9689521193504333, 'rewards/lexical_plausibility/std': 0.08781670778989792, 'rewards/judge_quality/mean': 0.5387500524520874, 'rewards/judge_quality/std': 0.3484019637107849, 'rewards/repeat_penalty/mean': 0.9807029962539673, 'rewards/repeat_penalty/std': 0.024172764271497726, 'rewards/repeat_floor/mean': 0.9971054196357727, 'rewards/repeat_floor/std': 0.0036259046755731106, 'rewards/near_duplicate_penalty/mean': 0.9807029962539673, 'rewards/near_duplicate_penalty/std': 0.024172764271497726, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2465665340423584, 'rewards/total_composite/std': 0.2989174425601959, 'reward': 0.2465665340423584, 'reward_std': 0.2989174425601959, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20465432107448578, 'sampling/sampling_logp_difference/max': 1.364037036895752, 'sampling/importance_sampling_ratio/min': 0.25562670826911926, 'sampling/importance_sampling_ratio/mean': 1.0323041677474976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.524191364645958, 'clip_ratio/low_mean': 0.045112782157957554, 'clip_ratio/low_min': 0.045112782157957554, 'clip_ratio/high_mean': 0.10233794245868921, 'clip_ratio/high_max': 0.10233794245868921, 'clip_ratio/region_mean': 0.14745072461664677, 'reward_total_mean': 0.2465665340423584, 'reward_meter_mean': 0.40104883909225464, 'reward_meter_std': 0.4383257031440735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9689521193504333, 'reward_lexical_plausibility_std': 0.08781670778989792, 'reward_repeat_penalty_mean': 0.9807029962539673, 'reward_repeat_penalty_std': 0.024172764271497726, 'reward_repeat_floor_mean': 0.9971054196357727, 'reward_repeat_floor_std': 0.0036259046755731106, 'reward_near_duplicate_penalty_mean': 0.9807029962539673, 'reward_near_duplicate_penalty_std': 0.024172764271497726, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5387500524520874, 'reward_judge_quality_std': 0.3484019637107849, 'reward_total_composite_mean': 0.2465665340423584, 'reward_total_composite_std': 0.2989174425601959, 'epoch': 0.01}

  8%|▊         | 253/3300 [55:59<26:21:23, 31.14s/it]INFO 04-12 16:54:26 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 254/3300 [56:07<20:25:31, 24.14s/it]2026-04-12 16:54:33,618 | INFO | train_grpo_train | metrics_logged mode=train step=254

                                                     
{'loss': 0.1065, 'grad_norm': 26.027971267700195, 'learning_rate': 9.233333333333334e-06, 'num_tokens': 474076.0, 'completions/mean_length': 17.75, 'completions/min_length': 15.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.86481773853302, 'rewards/meter/std': 0.3432632088661194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6974999904632568, 'rewards/judge_quality/std': 0.3259163200855255, 'rewards/repeat_penalty/mean': 0.7214862108230591, 'rewards/repeat_penalty/std': 0.05688679963350296, 'rewards/repeat_floor/mean': 0.9792972803115845, 'rewards/repeat_floor/std': 0.011377356946468353, 'rewards/near_duplicate_penalty/mean': 0.9619816541671753, 'rewards/near_duplicate_penalty/std': 0.07584905624389648, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6339339017868042, 'rewards/total_composite/std': 0.38137876987457275, 'reward': 0.6339339017868042, 'reward_std': 0.38137876987457275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17500117421150208, 'sampling/sampling_logp_difference/max': 1.4457502365112305, 'sampling/importance_sampling_ratio/min': 0.2355692833662033, 'sampling/importance_sampling_ratio/mean': 0.9843676686286926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1645722165703773, 'clip_ratio/low_mean': 0.05166666768491268, 'clip_ratio/low_min': 0.05166666768491268, 'clip_ratio/high_mean': 0.09493517922237515, 'clip_ratio/high_max': 0.09493517922237515, 'clip_ratio/region_mean': 0.14660184690728784, 'reward_total_mean': 0.6339339017868042, 'reward_meter_mean': 0.86481773853302, 'reward_meter_std': 0.3432632088661194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7214862108230591, 'reward_repeat_penalty_std': 0.05688679963350296, 'reward_repeat_floor_mean': 0.9792972803115845, 'reward_repeat_floor_std': 0.011377356946468353, 'reward_near_duplicate_penalty_mean': 0.9619816541671753, 'reward_near_duplicate_penalty_std': 0.07584905624389648, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6974999904632568, 'reward_judge_quality_std': 0.3259163200855255, 'reward_total_composite_mean': 0.6339339017868042, 'reward_total_composite_std': 0.38137876987457275, 'epoch': 0.01}

  8%|▊         | 254/3300 [56:07<20:25:31, 24.14s/it]INFO 04-12 16:54:33 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 255/3300 [56:14<16:11:06, 19.14s/it]2026-04-12 16:54:41,071 | INFO | train_grpo_train | metrics_logged mode=train step=255

                                                     
{'loss': 0.1081, 'grad_norm': 22.420490264892578, 'learning_rate': 9.23030303030303e-06, 'num_tokens': 475483.0, 'completions/mean_length': 19.875, 'completions/min_length': 16.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5202906131744385, 'rewards/meter/std': 0.3947794735431671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7799999713897705, 'rewards/judge_quality/std': 0.3071993589401245, 'rewards/repeat_penalty/mean': 0.7077559232711792, 'rewards/repeat_penalty/std': 0.062000926584005356, 'rewards/repeat_floor/mean': 0.9765511751174927, 'rewards/repeat_floor/std': 0.012400192208588123, 'rewards/near_duplicate_penalty/mean': 0.9436745643615723, 'rewards/near_duplicate_penalty/std': 0.08266790211200714, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3250795900821686, 'rewards/total_composite/std': 0.2747003734111786, 'reward': 0.3250795900821686, 'reward_std': 0.2747003734111786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14918068051338196, 'sampling/sampling_logp_difference/max': 1.1693806648254395, 'sampling/importance_sampling_ratio/min': 0.3105592131614685, 'sampling/importance_sampling_ratio/mean': 1.032403826713562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9908276125788689, 'clip_ratio/low_mean': 0.09877840150147676, 'clip_ratio/low_min': 0.09877840150147676, 'clip_ratio/high_mean': 0.02878289483487606, 'clip_ratio/high_max': 0.02878289483487606, 'clip_ratio/region_mean': 0.12756129633635283, 'reward_total_mean': 0.3250795900821686, 'reward_meter_mean': 0.5202906131744385, 'reward_meter_std': 0.3947794735431671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7077559232711792, 'reward_repeat_penalty_std': 0.062000926584005356, 'reward_repeat_floor_mean': 0.9765511751174927, 'reward_repeat_floor_std': 0.012400192208588123, 'reward_near_duplicate_penalty_mean': 0.9436745643615723, 'reward_near_duplicate_penalty_std': 0.08266790211200714, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7799999713897705, 'reward_judge_quality_std': 0.3071993589401245, 'reward_total_composite_mean': 0.3250795900821686, 'reward_total_composite_std': 0.2747003734111786, 'epoch': 0.01}

  8%|▊         | 255/3300 [56:14<16:11:06, 19.14s/it]INFO 04-12 16:54:41 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 256/3300 [56:28<14:47:36, 17.50s/it]2026-04-12 16:54:54,743 | INFO | train_grpo_train | metrics_logged mode=train step=256

                                                     
{'loss': -0.144, 'grad_norm': 3.8688511848449707, 'learning_rate': 9.227272727272728e-06, 'num_tokens': 477640.0, 'completions/mean_length': 223.625, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 127.5, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.9472206234931946, 'rewards/meter/std': 0.07846547663211823, 'rewards/count_adherence/mean': 0.7708333134651184, 'rewards/count_adherence/std': 0.19795581698417664, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9312499761581421, 'rewards/count_floor/std': 0.05938674136996269, 'rewards/lexical_plausibility/mean': 0.9372028112411499, 'rewards/lexical_plausibility/std': 0.1017853245139122, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.6489769816398621, 'rewards/repeat_penalty/std': 0.40324708819389343, 'rewards/repeat_floor/mean': 0.958721399307251, 'rewards/repeat_floor/std': 0.05086202174425125, 'rewards/near_duplicate_penalty/mean': 0.9152227640151978, 'rewards/near_duplicate_penalty/std': 0.11003494262695312, 'rewards/opening_diversity/mean': 0.9642857313156128, 'rewards/opening_diversity/std': 0.0874817818403244, 'rewards/distinct_2/mean': 0.7064533233642578, 'rewards/distinct_2/std': 0.34807756543159485, 'rewards/total_composite/mean': 0.16540731489658356, 'rewards/total_composite/std': 0.13221320509910583, 'reward': 0.16540731489658356, 'reward_std': 0.13221320509910583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16439630091190338, 'sampling/sampling_logp_difference/max': 1.6514892578125, 'sampling/importance_sampling_ratio/min': 0.19176411628723145, 'sampling/importance_sampling_ratio/mean': 1.0185129642486572, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.504008173942566, 'clip_ratio/low_mean': 0.07231230568140745, 'clip_ratio/low_min': 0.07231230568140745, 'clip_ratio/high_mean': 0.04384084418416023, 'clip_ratio/high_max': 0.04384084418416023, 'clip_ratio/region_mean': 0.11615314986556768, 'reward_total_mean': 0.16540731489658356, 'reward_meter_mean': 0.9472206234931946, 'reward_meter_std': 0.07846547663211823, 'reward_count_adherence_mean': 0.7708333134651184, 'reward_count_adherence_std': 0.19795581698417664, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9312499761581421, 'reward_count_floor_std': 0.05938674136996269, 'reward_lexical_plausibility_mean': 0.9372028112411499, 'reward_lexical_plausibility_std': 0.1017853245139122, 'reward_repeat_penalty_mean': 0.6489769816398621, 'reward_repeat_penalty_std': 0.40324708819389343, 'reward_repeat_floor_mean': 0.958721399307251, 'reward_repeat_floor_std': 0.05086202174425125, 'reward_near_duplicate_penalty_mean': 0.9152227640151978, 'reward_near_duplicate_penalty_std': 0.11003494262695312, 'reward_opening_diversity_mean': 0.9642857313156128, 'reward_opening_diversity_std': 0.0874817818403244, 'reward_distinct_2_mean': 0.7064533233642578, 'reward_distinct_2_std': 0.34807756543159485, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.16540731489658356, 'reward_total_composite_std': 0.13221320509910583, 'epoch': 0.01}

  8%|▊         | 256/3300 [56:28<14:47:36, 17.50s/it]INFO 04-12 16:54:54 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 257/3300 [56:36<12:21:36, 14.62s/it]2026-04-12 16:55:02,663 | INFO | train_grpo_train | metrics_logged mode=train step=257

                                                     
{'loss': -0.0232, 'grad_norm': 18.625070571899414, 'learning_rate': 9.224242424242424e-06, 'num_tokens': 479142.0, 'completions/mean_length': 33.75, 'completions/min_length': 27.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7656682133674622, 'rewards/meter/std': 0.3254414200782776, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9828088283538818, 'rewards/repeat_penalty/std': 0.03942342847585678, 'rewards/repeat_floor/mean': 0.9974213242530823, 'rewards/repeat_floor/std': 0.0059135062620043755, 'rewards/near_duplicate_penalty/mean': 0.9828088283538818, 'rewards/near_duplicate_penalty/std': 0.03942342847585678, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4802744388580322, 'rewards/total_composite/std': 0.27702468633651733, 'reward': 0.4802744388580322, 'reward_std': 0.27702465653419495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21323837339878082, 'sampling/sampling_logp_difference/max': 1.6583452224731445, 'sampling/importance_sampling_ratio/min': 0.1904538869857788, 'sampling/importance_sampling_ratio/mean': 1.0013370513916016, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4175041690468788, 'clip_ratio/low_mean': 0.11052919365465641, 'clip_ratio/low_min': 0.11052919365465641, 'clip_ratio/high_mean': 0.04904306307435036, 'clip_ratio/high_max': 0.04904306307435036, 'clip_ratio/region_mean': 0.15957225672900677, 'reward_total_mean': 0.4802744388580322, 'reward_meter_mean': 0.7656682133674622, 'reward_meter_std': 0.3254414200782776, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9828088283538818, 'reward_repeat_penalty_std': 0.03942342847585678, 'reward_repeat_floor_mean': 0.9974213242530823, 'reward_repeat_floor_std': 0.0059135062620043755, 'reward_near_duplicate_penalty_mean': 0.9828088283538818, 'reward_near_duplicate_penalty_std': 0.03942342847585678, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.4802744388580322, 'reward_total_composite_std': 0.27702468633651733, 'epoch': 0.01}

  8%|▊         | 257/3300 [56:36<12:21:36, 14.62s/it]INFO 04-12 16:55:02 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 258/3300 [56:45<11:00:26, 13.03s/it]2026-04-12 16:55:11,965 | INFO | train_grpo_train | metrics_logged mode=train step=258

                                                     
{'loss': 0.0868, 'grad_norm': 18.444421768188477, 'learning_rate': 9.221212121212123e-06, 'num_tokens': 481072.0, 'completions/mean_length': 74.25, 'completions/min_length': 61.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.6494839787483215, 'rewards/meter/std': 0.32895997166633606, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5249999761581421, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9731181859970093, 'rewards/repeat_penalty/std': 0.038941849023103714, 'rewards/repeat_floor/mean': 0.9965662360191345, 'rewards/repeat_floor/std': 0.004189968574792147, 'rewards/near_duplicate_penalty/mean': 0.9837016463279724, 'rewards/near_duplicate_penalty/std': 0.01104305312037468, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9890109896659851, 'rewards/distinct_2/std': 0.031081615015864372, 'rewards/total_composite/mean': 0.3533145785331726, 'rewards/total_composite/std': 0.2352585643529892, 'reward': 0.3533145785331726, 'reward_std': 0.235258549451828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21634989976882935, 'sampling/sampling_logp_difference/max': 3.4774093627929688, 'sampling/importance_sampling_ratio/min': 0.03088732436299324, 'sampling/importance_sampling_ratio/mean': 0.9958537817001343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4792455434799194, 'clip_ratio/low_mean': 0.08586963266134262, 'clip_ratio/low_min': 0.08586963266134262, 'clip_ratio/high_mean': 0.09246682561933994, 'clip_ratio/high_max': 0.09246682561933994, 'clip_ratio/region_mean': 0.17833645828068256, 'reward_total_mean': 0.3533145785331726, 'reward_meter_mean': 0.6494839787483215, 'reward_meter_std': 0.32895997166633606, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9731181859970093, 'reward_repeat_penalty_std': 0.038941849023103714, 'reward_repeat_floor_mean': 0.9965662360191345, 'reward_repeat_floor_std': 0.004189968574792147, 'reward_near_duplicate_penalty_mean': 0.9837016463279724, 'reward_near_duplicate_penalty_std': 0.01104305312037468, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9890109896659851, 'reward_distinct_2_std': 0.031081615015864372, 'reward_judge_quality_mean': 0.5249999761581421, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.3533145785331726, 'reward_total_composite_std': 0.2352585643529892, 'epoch': 0.01}

  8%|▊         | 258/3300 [56:45<11:00:26, 13.03s/it]INFO 04-12 16:55:12 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 259/3300 [56:54<9:54:24, 11.73s/it] 2026-04-12 16:55:20,662 | INFO | train_grpo_train | metrics_logged mode=train step=259

                                                    
{'loss': 0.0318, 'grad_norm': 13.94146728515625, 'learning_rate': 9.21818181818182e-06, 'num_tokens': 482844.0, 'completions/mean_length': 46.5, 'completions/min_length': 37.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7833342552185059, 'rewards/meter/std': 0.3319782018661499, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.3142678737640381, 'rewards/repeat_penalty/mean': 0.9918050169944763, 'rewards/repeat_penalty/std': 0.014694659039378166, 'rewards/repeat_floor/mean': 0.9987707138061523, 'rewards/repeat_floor/std': 0.002204208169132471, 'rewards/near_duplicate_penalty/mean': 0.9918050169944763, 'rewards/near_duplicate_penalty/std': 0.014694659039378166, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4246077835559845, 'rewards/total_composite/std': 0.32079240679740906, 'reward': 0.4246077835559845, 'reward_std': 0.32079240679740906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18138344585895538, 'sampling/sampling_logp_difference/max': 2.3204782009124756, 'sampling/importance_sampling_ratio/min': 0.09822659939527512, 'sampling/importance_sampling_ratio/mean': 1.020614504814148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3293571323156357, 'clip_ratio/low_mean': 0.11445644684135914, 'clip_ratio/low_min': 0.11445644684135914, 'clip_ratio/high_mean': 0.08980211615562439, 'clip_ratio/high_max': 0.08980211615562439, 'clip_ratio/region_mean': 0.20425856299698353, 'reward_total_mean': 0.4246077835559845, 'reward_meter_mean': 0.7833342552185059, 'reward_meter_std': 0.3319782018661499, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9918050169944763, 'reward_repeat_penalty_std': 0.014694659039378166, 'reward_repeat_floor_mean': 0.9987707138061523, 'reward_repeat_floor_std': 0.002204208169132471, 'reward_near_duplicate_penalty_mean': 0.9918050169944763, 'reward_near_duplicate_penalty_std': 0.014694659039378166, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.3142678737640381, 'reward_total_composite_mean': 0.4246077835559845, 'reward_total_composite_std': 0.32079240679740906, 'epoch': 0.01}

  8%|▊         | 259/3300 [56:54<9:54:24, 11.73s/it]INFO 04-12 16:55:20 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 260/3300 [57:02<9:02:44, 10.71s/it]2026-04-12 16:55:29,004 | INFO | train_grpo_train | metrics_logged mode=train step=260

                                                    
{'loss': 0.1219, 'grad_norm': 21.54400634765625, 'learning_rate': 9.215151515151515e-06, 'num_tokens': 484488.0, 'completions/mean_length': 38.5, 'completions/min_length': 30.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.35186442732810974, 'rewards/meter/std': 0.2855783700942993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5249999761581421, 'rewards/judge_quality/std': 0.13887302577495575, 'rewards/repeat_penalty/mean': 0.963065505027771, 'rewards/repeat_penalty/std': 0.04576059803366661, 'rewards/repeat_floor/mean': 0.9949644804000854, 'rewards/repeat_floor/std': 0.005850640591233969, 'rewards/near_duplicate_penalty/mean': 0.9721989631652832, 'rewards/near_duplicate_penalty/std': 0.030996691435575485, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.15843889117240906, 'rewards/total_composite/std': 0.11046849191188812, 'reward': 0.15843889117240906, 'reward_std': 0.11046849191188812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18502263724803925, 'sampling/sampling_logp_difference/max': 1.6262269020080566, 'sampling/importance_sampling_ratio/min': 0.19667023420333862, 'sampling/importance_sampling_ratio/mean': 1.03053879737854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3729137480258942, 'clip_ratio/low_mean': 0.0936553031206131, 'clip_ratio/low_min': 0.0936553031206131, 'clip_ratio/high_mean': 0.0783494608476758, 'clip_ratio/high_max': 0.0783494608476758, 'clip_ratio/region_mean': 0.1720047639682889, 'reward_total_mean': 0.15843889117240906, 'reward_meter_mean': 0.35186442732810974, 'reward_meter_std': 0.2855783700942993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.963065505027771, 'reward_repeat_penalty_std': 0.04576059803366661, 'reward_repeat_floor_mean': 0.9949644804000854, 'reward_repeat_floor_std': 0.005850640591233969, 'reward_near_duplicate_penalty_mean': 0.9721989631652832, 'reward_near_duplicate_penalty_std': 0.030996691435575485, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.5249999761581421, 'reward_judge_quality_std': 0.13887302577495575, 'reward_total_composite_mean': 0.15843889117240906, 'reward_total_composite_std': 0.11046849191188812, 'epoch': 0.01}

  8%|▊         | 260/3300 [57:02<9:02:44, 10.71s/it]INFO 04-12 16:55:29 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 261/3300 [57:10<8:10:08,  9.68s/it]2026-04-12 16:55:36,266 | INFO | train_grpo_train | metrics_logged mode=train step=261

                                                    
{'loss': 0.0201, 'grad_norm': 25.051698684692383, 'learning_rate': 9.212121212121213e-06, 'num_tokens': 485824.0, 'completions/mean_length': 17.0, 'completions/min_length': 14.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.826312780380249, 'rewards/meter/std': 0.24248762428760529, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.793749988079071, 'rewards/judge_quality/std': 0.2447119504213333, 'rewards/repeat_penalty/mean': 0.7389006614685059, 'rewards/repeat_penalty/std': 0.024166369810700417, 'rewards/repeat_floor/mean': 0.9827800989151001, 'rewards/repeat_floor/std': 0.004833268467336893, 'rewards/near_duplicate_penalty/mean': 0.9852008819580078, 'rewards/near_duplicate_penalty/std': 0.03222181275486946, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6345161199569702, 'rewards/total_composite/std': 0.27586397528648376, 'reward': 0.6345161199569702, 'reward_std': 0.27586397528648376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19997957348823547, 'sampling/sampling_logp_difference/max': 1.5830087661743164, 'sampling/importance_sampling_ratio/min': 0.20535629987716675, 'sampling/importance_sampling_ratio/mean': 1.0394620895385742, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.051232859492302, 'clip_ratio/low_mean': 0.1126575656235218, 'clip_ratio/low_min': 0.1126575656235218, 'clip_ratio/high_mean': 0.09171235375106335, 'clip_ratio/high_max': 0.09171235375106335, 'clip_ratio/region_mean': 0.20436991937458515, 'reward_total_mean': 0.6345161199569702, 'reward_meter_mean': 0.826312780380249, 'reward_meter_std': 0.24248762428760529, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7389006614685059, 'reward_repeat_penalty_std': 0.024166369810700417, 'reward_repeat_floor_mean': 0.9827800989151001, 'reward_repeat_floor_std': 0.004833268467336893, 'reward_near_duplicate_penalty_mean': 0.9852008819580078, 'reward_near_duplicate_penalty_std': 0.03222181275486946, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.793749988079071, 'reward_judge_quality_std': 0.2447119504213333, 'reward_total_composite_mean': 0.6345161199569702, 'reward_total_composite_std': 0.27586397528648376, 'epoch': 0.01}

  8%|▊         | 261/3300 [57:10<8:10:08,  9.68s/it]INFO 04-12 16:55:36 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 262/3300 [57:20<8:23:34,  9.95s/it]2026-04-12 16:55:46,838 | INFO | train_grpo_train | metrics_logged mode=train step=262

                                                    
{'loss': 0.1006, 'grad_norm': 8.600594520568848, 'learning_rate': 9.20909090909091e-06, 'num_tokens': 488790.0, 'completions/mean_length': 160.75, 'completions/min_length': 135.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.75, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.6169558167457581, 'rewards/meter/std': 0.329196959733963, 'rewards/count_adherence/mean': 0.9027777910232544, 'rewards/count_adherence/std': 0.07120776921510696, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9708333611488342, 'rewards/count_floor/std': 0.021362334489822388, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.7664869427680969, 'rewards/repeat_penalty/std': 0.30370059609413147, 'rewards/repeat_floor/mean': 0.9746162295341492, 'rewards/repeat_floor/std': 0.03300850838422775, 'rewards/near_duplicate_penalty/mean': 0.9478528499603271, 'rewards/near_duplicate_penalty/std': 0.05683739483356476, 'rewards/opening_diversity/mean': 0.996874988079071, 'rewards/opening_diversity/std': 0.008838826790452003, 'rewards/distinct_2/mean': 0.8069530129432678, 'rewards/distinct_2/std': 0.2731459438800812, 'rewards/total_composite/mean': 0.2254660427570343, 'rewards/total_composite/std': 0.13885778188705444, 'reward': 0.2254660427570343, 'reward_std': 0.13885776698589325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1995864361524582, 'sampling/sampling_logp_difference/max': 2.6709909439086914, 'sampling/importance_sampling_ratio/min': 0.06918363273143768, 'sampling/importance_sampling_ratio/mean': 1.0171693563461304, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5021056979894638, 'clip_ratio/low_mean': 0.07800502516329288, 'clip_ratio/low_min': 0.07800502516329288, 'clip_ratio/high_mean': 0.11030659079551697, 'clip_ratio/high_max': 0.11030659079551697, 'clip_ratio/region_mean': 0.18831161595880985, 'reward_total_mean': 0.2254660427570343, 'reward_meter_mean': 0.6169558167457581, 'reward_meter_std': 0.329196959733963, 'reward_count_adherence_mean': 0.9027777910232544, 'reward_count_adherence_std': 0.07120776921510696, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9708333611488342, 'reward_count_floor_std': 0.021362334489822388, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7664869427680969, 'reward_repeat_penalty_std': 0.30370059609413147, 'reward_repeat_floor_mean': 0.9746162295341492, 'reward_repeat_floor_std': 0.03300850838422775, 'reward_near_duplicate_penalty_mean': 0.9478528499603271, 'reward_near_duplicate_penalty_std': 0.05683739483356476, 'reward_opening_diversity_mean': 0.996874988079071, 'reward_opening_diversity_std': 0.008838826790452003, 'reward_distinct_2_mean': 0.8069530129432678, 'reward_distinct_2_std': 0.2731459438800812, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.2254660427570343, 'reward_total_composite_std': 0.13885778188705444, 'epoch': 0.01}

  8%|▊         | 262/3300 [57:20<8:23:34,  9.95s/it]INFO 04-12 16:55:47 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 263/3300 [57:28<7:52:14,  9.33s/it]2026-04-12 16:55:54,733 | INFO | train_grpo_train | metrics_logged mode=train step=263

                                                    
{'loss': -0.0171, 'grad_norm': 16.67426300048828, 'learning_rate': 9.206060606060607e-06, 'num_tokens': 490132.0, 'completions/mean_length': 16.75, 'completions/min_length': 15.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.3862903118133545, 'rewards/meter/std': 0.5008962154388428, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9975490570068359, 'rewards/lexical_plausibility/std': 0.006932419259101152, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.3358757197856903, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.09208157658576965, 'rewards/total_composite/std': 0.11924655735492706, 'reward': 0.09208157658576965, 'reward_std': 0.11924654990434647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15335243940353394, 'sampling/sampling_logp_difference/max': 1.4169368743896484, 'sampling/importance_sampling_ratio/min': 0.24245555698871613, 'sampling/importance_sampling_ratio/mean': 1.0290390253067017, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0767432823777199, 'clip_ratio/low_mean': 0.08767361240461469, 'clip_ratio/low_min': 0.08767361240461469, 'clip_ratio/high_mean': 0.08931028470396996, 'clip_ratio/high_max': 0.08931028470396996, 'clip_ratio/region_mean': 0.17698389710858464, 'reward_total_mean': 0.09208157658576965, 'reward_meter_mean': 0.3862903118133545, 'reward_meter_std': 0.5008962154388428, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9975490570068359, 'reward_lexical_plausibility_std': 0.006932419259101152, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.3358757197856903, 'reward_total_composite_mean': 0.09208157658576965, 'reward_total_composite_std': 0.11924655735492706, 'epoch': 0.01}

  8%|▊         | 263/3300 [57:28<7:52:14,  9.33s/it]INFO 04-12 16:55:54 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 264/3300 [57:37<7:51:33,  9.32s/it]2026-04-12 16:56:04,028 | INFO | train_grpo_train | metrics_logged mode=train step=264

                                                    
{'loss': 0.0375, 'grad_norm': 18.884525299072266, 'learning_rate': 9.203030303030304e-06, 'num_tokens': 491856.0, 'completions/mean_length': 49.5, 'completions/min_length': 26.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7616921663284302, 'rewards/meter/std': 0.3500820994377136, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9848946332931519, 'rewards/repeat_penalty/std': 0.03017011098563671, 'rewards/repeat_floor/mean': 0.9980955123901367, 'rewards/repeat_floor/std': 0.003537016222253442, 'rewards/near_duplicate_penalty/mean': 0.991282045841217, 'rewards/near_duplicate_penalty/std': 0.013199453242123127, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.3263661861419678, 'rewards/total_composite/std': 0.17969068884849548, 'reward': 0.3263661861419678, 'reward_std': 0.17969068884849548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16969145834445953, 'sampling/sampling_logp_difference/max': 1.3060121536254883, 'sampling/importance_sampling_ratio/min': 0.2708982229232788, 'sampling/importance_sampling_ratio/mean': 1.0045783519744873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.137873463332653, 'clip_ratio/low_mean': 0.06228146981447935, 'clip_ratio/low_min': 0.06228146981447935, 'clip_ratio/high_mean': 0.08988671563565731, 'clip_ratio/high_max': 0.08988671563565731, 'clip_ratio/region_mean': 0.15216818545013666, 'reward_total_mean': 0.3263661861419678, 'reward_meter_mean': 0.7616921663284302, 'reward_meter_std': 0.3500820994377136, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9848946332931519, 'reward_repeat_penalty_std': 0.03017011098563671, 'reward_repeat_floor_mean': 0.9980955123901367, 'reward_repeat_floor_std': 0.003537016222253442, 'reward_near_duplicate_penalty_mean': 0.991282045841217, 'reward_near_duplicate_penalty_std': 0.013199453242123127, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.3263661861419678, 'reward_total_composite_std': 0.17969068884849548, 'epoch': 0.01}

  8%|▊         | 264/3300 [57:37<7:51:33,  9.32s/it]INFO 04-12 16:56:04 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 265/3300 [57:51<8:52:48, 10.53s/it]2026-04-12 16:56:17,394 | INFO | train_grpo_train | metrics_logged mode=train step=265

                                                    
{'loss': -0.0388, 'grad_norm': 6.325990200042725, 'learning_rate': 9.200000000000002e-06, 'num_tokens': 493720.0, 'completions/mean_length': 119.0, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 62.857147216796875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.6770095825195312, 'rewards/meter/std': 0.42658403515815735, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 0.976481556892395, 'rewards/lexical_plausibility/std': 0.06652020663022995, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.1642080545425415, 'rewards/repeat_penalty/mean': 0.8401646614074707, 'rewards/repeat_penalty/std': 0.3135274052619934, 'rewards/repeat_floor/mean': 0.9777703881263733, 'rewards/repeat_floor/std': 0.049781881272792816, 'rewards/near_duplicate_penalty/mean': 0.9321873188018799, 'rewards/near_duplicate_penalty/std': 0.14401811361312866, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9076923131942749, 'rewards/distinct_2/std': 0.261085569858551, 'rewards/total_composite/mean': 0.2789556682109833, 'rewards/total_composite/std': 0.2078925371170044, 'reward': 0.2789556682109833, 'reward_std': 0.2078925222158432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1912977397441864, 'sampling/sampling_logp_difference/max': 1.1942138671875, 'sampling/importance_sampling_ratio/min': 0.3029420077800751, 'sampling/importance_sampling_ratio/mean': 1.061216950416565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5955148190259933, 'clip_ratio/low_mean': 0.03898264188319445, 'clip_ratio/low_min': 0.03898264188319445, 'clip_ratio/high_mean': 0.11994142830371857, 'clip_ratio/high_max': 0.11994142830371857, 'clip_ratio/region_mean': 0.158924070186913, 'reward_total_mean': 0.2789556682109833, 'reward_meter_mean': 0.6770095825195312, 'reward_meter_std': 0.42658403515815735, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 0.976481556892395, 'reward_lexical_plausibility_std': 0.06652020663022995, 'reward_repeat_penalty_mean': 0.8401646614074707, 'reward_repeat_penalty_std': 0.3135274052619934, 'reward_repeat_floor_mean': 0.9777703881263733, 'reward_repeat_floor_std': 0.049781881272792816, 'reward_near_duplicate_penalty_mean': 0.9321873188018799, 'reward_near_duplicate_penalty_std': 0.14401811361312866, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9076923131942749, 'reward_distinct_2_std': 0.261085569858551, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.1642080545425415, 'reward_total_composite_mean': 0.2789556682109833, 'reward_total_composite_std': 0.2078925371170044, 'epoch': 0.01}

  8%|▊         | 265/3300 [57:51<8:52:48, 10.53s/it]INFO 04-12 16:56:17 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 266/3300 [58:04<9:39:23, 11.46s/it]2026-04-12 16:56:31,008 | INFO | train_grpo_train | metrics_logged mode=train step=266

                                                    
{'loss': -0.0126, 'grad_norm': 4.99873161315918, 'learning_rate': 9.196969696969697e-06, 'num_tokens': 495166.0, 'completions/mean_length': 81.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.285715103149414, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.5839483737945557, 'rewards/meter/std': 0.45508986711502075, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9637843370437622, 'rewards/lexical_plausibility/std': 0.08021838963031769, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.32762131094932556, 'rewards/repeat_penalty/mean': 0.7245987057685852, 'rewards/repeat_penalty/std': 0.14216618239879608, 'rewards/repeat_floor/mean': 0.9759916067123413, 'rewards/repeat_floor/std': 0.01924424059689045, 'rewards/near_duplicate_penalty/mean': 0.9379336833953857, 'rewards/near_duplicate_penalty/std': 0.09077394753694534, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.3220638632774353, 'rewards/total_composite/std': 0.37391313910484314, 'reward': 0.3220638632774353, 'reward_std': 0.37391313910484314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22514304518699646, 'sampling/sampling_logp_difference/max': 1.2849922180175781, 'sampling/importance_sampling_ratio/min': 0.2766527533531189, 'sampling/importance_sampling_ratio/mean': 1.002685785293579, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8283866345882416, 'clip_ratio/low_mean': 0.10828283056616783, 'clip_ratio/low_min': 0.10828283056616783, 'clip_ratio/high_mean': 0.06835317704826593, 'clip_ratio/high_max': 0.06835317704826593, 'clip_ratio/region_mean': 0.17663600761443377, 'reward_total_mean': 0.3220638632774353, 'reward_meter_mean': 0.5839483737945557, 'reward_meter_std': 0.45508986711502075, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9637843370437622, 'reward_lexical_plausibility_std': 0.08021838963031769, 'reward_repeat_penalty_mean': 0.7245987057685852, 'reward_repeat_penalty_std': 0.14216618239879608, 'reward_repeat_floor_mean': 0.9759916067123413, 'reward_repeat_floor_std': 0.01924424059689045, 'reward_near_duplicate_penalty_mean': 0.9379336833953857, 'reward_near_duplicate_penalty_std': 0.09077394753694534, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.32762131094932556, 'reward_total_composite_mean': 0.3220638632774353, 'reward_total_composite_std': 0.37391313910484314, 'epoch': 0.01}

  8%|▊         | 266/3300 [58:04<9:39:23, 11.46s/it]INFO 04-12 16:56:31 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 267/3300 [58:12<8:43:50, 10.36s/it]2026-04-12 16:56:38,816 | INFO | train_grpo_train | metrics_logged mode=train step=267

                                                    
{'loss': -0.0225, 'grad_norm': 20.03293228149414, 'learning_rate': 9.193939393939395e-06, 'num_tokens': 496774.0, 'completions/mean_length': 38.0, 'completions/min_length': 30.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5126917362213135, 'rewards/meter/std': 0.48572054505348206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8600000143051147, 'rewards/judge_quality/std': 0.2066052109003067, 'rewards/repeat_penalty/mean': 0.9858925938606262, 'rewards/repeat_penalty/std': 0.024782583117485046, 'rewards/repeat_floor/mean': 0.9984083771705627, 'rewards/repeat_floor/std': 0.002445716643705964, 'rewards/near_duplicate_penalty/mean': 0.9946338534355164, 'rewards/near_duplicate_penalty/std': 0.01049359142780304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.4760437607765198, 'rewards/total_composite/std': 0.4552389979362488, 'reward': 0.4760437607765198, 'reward_std': 0.4552389979362488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19471172988414764, 'sampling/sampling_logp_difference/max': 1.7212409973144531, 'sampling/importance_sampling_ratio/min': 0.17884406447410583, 'sampling/importance_sampling_ratio/mean': 1.0050082206726074, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.06889970600605, 'clip_ratio/low_mean': 0.0628314409404993, 'clip_ratio/low_min': 0.0628314409404993, 'clip_ratio/high_mean': 0.11642867513000965, 'clip_ratio/high_max': 0.11642867513000965, 'clip_ratio/region_mean': 0.17926011607050896, 'reward_total_mean': 0.4760437607765198, 'reward_meter_mean': 0.5126917362213135, 'reward_meter_std': 0.48572054505348206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9858925938606262, 'reward_repeat_penalty_std': 0.024782583117485046, 'reward_repeat_floor_mean': 0.9984083771705627, 'reward_repeat_floor_std': 0.002445716643705964, 'reward_near_duplicate_penalty_mean': 0.9946338534355164, 'reward_near_duplicate_penalty_std': 0.01049359142780304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.8600000143051147, 'reward_judge_quality_std': 0.2066052109003067, 'reward_total_composite_mean': 0.4760437607765198, 'reward_total_composite_std': 0.4552389979362488, 'epoch': 0.01}

  8%|▊         | 267/3300 [58:12<8:43:50, 10.36s/it]INFO 04-12 16:56:39 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 268/3300 [58:21<8:27:14, 10.04s/it]2026-04-12 16:56:48,095 | INFO | train_grpo_train | metrics_logged mode=train step=268

                                                    
{'loss': 0.3745, 'grad_norm': 11.29027271270752, 'learning_rate': 9.190909090909092e-06, 'num_tokens': 498694.0, 'completions/mean_length': 65.0, 'completions/min_length': 18.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.48991847038269043, 'rewards/meter/std': 0.3096616268157959, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3471825420856476, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.10415478050708771, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.8983121514320374, 'rewards/repeat_penalty/std': 0.10661174356937408, 'rewards/repeat_floor/mean': 0.992123007774353, 'rewards/repeat_floor/std': 0.006823556497693062, 'rewards/near_duplicate_penalty/mean': 0.9883452653884888, 'rewards/near_duplicate_penalty/std': 0.012447234243154526, 'rewards/opening_diversity/mean': 0.9296875, 'rewards/opening_diversity/std': 0.1129826009273529, 'rewards/distinct_2/mean': 0.9787768721580505, 'rewards/distinct_2/std': 0.032197922468185425, 'rewards/total_composite/mean': 0.19729964435100555, 'rewards/total_composite/std': 0.20176763832569122, 'reward': 0.19729964435100555, 'reward_std': 0.20176765322685242, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1991879642009735, 'sampling/sampling_logp_difference/max': 1.4274108409881592, 'sampling/importance_sampling_ratio/min': 0.23992934823036194, 'sampling/importance_sampling_ratio/mean': 0.9962937831878662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4438332840800285, 'clip_ratio/low_mean': 0.1417461894452572, 'clip_ratio/low_min': 0.1417461894452572, 'clip_ratio/high_mean': 0.05208333395421505, 'clip_ratio/high_max': 0.05208333395421505, 'clip_ratio/region_mean': 0.19382952339947224, 'reward_total_mean': 0.19729964435100555, 'reward_meter_mean': 0.48991847038269043, 'reward_meter_std': 0.3096616268157959, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3471825420856476, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.10415478050708771, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8983121514320374, 'reward_repeat_penalty_std': 0.10661174356937408, 'reward_repeat_floor_mean': 0.992123007774353, 'reward_repeat_floor_std': 0.006823556497693062, 'reward_near_duplicate_penalty_mean': 0.9883452653884888, 'reward_near_duplicate_penalty_std': 0.012447234243154526, 'reward_opening_diversity_mean': 0.9296875, 'reward_opening_diversity_std': 0.1129826009273529, 'reward_distinct_2_mean': 0.9787768721580505, 'reward_distinct_2_std': 0.032197922468185425, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.19729964435100555, 'reward_total_composite_std': 0.20176763832569122, 'epoch': 0.01}

  8%|▊         | 268/3300 [58:21<8:27:14, 10.04s/it]INFO 04-12 16:56:48 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 269/3300 [58:36<9:31:11, 11.31s/it]2026-04-12 16:57:02,364 | INFO | train_grpo_train | metrics_logged mode=train step=269

                                                    
{'loss': -0.0673, 'grad_norm': 4.49676513671875, 'learning_rate': 9.187878787878789e-06, 'num_tokens': 501596.0, 'completions/mean_length': 210.75, 'completions/min_length': 135.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 167.71429443359375, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.7809507846832275, 'rewards/meter/std': 0.33067092299461365, 'rewards/count_adherence/mean': 0.8571429252624512, 'rewards/count_adherence/std': 0.1870439052581787, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9571428298950195, 'rewards/count_floor/std': 0.05611317232251167, 'rewards/lexical_plausibility/mean': 0.9653884172439575, 'rewards/lexical_plausibility/std': 0.06582440435886383, 'rewards/judge_quality/mean': 0.2499999850988388, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.8893945217132568, 'rewards/repeat_penalty/std': 0.18940770626068115, 'rewards/repeat_floor/mean': 0.9884815216064453, 'rewards/repeat_floor/std': 0.019608257338404655, 'rewards/near_duplicate_penalty/mean': 0.9775195121765137, 'rewards/near_duplicate_penalty/std': 0.034854110330343246, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9094846248626709, 'rewards/distinct_2/std': 0.1604984998703003, 'rewards/total_composite/mean': 0.2199091911315918, 'rewards/total_composite/std': 0.17031200230121613, 'reward': 0.2199091911315918, 'reward_std': 0.17031201720237732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20452137291431427, 'sampling/sampling_logp_difference/max': 1.7001419067382812, 'sampling/importance_sampling_ratio/min': 0.18265759944915771, 'sampling/importance_sampling_ratio/mean': 1.036447525024414, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5784726440906525, 'clip_ratio/low_mean': 0.04589068330824375, 'clip_ratio/low_min': 0.04589068330824375, 'clip_ratio/high_mean': 0.1026510912925005, 'clip_ratio/high_max': 0.1026510912925005, 'clip_ratio/region_mean': 0.14854177460074425, 'reward_total_mean': 0.2199091911315918, 'reward_meter_mean': 0.7809507846832275, 'reward_meter_std': 0.33067092299461365, 'reward_count_adherence_mean': 0.8571429252624512, 'reward_count_adherence_std': 0.1870439052581787, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9571428298950195, 'reward_count_floor_std': 0.05611317232251167, 'reward_lexical_plausibility_mean': 0.9653884172439575, 'reward_lexical_plausibility_std': 0.06582440435886383, 'reward_repeat_penalty_mean': 0.8893945217132568, 'reward_repeat_penalty_std': 0.18940770626068115, 'reward_repeat_floor_mean': 0.9884815216064453, 'reward_repeat_floor_std': 0.019608257338404655, 'reward_near_duplicate_penalty_mean': 0.9775195121765137, 'reward_near_duplicate_penalty_std': 0.034854110330343246, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9094846248626709, 'reward_distinct_2_std': 0.1604984998703003, 'reward_judge_quality_mean': 0.2499999850988388, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.2199091911315918, 'reward_total_composite_std': 0.17031200230121613, 'epoch': 0.01}

  8%|▊         | 269/3300 [58:36<9:31:11, 11.31s/it]INFO 04-12 16:57:02 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 270/3300 [58:49<10:06:20, 12.01s/it]2026-04-12 16:57:16,004 | INFO | train_grpo_train | metrics_logged mode=train step=270

                                                     
{'loss': -0.1885, 'grad_norm': 4.5092997550964355, 'learning_rate': 9.184848484848485e-06, 'num_tokens': 504535.0, 'completions/mean_length': 203.375, 'completions/min_length': 131.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 159.2857208251953, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.9101657867431641, 'rewards/meter/std': 0.2245728224515915, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9678571224212646, 'rewards/count_floor/std': 0.01983901485800743, 'rewards/lexical_plausibility/mean': 0.9549747705459595, 'rewards/lexical_plausibility/std': 0.1273505836725235, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.2886304259300232, 'rewards/repeat_penalty/std': 0.35808151960372925, 'rewards/repeat_floor/mean': 0.8897981643676758, 'rewards/repeat_floor/std': 0.07292575389146805, 'rewards/near_duplicate_penalty/mean': 0.7278239130973816, 'rewards/near_duplicate_penalty/std': 0.19873596727848053, 'rewards/opening_diversity/mean': 0.7885044813156128, 'rewards/opening_diversity/std': 0.2914164960384369, 'rewards/distinct_2/mean': 0.3701585531234741, 'rewards/distinct_2/std': 0.3632041811943054, 'rewards/total_composite/mean': 0.2062847912311554, 'rewards/total_composite/std': 0.12853997945785522, 'reward': 0.2062847912311554, 'reward_std': 0.12853999435901642, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12081237137317657, 'sampling/sampling_logp_difference/max': 1.8387348651885986, 'sampling/importance_sampling_ratio/min': 0.15901848673820496, 'sampling/importance_sampling_ratio/mean': 1.0191649198532104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9774073697626591, 'clip_ratio/low_mean': 0.034974793903529644, 'clip_ratio/low_min': 0.034974793903529644, 'clip_ratio/high_mean': 0.05621932726353407, 'clip_ratio/high_max': 0.05621932726353407, 'clip_ratio/region_mean': 0.09119412116706371, 'reward_total_mean': 0.2062847912311554, 'reward_meter_mean': 0.9101657867431641, 'reward_meter_std': 0.2245728224515915, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9678571224212646, 'reward_count_floor_std': 0.01983901485800743, 'reward_lexical_plausibility_mean': 0.9549747705459595, 'reward_lexical_plausibility_std': 0.1273505836725235, 'reward_repeat_penalty_mean': 0.2886304259300232, 'reward_repeat_penalty_std': 0.35808151960372925, 'reward_repeat_floor_mean': 0.8897981643676758, 'reward_repeat_floor_std': 0.07292575389146805, 'reward_near_duplicate_penalty_mean': 0.7278239130973816, 'reward_near_duplicate_penalty_std': 0.19873596727848053, 'reward_opening_diversity_mean': 0.7885044813156128, 'reward_opening_diversity_std': 0.2914164960384369, 'reward_distinct_2_mean': 0.3701585531234741, 'reward_distinct_2_std': 0.3632041811943054, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.2062847912311554, 'reward_total_composite_std': 0.12853997945785522, 'epoch': 0.01}

  8%|▊         | 270/3300 [58:49<10:06:20, 12.01s/it]INFO 04-12 16:57:16 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 271/3300 [58:58<9:19:10, 11.08s/it] 2026-04-12 16:57:24,909 | INFO | train_grpo_train | metrics_logged mode=train step=271

                                                    
{'loss': 0.0402, 'grad_norm': 15.954353332519531, 'learning_rate': 9.181818181818184e-06, 'num_tokens': 506323.0, 'completions/mean_length': 49.5, 'completions/min_length': 41.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8890167474746704, 'rewards/meter/std': 0.2670239210128784, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9825080633163452, 'rewards/repeat_penalty/std': 0.026309307664632797, 'rewards/repeat_floor/mean': 0.9973762035369873, 'rewards/repeat_floor/std': 0.003946396987885237, 'rewards/near_duplicate_penalty/mean': 0.9825080633163452, 'rewards/near_duplicate_penalty/std': 0.026309307664632797, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3503352403640747, 'rewards/total_composite/std': 0.11290138959884644, 'reward': 0.3503352403640747, 'reward_std': 0.11290138214826584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18417158722877502, 'sampling/sampling_logp_difference/max': 1.2668890953063965, 'sampling/importance_sampling_ratio/min': 0.28170663118362427, 'sampling/importance_sampling_ratio/mean': 1.0386143922805786, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9212061315774918, 'clip_ratio/low_mean': 0.08594587538391352, 'clip_ratio/low_min': 0.08594587538391352, 'clip_ratio/high_mean': 0.06211677938699722, 'clip_ratio/high_max': 0.06211677938699722, 'clip_ratio/region_mean': 0.14806265477091074, 'reward_total_mean': 0.3503352403640747, 'reward_meter_mean': 0.8890167474746704, 'reward_meter_std': 0.2670239210128784, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9825080633163452, 'reward_repeat_penalty_std': 0.026309307664632797, 'reward_repeat_floor_mean': 0.9973762035369873, 'reward_repeat_floor_std': 0.003946396987885237, 'reward_near_duplicate_penalty_mean': 0.9825080633163452, 'reward_near_duplicate_penalty_std': 0.026309307664632797, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.3503352403640747, 'reward_total_composite_std': 0.11290138959884644, 'epoch': 0.01}

  8%|▊         | 271/3300 [58:58<9:19:10, 11.08s/it]INFO 04-12 16:57:25 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 272/3300 [59:06<8:33:38, 10.18s/it]2026-04-12 16:57:32,989 | INFO | train_grpo_train | metrics_logged mode=train step=272

                                                    
{'loss': -0.0054, 'grad_norm': 17.56914710998535, 'learning_rate': 9.178787878787879e-06, 'num_tokens': 507864.0, 'completions/mean_length': 37.625, 'completions/min_length': 28.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5218707323074341, 'rewards/meter/std': 0.42126917839050293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6212499737739563, 'rewards/judge_quality/std': 0.2366696149110794, 'rewards/repeat_penalty/mean': 0.9531925916671753, 'rewards/repeat_penalty/std': 0.060648027807474136, 'rewards/repeat_floor/mean': 0.9934064745903015, 'rewards/repeat_floor/std': 0.008111451752483845, 'rewards/near_duplicate_penalty/mean': 0.961812436580658, 'rewards/near_duplicate_penalty/std': 0.042382217943668365, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.27356940507888794, 'rewards/total_composite/std': 0.18963998556137085, 'reward': 0.27356940507888794, 'reward_std': 0.18964000046253204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18089821934700012, 'sampling/sampling_logp_difference/max': 1.433882713317871, 'sampling/importance_sampling_ratio/min': 0.23838156461715698, 'sampling/importance_sampling_ratio/mean': 1.0495569705963135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7605403512716293, 'clip_ratio/low_mean': 0.07577519584447145, 'clip_ratio/low_min': 0.07577519584447145, 'clip_ratio/high_mean': 0.09659185446798801, 'clip_ratio/high_max': 0.09659185446798801, 'clip_ratio/region_mean': 0.17236705031245947, 'reward_total_mean': 0.27356940507888794, 'reward_meter_mean': 0.5218707323074341, 'reward_meter_std': 0.42126917839050293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9531925916671753, 'reward_repeat_penalty_std': 0.060648027807474136, 'reward_repeat_floor_mean': 0.9934064745903015, 'reward_repeat_floor_std': 0.008111451752483845, 'reward_near_duplicate_penalty_mean': 0.961812436580658, 'reward_near_duplicate_penalty_std': 0.042382217943668365, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.6212499737739563, 'reward_judge_quality_std': 0.2366696149110794, 'reward_total_composite_mean': 0.27356940507888794, 'reward_total_composite_std': 0.18963998556137085, 'epoch': 0.01}

  8%|▊         | 272/3300 [59:06<8:33:38, 10.18s/it]INFO 04-12 16:57:33 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 273/3300 [59:15<8:08:26,  9.68s/it]2026-04-12 16:57:41,514 | INFO | train_grpo_train | metrics_logged mode=train step=273

                                                    
{'loss': 0.0331, 'grad_norm': 14.007634162902832, 'learning_rate': 9.175757575757576e-06, 'num_tokens': 509639.0, 'completions/mean_length': 53.875, 'completions/min_length': 35.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7498190402984619, 'rewards/meter/std': 0.2539932429790497, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 0.9226579666137695, 'rewards/lexical_plausibility/std': 0.21875634789466858, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.7688454389572144, 'rewards/repeat_penalty/std': 0.32037031650543213, 'rewards/repeat_floor/mean': 0.9698281288146973, 'rewards/repeat_floor/std': 0.053871382027864456, 'rewards/near_duplicate_penalty/mean': 0.9199416637420654, 'rewards/near_duplicate_penalty/std': 0.14942771196365356, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.18898223340511322, 'rewards/distinct_2/mean': 0.8815209865570068, 'rewards/distinct_2/std': 0.24810031056404114, 'rewards/total_composite/mean': 0.20851707458496094, 'rewards/total_composite/std': 0.12288745492696762, 'reward': 0.20851707458496094, 'reward_std': 0.12288743257522583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1780700981616974, 'sampling/sampling_logp_difference/max': 1.558349370956421, 'sampling/importance_sampling_ratio/min': 0.21048320829868317, 'sampling/importance_sampling_ratio/mean': 1.0074797868728638, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.261663943529129, 'clip_ratio/low_mean': 0.06944405287504196, 'clip_ratio/low_min': 0.06944405287504196, 'clip_ratio/high_mean': 0.1003403440117836, 'clip_ratio/high_max': 0.1003403440117836, 'clip_ratio/region_mean': 0.16978439688682556, 'reward_total_mean': 0.20851707458496094, 'reward_meter_mean': 0.7498190402984619, 'reward_meter_std': 0.2539932429790497, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 0.9226579666137695, 'reward_lexical_plausibility_std': 0.21875634789466858, 'reward_repeat_penalty_mean': 0.7688454389572144, 'reward_repeat_penalty_std': 0.32037031650543213, 'reward_repeat_floor_mean': 0.9698281288146973, 'reward_repeat_floor_std': 0.053871382027864456, 'reward_near_duplicate_penalty_mean': 0.9199416637420654, 'reward_near_duplicate_penalty_std': 0.14942771196365356, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.18898223340511322, 'reward_distinct_2_mean': 0.8815209865570068, 'reward_distinct_2_std': 0.24810031056404114, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.20851707458496094, 'reward_total_composite_std': 0.12288745492696762, 'epoch': 0.01}

  8%|▊         | 273/3300 [59:15<8:08:26,  9.68s/it]INFO 04-12 16:57:41 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 274/3300 [59:23<7:45:59,  9.24s/it]2026-04-12 16:57:49,722 | INFO | train_grpo_train | metrics_logged mode=train step=274

                                                    
{'loss': -0.0698, 'grad_norm': 16.908870697021484, 'learning_rate': 9.172727272727274e-06, 'num_tokens': 511364.0, 'completions/mean_length': 45.625, 'completions/min_length': 36.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8064875602722168, 'rewards/meter/std': 0.2932147979736328, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9230809807777405, 'rewards/lexical_plausibility/std': 0.07844741642475128, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.35847893357276917, 'rewards/repeat_penalty/mean': 0.8598893284797668, 'rewards/repeat_penalty/std': 0.3481849730014801, 'rewards/repeat_floor/mean': 0.9718217849731445, 'rewards/repeat_floor/std': 0.07245302945375443, 'rewards/near_duplicate_penalty/mean': 0.9046551585197449, 'rewards/near_duplicate_penalty/std': 0.22198894619941711, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.3535533845424652, 'rewards/distinct_2/mean': 0.9291498064994812, 'rewards/distinct_2/std': 0.2003946453332901, 'rewards/total_composite/mean': 0.44291114807128906, 'rewards/total_composite/std': 0.33945468068122864, 'reward': 0.44291114807128906, 'reward_std': 0.33945468068122864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22264893352985382, 'sampling/sampling_logp_difference/max': 2.238255500793457, 'sampling/importance_sampling_ratio/min': 0.10664438456296921, 'sampling/importance_sampling_ratio/mean': 1.028499960899353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9048881530761719, 'clip_ratio/low_mean': 0.12737103272229433, 'clip_ratio/low_min': 0.12737103272229433, 'clip_ratio/high_mean': 0.059583334252238274, 'clip_ratio/high_max': 0.059583334252238274, 'clip_ratio/region_mean': 0.1869543669745326, 'reward_total_mean': 0.44291114807128906, 'reward_meter_mean': 0.8064875602722168, 'reward_meter_std': 0.2932147979736328, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9230809807777405, 'reward_lexical_plausibility_std': 0.07844741642475128, 'reward_repeat_penalty_mean': 0.8598893284797668, 'reward_repeat_penalty_std': 0.3481849730014801, 'reward_repeat_floor_mean': 0.9718217849731445, 'reward_repeat_floor_std': 0.07245302945375443, 'reward_near_duplicate_penalty_mean': 0.9046551585197449, 'reward_near_duplicate_penalty_std': 0.22198894619941711, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.3535533845424652, 'reward_distinct_2_mean': 0.9291498064994812, 'reward_distinct_2_std': 0.2003946453332901, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.35847893357276917, 'reward_total_composite_mean': 0.44291114807128906, 'reward_total_composite_std': 0.33945468068122864, 'epoch': 0.01}

  8%|▊         | 274/3300 [59:23<7:45:59,  9.24s/it]INFO 04-12 16:57:49 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 275/3300 [59:32<7:34:52,  9.02s/it]2026-04-12 16:57:58,238 | INFO | train_grpo_train | metrics_logged mode=train step=275

                                                    
{'loss': 0.0341, 'grad_norm': 16.079748153686523, 'learning_rate': 9.169696969696971e-06, 'num_tokens': 512948.0, 'completions/mean_length': 35.0, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7857391834259033, 'rewards/meter/std': 0.31257539987564087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9848056435585022, 'rewards/repeat_penalty/std': 0.041581347584724426, 'rewards/repeat_floor/mean': 0.9983481168746948, 'rewards/repeat_floor/std': 0.004463964607566595, 'rewards/near_duplicate_penalty/mean': 0.995775043964386, 'rewards/near_duplicate_penalty/std': 0.010607925243675709, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.4544386863708496, 'rewards/total_composite/std': 0.30342113971710205, 'reward': 0.4544386863708496, 'reward_std': 0.30342110991477966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16112521290779114, 'sampling/sampling_logp_difference/max': 1.8420591354370117, 'sampling/importance_sampling_ratio/min': 0.15849074721336365, 'sampling/importance_sampling_ratio/mean': 1.0035245418548584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1782893761992455, 'clip_ratio/low_mean': 0.1222641235217452, 'clip_ratio/low_min': 0.1222641235217452, 'clip_ratio/high_mean': 0.03282828349620104, 'clip_ratio/high_max': 0.03282828349620104, 'clip_ratio/region_mean': 0.15509240701794624, 'reward_total_mean': 0.4544386863708496, 'reward_meter_mean': 0.7857391834259033, 'reward_meter_std': 0.31257539987564087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9848056435585022, 'reward_repeat_penalty_std': 0.041581347584724426, 'reward_repeat_floor_mean': 0.9983481168746948, 'reward_repeat_floor_std': 0.004463964607566595, 'reward_near_duplicate_penalty_mean': 0.995775043964386, 'reward_near_duplicate_penalty_std': 0.010607925243675709, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.4544386863708496, 'reward_total_composite_std': 0.30342113971710205, 'epoch': 0.01}

  8%|▊         | 275/3300 [59:32<7:34:52,  9.02s/it]INFO 04-12 16:57:58 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 276/3300 [59:40<7:32:18,  8.97s/it]2026-04-12 16:58:07,099 | INFO | train_grpo_train | metrics_logged mode=train step=276

                                                    
{'loss': 0.0716, 'grad_norm': 21.031532287597656, 'learning_rate': 9.166666666666666e-06, 'num_tokens': 514792.0, 'completions/mean_length': 67.5, 'completions/min_length': 63.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.42907848954200745, 'rewards/meter/std': 0.23413845896720886, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9925000071525574, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9666136503219604, 'rewards/lexical_plausibility/std': 0.06906955689191818, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9715597033500671, 'rewards/repeat_penalty/std': 0.00994861125946045, 'rewards/repeat_floor/mean': 0.9957339763641357, 'rewards/repeat_floor/std': 0.001492294017225504, 'rewards/near_duplicate_penalty/mean': 0.9715597033500671, 'rewards/near_duplicate_penalty/std': 0.00994861125946045, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17563453316688538, 'rewards/total_composite/std': 0.1097688153386116, 'reward': 0.17563453316688538, 'reward_std': 0.109768807888031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2141132801771164, 'sampling/sampling_logp_difference/max': 2.7121477127075195, 'sampling/importance_sampling_ratio/min': 0.06639406085014343, 'sampling/importance_sampling_ratio/mean': 0.9953755736351013, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0091766491532326, 'clip_ratio/low_mean': 0.09259961731731892, 'clip_ratio/low_min': 0.09259961731731892, 'clip_ratio/high_mean': 0.06883588805794716, 'clip_ratio/high_max': 0.06883588805794716, 'clip_ratio/region_mean': 0.16143550537526608, 'reward_total_mean': 0.17563453316688538, 'reward_meter_mean': 0.42907848954200745, 'reward_meter_std': 0.23413845896720886, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9925000071525574, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9666136503219604, 'reward_lexical_plausibility_std': 0.06906955689191818, 'reward_repeat_penalty_mean': 0.9715597033500671, 'reward_repeat_penalty_std': 0.00994861125946045, 'reward_repeat_floor_mean': 0.9957339763641357, 'reward_repeat_floor_std': 0.001492294017225504, 'reward_near_duplicate_penalty_mean': 0.9715597033500671, 'reward_near_duplicate_penalty_std': 0.00994861125946045, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.17563453316688538, 'reward_total_composite_std': 0.1097688153386116, 'epoch': 0.01}

  8%|▊         | 276/3300 [59:40<7:32:18,  8.97s/it]INFO 04-12 16:58:07 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 277/3300 [59:51<8:00:11,  9.53s/it]2026-04-12 16:58:17,933 | INFO | train_grpo_train | metrics_logged mode=train step=277

                                                    
{'loss': 0.0255, 'grad_norm': 10.240793228149414, 'learning_rate': 9.163636363636365e-06, 'num_tokens': 517661.0, 'completions/mean_length': 159.625, 'completions/min_length': 135.0, 'completions/max_length': 195.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.625, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 195.0, 'rewards/meter/mean': 0.728797435760498, 'rewards/meter/std': 0.3267737925052643, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9571428298950195, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.8291047215461731, 'rewards/repeat_penalty/std': 0.14760397374629974, 'rewards/repeat_floor/mean': 0.9821028709411621, 'rewards/repeat_floor/std': 0.014180499128997326, 'rewards/near_duplicate_penalty/mean': 0.9613385200500488, 'rewards/near_duplicate_penalty/std': 0.019460754469037056, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8655791878700256, 'rewards/distinct_2/std': 0.12903448939323425, 'rewards/total_composite/mean': 0.2510860562324524, 'rewards/total_composite/std': 0.11653144657611847, 'reward': 0.2510860562324524, 'reward_std': 0.11653143912553787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19137711822986603, 'sampling/sampling_logp_difference/max': 2.0966250896453857, 'sampling/importance_sampling_ratio/min': 0.12287040799856186, 'sampling/importance_sampling_ratio/mean': 1.022000789642334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4952750951051712, 'clip_ratio/low_mean': 0.06852423213422298, 'clip_ratio/low_min': 0.06852423213422298, 'clip_ratio/high_mean': 0.10699446965008974, 'clip_ratio/high_max': 0.10699446965008974, 'clip_ratio/region_mean': 0.17551870178431273, 'reward_total_mean': 0.2510860562324524, 'reward_meter_mean': 0.728797435760498, 'reward_meter_std': 0.3267737925052643, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9571428298950195, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8291047215461731, 'reward_repeat_penalty_std': 0.14760397374629974, 'reward_repeat_floor_mean': 0.9821028709411621, 'reward_repeat_floor_std': 0.014180499128997326, 'reward_near_duplicate_penalty_mean': 0.9613385200500488, 'reward_near_duplicate_penalty_std': 0.019460754469037056, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8655791878700256, 'reward_distinct_2_std': 0.12903448939323425, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2510860562324524, 'reward_total_composite_std': 0.11653144657611847, 'epoch': 0.01}

  8%|▊         | 277/3300 [59:51<8:00:11,  9.53s/it]INFO 04-12 16:58:18 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 278/3300 [1:00:00<7:42:50,  9.19s/it]2026-04-12 16:58:26,348 | INFO | train_grpo_train | metrics_logged mode=train step=278

                                                      
{'loss': 0.0028, 'grad_norm': 17.15207862854004, 'learning_rate': 9.160606060606061e-06, 'num_tokens': 519255.0, 'completions/mean_length': 38.25, 'completions/min_length': 30.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6188963651657104, 'rewards/meter/std': 0.2938784956932068, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6712499856948853, 'rewards/judge_quality/std': 0.21216151118278503, 'rewards/repeat_penalty/mean': 0.9941428303718567, 'rewards/repeat_penalty/std': 0.012708526104688644, 'rewards/repeat_floor/mean': 0.9991214275360107, 'rewards/repeat_floor/std': 0.0019062899518758059, 'rewards/near_duplicate_penalty/mean': 0.9941428303718567, 'rewards/near_duplicate_penalty/std': 0.012708526104688644, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4416208267211914, 'rewards/total_composite/std': 0.2849549353122711, 'reward': 0.4416208267211914, 'reward_std': 0.2849549651145935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17788735032081604, 'sampling/sampling_logp_difference/max': 2.385617256164551, 'sampling/importance_sampling_ratio/min': 0.09203215688467026, 'sampling/importance_sampling_ratio/mean': 1.0471211671829224, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2068944796919823, 'clip_ratio/low_mean': 0.10344893299043179, 'clip_ratio/low_min': 0.10344893299043179, 'clip_ratio/high_mean': 0.05647659860551357, 'clip_ratio/high_max': 0.05647659860551357, 'clip_ratio/region_mean': 0.15992553159594536, 'reward_total_mean': 0.4416208267211914, 'reward_meter_mean': 0.6188963651657104, 'reward_meter_std': 0.2938784956932068, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9941428303718567, 'reward_repeat_penalty_std': 0.012708526104688644, 'reward_repeat_floor_mean': 0.9991214275360107, 'reward_repeat_floor_std': 0.0019062899518758059, 'reward_near_duplicate_penalty_mean': 0.9941428303718567, 'reward_near_duplicate_penalty_std': 0.012708526104688644, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6712499856948853, 'reward_judge_quality_std': 0.21216151118278503, 'reward_total_composite_mean': 0.4416208267211914, 'reward_total_composite_std': 0.2849549353122711, 'epoch': 0.01}

  8%|▊         | 278/3300 [1:00:00<7:42:50,  9.19s/it]INFO 04-12 16:58:26 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 279/3300 [1:00:14<8:56:04, 10.65s/it]2026-04-12 16:58:40,372 | INFO | train_grpo_train | metrics_logged mode=train step=279

                                                      
{'loss': -0.0698, 'grad_norm': 6.8083109855651855, 'learning_rate': 9.157575757575758e-06, 'num_tokens': 521177.0, 'completions/mean_length': 137.25, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.71428680419922, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.5831443667411804, 'rewards/meter/std': 0.38018566370010376, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9480080008506775, 'rewards/lexical_plausibility/std': 0.14705561101436615, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9823733568191528, 'rewards/repeat_penalty/std': 0.022886598482728004, 'rewards/repeat_floor/mean': 0.9979701042175293, 'rewards/repeat_floor/std': 0.0023654166143387556, 'rewards/near_duplicate_penalty/mean': 0.9927911758422852, 'rewards/near_duplicate_penalty/std': 0.007008022628724575, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9894605875015259, 'rewards/distinct_2/std': 0.01951758563518524, 'rewards/total_composite/mean': 0.21584241092205048, 'rewards/total_composite/std': 0.13955256342887878, 'reward': 0.21584241092205048, 'reward_std': 0.13955256342887878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20471937954425812, 'sampling/sampling_logp_difference/max': 2.334380626678467, 'sampling/importance_sampling_ratio/min': 0.09687045961618423, 'sampling/importance_sampling_ratio/mean': 1.0382238626480103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9867303967475891, 'clip_ratio/low_mean': 0.05974718276411295, 'clip_ratio/low_min': 0.05974718276411295, 'clip_ratio/high_mean': 0.1124513577669859, 'clip_ratio/high_max': 0.1124513577669859, 'clip_ratio/region_mean': 0.17219854053109884, 'reward_total_mean': 0.21584241092205048, 'reward_meter_mean': 0.5831443667411804, 'reward_meter_std': 0.38018566370010376, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9480080008506775, 'reward_lexical_plausibility_std': 0.14705561101436615, 'reward_repeat_penalty_mean': 0.9823733568191528, 'reward_repeat_penalty_std': 0.022886598482728004, 'reward_repeat_floor_mean': 0.9979701042175293, 'reward_repeat_floor_std': 0.0023654166143387556, 'reward_near_duplicate_penalty_mean': 0.9927911758422852, 'reward_near_duplicate_penalty_std': 0.007008022628724575, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9894605875015259, 'reward_distinct_2_std': 0.01951758563518524, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.21584241092205048, 'reward_total_composite_std': 0.13955256342887878, 'epoch': 0.01}

  8%|▊         | 279/3300 [1:00:14<8:56:04, 10.65s/it]INFO 04-12 16:58:40 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 280/3300 [1:00:23<8:34:38, 10.22s/it]2026-04-12 16:58:49,621 | INFO | train_grpo_train | metrics_logged mode=train step=280

                                                      
{'loss': 0.0238, 'grad_norm': 14.027534484863281, 'learning_rate': 9.154545454545455e-06, 'num_tokens': 523216.0, 'completions/mean_length': 66.875, 'completions/min_length': 52.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.6803208589553833, 'rewards/meter/std': 0.34330227971076965, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.05580177903175354, 'rewards/lexical_plausibility/mean': 0.9822304248809814, 'rewards/lexical_plausibility/std': 0.0502600334584713, 'rewards/judge_quality/mean': 0.4750000238418579, 'rewards/judge_quality/std': 0.27536988258361816, 'rewards/repeat_penalty/mean': 0.9405751824378967, 'rewards/repeat_penalty/std': 0.13471969962120056, 'rewards/repeat_floor/mean': 0.9933263659477234, 'rewards/repeat_floor/std': 0.01501352060586214, 'rewards/near_duplicate_penalty/mean': 0.9833229780197144, 'rewards/near_duplicate_penalty/std': 0.03350048512220383, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9588519334793091, 'rewards/distinct_2/std': 0.09674642980098724, 'rewards/total_composite/mean': 0.3439634144306183, 'rewards/total_composite/std': 0.2557860016822815, 'reward': 0.3439634144306183, 'reward_std': 0.2557860314846039, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18119318783283234, 'sampling/sampling_logp_difference/max': 2.4829185009002686, 'sampling/importance_sampling_ratio/min': 0.08349917083978653, 'sampling/importance_sampling_ratio/mean': 1.0178357362747192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5650077611207962, 'clip_ratio/low_mean': 0.10909930150955915, 'clip_ratio/low_min': 0.10909930150955915, 'clip_ratio/high_mean': 0.03057802841067314, 'clip_ratio/high_max': 0.03057802841067314, 'clip_ratio/region_mean': 0.1396773299202323, 'reward_total_mean': 0.3439634144306183, 'reward_meter_mean': 0.6803208589553833, 'reward_meter_std': 0.34330227971076965, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.05580177903175354, 'reward_lexical_plausibility_mean': 0.9822304248809814, 'reward_lexical_plausibility_std': 0.0502600334584713, 'reward_repeat_penalty_mean': 0.9405751824378967, 'reward_repeat_penalty_std': 0.13471969962120056, 'reward_repeat_floor_mean': 0.9933263659477234, 'reward_repeat_floor_std': 0.01501352060586214, 'reward_near_duplicate_penalty_mean': 0.9833229780197144, 'reward_near_duplicate_penalty_std': 0.03350048512220383, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9588519334793091, 'reward_distinct_2_std': 0.09674642980098724, 'reward_judge_quality_mean': 0.4750000238418579, 'reward_judge_quality_std': 0.27536988258361816, 'reward_total_composite_mean': 0.3439634144306183, 'reward_total_composite_std': 0.2557860016822815, 'epoch': 0.01}

  8%|▊         | 280/3300 [1:00:23<8:34:38, 10.22s/it]INFO 04-12 16:58:49 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 281/3300 [1:00:32<8:12:55,  9.80s/it]2026-04-12 16:58:58,420 | INFO | train_grpo_train | metrics_logged mode=train step=281

                                                      
{'loss': -0.0399, 'grad_norm': 15.764311790466309, 'learning_rate': 9.151515151515153e-06, 'num_tokens': 524735.0, 'completions/mean_length': 34.875, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7278113961219788, 'rewards/meter/std': 0.38084539771080017, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.9337247610092163, 'rewards/repeat_penalty/std': 0.1624285876750946, 'rewards/repeat_floor/mean': 0.9924181699752808, 'rewards/repeat_floor/std': 0.01770925335586071, 'rewards/near_duplicate_penalty/mean': 0.9755292534828186, 'rewards/near_duplicate_penalty/std': 0.04505830630660057, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9773755669593811, 'rewards/distinct_2/std': 0.06399156153202057, 'rewards/total_composite/mean': 0.4392854571342468, 'rewards/total_composite/std': 0.32982638478279114, 'reward': 0.4392854571342468, 'reward_std': 0.32982638478279114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16213636100292206, 'sampling/sampling_logp_difference/max': 1.188344955444336, 'sampling/importance_sampling_ratio/min': 0.30472517013549805, 'sampling/importance_sampling_ratio/mean': 1.0190047025680542, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9890183359384537, 'clip_ratio/low_mean': 0.07817160151898861, 'clip_ratio/low_min': 0.07817160151898861, 'clip_ratio/high_mean': 0.04904306307435036, 'clip_ratio/high_max': 0.04904306307435036, 'clip_ratio/region_mean': 0.12721466459333897, 'reward_total_mean': 0.4392854571342468, 'reward_meter_mean': 0.7278113961219788, 'reward_meter_std': 0.38084539771080017, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9337247610092163, 'reward_repeat_penalty_std': 0.1624285876750946, 'reward_repeat_floor_mean': 0.9924181699752808, 'reward_repeat_floor_std': 0.01770925335586071, 'reward_near_duplicate_penalty_mean': 0.9755292534828186, 'reward_near_duplicate_penalty_std': 0.04505830630660057, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9773755669593811, 'reward_distinct_2_std': 0.06399156153202057, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.4392854571342468, 'reward_total_composite_std': 0.32982638478279114, 'epoch': 0.01}

  9%|▊         | 281/3300 [1:00:32<8:12:55,  9.80s/it]INFO 04-12 16:58:58 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 282/3300 [1:00:41<7:59:38,  9.54s/it]2026-04-12 16:59:07,335 | INFO | train_grpo_train | metrics_logged mode=train step=282

                                                      
{'loss': 0.0643, 'grad_norm': 11.418998718261719, 'learning_rate': 9.148484848484848e-06, 'num_tokens': 526504.0, 'completions/mean_length': 41.125, 'completions/min_length': 35.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.38451939821243286, 'rewards/meter/std': 0.39363592863082886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5337499976158142, 'rewards/judge_quality/std': 0.268537700176239, 'rewards/repeat_penalty/mean': 0.8803739547729492, 'rewards/repeat_penalty/std': 0.3193255364894867, 'rewards/repeat_floor/mean': 0.9827873706817627, 'rewards/repeat_floor/std': 0.045831385999917984, 'rewards/near_duplicate_penalty/mean': 0.9524049758911133, 'rewards/near_duplicate_penalty/std': 0.11582916975021362, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.908906877040863, 'rewards/distinct_2/std': 0.2576502561569214, 'rewards/total_composite/mean': 0.23270350694656372, 'rewards/total_composite/std': 0.3042370080947876, 'reward': 0.23270350694656372, 'reward_std': 0.3042370080947876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18904976546764374, 'sampling/sampling_logp_difference/max': 1.35774564743042, 'sampling/importance_sampling_ratio/min': 0.25724002718925476, 'sampling/importance_sampling_ratio/mean': 1.0087424516677856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3395363837480545, 'clip_ratio/low_mean': 0.10560396127402782, 'clip_ratio/low_min': 0.10560396127402782, 'clip_ratio/high_mean': 0.08660891093313694, 'clip_ratio/high_max': 0.08660891093313694, 'clip_ratio/region_mean': 0.19221287220716476, 'reward_total_mean': 0.23270350694656372, 'reward_meter_mean': 0.38451939821243286, 'reward_meter_std': 0.39363592863082886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8803739547729492, 'reward_repeat_penalty_std': 0.3193255364894867, 'reward_repeat_floor_mean': 0.9827873706817627, 'reward_repeat_floor_std': 0.045831385999917984, 'reward_near_duplicate_penalty_mean': 0.9524049758911133, 'reward_near_duplicate_penalty_std': 0.11582916975021362, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.908906877040863, 'reward_distinct_2_std': 0.2576502561569214, 'reward_judge_quality_mean': 0.5337499976158142, 'reward_judge_quality_std': 0.268537700176239, 'reward_total_composite_mean': 0.23270350694656372, 'reward_total_composite_std': 0.3042370080947876, 'epoch': 0.01}

  9%|▊         | 282/3300 [1:00:41<7:59:38,  9.54s/it]INFO 04-12 16:59:07 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 283/3300 [1:00:48<7:30:57,  8.97s/it]2026-04-12 16:59:14,981 | INFO | train_grpo_train | metrics_logged mode=train step=283

                                                      
{'loss': 0.0162, 'grad_norm': 26.292882919311523, 'learning_rate': 9.145454545454546e-06, 'num_tokens': 527925.0, 'completions/mean_length': 19.625, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8695434331893921, 'rewards/meter/std': 0.3494793772697449, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.3620477616786957, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5417861938476562, 'rewards/total_composite/std': 0.4000450372695923, 'reward': 0.5417861938476562, 'reward_std': 0.4000450372695923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22980155050754547, 'sampling/sampling_logp_difference/max': 1.767982006072998, 'sampling/importance_sampling_ratio/min': 0.1706770658493042, 'sampling/importance_sampling_ratio/mean': 1.0327543020248413, 'sampling/importance_sampling_ratio/max': 1.93062424659729, 'entropy': 1.9435898512601852, 'clip_ratio/low_mean': 0.07785870879888535, 'clip_ratio/low_min': 0.07785870879888535, 'clip_ratio/high_mean': 0.09802359342575073, 'clip_ratio/high_max': 0.09802359342575073, 'clip_ratio/region_mean': 0.17588230222463608, 'reward_total_mean': 0.5417861938476562, 'reward_meter_mean': 0.8695434331893921, 'reward_meter_std': 0.3494793772697449, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.3620477616786957, 'reward_total_composite_mean': 0.5417861938476562, 'reward_total_composite_std': 0.4000450372695923, 'epoch': 0.01}

  9%|▊         | 283/3300 [1:00:48<7:30:57,  8.97s/it]INFO 04-12 16:59:15 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 284/3300 [1:00:58<7:49:03,  9.33s/it]2026-04-12 16:59:25,169 | INFO | train_grpo_train | metrics_logged mode=train step=284

                                                      
{'loss': 0.0677, 'grad_norm': 8.83140754699707, 'learning_rate': 9.142424242424243e-06, 'num_tokens': 530039.0, 'completions/mean_length': 83.25, 'completions/min_length': 55.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.25, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9847642183303833, 'rewards/meter/std': 0.011187692172825336, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.4587499797344208, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9827767610549927, 'rewards/repeat_penalty/std': 0.019196975976228714, 'rewards/repeat_floor/mean': 0.9979097843170166, 'rewards/repeat_floor/std': 0.002083313185721636, 'rewards/near_duplicate_penalty/mean': 0.9911524057388306, 'rewards/near_duplicate_penalty/std': 0.008081563748419285, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9915212988853455, 'rewards/distinct_2/std': 0.015731029212474823, 'rewards/total_composite/mean': 0.4405641555786133, 'rewards/total_composite/std': 0.19322501122951508, 'reward': 0.4405641555786133, 'reward_std': 0.19322501122951508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2049863040447235, 'sampling/sampling_logp_difference/max': 3.1022140979766846, 'sampling/importance_sampling_ratio/min': 0.044949568808078766, 'sampling/importance_sampling_ratio/mean': 1.041797161102295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.764414131641388, 'clip_ratio/low_mean': 0.11656825989484787, 'clip_ratio/low_min': 0.11656825989484787, 'clip_ratio/high_mean': 0.0931461350992322, 'clip_ratio/high_max': 0.0931461350992322, 'clip_ratio/region_mean': 0.20971439499408007, 'reward_total_mean': 0.4405641555786133, 'reward_meter_mean': 0.9847642183303833, 'reward_meter_std': 0.011187692172825336, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.9827767610549927, 'reward_repeat_penalty_std': 0.019196975976228714, 'reward_repeat_floor_mean': 0.9979097843170166, 'reward_repeat_floor_std': 0.002083313185721636, 'reward_near_duplicate_penalty_mean': 0.9911524057388306, 'reward_near_duplicate_penalty_std': 0.008081563748419285, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9915212988853455, 'reward_distinct_2_std': 0.015731029212474823, 'reward_judge_quality_mean': 0.4587499797344208, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.4405641555786133, 'reward_total_composite_std': 0.19322501122951508, 'epoch': 0.01}

  9%|▊         | 284/3300 [1:00:59<7:49:03,  9.33s/it]INFO 04-12 16:59:25 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 285/3300 [1:01:12<8:50:35, 10.56s/it]2026-04-12 16:59:38,583 | INFO | train_grpo_train | metrics_logged mode=train step=285

                                                      
{'loss': -0.0905, 'grad_norm': 3.715938091278076, 'learning_rate': 9.13939393939394e-06, 'num_tokens': 531627.0, 'completions/mean_length': 163.5, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 47.333335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8180938959121704, 'rewards/meter/std': 0.2964673340320587, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.910395622253418, 'rewards/lexical_plausibility/std': 0.15763933956623077, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.27458736300468445, 'rewards/repeat_penalty/mean': 0.908223032951355, 'rewards/repeat_penalty/std': 0.1296989619731903, 'rewards/repeat_floor/mean': 0.992796778678894, 'rewards/repeat_floor/std': 0.009746378287672997, 'rewards/near_duplicate_penalty/mean': 0.9912854433059692, 'rewards/near_duplicate_penalty/std': 0.012134880758821964, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9761550426483154, 'rewards/distinct_2/std': 0.048689957708120346, 'rewards/total_composite/mean': 0.3021094501018524, 'rewards/total_composite/std': 0.2318974882364273, 'reward': 0.3021094501018524, 'reward_std': 0.2318974733352661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23120664060115814, 'sampling/sampling_logp_difference/max': 1.9547584056854248, 'sampling/importance_sampling_ratio/min': 0.1415986716747284, 'sampling/importance_sampling_ratio/mean': 1.0576003789901733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7501201778650284, 'clip_ratio/low_mean': 0.02187499962747097, 'clip_ratio/low_min': 0.02187499962747097, 'clip_ratio/high_mean': 0.1224695174023509, 'clip_ratio/high_max': 0.1224695174023509, 'clip_ratio/region_mean': 0.14434451702982187, 'reward_total_mean': 0.3021094501018524, 'reward_meter_mean': 0.8180938959121704, 'reward_meter_std': 0.2964673340320587, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.910395622253418, 'reward_lexical_plausibility_std': 0.15763933956623077, 'reward_repeat_penalty_mean': 0.908223032951355, 'reward_repeat_penalty_std': 0.1296989619731903, 'reward_repeat_floor_mean': 0.992796778678894, 'reward_repeat_floor_std': 0.009746378287672997, 'reward_near_duplicate_penalty_mean': 0.9912854433059692, 'reward_near_duplicate_penalty_std': 0.012134880758821964, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9761550426483154, 'reward_distinct_2_std': 0.048689957708120346, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.27458736300468445, 'reward_total_composite_mean': 0.3021094501018524, 'reward_total_composite_std': 0.2318974882364273, 'epoch': 0.01}

  9%|▊         | 285/3300 [1:01:12<8:50:35, 10.56s/it]INFO 04-12 16:59:38 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 286/3300 [1:01:26<9:50:01, 11.75s/it]2026-04-12 16:59:53,102 | INFO | train_grpo_train | metrics_logged mode=train step=286

                                                      
{'loss': -0.2538, 'grad_norm': 4.339982509613037, 'learning_rate': 9.136363636363637e-06, 'num_tokens': 534582.0, 'completions/mean_length': 229.375, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 189.00001525878906, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 242.0, 'rewards/meter/mean': 0.975068211555481, 'rewards/meter/std': 0.025850258767604828, 'rewards/count_adherence/mean': 0.7638888955116272, 'rewards/count_adherence/std': 0.2749859690666199, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9291666746139526, 'rewards/count_floor/std': 0.0824957937002182, 'rewards/lexical_plausibility/mean': 0.9418684244155884, 'rewards/lexical_plausibility/std': 0.13166607916355133, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.6418706178665161, 'rewards/repeat_penalty/std': 0.3198451101779938, 'rewards/repeat_floor/mean': 0.9624670147895813, 'rewards/repeat_floor/std': 0.036328624933958054, 'rewards/near_duplicate_penalty/mean': 0.941243052482605, 'rewards/near_duplicate_penalty/std': 0.06054975837469101, 'rewards/opening_diversity/mean': 0.9635416865348816, 'rewards/opening_diversity/std': 0.06298425048589706, 'rewards/distinct_2/mean': 0.7052004337310791, 'rewards/distinct_2/std': 0.2991887032985687, 'rewards/total_composite/mean': 0.23565421998500824, 'rewards/total_composite/std': 0.12065204232931137, 'reward': 0.23565421998500824, 'reward_std': 0.12065203487873077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18805421888828278, 'sampling/sampling_logp_difference/max': 1.7813348770141602, 'sampling/importance_sampling_ratio/min': 0.1684131920337677, 'sampling/importance_sampling_ratio/mean': 1.0380291938781738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7005761861801147, 'clip_ratio/low_mean': 0.037321277894079685, 'clip_ratio/low_min': 0.037321277894079685, 'clip_ratio/high_mean': 0.09950423613190651, 'clip_ratio/high_max': 0.09950423613190651, 'clip_ratio/region_mean': 0.1368255140259862, 'reward_total_mean': 0.23565421998500824, 'reward_meter_mean': 0.975068211555481, 'reward_meter_std': 0.025850258767604828, 'reward_count_adherence_mean': 0.7638888955116272, 'reward_count_adherence_std': 0.2749859690666199, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9291666746139526, 'reward_count_floor_std': 0.0824957937002182, 'reward_lexical_plausibility_mean': 0.9418684244155884, 'reward_lexical_plausibility_std': 0.13166607916355133, 'reward_repeat_penalty_mean': 0.6418706178665161, 'reward_repeat_penalty_std': 0.3198451101779938, 'reward_repeat_floor_mean': 0.9624670147895813, 'reward_repeat_floor_std': 0.036328624933958054, 'reward_near_duplicate_penalty_mean': 0.941243052482605, 'reward_near_duplicate_penalty_std': 0.06054975837469101, 'reward_opening_diversity_mean': 0.9635416865348816, 'reward_opening_diversity_std': 0.06298425048589706, 'reward_distinct_2_mean': 0.7052004337310791, 'reward_distinct_2_std': 0.2991887032985687, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.23565421998500824, 'reward_total_composite_std': 0.12065204232931137, 'epoch': 0.02}

  9%|▊         | 286/3300 [1:01:26<9:50:01, 11.75s/it]INFO 04-12 16:59:53 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 287/3300 [1:01:35<9:08:00, 10.91s/it]2026-04-12 17:00:02,067 | INFO | train_grpo_train | metrics_logged mode=train step=287

                                                      
{'loss': 0.1296, 'grad_norm': 10.649018287658691, 'learning_rate': 9.133333333333335e-06, 'num_tokens': 536727.0, 'completions/mean_length': 82.125, 'completions/min_length': 70.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9153368473052979, 'rewards/meter/std': 0.12883660197257996, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.933342456817627, 'rewards/repeat_penalty/std': 0.07116208970546722, 'rewards/repeat_floor/mean': 0.9924843907356262, 'rewards/repeat_floor/std': 0.007308326195925474, 'rewards/near_duplicate_penalty/mean': 0.9770587682723999, 'rewards/near_duplicate_penalty/std': 0.01574035920202732, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9547288417816162, 'rewards/distinct_2/std': 0.06276123225688934, 'rewards/total_composite/mean': 0.5432270765304565, 'rewards/total_composite/std': 0.2151651531457901, 'reward': 0.5432270765304565, 'reward_std': 0.2151651382446289, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19400514662265778, 'sampling/sampling_logp_difference/max': 2.0500638484954834, 'sampling/importance_sampling_ratio/min': 0.12872667610645294, 'sampling/importance_sampling_ratio/mean': 1.0157383680343628, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5051335841417313, 'clip_ratio/low_mean': 0.10248292610049248, 'clip_ratio/low_min': 0.10248292610049248, 'clip_ratio/high_mean': 0.07777336798608303, 'clip_ratio/high_max': 0.07777336798608303, 'clip_ratio/region_mean': 0.1802562940865755, 'reward_total_mean': 0.5432270765304565, 'reward_meter_mean': 0.9153368473052979, 'reward_meter_std': 0.12883660197257996, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.933342456817627, 'reward_repeat_penalty_std': 0.07116208970546722, 'reward_repeat_floor_mean': 0.9924843907356262, 'reward_repeat_floor_std': 0.007308326195925474, 'reward_near_duplicate_penalty_mean': 0.9770587682723999, 'reward_near_duplicate_penalty_std': 0.01574035920202732, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9547288417816162, 'reward_distinct_2_std': 0.06276123225688934, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.5432270765304565, 'reward_total_composite_std': 0.2151651531457901, 'epoch': 0.02}

  9%|▊         | 287/3300 [1:01:35<9:08:00, 10.91s/it]INFO 04-12 17:00:02 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 288/3300 [1:01:43<8:22:38, 10.01s/it]2026-04-12 17:00:09,978 | INFO | train_grpo_train | metrics_logged mode=train step=288

                                                      
{'loss': 0.0657, 'grad_norm': 20.441247940063477, 'learning_rate': 9.130303030303032e-06, 'num_tokens': 538288.0, 'completions/mean_length': 31.125, 'completions/min_length': 28.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.691908597946167, 'rewards/meter/std': 0.3260622024536133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9883469343185425, 'rewards/repeat_penalty/std': 0.018376260995864868, 'rewards/repeat_floor/mean': 0.9982520341873169, 'rewards/repeat_floor/std': 0.002756433328613639, 'rewards/near_duplicate_penalty/mean': 0.9883469343185425, 'rewards/near_duplicate_penalty/std': 0.018376260995864868, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3841954171657562, 'rewards/total_composite/std': 0.2176816463470459, 'reward': 0.3841954171657562, 'reward_std': 0.2176816761493683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17679233849048615, 'sampling/sampling_logp_difference/max': 2.426983594894409, 'sampling/importance_sampling_ratio/min': 0.08830278366804123, 'sampling/importance_sampling_ratio/mean': 1.0266616344451904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1910634189844131, 'clip_ratio/low_mean': 0.07264610566198826, 'clip_ratio/low_min': 0.07264610566198826, 'clip_ratio/high_mean': 0.08435960672795773, 'clip_ratio/high_max': 0.08435960672795773, 'clip_ratio/region_mean': 0.15700571238994598, 'reward_total_mean': 0.3841954171657562, 'reward_meter_mean': 0.691908597946167, 'reward_meter_std': 0.3260622024536133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9883469343185425, 'reward_repeat_penalty_std': 0.018376260995864868, 'reward_repeat_floor_mean': 0.9982520341873169, 'reward_repeat_floor_std': 0.002756433328613639, 'reward_near_duplicate_penalty_mean': 0.9883469343185425, 'reward_near_duplicate_penalty_std': 0.018376260995864868, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.3841954171657562, 'reward_total_composite_std': 0.2176816463470459, 'epoch': 0.02}

  9%|▊         | 288/3300 [1:01:43<8:22:38, 10.01s/it]INFO 04-12 17:00:10 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 289/3300 [1:01:52<7:56:09,  9.49s/it]2026-04-12 17:00:18,243 | INFO | train_grpo_train | metrics_logged mode=train step=289

                                                      
{'loss': 0.0171, 'grad_norm': 14.392022132873535, 'learning_rate': 9.127272727272727e-06, 'num_tokens': 539931.0, 'completions/mean_length': 40.375, 'completions/min_length': 33.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7644199132919312, 'rewards/meter/std': 0.4023491442203522, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.04419417306780815, 'rewards/judge_quality/mean': 0.7387499809265137, 'rewards/judge_quality/std': 0.2687770128250122, 'rewards/repeat_penalty/mean': 0.9844462871551514, 'rewards/repeat_penalty/std': 0.03247229754924774, 'rewards/repeat_floor/mean': 0.9983397722244263, 'rewards/repeat_floor/std': 0.0030537107959389687, 'rewards/near_duplicate_penalty/mean': 0.9957189559936523, 'rewards/near_duplicate_penalty/std': 0.007621096912771463, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.5681802034378052, 'rewards/total_composite/std': 0.38565996289253235, 'reward': 0.5681802034378052, 'reward_std': 0.38565999269485474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18957798182964325, 'sampling/sampling_logp_difference/max': 1.7113916873931885, 'sampling/importance_sampling_ratio/min': 0.18061426281929016, 'sampling/importance_sampling_ratio/mean': 1.038847804069519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5313074737787247, 'clip_ratio/low_mean': 0.10078554414212704, 'clip_ratio/low_min': 0.10078554414212704, 'clip_ratio/high_mean': 0.09208153188228607, 'clip_ratio/high_max': 0.09208153188228607, 'clip_ratio/region_mean': 0.1928670760244131, 'reward_total_mean': 0.5681802034378052, 'reward_meter_mean': 0.7644199132919312, 'reward_meter_std': 0.4023491442203522, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.04419417306780815, 'reward_repeat_penalty_mean': 0.9844462871551514, 'reward_repeat_penalty_std': 0.03247229754924774, 'reward_repeat_floor_mean': 0.9983397722244263, 'reward_repeat_floor_std': 0.0030537107959389687, 'reward_near_duplicate_penalty_mean': 0.9957189559936523, 'reward_near_duplicate_penalty_std': 0.007621096912771463, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.7387499809265137, 'reward_judge_quality_std': 0.2687770128250122, 'reward_total_composite_mean': 0.5681802034378052, 'reward_total_composite_std': 0.38565996289253235, 'epoch': 0.02}

  9%|▉         | 289/3300 [1:01:52<7:56:09,  9.49s/it]INFO 04-12 17:00:18 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 290/3300 [1:01:59<7:20:56,  8.79s/it]2026-04-12 17:00:25,403 | INFO | train_grpo_train | metrics_logged mode=train step=290

                                                      
{'loss': -0.0533, 'grad_norm': 18.467220306396484, 'learning_rate': 9.124242424242425e-06, 'num_tokens': 541251.0, 'completions/mean_length': 19.0, 'completions/min_length': 13.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.0, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9261953830718994, 'rewards/meter/std': 0.07821160554885864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9237500429153442, 'rewards/judge_quality/std': 0.01060659158974886, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.8428959846496582, 'rewards/total_composite/std': 0.07375600188970566, 'reward': 0.8428959846496582, 'reward_std': 0.07375600188970566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1571652740240097, 'sampling/sampling_logp_difference/max': 1.5724208354949951, 'sampling/importance_sampling_ratio/min': 0.20754215121269226, 'sampling/importance_sampling_ratio/mean': 1.011788249015808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9865400865674019, 'clip_ratio/low_mean': 0.07368983514606953, 'clip_ratio/low_min': 0.07368983514606953, 'clip_ratio/high_mean': 0.07373120496049523, 'clip_ratio/high_max': 0.07373120496049523, 'clip_ratio/region_mean': 0.14742104010656476, 'reward_total_mean': 0.8428959846496582, 'reward_meter_mean': 0.9261953830718994, 'reward_meter_std': 0.07821160554885864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9237500429153442, 'reward_judge_quality_std': 0.01060659158974886, 'reward_total_composite_mean': 0.8428959846496582, 'reward_total_composite_std': 0.07375600188970566, 'epoch': 0.02}

  9%|▉         | 290/3300 [1:01:59<7:20:56,  8.79s/it]INFO 04-12 17:00:25 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 291/3300 [1:02:10<7:56:09,  9.49s/it]2026-04-12 17:00:36,545 | INFO | train_grpo_train | metrics_logged mode=train step=291

                                                      
{'loss': 0.0427, 'grad_norm': 12.23320484161377, 'learning_rate': 9.121212121212122e-06, 'num_tokens': 543317.0, 'completions/mean_length': 80.25, 'completions/min_length': 68.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.25, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.8204764127731323, 'rewards/meter/std': 0.22252389788627625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9235402345657349, 'rewards/repeat_penalty/std': 0.0735095664858818, 'rewards/repeat_floor/mean': 0.9913394451141357, 'rewards/repeat_floor/std': 0.007823782041668892, 'rewards/near_duplicate_penalty/mean': 0.970642626285553, 'rewards/near_duplicate_penalty/std': 0.02177763730287552, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.957909345626831, 'rewards/distinct_2/std': 0.05162857845425606, 'rewards/total_composite/mean': 0.26884764432907104, 'rewards/total_composite/std': 0.13277824223041534, 'reward': 0.26884764432907104, 'reward_std': 0.13277824223041534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21311388909816742, 'sampling/sampling_logp_difference/max': 2.26963472366333, 'sampling/importance_sampling_ratio/min': 0.10334992408752441, 'sampling/importance_sampling_ratio/mean': 1.0090792179107666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5345431864261627, 'clip_ratio/low_mean': 0.06731525808572769, 'clip_ratio/low_min': 0.06731525808572769, 'clip_ratio/high_mean': 0.12313556112349033, 'clip_ratio/high_max': 0.12313556112349033, 'clip_ratio/region_mean': 0.19045081920921803, 'reward_total_mean': 0.26884764432907104, 'reward_meter_mean': 0.8204764127731323, 'reward_meter_std': 0.22252389788627625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9235402345657349, 'reward_repeat_penalty_std': 0.0735095664858818, 'reward_repeat_floor_mean': 0.9913394451141357, 'reward_repeat_floor_std': 0.007823782041668892, 'reward_near_duplicate_penalty_mean': 0.970642626285553, 'reward_near_duplicate_penalty_std': 0.02177763730287552, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.957909345626831, 'reward_distinct_2_std': 0.05162857845425606, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.26884764432907104, 'reward_total_composite_std': 0.13277824223041534, 'epoch': 0.02}

  9%|▉         | 291/3300 [1:02:10<7:56:09,  9.49s/it]INFO 04-12 17:00:36 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 292/3300 [1:02:22<8:35:53, 10.29s/it]2026-04-12 17:00:48,690 | INFO | train_grpo_train | metrics_logged mode=train step=292

                                                      
{'loss': 0.0966, 'grad_norm': 5.87149715423584, 'learning_rate': 9.118181818181819e-06, 'num_tokens': 547735.0, 'completions/mean_length': 290.25, 'completions/min_length': 245.0, 'completions/max_length': 337.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 290.25, 'completions/min_terminated_length': 245.0, 'completions/max_terminated_length': 337.0, 'rewards/meter/mean': 0.7545753717422485, 'rewards/meter/std': 0.25584375858306885, 'rewards/count_adherence/mean': 0.8392857313156128, 'rewards/count_adherence/std': 0.06331465393304825, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.951785683631897, 'rewards/count_floor/std': 0.018994387239217758, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/repeat_penalty/mean': 0.5786294937133789, 'rewards/repeat_penalty/std': 0.365617036819458, 'rewards/repeat_floor/mean': 0.9374609589576721, 'rewards/repeat_floor/std': 0.08615387231111526, 'rewards/near_duplicate_penalty/mean': 0.8657467365264893, 'rewards/near_duplicate_penalty/std': 0.26059040427207947, 'rewards/opening_diversity/mean': 0.880859375, 'rewards/opening_diversity/std': 0.33698058128356934, 'rewards/distinct_2/mean': 0.608304500579834, 'rewards/distinct_2/std': 0.36024758219718933, 'rewards/total_composite/mean': 0.25096577405929565, 'rewards/total_composite/std': 0.16703999042510986, 'reward': 0.25096577405929565, 'reward_std': 0.16703997552394867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16984857618808746, 'sampling/sampling_logp_difference/max': 2.166409492492676, 'sampling/importance_sampling_ratio/min': 0.11458830535411835, 'sampling/importance_sampling_ratio/mean': 1.0080246925354004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3792203925549984, 'clip_ratio/low_mean': 0.0506894551217556, 'clip_ratio/low_min': 0.0506894551217556, 'clip_ratio/high_mean': 0.10293518006801605, 'clip_ratio/high_max': 0.10293518006801605, 'clip_ratio/region_mean': 0.15362463518977165, 'reward_total_mean': 0.25096577405929565, 'reward_meter_mean': 0.7545753717422485, 'reward_meter_std': 0.25584375858306885, 'reward_count_adherence_mean': 0.8392857313156128, 'reward_count_adherence_std': 0.06331465393304825, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.951785683631897, 'reward_count_floor_std': 0.018994387239217758, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5786294937133789, 'reward_repeat_penalty_std': 0.365617036819458, 'reward_repeat_floor_mean': 0.9374609589576721, 'reward_repeat_floor_std': 0.08615387231111526, 'reward_near_duplicate_penalty_mean': 0.8657467365264893, 'reward_near_duplicate_penalty_std': 0.26059040427207947, 'reward_opening_diversity_mean': 0.880859375, 'reward_opening_diversity_std': 0.33698058128356934, 'reward_distinct_2_mean': 0.608304500579834, 'reward_distinct_2_std': 0.36024758219718933, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.25096577405929565, 'reward_total_composite_std': 0.16703999042510986, 'epoch': 0.02}

  9%|▉         | 292/3300 [1:02:22<8:35:53, 10.29s/it]INFO 04-12 17:00:48 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 293/3300 [1:02:32<8:24:17, 10.06s/it]2026-04-12 17:00:58,219 | INFO | train_grpo_train | metrics_logged mode=train step=293

                                                      
{'loss': 0.0856, 'grad_norm': 13.421359062194824, 'learning_rate': 9.115151515151516e-06, 'num_tokens': 549806.0, 'completions/mean_length': 69.875, 'completions/min_length': 55.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.5824167728424072, 'rewards/meter/std': 0.3231014311313629, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9687185883522034, 'rewards/repeat_penalty/std': 0.02509327046573162, 'rewards/repeat_floor/mean': 0.9962530136108398, 'rewards/repeat_floor/std': 0.0025091669522225857, 'rewards/near_duplicate_penalty/mean': 0.9849103689193726, 'rewards/near_duplicate_penalty/std': 0.005877811461687088, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9835164546966553, 'rewards/distinct_2/std': 0.022749414667487144, 'rewards/total_composite/mean': 0.2556416094303131, 'rewards/total_composite/std': 0.12598897516727448, 'reward': 0.2556416094303131, 'reward_std': 0.12598896026611328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19969646632671356, 'sampling/sampling_logp_difference/max': 2.7699227333068848, 'sampling/importance_sampling_ratio/min': 0.06266684830188751, 'sampling/importance_sampling_ratio/mean': 1.011916995048523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4582985788583755, 'clip_ratio/low_mean': 0.08662841096520424, 'clip_ratio/low_min': 0.08662841096520424, 'clip_ratio/high_mean': 0.09717698208987713, 'clip_ratio/high_max': 0.09717698208987713, 'clip_ratio/region_mean': 0.18380539305508137, 'reward_total_mean': 0.2556416094303131, 'reward_meter_mean': 0.5824167728424072, 'reward_meter_std': 0.3231014311313629, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9687185883522034, 'reward_repeat_penalty_std': 0.02509327046573162, 'reward_repeat_floor_mean': 0.9962530136108398, 'reward_repeat_floor_std': 0.0025091669522225857, 'reward_near_duplicate_penalty_mean': 0.9849103689193726, 'reward_near_duplicate_penalty_std': 0.005877811461687088, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9835164546966553, 'reward_distinct_2_std': 0.022749414667487144, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.2556416094303131, 'reward_total_composite_std': 0.12598897516727448, 'epoch': 0.02}

  9%|▉         | 293/3300 [1:02:32<8:24:17, 10.06s/it]INFO 04-12 17:00:58 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 294/3300 [1:02:41<8:11:18,  9.81s/it]2026-04-12 17:01:07,429 | INFO | train_grpo_train | metrics_logged mode=train step=294

                                                      
{'loss': 0.27, 'grad_norm': 10.469935417175293, 'learning_rate': 9.112121212121214e-06, 'num_tokens': 551551.0, 'completions/mean_length': 68.125, 'completions/min_length': 43.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.125, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.4667671322822571, 'rewards/meter/std': 0.3156149685382843, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9322812557220459, 'rewards/lexical_plausibility/std': 0.16771572828292847, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.942449688911438, 'rewards/repeat_penalty/std': 0.13447071611881256, 'rewards/repeat_floor/mean': 0.9921880960464478, 'rewards/repeat_floor/std': 0.01786406897008419, 'rewards/near_duplicate_penalty/mean': 0.9650149345397949, 'rewards/near_duplicate_penalty/std': 0.07132438570261002, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9715099334716797, 'rewards/distinct_2/std': 0.08058197051286697, 'rewards/total_composite/mean': 0.17989294230937958, 'rewards/total_composite/std': 0.13754160702228546, 'reward': 0.17989294230937958, 'reward_std': 0.13754160702228546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19069939851760864, 'sampling/sampling_logp_difference/max': 1.7069435119628906, 'sampling/importance_sampling_ratio/min': 0.18141944706439972, 'sampling/importance_sampling_ratio/mean': 1.0364320278167725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1089000403881073, 'clip_ratio/low_mean': 0.10851164069026709, 'clip_ratio/low_min': 0.10851164069026709, 'clip_ratio/high_mean': 0.06903649494051933, 'clip_ratio/high_max': 0.06903649494051933, 'clip_ratio/region_mean': 0.17754813563078642, 'reward_total_mean': 0.17989294230937958, 'reward_meter_mean': 0.4667671322822571, 'reward_meter_std': 0.3156149685382843, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9322812557220459, 'reward_lexical_plausibility_std': 0.16771572828292847, 'reward_repeat_penalty_mean': 0.942449688911438, 'reward_repeat_penalty_std': 0.13447071611881256, 'reward_repeat_floor_mean': 0.9921880960464478, 'reward_repeat_floor_std': 0.01786406897008419, 'reward_near_duplicate_penalty_mean': 0.9650149345397949, 'reward_near_duplicate_penalty_std': 0.07132438570261002, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9715099334716797, 'reward_distinct_2_std': 0.08058197051286697, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.17989294230937958, 'reward_total_composite_std': 0.13754160702228546, 'epoch': 0.02}

  9%|▉         | 294/3300 [1:02:41<8:11:18,  9.81s/it]INFO 04-12 17:01:07 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 295/3300 [1:02:52<8:36:32, 10.31s/it]2026-04-12 17:01:18,927 | INFO | train_grpo_train | metrics_logged mode=train step=295

                                                      
{'loss': 0.0718, 'grad_norm': 8.419734001159668, 'learning_rate': 9.10909090909091e-06, 'num_tokens': 554948.0, 'completions/mean_length': 230.625, 'completions/min_length': 186.0, 'completions/max_length': 267.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.625, 'completions/min_terminated_length': 186.0, 'completions/max_terminated_length': 267.0, 'rewards/meter/mean': 0.6822469234466553, 'rewards/meter/std': 0.26969632506370544, 'rewards/count_adherence/mean': 0.8229166269302368, 'rewards/count_adherence/std': 0.05340583249926567, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9468749761581421, 'rewards/count_floor/std': 0.01602174900472164, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.49954450130462646, 'rewards/repeat_penalty/std': 0.3762837052345276, 'rewards/repeat_floor/mean': 0.9458256959915161, 'rewards/repeat_floor/std': 0.04378613457083702, 'rewards/near_duplicate_penalty/mean': 0.9092711210250854, 'rewards/near_duplicate_penalty/std': 0.07800742983818054, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5492782592773438, 'rewards/distinct_2/std': 0.3753294348716736, 'rewards/total_composite/mean': 0.0907188206911087, 'rewards/total_composite/std': 0.03504229336977005, 'reward': 0.0907188206911087, 'reward_std': 0.03504229336977005, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17964164912700653, 'sampling/sampling_logp_difference/max': 2.6832141876220703, 'sampling/importance_sampling_ratio/min': 0.0683431327342987, 'sampling/importance_sampling_ratio/mean': 1.016755223274231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.636583685874939, 'clip_ratio/low_mean': 0.044303144328296185, 'clip_ratio/low_min': 0.044303144328296185, 'clip_ratio/high_mean': 0.11310102976858616, 'clip_ratio/high_max': 0.11310102976858616, 'clip_ratio/region_mean': 0.15740417409688234, 'reward_total_mean': 0.0907188206911087, 'reward_meter_mean': 0.6822469234466553, 'reward_meter_std': 0.26969632506370544, 'reward_count_adherence_mean': 0.8229166269302368, 'reward_count_adherence_std': 0.05340583249926567, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9468749761581421, 'reward_count_floor_std': 0.01602174900472164, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.49954450130462646, 'reward_repeat_penalty_std': 0.3762837052345276, 'reward_repeat_floor_mean': 0.9458256959915161, 'reward_repeat_floor_std': 0.04378613457083702, 'reward_near_duplicate_penalty_mean': 0.9092711210250854, 'reward_near_duplicate_penalty_std': 0.07800742983818054, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5492782592773438, 'reward_distinct_2_std': 0.3753294348716736, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.0907188206911087, 'reward_total_composite_std': 0.03504229336977005, 'epoch': 0.02}

  9%|▉         | 295/3300 [1:02:52<8:36:32, 10.31s/it]INFO 04-12 17:01:19 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 296/3300 [1:03:01<8:16:39,  9.92s/it]2026-04-12 17:01:27,932 | INFO | train_grpo_train | metrics_logged mode=train step=296

                                                      
{'loss': 0.0595, 'grad_norm': 20.325502395629883, 'learning_rate': 9.106060606060606e-06, 'num_tokens': 556612.0, 'completions/mean_length': 37.0, 'completions/min_length': 31.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6823065280914307, 'rewards/meter/std': 0.3554304838180542, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.9453717470169067, 'rewards/repeat_penalty/std': 0.12018889933824539, 'rewards/repeat_floor/mean': 0.9951096773147583, 'rewards/repeat_floor/std': 0.010345329530537128, 'rewards/near_duplicate_penalty/mean': 0.9914365410804749, 'rewards/near_duplicate_penalty/std': 0.017113231122493744, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.03560846298933029, 'rewards/total_composite/mean': 0.33624467253685, 'rewards/total_composite/std': 0.21759910881519318, 'reward': 0.33624467253685, 'reward_std': 0.21759910881519318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17693468928337097, 'sampling/sampling_logp_difference/max': 1.5612378120422363, 'sampling/importance_sampling_ratio/min': 0.20987612009048462, 'sampling/importance_sampling_ratio/mean': 1.0141507387161255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.406125545501709, 'clip_ratio/low_mean': 0.060963901691138744, 'clip_ratio/low_min': 0.060963901691138744, 'clip_ratio/high_mean': 0.1305535202845931, 'clip_ratio/high_max': 0.1305535202845931, 'clip_ratio/region_mean': 0.19151742197573185, 'reward_total_mean': 0.33624467253685, 'reward_meter_mean': 0.6823065280914307, 'reward_meter_std': 0.3554304838180542, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9453717470169067, 'reward_repeat_penalty_std': 0.12018889933824539, 'reward_repeat_floor_mean': 0.9951096773147583, 'reward_repeat_floor_std': 0.010345329530537128, 'reward_near_duplicate_penalty_mean': 0.9914365410804749, 'reward_near_duplicate_penalty_std': 0.017113231122493744, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.03560846298933029, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.33624467253685, 'reward_total_composite_std': 0.21759910881519318, 'epoch': 0.02}

  9%|▉         | 296/3300 [1:03:01<8:16:39,  9.92s/it]INFO 04-12 17:01:28 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 297/3300 [1:03:10<7:58:22,  9.56s/it]2026-04-12 17:01:36,642 | INFO | train_grpo_train | metrics_logged mode=train step=297

                                                      
{'loss': 0.1529, 'grad_norm': 18.864351272583008, 'learning_rate': 9.103030303030304e-06, 'num_tokens': 558371.0, 'completions/mean_length': 53.875, 'completions/min_length': 22.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6778982877731323, 'rewards/meter/std': 0.3530502915382385, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8519001007080078, 'rewards/repeat_penalty/std': 0.17187410593032837, 'rewards/repeat_floor/mean': 0.9848721027374268, 'rewards/repeat_floor/std': 0.017985185608267784, 'rewards/near_duplicate_penalty/mean': 0.9594448804855347, 'rewards/near_duplicate_penalty/std': 0.04639763757586479, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.941170871257782, 'rewards/distinct_2/std': 0.09881649166345596, 'rewards/total_composite/mean': 0.25974708795547485, 'rewards/total_composite/std': 0.11362625658512115, 'reward': 0.25974708795547485, 'reward_std': 0.11362624913454056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17619004845619202, 'sampling/sampling_logp_difference/max': 1.391514778137207, 'sampling/importance_sampling_ratio/min': 0.27641117572784424, 'sampling/importance_sampling_ratio/mean': 1.0293015241622925, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2970117777585983, 'clip_ratio/low_mean': 0.059519230388104916, 'clip_ratio/low_min': 0.059519230388104916, 'clip_ratio/high_mean': 0.086685910820961, 'clip_ratio/high_max': 0.086685910820961, 'clip_ratio/region_mean': 0.14620514120906591, 'reward_total_mean': 0.25974708795547485, 'reward_meter_mean': 0.6778982877731323, 'reward_meter_std': 0.3530502915382385, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8519001007080078, 'reward_repeat_penalty_std': 0.17187410593032837, 'reward_repeat_floor_mean': 0.9848721027374268, 'reward_repeat_floor_std': 0.017985185608267784, 'reward_near_duplicate_penalty_mean': 0.9594448804855347, 'reward_near_duplicate_penalty_std': 0.04639763757586479, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.941170871257782, 'reward_distinct_2_std': 0.09881649166345596, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.25974708795547485, 'reward_total_composite_std': 0.11362625658512115, 'epoch': 0.02}

  9%|▉         | 297/3300 [1:03:10<7:58:22,  9.56s/it]INFO 04-12 17:01:36 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 298/3300 [1:03:18<7:37:58,  9.15s/it]2026-04-12 17:01:44,851 | INFO | train_grpo_train | metrics_logged mode=train step=298

                                                      
{'loss': 0.1765, 'grad_norm': 16.459354400634766, 'learning_rate': 9.100000000000001e-06, 'num_tokens': 559888.0, 'completions/mean_length': 36.625, 'completions/min_length': 28.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.3188633918762207, 'rewards/meter/std': 0.39131543040275574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9864082336425781, 'rewards/lexical_plausibility/std': 0.0384434312582016, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9565424919128418, 'rewards/repeat_penalty/std': 0.08799579739570618, 'rewards/repeat_floor/mean': 0.9944190979003906, 'rewards/repeat_floor/std': 0.010565795004367828, 'rewards/near_duplicate_penalty/mean': 0.9756143093109131, 'rewards/near_duplicate_penalty/std': 0.03482674062252045, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.14072012901306152, 'rewards/total_composite/std': 0.17767742276191711, 'reward': 0.14072012901306152, 'reward_std': 0.17767740786075592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22021010518074036, 'sampling/sampling_logp_difference/max': 1.7613716125488281, 'sampling/importance_sampling_ratio/min': 0.17180904746055603, 'sampling/importance_sampling_ratio/mean': 1.0427734851837158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9960065633058548, 'clip_ratio/low_mean': 0.155930295586586, 'clip_ratio/low_min': 0.155930295586586, 'clip_ratio/high_mean': 0.05565476417541504, 'clip_ratio/high_max': 0.05565476417541504, 'clip_ratio/region_mean': 0.21158505976200104, 'reward_total_mean': 0.14072012901306152, 'reward_meter_mean': 0.3188633918762207, 'reward_meter_std': 0.39131543040275574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9864082336425781, 'reward_lexical_plausibility_std': 0.0384434312582016, 'reward_repeat_penalty_mean': 0.9565424919128418, 'reward_repeat_penalty_std': 0.08799579739570618, 'reward_repeat_floor_mean': 0.9944190979003906, 'reward_repeat_floor_std': 0.010565795004367828, 'reward_near_duplicate_penalty_mean': 0.9756143093109131, 'reward_near_duplicate_penalty_std': 0.03482674062252045, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.14072012901306152, 'reward_total_composite_std': 0.17767742276191711, 'epoch': 0.02}

  9%|▉         | 298/3300 [1:03:18<7:37:58,  9.15s/it]INFO 04-12 17:01:45 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 299/3300 [1:03:27<7:32:59,  9.06s/it]2026-04-12 17:01:53,684 | INFO | train_grpo_train | metrics_logged mode=train step=299

                                                      
{'loss': -0.0105, 'grad_norm': 22.123252868652344, 'learning_rate': 9.096969696969698e-06, 'num_tokens': 561547.0, 'completions/mean_length': 40.375, 'completions/min_length': 32.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8137967586517334, 'rewards/meter/std': 0.3468204438686371, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.685253381729126, 'rewards/repeat_penalty/std': 0.23621059954166412, 'rewards/repeat_floor/mean': 0.9616619944572449, 'rewards/repeat_floor/std': 0.029687771573662758, 'rewards/near_duplicate_penalty/mean': 0.872746467590332, 'rewards/near_duplicate_penalty/std': 0.0836421325802803, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.8694448471069336, 'rewards/distinct_2/std': 0.14027510583400726, 'rewards/total_composite/mean': 0.3305690884590149, 'rewards/total_composite/std': 0.1938231736421585, 'reward': 0.3305690884590149, 'reward_std': 0.19382315874099731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.188039630651474, 'sampling/sampling_logp_difference/max': 1.361454963684082, 'sampling/importance_sampling_ratio/min': 0.2562876045703888, 'sampling/importance_sampling_ratio/mean': 1.0174654722213745, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6063385754823685, 'clip_ratio/low_mean': 0.04625741112977266, 'clip_ratio/low_min': 0.04625741112977266, 'clip_ratio/high_mean': 0.11450192146003246, 'clip_ratio/high_max': 0.11450192146003246, 'clip_ratio/region_mean': 0.16075933258980513, 'reward_total_mean': 0.3305690884590149, 'reward_meter_mean': 0.8137967586517334, 'reward_meter_std': 0.3468204438686371, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.685253381729126, 'reward_repeat_penalty_std': 0.23621059954166412, 'reward_repeat_floor_mean': 0.9616619944572449, 'reward_repeat_floor_std': 0.029687771573662758, 'reward_near_duplicate_penalty_mean': 0.872746467590332, 'reward_near_duplicate_penalty_std': 0.0836421325802803, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.8694448471069336, 'reward_distinct_2_std': 0.14027510583400726, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.3305690884590149, 'reward_total_composite_std': 0.1938231736421585, 'epoch': 0.02}

  9%|▉         | 299/3300 [1:03:27<7:32:59,  9.06s/it]INFO 04-12 17:01:53 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 300/3300 [1:03:37<7:42:20,  9.25s/it]2026-04-12 17:02:03,373 | INFO | train_grpo_train | metrics_logged mode=train step=300

                                                      
{'loss': 0.1294, 'grad_norm': 10.732985496520996, 'learning_rate': 9.093939393939395e-06, 'num_tokens': 563829.0, 'completions/mean_length': 102.25, 'completions/min_length': 87.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.25, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9553652405738831, 'rewards/meter/std': 0.062410105019807816, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9925000071525574, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9989224076271057, 'rewards/lexical_plausibility/std': 0.0030478707049041986, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.6932989358901978, 'rewards/repeat_penalty/std': 0.2703261375427246, 'rewards/repeat_floor/mean': 0.9686387777328491, 'rewards/repeat_floor/std': 0.02692396752536297, 'rewards/near_duplicate_penalty/mean': 0.9303176999092102, 'rewards/near_duplicate_penalty/std': 0.053814664483070374, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.7968460321426392, 'rewards/distinct_2/std': 0.17083317041397095, 'rewards/total_composite/mean': 0.31033384799957275, 'rewards/total_composite/std': 0.09578508138656616, 'reward': 0.31033384799957275, 'reward_std': 0.09578507393598557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1884276568889618, 'sampling/sampling_logp_difference/max': 1.5811958312988281, 'sampling/importance_sampling_ratio/min': 0.2057289183139801, 'sampling/importance_sampling_ratio/mean': 1.0302212238311768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8757694959640503, 'clip_ratio/low_mean': 0.04108297452330589, 'clip_ratio/low_min': 0.04108297452330589, 'clip_ratio/high_mean': 0.10998972225934267, 'clip_ratio/high_max': 0.10998972225934267, 'clip_ratio/region_mean': 0.15107269678264856, 'reward_total_mean': 0.31033384799957275, 'reward_meter_mean': 0.9553652405738831, 'reward_meter_std': 0.062410105019807816, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9925000071525574, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9989224076271057, 'reward_lexical_plausibility_std': 0.0030478707049041986, 'reward_repeat_penalty_mean': 0.6932989358901978, 'reward_repeat_penalty_std': 0.2703261375427246, 'reward_repeat_floor_mean': 0.9686387777328491, 'reward_repeat_floor_std': 0.02692396752536297, 'reward_near_duplicate_penalty_mean': 0.9303176999092102, 'reward_near_duplicate_penalty_std': 0.053814664483070374, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.7968460321426392, 'reward_distinct_2_std': 0.17083317041397095, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.31033384799957275, 'reward_total_composite_std': 0.09578508138656616, 'epoch': 0.02}

  9%|▉         | 300/3300 [1:03:37<7:42:20,  9.25s/it]INFO 04-12 17:02:03 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:09<00:52,  4.74s/it]

 23%|██▎       | 3/13 [00:15<00:54,  5.44s/it]

 31%|███       | 4/13 [00:27<01:09,  7.75s/it]

 38%|███▊      | 5/13 [00:36<01:06,  8.27s/it]

 46%|████▌     | 6/13 [00:47<01:02,  8.97s/it]

 54%|█████▍    | 7/13 [00:56<00:54,  9.05s/it]

 62%|██████▏   | 8/13 [01:08<00:49,  9.89s/it]

 69%|██████▉   | 9/13 [01:19<00:40, 10.24s/it]

 77%|███████▋  | 10/13 [01:27<00:29,  9.67s/it]

 85%|████████▍ | 11/13 [01:36<00:19,  9.50s/it]

 92%|█████████▏| 12/13 [01:48<00:10, 10.25s/it]

100%|██████████| 13/13 [01:59<00:00, 10.59s/it]2026-04-12 17:04:15,258 | INFO | train_grpo_train | metrics_logged mode=eval step=300

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 131.8575, 'eval_samples_per_second': 0.789, 'eval_steps_per_second': 0.099, 'eval_num_tokens': 563829.0, 'eval_completions/mean_length': 162.1346153846154, 'eval_completions/min_length': 34.0, 'eval_completions/max_length': 393.46153846153845, 'eval_completions/clipped_ratio': 0.07692307692307693, 'eval_completions/mean_terminated_length': 132.56273181621845, 'eval_completions/min_terminated_length': 34.0, 'eval_completions/max_terminated_length': 286.38461538461536, 'eval_rewards/meter/mean': 0.5356990442826197, 'eval_rewards/meter/std': 0.3575832729156201, 'eval_rewards/count_adherence/mean': 0.8707482906488272, 'eval_rewards/count_adherence/std': 0.149359007867483, 'eval_rewards/arabic_clean/mean': 0.8173076923076923, 'eval_rewards/arabic_clean/std': 0.3392545282840729, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9971153827813956, 'eval_rewards/arabic_floor/std': 0.00534126850274893, 'eval_rewards/count_floor/mean': 0.9612244780247028, 'eval_rewards/count_floor/std': 0.044807702159652345, 'eval_rewards/lexical_plausibility/mean': 0.9745920758980972, 'eval_rewards/lexical_plausibility/std': 0.05579023715108633, 'eval_rewards/judge_quality/mean': 0.3935576814871568, 'eval_rewards/judge_quality/std': 0.2108941158423057, 'eval_rewards/repeat_penalty/mean': 0.8596907533132113, 'eval_rewards/repeat_penalty/std': 0.20423086933218515, 'eval_rewards/repeat_floor/mean': 0.9841216344099778, 'eval_rewards/repeat_floor/std': 0.025210401222396355, 'eval_rewards/near_duplicate_penalty/mean': 0.9630640974411597, 'eval_rewards/near_duplicate_penalty/std': 0.05644430162814947, 'eval_rewards/opening_diversity/mean': 0.9803355702987084, 'eval_rewards/opening_diversity/std': 0.0391376565855283, 'eval_rewards/distinct_2/mean': 0.8982431384233328, 'eval_rewards/distinct_2/std': 0.17889117420865938, 'eval_rewards/total_composite/mean': 0.19682092506151933, 'eval_rewards/total_composite/std': 0.19787883586608446, 'eval_reward': 0.19682092506151933, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.12074239552021027, 'eval_sampling/sampling_logp_difference/max': 1.240494728088379, 'eval_sampling/importance_sampling_ratio/min': 0.2920835820528177, 'eval_sampling/importance_sampling_ratio/mean': 1.0341917826579168, 'eval_sampling/importance_sampling_ratio/max': 1.6474675215207613, 'eval_entropy': 1.6511328403766339, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.19682092506151933, 'eval_reward_meter_mean': 0.5356990442826197, 'eval_reward_meter_std': 0.3575832729156201, 'eval_reward_count_adherence_mean': 0.8707482906488272, 'eval_reward_count_adherence_std': 0.149359007867483, 'eval_reward_arabic_clean_mean': 0.8173076923076923, 'eval_reward_arabic_clean_std': 0.3392545282840729, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9971153827813956, 'eval_reward_arabic_floor_std': 0.00534126850274893, 'eval_reward_count_floor_mean': 0.9612244780247028, 'eval_reward_count_floor_std': 0.044807702159652345, 'eval_reward_lexical_plausibility_mean': 0.9745920758980972, 'eval_reward_lexical_plausibility_std': 0.05579023715108633, 'eval_reward_repeat_penalty_mean': 0.8596907533132113, 'eval_reward_repeat_penalty_std': 0.20423086933218515, 'eval_reward_repeat_floor_mean': 0.9841216344099778, 'eval_reward_repeat_floor_std': 0.025210401222396355, 'eval_reward_near_duplicate_penalty_mean': 0.9630640974411597, 'eval_reward_near_duplicate_penalty_std': 0.05644430162814947, 'eval_reward_opening_diversity_mean': 0.9803355702987084, 'eval_reward_opening_diversity_std': 0.0391376565855283, 'eval_reward_distinct_2_mean': 0.8982431384233328, 'eval_reward_distinct_2_std': 0.17889117420865938, 'eval_reward_judge_quality_mean': 0.3935576814871568, 'eval_reward_judge_quality_std': 0.2108941158423057, 'eval_reward_total_composite_mean': 0.19682092506151933, 'eval_reward_total_composite_std': 0.19787883586608446, 'epoch': 0.02}

  9%|▉         | 300/3300 [1:05:49<7:42:20,  9.25s/it]

100%|██████████| 13/13 [01:59<00:00, 10.59s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

  9%|▉         | 301/3300 [1:05:59<41:03:35, 49.29s/it]2026-04-12 17:04:26,090 | INFO | train_grpo_train | metrics_logged mode=train step=301

                                                       
{'loss': -0.0099, 'grad_norm': 14.73071002960205, 'learning_rate': 9.090909090909091e-06, 'num_tokens': 565447.0, 'completions/mean_length': 31.25, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9344687461853027, 'rewards/meter/std': 0.14397835731506348, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.528749942779541, 'rewards/judge_quality/std': 0.18333712220191956, 'rewards/repeat_penalty/mean': 0.9480298757553101, 'rewards/repeat_penalty/std': 0.02693019062280655, 'rewards/repeat_floor/mean': 0.9922044277191162, 'rewards/repeat_floor/std': 0.00403952831402421, 'rewards/near_duplicate_penalty/mean': 0.9480298757553101, 'rewards/near_duplicate_penalty/std': 0.02693019062280655, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5002779960632324, 'rewards/total_composite/std': 0.2085338681936264, 'reward': 0.5002779960632324, 'reward_std': 0.2085338681936264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14948254823684692, 'sampling/sampling_logp_difference/max': 1.0367588996887207, 'sampling/importance_sampling_ratio/min': 0.3546021282672882, 'sampling/importance_sampling_ratio/mean': 1.0220539569854736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9874745160341263, 'clip_ratio/low_mean': 0.0835039522498846, 'clip_ratio/low_min': 0.0835039522498846, 'clip_ratio/high_mean': 0.02812500041909516, 'clip_ratio/high_max': 0.02812500041909516, 'clip_ratio/region_mean': 0.11162895266897976, 'reward_total_mean': 0.5002779960632324, 'reward_meter_mean': 0.9344687461853027, 'reward_meter_std': 0.14397835731506348, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9480298757553101, 'reward_repeat_penalty_std': 0.02693019062280655, 'reward_repeat_floor_mean': 0.9922044277191162, 'reward_repeat_floor_std': 0.00403952831402421, 'reward_near_duplicate_penalty_mean': 0.9480298757553101, 'reward_near_duplicate_penalty_std': 0.02693019062280655, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.528749942779541, 'reward_judge_quality_std': 0.18333712220191956, 'reward_total_composite_mean': 0.5002779960632324, 'reward_total_composite_std': 0.2085338681936264, 'epoch': 0.02}

  9%|▉         | 301/3300 [1:05:59<41:03:35, 49.29s/it]INFO 04-12 17:04:26 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 302/3300 [1:06:07<30:33:42, 36.70s/it]2026-04-12 17:04:33,412 | INFO | train_grpo_train | metrics_logged mode=train step=302

                                                       
{'loss': 0.0162, 'grad_norm': 20.9350643157959, 'learning_rate': 9.087878787878788e-06, 'num_tokens': 566780.0, 'completions/mean_length': 17.625, 'completions/min_length': 15.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.9744971990585327, 'rewards/meter/std': 0.03883900120854378, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6812500357627869, 'rewards/judge_quality/std': 0.33523714542388916, 'rewards/repeat_penalty/mean': 0.7469173073768616, 'rewards/repeat_penalty/std': 0.00871915090829134, 'rewards/repeat_floor/mean': 0.9843834638595581, 'rewards/repeat_floor/std': 0.0017438469221815467, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6600235104560852, 'rewards/total_composite/std': 0.32864198088645935, 'reward': 0.6600235104560852, 'reward_std': 0.32864195108413696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16773949563503265, 'sampling/sampling_logp_difference/max': 1.3249053955078125, 'sampling/importance_sampling_ratio/min': 0.26582810282707214, 'sampling/importance_sampling_ratio/mean': 1.074419379234314, 'sampling/importance_sampling_ratio/max': 1.939653992652893, 'entropy': 1.5273366421461105, 'clip_ratio/low_mean': 0.09027777845039964, 'clip_ratio/low_min': 0.09027777845039964, 'clip_ratio/high_mean': 0.08864217903465033, 'clip_ratio/high_max': 0.08864217903465033, 'clip_ratio/region_mean': 0.17891995748504996, 'reward_total_mean': 0.6600235104560852, 'reward_meter_mean': 0.9744971990585327, 'reward_meter_std': 0.03883900120854378, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7469173073768616, 'reward_repeat_penalty_std': 0.00871915090829134, 'reward_repeat_floor_mean': 0.9843834638595581, 'reward_repeat_floor_std': 0.0017438469221815467, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6812500357627869, 'reward_judge_quality_std': 0.33523714542388916, 'reward_total_composite_mean': 0.6600235104560852, 'reward_total_composite_std': 0.32864198088645935, 'epoch': 0.02}

  9%|▉         | 302/3300 [1:06:07<30:33:42, 36.70s/it]INFO 04-12 17:04:33 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 303/3300 [1:06:14<23:16:01, 27.95s/it]2026-04-12 17:04:40,946 | INFO | train_grpo_train | metrics_logged mode=train step=303

                                                       
{'loss': -0.016, 'grad_norm': 22.582056045532227, 'learning_rate': 9.084848484848486e-06, 'num_tokens': 568200.0, 'completions/mean_length': 19.5, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8606141209602356, 'rewards/meter/std': 0.3475630581378937, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7275000214576721, 'rewards/judge_quality/std': 0.3564407229423523, 'rewards/repeat_penalty/mean': 0.692996621131897, 'rewards/repeat_penalty/std': 0.14628127217292786, 'rewards/repeat_floor/mean': 0.9733902215957642, 'rewards/repeat_floor/std': 0.029845232143998146, 'rewards/near_duplicate_penalty/mean': 0.9354219436645508, 'rewards/near_duplicate_penalty/std': 0.1628876030445099, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.5908148288726807, 'rewards/total_composite/std': 0.4106263220310211, 'reward': 0.5908148288726807, 'reward_std': 0.4106263220310211, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19459471106529236, 'sampling/sampling_logp_difference/max': 1.9386756420135498, 'sampling/importance_sampling_ratio/min': 0.14389438927173615, 'sampling/importance_sampling_ratio/mean': 0.9892681837081909, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3566118329763412, 'clip_ratio/low_mean': 0.0390625, 'clip_ratio/low_min': 0.0390625, 'clip_ratio/high_mean': 0.09017220605164766, 'clip_ratio/high_max': 0.09017220605164766, 'clip_ratio/region_mean': 0.12923470605164766, 'reward_total_mean': 0.5908148288726807, 'reward_meter_mean': 0.8606141209602356, 'reward_meter_std': 0.3475630581378937, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.692996621131897, 'reward_repeat_penalty_std': 0.14628127217292786, 'reward_repeat_floor_mean': 0.9733902215957642, 'reward_repeat_floor_std': 0.029845232143998146, 'reward_near_duplicate_penalty_mean': 0.9354219436645508, 'reward_near_duplicate_penalty_std': 0.1628876030445099, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.7275000214576721, 'reward_judge_quality_std': 0.3564407229423523, 'reward_total_composite_mean': 0.5908148288726807, 'reward_total_composite_std': 0.4106263220310211, 'epoch': 0.02}

  9%|▉         | 303/3300 [1:06:14<23:16:01, 27.95s/it]INFO 04-12 17:04:41 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 304/3300 [1:06:26<19:15:01, 23.13s/it]2026-04-12 17:04:52,838 | INFO | train_grpo_train | metrics_logged mode=train step=304

                                                       
{'loss': -0.074, 'grad_norm': 9.789715766906738, 'learning_rate': 9.081818181818183e-06, 'num_tokens': 571499.0, 'completions/mean_length': 197.375, 'completions/min_length': 13.0, 'completions/max_length': 268.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.375, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 268.0, 'rewards/meter/mean': 0.621332049369812, 'rewards/meter/std': 0.2757330536842346, 'rewards/count_adherence/mean': 0.6538461446762085, 'rewards/count_adherence/std': 0.2501056492328644, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8961538672447205, 'rewards/count_floor/std': 0.07503168284893036, 'rewards/lexical_plausibility/mean': 0.9904966354370117, 'rewards/lexical_plausibility/std': 0.02687946893274784, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.8952893018722534, 'rewards/repeat_penalty/std': 0.0984295904636383, 'rewards/repeat_floor/mean': 0.9903676509857178, 'rewards/repeat_floor/std': 0.007681592833250761, 'rewards/near_duplicate_penalty/mean': 0.9749888777732849, 'rewards/near_duplicate_penalty/std': 0.018389739096164703, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9554927349090576, 'rewards/distinct_2/std': 0.04836355149745941, 'rewards/total_composite/mean': 0.15165047347545624, 'rewards/total_composite/std': 0.12145166099071503, 'reward': 0.15165047347545624, 'reward_std': 0.12145166099071503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2635415494441986, 'sampling/sampling_logp_difference/max': 2.6532626152038574, 'sampling/importance_sampling_ratio/min': 0.07042107731103897, 'sampling/importance_sampling_ratio/mean': 0.990336537361145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.568487524986267, 'clip_ratio/low_mean': 0.12380932085216045, 'clip_ratio/low_min': 0.12380932085216045, 'clip_ratio/high_mean': 0.12513156607747078, 'clip_ratio/high_max': 0.12513156607747078, 'clip_ratio/region_mean': 0.24894088692963123, 'reward_total_mean': 0.15165047347545624, 'reward_meter_mean': 0.621332049369812, 'reward_meter_std': 0.2757330536842346, 'reward_count_adherence_mean': 0.6538461446762085, 'reward_count_adherence_std': 0.2501056492328644, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8961538672447205, 'reward_count_floor_std': 0.07503168284893036, 'reward_lexical_plausibility_mean': 0.9904966354370117, 'reward_lexical_plausibility_std': 0.02687946893274784, 'reward_repeat_penalty_mean': 0.8952893018722534, 'reward_repeat_penalty_std': 0.0984295904636383, 'reward_repeat_floor_mean': 0.9903676509857178, 'reward_repeat_floor_std': 0.007681592833250761, 'reward_near_duplicate_penalty_mean': 0.9749888777732849, 'reward_near_duplicate_penalty_std': 0.018389739096164703, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9554927349090576, 'reward_distinct_2_std': 0.04836355149745941, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.15165047347545624, 'reward_total_composite_std': 0.12145166099071503, 'epoch': 0.02}

  9%|▉         | 304/3300 [1:06:26<19:15:01, 23.13s/it]INFO 04-12 17:04:53 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 305/3300 [1:06:33<15:13:26, 18.30s/it]2026-04-12 17:04:59,862 | INFO | train_grpo_train | metrics_logged mode=train step=305

                                                       
{'loss': -0.0101, 'grad_norm': 19.34548568725586, 'learning_rate': 9.078787878787878e-06, 'num_tokens': 572837.0, 'completions/mean_length': 16.25, 'completions/min_length': 16.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.9452635645866394, 'rewards/meter/std': 0.13600192964076996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3462500274181366, 'rewards/judge_quality/std': 0.3634728491306305, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2931850552558899, 'rewards/total_composite/std': 0.2865670323371887, 'reward': 0.2931850552558899, 'reward_std': 0.2865670323371887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06238498538732529, 'sampling/sampling_logp_difference/max': 0.9791786670684814, 'sampling/importance_sampling_ratio/min': 0.375619500875473, 'sampling/importance_sampling_ratio/mean': 1.0298633575439453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4287057965993881, 'clip_ratio/low_mean': 0.0390625, 'clip_ratio/low_min': 0.0390625, 'clip_ratio/high_mean': 0.0069444444961845875, 'clip_ratio/high_max': 0.0069444444961845875, 'clip_ratio/region_mean': 0.04600694449618459, 'reward_total_mean': 0.2931850552558899, 'reward_meter_mean': 0.9452635645866394, 'reward_meter_std': 0.13600192964076996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3462500274181366, 'reward_judge_quality_std': 0.3634728491306305, 'reward_total_composite_mean': 0.2931850552558899, 'reward_total_composite_std': 0.2865670323371887, 'epoch': 0.02}

  9%|▉         | 305/3300 [1:06:33<15:13:26, 18.30s/it]INFO 04-12 17:05:00 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 306/3300 [1:06:41<12:39:58, 15.23s/it]2026-04-12 17:05:07,936 | INFO | train_grpo_train | metrics_logged mode=train step=306

                                                       
{'loss': 0.0464, 'grad_norm': 19.623510360717773, 'learning_rate': 9.075757575757577e-06, 'num_tokens': 574345.0, 'completions/mean_length': 30.5, 'completions/min_length': 26.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9649000763893127, 'rewards/meter/std': 0.03835184872150421, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6424999833106995, 'rewards/judge_quality/std': 0.27395257353782654, 'rewards/repeat_penalty/mean': 0.9656938314437866, 'rewards/repeat_penalty/std': 0.02919146418571472, 'rewards/repeat_floor/mean': 0.9948540925979614, 'rewards/repeat_floor/std': 0.004378714598715305, 'rewards/near_duplicate_penalty/mean': 0.9656938314437866, 'rewards/near_duplicate_penalty/std': 0.02919146418571472, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6089813709259033, 'rewards/total_composite/std': 0.2839276194572449, 'reward': 0.6089813709259033, 'reward_std': 0.2839275896549225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17152456939220428, 'sampling/sampling_logp_difference/max': 1.2357168197631836, 'sampling/importance_sampling_ratio/min': 0.2906263768672943, 'sampling/importance_sampling_ratio/mean': 1.0257127285003662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.151340663433075, 'clip_ratio/low_mean': 0.07075957581400871, 'clip_ratio/low_min': 0.07075957581400871, 'clip_ratio/high_mean': 0.07881645485758781, 'clip_ratio/high_max': 0.07881645485758781, 'clip_ratio/region_mean': 0.14957603067159653, 'reward_total_mean': 0.6089813709259033, 'reward_meter_mean': 0.9649000763893127, 'reward_meter_std': 0.03835184872150421, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9656938314437866, 'reward_repeat_penalty_std': 0.02919146418571472, 'reward_repeat_floor_mean': 0.9948540925979614, 'reward_repeat_floor_std': 0.004378714598715305, 'reward_near_duplicate_penalty_mean': 0.9656938314437866, 'reward_near_duplicate_penalty_std': 0.02919146418571472, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6424999833106995, 'reward_judge_quality_std': 0.27395257353782654, 'reward_total_composite_mean': 0.6089813709259033, 'reward_total_composite_std': 0.2839276194572449, 'epoch': 0.02}

  9%|▉         | 306/3300 [1:06:41<12:39:58, 15.23s/it]INFO 04-12 17:05:08 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 307/3300 [1:06:50<11:02:22, 13.28s/it]2026-04-12 17:05:16,654 | INFO | train_grpo_train | metrics_logged mode=train step=307

                                                       
{'loss': -0.0074, 'grad_norm': 15.526321411132812, 'learning_rate': 9.072727272727273e-06, 'num_tokens': 575903.0, 'completions/mean_length': 42.75, 'completions/min_length': 28.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.2503531873226166, 'rewards/meter/std': 0.32325610518455505, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5837500095367432, 'rewards/judge_quality/std': 0.2719736099243164, 'rewards/repeat_penalty/mean': 0.9855555295944214, 'rewards/repeat_penalty/std': 0.017932593822479248, 'rewards/repeat_floor/mean': 0.9978333115577698, 'rewards/repeat_floor/std': 0.002689896384254098, 'rewards/near_duplicate_penalty/mean': 0.9855555295944214, 'rewards/near_duplicate_penalty/std': 0.017932593822479248, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12381266057491302, 'rewards/total_composite/std': 0.15705926716327667, 'reward': 0.12381266057491302, 'reward_std': 0.15705928206443787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19509024918079376, 'sampling/sampling_logp_difference/max': 1.6404211521148682, 'sampling/importance_sampling_ratio/min': 0.237303227186203, 'sampling/importance_sampling_ratio/mean': 1.0305685997009277, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5318706631660461, 'clip_ratio/low_mean': 0.11107687465846539, 'clip_ratio/low_min': 0.11107687465846539, 'clip_ratio/high_mean': 0.05315170995891094, 'clip_ratio/high_max': 0.05315170995891094, 'clip_ratio/region_mean': 0.16422858461737633, 'reward_total_mean': 0.12381266057491302, 'reward_meter_mean': 0.2503531873226166, 'reward_meter_std': 0.32325610518455505, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9855555295944214, 'reward_repeat_penalty_std': 0.017932593822479248, 'reward_repeat_floor_mean': 0.9978333115577698, 'reward_repeat_floor_std': 0.002689896384254098, 'reward_near_duplicate_penalty_mean': 0.9855555295944214, 'reward_near_duplicate_penalty_std': 0.017932593822479248, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5837500095367432, 'reward_judge_quality_std': 0.2719736099243164, 'reward_total_composite_mean': 0.12381266057491302, 'reward_total_composite_std': 0.15705926716327667, 'epoch': 0.02}

  9%|▉         | 307/3300 [1:06:50<11:02:22, 13.28s/it]INFO 04-12 17:05:16 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 308/3300 [1:06:59<9:53:36, 11.90s/it] 2026-04-12 17:05:25,354 | INFO | train_grpo_train | metrics_logged mode=train step=308

                                                      
{'loss': 0.001, 'grad_norm': 15.640469551086426, 'learning_rate': 9.06969696969697e-06, 'num_tokens': 577647.0, 'completions/mean_length': 40.0, 'completions/min_length': 35.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9640589952468872, 'rewards/meter/std': 0.05266324058175087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.1164964810013771, 'rewards/repeat_penalty/mean': 0.9954004287719727, 'rewards/repeat_penalty/std': 0.013009531423449516, 'rewards/repeat_floor/mean': 0.9993100762367249, 'rewards/repeat_floor/std': 0.0019514414016157389, 'rewards/near_duplicate_penalty/mean': 0.9954004287719727, 'rewards/near_duplicate_penalty/std': 0.013009531423449516, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4585379958152771, 'rewards/total_composite/std': 0.12229236215353012, 'reward': 0.4585379958152771, 'reward_std': 0.12229236960411072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1859753131866455, 'sampling/sampling_logp_difference/max': 1.691523551940918, 'sampling/importance_sampling_ratio/min': 0.18423861265182495, 'sampling/importance_sampling_ratio/mean': 1.0469082593917847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8675436079502106, 'clip_ratio/low_mean': 0.1685184296220541, 'clip_ratio/low_min': 0.1685184296220541, 'clip_ratio/high_mean': 0.028846153989434242, 'clip_ratio/high_max': 0.028846153989434242, 'clip_ratio/region_mean': 0.19736458361148834, 'reward_total_mean': 0.4585379958152771, 'reward_meter_mean': 0.9640589952468872, 'reward_meter_std': 0.05266324058175087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9954004287719727, 'reward_repeat_penalty_std': 0.013009531423449516, 'reward_repeat_floor_mean': 0.9993100762367249, 'reward_repeat_floor_std': 0.0019514414016157389, 'reward_near_duplicate_penalty_mean': 0.9954004287719727, 'reward_near_duplicate_penalty_std': 0.013009531423449516, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.1164964810013771, 'reward_total_composite_mean': 0.4585379958152771, 'reward_total_composite_std': 0.12229236215353012, 'epoch': 0.02}

  9%|▉         | 308/3300 [1:06:59<9:53:36, 11.90s/it]INFO 04-12 17:05:25 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 309/3300 [1:07:09<9:34:47, 11.53s/it]2026-04-12 17:05:36,011 | INFO | train_grpo_train | metrics_logged mode=train step=309

                                                      
{'loss': -0.0282, 'grad_norm': 7.242228031158447, 'learning_rate': 9.066666666666667e-06, 'num_tokens': 580498.0, 'completions/mean_length': 172.375, 'completions/min_length': 141.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 172.375, 'completions/min_terminated_length': 141.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.49982455372810364, 'rewards/meter/std': 0.3290709853172302, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.0927247703075409, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.027817437425255775, 'rewards/lexical_plausibility/mean': 0.9972826242446899, 'rewards/lexical_plausibility/std': 0.007685941644012928, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.08864052593708038, 'rewards/repeat_penalty/mean': 0.5439807772636414, 'rewards/repeat_penalty/std': 0.33551767468452454, 'rewards/repeat_floor/mean': 0.9450145363807678, 'rewards/repeat_floor/std': 0.05226676166057587, 'rewards/near_duplicate_penalty/mean': 0.8956697583198547, 'rewards/near_duplicate_penalty/std': 0.12272150069475174, 'rewards/opening_diversity/mean': 0.9176136255264282, 'rewards/opening_diversity/std': 0.23302382230758667, 'rewards/distinct_2/mean': 0.6178585290908813, 'rewards/distinct_2/std': 0.28116124868392944, 'rewards/total_composite/mean': 0.07592283189296722, 'rewards/total_composite/std': 0.04808022081851959, 'reward': 0.07592283189296722, 'reward_std': 0.04808022454380989, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18043243885040283, 'sampling/sampling_logp_difference/max': 1.9903476238250732, 'sampling/importance_sampling_ratio/min': 0.13664790987968445, 'sampling/importance_sampling_ratio/mean': 1.0302400588989258, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1103629022836685, 'clip_ratio/low_mean': 0.08400345873087645, 'clip_ratio/low_min': 0.08400345873087645, 'clip_ratio/high_mean': 0.07768644392490387, 'clip_ratio/high_max': 0.07768644392490387, 'clip_ratio/region_mean': 0.16168990265578032, 'reward_total_mean': 0.07592283189296722, 'reward_meter_mean': 0.49982455372810364, 'reward_meter_std': 0.3290709853172302, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.0927247703075409, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.027817437425255775, 'reward_lexical_plausibility_mean': 0.9972826242446899, 'reward_lexical_plausibility_std': 0.007685941644012928, 'reward_repeat_penalty_mean': 0.5439807772636414, 'reward_repeat_penalty_std': 0.33551767468452454, 'reward_repeat_floor_mean': 0.9450145363807678, 'reward_repeat_floor_std': 0.05226676166057587, 'reward_near_duplicate_penalty_mean': 0.8956697583198547, 'reward_near_duplicate_penalty_std': 0.12272150069475174, 'reward_opening_diversity_mean': 0.9176136255264282, 'reward_opening_diversity_std': 0.23302382230758667, 'reward_distinct_2_mean': 0.6178585290908813, 'reward_distinct_2_std': 0.28116124868392944, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.08864052593708038, 'reward_total_composite_mean': 0.07592283189296722, 'reward_total_composite_std': 0.04808022081851959, 'epoch': 0.02}

  9%|▉         | 309/3300 [1:07:09<9:34:47, 11.53s/it]INFO 04-12 17:05:36 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 310/3300 [1:07:17<8:37:46, 10.39s/it]2026-04-12 17:05:43,739 | INFO | train_grpo_train | metrics_logged mode=train step=310

                                                      
{'loss': 0.011, 'grad_norm': 19.942989349365234, 'learning_rate': 9.063636363636365e-06, 'num_tokens': 581819.0, 'completions/mean_length': 15.125, 'completions/min_length': 14.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.125, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.9323422908782959, 'rewards/meter/std': 0.11280333995819092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7914478778839111, 'rewards/total_composite/std': 0.17643626034259796, 'reward': 0.7914478778839111, 'reward_std': 0.17643627524375916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1103106439113617, 'sampling/sampling_logp_difference/max': 1.2927289009094238, 'sampling/importance_sampling_ratio/min': 0.274520605802536, 'sampling/importance_sampling_ratio/mean': 0.9970378875732422, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8194018229842186, 'clip_ratio/low_mean': 0.023634454235434532, 'clip_ratio/low_min': 0.023634454235434532, 'clip_ratio/high_mean': 0.057440477423369884, 'clip_ratio/high_max': 0.057440477423369884, 'clip_ratio/region_mean': 0.08107493165880442, 'reward_total_mean': 0.7914478778839111, 'reward_meter_mean': 0.9323422908782959, 'reward_meter_std': 0.11280333995819092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.7914478778839111, 'reward_total_composite_std': 0.17643626034259796, 'epoch': 0.02}

  9%|▉         | 310/3300 [1:07:17<8:37:46, 10.39s/it]INFO 04-12 17:05:43 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 311/3300 [1:07:28<8:40:11, 10.44s/it]2026-04-12 17:05:54,306 | INFO | train_grpo_train | metrics_logged mode=train step=311

                                                      
{'loss': 0.1243, 'grad_norm': 23.695682525634766, 'learning_rate': 9.06060606060606e-06, 'num_tokens': 584255.0, 'completions/mean_length': 127.5, 'completions/min_length': 99.0, 'completions/max_length': 168.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.5, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 168.0, 'rewards/meter/mean': 0.7575584650039673, 'rewards/meter/std': 0.12200148403644562, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9357142448425293, 'rewards/count_floor/std': 0.02290808968245983, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9072020053863525, 'rewards/repeat_penalty/std': 0.08781477808952332, 'rewards/repeat_floor/mean': 0.9906822443008423, 'rewards/repeat_floor/std': 0.0062971110455691814, 'rewards/near_duplicate_penalty/mean': 0.9672962427139282, 'rewards/near_duplicate_penalty/std': 0.007038149982690811, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9509758353233337, 'rewards/distinct_2/std': 0.06346998363733292, 'rewards/total_composite/mean': 0.2319379448890686, 'rewards/total_composite/std': 0.06914785504341125, 'reward': 0.2319379448890686, 'reward_std': 0.06914784759283066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2435443103313446, 'sampling/sampling_logp_difference/max': 3.416959762573242, 'sampling/importance_sampling_ratio/min': 0.03281204029917717, 'sampling/importance_sampling_ratio/mean': 1.0153148174285889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4762604981660843, 'clip_ratio/low_mean': 0.05055607669055462, 'clip_ratio/low_min': 0.05055607669055462, 'clip_ratio/high_mean': 0.153734028339386, 'clip_ratio/high_max': 0.153734028339386, 'clip_ratio/region_mean': 0.2042901050299406, 'reward_total_mean': 0.2319379448890686, 'reward_meter_mean': 0.7575584650039673, 'reward_meter_std': 0.12200148403644562, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9357142448425293, 'reward_count_floor_std': 0.02290808968245983, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9072020053863525, 'reward_repeat_penalty_std': 0.08781477808952332, 'reward_repeat_floor_mean': 0.9906822443008423, 'reward_repeat_floor_std': 0.0062971110455691814, 'reward_near_duplicate_penalty_mean': 0.9672962427139282, 'reward_near_duplicate_penalty_std': 0.007038149982690811, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9509758353233337, 'reward_distinct_2_std': 0.06346998363733292, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.2319379448890686, 'reward_total_composite_std': 0.06914785504341125, 'epoch': 0.02}

  9%|▉         | 311/3300 [1:07:28<8:40:11, 10.44s/it]INFO 04-12 17:05:54 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 312/3300 [1:07:37<8:23:15, 10.11s/it]2026-04-12 17:06:03,628 | INFO | train_grpo_train | metrics_logged mode=train step=312

                                                      
{'loss': 0.0475, 'grad_norm': 15.920761108398438, 'learning_rate': 9.057575757575759e-06, 'num_tokens': 586386.0, 'completions/mean_length': 85.375, 'completions/min_length': 74.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.41423481702804565, 'rewards/meter/std': 0.297497034072876, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.9562499523162842, 'rewards/count_floor/std': 0.01767767407000065, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.20364098250865936, 'rewards/repeat_penalty/mean': 0.9256150126457214, 'rewards/repeat_penalty/std': 0.17002542316913605, 'rewards/repeat_floor/mean': 0.9923303127288818, 'rewards/repeat_floor/std': 0.015654226765036583, 'rewards/near_duplicate_penalty/mean': 0.9803370833396912, 'rewards/near_duplicate_penalty/std': 0.016884636133909225, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9475524425506592, 'rewards/distinct_2/std': 0.14834408462047577, 'rewards/total_composite/mean': 0.20512068271636963, 'rewards/total_composite/std': 0.18300987780094147, 'reward': 0.20512068271636963, 'reward_std': 0.18300987780094147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21955280005931854, 'sampling/sampling_logp_difference/max': 2.5457582473754883, 'sampling/importance_sampling_ratio/min': 0.07841357588768005, 'sampling/importance_sampling_ratio/mean': 1.0039602518081665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.186564490199089, 'clip_ratio/low_mean': 0.13033979572355747, 'clip_ratio/low_min': 0.13033979572355747, 'clip_ratio/high_mean': 0.061771913431584835, 'clip_ratio/high_max': 0.061771913431584835, 'clip_ratio/region_mean': 0.1921117091551423, 'reward_total_mean': 0.20512068271636963, 'reward_meter_mean': 0.41423481702804565, 'reward_meter_std': 0.297497034072876, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.9562499523162842, 'reward_count_floor_std': 0.01767767407000065, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9256150126457214, 'reward_repeat_penalty_std': 0.17002542316913605, 'reward_repeat_floor_mean': 0.9923303127288818, 'reward_repeat_floor_std': 0.015654226765036583, 'reward_near_duplicate_penalty_mean': 0.9803370833396912, 'reward_near_duplicate_penalty_std': 0.016884636133909225, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9475524425506592, 'reward_distinct_2_std': 0.14834408462047577, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.20364098250865936, 'reward_total_composite_mean': 0.20512068271636963, 'reward_total_composite_std': 0.18300987780094147, 'epoch': 0.02}

  9%|▉         | 312/3300 [1:07:37<8:23:15, 10.11s/it]INFO 04-12 17:06:03 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 313/3300 [1:07:45<7:50:27,  9.45s/it]2026-04-12 17:06:11,543 | INFO | train_grpo_train | metrics_logged mode=train step=313

                                                      
{'loss': -0.013, 'grad_norm': 27.450632095336914, 'learning_rate': 9.054545454545455e-06, 'num_tokens': 587671.0, 'completions/mean_length': 15.625, 'completions/min_length': 10.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.625, 'completions/min_terminated_length': 10.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.5798760652542114, 'rewards/meter/std': 0.4535468816757202, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9246323108673096, 'rewards/lexical_plausibility/std': 0.14093062281608582, 'rewards/judge_quality/mean': 0.6687500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38387513160705566, 'rewards/total_composite/std': 0.3713848888874054, 'reward': 0.38387513160705566, 'reward_std': 0.371384859085083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14199897646903992, 'sampling/sampling_logp_difference/max': 1.3993990421295166, 'sampling/importance_sampling_ratio/min': 0.24674519896507263, 'sampling/importance_sampling_ratio/mean': 0.9715681672096252, 'sampling/importance_sampling_ratio/max': 1.72029447555542, 'entropy': 1.186026468873024, 'clip_ratio/low_mean': 0.041165413334965706, 'clip_ratio/low_min': 0.041165413334965706, 'clip_ratio/high_mean': 0.09990809019654989, 'clip_ratio/high_max': 0.09990809019654989, 'clip_ratio/region_mean': 0.1410735035315156, 'reward_total_mean': 0.38387513160705566, 'reward_meter_mean': 0.5798760652542114, 'reward_meter_std': 0.4535468816757202, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9246323108673096, 'reward_lexical_plausibility_std': 0.14093062281608582, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6687500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.38387513160705566, 'reward_total_composite_std': 0.3713848888874054, 'epoch': 0.02}

  9%|▉         | 313/3300 [1:07:45<7:50:27,  9.45s/it]INFO 04-12 17:06:11 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 314/3300 [1:07:53<7:26:41,  8.98s/it]2026-04-12 17:06:19,412 | INFO | train_grpo_train | metrics_logged mode=train step=314

                                                      
{'loss': 0.0362, 'grad_norm': 16.598642349243164, 'learning_rate': 9.051515151515152e-06, 'num_tokens': 589168.0, 'completions/mean_length': 38.125, 'completions/min_length': 27.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6301953792572021, 'rewards/meter/std': 0.41348692774772644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.20364098250865936, 'rewards/repeat_penalty/mean': 0.9459177851676941, 'rewards/repeat_penalty/std': 0.0359596312046051, 'rewards/repeat_floor/mean': 0.9924371242523193, 'rewards/repeat_floor/std': 0.005281753372400999, 'rewards/near_duplicate_penalty/mean': 0.9550752639770508, 'rewards/near_duplicate_penalty/std': 0.039529163390398026, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.3598251938819885, 'rewards/total_composite/std': 0.28503474593162537, 'reward': 0.3598251938819885, 'reward_std': 0.28503474593162537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18062487244606018, 'sampling/sampling_logp_difference/max': 1.8008413314819336, 'sampling/importance_sampling_ratio/min': 0.16515988111495972, 'sampling/importance_sampling_ratio/mean': 1.0134027004241943, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2614011317491531, 'clip_ratio/low_mean': 0.07839708961546421, 'clip_ratio/low_min': 0.07839708961546421, 'clip_ratio/high_mean': 0.08655215473845601, 'clip_ratio/high_max': 0.08655215473845601, 'clip_ratio/region_mean': 0.16494924435392022, 'reward_total_mean': 0.3598251938819885, 'reward_meter_mean': 0.6301953792572021, 'reward_meter_std': 0.41348692774772644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9459177851676941, 'reward_repeat_penalty_std': 0.0359596312046051, 'reward_repeat_floor_mean': 0.9924371242523193, 'reward_repeat_floor_std': 0.005281753372400999, 'reward_near_duplicate_penalty_mean': 0.9550752639770508, 'reward_near_duplicate_penalty_std': 0.039529163390398026, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.20364098250865936, 'reward_total_composite_mean': 0.3598251938819885, 'reward_total_composite_std': 0.28503474593162537, 'epoch': 0.02}

 10%|▉         | 314/3300 [1:07:53<7:26:41,  8.98s/it]INFO 04-12 17:06:19 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 315/3300 [1:08:02<7:37:30,  9.20s/it]2026-04-12 17:06:29,122 | INFO | train_grpo_train | metrics_logged mode=train step=315

                                                      
{'loss': 0.0764, 'grad_norm': 12.955530166625977, 'learning_rate': 9.04848484848485e-06, 'num_tokens': 591191.0, 'completions/mean_length': 70.875, 'completions/min_length': 63.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.5339852571487427, 'rewards/meter/std': 0.3451843857765198, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9940476417541504, 'rewards/lexical_plausibility/std': 0.016835875809192657, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.13024701178073883, 'rewards/repeat_penalty/mean': 0.9954635500907898, 'rewards/repeat_penalty/std': 0.005026623606681824, 'rewards/repeat_floor/mean': 0.9993195533752441, 'rewards/repeat_floor/std': 0.0007539888611063361, 'rewards/near_duplicate_penalty/mean': 0.9954635500907898, 'rewards/near_duplicate_penalty/std': 0.005026623606681824, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16002997756004333, 'rewards/total_composite/std': 0.13243502378463745, 'reward': 0.16002997756004333, 'reward_std': 0.13243502378463745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24005933105945587, 'sampling/sampling_logp_difference/max': 2.231555938720703, 'sampling/importance_sampling_ratio/min': 0.10736124962568283, 'sampling/importance_sampling_ratio/mean': 1.0492222309112549, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.608979359269142, 'clip_ratio/low_mean': 0.16440259478986263, 'clip_ratio/low_min': 0.16440259478986263, 'clip_ratio/high_mean': 0.051913054659962654, 'clip_ratio/high_max': 0.051913054659962654, 'clip_ratio/region_mean': 0.2163156494498253, 'reward_total_mean': 0.16002997756004333, 'reward_meter_mean': 0.5339852571487427, 'reward_meter_std': 0.3451843857765198, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9940476417541504, 'reward_lexical_plausibility_std': 0.016835875809192657, 'reward_repeat_penalty_mean': 0.9954635500907898, 'reward_repeat_penalty_std': 0.005026623606681824, 'reward_repeat_floor_mean': 0.9993195533752441, 'reward_repeat_floor_std': 0.0007539888611063361, 'reward_near_duplicate_penalty_mean': 0.9954635500907898, 'reward_near_duplicate_penalty_std': 0.005026623606681824, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.13024701178073883, 'reward_total_composite_mean': 0.16002997756004333, 'reward_total_composite_std': 0.13243502378463745, 'epoch': 0.02}

 10%|▉         | 315/3300 [1:08:02<7:37:30,  9.20s/it]INFO 04-12 17:06:29 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 316/3300 [1:08:12<7:46:40,  9.38s/it]2026-04-12 17:06:38,944 | INFO | train_grpo_train | metrics_logged mode=train step=316

                                                      
{'loss': 0.0369, 'grad_norm': 14.941880226135254, 'learning_rate': 9.045454545454546e-06, 'num_tokens': 593621.0, 'completions/mean_length': 116.75, 'completions/min_length': 102.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.75, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.6072092652320862, 'rewards/meter/std': 0.1961025595664978, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9142857193946838, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9895833730697632, 'rewards/lexical_plausibility/std': 0.029462775215506554, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.8918505907058716, 'rewards/repeat_penalty/std': 0.11732038855552673, 'rewards/repeat_floor/mean': 0.9889510273933411, 'rewards/repeat_floor/std': 0.011997546069324017, 'rewards/near_duplicate_penalty/mean': 0.9743615388870239, 'rewards/near_duplicate_penalty/std': 0.023420952260494232, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9199645519256592, 'rewards/distinct_2/std': 0.09881288558244705, 'rewards/total_composite/mean': 0.2331920862197876, 'rewards/total_composite/std': 0.15198205411434174, 'reward': 0.2331920862197876, 'reward_std': 0.15198205411434174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23752768337726593, 'sampling/sampling_logp_difference/max': 2.125997543334961, 'sampling/importance_sampling_ratio/min': 0.11931388080120087, 'sampling/importance_sampling_ratio/mean': 1.0102297067642212, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.48250912129879, 'clip_ratio/low_mean': 0.09028875920921564, 'clip_ratio/low_min': 0.09028875920921564, 'clip_ratio/high_mean': 0.09031237103044987, 'clip_ratio/high_max': 0.09031237103044987, 'clip_ratio/region_mean': 0.1806011302396655, 'reward_total_mean': 0.2331920862197876, 'reward_meter_mean': 0.6072092652320862, 'reward_meter_std': 0.1961025595664978, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9142857193946838, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9895833730697632, 'reward_lexical_plausibility_std': 0.029462775215506554, 'reward_repeat_penalty_mean': 0.8918505907058716, 'reward_repeat_penalty_std': 0.11732038855552673, 'reward_repeat_floor_mean': 0.9889510273933411, 'reward_repeat_floor_std': 0.011997546069324017, 'reward_near_duplicate_penalty_mean': 0.9743615388870239, 'reward_near_duplicate_penalty_std': 0.023420952260494232, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9199645519256592, 'reward_distinct_2_std': 0.09881288558244705, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.2331920862197876, 'reward_total_composite_std': 0.15198205411434174, 'epoch': 0.02}

 10%|▉         | 316/3300 [1:08:12<7:46:40,  9.38s/it]INFO 04-12 17:06:39 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 317/3300 [1:08:25<8:41:27, 10.49s/it]2026-04-12 17:06:52,017 | INFO | train_grpo_train | metrics_logged mode=train step=317

                                                      
{'loss': -0.0685, 'grad_norm': 7.577369689941406, 'learning_rate': 9.042424242424244e-06, 'num_tokens': 595465.0, 'completions/mean_length': 109.5, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.000003814697266, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.42767134308815, 'rewards/meter/std': 0.3034510612487793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9421781897544861, 'rewards/lexical_plausibility/std': 0.16354475915431976, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.9288090467453003, 'rewards/repeat_penalty/std': 0.06002446636557579, 'rewards/repeat_floor/mean': 0.9925997257232666, 'rewards/repeat_floor/std': 0.0061640311032533646, 'rewards/near_duplicate_penalty/mean': 0.9852910041809082, 'rewards/near_duplicate_penalty/std': 0.013444710522890091, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9422902464866638, 'rewards/distinct_2/std': 0.05239473283290863, 'rewards/total_composite/mean': 0.26846200227737427, 'rewards/total_composite/std': 0.31822338700294495, 'reward': 0.26846200227737427, 'reward_std': 0.31822338700294495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1445298194885254, 'sampling/sampling_logp_difference/max': 1.5252630710601807, 'sampling/importance_sampling_ratio/min': 0.21756382286548615, 'sampling/importance_sampling_ratio/mean': 1.0262278318405151, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9528652653098106, 'clip_ratio/low_mean': 0.09752868209034204, 'clip_ratio/low_min': 0.09752868209034204, 'clip_ratio/high_mean': 0.03143939469009638, 'clip_ratio/high_max': 0.03143939469009638, 'clip_ratio/region_mean': 0.12896807678043842, 'reward_total_mean': 0.26846200227737427, 'reward_meter_mean': 0.42767134308815, 'reward_meter_std': 0.3034510612487793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9421781897544861, 'reward_lexical_plausibility_std': 0.16354475915431976, 'reward_repeat_penalty_mean': 0.9288090467453003, 'reward_repeat_penalty_std': 0.06002446636557579, 'reward_repeat_floor_mean': 0.9925997257232666, 'reward_repeat_floor_std': 0.0061640311032533646, 'reward_near_duplicate_penalty_mean': 0.9852910041809082, 'reward_near_duplicate_penalty_std': 0.013444710522890091, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9422902464866638, 'reward_distinct_2_std': 0.05239473283290863, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.26846200227737427, 'reward_total_composite_std': 0.31822338700294495, 'epoch': 0.02}

 10%|▉         | 317/3300 [1:08:25<8:41:27, 10.49s/it]INFO 04-12 17:06:52 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 318/3300 [1:08:35<8:21:39, 10.09s/it]2026-04-12 17:07:01,186 | INFO | train_grpo_train | metrics_logged mode=train step=318

                                                      
{'loss': -0.0155, 'grad_norm': 11.436017990112305, 'learning_rate': 9.03939393939394e-06, 'num_tokens': 597615.0, 'completions/mean_length': 82.75, 'completions/min_length': 66.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.2961556315422058, 'rewards/meter/std': 0.32202163338661194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4000000059604645, 'rewards/judge_quality/std': 0.22038927674293518, 'rewards/repeat_penalty/mean': 0.9850220680236816, 'rewards/repeat_penalty/std': 0.013130299746990204, 'rewards/repeat_floor/mean': 0.9980374574661255, 'rewards/repeat_floor/std': 0.0012959600426256657, 'rewards/near_duplicate_penalty/mean': 0.9898011088371277, 'rewards/near_duplicate_penalty/std': 0.006006290670484304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.995192289352417, 'rewards/distinct_2/std': 0.013598216697573662, 'rewards/total_composite/mean': 0.1574200987815857, 'rewards/total_composite/std': 0.21431198716163635, 'reward': 0.1574200987815857, 'reward_std': 0.21431197226047516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20027244091033936, 'sampling/sampling_logp_difference/max': 1.806778907775879, 'sampling/importance_sampling_ratio/min': 0.1641821265220642, 'sampling/importance_sampling_ratio/mean': 1.0516811609268188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7966539114713669, 'clip_ratio/low_mean': 0.10688901133835316, 'clip_ratio/low_min': 0.10688901133835316, 'clip_ratio/high_mean': 0.03520032297819853, 'clip_ratio/high_max': 0.03520032297819853, 'clip_ratio/region_mean': 0.14208933431655169, 'reward_total_mean': 0.1574200987815857, 'reward_meter_mean': 0.2961556315422058, 'reward_meter_std': 0.32202163338661194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9850220680236816, 'reward_repeat_penalty_std': 0.013130299746990204, 'reward_repeat_floor_mean': 0.9980374574661255, 'reward_repeat_floor_std': 0.0012959600426256657, 'reward_near_duplicate_penalty_mean': 0.9898011088371277, 'reward_near_duplicate_penalty_std': 0.006006290670484304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.995192289352417, 'reward_distinct_2_std': 0.013598216697573662, 'reward_judge_quality_mean': 0.4000000059604645, 'reward_judge_quality_std': 0.22038927674293518, 'reward_total_composite_mean': 0.1574200987815857, 'reward_total_composite_std': 0.21431198716163635, 'epoch': 0.02}

 10%|▉         | 318/3300 [1:08:35<8:21:39, 10.09s/it]INFO 04-12 17:07:01 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 319/3300 [1:08:43<7:53:50,  9.54s/it]2026-04-12 17:07:09,423 | INFO | train_grpo_train | metrics_logged mode=train step=319

                                                      
{'loss': 0.0669, 'grad_norm': 30.49716567993164, 'learning_rate': 9.036363636363638e-06, 'num_tokens': 598934.0, 'completions/mean_length': 18.875, 'completions/min_length': 16.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7936312556266785, 'rewards/meter/std': 0.3362228274345398, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5063081383705139, 'rewards/total_composite/std': 0.2949349880218506, 'reward': 0.5063081383705139, 'reward_std': 0.2949349880218506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2095964252948761, 'sampling/sampling_logp_difference/max': 1.1650099754333496, 'sampling/importance_sampling_ratio/min': 0.31191954016685486, 'sampling/importance_sampling_ratio/mean': 1.0343742370605469, 'sampling/importance_sampling_ratio/max': 1.9709964990615845, 'entropy': 1.7813816964626312, 'clip_ratio/low_mean': 0.09983944287523627, 'clip_ratio/low_min': 0.09983944287523627, 'clip_ratio/high_mean': 0.07778637856245041, 'clip_ratio/high_max': 0.07778637856245041, 'clip_ratio/region_mean': 0.17762582143768668, 'reward_total_mean': 0.5063081383705139, 'reward_meter_mean': 0.7936312556266785, 'reward_meter_std': 0.3362228274345398, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.5063081383705139, 'reward_total_composite_std': 0.2949349880218506, 'epoch': 0.02}

 10%|▉         | 319/3300 [1:08:43<7:53:50,  9.54s/it]INFO 04-12 17:07:09 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 320/3300 [1:08:51<7:41:17,  9.29s/it]2026-04-12 17:07:18,145 | INFO | train_grpo_train | metrics_logged mode=train step=320

                                                      
{'loss': -0.0525, 'grad_norm': 13.675727844238281, 'learning_rate': 9.033333333333334e-06, 'num_tokens': 600460.0, 'completions/mean_length': 36.75, 'completions/min_length': 26.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7165847420692444, 'rewards/meter/std': 0.35428208112716675, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9961137771606445, 'rewards/repeat_penalty/std': 0.010145732201635838, 'rewards/repeat_floor/mean': 0.9994170665740967, 'rewards/repeat_floor/std': 0.0015218672342598438, 'rewards/near_duplicate_penalty/mean': 0.9961137771606445, 'rewards/near_duplicate_penalty/std': 0.010145732201635838, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3932347297668457, 'rewards/total_composite/std': 0.3415294289588928, 'reward': 0.3932347297668457, 'reward_std': 0.34152939915657043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17731329798698425, 'sampling/sampling_logp_difference/max': 1.387129783630371, 'sampling/importance_sampling_ratio/min': 0.2497912496328354, 'sampling/importance_sampling_ratio/mean': 1.0123792886734009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4262383878231049, 'clip_ratio/low_mean': 0.11775892227888107, 'clip_ratio/low_min': 0.11775892227888107, 'clip_ratio/high_mean': 0.06192527338862419, 'clip_ratio/high_max': 0.06192527338862419, 'clip_ratio/region_mean': 0.17968419566750526, 'reward_total_mean': 0.3932347297668457, 'reward_meter_mean': 0.7165847420692444, 'reward_meter_std': 0.35428208112716675, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9961137771606445, 'reward_repeat_penalty_std': 0.010145732201635838, 'reward_repeat_floor_mean': 0.9994170665740967, 'reward_repeat_floor_std': 0.0015218672342598438, 'reward_near_duplicate_penalty_mean': 0.9961137771606445, 'reward_near_duplicate_penalty_std': 0.010145732201635838, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.3932347297668457, 'reward_total_composite_std': 0.3415294289588928, 'epoch': 0.02}

 10%|▉         | 320/3300 [1:08:52<7:41:17,  9.29s/it]INFO 04-12 17:07:18 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 321/3300 [1:09:07<9:10:18, 11.08s/it]2026-04-12 17:07:33,404 | INFO | train_grpo_train | metrics_logged mode=train step=321

                                                      
{'loss': -0.04, 'grad_norm': 5.024820804595947, 'learning_rate': 9.030303030303031e-06, 'num_tokens': 601934.0, 'completions/mean_length': 90.25, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 30.000001907348633, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5445935130119324, 'rewards/meter/std': 0.3765180706977844, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9750847220420837, 'rewards/lexical_plausibility/std': 0.060039129108190536, 'rewards/judge_quality/mean': 0.5600000619888306, 'rewards/judge_quality/std': 0.4012124836444855, 'rewards/repeat_penalty/mean': 0.9535180330276489, 'rewards/repeat_penalty/std': 0.08469787240028381, 'rewards/repeat_floor/mean': 0.9960237145423889, 'rewards/repeat_floor/std': 0.00515634985640645, 'rewards/near_duplicate_penalty/mean': 0.9880056381225586, 'rewards/near_duplicate_penalty/std': 0.013819819316267967, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9966430068016052, 'rewards/distinct_2/std': 0.009495011530816555, 'rewards/total_composite/mean': 0.27255403995513916, 'rewards/total_composite/std': 0.2881576120853424, 'reward': 0.27255403995513916, 'reward_std': 0.2881576120853424, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17371590435504913, 'sampling/sampling_logp_difference/max': 2.350584030151367, 'sampling/importance_sampling_ratio/min': 0.09531348198652267, 'sampling/importance_sampling_ratio/mean': 1.026108980178833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1391584053635597, 'clip_ratio/low_mean': 0.09603851195424795, 'clip_ratio/low_min': 0.09603851195424795, 'clip_ratio/high_mean': 0.03629032149910927, 'clip_ratio/high_max': 0.03629032149910927, 'clip_ratio/region_mean': 0.13232883345335722, 'reward_total_mean': 0.27255403995513916, 'reward_meter_mean': 0.5445935130119324, 'reward_meter_std': 0.3765180706977844, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9750847220420837, 'reward_lexical_plausibility_std': 0.060039129108190536, 'reward_repeat_penalty_mean': 0.9535180330276489, 'reward_repeat_penalty_std': 0.08469787240028381, 'reward_repeat_floor_mean': 0.9960237145423889, 'reward_repeat_floor_std': 0.00515634985640645, 'reward_near_duplicate_penalty_mean': 0.9880056381225586, 'reward_near_duplicate_penalty_std': 0.013819819316267967, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9966430068016052, 'reward_distinct_2_std': 0.009495011530816555, 'reward_judge_quality_mean': 0.5600000619888306, 'reward_judge_quality_std': 0.4012124836444855, 'reward_total_composite_mean': 0.27255403995513916, 'reward_total_composite_std': 0.2881576120853424, 'epoch': 0.02}

 10%|▉         | 321/3300 [1:09:07<9:10:18, 11.08s/it]INFO 04-12 17:07:33 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 322/3300 [1:09:21<9:50:37, 11.90s/it]2026-04-12 17:07:47,207 | INFO | train_grpo_train | metrics_logged mode=train step=322

                                                      
{'loss': -0.1657, 'grad_norm': 2.980701208114624, 'learning_rate': 9.027272727272728e-06, 'num_tokens': 603788.0, 'completions/mean_length': 194.75, 'completions/min_length': 75.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.7815793752670288, 'rewards/meter/std': 0.33709579706192017, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.08017838001251221, 'rewards/lexical_plausibility/mean': 0.9196681976318359, 'rewards/lexical_plausibility/std': 0.1511790156364441, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.9904124736785889, 'rewards/repeat_penalty/std': 0.013655727729201317, 'rewards/repeat_floor/mean': 0.9988014698028564, 'rewards/repeat_floor/std': 0.0014115042285993695, 'rewards/near_duplicate_penalty/mean': 0.9944649934768677, 'rewards/near_duplicate_penalty/std': 0.004200012423098087, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9959083795547485, 'rewards/distinct_2/std': 0.011572934687137604, 'rewards/total_composite/mean': 0.2697392702102661, 'rewards/total_composite/std': 0.1861606240272522, 'reward': 0.2697392702102661, 'reward_std': 0.1861606240272522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2009381651878357, 'sampling/sampling_logp_difference/max': 1.1313667297363281, 'sampling/importance_sampling_ratio/min': 0.3225920796394348, 'sampling/importance_sampling_ratio/mean': 1.0548303127288818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1985997557640076, 'clip_ratio/low_mean': 0.02777777798473835, 'clip_ratio/low_min': 0.02777777798473835, 'clip_ratio/high_mean': 0.12069774232804775, 'clip_ratio/high_max': 0.12069774232804775, 'clip_ratio/region_mean': 0.1484755203127861, 'reward_total_mean': 0.2697392702102661, 'reward_meter_mean': 0.7815793752670288, 'reward_meter_std': 0.33709579706192017, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.08017838001251221, 'reward_lexical_plausibility_mean': 0.9196681976318359, 'reward_lexical_plausibility_std': 0.1511790156364441, 'reward_repeat_penalty_mean': 0.9904124736785889, 'reward_repeat_penalty_std': 0.013655727729201317, 'reward_repeat_floor_mean': 0.9988014698028564, 'reward_repeat_floor_std': 0.0014115042285993695, 'reward_near_duplicate_penalty_mean': 0.9944649934768677, 'reward_near_duplicate_penalty_std': 0.004200012423098087, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9959083795547485, 'reward_distinct_2_std': 0.011572934687137604, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.2697392702102661, 'reward_total_composite_std': 0.1861606240272522, 'epoch': 0.02}

 10%|▉         | 322/3300 [1:09:21<9:50:37, 11.90s/it]INFO 04-12 17:07:47 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 323/3300 [1:09:29<9:03:27, 10.95s/it]2026-04-12 17:07:55,953 | INFO | train_grpo_train | metrics_logged mode=train step=323

                                                      
{'loss': 0.0405, 'grad_norm': 16.94824981689453, 'learning_rate': 9.024242424242426e-06, 'num_tokens': 605687.0, 'completions/mean_length': 44.375, 'completions/min_length': 38.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8522233963012695, 'rewards/meter/std': 0.18364009261131287, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.897731602191925, 'rewards/repeat_penalty/std': 0.18885688483715057, 'rewards/repeat_floor/mean': 0.9878875017166138, 'rewards/repeat_floor/std': 0.022107074037194252, 'rewards/near_duplicate_penalty/mean': 0.9626564979553223, 'rewards/near_duplicate_penalty/std': 0.06179862096905708, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9484890103340149, 'rewards/distinct_2/std': 0.09566201269626617, 'rewards/total_composite/mean': 0.3616304099559784, 'rewards/total_composite/std': 0.08909326046705246, 'reward': 0.3616304099559784, 'reward_std': 0.08909326046705246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24145522713661194, 'sampling/sampling_logp_difference/max': 1.7497434616088867, 'sampling/importance_sampling_ratio/min': 0.17381852865219116, 'sampling/importance_sampling_ratio/mean': 1.0306921005249023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.072322279214859, 'clip_ratio/low_mean': 0.07108516525477171, 'clip_ratio/low_min': 0.07108516525477171, 'clip_ratio/high_mean': 0.1286225551739335, 'clip_ratio/high_max': 0.1286225551739335, 'clip_ratio/region_mean': 0.19970772042870522, 'reward_total_mean': 0.3616304099559784, 'reward_meter_mean': 0.8522233963012695, 'reward_meter_std': 0.18364009261131287, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.897731602191925, 'reward_repeat_penalty_std': 0.18885688483715057, 'reward_repeat_floor_mean': 0.9878875017166138, 'reward_repeat_floor_std': 0.022107074037194252, 'reward_near_duplicate_penalty_mean': 0.9626564979553223, 'reward_near_duplicate_penalty_std': 0.06179862096905708, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9484890103340149, 'reward_distinct_2_std': 0.09566201269626617, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.3616304099559784, 'reward_total_composite_std': 0.08909326046705246, 'epoch': 0.02}

 10%|▉         | 323/3300 [1:09:29<9:03:27, 10.95s/it]INFO 04-12 17:07:56 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 324/3300 [1:09:37<8:14:00,  9.96s/it]2026-04-12 17:08:03,594 | INFO | train_grpo_train | metrics_logged mode=train step=324

                                                      
{'loss': 0.0265, 'grad_norm': 19.082670211791992, 'learning_rate': 9.021212121212121e-06, 'num_tokens': 607056.0, 'completions/mean_length': 22.125, 'completions/min_length': 17.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6404979825019836, 'rewards/meter/std': 0.43355244398117065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.6094648241996765, 'rewards/repeat_penalty/std': 0.12628398835659027, 'rewards/repeat_floor/mean': 0.9569916725158691, 'rewards/repeat_floor/std': 0.025063540786504745, 'rewards/near_duplicate_penalty/mean': 0.8260983228683472, 'rewards/near_duplicate_penalty/std': 0.14450305700302124, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.2601917088031769, 'rewards/total_composite/std': 0.19473481178283691, 'reward': 0.2601917088031769, 'reward_std': 0.19473479688167572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16124628484249115, 'sampling/sampling_logp_difference/max': 1.5426025390625, 'sampling/importance_sampling_ratio/min': 0.21382389962673187, 'sampling/importance_sampling_ratio/mean': 0.9949864745140076, 'sampling/importance_sampling_ratio/max': 1.9606815576553345, 'entropy': 1.056018553674221, 'clip_ratio/low_mean': 0.07366867363452911, 'clip_ratio/low_min': 0.07366867363452911, 'clip_ratio/high_mean': 0.09975808765739202, 'clip_ratio/high_max': 0.09975808765739202, 'clip_ratio/region_mean': 0.17342676129192114, 'reward_total_mean': 0.2601917088031769, 'reward_meter_mean': 0.6404979825019836, 'reward_meter_std': 0.43355244398117065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.6094648241996765, 'reward_repeat_penalty_std': 0.12628398835659027, 'reward_repeat_floor_mean': 0.9569916725158691, 'reward_repeat_floor_std': 0.025063540786504745, 'reward_near_duplicate_penalty_mean': 0.8260983228683472, 'reward_near_duplicate_penalty_std': 0.14450305700302124, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.2601917088031769, 'reward_total_composite_std': 0.19473481178283691, 'epoch': 0.02}

 10%|▉         | 324/3300 [1:09:37<8:14:00,  9.96s/it]INFO 04-12 17:08:03 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 325/3300 [1:09:50<9:03:36, 10.96s/it]2026-04-12 17:08:16,900 | INFO | train_grpo_train | metrics_logged mode=train step=325

                                                      
{'loss': -0.0799, 'grad_norm': 3.325479030609131, 'learning_rate': 9.01818181818182e-06, 'num_tokens': 608534.0, 'completions/mean_length': 155.75, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.45075857639312744, 'rewards/meter/std': 0.40244653820991516, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.879709005355835, 'rewards/lexical_plausibility/std': 0.1780509501695633, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/repeat_penalty/mean': 0.9158079624176025, 'rewards/repeat_penalty/std': 0.10471590608358383, 'rewards/repeat_floor/mean': 0.9929962158203125, 'rewards/repeat_floor/std': 0.005952437408268452, 'rewards/near_duplicate_penalty/mean': 0.9783079624176025, 'rewards/near_duplicate_penalty/std': 0.025911886245012283, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1340664029121399, 'rewards/total_composite/std': 0.13815456628799438, 'reward': 0.1340664029121399, 'reward_std': 0.13815458118915558, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23885197937488556, 'sampling/sampling_logp_difference/max': 1.6627397537231445, 'sampling/importance_sampling_ratio/min': 0.18961875140666962, 'sampling/importance_sampling_ratio/mean': 1.021572232246399, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7420674115419388, 'clip_ratio/low_mean': 0.04676120541989803, 'clip_ratio/low_min': 0.04676120541989803, 'clip_ratio/high_mean': 0.06735125556588173, 'clip_ratio/high_max': 0.06735125556588173, 'clip_ratio/region_mean': 0.11411246098577976, 'reward_total_mean': 0.1340664029121399, 'reward_meter_mean': 0.45075857639312744, 'reward_meter_std': 0.40244653820991516, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.879709005355835, 'reward_lexical_plausibility_std': 0.1780509501695633, 'reward_repeat_penalty_mean': 0.9158079624176025, 'reward_repeat_penalty_std': 0.10471590608358383, 'reward_repeat_floor_mean': 0.9929962158203125, 'reward_repeat_floor_std': 0.005952437408268452, 'reward_near_duplicate_penalty_mean': 0.9783079624176025, 'reward_near_duplicate_penalty_std': 0.025911886245012283, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.1340664029121399, 'reward_total_composite_std': 0.13815456628799438, 'epoch': 0.02}

 10%|▉         | 325/3300 [1:09:50<9:03:36, 10.96s/it]INFO 04-12 17:08:17 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 326/3300 [1:10:01<9:00:20, 10.90s/it]2026-04-12 17:08:27,656 | INFO | train_grpo_train | metrics_logged mode=train step=326

                                                      
{'loss': 0.0237, 'grad_norm': 13.654400825500488, 'learning_rate': 9.015151515151516e-06, 'num_tokens': 610100.0, 'completions/mean_length': 47.75, 'completions/min_length': 42.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5693294405937195, 'rewards/meter/std': 0.411795049905777, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.991178035736084, 'rewards/repeat_penalty/std': 0.013867231085896492, 'rewards/repeat_floor/mean': 0.9986767172813416, 'rewards/repeat_floor/std': 0.0020800859201699495, 'rewards/near_duplicate_penalty/mean': 0.991178035736084, 'rewards/near_duplicate_penalty/std': 0.013867231085896492, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2887307107448578, 'rewards/total_composite/std': 0.27768048644065857, 'reward': 0.2887307107448578, 'reward_std': 0.2776804566383362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20421560108661652, 'sampling/sampling_logp_difference/max': 1.4174118041992188, 'sampling/importance_sampling_ratio/min': 0.24234043061733246, 'sampling/importance_sampling_ratio/mean': 1.030893325805664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.896958127617836, 'clip_ratio/low_mean': 0.06497407425194979, 'clip_ratio/low_min': 0.06497407425194979, 'clip_ratio/high_mean': 0.08649432007223368, 'clip_ratio/high_max': 0.08649432007223368, 'clip_ratio/region_mean': 0.15146839432418346, 'reward_total_mean': 0.2887307107448578, 'reward_meter_mean': 0.5693294405937195, 'reward_meter_std': 0.411795049905777, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.991178035736084, 'reward_repeat_penalty_std': 0.013867231085896492, 'reward_repeat_floor_mean': 0.9986767172813416, 'reward_repeat_floor_std': 0.0020800859201699495, 'reward_near_duplicate_penalty_mean': 0.991178035736084, 'reward_near_duplicate_penalty_std': 0.013867231085896492, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.2887307107448578, 'reward_total_composite_std': 0.27768048644065857, 'epoch': 0.02}

 10%|▉         | 326/3300 [1:10:01<9:00:20, 10.90s/it]INFO 04-12 17:08:27 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 327/3300 [1:10:15<9:46:03, 11.83s/it]2026-04-12 17:08:41,647 | INFO | train_grpo_train | metrics_logged mode=train step=327

                                                      
{'loss': -0.0724, 'grad_norm': 6.321916103363037, 'learning_rate': 9.012121212121213e-06, 'num_tokens': 611672.0, 'completions/mean_length': 100.5, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.71428680419922, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4724549353122711, 'rewards/meter/std': 0.4008066952228546, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.942983865737915, 'rewards/lexical_plausibility/std': 0.14270249009132385, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.24706491827964783, 'rewards/repeat_penalty/mean': 0.9759707450866699, 'rewards/repeat_penalty/std': 0.03519072011113167, 'rewards/repeat_floor/mean': 0.9968447685241699, 'rewards/repeat_floor/std': 0.004106494132429361, 'rewards/near_duplicate_penalty/mean': 0.9839820861816406, 'rewards/near_duplicate_penalty/std': 0.01587345078587532, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.2145029902458191, 'rewards/total_composite/std': 0.22043871879577637, 'reward': 0.2145029902458191, 'reward_std': 0.22043870389461517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19373220205307007, 'sampling/sampling_logp_difference/max': 1.6452445983886719, 'sampling/importance_sampling_ratio/min': 0.19296535849571228, 'sampling/importance_sampling_ratio/mean': 1.0285770893096924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7601165026426315, 'clip_ratio/low_mean': 0.038460261188447475, 'clip_ratio/low_min': 0.038460261188447475, 'clip_ratio/high_mean': 0.07752238027751446, 'clip_ratio/high_max': 0.07752238027751446, 'clip_ratio/region_mean': 0.11598264146596193, 'reward_total_mean': 0.2145029902458191, 'reward_meter_mean': 0.4724549353122711, 'reward_meter_std': 0.4008066952228546, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.942983865737915, 'reward_lexical_plausibility_std': 0.14270249009132385, 'reward_repeat_penalty_mean': 0.9759707450866699, 'reward_repeat_penalty_std': 0.03519072011113167, 'reward_repeat_floor_mean': 0.9968447685241699, 'reward_repeat_floor_std': 0.004106494132429361, 'reward_near_duplicate_penalty_mean': 0.9839820861816406, 'reward_near_duplicate_penalty_std': 0.01587345078587532, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.24706491827964783, 'reward_total_composite_mean': 0.2145029902458191, 'reward_total_composite_std': 0.22043871879577637, 'epoch': 0.02}

 10%|▉         | 327/3300 [1:10:15<9:46:03, 11.83s/it]INFO 04-12 17:08:41 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 328/3300 [1:10:29<10:16:27, 12.45s/it]2026-04-12 17:08:55,532 | INFO | train_grpo_train | metrics_logged mode=train step=328

                                                       
{'loss': -0.0714, 'grad_norm': 3.29329252243042, 'learning_rate': 9.00909090909091e-06, 'num_tokens': 613329.0, 'completions/mean_length': 162.125, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5958593487739563, 'rewards/meter/std': 0.4086369574069977, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8750607967376709, 'rewards/lexical_plausibility/std': 0.16764959692955017, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9930334091186523, 'rewards/repeat_penalty/std': 0.013267184607684612, 'rewards/repeat_floor/mean': 0.9989550113677979, 'rewards/repeat_floor/std': 0.0019900870975106955, 'rewards/near_duplicate_penalty/mean': 0.9930334091186523, 'rewards/near_duplicate_penalty/std': 0.013267184607684612, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19685010612010956, 'rewards/total_composite/std': 0.17298580706119537, 'reward': 0.19685010612010956, 'reward_std': 0.17298579216003418, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2268337905406952, 'sampling/sampling_logp_difference/max': 1.304356575012207, 'sampling/importance_sampling_ratio/min': 0.2713470757007599, 'sampling/importance_sampling_ratio/mean': 1.0349611043930054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3360288888216019, 'clip_ratio/low_mean': 0.057386364787817, 'clip_ratio/low_min': 0.057386364787817, 'clip_ratio/high_mean': 0.10669773630797863, 'clip_ratio/high_max': 0.10669773630797863, 'clip_ratio/region_mean': 0.16408410109579563, 'reward_total_mean': 0.19685010612010956, 'reward_meter_mean': 0.5958593487739563, 'reward_meter_std': 0.4086369574069977, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8750607967376709, 'reward_lexical_plausibility_std': 0.16764959692955017, 'reward_repeat_penalty_mean': 0.9930334091186523, 'reward_repeat_penalty_std': 0.013267184607684612, 'reward_repeat_floor_mean': 0.9989550113677979, 'reward_repeat_floor_std': 0.0019900870975106955, 'reward_near_duplicate_penalty_mean': 0.9930334091186523, 'reward_near_duplicate_penalty_std': 0.013267184607684612, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.19685010612010956, 'reward_total_composite_std': 0.17298580706119537, 'epoch': 0.02}

 10%|▉         | 328/3300 [1:10:29<10:16:27, 12.45s/it]INFO 04-12 17:08:55 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 329/3300 [1:10:37<9:16:28, 11.24s/it] 2026-04-12 17:09:03,952 | INFO | train_grpo_train | metrics_logged mode=train step=329

                                                      
{'loss': 0.0155, 'grad_norm': 12.86247730255127, 'learning_rate': 9.006060606060607e-06, 'num_tokens': 614868.0, 'completions/mean_length': 39.375, 'completions/min_length': 31.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5450085401535034, 'rewards/meter/std': 0.4167117476463318, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6337499618530273, 'rewards/judge_quality/std': 0.3177122473716736, 'rewards/repeat_penalty/mean': 0.9742398858070374, 'rewards/repeat_penalty/std': 0.06834931671619415, 'rewards/repeat_floor/mean': 0.9966607093811035, 'rewards/repeat_floor/std': 0.008771215565502644, 'rewards/near_duplicate_penalty/mean': 0.9849497079849243, 'rewards/near_duplicate_penalty/std': 0.038151029497385025, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9879807829856873, 'rewards/distinct_2/std': 0.03399552404880524, 'rewards/total_composite/mean': 0.3427734971046448, 'rewards/total_composite/std': 0.361762136220932, 'reward': 0.3427734971046448, 'reward_std': 0.3617621064186096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1714494675397873, 'sampling/sampling_logp_difference/max': 1.2371959686279297, 'sampling/importance_sampling_ratio/min': 0.29019680619239807, 'sampling/importance_sampling_ratio/mean': 1.01372230052948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5177952349185944, 'clip_ratio/low_mean': 0.07150704646483064, 'clip_ratio/low_min': 0.07150704646483064, 'clip_ratio/high_mean': 0.06821303069591522, 'clip_ratio/high_max': 0.06821303069591522, 'clip_ratio/region_mean': 0.13972007716074586, 'reward_total_mean': 0.3427734971046448, 'reward_meter_mean': 0.5450085401535034, 'reward_meter_std': 0.4167117476463318, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9742398858070374, 'reward_repeat_penalty_std': 0.06834931671619415, 'reward_repeat_floor_mean': 0.9966607093811035, 'reward_repeat_floor_std': 0.008771215565502644, 'reward_near_duplicate_penalty_mean': 0.9849497079849243, 'reward_near_duplicate_penalty_std': 0.038151029497385025, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9879807829856873, 'reward_distinct_2_std': 0.03399552404880524, 'reward_judge_quality_mean': 0.6337499618530273, 'reward_judge_quality_std': 0.3177122473716736, 'reward_total_composite_mean': 0.3427734971046448, 'reward_total_composite_std': 0.361762136220932, 'epoch': 0.02}

 10%|▉         | 329/3300 [1:10:37<9:16:28, 11.24s/it]INFO 04-12 17:09:04 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 330/3300 [1:10:48<9:06:13, 11.03s/it]2026-04-12 17:09:14,512 | INFO | train_grpo_train | metrics_logged mode=train step=330

                                                      
{'loss': 0.0333, 'grad_norm': 7.601152420043945, 'learning_rate': 9.003030303030303e-06, 'num_tokens': 617814.0, 'completions/mean_length': 175.25, 'completions/min_length': 148.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 175.25, 'completions/min_terminated_length': 148.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.6623580455780029, 'rewards/meter/std': 0.3066847026348114, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9142857193946838, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.6902416944503784, 'rewards/repeat_penalty/std': 0.2966504991054535, 'rewards/repeat_floor/mean': 0.9670842885971069, 'rewards/repeat_floor/std': 0.03169690817594528, 'rewards/near_duplicate_penalty/mean': 0.9358048439025879, 'rewards/near_duplicate_penalty/std': 0.06196501478552818, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7412613034248352, 'rewards/distinct_2/std': 0.25303134322166443, 'rewards/total_composite/mean': 0.19090114533901215, 'rewards/total_composite/std': 0.11105522513389587, 'reward': 0.19090114533901215, 'reward_std': 0.11105521768331528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18563103675842285, 'sampling/sampling_logp_difference/max': 3.275938034057617, 'sampling/importance_sampling_ratio/min': 0.03778141364455223, 'sampling/importance_sampling_ratio/mean': 1.0296766757965088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8779809772968292, 'clip_ratio/low_mean': 0.05767340213060379, 'clip_ratio/low_min': 0.05767340213060379, 'clip_ratio/high_mean': 0.10809224285185337, 'clip_ratio/high_max': 0.10809224285185337, 'clip_ratio/region_mean': 0.16576564498245716, 'reward_total_mean': 0.19090114533901215, 'reward_meter_mean': 0.6623580455780029, 'reward_meter_std': 0.3066847026348114, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9142857193946838, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6902416944503784, 'reward_repeat_penalty_std': 0.2966504991054535, 'reward_repeat_floor_mean': 0.9670842885971069, 'reward_repeat_floor_std': 0.03169690817594528, 'reward_near_duplicate_penalty_mean': 0.9358048439025879, 'reward_near_duplicate_penalty_std': 0.06196501478552818, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7412613034248352, 'reward_distinct_2_std': 0.25303134322166443, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.19090114533901215, 'reward_total_composite_std': 0.11105522513389587, 'epoch': 0.02}

 10%|█         | 330/3300 [1:10:48<9:06:13, 11.03s/it]INFO 04-12 17:09:14 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 331/3300 [1:10:58<8:54:33, 10.80s/it]2026-04-12 17:09:24,775 | INFO | train_grpo_train | metrics_logged mode=train step=331

                                                      
{'loss': -0.0227, 'grad_norm': 9.477758407592773, 'learning_rate': 9e-06, 'num_tokens': 620345.0, 'completions/mean_length': 119.375, 'completions/min_length': 98.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.375, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.5467900633811951, 'rewards/meter/std': 0.36305561661720276, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.9762307405471802, 'rewards/repeat_penalty/std': 0.04083777964115143, 'rewards/repeat_floor/mean': 0.9972497820854187, 'rewards/repeat_floor/std': 0.004304586909711361, 'rewards/near_duplicate_penalty/mean': 0.9903911352157593, 'rewards/near_duplicate_penalty/std': 0.00962204672396183, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9854571223258972, 'rewards/distinct_2/std': 0.03325406089425087, 'rewards/total_composite/mean': 0.19826576113700867, 'rewards/total_composite/std': 0.12650547921657562, 'reward': 0.19826576113700867, 'reward_std': 0.12650547921657562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20425374805927277, 'sampling/sampling_logp_difference/max': 1.7474346160888672, 'sampling/importance_sampling_ratio/min': 0.17422030866146088, 'sampling/importance_sampling_ratio/mean': 1.0514130592346191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1810272932052612, 'clip_ratio/low_mean': 0.10374086908996105, 'clip_ratio/low_min': 0.10374086908996105, 'clip_ratio/high_mean': 0.08999050967395306, 'clip_ratio/high_max': 0.08999050967395306, 'clip_ratio/region_mean': 0.1937313787639141, 'reward_total_mean': 0.19826576113700867, 'reward_meter_mean': 0.5467900633811951, 'reward_meter_std': 0.36305561661720276, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9762307405471802, 'reward_repeat_penalty_std': 0.04083777964115143, 'reward_repeat_floor_mean': 0.9972497820854187, 'reward_repeat_floor_std': 0.004304586909711361, 'reward_near_duplicate_penalty_mean': 0.9903911352157593, 'reward_near_duplicate_penalty_std': 0.00962204672396183, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9854571223258972, 'reward_distinct_2_std': 0.03325406089425087, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.19826576113700867, 'reward_total_composite_std': 0.12650547921657562, 'epoch': 0.02}

 10%|█         | 331/3300 [1:10:58<8:54:33, 10.80s/it]INFO 04-12 17:09:25 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 332/3300 [1:11:06<8:10:51,  9.92s/it]2026-04-12 17:09:32,645 | INFO | train_grpo_train | metrics_logged mode=train step=332

                                                      
{'loss': 0.0621, 'grad_norm': 15.719371795654297, 'learning_rate': 8.996969696969697e-06, 'num_tokens': 621960.0, 'completions/mean_length': 44.875, 'completions/min_length': 38.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7392489314079285, 'rewards/meter/std': 0.4448111951351166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.2503390610218048, 'rewards/repeat_penalty/mean': 0.8899849057197571, 'rewards/repeat_penalty/std': 0.1520090252161026, 'rewards/repeat_floor/mean': 0.9863828420639038, 'rewards/repeat_floor/std': 0.01840312033891678, 'rewards/near_duplicate_penalty/mean': 0.950941801071167, 'rewards/near_duplicate_penalty/std': 0.058040469884872437, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9304618835449219, 'rewards/distinct_2/std': 0.11295775324106216, 'rewards/total_composite/mean': 0.4814991354942322, 'rewards/total_composite/std': 0.3474137485027313, 'reward': 0.4814991354942322, 'reward_std': 0.3474137485027313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18538926541805267, 'sampling/sampling_logp_difference/max': 1.3385437726974487, 'sampling/importance_sampling_ratio/min': 0.26222726702690125, 'sampling/importance_sampling_ratio/mean': 1.021314263343811, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4954206347465515, 'clip_ratio/low_mean': 0.09325589146465063, 'clip_ratio/low_min': 0.09325589146465063, 'clip_ratio/high_mean': 0.07908442988991737, 'clip_ratio/high_max': 0.07908442988991737, 'clip_ratio/region_mean': 0.172340321354568, 'reward_total_mean': 0.4814991354942322, 'reward_meter_mean': 0.7392489314079285, 'reward_meter_std': 0.4448111951351166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8899849057197571, 'reward_repeat_penalty_std': 0.1520090252161026, 'reward_repeat_floor_mean': 0.9863828420639038, 'reward_repeat_floor_std': 0.01840312033891678, 'reward_near_duplicate_penalty_mean': 0.950941801071167, 'reward_near_duplicate_penalty_std': 0.058040469884872437, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9304618835449219, 'reward_distinct_2_std': 0.11295775324106216, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.2503390610218048, 'reward_total_composite_mean': 0.4814991354942322, 'reward_total_composite_std': 0.3474137485027313, 'epoch': 0.02}

 10%|█         | 332/3300 [1:11:06<8:10:51,  9.92s/it]INFO 04-12 17:09:32 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 333/3300 [1:11:19<8:58:01, 10.88s/it]2026-04-12 17:09:45,757 | INFO | train_grpo_train | metrics_logged mode=train step=333

                                                      
{'loss': -0.037, 'grad_norm': 6.324812412261963, 'learning_rate': 8.993939393939395e-06, 'num_tokens': 623609.0, 'completions/mean_length': 97.125, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.85714340209961, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.3344654142856598, 'rewards/meter/std': 0.3455292284488678, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9542313814163208, 'rewards/lexical_plausibility/std': 0.12945318222045898, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.3374272584915161, 'rewards/repeat_penalty/mean': 0.8774904012680054, 'rewards/repeat_penalty/std': 0.26813358068466187, 'rewards/repeat_floor/mean': 0.9819175004959106, 'rewards/repeat_floor/std': 0.03940322995185852, 'rewards/near_duplicate_penalty/mean': 0.9334884285926819, 'rewards/near_duplicate_penalty/std': 0.11620126664638519, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9307692050933838, 'rewards/distinct_2/std': 0.19581419229507446, 'rewards/total_composite/mean': 0.16200512647628784, 'rewards/total_composite/std': 0.15872851014137268, 'reward': 0.16200512647628784, 'reward_std': 0.15872851014137268, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20526380836963654, 'sampling/sampling_logp_difference/max': 2.0141396522521973, 'sampling/importance_sampling_ratio/min': 0.13343514502048492, 'sampling/importance_sampling_ratio/mean': 1.0475233793258667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.45550936460495, 'clip_ratio/low_mean': 0.08114611357450485, 'clip_ratio/low_min': 0.08114611357450485, 'clip_ratio/high_mean': 0.08637686166912317, 'clip_ratio/high_max': 0.08637686166912317, 'clip_ratio/region_mean': 0.16752297524362803, 'reward_total_mean': 0.16200512647628784, 'reward_meter_mean': 0.3344654142856598, 'reward_meter_std': 0.3455292284488678, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9542313814163208, 'reward_lexical_plausibility_std': 0.12945318222045898, 'reward_repeat_penalty_mean': 0.8774904012680054, 'reward_repeat_penalty_std': 0.26813358068466187, 'reward_repeat_floor_mean': 0.9819175004959106, 'reward_repeat_floor_std': 0.03940322995185852, 'reward_near_duplicate_penalty_mean': 0.9334884285926819, 'reward_near_duplicate_penalty_std': 0.11620126664638519, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9307692050933838, 'reward_distinct_2_std': 0.19581419229507446, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.3374272584915161, 'reward_total_composite_mean': 0.16200512647628784, 'reward_total_composite_std': 0.15872851014137268, 'epoch': 0.02}

 10%|█         | 333/3300 [1:11:19<8:58:01, 10.88s/it]INFO 04-12 17:09:45 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 334/3300 [1:11:27<8:12:30,  9.96s/it]2026-04-12 17:09:53,582 | INFO | train_grpo_train | metrics_logged mode=train step=334

                                                      
{'loss': 0.0261, 'grad_norm': 15.41718578338623, 'learning_rate': 8.990909090909092e-06, 'num_tokens': 625241.0, 'completions/mean_length': 32.0, 'completions/min_length': 28.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.3763948082923889, 'rewards/meter/std': 0.33724334836006165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9312499761581421, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.8744696378707886, 'rewards/repeat_penalty/std': 0.0550113208591938, 'rewards/repeat_floor/mean': 0.9811704158782959, 'rewards/repeat_floor/std': 0.00825169775635004, 'rewards/near_duplicate_penalty/mean': 0.8744696378707886, 'rewards/near_duplicate_penalty/std': 0.0550113208591938, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3434305787086487, 'rewards/total_composite/std': 0.30541110038757324, 'reward': 0.3434305787086487, 'reward_std': 0.30541110038757324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11281287670135498, 'sampling/sampling_logp_difference/max': 2.121975898742676, 'sampling/importance_sampling_ratio/min': 0.11979468911886215, 'sampling/importance_sampling_ratio/mean': 0.995778501033783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49717994406819344, 'clip_ratio/low_mean': 0.0692060450091958, 'clip_ratio/low_min': 0.0692060450091958, 'clip_ratio/high_mean': 0.03940886817872524, 'clip_ratio/high_max': 0.03940886817872524, 'clip_ratio/region_mean': 0.10861491318792105, 'reward_total_mean': 0.3434305787086487, 'reward_meter_mean': 0.3763948082923889, 'reward_meter_std': 0.33724334836006165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8744696378707886, 'reward_repeat_penalty_std': 0.0550113208591938, 'reward_repeat_floor_mean': 0.9811704158782959, 'reward_repeat_floor_std': 0.00825169775635004, 'reward_near_duplicate_penalty_mean': 0.8744696378707886, 'reward_near_duplicate_penalty_std': 0.0550113208591938, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9312499761581421, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.3434305787086487, 'reward_total_composite_std': 0.30541110038757324, 'epoch': 0.02}

 10%|█         | 334/3300 [1:11:27<8:12:30,  9.96s/it]INFO 04-12 17:09:53 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 335/3300 [1:11:41<9:11:03, 11.15s/it]2026-04-12 17:10:07,507 | INFO | train_grpo_train | metrics_logged mode=train step=335

                                                      
{'loss': -0.0462, 'grad_norm': 6.4609808921813965, 'learning_rate': 8.98787878787879e-06, 'num_tokens': 627326.0, 'completions/mean_length': 139.625, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.42857360839844, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.5335412621498108, 'rewards/meter/std': 0.4433954656124115, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9291043281555176, 'rewards/lexical_plausibility/std': 0.12765920162200928, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.9270902872085571, 'rewards/repeat_penalty/std': 0.060830820351839066, 'rewards/repeat_floor/mean': 0.991856038570404, 'rewards/repeat_floor/std': 0.006524866912513971, 'rewards/near_duplicate_penalty/mean': 0.9762570858001709, 'rewards/near_duplicate_penalty/std': 0.0161101333796978, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9490830898284912, 'rewards/distinct_2/std': 0.049733735620975494, 'rewards/total_composite/mean': 0.17774923145771027, 'rewards/total_composite/std': 0.1678026020526886, 'reward': 0.17774923145771027, 'reward_std': 0.1678025871515274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20937727391719818, 'sampling/sampling_logp_difference/max': 2.242918014526367, 'sampling/importance_sampling_ratio/min': 0.10614831000566483, 'sampling/importance_sampling_ratio/mean': 1.030953049659729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.594652384519577, 'clip_ratio/low_mean': 0.07543732784688473, 'clip_ratio/low_min': 0.07543732784688473, 'clip_ratio/high_mean': 0.06352804228663445, 'clip_ratio/high_max': 0.06352804228663445, 'clip_ratio/region_mean': 0.13896537013351917, 'reward_total_mean': 0.17774923145771027, 'reward_meter_mean': 0.5335412621498108, 'reward_meter_std': 0.4433954656124115, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9291043281555176, 'reward_lexical_plausibility_std': 0.12765920162200928, 'reward_repeat_penalty_mean': 0.9270902872085571, 'reward_repeat_penalty_std': 0.060830820351839066, 'reward_repeat_floor_mean': 0.991856038570404, 'reward_repeat_floor_std': 0.006524866912513971, 'reward_near_duplicate_penalty_mean': 0.9762570858001709, 'reward_near_duplicate_penalty_std': 0.0161101333796978, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9490830898284912, 'reward_distinct_2_std': 0.049733735620975494, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.17774923145771027, 'reward_total_composite_std': 0.1678026020526886, 'epoch': 0.02}

 10%|█         | 335/3300 [1:11:41<9:11:03, 11.15s/it]INFO 04-12 17:10:07 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 336/3300 [1:11:50<8:40:18, 10.53s/it]2026-04-12 17:10:16,595 | INFO | train_grpo_train | metrics_logged mode=train step=336

                                                      
{'loss': 0.0914, 'grad_norm': 10.081744194030762, 'learning_rate': 8.984848484848485e-06, 'num_tokens': 629413.0, 'completions/mean_length': 91.875, 'completions/min_length': 69.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.875, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9746475219726562, 'rewards/meter/std': 0.024971460923552513, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48249998688697815, 'rewards/judge_quality/std': 0.26222947239875793, 'rewards/repeat_penalty/mean': 0.927332878112793, 'rewards/repeat_penalty/std': 0.09314508736133575, 'rewards/repeat_floor/mean': 0.9919474124908447, 'rewards/repeat_floor/std': 0.009451538324356079, 'rewards/near_duplicate_penalty/mean': 0.9778405427932739, 'rewards/near_duplicate_penalty/std': 0.012873595580458641, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9474591016769409, 'rewards/distinct_2/std': 0.08519522100687027, 'rewards/total_composite/mean': 0.468796968460083, 'rewards/total_composite/std': 0.2625926434993744, 'reward': 0.468796968460083, 'reward_std': 0.2625926434993744, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2029363065958023, 'sampling/sampling_logp_difference/max': 2.309131622314453, 'sampling/importance_sampling_ratio/min': 0.09934748709201813, 'sampling/importance_sampling_ratio/mean': 1.038577914237976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.144473224878311, 'clip_ratio/low_mean': 0.1315344125032425, 'clip_ratio/low_min': 0.1315344125032425, 'clip_ratio/high_mean': 0.03803516924381256, 'clip_ratio/high_max': 0.03803516924381256, 'clip_ratio/region_mean': 0.16956958174705505, 'reward_total_mean': 0.468796968460083, 'reward_meter_mean': 0.9746475219726562, 'reward_meter_std': 0.024971460923552513, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.927332878112793, 'reward_repeat_penalty_std': 0.09314508736133575, 'reward_repeat_floor_mean': 0.9919474124908447, 'reward_repeat_floor_std': 0.009451538324356079, 'reward_near_duplicate_penalty_mean': 0.9778405427932739, 'reward_near_duplicate_penalty_std': 0.012873595580458641, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9474591016769409, 'reward_distinct_2_std': 0.08519522100687027, 'reward_judge_quality_mean': 0.48249998688697815, 'reward_judge_quality_std': 0.26222947239875793, 'reward_total_composite_mean': 0.468796968460083, 'reward_total_composite_std': 0.2625926434993744, 'epoch': 0.02}

 10%|█         | 336/3300 [1:11:50<8:40:18, 10.53s/it]INFO 04-12 17:10:16 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 337/3300 [1:11:57<7:52:57,  9.58s/it]2026-04-12 17:10:23,943 | INFO | train_grpo_train | metrics_logged mode=train step=337

                                                      
{'loss': 0.117, 'grad_norm': 23.095705032348633, 'learning_rate': 8.981818181818182e-06, 'num_tokens': 630798.0, 'completions/mean_length': 22.125, 'completions/min_length': 18.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6356936693191528, 'rewards/meter/std': 0.4879623353481293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.831250011920929, 'rewards/judge_quality/std': 0.27559998631477356, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49132412672042847, 'rewards/total_composite/std': 0.4544740319252014, 'reward': 0.49132412672042847, 'reward_std': 0.4544740319252014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13992170989513397, 'sampling/sampling_logp_difference/max': 1.1954584121704102, 'sampling/importance_sampling_ratio/min': 0.30256521701812744, 'sampling/importance_sampling_ratio/mean': 1.0445661544799805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.285483255982399, 'clip_ratio/low_mean': 0.07399891968816519, 'clip_ratio/low_min': 0.07399891968816519, 'clip_ratio/high_mean': 0.07162698544561863, 'clip_ratio/high_max': 0.07162698544561863, 'clip_ratio/region_mean': 0.14562590513378382, 'reward_total_mean': 0.49132412672042847, 'reward_meter_mean': 0.6356936693191528, 'reward_meter_std': 0.4879623353481293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.831250011920929, 'reward_judge_quality_std': 0.27559998631477356, 'reward_total_composite_mean': 0.49132412672042847, 'reward_total_composite_std': 0.4544740319252014, 'epoch': 0.02}

 10%|█         | 337/3300 [1:11:57<7:52:57,  9.58s/it]INFO 04-12 17:10:24 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 338/3300 [1:12:06<7:36:00,  9.24s/it]2026-04-12 17:10:32,387 | INFO | train_grpo_train | metrics_logged mode=train step=338

                                                      
{'loss': 0.0889, 'grad_norm': 18.370445251464844, 'learning_rate': 8.97878787878788e-06, 'num_tokens': 632504.0, 'completions/mean_length': 32.25, 'completions/min_length': 29.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.305332213640213, 'rewards/meter/std': 0.2746565639972687, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.2655990719795227, 'rewards/repeat_penalty/mean': 0.9727151989936829, 'rewards/repeat_penalty/std': 0.039194636046886444, 'rewards/repeat_floor/mean': 0.9963968992233276, 'rewards/repeat_floor/std': 0.004689360503107309, 'rewards/near_duplicate_penalty/mean': 0.9814739227294922, 'rewards/near_duplicate_penalty/std': 0.02053259313106537, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.25843459367752075, 'rewards/total_composite/std': 0.2698877453804016, 'reward': 0.25843459367752075, 'reward_std': 0.2698877155780792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1457429975271225, 'sampling/sampling_logp_difference/max': 2.2974085807800293, 'sampling/importance_sampling_ratio/min': 0.10051899403333664, 'sampling/importance_sampling_ratio/mean': 0.991886556148529, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8547821119427681, 'clip_ratio/low_mean': 0.07054067309945822, 'clip_ratio/low_min': 0.07054067309945822, 'clip_ratio/high_mean': 0.08742121048271656, 'clip_ratio/high_max': 0.08742121048271656, 'clip_ratio/region_mean': 0.15796188358217478, 'reward_total_mean': 0.25843459367752075, 'reward_meter_mean': 0.305332213640213, 'reward_meter_std': 0.2746565639972687, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9727151989936829, 'reward_repeat_penalty_std': 0.039194636046886444, 'reward_repeat_floor_mean': 0.9963968992233276, 'reward_repeat_floor_std': 0.004689360503107309, 'reward_near_duplicate_penalty_mean': 0.9814739227294922, 'reward_near_duplicate_penalty_std': 0.02053259313106537, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.2655990719795227, 'reward_total_composite_mean': 0.25843459367752075, 'reward_total_composite_std': 0.2698877453804016, 'epoch': 0.02}

 10%|█         | 338/3300 [1:12:06<7:36:00,  9.24s/it]INFO 04-12 17:10:32 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 339/3300 [1:12:17<8:08:26,  9.90s/it]2026-04-12 17:10:43,827 | INFO | train_grpo_train | metrics_logged mode=train step=339

                                                      
{'loss': 0.1405, 'grad_norm': 5.842377185821533, 'learning_rate': 8.975757575757577e-06, 'num_tokens': 636055.0, 'completions/mean_length': 232.875, 'completions/min_length': 101.0, 'completions/max_length': 264.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 232.875, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 264.0, 'rewards/meter/mean': 0.7700097560882568, 'rewards/meter/std': 0.2653079628944397, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.08864052593708038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9175000190734863, 'rewards/count_floor/std': 0.026592157781124115, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.612265944480896, 'rewards/repeat_penalty/std': 0.3063110113143921, 'rewards/repeat_floor/mean': 0.9568666219711304, 'rewards/repeat_floor/std': 0.03775676712393761, 'rewards/near_duplicate_penalty/mean': 0.92043137550354, 'rewards/near_duplicate_penalty/std': 0.07240410894155502, 'rewards/opening_diversity/mean': 0.990384578704834, 'rewards/opening_diversity/std': 0.027196412906050682, 'rewards/distinct_2/mean': 0.6597642302513123, 'rewards/distinct_2/std': 0.2927972972393036, 'rewards/total_composite/mean': 0.19891105592250824, 'rewards/total_composite/std': 0.09404581040143967, 'reward': 0.19891105592250824, 'reward_std': 0.09404580295085907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15775896608829498, 'sampling/sampling_logp_difference/max': 1.7183837890625, 'sampling/importance_sampling_ratio/min': 0.17935580015182495, 'sampling/importance_sampling_ratio/mean': 1.028220534324646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4751836135983467, 'clip_ratio/low_mean': 0.05240679858252406, 'clip_ratio/low_min': 0.05240679858252406, 'clip_ratio/high_mean': 0.09283824265003204, 'clip_ratio/high_max': 0.09283824265003204, 'clip_ratio/region_mean': 0.1452450412325561, 'reward_total_mean': 0.19891105592250824, 'reward_meter_mean': 0.7700097560882568, 'reward_meter_std': 0.2653079628944397, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.08864052593708038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9175000190734863, 'reward_count_floor_std': 0.026592157781124115, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.612265944480896, 'reward_repeat_penalty_std': 0.3063110113143921, 'reward_repeat_floor_mean': 0.9568666219711304, 'reward_repeat_floor_std': 0.03775676712393761, 'reward_near_duplicate_penalty_mean': 0.92043137550354, 'reward_near_duplicate_penalty_std': 0.07240410894155502, 'reward_opening_diversity_mean': 0.990384578704834, 'reward_opening_diversity_std': 0.027196412906050682, 'reward_distinct_2_mean': 0.6597642302513123, 'reward_distinct_2_std': 0.2927972972393036, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.19891105592250824, 'reward_total_composite_std': 0.09404581040143967, 'epoch': 0.02}

 10%|█         | 339/3300 [1:12:17<8:08:26,  9.90s/it]INFO 04-12 17:10:44 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 340/3300 [1:12:31<9:03:58, 11.03s/it]2026-04-12 17:10:57,485 | INFO | train_grpo_train | metrics_logged mode=train step=340

                                                      
{'loss': -0.1115, 'grad_norm': 4.992883682250977, 'learning_rate': 8.972727272727272e-06, 'num_tokens': 638043.0, 'completions/mean_length': 135.5, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.71428680419922, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.6432020664215088, 'rewards/meter/std': 0.38517627120018005, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.971875011920929, 'rewards/count_floor/std': 0.05580177903175354, 'rewards/lexical_plausibility/mean': 0.9695316553115845, 'rewards/lexical_plausibility/std': 0.08617749065160751, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9816516637802124, 'rewards/repeat_penalty/std': 0.025329409167170525, 'rewards/repeat_floor/mean': 0.9978464245796204, 'rewards/repeat_floor/std': 0.002785325050354004, 'rewards/near_duplicate_penalty/mean': 0.9918843507766724, 'rewards/near_duplicate_penalty/std': 0.009686465375125408, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9895973801612854, 'rewards/distinct_2/std': 0.01927127316594124, 'rewards/total_composite/mean': 0.2468227744102478, 'rewards/total_composite/std': 0.14359211921691895, 'reward': 0.2468227744102478, 'reward_std': 0.14359211921691895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2102813720703125, 'sampling/sampling_logp_difference/max': 2.0120761394500732, 'sampling/importance_sampling_ratio/min': 0.13371077179908752, 'sampling/importance_sampling_ratio/mean': 1.0276657342910767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.608161985874176, 'clip_ratio/low_mean': 0.051081934943795204, 'clip_ratio/low_min': 0.051081934943795204, 'clip_ratio/high_mean': 0.11457800958305597, 'clip_ratio/high_max': 0.11457800958305597, 'clip_ratio/region_mean': 0.16565994452685118, 'reward_total_mean': 0.2468227744102478, 'reward_meter_mean': 0.6432020664215088, 'reward_meter_std': 0.38517627120018005, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.971875011920929, 'reward_count_floor_std': 0.05580177903175354, 'reward_lexical_plausibility_mean': 0.9695316553115845, 'reward_lexical_plausibility_std': 0.08617749065160751, 'reward_repeat_penalty_mean': 0.9816516637802124, 'reward_repeat_penalty_std': 0.025329409167170525, 'reward_repeat_floor_mean': 0.9978464245796204, 'reward_repeat_floor_std': 0.002785325050354004, 'reward_near_duplicate_penalty_mean': 0.9918843507766724, 'reward_near_duplicate_penalty_std': 0.009686465375125408, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9895973801612854, 'reward_distinct_2_std': 0.01927127316594124, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.2468227744102478, 'reward_total_composite_std': 0.14359211921691895, 'epoch': 0.02}

 10%|█         | 340/3300 [1:12:31<9:03:58, 11.03s/it]INFO 04-12 17:10:57 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 341/3300 [1:12:39<8:23:35, 10.21s/it]2026-04-12 17:11:05,799 | INFO | train_grpo_train | metrics_logged mode=train step=341

                                                      
{'loss': -0.0248, 'grad_norm': 14.611605644226074, 'learning_rate': 8.969696969696971e-06, 'num_tokens': 639705.0, 'completions/mean_length': 37.75, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9741537570953369, 'rewards/meter/std': 0.018269570544362068, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.9834984540939331, 'rewards/repeat_penalty/std': 0.04450032487511635, 'rewards/repeat_floor/mean': 0.9979462623596191, 'rewards/repeat_floor/std': 0.005484189372509718, 'rewards/near_duplicate_penalty/mean': 0.9913252592086792, 'rewards/near_duplicate_penalty/std': 0.022403746843338013, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.37501412630081177, 'rewards/total_composite/std': 0.23986515402793884, 'reward': 0.37501412630081177, 'reward_std': 0.23986515402793884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19988596439361572, 'sampling/sampling_logp_difference/max': 1.6068944931030273, 'sampling/importance_sampling_ratio/min': 0.20050933957099915, 'sampling/importance_sampling_ratio/mean': 0.9957857728004456, 'sampling/importance_sampling_ratio/max': 1.8379789590835571, 'entropy': 1.989799588918686, 'clip_ratio/low_mean': 0.09782019723206758, 'clip_ratio/low_min': 0.09782019723206758, 'clip_ratio/high_mean': 0.053601814433932304, 'clip_ratio/high_max': 0.053601814433932304, 'clip_ratio/region_mean': 0.1514220116659999, 'reward_total_mean': 0.37501412630081177, 'reward_meter_mean': 0.9741537570953369, 'reward_meter_std': 0.018269570544362068, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9834984540939331, 'reward_repeat_penalty_std': 0.04450032487511635, 'reward_repeat_floor_mean': 0.9979462623596191, 'reward_repeat_floor_std': 0.005484189372509718, 'reward_near_duplicate_penalty_mean': 0.9913252592086792, 'reward_near_duplicate_penalty_std': 0.022403746843338013, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.37501412630081177, 'reward_total_composite_std': 0.23986515402793884, 'epoch': 0.02}

 10%|█         | 341/3300 [1:12:39<8:23:35, 10.21s/it]INFO 04-12 17:11:06 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 342/3300 [1:12:53<9:12:06, 11.20s/it]2026-04-12 17:11:19,299 | INFO | train_grpo_train | metrics_logged mode=train step=342

                                                      
{'loss': -0.0887, 'grad_norm': 5.5700602531433105, 'learning_rate': 8.966666666666667e-06, 'num_tokens': 641231.0, 'completions/mean_length': 97.75, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.57143020629883, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8387867212295532, 'rewards/meter/std': 0.33959174156188965, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9402444958686829, 'rewards/lexical_plausibility/std': 0.11721032857894897, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.9470990896224976, 'rewards/repeat_penalty/std': 0.11002103984355927, 'rewards/repeat_floor/mean': 0.9949995279312134, 'rewards/repeat_floor/std': 0.008322611451148987, 'rewards/near_duplicate_penalty/mean': 0.9852362871170044, 'rewards/near_duplicate_penalty/std': 0.013277208432555199, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.2696087956428528, 'rewards/total_composite/std': 0.18130116164684296, 'reward': 0.2696087956428528, 'reward_std': 0.18130116164684296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2047036588191986, 'sampling/sampling_logp_difference/max': 2.1100409030914307, 'sampling/importance_sampling_ratio/min': 0.1212330088019371, 'sampling/importance_sampling_ratio/mean': 1.0251104831695557, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3973510414361954, 'clip_ratio/low_mean': 0.08008603379130363, 'clip_ratio/low_min': 0.08008603379130363, 'clip_ratio/high_mean': 0.07942503038793802, 'clip_ratio/high_max': 0.07942503038793802, 'clip_ratio/region_mean': 0.15951106417924166, 'reward_total_mean': 0.2696087956428528, 'reward_meter_mean': 0.8387867212295532, 'reward_meter_std': 0.33959174156188965, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9402444958686829, 'reward_lexical_plausibility_std': 0.11721032857894897, 'reward_repeat_penalty_mean': 0.9470990896224976, 'reward_repeat_penalty_std': 0.11002103984355927, 'reward_repeat_floor_mean': 0.9949995279312134, 'reward_repeat_floor_std': 0.008322611451148987, 'reward_near_duplicate_penalty_mean': 0.9852362871170044, 'reward_near_duplicate_penalty_std': 0.013277208432555199, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.2696087956428528, 'reward_total_composite_std': 0.18130116164684296, 'epoch': 0.02}

 10%|█         | 342/3300 [1:12:53<9:12:06, 11.20s/it]INFO 04-12 17:11:19 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 343/3300 [1:13:00<8:20:28, 10.16s/it]2026-04-12 17:11:27,017 | INFO | train_grpo_train | metrics_logged mode=train step=343

                                                      
{'loss': -0.119, 'grad_norm': 30.251667022705078, 'learning_rate': 8.963636363636364e-06, 'num_tokens': 642400.0, 'completions/mean_length': 18.125, 'completions/min_length': 14.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.125, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.3097136318683624, 'rewards/meter/std': 0.4110397696495056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6549999713897705, 'rewards/judge_quality/std': 0.30603456497192383, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20622941851615906, 'rewards/total_composite/std': 0.3088410496711731, 'reward': 0.20622941851615906, 'reward_std': 0.3088410496711731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1320803165435791, 'sampling/sampling_logp_difference/max': 1.0534849166870117, 'sampling/importance_sampling_ratio/min': 0.34872034192085266, 'sampling/importance_sampling_ratio/mean': 0.9990514516830444, 'sampling/importance_sampling_ratio/max': 1.5801866054534912, 'entropy': 1.2459015175700188, 'clip_ratio/low_mean': 0.08357597421854734, 'clip_ratio/low_min': 0.08357597421854734, 'clip_ratio/high_mean': 0.028125000186264515, 'clip_ratio/high_max': 0.028125000186264515, 'clip_ratio/region_mean': 0.11170097440481186, 'reward_total_mean': 0.20622941851615906, 'reward_meter_mean': 0.3097136318683624, 'reward_meter_std': 0.4110397696495056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6549999713897705, 'reward_judge_quality_std': 0.30603456497192383, 'reward_total_composite_mean': 0.20622941851615906, 'reward_total_composite_std': 0.3088410496711731, 'epoch': 0.02}

 10%|█         | 343/3300 [1:13:00<8:20:28, 10.16s/it]INFO 04-12 17:11:27 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 344/3300 [1:13:09<7:52:11,  9.58s/it]2026-04-12 17:11:35,270 | INFO | train_grpo_train | metrics_logged mode=train step=344

                                                      
{'loss': -0.0041, 'grad_norm': 20.964290618896484, 'learning_rate': 8.960606060606061e-06, 'num_tokens': 643931.0, 'completions/mean_length': 31.375, 'completions/min_length': 28.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6057028770446777, 'rewards/meter/std': 0.3872171640396118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9916572570800781, 'rewards/lexical_plausibility/std': 0.023596882820129395, 'rewards/judge_quality/mean': 0.5337499976158142, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/repeat_penalty/mean': 0.9896930456161499, 'rewards/repeat_penalty/std': 0.017976956441998482, 'rewards/repeat_floor/mean': 0.9984539747238159, 'rewards/repeat_floor/std': 0.0026965467259287834, 'rewards/near_duplicate_penalty/mean': 0.9896930456161499, 'rewards/near_duplicate_penalty/std': 0.017976956441998482, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3533518314361572, 'rewards/total_composite/std': 0.29671037197113037, 'reward': 0.3533518314361572, 'reward_std': 0.29671037197113037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18388763070106506, 'sampling/sampling_logp_difference/max': 2.3177013397216797, 'sampling/importance_sampling_ratio/min': 0.09849974513053894, 'sampling/importance_sampling_ratio/mean': 1.0307540893554688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1475198790431023, 'clip_ratio/low_mean': 0.09939177334308624, 'clip_ratio/low_min': 0.09939177334308624, 'clip_ratio/high_mean': 0.09572322852909565, 'clip_ratio/high_max': 0.09572322852909565, 'clip_ratio/region_mean': 0.1951150018721819, 'reward_total_mean': 0.3533518314361572, 'reward_meter_mean': 0.6057028770446777, 'reward_meter_std': 0.3872171640396118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9916572570800781, 'reward_lexical_plausibility_std': 0.023596882820129395, 'reward_repeat_penalty_mean': 0.9896930456161499, 'reward_repeat_penalty_std': 0.017976956441998482, 'reward_repeat_floor_mean': 0.9984539747238159, 'reward_repeat_floor_std': 0.0026965467259287834, 'reward_near_duplicate_penalty_mean': 0.9896930456161499, 'reward_near_duplicate_penalty_std': 0.017976956441998482, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5337499976158142, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.3533518314361572, 'reward_total_composite_std': 0.29671037197113037, 'epoch': 0.02}

 10%|█         | 344/3300 [1:13:09<7:52:11,  9.58s/it]INFO 04-12 17:11:35 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 345/3300 [1:13:17<7:30:08,  9.14s/it]2026-04-12 17:11:43,374 | INFO | train_grpo_train | metrics_logged mode=train step=345

                                                      
{'loss': 0.2317, 'grad_norm': 15.436206817626953, 'learning_rate': 8.957575757575758e-06, 'num_tokens': 645505.0, 'completions/mean_length': 44.75, 'completions/min_length': 32.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.422957181930542, 'rewards/meter/std': 0.37584730982780457, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9929937124252319, 'rewards/repeat_penalty/std': 0.01794723980128765, 'rewards/repeat_floor/mean': 0.9989490509033203, 'rewards/repeat_floor/std': 0.0026920896489173174, 'rewards/near_duplicate_penalty/mean': 0.9929937124252319, 'rewards/near_duplicate_penalty/std': 0.01794723980128765, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1952688992023468, 'rewards/total_composite/std': 0.1663089394569397, 'reward': 0.1952688992023468, 'reward_std': 0.1663089245557785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1889335960149765, 'sampling/sampling_logp_difference/max': 1.4973969459533691, 'sampling/importance_sampling_ratio/min': 0.2237117439508438, 'sampling/importance_sampling_ratio/mean': 1.0012470483779907, 'sampling/importance_sampling_ratio/max': 1.9571013450622559, 'entropy': 1.4239961951971054, 'clip_ratio/low_mean': 0.0805791518650949, 'clip_ratio/low_min': 0.0805791518650949, 'clip_ratio/high_mean': 0.056426698341965675, 'clip_ratio/high_max': 0.056426698341965675, 'clip_ratio/region_mean': 0.13700585020706058, 'reward_total_mean': 0.1952688992023468, 'reward_meter_mean': 0.422957181930542, 'reward_meter_std': 0.37584730982780457, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9929937124252319, 'reward_repeat_penalty_std': 0.01794723980128765, 'reward_repeat_floor_mean': 0.9989490509033203, 'reward_repeat_floor_std': 0.0026920896489173174, 'reward_near_duplicate_penalty_mean': 0.9929937124252319, 'reward_near_duplicate_penalty_std': 0.01794723980128765, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.1952688992023468, 'reward_total_composite_std': 0.1663089394569397, 'epoch': 0.02}

 10%|█         | 345/3300 [1:13:17<7:30:08,  9.14s/it]INFO 04-12 17:11:43 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 346/3300 [1:13:30<8:37:03, 10.50s/it]2026-04-12 17:11:57,056 | INFO | train_grpo_train | metrics_logged mode=train step=346

                                                      
{'loss': -0.2181, 'grad_norm': 2.4873247146606445, 'learning_rate': 8.954545454545456e-06, 'num_tokens': 647715.0, 'completions/mean_length': 225.25, 'completions/min_length': 118.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 129.6666717529297, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.8998244404792786, 'rewards/meter/std': 0.1816522479057312, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9143241047859192, 'rewards/lexical_plausibility/std': 0.16443945467472076, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.852328896522522, 'rewards/repeat_penalty/std': 0.2978889048099518, 'rewards/repeat_floor/mean': 0.9822715520858765, 'rewards/repeat_floor/std': 0.03648386895656586, 'rewards/near_duplicate_penalty/mean': 0.9550489187240601, 'rewards/near_duplicate_penalty/std': 0.08249100297689438, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8987692594528198, 'rewards/distinct_2/std': 0.2097814381122589, 'rewards/total_composite/mean': 0.29176658391952515, 'rewards/total_composite/std': 0.16676844656467438, 'reward': 0.29176658391952515, 'reward_std': 0.16676846146583557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19190363585948944, 'sampling/sampling_logp_difference/max': 1.6605510711669922, 'sampling/importance_sampling_ratio/min': 0.19003422558307648, 'sampling/importance_sampling_ratio/mean': 1.0319976806640625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.504880130290985, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.1571525987237692, 'clip_ratio/high_max': 0.1571525987237692, 'clip_ratio/region_mean': 0.1571525987237692, 'reward_total_mean': 0.29176658391952515, 'reward_meter_mean': 0.8998244404792786, 'reward_meter_std': 0.1816522479057312, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9143241047859192, 'reward_lexical_plausibility_std': 0.16443945467472076, 'reward_repeat_penalty_mean': 0.852328896522522, 'reward_repeat_penalty_std': 0.2978889048099518, 'reward_repeat_floor_mean': 0.9822715520858765, 'reward_repeat_floor_std': 0.03648386895656586, 'reward_near_duplicate_penalty_mean': 0.9550489187240601, 'reward_near_duplicate_penalty_std': 0.08249100297689438, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8987692594528198, 'reward_distinct_2_std': 0.2097814381122589, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.29176658391952515, 'reward_total_composite_std': 0.16676844656467438, 'epoch': 0.02}

 10%|█         | 346/3300 [1:13:30<8:37:03, 10.50s/it]INFO 04-12 17:11:57 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 347/3300 [1:13:38<7:47:58,  9.51s/it]2026-04-12 17:12:04,245 | INFO | train_grpo_train | metrics_logged mode=train step=347

                                                      
{'loss': -0.0315, 'grad_norm': 29.22102165222168, 'learning_rate': 8.951515151515153e-06, 'num_tokens': 649136.0, 'completions/mean_length': 18.625, 'completions/min_length': 14.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.6110738515853882, 'rewards/meter/std': 0.48206010460853577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.9387500286102295, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5648782849311829, 'rewards/total_composite/std': 0.44622117280960083, 'reward': 0.5648782849311829, 'reward_std': 0.44622117280960083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18010105192661285, 'sampling/sampling_logp_difference/max': 1.4136180877685547, 'sampling/importance_sampling_ratio/min': 0.24326156079769135, 'sampling/importance_sampling_ratio/mean': 1.022800326347351, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3609361499547958, 'clip_ratio/low_mean': 0.044903927482664585, 'clip_ratio/low_min': 0.044903927482664585, 'clip_ratio/high_mean': 0.11353801470249891, 'clip_ratio/high_max': 0.11353801470249891, 'clip_ratio/region_mean': 0.1584419421851635, 'reward_total_mean': 0.5648782849311829, 'reward_meter_mean': 0.6110738515853882, 'reward_meter_std': 0.48206010460853577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9387500286102295, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.5648782849311829, 'reward_total_composite_std': 0.44622117280960083, 'epoch': 0.02}

 11%|█         | 347/3300 [1:13:38<7:47:58,  9.51s/it]INFO 04-12 17:12:04 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 348/3300 [1:13:46<7:26:11,  9.07s/it]2026-04-12 17:12:12,288 | INFO | train_grpo_train | metrics_logged mode=train step=348

                                                      
{'loss': 0.1019, 'grad_norm': 16.159345626831055, 'learning_rate': 8.94848484848485e-06, 'num_tokens': 650959.0, 'completions/mean_length': 37.875, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6884497404098511, 'rewards/meter/std': 0.3094226121902466, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3046778440475464, 'rewards/repeat_penalty/mean': 0.8826597929000854, 'rewards/repeat_penalty/std': 0.2535805404186249, 'rewards/repeat_floor/mean': 0.9875184297561646, 'rewards/repeat_floor/std': 0.024994323030114174, 'rewards/near_duplicate_penalty/mean': 0.9620068073272705, 'rewards/near_duplicate_penalty/std': 0.056380949914455414, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9454711675643921, 'rewards/distinct_2/std': 0.12832939624786377, 'rewards/total_composite/mean': 0.4540812373161316, 'rewards/total_composite/std': 0.35023945569992065, 'reward': 0.4540812373161316, 'reward_std': 0.35023942589759827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1531946212053299, 'sampling/sampling_logp_difference/max': 1.2673401832580566, 'sampling/importance_sampling_ratio/min': 0.2815795838832855, 'sampling/importance_sampling_ratio/mean': 1.0411983728408813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1304955556988716, 'clip_ratio/low_mean': 0.094660934060812, 'clip_ratio/low_min': 0.094660934060812, 'clip_ratio/high_mean': 0.03898809617385268, 'clip_ratio/high_max': 0.03898809617385268, 'clip_ratio/region_mean': 0.13364903023466468, 'reward_total_mean': 0.4540812373161316, 'reward_meter_mean': 0.6884497404098511, 'reward_meter_std': 0.3094226121902466, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8826597929000854, 'reward_repeat_penalty_std': 0.2535805404186249, 'reward_repeat_floor_mean': 0.9875184297561646, 'reward_repeat_floor_std': 0.024994323030114174, 'reward_near_duplicate_penalty_mean': 0.9620068073272705, 'reward_near_duplicate_penalty_std': 0.056380949914455414, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9454711675643921, 'reward_distinct_2_std': 0.12832939624786377, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3046778440475464, 'reward_total_composite_mean': 0.4540812373161316, 'reward_total_composite_std': 0.35023945569992065, 'epoch': 0.02}

 11%|█         | 348/3300 [1:13:46<7:26:11,  9.07s/it]INFO 04-12 17:12:12 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 349/3300 [1:13:54<7:15:34,  8.86s/it]2026-04-12 17:12:20,649 | INFO | train_grpo_train | metrics_logged mode=train step=349

                                                      
{'loss': 0.0321, 'grad_norm': 16.422067642211914, 'learning_rate': 8.945454545454546e-06, 'num_tokens': 652502.0, 'completions/mean_length': 40.875, 'completions/min_length': 37.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7101024389266968, 'rewards/meter/std': 0.38420072197914124, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5249999761581421, 'rewards/judge_quality/std': 0.23628070950508118, 'rewards/repeat_penalty/mean': 0.9426225423812866, 'rewards/repeat_penalty/std': 0.14379814267158508, 'rewards/repeat_floor/mean': 0.9940041899681091, 'rewards/repeat_floor/std': 0.014206590130925179, 'rewards/near_duplicate_penalty/mean': 0.9830174446105957, 'rewards/near_duplicate_penalty/std': 0.030583709478378296, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.3415525257587433, 'rewards/total_composite/std': 0.24239085614681244, 'reward': 0.3415525257587433, 'reward_std': 0.24239085614681244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17797724902629852, 'sampling/sampling_logp_difference/max': 1.420872688293457, 'sampling/importance_sampling_ratio/min': 0.24150317907333374, 'sampling/importance_sampling_ratio/mean': 1.0511341094970703, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.652133971452713, 'clip_ratio/low_mean': 0.04242530930787325, 'clip_ratio/low_min': 0.04242530930787325, 'clip_ratio/high_mean': 0.0944294249638915, 'clip_ratio/high_max': 0.0944294249638915, 'clip_ratio/region_mean': 0.13685473427176476, 'reward_total_mean': 0.3415525257587433, 'reward_meter_mean': 0.7101024389266968, 'reward_meter_std': 0.38420072197914124, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9426225423812866, 'reward_repeat_penalty_std': 0.14379814267158508, 'reward_repeat_floor_mean': 0.9940041899681091, 'reward_repeat_floor_std': 0.014206590130925179, 'reward_near_duplicate_penalty_mean': 0.9830174446105957, 'reward_near_duplicate_penalty_std': 0.030583709478378296, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.5249999761581421, 'reward_judge_quality_std': 0.23628070950508118, 'reward_total_composite_mean': 0.3415525257587433, 'reward_total_composite_std': 0.24239085614681244, 'epoch': 0.02}

 11%|█         | 349/3300 [1:13:54<7:15:34,  8.86s/it]INFO 04-12 17:12:20 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 350/3300 [1:14:04<7:33:14,  9.22s/it]2026-04-12 17:12:30,711 | INFO | train_grpo_train | metrics_logged mode=train step=350

                                                      
{'loss': 0.1157, 'grad_norm': 12.64472484588623, 'learning_rate': 8.942424242424243e-06, 'num_tokens': 654296.0, 'completions/mean_length': 60.25, 'completions/min_length': 50.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.5088467001914978, 'rewards/meter/std': 0.29301485419273376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9579846262931824, 'rewards/repeat_penalty/std': 0.06559336185455322, 'rewards/repeat_floor/mean': 0.9949827194213867, 'rewards/repeat_floor/std': 0.00734906317666173, 'rewards/near_duplicate_penalty/mean': 0.981090784072876, 'rewards/near_duplicate_penalty/std': 0.023007508367300034, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9757676720619202, 'rewards/distinct_2/std': 0.051502835005521774, 'rewards/total_composite/mean': 0.16552507877349854, 'rewards/total_composite/std': 0.1121489629149437, 'reward': 0.16552507877349854, 'reward_std': 0.1121489554643631, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18388310074806213, 'sampling/sampling_logp_difference/max': 1.8387489318847656, 'sampling/importance_sampling_ratio/min': 0.15901625156402588, 'sampling/importance_sampling_ratio/mean': 1.0285263061523438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.589398980140686, 'clip_ratio/low_mean': 0.09163669403642416, 'clip_ratio/low_min': 0.09163669403642416, 'clip_ratio/high_mean': 0.06783670000731945, 'clip_ratio/high_max': 0.06783670000731945, 'clip_ratio/region_mean': 0.1594733940437436, 'reward_total_mean': 0.16552507877349854, 'reward_meter_mean': 0.5088467001914978, 'reward_meter_std': 0.29301485419273376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9579846262931824, 'reward_repeat_penalty_std': 0.06559336185455322, 'reward_repeat_floor_mean': 0.9949827194213867, 'reward_repeat_floor_std': 0.00734906317666173, 'reward_near_duplicate_penalty_mean': 0.981090784072876, 'reward_near_duplicate_penalty_std': 0.023007508367300034, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9757676720619202, 'reward_distinct_2_std': 0.051502835005521774, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.16552507877349854, 'reward_total_composite_std': 0.1121489629149437, 'epoch': 0.02}

 11%|█         | 350/3300 [1:14:04<7:33:14,  9.22s/it]INFO 04-12 17:12:30 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:10<00:56,  5.17s/it]

 23%|██▎       | 3/13 [00:17<00:59,  5.99s/it]

 31%|███       | 4/13 [00:29<01:13,  8.20s/it]

 38%|███▊      | 5/13 [00:41<01:16,  9.54s/it]

 46%|████▌     | 6/13 [00:53<01:13, 10.49s/it]

 54%|█████▍    | 7/13 [01:05<01:04, 10.82s/it]

 62%|██████▏   | 8/13 [01:17<00:56, 11.33s/it]

 69%|██████▉   | 9/13 [01:29<00:45, 11.36s/it]

 77%|███████▋  | 10/13 [01:40<00:34, 11.41s/it]

 85%|████████▍ | 11/13 [01:50<00:22, 11.05s/it]

 92%|█████████▏| 12/13 [02:02<00:11, 11.29s/it]

100%|██████████| 13/13 [02:14<00:00, 11.36s/it]2026-04-12 17:14:55,503 | INFO | train_grpo_train | metrics_logged mode=eval step=350

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 144.7598, 'eval_samples_per_second': 0.718, 'eval_steps_per_second': 0.09, 'eval_num_tokens': 654296.0, 'eval_completions/mean_length': 191.83653846153845, 'eval_completions/min_length': 38.46153846153846, 'eval_completions/max_length': 456.15384615384613, 'eval_completions/clipped_ratio': 0.10576923076923077, 'eval_completions/mean_terminated_length': 153.97482182429388, 'eval_completions/min_terminated_length': 38.46153846153846, 'eval_completions/max_terminated_length': 326.38461538461536, 'eval_rewards/meter/mean': 0.6187796180064862, 'eval_rewards/meter/std': 0.3389004170894623, 'eval_rewards/count_adherence/mean': 0.8130821356406579, 'eval_rewards/count_adherence/std': 0.17862427234649658, 'eval_rewards/arabic_clean/mean': 0.7884615384615384, 'eval_rewards/arabic_clean/std': 0.4020594083345853, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9942307655627911, 'eval_rewards/arabic_floor/std': 0.01350019184442667, 'eval_rewards/count_floor/mean': 0.9439246425261865, 'eval_rewards/count_floor/std': 0.05358728565848791, 'eval_rewards/lexical_plausibility/mean': 0.9635053460414593, 'eval_rewards/lexical_plausibility/std': 0.08824307564646006, 'eval_rewards/judge_quality/mean': 0.3454807652876927, 'eval_rewards/judge_quality/std': 0.21029764815018728, 'eval_rewards/repeat_penalty/mean': 0.7776802594845111, 'eval_rewards/repeat_penalty/std': 0.29088544043210834, 'eval_rewards/repeat_floor/mean': 0.9749004519902743, 'eval_rewards/repeat_floor/std': 0.03472641851896277, 'eval_rewards/near_duplicate_penalty/mean': 0.9483849773040185, 'eval_rewards/near_duplicate_penalty/std': 0.06695309152396825, 'eval_rewards/opening_diversity/mean': 0.9860931359804593, 'eval_rewards/opening_diversity/std': 0.03646333916829182, 'eval_rewards/distinct_2/mean': 0.8164123571836032, 'eval_rewards/distinct_2/std': 0.27630343202214974, 'eval_rewards/total_composite/mean': 0.20006298101865327, 'eval_rewards/total_composite/std': 0.18831274658441544, 'eval_reward': 0.20006298101865327, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1004102614063483, 'eval_sampling/sampling_logp_difference/max': 1.3372466747577374, 'eval_sampling/importance_sampling_ratio/min': 0.26711086011849916, 'eval_sampling/importance_sampling_ratio/mean': 1.0299167449657733, 'eval_sampling/importance_sampling_ratio/max': 1.6205786374899058, 'eval_entropy': 1.29667682831104, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.20006298101865327, 'eval_reward_meter_mean': 0.6187796180064862, 'eval_reward_meter_std': 0.3389004170894623, 'eval_reward_count_adherence_mean': 0.8130821356406579, 'eval_reward_count_adherence_std': 0.17862427234649658, 'eval_reward_arabic_clean_mean': 0.7884615384615384, 'eval_reward_arabic_clean_std': 0.4020594083345853, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9942307655627911, 'eval_reward_arabic_floor_std': 0.01350019184442667, 'eval_reward_count_floor_mean': 0.9439246425261865, 'eval_reward_count_floor_std': 0.05358728565848791, 'eval_reward_lexical_plausibility_mean': 0.9635053460414593, 'eval_reward_lexical_plausibility_std': 0.08824307564646006, 'eval_reward_repeat_penalty_mean': 0.7776802594845111, 'eval_reward_repeat_penalty_std': 0.29088544043210834, 'eval_reward_repeat_floor_mean': 0.9749004519902743, 'eval_reward_repeat_floor_std': 0.03472641851896277, 'eval_reward_near_duplicate_penalty_mean': 0.9483849773040185, 'eval_reward_near_duplicate_penalty_std': 0.06695309152396825, 'eval_reward_opening_diversity_mean': 0.9860931359804593, 'eval_reward_opening_diversity_std': 0.03646333916829182, 'eval_reward_distinct_2_mean': 0.8164123571836032, 'eval_reward_distinct_2_std': 0.27630343202214974, 'eval_reward_judge_quality_mean': 0.3454807652876927, 'eval_reward_judge_quality_std': 0.21029764815018728, 'eval_reward_total_composite_mean': 0.20006298101865327, 'eval_reward_total_composite_std': 0.18831274658441544, 'epoch': 0.02}

 11%|█         | 350/3300 [1:16:29<7:33:14,  9.22s/it]

100%|██████████| 13/13 [02:14<00:00, 11.36s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 11%|█         | 351/3300 [1:16:42<44:08:19, 53.88s/it]2026-04-12 17:15:08,810 | INFO | train_grpo_train | metrics_logged mode=train step=351

                                                       
{'loss': 0.0068, 'grad_norm': 10.76603889465332, 'learning_rate': 8.93939393939394e-06, 'num_tokens': 656400.0, 'completions/mean_length': 89.0, 'completions/min_length': 64.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.5040855407714844, 'rewards/meter/std': 0.3973883092403412, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.7608175277709961, 'rewards/repeat_penalty/std': 0.17903673648834229, 'rewards/repeat_floor/mean': 0.9755060076713562, 'rewards/repeat_floor/std': 0.01765604317188263, 'rewards/near_duplicate_penalty/mean': 0.9422729015350342, 'rewards/near_duplicate_penalty/std': 0.04892025515437126, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8240566253662109, 'rewards/distinct_2/std': 0.12574154138565063, 'rewards/total_composite/mean': 0.19187957048416138, 'rewards/total_composite/std': 0.15353643894195557, 'reward': 0.19187957048416138, 'reward_std': 0.15353643894195557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18777593970298767, 'sampling/sampling_logp_difference/max': 1.54642915725708, 'sampling/importance_sampling_ratio/min': 0.21300722658634186, 'sampling/importance_sampling_ratio/mean': 1.0208381414413452, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6547548472881317, 'clip_ratio/low_mean': 0.07881247438490391, 'clip_ratio/low_min': 0.07881247438490391, 'clip_ratio/high_mean': 0.11248178035020828, 'clip_ratio/high_max': 0.11248178035020828, 'clip_ratio/region_mean': 0.1912942547351122, 'reward_total_mean': 0.19187957048416138, 'reward_meter_mean': 0.5040855407714844, 'reward_meter_std': 0.3973883092403412, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7608175277709961, 'reward_repeat_penalty_std': 0.17903673648834229, 'reward_repeat_floor_mean': 0.9755060076713562, 'reward_repeat_floor_std': 0.01765604317188263, 'reward_near_duplicate_penalty_mean': 0.9422729015350342, 'reward_near_duplicate_penalty_std': 0.04892025515437126, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8240566253662109, 'reward_distinct_2_std': 0.12574154138565063, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.19187957048416138, 'reward_total_composite_std': 0.15353643894195557, 'epoch': 0.02}

 11%|█         | 351/3300 [1:16:42<44:08:19, 53.88s/it]INFO 04-12 17:15:09 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 352/3300 [1:16:50<32:55:29, 40.21s/it]2026-04-12 17:15:17,111 | INFO | train_grpo_train | metrics_logged mode=train step=352

                                                       
{'loss': -0.1344, 'grad_norm': 19.398202896118164, 'learning_rate': 8.936363636363638e-06, 'num_tokens': 657769.0, 'completions/mean_length': 24.125, 'completions/min_length': 17.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9638572335243225, 'rewards/meter/std': 0.07437514513731003, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.3620477616786957, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.631958544254303, 'rewards/total_composite/std': 0.34043869376182556, 'reward': 0.631958544254303, 'reward_std': 0.3404386639595032, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18029649555683136, 'sampling/sampling_logp_difference/max': 1.8325433731079102, 'sampling/importance_sampling_ratio/min': 0.1600060909986496, 'sampling/importance_sampling_ratio/mean': 1.0404987335205078, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6146266758441925, 'clip_ratio/low_mean': 0.08503377623856068, 'clip_ratio/low_min': 0.08503377623856068, 'clip_ratio/high_mean': 0.11602690163999796, 'clip_ratio/high_max': 0.11602690163999796, 'clip_ratio/region_mean': 0.20106067787855864, 'reward_total_mean': 0.631958544254303, 'reward_meter_mean': 0.9638572335243225, 'reward_meter_std': 0.07437514513731003, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.3620477616786957, 'reward_total_composite_mean': 0.631958544254303, 'reward_total_composite_std': 0.34043869376182556, 'epoch': 0.02}

 11%|█         | 352/3300 [1:16:50<32:55:29, 40.21s/it]INFO 04-12 17:15:17 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 353/3300 [1:16:59<25:13:35, 30.82s/it]2026-04-12 17:15:26,031 | INFO | train_grpo_train | metrics_logged mode=train step=353

                                                       
{'loss': 0.015, 'grad_norm': 16.73836326599121, 'learning_rate': 8.933333333333333e-06, 'num_tokens': 659380.0, 'completions/mean_length': 27.375, 'completions/min_length': 21.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6578106880187988, 'rewards/meter/std': 0.3674665093421936, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9607843160629272, 'rewards/lexical_plausibility/std': 0.11091870814561844, 'rewards/judge_quality/mean': 0.8762500286102295, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5581198334693909, 'rewards/total_composite/std': 0.34355735778808594, 'reward': 0.5581198334693909, 'reward_std': 0.3435573875904083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1278843730688095, 'sampling/sampling_logp_difference/max': 1.219208002090454, 'sampling/importance_sampling_ratio/min': 0.29546409845352173, 'sampling/importance_sampling_ratio/mean': 0.9870318174362183, 'sampling/importance_sampling_ratio/max': 1.876649022102356, 'entropy': 1.0078953579068184, 'clip_ratio/low_mean': 0.06661266321316361, 'clip_ratio/low_min': 0.06661266321316361, 'clip_ratio/high_mean': 0.039164246525615454, 'clip_ratio/high_max': 0.039164246525615454, 'clip_ratio/region_mean': 0.10577690973877907, 'reward_total_mean': 0.5581198334693909, 'reward_meter_mean': 0.6578106880187988, 'reward_meter_std': 0.3674665093421936, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9607843160629272, 'reward_lexical_plausibility_std': 0.11091870814561844, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8762500286102295, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5581198334693909, 'reward_total_composite_std': 0.34355735778808594, 'epoch': 0.02}

 11%|█         | 353/3300 [1:16:59<25:13:35, 30.82s/it]INFO 04-12 17:15:26 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 354/3300 [1:17:09<19:57:23, 24.39s/it]2026-04-12 17:15:35,417 | INFO | train_grpo_train | metrics_logged mode=train step=354

                                                       
{'loss': 0.0475, 'grad_norm': 15.901016235351562, 'learning_rate': 8.930303030303032e-06, 'num_tokens': 661002.0, 'completions/mean_length': 40.75, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6143783926963806, 'rewards/meter/std': 0.3755393624305725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6262499690055847, 'rewards/judge_quality/std': 0.2602711617946625, 'rewards/repeat_penalty/mean': 0.9787507057189941, 'rewards/repeat_penalty/std': 0.017217284068465233, 'rewards/repeat_floor/mean': 0.9968125820159912, 'rewards/repeat_floor/std': 0.0025825840421020985, 'rewards/near_duplicate_penalty/mean': 0.9787507057189941, 'rewards/near_duplicate_penalty/std': 0.017217284068465233, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38773494958877563, 'rewards/total_composite/std': 0.2940836548805237, 'reward': 0.38773494958877563, 'reward_std': 0.2940836548805237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14339789748191833, 'sampling/sampling_logp_difference/max': 1.6837506294250488, 'sampling/importance_sampling_ratio/min': 0.18567626178264618, 'sampling/importance_sampling_ratio/mean': 1.0196411609649658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.055018775165081, 'clip_ratio/low_mean': 0.05689931521192193, 'clip_ratio/low_min': 0.05689931521192193, 'clip_ratio/high_mean': 0.041463456116616726, 'clip_ratio/high_max': 0.041463456116616726, 'clip_ratio/region_mean': 0.09836277132853866, 'reward_total_mean': 0.38773494958877563, 'reward_meter_mean': 0.6143783926963806, 'reward_meter_std': 0.3755393624305725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9787507057189941, 'reward_repeat_penalty_std': 0.017217284068465233, 'reward_repeat_floor_mean': 0.9968125820159912, 'reward_repeat_floor_std': 0.0025825840421020985, 'reward_near_duplicate_penalty_mean': 0.9787507057189941, 'reward_near_duplicate_penalty_std': 0.017217284068465233, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6262499690055847, 'reward_judge_quality_std': 0.2602711617946625, 'reward_total_composite_mean': 0.38773494958877563, 'reward_total_composite_std': 0.2940836548805237, 'epoch': 0.02}

 11%|█         | 354/3300 [1:17:09<19:57:23, 24.39s/it]INFO 04-12 17:15:35 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 355/3300 [1:17:19<16:26:13, 20.09s/it]2026-04-12 17:15:45,487 | INFO | train_grpo_train | metrics_logged mode=train step=355

                                                       
{'loss': 0.0311, 'grad_norm': 9.03620719909668, 'learning_rate': 8.927272727272728e-06, 'num_tokens': 663476.0, 'completions/mean_length': 115.25, 'completions/min_length': 92.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.25, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.6599809527397156, 'rewards/meter/std': 0.3750666379928589, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9775000214576721, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.8529702425003052, 'rewards/repeat_penalty/std': 0.15573084354400635, 'rewards/repeat_floor/mean': 0.9831531643867493, 'rewards/repeat_floor/std': 0.01928114704787731, 'rewards/near_duplicate_penalty/mean': 0.9558294415473938, 'rewards/near_duplicate_penalty/std': 0.05347747728228569, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8864303827285767, 'rewards/distinct_2/std': 0.1275208294391632, 'rewards/total_composite/mean': 0.17257274687290192, 'rewards/total_composite/std': 0.1348387449979782, 'reward': 0.17257274687290192, 'reward_std': 0.1348387449979782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1767425537109375, 'sampling/sampling_logp_difference/max': 1.6581840515136719, 'sampling/importance_sampling_ratio/min': 0.19048456847667694, 'sampling/importance_sampling_ratio/mean': 1.031744360923767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7184740453958511, 'clip_ratio/low_mean': 0.11004908289760351, 'clip_ratio/low_min': 0.11004908289760351, 'clip_ratio/high_mean': 0.05548185668885708, 'clip_ratio/high_max': 0.05548185668885708, 'clip_ratio/region_mean': 0.1655309395864606, 'reward_total_mean': 0.17257274687290192, 'reward_meter_mean': 0.6599809527397156, 'reward_meter_std': 0.3750666379928589, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9775000214576721, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.8529702425003052, 'reward_repeat_penalty_std': 0.15573084354400635, 'reward_repeat_floor_mean': 0.9831531643867493, 'reward_repeat_floor_std': 0.01928114704787731, 'reward_near_duplicate_penalty_mean': 0.9558294415473938, 'reward_near_duplicate_penalty_std': 0.05347747728228569, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8864303827285767, 'reward_distinct_2_std': 0.1275208294391632, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.17257274687290192, 'reward_total_composite_std': 0.1348387449979782, 'epoch': 0.02}

 11%|█         | 355/3300 [1:17:19<16:26:13, 20.09s/it]INFO 04-12 17:15:45 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 356/3300 [1:17:33<15:02:20, 18.39s/it]2026-04-12 17:15:59,887 | INFO | train_grpo_train | metrics_logged mode=train step=356

                                                       
{'loss': -0.0711, 'grad_norm': 5.177454948425293, 'learning_rate': 8.924242424242425e-06, 'num_tokens': 665124.0, 'completions/mean_length': 108.0, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.28571701049805, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7517337799072266, 'rewards/meter/std': 0.4055202305316925, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9301379323005676, 'rewards/lexical_plausibility/std': 0.16398213803768158, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.23868316411972046, 'rewards/repeat_penalty/mean': 0.95414799451828, 'rewards/repeat_penalty/std': 0.0918140634894371, 'rewards/repeat_floor/mean': 0.9965232014656067, 'rewards/repeat_floor/std': 0.006406393367797136, 'rewards/near_duplicate_penalty/mean': 0.9957316517829895, 'rewards/near_duplicate_penalty/std': 0.011526302434504032, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.37034597992897034, 'rewards/total_composite/std': 0.26873478293418884, 'reward': 0.37034597992897034, 'reward_std': 0.26873478293418884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16878367960453033, 'sampling/sampling_logp_difference/max': 1.2176623344421387, 'sampling/importance_sampling_ratio/min': 0.2959211468696594, 'sampling/importance_sampling_ratio/mean': 1.0150970220565796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3488233089447021, 'clip_ratio/low_mean': 0.05135582014918327, 'clip_ratio/low_min': 0.05135582014918327, 'clip_ratio/high_mean': 0.07164111733436584, 'clip_ratio/high_max': 0.07164111733436584, 'clip_ratio/region_mean': 0.12299693748354912, 'reward_total_mean': 0.37034597992897034, 'reward_meter_mean': 0.7517337799072266, 'reward_meter_std': 0.4055202305316925, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9301379323005676, 'reward_lexical_plausibility_std': 0.16398213803768158, 'reward_repeat_penalty_mean': 0.95414799451828, 'reward_repeat_penalty_std': 0.0918140634894371, 'reward_repeat_floor_mean': 0.9965232014656067, 'reward_repeat_floor_std': 0.006406393367797136, 'reward_near_duplicate_penalty_mean': 0.9957316517829895, 'reward_near_duplicate_penalty_std': 0.011526302434504032, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.23868316411972046, 'reward_total_composite_mean': 0.37034597992897034, 'reward_total_composite_std': 0.26873478293418884, 'epoch': 0.02}

 11%|█         | 356/3300 [1:17:33<15:02:20, 18.39s/it]INFO 04-12 17:16:00 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 357/3300 [1:17:41<12:23:25, 15.16s/it]2026-04-12 17:16:07,499 | INFO | train_grpo_train | metrics_logged mode=train step=357

                                                       
{'loss': 0.0646, 'grad_norm': 31.852699279785156, 'learning_rate': 8.921212121212122e-06, 'num_tokens': 666383.0, 'completions/mean_length': 16.375, 'completions/min_length': 15.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.375, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.7171890735626221, 'rewards/meter/std': 0.3812980353832245, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7224999666213989, 'rewards/judge_quality/std': 0.28272905945777893, 'rewards/repeat_penalty/mean': 0.7396639585494995, 'rewards/repeat_penalty/std': 0.02923482470214367, 'rewards/repeat_floor/mean': 0.9829328060150146, 'rewards/repeat_floor/std': 0.005846968851983547, 'rewards/near_duplicate_penalty/mean': 0.9862185716629028, 'rewards/near_duplicate_penalty/std': 0.03897976875305176, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5214267373085022, 'rewards/total_composite/std': 0.3253229260444641, 'reward': 0.5214267373085022, 'reward_std': 0.3253229260444641, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20550617575645447, 'sampling/sampling_logp_difference/max': 1.5543746948242188, 'sampling/importance_sampling_ratio/min': 0.21132147312164307, 'sampling/importance_sampling_ratio/mean': 1.0072880983352661, 'sampling/importance_sampling_ratio/max': 1.886308193206787, 'entropy': 1.1666172742843628, 'clip_ratio/low_mean': 0.06274509895592928, 'clip_ratio/low_min': 0.06274509895592928, 'clip_ratio/high_mean': 0.06093750149011612, 'clip_ratio/high_max': 0.06093750149011612, 'clip_ratio/region_mean': 0.1236826004460454, 'reward_total_mean': 0.5214267373085022, 'reward_meter_mean': 0.7171890735626221, 'reward_meter_std': 0.3812980353832245, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7396639585494995, 'reward_repeat_penalty_std': 0.02923482470214367, 'reward_repeat_floor_mean': 0.9829328060150146, 'reward_repeat_floor_std': 0.005846968851983547, 'reward_near_duplicate_penalty_mean': 0.9862185716629028, 'reward_near_duplicate_penalty_std': 0.03897976875305176, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7224999666213989, 'reward_judge_quality_std': 0.28272905945777893, 'reward_total_composite_mean': 0.5214267373085022, 'reward_total_composite_std': 0.3253229260444641, 'epoch': 0.02}

 11%|█         | 357/3300 [1:17:41<12:23:25, 15.16s/it]INFO 04-12 17:16:07 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 358/3300 [1:17:50<10:49:51, 13.25s/it]2026-04-12 17:16:16,312 | INFO | train_grpo_train | metrics_logged mode=train step=358

                                                       
{'loss': 0.125, 'grad_norm': 13.7113037109375, 'learning_rate': 8.91818181818182e-06, 'num_tokens': 667931.0, 'completions/mean_length': 42.5, 'completions/min_length': 34.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8698437213897705, 'rewards/meter/std': 0.25244367122650146, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.29980650544166565, 'rewards/repeat_penalty/mean': 0.9122350215911865, 'rewards/repeat_penalty/std': 0.14125025272369385, 'rewards/repeat_floor/mean': 0.9896342754364014, 'rewards/repeat_floor/std': 0.0170968696475029, 'rewards/near_duplicate_penalty/mean': 0.968704104423523, 'rewards/near_duplicate_penalty/std': 0.05550863593816757, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.936985194683075, 'rewards/distinct_2/std': 0.10052327811717987, 'rewards/total_composite/mean': 0.5089001655578613, 'rewards/total_composite/std': 0.34366846084594727, 'reward': 0.5089001655578613, 'reward_std': 0.34366849064826965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1463201493024826, 'sampling/sampling_logp_difference/max': 1.3036623001098633, 'sampling/importance_sampling_ratio/min': 0.2715355455875397, 'sampling/importance_sampling_ratio/mean': 1.0084954500198364, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0244826599955559, 'clip_ratio/low_mean': 0.061588266398757696, 'clip_ratio/low_min': 0.061588266398757696, 'clip_ratio/high_mean': 0.05237369425594807, 'clip_ratio/high_max': 0.05237369425594807, 'clip_ratio/region_mean': 0.11396196065470576, 'reward_total_mean': 0.5089001655578613, 'reward_meter_mean': 0.8698437213897705, 'reward_meter_std': 0.25244367122650146, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9122350215911865, 'reward_repeat_penalty_std': 0.14125025272369385, 'reward_repeat_floor_mean': 0.9896342754364014, 'reward_repeat_floor_std': 0.0170968696475029, 'reward_near_duplicate_penalty_mean': 0.968704104423523, 'reward_near_duplicate_penalty_std': 0.05550863593816757, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.936985194683075, 'reward_distinct_2_std': 0.10052327811717987, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.29980650544166565, 'reward_total_composite_mean': 0.5089001655578613, 'reward_total_composite_std': 0.34366846084594727, 'epoch': 0.02}

 11%|█         | 358/3300 [1:17:50<10:49:51, 13.25s/it]INFO 04-12 17:16:16 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 359/3300 [1:18:00<10:06:18, 12.37s/it]2026-04-12 17:16:26,620 | INFO | train_grpo_train | metrics_logged mode=train step=359

                                                       
{'loss': 0.1407, 'grad_norm': 9.80252742767334, 'learning_rate': 8.915151515151515e-06, 'num_tokens': 669830.0, 'completions/mean_length': 78.375, 'completions/min_length': 47.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9786201119422913, 'rewards/meter/std': 0.015729490667581558, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9513295888900757, 'rewards/repeat_penalty/std': 0.048765815794467926, 'rewards/repeat_floor/mean': 0.9945008754730225, 'rewards/repeat_floor/std': 0.005266440100967884, 'rewards/near_duplicate_penalty/mean': 0.9827135801315308, 'rewards/near_duplicate_penalty/std': 0.014589764177799225, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9677096009254456, 'rewards/distinct_2/std': 0.038812536746263504, 'rewards/total_composite/mean': 0.3553292751312256, 'rewards/total_composite/std': 0.04220420494675636, 'reward': 0.3553292751312256, 'reward_std': 0.04220420867204666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16266827285289764, 'sampling/sampling_logp_difference/max': 2.425251007080078, 'sampling/importance_sampling_ratio/min': 0.08845590800046921, 'sampling/importance_sampling_ratio/mean': 1.0338057279586792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.546302244067192, 'clip_ratio/low_mean': 0.10199844837188721, 'clip_ratio/low_min': 0.10199844837188721, 'clip_ratio/high_mean': 0.05682884901762009, 'clip_ratio/high_max': 0.05682884901762009, 'clip_ratio/region_mean': 0.1588272973895073, 'reward_total_mean': 0.3553292751312256, 'reward_meter_mean': 0.9786201119422913, 'reward_meter_std': 0.015729490667581558, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9513295888900757, 'reward_repeat_penalty_std': 0.048765815794467926, 'reward_repeat_floor_mean': 0.9945008754730225, 'reward_repeat_floor_std': 0.005266440100967884, 'reward_near_duplicate_penalty_mean': 0.9827135801315308, 'reward_near_duplicate_penalty_std': 0.014589764177799225, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9677096009254456, 'reward_distinct_2_std': 0.038812536746263504, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.3553292751312256, 'reward_total_composite_std': 0.04220420494675636, 'epoch': 0.02}

 11%|█         | 359/3300 [1:18:00<10:06:18, 12.37s/it]INFO 04-12 17:16:26 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 360/3300 [1:18:10<9:29:38, 11.63s/it] 2026-04-12 17:16:36,510 | INFO | train_grpo_train | metrics_logged mode=train step=360

                                                      
{'loss': 0.0458, 'grad_norm': 8.05918025970459, 'learning_rate': 8.912121212121214e-06, 'num_tokens': 672380.0, 'completions/mean_length': 131.75, 'completions/min_length': 121.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.75, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.9727927446365356, 'rewards/meter/std': 0.05623193830251694, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.6192942261695862, 'rewards/repeat_penalty/std': 0.23883813619613647, 'rewards/repeat_floor/mean': 0.9574646949768066, 'rewards/repeat_floor/std': 0.027545826509594917, 'rewards/near_duplicate_penalty/mean': 0.9113891124725342, 'rewards/near_duplicate_penalty/std': 0.05591447278857231, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.695903480052948, 'rewards/distinct_2/std': 0.19315484166145325, 'rewards/total_composite/mean': 0.2731468379497528, 'rewards/total_composite/std': 0.13597573339939117, 'reward': 0.2731468379497528, 'reward_std': 0.13597573339939117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1425296515226364, 'sampling/sampling_logp_difference/max': 2.013340473175049, 'sampling/importance_sampling_ratio/min': 0.13354183733463287, 'sampling/importance_sampling_ratio/mean': 1.0411529541015625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2808602526783943, 'clip_ratio/low_mean': 0.04793283250182867, 'clip_ratio/low_min': 0.04793283250182867, 'clip_ratio/high_mean': 0.09201209712773561, 'clip_ratio/high_max': 0.09201209712773561, 'clip_ratio/region_mean': 0.13994492962956429, 'reward_total_mean': 0.2731468379497528, 'reward_meter_mean': 0.9727927446365356, 'reward_meter_std': 0.05623193830251694, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6192942261695862, 'reward_repeat_penalty_std': 0.23883813619613647, 'reward_repeat_floor_mean': 0.9574646949768066, 'reward_repeat_floor_std': 0.027545826509594917, 'reward_near_duplicate_penalty_mean': 0.9113891124725342, 'reward_near_duplicate_penalty_std': 0.05591447278857231, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.695903480052948, 'reward_distinct_2_std': 0.19315484166145325, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2731468379497528, 'reward_total_composite_std': 0.13597573339939117, 'epoch': 0.02}

 11%|█         | 360/3300 [1:18:10<9:29:38, 11.63s/it]INFO 04-12 17:16:36 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 361/3300 [1:18:18<8:37:45, 10.57s/it]2026-04-12 17:16:44,616 | INFO | train_grpo_train | metrics_logged mode=train step=361

                                                      
{'loss': 0.0623, 'grad_norm': 12.581989288330078, 'learning_rate': 8.90909090909091e-06, 'num_tokens': 674011.0, 'completions/mean_length': 41.875, 'completions/min_length': 39.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9363856315612793, 'rewards/meter/std': 0.03530038893222809, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.2796298861503601, 'rewards/repeat_penalty/mean': 0.9238135814666748, 'rewards/repeat_penalty/std': 0.09648194909095764, 'rewards/repeat_floor/mean': 0.9905943274497986, 'rewards/repeat_floor/std': 0.01162633579224348, 'rewards/near_duplicate_penalty/mean': 0.9627256393432617, 'rewards/near_duplicate_penalty/std': 0.04462045058608055, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.95761638879776, 'rewards/distinct_2/std': 0.06285658478736877, 'rewards/total_composite/mean': 0.4604656398296356, 'rewards/total_composite/std': 0.2795504629611969, 'reward': 0.4604656398296356, 'reward_std': 0.2795504331588745, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16266760230064392, 'sampling/sampling_logp_difference/max': 1.2708740234375, 'sampling/importance_sampling_ratio/min': 0.28058627247810364, 'sampling/importance_sampling_ratio/mean': 1.0118446350097656, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2341424077749252, 'clip_ratio/low_mean': 0.12048909720033407, 'clip_ratio/low_min': 0.12048909720033407, 'clip_ratio/high_mean': 0.03749999962747097, 'clip_ratio/high_max': 0.03749999962747097, 'clip_ratio/region_mean': 0.15798909682780504, 'reward_total_mean': 0.4604656398296356, 'reward_meter_mean': 0.9363856315612793, 'reward_meter_std': 0.03530038893222809, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9238135814666748, 'reward_repeat_penalty_std': 0.09648194909095764, 'reward_repeat_floor_mean': 0.9905943274497986, 'reward_repeat_floor_std': 0.01162633579224348, 'reward_near_duplicate_penalty_mean': 0.9627256393432617, 'reward_near_duplicate_penalty_std': 0.04462045058608055, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.95761638879776, 'reward_distinct_2_std': 0.06285658478736877, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.2796298861503601, 'reward_total_composite_mean': 0.4604656398296356, 'reward_total_composite_std': 0.2795504629611969, 'epoch': 0.02}

 11%|█         | 361/3300 [1:18:18<8:37:45, 10.57s/it]INFO 04-12 17:16:44 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 362/3300 [1:18:28<8:32:24, 10.46s/it]2026-04-12 17:16:54,834 | INFO | train_grpo_train | metrics_logged mode=train step=362

                                                      
{'loss': 0.0213, 'grad_norm': 8.42110824584961, 'learning_rate': 8.906060606060607e-06, 'num_tokens': 676290.0, 'completions/mean_length': 108.875, 'completions/min_length': 101.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.875, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.8600841760635376, 'rewards/meter/std': 0.2729659974575043, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9075525999069214, 'rewards/repeat_penalty/std': 0.09921245276927948, 'rewards/repeat_floor/mean': 0.9899107217788696, 'rewards/repeat_floor/std': 0.010220066644251347, 'rewards/near_duplicate_penalty/mean': 0.9748443961143494, 'rewards/near_duplicate_penalty/std': 0.017743341624736786, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9298232793807983, 'rewards/distinct_2/std': 0.08806470781564713, 'rewards/total_composite/mean': 0.34355881810188293, 'rewards/total_composite/std': 0.12731558084487915, 'reward': 0.34355881810188293, 'reward_std': 0.12731556594371796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1571449637413025, 'sampling/sampling_logp_difference/max': 1.6364707946777344, 'sampling/importance_sampling_ratio/min': 0.1946658343076706, 'sampling/importance_sampling_ratio/mean': 1.032671570777893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5461475998163223, 'clip_ratio/low_mean': 0.062248033471405506, 'clip_ratio/low_min': 0.062248033471405506, 'clip_ratio/high_mean': 0.07236636523157358, 'clip_ratio/high_max': 0.07236636523157358, 'clip_ratio/region_mean': 0.1346143987029791, 'reward_total_mean': 0.34355881810188293, 'reward_meter_mean': 0.8600841760635376, 'reward_meter_std': 0.2729659974575043, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9075525999069214, 'reward_repeat_penalty_std': 0.09921245276927948, 'reward_repeat_floor_mean': 0.9899107217788696, 'reward_repeat_floor_std': 0.010220066644251347, 'reward_near_duplicate_penalty_mean': 0.9748443961143494, 'reward_near_duplicate_penalty_std': 0.017743341624736786, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9298232793807983, 'reward_distinct_2_std': 0.08806470781564713, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.34355881810188293, 'reward_total_composite_std': 0.12731558084487915, 'epoch': 0.02}

 11%|█         | 362/3300 [1:18:28<8:32:24, 10.46s/it]INFO 04-12 17:16:55 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 363/3300 [1:18:43<9:36:24, 11.78s/it]2026-04-12 17:17:09,669 | INFO | train_grpo_train | metrics_logged mode=train step=363

                                                      
{'loss': -0.0513, 'grad_norm': 3.1529197692871094, 'learning_rate': 8.903030303030304e-06, 'num_tokens': 679959.0, 'completions/mean_length': 355.625, 'completions/min_length': 242.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 303.5, 'completions/min_terminated_length': 242.0, 'completions/max_terminated_length': 397.0, 'rewards/meter/mean': 0.8714945912361145, 'rewards/meter/std': 0.18211692571640015, 'rewards/count_adherence/mean': 0.6388888955116272, 'rewards/count_adherence/std': 0.27054065465927124, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8916666507720947, 'rewards/count_floor/std': 0.08116218447685242, 'rewards/lexical_plausibility/mean': 0.8723778128623962, 'rewards/lexical_plausibility/std': 0.1827472448348999, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.3893314599990845, 'rewards/repeat_penalty/std': 0.4064381420612335, 'rewards/repeat_floor/mean': 0.916541337966919, 'rewards/repeat_floor/std': 0.06046004220843315, 'rewards/near_duplicate_penalty/mean': 0.8059448003768921, 'rewards/near_duplicate_penalty/std': 0.15235379338264465, 'rewards/opening_diversity/mean': 0.9285511374473572, 'rewards/opening_diversity/std': 0.11984875798225403, 'rewards/distinct_2/mean': 0.4437389373779297, 'rewards/distinct_2/std': 0.39232203364372253, 'rewards/total_composite/mean': 0.1308785378932953, 'rewards/total_composite/std': 0.11467240005731583, 'reward': 0.1308785378932953, 'reward_std': 0.11467240005731583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10700076073408127, 'sampling/sampling_logp_difference/max': 1.3937110900878906, 'sampling/importance_sampling_ratio/min': 0.24815267324447632, 'sampling/importance_sampling_ratio/mean': 1.0298869609832764, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8040526509284973, 'clip_ratio/low_mean': 0.04215240245684981, 'clip_ratio/low_min': 0.04215240245684981, 'clip_ratio/high_mean': 0.014462810009717941, 'clip_ratio/high_max': 0.014462810009717941, 'clip_ratio/region_mean': 0.056615212466567755, 'reward_total_mean': 0.1308785378932953, 'reward_meter_mean': 0.8714945912361145, 'reward_meter_std': 0.18211692571640015, 'reward_count_adherence_mean': 0.6388888955116272, 'reward_count_adherence_std': 0.27054065465927124, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8916666507720947, 'reward_count_floor_std': 0.08116218447685242, 'reward_lexical_plausibility_mean': 0.8723778128623962, 'reward_lexical_plausibility_std': 0.1827472448348999, 'reward_repeat_penalty_mean': 0.3893314599990845, 'reward_repeat_penalty_std': 0.4064381420612335, 'reward_repeat_floor_mean': 0.916541337966919, 'reward_repeat_floor_std': 0.06046004220843315, 'reward_near_duplicate_penalty_mean': 0.8059448003768921, 'reward_near_duplicate_penalty_std': 0.15235379338264465, 'reward_opening_diversity_mean': 0.9285511374473572, 'reward_opening_diversity_std': 0.11984875798225403, 'reward_distinct_2_mean': 0.4437389373779297, 'reward_distinct_2_std': 0.39232203364372253, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.1308785378932953, 'reward_total_composite_std': 0.11467240005731583, 'epoch': 0.02}

 11%|█         | 363/3300 [1:18:43<9:36:24, 11.78s/it]INFO 04-12 17:17:09 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 364/3300 [1:18:53<9:11:26, 11.27s/it]2026-04-12 17:17:19,761 | INFO | train_grpo_train | metrics_logged mode=train step=364

                                                      
{'loss': -0.0034, 'grad_norm': 14.464784622192383, 'learning_rate': 8.900000000000001e-06, 'num_tokens': 681824.0, 'completions/mean_length': 61.125, 'completions/min_length': 54.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.8400001525878906, 'rewards/meter/std': 0.2184767723083496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9641179442405701, 'rewards/repeat_penalty/std': 0.02980227954685688, 'rewards/repeat_floor/mean': 0.9954157471656799, 'rewards/repeat_floor/std': 0.003415758488699794, 'rewards/near_duplicate_penalty/mean': 0.9778327941894531, 'rewards/near_duplicate_penalty/std': 0.01799760013818741, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9860093593597412, 'rewards/distinct_2/std': 0.025911342352628708, 'rewards/total_composite/mean': 0.30566051602363586, 'rewards/total_composite/std': 0.10472958534955978, 'reward': 0.30566051602363586, 'reward_std': 0.10472958534955978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1903400421142578, 'sampling/sampling_logp_difference/max': 1.529982089996338, 'sampling/importance_sampling_ratio/min': 0.21653956174850464, 'sampling/importance_sampling_ratio/mean': 1.018998146057129, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4020500034093857, 'clip_ratio/low_mean': 0.06268477812409401, 'clip_ratio/low_min': 0.06268477812409401, 'clip_ratio/high_mean': 0.1172677455469966, 'clip_ratio/high_max': 0.1172677455469966, 'clip_ratio/region_mean': 0.1799525236710906, 'reward_total_mean': 0.30566051602363586, 'reward_meter_mean': 0.8400001525878906, 'reward_meter_std': 0.2184767723083496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9641179442405701, 'reward_repeat_penalty_std': 0.02980227954685688, 'reward_repeat_floor_mean': 0.9954157471656799, 'reward_repeat_floor_std': 0.003415758488699794, 'reward_near_duplicate_penalty_mean': 0.9778327941894531, 'reward_near_duplicate_penalty_std': 0.01799760013818741, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9860093593597412, 'reward_distinct_2_std': 0.025911342352628708, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.30566051602363586, 'reward_total_composite_std': 0.10472958534955978, 'epoch': 0.02}

 11%|█         | 364/3300 [1:18:53<9:11:26, 11.27s/it]INFO 04-12 17:17:20 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 365/3300 [1:19:03<8:44:21, 10.72s/it]2026-04-12 17:17:29,194 | INFO | train_grpo_train | metrics_logged mode=train step=365

                                                      
{'loss': 0.0845, 'grad_norm': 11.75033950805664, 'learning_rate': 8.896969696969697e-06, 'num_tokens': 683438.0, 'completions/mean_length': 48.75, 'completions/min_length': 37.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9052804708480835, 'rewards/meter/std': 0.1998172402381897, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.960798978805542, 'rewards/repeat_penalty/std': 0.05375628545880318, 'rewards/repeat_floor/mean': 0.9946165084838867, 'rewards/repeat_floor/std': 0.007642010226845741, 'rewards/near_duplicate_penalty/mean': 0.9693548083305359, 'rewards/near_duplicate_penalty/std': 0.049885813146829605, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.3255195617675781, 'rewards/total_composite/std': 0.12233838438987732, 'reward': 0.3255195617675781, 'reward_std': 0.12233838438987732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1449141502380371, 'sampling/sampling_logp_difference/max': 2.0596542358398438, 'sampling/importance_sampling_ratio/min': 0.12749804556369781, 'sampling/importance_sampling_ratio/mean': 1.0061140060424805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3844047337770462, 'clip_ratio/low_mean': 0.035484074614942074, 'clip_ratio/low_min': 0.035484074614942074, 'clip_ratio/high_mean': 0.10205406416207552, 'clip_ratio/high_max': 0.10205406416207552, 'clip_ratio/region_mean': 0.1375381387770176, 'reward_total_mean': 0.3255195617675781, 'reward_meter_mean': 0.9052804708480835, 'reward_meter_std': 0.1998172402381897, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.960798978805542, 'reward_repeat_penalty_std': 0.05375628545880318, 'reward_repeat_floor_mean': 0.9946165084838867, 'reward_repeat_floor_std': 0.007642010226845741, 'reward_near_duplicate_penalty_mean': 0.9693548083305359, 'reward_near_duplicate_penalty_std': 0.049885813146829605, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.3255195617675781, 'reward_total_composite_std': 0.12233838438987732, 'epoch': 0.02}

 11%|█         | 365/3300 [1:19:03<8:44:21, 10.72s/it]INFO 04-12 17:17:29 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 366/3300 [1:19:12<8:21:32, 10.26s/it]2026-04-12 17:17:38,374 | INFO | train_grpo_train | metrics_logged mode=train step=366

                                                      
{'loss': 0.0536, 'grad_norm': 15.162606239318848, 'learning_rate': 8.893939393939394e-06, 'num_tokens': 685036.0, 'completions/mean_length': 48.75, 'completions/min_length': 43.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7743692994117737, 'rewards/meter/std': 0.3453243672847748, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.30789607763290405, 'rewards/repeat_penalty/mean': 0.9565794467926025, 'rewards/repeat_penalty/std': 0.08036018162965775, 'rewards/repeat_floor/mean': 0.994333028793335, 'rewards/repeat_floor/std': 0.011103691533207893, 'rewards/near_duplicate_penalty/mean': 0.9729592204093933, 'rewards/near_duplicate_penalty/std': 0.06053483858704567, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9821012020111084, 'rewards/distinct_2/std': 0.03315402567386627, 'rewards/total_composite/mean': 0.46257612109184265, 'rewards/total_composite/std': 0.29698050022125244, 'reward': 0.46257612109184265, 'reward_std': 0.29698050022125244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1491256058216095, 'sampling/sampling_logp_difference/max': 1.3192739486694336, 'sampling/importance_sampling_ratio/min': 0.26732930541038513, 'sampling/importance_sampling_ratio/mean': 1.0132737159729004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2742609530687332, 'clip_ratio/low_mean': 0.09035304561257362, 'clip_ratio/low_min': 0.09035304561257362, 'clip_ratio/high_mean': 0.07123877294361591, 'clip_ratio/high_max': 0.07123877294361591, 'clip_ratio/region_mean': 0.16159181855618954, 'reward_total_mean': 0.46257612109184265, 'reward_meter_mean': 0.7743692994117737, 'reward_meter_std': 0.3453243672847748, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9565794467926025, 'reward_repeat_penalty_std': 0.08036018162965775, 'reward_repeat_floor_mean': 0.994333028793335, 'reward_repeat_floor_std': 0.011103691533207893, 'reward_near_duplicate_penalty_mean': 0.9729592204093933, 'reward_near_duplicate_penalty_std': 0.06053483858704567, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9821012020111084, 'reward_distinct_2_std': 0.03315402567386627, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.30789607763290405, 'reward_total_composite_mean': 0.46257612109184265, 'reward_total_composite_std': 0.29698050022125244, 'epoch': 0.02}

 11%|█         | 366/3300 [1:19:12<8:21:32, 10.26s/it]INFO 04-12 17:17:38 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 367/3300 [1:19:21<8:07:52,  9.98s/it]2026-04-12 17:17:47,709 | INFO | train_grpo_train | metrics_logged mode=train step=367

                                                      
{'loss': -0.0082, 'grad_norm': 14.321744918823242, 'learning_rate': 8.890909090909091e-06, 'num_tokens': 686909.0, 'completions/mean_length': 74.125, 'completions/min_length': 60.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5353145003318787, 'rewards/meter/std': 0.34504234790802, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.04008918255567551, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.768750011920929, 'rewards/judge_quality/std': 0.21767853200435638, 'rewards/repeat_penalty/mean': 0.9708654284477234, 'rewards/repeat_penalty/std': 0.030411241576075554, 'rewards/repeat_floor/mean': 0.9964196085929871, 'rewards/repeat_floor/std': 0.003245903877541423, 'rewards/near_duplicate_penalty/mean': 0.9844350218772888, 'rewards/near_duplicate_penalty/std': 0.012171246111392975, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9861596822738647, 'rewards/distinct_2/std': 0.026049774140119553, 'rewards/total_composite/mean': 0.42354917526245117, 'rewards/total_composite/std': 0.32641151547431946, 'reward': 0.42354917526245117, 'reward_std': 0.32641151547431946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18299533426761627, 'sampling/sampling_logp_difference/max': 1.67714262008667, 'sampling/importance_sampling_ratio/min': 0.18690727651119232, 'sampling/importance_sampling_ratio/mean': 1.0056891441345215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9311363324522972, 'clip_ratio/low_mean': 0.09001205768436193, 'clip_ratio/low_min': 0.09001205768436193, 'clip_ratio/high_mean': 0.08259473368525505, 'clip_ratio/high_max': 0.08259473368525505, 'clip_ratio/region_mean': 0.17260679136961699, 'reward_total_mean': 0.42354917526245117, 'reward_meter_mean': 0.5353145003318787, 'reward_meter_std': 0.34504234790802, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.04008918255567551, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9708654284477234, 'reward_repeat_penalty_std': 0.030411241576075554, 'reward_repeat_floor_mean': 0.9964196085929871, 'reward_repeat_floor_std': 0.003245903877541423, 'reward_near_duplicate_penalty_mean': 0.9844350218772888, 'reward_near_duplicate_penalty_std': 0.012171246111392975, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9861596822738647, 'reward_distinct_2_std': 0.026049774140119553, 'reward_judge_quality_mean': 0.768750011920929, 'reward_judge_quality_std': 0.21767853200435638, 'reward_total_composite_mean': 0.42354917526245117, 'reward_total_composite_std': 0.32641151547431946, 'epoch': 0.02}

 11%|█         | 367/3300 [1:19:21<8:07:52,  9.98s/it]INFO 04-12 17:17:47 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 368/3300 [1:19:29<7:41:47,  9.45s/it]2026-04-12 17:17:55,919 | INFO | train_grpo_train | metrics_logged mode=train step=368

                                                      
{'loss': 0.1037, 'grad_norm': 24.614883422851562, 'learning_rate': 8.887878787878789e-06, 'num_tokens': 688344.0, 'completions/mean_length': 25.375, 'completions/min_length': 18.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.750693142414093, 'rewards/meter/std': 0.4268879294395447, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6675000190734863, 'rewards/judge_quality/std': 0.38975268602371216, 'rewards/repeat_penalty/mean': 0.7699704170227051, 'rewards/repeat_penalty/std': 0.0974021628499031, 'rewards/repeat_floor/mean': 0.9846190810203552, 'rewards/repeat_floor/std': 0.008517809212207794, 'rewards/near_duplicate_penalty/mean': 0.9849606156349182, 'rewards/near_duplicate_penalty/std': 0.039305027574300766, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5898723602294922, 'rewards/total_composite/std': 0.44210949540138245, 'reward': 0.5898723602294922, 'reward_std': 0.44210949540138245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18786954879760742, 'sampling/sampling_logp_difference/max': 1.6972036361694336, 'sampling/importance_sampling_ratio/min': 0.1831950843334198, 'sampling/importance_sampling_ratio/mean': 1.0439058542251587, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3269711285829544, 'clip_ratio/low_mean': 0.07433712109923363, 'clip_ratio/low_min': 0.07433712109923363, 'clip_ratio/high_mean': 0.09368357481434941, 'clip_ratio/high_max': 0.09368357481434941, 'clip_ratio/region_mean': 0.16802069591358304, 'reward_total_mean': 0.5898723602294922, 'reward_meter_mean': 0.750693142414093, 'reward_meter_std': 0.4268879294395447, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7699704170227051, 'reward_repeat_penalty_std': 0.0974021628499031, 'reward_repeat_floor_mean': 0.9846190810203552, 'reward_repeat_floor_std': 0.008517809212207794, 'reward_near_duplicate_penalty_mean': 0.9849606156349182, 'reward_near_duplicate_penalty_std': 0.039305027574300766, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6675000190734863, 'reward_judge_quality_std': 0.38975268602371216, 'reward_total_composite_mean': 0.5898723602294922, 'reward_total_composite_std': 0.44210949540138245, 'epoch': 0.02}

 11%|█         | 368/3300 [1:19:29<7:41:47,  9.45s/it]INFO 04-12 17:17:56 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 369/3300 [1:19:40<8:07:03,  9.97s/it]2026-04-12 17:18:07,104 | INFO | train_grpo_train | metrics_logged mode=train step=369

                                                      
{'loss': 0.1814, 'grad_norm': 11.213492393493652, 'learning_rate': 8.884848484848486e-06, 'num_tokens': 691255.0, 'completions/mean_length': 171.875, 'completions/min_length': 136.0, 'completions/max_length': 208.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 171.875, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 208.0, 'rewards/meter/mean': 0.5056368112564087, 'rewards/meter/std': 0.33909738063812256, 'rewards/count_adherence/mean': 0.734375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9203125238418579, 'rewards/count_floor/std': 0.013258260674774647, 'rewards/lexical_plausibility/mean': 0.9983552694320679, 'rewards/lexical_plausibility/std': 0.004652021918445826, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.8123867511749268, 'rewards/repeat_penalty/std': 0.11044201999902725, 'rewards/repeat_floor/mean': 0.9818981885910034, 'rewards/repeat_floor/std': 0.009863175451755524, 'rewards/near_duplicate_penalty/mean': 0.9661498665809631, 'rewards/near_duplicate_penalty/std': 0.016471514478325844, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8552853465080261, 'rewards/distinct_2/std': 0.09222674369812012, 'rewards/total_composite/mean': 0.13929015398025513, 'rewards/total_composite/std': 0.07933954149484634, 'reward': 0.13929015398025513, 'reward_std': 0.07933953404426575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21959906816482544, 'sampling/sampling_logp_difference/max': 2.517941474914551, 'sampling/importance_sampling_ratio/min': 0.08062540739774704, 'sampling/importance_sampling_ratio/mean': 1.0156633853912354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2597133070230484, 'clip_ratio/low_mean': 0.09509143978357315, 'clip_ratio/low_min': 0.09509143978357315, 'clip_ratio/high_mean': 0.09948217123746872, 'clip_ratio/high_max': 0.09948217123746872, 'clip_ratio/region_mean': 0.19457361102104187, 'reward_total_mean': 0.13929015398025513, 'reward_meter_mean': 0.5056368112564087, 'reward_meter_std': 0.33909738063812256, 'reward_count_adherence_mean': 0.734375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9203125238418579, 'reward_count_floor_std': 0.013258260674774647, 'reward_lexical_plausibility_mean': 0.9983552694320679, 'reward_lexical_plausibility_std': 0.004652021918445826, 'reward_repeat_penalty_mean': 0.8123867511749268, 'reward_repeat_penalty_std': 0.11044201999902725, 'reward_repeat_floor_mean': 0.9818981885910034, 'reward_repeat_floor_std': 0.009863175451755524, 'reward_near_duplicate_penalty_mean': 0.9661498665809631, 'reward_near_duplicate_penalty_std': 0.016471514478325844, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8552853465080261, 'reward_distinct_2_std': 0.09222674369812012, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.13929015398025513, 'reward_total_composite_std': 0.07933954149484634, 'epoch': 0.02}

 11%|█         | 369/3300 [1:19:40<8:07:03,  9.97s/it]INFO 04-12 17:18:07 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 370/3300 [1:19:51<8:08:26, 10.00s/it]2026-04-12 17:18:17,180 | INFO | train_grpo_train | metrics_logged mode=train step=370

                                                      
{'loss': 0.0645, 'grad_norm': 10.120687484741211, 'learning_rate': 8.881818181818183e-06, 'num_tokens': 693198.0, 'completions/mean_length': 64.875, 'completions/min_length': 49.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.7698711156845093, 'rewards/meter/std': 0.3382386863231659, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390502154827118, 'rewards/repeat_penalty/mean': 0.9402925968170166, 'rewards/repeat_penalty/std': 0.03833829611539841, 'rewards/repeat_floor/mean': 0.992631733417511, 'rewards/repeat_floor/std': 0.004147879313677549, 'rewards/near_duplicate_penalty/mean': 0.9681292772293091, 'rewards/near_duplicate_penalty/std': 0.017739932984113693, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.97124844789505, 'rewards/distinct_2/std': 0.035433944314718246, 'rewards/total_composite/mean': 0.34389227628707886, 'rewards/total_composite/std': 0.26905587315559387, 'reward': 0.34389227628707886, 'reward_std': 0.2690558433532715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17440679669380188, 'sampling/sampling_logp_difference/max': 2.0173730850219727, 'sampling/importance_sampling_ratio/min': 0.13300441205501556, 'sampling/importance_sampling_ratio/mean': 1.0081368684768677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5562718659639359, 'clip_ratio/low_mean': 0.10634280368685722, 'clip_ratio/low_min': 0.10634280368685722, 'clip_ratio/high_mean': 0.09382317028939724, 'clip_ratio/high_max': 0.09382317028939724, 'clip_ratio/region_mean': 0.20016597397625446, 'reward_total_mean': 0.34389227628707886, 'reward_meter_mean': 0.7698711156845093, 'reward_meter_std': 0.3382386863231659, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9402925968170166, 'reward_repeat_penalty_std': 0.03833829611539841, 'reward_repeat_floor_mean': 0.992631733417511, 'reward_repeat_floor_std': 0.004147879313677549, 'reward_near_duplicate_penalty_mean': 0.9681292772293091, 'reward_near_duplicate_penalty_std': 0.017739932984113693, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.97124844789505, 'reward_distinct_2_std': 0.035433944314718246, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390502154827118, 'reward_total_composite_mean': 0.34389227628707886, 'reward_total_composite_std': 0.26905587315559387, 'epoch': 0.02}

 11%|█         | 370/3300 [1:19:51<8:08:26, 10.00s/it]INFO 04-12 17:18:17 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 371/3300 [1:20:03<8:45:26, 10.76s/it]2026-04-12 17:18:29,723 | INFO | train_grpo_train | metrics_logged mode=train step=371

                                                      
{'loss': 0.0797, 'grad_norm': 4.831717014312744, 'learning_rate': 8.87878787878788e-06, 'num_tokens': 696939.0, 'completions/mean_length': 277.625, 'completions/min_length': 228.0, 'completions/max_length': 388.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 277.625, 'completions/min_terminated_length': 228.0, 'completions/max_terminated_length': 388.0, 'rewards/meter/mean': 0.7116806507110596, 'rewards/meter/std': 0.41652625799179077, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.2507132589817047, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8799999952316284, 'rewards/count_floor/std': 0.07521399110555649, 'rewards/lexical_plausibility/mean': 0.9736841917037964, 'rewards/lexical_plausibility/std': 0.07443228363990784, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.18465471267700195, 'rewards/repeat_penalty/std': 0.23464816808700562, 'rewards/repeat_floor/mean': 0.8778722882270813, 'rewards/repeat_floor/std': 0.06475692987442017, 'rewards/near_duplicate_penalty/mean': 0.7020835876464844, 'rewards/near_duplicate_penalty/std': 0.27087485790252686, 'rewards/opening_diversity/mean': 0.8760069608688354, 'rewards/opening_diversity/std': 0.16104254126548767, 'rewards/distinct_2/mean': 0.22221457958221436, 'rewards/distinct_2/std': 0.25585976243019104, 'rewards/total_composite/mean': 0.09522828459739685, 'rewards/total_composite/std': 0.06214176118373871, 'reward': 0.09522828459739685, 'reward_std': 0.06214176118373871, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10685379058122635, 'sampling/sampling_logp_difference/max': 4.284203052520752, 'sampling/importance_sampling_ratio/min': 0.013784602284431458, 'sampling/importance_sampling_ratio/mean': 1.006848692893982, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8036476597189903, 'clip_ratio/low_mean': 0.014815543312579393, 'clip_ratio/low_min': 0.014815543312579393, 'clip_ratio/high_mean': 0.08764226967468858, 'clip_ratio/high_max': 0.08764226967468858, 'clip_ratio/region_mean': 0.10245781298726797, 'reward_total_mean': 0.09522828459739685, 'reward_meter_mean': 0.7116806507110596, 'reward_meter_std': 0.41652625799179077, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.2507132589817047, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8799999952316284, 'reward_count_floor_std': 0.07521399110555649, 'reward_lexical_plausibility_mean': 0.9736841917037964, 'reward_lexical_plausibility_std': 0.07443228363990784, 'reward_repeat_penalty_mean': 0.18465471267700195, 'reward_repeat_penalty_std': 0.23464816808700562, 'reward_repeat_floor_mean': 0.8778722882270813, 'reward_repeat_floor_std': 0.06475692987442017, 'reward_near_duplicate_penalty_mean': 0.7020835876464844, 'reward_near_duplicate_penalty_std': 0.27087485790252686, 'reward_opening_diversity_mean': 0.8760069608688354, 'reward_opening_diversity_std': 0.16104254126548767, 'reward_distinct_2_mean': 0.22221457958221436, 'reward_distinct_2_std': 0.25585976243019104, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.09522828459739685, 'reward_total_composite_std': 0.06214176118373871, 'epoch': 0.02}

 11%|█         | 371/3300 [1:20:03<8:45:26, 10.76s/it]INFO 04-12 17:18:29 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 372/3300 [1:20:13<8:40:00, 10.66s/it]2026-04-12 17:18:40,127 | INFO | train_grpo_train | metrics_logged mode=train step=372

                                                      
{'loss': 0.1003, 'grad_norm': 12.382516860961914, 'learning_rate': 8.875757575757576e-06, 'num_tokens': 698989.0, 'completions/mean_length': 84.25, 'completions/min_length': 70.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.7467271089553833, 'rewards/meter/std': 0.26991549134254456, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.8503808379173279, 'rewards/repeat_penalty/std': 0.17994295060634613, 'rewards/repeat_floor/mean': 0.9820507168769836, 'rewards/repeat_floor/std': 0.02148190326988697, 'rewards/near_duplicate_penalty/mean': 0.9449970722198486, 'rewards/near_duplicate_penalty/std': 0.053348246961832047, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8974431753158569, 'rewards/distinct_2/std': 0.13926663994789124, 'rewards/total_composite/mean': 0.25923699140548706, 'rewards/total_composite/std': 0.08671414852142334, 'reward': 0.25923699140548706, 'reward_std': 0.08671414852142334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.169866144657135, 'sampling/sampling_logp_difference/max': 1.5025932788848877, 'sampling/importance_sampling_ratio/min': 0.22255228459835052, 'sampling/importance_sampling_ratio/mean': 1.0178216695785522, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5575363039970398, 'clip_ratio/low_mean': 0.04029753431677818, 'clip_ratio/low_min': 0.04029753431677818, 'clip_ratio/high_mean': 0.09936683066189289, 'clip_ratio/high_max': 0.09936683066189289, 'clip_ratio/region_mean': 0.13966436497867107, 'reward_total_mean': 0.25923699140548706, 'reward_meter_mean': 0.7467271089553833, 'reward_meter_std': 0.26991549134254456, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8503808379173279, 'reward_repeat_penalty_std': 0.17994295060634613, 'reward_repeat_floor_mean': 0.9820507168769836, 'reward_repeat_floor_std': 0.02148190326988697, 'reward_near_duplicate_penalty_mean': 0.9449970722198486, 'reward_near_duplicate_penalty_std': 0.053348246961832047, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8974431753158569, 'reward_distinct_2_std': 0.13926663994789124, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.25923699140548706, 'reward_total_composite_std': 0.08671414852142334, 'epoch': 0.02}

 11%|█▏        | 372/3300 [1:20:13<8:40:00, 10.66s/it]INFO 04-12 17:18:40 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 373/3300 [1:20:23<8:16:56, 10.19s/it]2026-04-12 17:18:49,218 | INFO | train_grpo_train | metrics_logged mode=train step=373

                                                      
{'loss': 0.0601, 'grad_norm': 15.747854232788086, 'learning_rate': 8.872727272727275e-06, 'num_tokens': 700546.0, 'completions/mean_length': 39.625, 'completions/min_length': 33.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7566812038421631, 'rewards/meter/std': 0.29804617166519165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9986746907234192, 'rewards/repeat_penalty/std': 0.0019098890479654074, 'rewards/repeat_floor/mean': 0.9998011589050293, 'rewards/repeat_floor/std': 0.00028648527222685516, 'rewards/near_duplicate_penalty/mean': 0.9986746907234192, 'rewards/near_duplicate_penalty/std': 0.0019098890479654074, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2892093062400818, 'rewards/total_composite/std': 0.15111269056797028, 'reward': 0.2892093062400818, 'reward_std': 0.15111269056797028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2076290398836136, 'sampling/sampling_logp_difference/max': 1.383375644683838, 'sampling/importance_sampling_ratio/min': 0.2507307529449463, 'sampling/importance_sampling_ratio/mean': 1.050705075263977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0545389354228973, 'clip_ratio/low_mean': 0.054423414170742035, 'clip_ratio/low_min': 0.054423414170742035, 'clip_ratio/high_mean': 0.13461116328835487, 'clip_ratio/high_max': 0.13461116328835487, 'clip_ratio/region_mean': 0.1890345774590969, 'reward_total_mean': 0.2892093062400818, 'reward_meter_mean': 0.7566812038421631, 'reward_meter_std': 0.29804617166519165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9986746907234192, 'reward_repeat_penalty_std': 0.0019098890479654074, 'reward_repeat_floor_mean': 0.9998011589050293, 'reward_repeat_floor_std': 0.00028648527222685516, 'reward_near_duplicate_penalty_mean': 0.9986746907234192, 'reward_near_duplicate_penalty_std': 0.0019098890479654074, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.2892093062400818, 'reward_total_composite_std': 0.15111269056797028, 'epoch': 0.02}

 11%|█▏        | 373/3300 [1:20:23<8:16:56, 10.19s/it]INFO 04-12 17:18:49 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 374/3300 [1:20:31<7:50:17,  9.64s/it]2026-04-12 17:18:57,593 | INFO | train_grpo_train | metrics_logged mode=train step=374

                                                      
{'loss': 0.076, 'grad_norm': 18.713809967041016, 'learning_rate': 8.86969696969697e-06, 'num_tokens': 701973.0, 'completions/mean_length': 34.375, 'completions/min_length': 25.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8340343236923218, 'rewards/meter/std': 0.3362620770931244, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.612500011920929, 'rewards/judge_quality/std': 0.2548248767852783, 'rewards/repeat_penalty/mean': 0.9939566254615784, 'rewards/repeat_penalty/std': 0.009147074073553085, 'rewards/repeat_floor/mean': 0.9990935325622559, 'rewards/repeat_floor/std': 0.001372060040012002, 'rewards/near_duplicate_penalty/mean': 0.9939566254615784, 'rewards/near_duplicate_penalty/std': 0.009147074073553085, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45364952087402344, 'rewards/total_composite/std': 0.2479201704263687, 'reward': 0.45364952087402344, 'reward_std': 0.2479201704263687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16681362688541412, 'sampling/sampling_logp_difference/max': 1.346923589706421, 'sampling/importance_sampling_ratio/min': 0.2600390315055847, 'sampling/importance_sampling_ratio/mean': 1.0014628171920776, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.135002762079239, 'clip_ratio/low_mean': 0.09099704306572676, 'clip_ratio/low_min': 0.09099704306572676, 'clip_ratio/high_mean': 0.037361111491918564, 'clip_ratio/high_max': 0.037361111491918564, 'clip_ratio/region_mean': 0.12835815455764532, 'reward_total_mean': 0.45364952087402344, 'reward_meter_mean': 0.8340343236923218, 'reward_meter_std': 0.3362620770931244, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9939566254615784, 'reward_repeat_penalty_std': 0.009147074073553085, 'reward_repeat_floor_mean': 0.9990935325622559, 'reward_repeat_floor_std': 0.001372060040012002, 'reward_near_duplicate_penalty_mean': 0.9939566254615784, 'reward_near_duplicate_penalty_std': 0.009147074073553085, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.612500011920929, 'reward_judge_quality_std': 0.2548248767852783, 'reward_total_composite_mean': 0.45364952087402344, 'reward_total_composite_std': 0.2479201704263687, 'epoch': 0.02}

 11%|█▏        | 374/3300 [1:20:31<7:50:17,  9.64s/it]INFO 04-12 17:18:57 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 375/3300 [1:20:42<8:15:08, 10.16s/it]2026-04-12 17:19:08,946 | INFO | train_grpo_train | metrics_logged mode=train step=375

                                                      
{'loss': -0.2165, 'grad_norm': 4.994774341583252, 'learning_rate': 8.866666666666668e-06, 'num_tokens': 705503.0, 'completions/mean_length': 218.25, 'completions/min_length': 10.0, 'completions/max_length': 266.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 218.25, 'completions/min_terminated_length': 10.0, 'completions/max_terminated_length': 266.0, 'rewards/meter/mean': 0.5767972469329834, 'rewards/meter/std': 0.3385557532310486, 'rewards/count_adherence/mean': 0.6499999761581421, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8949999809265137, 'rewards/count_floor/std': 0.08017838001251221, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.7064532041549683, 'rewards/repeat_penalty/std': 0.2740466594696045, 'rewards/repeat_floor/mean': 0.9698563814163208, 'rewards/repeat_floor/std': 0.028440747410058975, 'rewards/near_duplicate_penalty/mean': 0.9573919773101807, 'rewards/near_duplicate_penalty/std': 0.03063259646296501, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7360846400260925, 'rewards/distinct_2/std': 0.2672129273414612, 'rewards/total_composite/mean': 0.15860283374786377, 'rewards/total_composite/std': 0.0946909487247467, 'reward': 0.15860283374786377, 'reward_std': 0.0946909487247467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16686995327472687, 'sampling/sampling_logp_difference/max': 2.11104154586792, 'sampling/importance_sampling_ratio/min': 0.12111175805330276, 'sampling/importance_sampling_ratio/mean': 1.0233567953109741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5053788647055626, 'clip_ratio/low_mean': 0.09331616386771202, 'clip_ratio/low_min': 0.09331616386771202, 'clip_ratio/high_mean': 0.07826313748955727, 'clip_ratio/high_max': 0.07826313748955727, 'clip_ratio/region_mean': 0.1715793013572693, 'reward_total_mean': 0.15860283374786377, 'reward_meter_mean': 0.5767972469329834, 'reward_meter_std': 0.3385557532310486, 'reward_count_adherence_mean': 0.6499999761581421, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8949999809265137, 'reward_count_floor_std': 0.08017838001251221, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7064532041549683, 'reward_repeat_penalty_std': 0.2740466594696045, 'reward_repeat_floor_mean': 0.9698563814163208, 'reward_repeat_floor_std': 0.028440747410058975, 'reward_near_duplicate_penalty_mean': 0.9573919773101807, 'reward_near_duplicate_penalty_std': 0.03063259646296501, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7360846400260925, 'reward_distinct_2_std': 0.2672129273414612, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.15860283374786377, 'reward_total_composite_std': 0.0946909487247467, 'epoch': 0.02}

 11%|█▏        | 375/3300 [1:20:42<8:15:08, 10.16s/it]INFO 04-12 17:19:09 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 376/3300 [1:20:51<7:55:39,  9.76s/it]2026-04-12 17:19:17,784 | INFO | train_grpo_train | metrics_logged mode=train step=376

                                                      
{'loss': -0.0054, 'grad_norm': 27.141374588012695, 'learning_rate': 8.863636363636365e-06, 'num_tokens': 707178.0, 'completions/mean_length': 44.375, 'completions/min_length': 34.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8870347738265991, 'rewards/meter/std': 0.18807987868785858, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6474999785423279, 'rewards/judge_quality/std': 0.2931479811668396, 'rewards/repeat_penalty/mean': 0.9769245386123657, 'rewards/repeat_penalty/std': 0.03607640415430069, 'rewards/repeat_floor/mean': 0.9973078966140747, 'rewards/repeat_floor/std': 0.0036129853688180447, 'rewards/near_duplicate_penalty/mean': 0.9897450804710388, 'rewards/near_duplicate_penalty/std': 0.016943814232945442, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.5531960725784302, 'rewards/total_composite/std': 0.2840522527694702, 'reward': 0.5531960725784302, 'reward_std': 0.2840522527694702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14562341570854187, 'sampling/sampling_logp_difference/max': 2.409628391265869, 'sampling/importance_sampling_ratio/min': 0.08984867483377457, 'sampling/importance_sampling_ratio/mean': 0.9764859676361084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.561041820794344, 'clip_ratio/low_mean': 0.06508746277540922, 'clip_ratio/low_min': 0.06508746277540922, 'clip_ratio/high_mean': 0.04916320415213704, 'clip_ratio/high_max': 0.04916320415213704, 'clip_ratio/region_mean': 0.11425066692754626, 'reward_total_mean': 0.5531960725784302, 'reward_meter_mean': 0.8870347738265991, 'reward_meter_std': 0.18807987868785858, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9769245386123657, 'reward_repeat_penalty_std': 0.03607640415430069, 'reward_repeat_floor_mean': 0.9973078966140747, 'reward_repeat_floor_std': 0.0036129853688180447, 'reward_near_duplicate_penalty_mean': 0.9897450804710388, 'reward_near_duplicate_penalty_std': 0.016943814232945442, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.6474999785423279, 'reward_judge_quality_std': 0.2931479811668396, 'reward_total_composite_mean': 0.5531960725784302, 'reward_total_composite_std': 0.2840522527694702, 'epoch': 0.02}

 11%|█▏        | 376/3300 [1:20:51<7:55:39,  9.76s/it]INFO 04-12 17:19:18 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 377/3300 [1:21:00<7:42:53,  9.50s/it]2026-04-12 17:19:26,685 | INFO | train_grpo_train | metrics_logged mode=train step=377

                                                      
{'loss': 0.046, 'grad_norm': 14.02459716796875, 'learning_rate': 8.860606060606062e-06, 'num_tokens': 708931.0, 'completions/mean_length': 45.125, 'completions/min_length': 40.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.728223979473114, 'rewards/meter/std': 0.37204641103744507, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.8173435926437378, 'rewards/repeat_penalty/std': 0.29813843965530396, 'rewards/repeat_floor/mean': 0.9741727113723755, 'rewards/repeat_floor/std': 0.044281184673309326, 'rewards/near_duplicate_penalty/mean': 0.9131169319152832, 'rewards/near_duplicate_penalty/std': 0.1279606819152832, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.909919023513794, 'rewards/distinct_2/std': 0.1458413451910019, 'rewards/total_composite/mean': 0.2686946988105774, 'rewards/total_composite/std': 0.28365370631217957, 'reward': 0.2686946988105774, 'reward_std': 0.2836536765098572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17564965784549713, 'sampling/sampling_logp_difference/max': 1.5391759872436523, 'sampling/importance_sampling_ratio/min': 0.21455782651901245, 'sampling/importance_sampling_ratio/mean': 0.9972003102302551, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4582325965166092, 'clip_ratio/low_mean': 0.07083836337551475, 'clip_ratio/low_min': 0.07083836337551475, 'clip_ratio/high_mean': 0.07301227748394012, 'clip_ratio/high_max': 0.07301227748394012, 'clip_ratio/region_mean': 0.14385064085945487, 'reward_total_mean': 0.2686946988105774, 'reward_meter_mean': 0.728223979473114, 'reward_meter_std': 0.37204641103744507, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8173435926437378, 'reward_repeat_penalty_std': 0.29813843965530396, 'reward_repeat_floor_mean': 0.9741727113723755, 'reward_repeat_floor_std': 0.044281184673309326, 'reward_near_duplicate_penalty_mean': 0.9131169319152832, 'reward_near_duplicate_penalty_std': 0.1279606819152832, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.909919023513794, 'reward_distinct_2_std': 0.1458413451910019, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.2686946988105774, 'reward_total_composite_std': 0.28365370631217957, 'epoch': 0.02}

 11%|█▏        | 377/3300 [1:21:00<7:42:53,  9.50s/it]INFO 04-12 17:19:26 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 378/3300 [1:21:08<7:25:26,  9.15s/it]2026-04-12 17:19:35,001 | INFO | train_grpo_train | metrics_logged mode=train step=378

                                                      
{'loss': 0.1539, 'grad_norm': 23.855758666992188, 'learning_rate': 8.857575757575758e-06, 'num_tokens': 710439.0, 'completions/mean_length': 30.5, 'completions/min_length': 20.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7858266830444336, 'rewards/meter/std': 0.3465760350227356, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7849999666213989, 'rewards/judge_quality/std': 0.2157379686832428, 'rewards/repeat_penalty/mean': 0.800424337387085, 'rewards/repeat_penalty/std': 0.12245331704616547, 'rewards/repeat_floor/mean': 0.9863779544830322, 'rewards/repeat_floor/std': 0.008887514472007751, 'rewards/near_duplicate_penalty/mean': 0.9841861128807068, 'rewards/near_duplicate_penalty/std': 0.02570686861872673, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6044842600822449, 'rewards/total_composite/std': 0.3213955760002136, 'reward': 0.6044842600822449, 'reward_std': 0.3213955760002136, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18083840608596802, 'sampling/sampling_logp_difference/max': 2.0914127826690674, 'sampling/importance_sampling_ratio/min': 0.12351251393556595, 'sampling/importance_sampling_ratio/mean': 1.051404595375061, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4470215737819672, 'clip_ratio/low_mean': 0.06036193436011672, 'clip_ratio/low_min': 0.06036193436011672, 'clip_ratio/high_mean': 0.07732909638434649, 'clip_ratio/high_max': 0.07732909638434649, 'clip_ratio/region_mean': 0.1376910307444632, 'reward_total_mean': 0.6044842600822449, 'reward_meter_mean': 0.7858266830444336, 'reward_meter_std': 0.3465760350227356, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.800424337387085, 'reward_repeat_penalty_std': 0.12245331704616547, 'reward_repeat_floor_mean': 0.9863779544830322, 'reward_repeat_floor_std': 0.008887514472007751, 'reward_near_duplicate_penalty_mean': 0.9841861128807068, 'reward_near_duplicate_penalty_std': 0.02570686861872673, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7849999666213989, 'reward_judge_quality_std': 0.2157379686832428, 'reward_total_composite_mean': 0.6044842600822449, 'reward_total_composite_std': 0.3213955760002136, 'epoch': 0.02}

 11%|█▏        | 378/3300 [1:21:08<7:25:26,  9.15s/it]INFO 04-12 17:19:35 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 379/3300 [1:21:18<7:26:48,  9.18s/it]2026-04-12 17:19:44,253 | INFO | train_grpo_train | metrics_logged mode=train step=379

                                                      
{'loss': 0.1904, 'grad_norm': 19.44912338256836, 'learning_rate': 8.854545454545455e-06, 'num_tokens': 712276.0, 'completions/mean_length': 60.625, 'completions/min_length': 29.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.41483208537101746, 'rewards/meter/std': 0.3286876380443573, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690448433160782, 'rewards/repeat_penalty/mean': 0.943665623664856, 'rewards/repeat_penalty/std': 0.08658132702112198, 'rewards/repeat_floor/mean': 0.9950630068778992, 'rewards/repeat_floor/std': 0.005955915432423353, 'rewards/near_duplicate_penalty/mean': 0.9872875809669495, 'rewards/near_duplicate_penalty/std': 0.017184389755129814, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9871652126312256, 'rewards/distinct_2/std': 0.02378297969698906, 'rewards/total_composite/mean': 0.1338021457195282, 'rewards/total_composite/std': 0.08030419796705246, 'reward': 0.1338021457195282, 'reward_std': 0.08030419051647186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2164975255727768, 'sampling/sampling_logp_difference/max': 2.004180431365967, 'sampling/importance_sampling_ratio/min': 0.1347707062959671, 'sampling/importance_sampling_ratio/mean': 1.0033055543899536, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.304855465888977, 'clip_ratio/low_mean': 0.08339879103004932, 'clip_ratio/low_min': 0.08339879103004932, 'clip_ratio/high_mean': 0.10393539629876614, 'clip_ratio/high_max': 0.10393539629876614, 'clip_ratio/region_mean': 0.18733418732881546, 'reward_total_mean': 0.1338021457195282, 'reward_meter_mean': 0.41483208537101746, 'reward_meter_std': 0.3286876380443573, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.943665623664856, 'reward_repeat_penalty_std': 0.08658132702112198, 'reward_repeat_floor_mean': 0.9950630068778992, 'reward_repeat_floor_std': 0.005955915432423353, 'reward_near_duplicate_penalty_mean': 0.9872875809669495, 'reward_near_duplicate_penalty_std': 0.017184389755129814, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9871652126312256, 'reward_distinct_2_std': 0.02378297969698906, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690448433160782, 'reward_total_composite_mean': 0.1338021457195282, 'reward_total_composite_std': 0.08030419796705246, 'epoch': 0.02}

 11%|█▏        | 379/3300 [1:21:18<7:26:48,  9.18s/it]INFO 04-12 17:19:44 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 380/3300 [1:21:27<7:24:42,  9.14s/it]2026-04-12 17:19:53,296 | INFO | train_grpo_train | metrics_logged mode=train step=380

                                                      
{'loss': 0.1374, 'grad_norm': 13.794597625732422, 'learning_rate': 8.851515151515152e-06, 'num_tokens': 713928.0, 'completions/mean_length': 45.5, 'completions/min_length': 36.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8509638905525208, 'rewards/meter/std': 0.2640708088874817, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8708540201187134, 'rewards/lexical_plausibility/std': 0.2625297009944916, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.25795555114746094, 'rewards/repeat_penalty/mean': 0.9555281400680542, 'rewards/repeat_penalty/std': 0.059608519077301025, 'rewards/repeat_floor/mean': 0.9946416020393372, 'rewards/repeat_floor/std': 0.007120018359273672, 'rewards/near_duplicate_penalty/mean': 0.9793872833251953, 'rewards/near_duplicate_penalty/std': 0.026510024443268776, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9748168587684631, 'rewards/distinct_2/std': 0.03521263226866722, 'rewards/total_composite/mean': 0.34486454725265503, 'rewards/total_composite/std': 0.2494838833808899, 'reward': 0.34486454725265503, 'reward_std': 0.2494838833808899, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1943255513906479, 'sampling/sampling_logp_difference/max': 1.334388256072998, 'sampling/importance_sampling_ratio/min': 0.2633192241191864, 'sampling/importance_sampling_ratio/mean': 1.0238460302352905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7199637442827225, 'clip_ratio/low_mean': 0.09679704904556274, 'clip_ratio/low_min': 0.09679704904556274, 'clip_ratio/high_mean': 0.09106815978884697, 'clip_ratio/high_max': 0.09106815978884697, 'clip_ratio/region_mean': 0.1878652088344097, 'reward_total_mean': 0.34486454725265503, 'reward_meter_mean': 0.8509638905525208, 'reward_meter_std': 0.2640708088874817, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8708540201187134, 'reward_lexical_plausibility_std': 0.2625297009944916, 'reward_repeat_penalty_mean': 0.9555281400680542, 'reward_repeat_penalty_std': 0.059608519077301025, 'reward_repeat_floor_mean': 0.9946416020393372, 'reward_repeat_floor_std': 0.007120018359273672, 'reward_near_duplicate_penalty_mean': 0.9793872833251953, 'reward_near_duplicate_penalty_std': 0.026510024443268776, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9748168587684631, 'reward_distinct_2_std': 0.03521263226866722, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.25795555114746094, 'reward_total_composite_mean': 0.34486454725265503, 'reward_total_composite_std': 0.2494838833808899, 'epoch': 0.02}

 12%|█▏        | 380/3300 [1:21:27<7:24:42,  9.14s/it]INFO 04-12 17:19:53 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 381/3300 [1:21:35<7:17:07,  8.99s/it]2026-04-12 17:20:01,923 | INFO | train_grpo_train | metrics_logged mode=train step=381

                                                      
{'loss': 0.0828, 'grad_norm': 20.097248077392578, 'learning_rate': 8.84848484848485e-06, 'num_tokens': 715643.0, 'completions/mean_length': 49.375, 'completions/min_length': 45.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7343670129776001, 'rewards/meter/std': 0.3419080078601837, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9616450071334839, 'rewards/lexical_plausibility/std': 0.07202228903770447, 'rewards/judge_quality/mean': 0.5725000500679016, 'rewards/judge_quality/std': 0.3767435848712921, 'rewards/repeat_penalty/mean': 0.9805794954299927, 'rewards/repeat_penalty/std': 0.013357391580939293, 'rewards/repeat_floor/mean': 0.9970868825912476, 'rewards/repeat_floor/std': 0.00200360594317317, 'rewards/near_duplicate_penalty/mean': 0.9805794954299927, 'rewards/near_duplicate_penalty/std': 0.013357391580939293, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49548688530921936, 'rewards/total_composite/std': 0.38653960824012756, 'reward': 0.49548688530921936, 'reward_std': 0.38653963804244995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20156219601631165, 'sampling/sampling_logp_difference/max': 2.3564653396606445, 'sampling/importance_sampling_ratio/min': 0.09475455433130264, 'sampling/importance_sampling_ratio/mean': 1.0000542402267456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0622112601995468, 'clip_ratio/low_mean': 0.10730747692286968, 'clip_ratio/low_min': 0.10730747692286968, 'clip_ratio/high_mean': 0.0692435372620821, 'clip_ratio/high_max': 0.0692435372620821, 'clip_ratio/region_mean': 0.17655101418495178, 'reward_total_mean': 0.49548688530921936, 'reward_meter_mean': 0.7343670129776001, 'reward_meter_std': 0.3419080078601837, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9616450071334839, 'reward_lexical_plausibility_std': 0.07202228903770447, 'reward_repeat_penalty_mean': 0.9805794954299927, 'reward_repeat_penalty_std': 0.013357391580939293, 'reward_repeat_floor_mean': 0.9970868825912476, 'reward_repeat_floor_std': 0.00200360594317317, 'reward_near_duplicate_penalty_mean': 0.9805794954299927, 'reward_near_duplicate_penalty_std': 0.013357391580939293, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5725000500679016, 'reward_judge_quality_std': 0.3767435848712921, 'reward_total_composite_mean': 0.49548688530921936, 'reward_total_composite_std': 0.38653960824012756, 'epoch': 0.02}

 12%|█▏        | 381/3300 [1:21:35<7:17:07,  8.99s/it]INFO 04-12 17:20:02 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 382/3300 [1:21:43<7:02:55,  8.70s/it]2026-04-12 17:20:09,950 | INFO | train_grpo_train | metrics_logged mode=train step=382

                                                      
{'loss': 0.1009, 'grad_norm': 16.714942932128906, 'learning_rate': 8.845454545454547e-06, 'num_tokens': 717150.0, 'completions/mean_length': 28.375, 'completions/min_length': 23.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8646761178970337, 'rewards/meter/std': 0.3436548709869385, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.2522328794002533, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.1157275140285492, 'rewards/repeat_floor/mean': 0.9887499809265137, 'rewards/repeat_floor/std': 0.0069436440244317055, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4472775459289551, 'rewards/total_composite/std': 0.30153030157089233, 'reward': 0.4472775459289551, 'reward_std': 0.30153030157089233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13552485406398773, 'sampling/sampling_logp_difference/max': 1.2284979820251465, 'sampling/importance_sampling_ratio/min': 0.292731910943985, 'sampling/importance_sampling_ratio/mean': 1.0380221605300903, 'sampling/importance_sampling_ratio/max': 1.8198275566101074, 'entropy': 0.9642823562026024, 'clip_ratio/low_mean': 0.1073402687907219, 'clip_ratio/low_min': 0.1073402687907219, 'clip_ratio/high_mean': 0.04222408030182123, 'clip_ratio/high_max': 0.04222408030182123, 'clip_ratio/region_mean': 0.14956434909254313, 'reward_total_mean': 0.4472775459289551, 'reward_meter_mean': 0.8646761178970337, 'reward_meter_std': 0.3436548709869385, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.1157275140285492, 'reward_repeat_floor_mean': 0.9887499809265137, 'reward_repeat_floor_std': 0.0069436440244317055, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.2522328794002533, 'reward_total_composite_mean': 0.4472775459289551, 'reward_total_composite_std': 0.30153030157089233, 'epoch': 0.02}

 12%|█▏        | 382/3300 [1:21:43<7:02:55,  8.70s/it]INFO 04-12 17:20:10 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 383/3300 [1:21:51<6:55:55,  8.56s/it]2026-04-12 17:20:18,175 | INFO | train_grpo_train | metrics_logged mode=train step=383

                                                      
{'loss': 0.1194, 'grad_norm': 13.620659828186035, 'learning_rate': 8.842424242424244e-06, 'num_tokens': 718727.0, 'completions/mean_length': 47.125, 'completions/min_length': 39.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7985309958457947, 'rewards/meter/std': 0.23880243301391602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9678767323493958, 'rewards/repeat_penalty/std': 0.047019246965646744, 'rewards/repeat_floor/mean': 0.995730996131897, 'rewards/repeat_floor/std': 0.006663989741355181, 'rewards/near_duplicate_penalty/mean': 0.9770342111587524, 'rewards/near_duplicate_penalty/std': 0.044951487332582474, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.3589385449886322, 'rewards/total_composite/std': 0.2025938332080841, 'reward': 0.3589385449886322, 'reward_std': 0.2025938332080841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16942140460014343, 'sampling/sampling_logp_difference/max': 1.446589469909668, 'sampling/importance_sampling_ratio/min': 0.2353716641664505, 'sampling/importance_sampling_ratio/mean': 1.0394906997680664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.68963922560215, 'clip_ratio/low_mean': 0.07068293169140816, 'clip_ratio/low_min': 0.07068293169140816, 'clip_ratio/high_mean': 0.04872140567749739, 'clip_ratio/high_max': 0.04872140567749739, 'clip_ratio/region_mean': 0.11940433736890554, 'reward_total_mean': 0.3589385449886322, 'reward_meter_mean': 0.7985309958457947, 'reward_meter_std': 0.23880243301391602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9678767323493958, 'reward_repeat_penalty_std': 0.047019246965646744, 'reward_repeat_floor_mean': 0.995730996131897, 'reward_repeat_floor_std': 0.006663989741355181, 'reward_near_duplicate_penalty_mean': 0.9770342111587524, 'reward_near_duplicate_penalty_std': 0.044951487332582474, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.3589385449886322, 'reward_total_composite_std': 0.2025938332080841, 'epoch': 0.02}

 12%|█▏        | 383/3300 [1:21:52<6:55:55,  8.56s/it]INFO 04-12 17:20:18 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 384/3300 [1:22:05<8:11:43, 10.12s/it]2026-04-12 17:20:31,938 | INFO | train_grpo_train | metrics_logged mode=train step=384

                                                      
{'loss': -0.0339, 'grad_norm': 2.7382514476776123, 'learning_rate': 8.83939393939394e-06, 'num_tokens': 720291.0, 'completions/mean_length': 228.5, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 58.400001525878906, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.41184407472610474, 'rewards/meter/std': 0.35498902201652527, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.7898761034011841, 'rewards/lexical_plausibility/std': 0.22936077415943146, 'rewards/judge_quality/mean': 0.27125000953674316, 'rewards/judge_quality/std': 0.2916180193424225, 'rewards/repeat_penalty/mean': 0.9973608255386353, 'rewards/repeat_penalty/std': 0.00566133763641119, 'rewards/repeat_floor/mean': 0.9996041059494019, 'rewards/repeat_floor/std': 0.000849198957439512, 'rewards/near_duplicate_penalty/mean': 0.9973608255386353, 'rewards/near_duplicate_penalty/std': 0.00566133763641119, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.09623901546001434, 'rewards/total_composite/std': 0.11577337980270386, 'reward': 0.09623901546001434, 'reward_std': 0.11577337980270386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1755877137184143, 'sampling/sampling_logp_difference/max': 1.018585205078125, 'sampling/importance_sampling_ratio/min': 0.36110547184944153, 'sampling/importance_sampling_ratio/mean': 1.0711219310760498, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5550226420164108, 'clip_ratio/low_mean': 0.027036355808377266, 'clip_ratio/low_min': 0.027036355808377266, 'clip_ratio/high_mean': 0.07140899263322353, 'clip_ratio/high_max': 0.07140899263322353, 'clip_ratio/region_mean': 0.0984453484416008, 'reward_total_mean': 0.09623901546001434, 'reward_meter_mean': 0.41184407472610474, 'reward_meter_std': 0.35498902201652527, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.7898761034011841, 'reward_lexical_plausibility_std': 0.22936077415943146, 'reward_repeat_penalty_mean': 0.9973608255386353, 'reward_repeat_penalty_std': 0.00566133763641119, 'reward_repeat_floor_mean': 0.9996041059494019, 'reward_repeat_floor_std': 0.000849198957439512, 'reward_near_duplicate_penalty_mean': 0.9973608255386353, 'reward_near_duplicate_penalty_std': 0.00566133763641119, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.27125000953674316, 'reward_judge_quality_std': 0.2916180193424225, 'reward_total_composite_mean': 0.09623901546001434, 'reward_total_composite_std': 0.11577337980270386, 'epoch': 0.02}

 12%|█▏        | 384/3300 [1:22:05<8:11:43, 10.12s/it]INFO 04-12 17:20:32 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 385/3300 [1:22:15<7:59:57,  9.88s/it]2026-04-12 17:20:41,260 | INFO | train_grpo_train | metrics_logged mode=train step=385

                                                      
{'loss': 0.077, 'grad_norm': 13.970919609069824, 'learning_rate': 8.836363636363637e-06, 'num_tokens': 721996.0, 'completions/mean_length': 48.125, 'completions/min_length': 39.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7287188768386841, 'rewards/meter/std': 0.2855376601219177, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.9306275248527527, 'rewards/repeat_penalty/std': 0.11843213438987732, 'rewards/repeat_floor/mean': 0.9910269975662231, 'rewards/repeat_floor/std': 0.01477543544024229, 'rewards/near_duplicate_penalty/mean': 0.9616802930831909, 'rewards/near_duplicate_penalty/std': 0.053322263062000275, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9641662836074829, 'rewards/distinct_2/std': 0.07792406529188156, 'rewards/total_composite/mean': 0.2882566452026367, 'rewards/total_composite/std': 0.19672976434230804, 'reward': 0.2882566452026367, 'reward_std': 0.19672976434230804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19441959261894226, 'sampling/sampling_logp_difference/max': 1.3446664810180664, 'sampling/importance_sampling_ratio/min': 0.2606266438961029, 'sampling/importance_sampling_ratio/mean': 1.051590085029602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9694604128599167, 'clip_ratio/low_mean': 0.09881034214049578, 'clip_ratio/low_min': 0.09881034214049578, 'clip_ratio/high_mean': 0.08061224780976772, 'clip_ratio/high_max': 0.08061224780976772, 'clip_ratio/region_mean': 0.1794225899502635, 'reward_total_mean': 0.2882566452026367, 'reward_meter_mean': 0.7287188768386841, 'reward_meter_std': 0.2855376601219177, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9306275248527527, 'reward_repeat_penalty_std': 0.11843213438987732, 'reward_repeat_floor_mean': 0.9910269975662231, 'reward_repeat_floor_std': 0.01477543544024229, 'reward_near_duplicate_penalty_mean': 0.9616802930831909, 'reward_near_duplicate_penalty_std': 0.053322263062000275, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9641662836074829, 'reward_distinct_2_std': 0.07792406529188156, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.2882566452026367, 'reward_total_composite_std': 0.19672976434230804, 'epoch': 0.02}

 12%|█▏        | 385/3300 [1:22:15<7:59:57,  9.88s/it]INFO 04-12 17:20:41 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 386/3300 [1:22:28<8:56:56, 11.06s/it]2026-04-12 17:20:55,060 | INFO | train_grpo_train | metrics_logged mode=train step=386

                                                      
{'loss': -0.0557, 'grad_norm': 5.858519554138184, 'learning_rate': 8.833333333333334e-06, 'num_tokens': 723502.0, 'completions/mean_length': 109.25, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6169261932373047, 'rewards/meter/std': 0.41295716166496277, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9399685859680176, 'rewards/lexical_plausibility/std': 0.16979444026947021, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9202920198440552, 'rewards/repeat_penalty/std': 0.07219918817281723, 'rewards/repeat_floor/mean': 0.991070568561554, 'rewards/repeat_floor/std': 0.007806537672877312, 'rewards/near_duplicate_penalty/mean': 0.9741638898849487, 'rewards/near_duplicate_penalty/std': 0.018550660461187363, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9438443183898926, 'rewards/distinct_2/std': 0.05826306343078613, 'rewards/total_composite/mean': 0.17364296317100525, 'rewards/total_composite/std': 0.1408962458372116, 'reward': 0.17364296317100525, 'reward_std': 0.14089623093605042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22827568650245667, 'sampling/sampling_logp_difference/max': 2.9266014099121094, 'sampling/importance_sampling_ratio/min': 0.05357882380485535, 'sampling/importance_sampling_ratio/mean': 1.0259442329406738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8930927515029907, 'clip_ratio/low_mean': 0.07026487775146961, 'clip_ratio/low_min': 0.07026487775146961, 'clip_ratio/high_mean': 0.08794041536748409, 'clip_ratio/high_max': 0.08794041536748409, 'clip_ratio/region_mean': 0.1582052931189537, 'reward_total_mean': 0.17364296317100525, 'reward_meter_mean': 0.6169261932373047, 'reward_meter_std': 0.41295716166496277, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9399685859680176, 'reward_lexical_plausibility_std': 0.16979444026947021, 'reward_repeat_penalty_mean': 0.9202920198440552, 'reward_repeat_penalty_std': 0.07219918817281723, 'reward_repeat_floor_mean': 0.991070568561554, 'reward_repeat_floor_std': 0.007806537672877312, 'reward_near_duplicate_penalty_mean': 0.9741638898849487, 'reward_near_duplicate_penalty_std': 0.018550660461187363, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9438443183898926, 'reward_distinct_2_std': 0.05826306343078613, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.17364296317100525, 'reward_total_composite_std': 0.1408962458372116, 'epoch': 0.02}

 12%|█▏        | 386/3300 [1:22:28<8:56:56, 11.06s/it]INFO 04-12 17:20:55 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 387/3300 [1:22:44<9:59:18, 12.34s/it]2026-04-12 17:21:10,412 | INFO | train_grpo_train | metrics_logged mode=train step=387

                                                      
{'loss': -0.1257, 'grad_norm': 3.3730335235595703, 'learning_rate': 8.830303030303031e-06, 'num_tokens': 725265.0, 'completions/mean_length': 183.375, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 73.83333587646484, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.640785276889801, 'rewards/meter/std': 0.40351882576942444, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 0.9212695956230164, 'rewards/lexical_plausibility/std': 0.14579758048057556, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.9264574646949768, 'rewards/repeat_penalty/std': 0.06898912787437439, 'rewards/repeat_floor/mean': 0.9922683238983154, 'rewards/repeat_floor/std': 0.0059776995331048965, 'rewards/near_duplicate_penalty/mean': 0.9750974178314209, 'rewards/near_duplicate_penalty/std': 0.017016621306538582, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.967501163482666, 'rewards/distinct_2/std': 0.05004063621163368, 'rewards/total_composite/mean': 0.20359057188034058, 'rewards/total_composite/std': 0.1862461119890213, 'reward': 0.20359057188034058, 'reward_std': 0.1862461119890213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18108008801937103, 'sampling/sampling_logp_difference/max': 1.8801088333129883, 'sampling/importance_sampling_ratio/min': 0.15257351100444794, 'sampling/importance_sampling_ratio/mean': 1.0532912015914917, 'sampling/importance_sampling_ratio/max': 1.9593380689620972, 'entropy': 1.5311155915260315, 'clip_ratio/low_mean': 0.0421570036560297, 'clip_ratio/low_min': 0.0421570036560297, 'clip_ratio/high_mean': 0.08029082231223583, 'clip_ratio/high_max': 0.08029082231223583, 'clip_ratio/region_mean': 0.12244782596826553, 'reward_total_mean': 0.20359057188034058, 'reward_meter_mean': 0.640785276889801, 'reward_meter_std': 0.40351882576942444, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 0.9212695956230164, 'reward_lexical_plausibility_std': 0.14579758048057556, 'reward_repeat_penalty_mean': 0.9264574646949768, 'reward_repeat_penalty_std': 0.06898912787437439, 'reward_repeat_floor_mean': 0.9922683238983154, 'reward_repeat_floor_std': 0.0059776995331048965, 'reward_near_duplicate_penalty_mean': 0.9750974178314209, 'reward_near_duplicate_penalty_std': 0.017016621306538582, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.967501163482666, 'reward_distinct_2_std': 0.05004063621163368, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.20359057188034058, 'reward_total_composite_std': 0.1862461119890213, 'epoch': 0.02}

 12%|█▏        | 387/3300 [1:22:44<9:59:18, 12.34s/it]INFO 04-12 17:21:10 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 388/3300 [1:22:59<10:45:31, 13.30s/it]2026-04-12 17:21:25,943 | INFO | train_grpo_train | metrics_logged mode=train step=388

                                                       
{'loss': -0.0383, 'grad_norm': 2.712648391723633, 'learning_rate': 8.827272727272727e-06, 'num_tokens': 726869.0, 'completions/mean_length': 228.5, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 58.400001525878906, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.6996219158172607, 'rewards/meter/std': 0.37979835271835327, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.8547126054763794, 'rewards/lexical_plausibility/std': 0.2040940821170807, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1927248239517212, 'rewards/repeat_penalty/mean': 0.9994698762893677, 'rewards/repeat_penalty/std': 0.0012265866389498115, 'rewards/repeat_floor/mean': 0.9999204874038696, 'rewards/repeat_floor/std': 0.00018398174142930657, 'rewards/near_duplicate_penalty/mean': 0.9994698762893677, 'rewards/near_duplicate_penalty/std': 0.0012265866389498115, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15015628933906555, 'rewards/total_composite/std': 0.15992803871631622, 'reward': 0.15015628933906555, 'reward_std': 0.15992803871631622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20559276640415192, 'sampling/sampling_logp_difference/max': 1.097541093826294, 'sampling/importance_sampling_ratio/min': 0.3336905837059021, 'sampling/importance_sampling_ratio/mean': 1.0503267049789429, 'sampling/importance_sampling_ratio/max': 1.9009308815002441, 'entropy': 1.6116873621940613, 'clip_ratio/low_mean': 0.054532794281840324, 'clip_ratio/low_min': 0.054532794281840324, 'clip_ratio/high_mean': 0.022411616519093513, 'clip_ratio/high_max': 0.022411616519093513, 'clip_ratio/region_mean': 0.07694441080093384, 'reward_total_mean': 0.15015628933906555, 'reward_meter_mean': 0.6996219158172607, 'reward_meter_std': 0.37979835271835327, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.8547126054763794, 'reward_lexical_plausibility_std': 0.2040940821170807, 'reward_repeat_penalty_mean': 0.9994698762893677, 'reward_repeat_penalty_std': 0.0012265866389498115, 'reward_repeat_floor_mean': 0.9999204874038696, 'reward_repeat_floor_std': 0.00018398174142930657, 'reward_near_duplicate_penalty_mean': 0.9994698762893677, 'reward_near_duplicate_penalty_std': 0.0012265866389498115, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1927248239517212, 'reward_total_composite_mean': 0.15015628933906555, 'reward_total_composite_std': 0.15992803871631622, 'epoch': 0.02}

 12%|█▏        | 388/3300 [1:22:59<10:45:31, 13.30s/it]INFO 04-12 17:21:26 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 389/3300 [1:23:13<10:55:11, 13.50s/it]2026-04-12 17:21:39,924 | INFO | train_grpo_train | metrics_logged mode=train step=389

                                                       
{'loss': -0.0573, 'grad_norm': 3.1303796768188477, 'learning_rate': 8.824242424242426e-06, 'num_tokens': 728568.0, 'completions/mean_length': 171.375, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 57.833335876464844, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.3571970462799072, 'rewards/meter/std': 0.3356514871120453, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 0.8955612182617188, 'rewards/lexical_plausibility/std': 0.19804127514362335, 'rewards/judge_quality/mean': 0.2837499976158142, 'rewards/judge_quality/std': 0.29271817207336426, 'rewards/repeat_penalty/mean': 0.9397079348564148, 'rewards/repeat_penalty/std': 0.08829059451818466, 'rewards/repeat_floor/mean': 0.9949350357055664, 'rewards/repeat_floor/std': 0.005885038059204817, 'rewards/near_duplicate_penalty/mean': 0.9908865690231323, 'rewards/near_duplicate_penalty/std': 0.009218401275575161, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797453880310059, 'rewards/distinct_2/std': 0.041432034224271774, 'rewards/total_composite/mean': 0.09466156363487244, 'rewards/total_composite/std': 0.09176289290189743, 'reward': 0.09466156363487244, 'reward_std': 0.09176289290189743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20467402040958405, 'sampling/sampling_logp_difference/max': 2.0913610458374023, 'sampling/importance_sampling_ratio/min': 0.12351889908313751, 'sampling/importance_sampling_ratio/mean': 1.0227696895599365, 'sampling/importance_sampling_ratio/max': 1.9052410125732422, 'entropy': 1.5064367949962616, 'clip_ratio/low_mean': 0.04087243415415287, 'clip_ratio/low_min': 0.04087243415415287, 'clip_ratio/high_mean': 0.10740681923925877, 'clip_ratio/high_max': 0.10740681923925877, 'clip_ratio/region_mean': 0.14827925339341164, 'reward_total_mean': 0.09466156363487244, 'reward_meter_mean': 0.3571970462799072, 'reward_meter_std': 0.3356514871120453, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 0.8955612182617188, 'reward_lexical_plausibility_std': 0.19804127514362335, 'reward_repeat_penalty_mean': 0.9397079348564148, 'reward_repeat_penalty_std': 0.08829059451818466, 'reward_repeat_floor_mean': 0.9949350357055664, 'reward_repeat_floor_std': 0.005885038059204817, 'reward_near_duplicate_penalty_mean': 0.9908865690231323, 'reward_near_duplicate_penalty_std': 0.009218401275575161, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797453880310059, 'reward_distinct_2_std': 0.041432034224271774, 'reward_judge_quality_mean': 0.2837499976158142, 'reward_judge_quality_std': 0.29271817207336426, 'reward_total_composite_mean': 0.09466156363487244, 'reward_total_composite_std': 0.09176289290189743, 'epoch': 0.02}

 12%|█▏        | 389/3300 [1:23:13<10:55:11, 13.50s/it]INFO 04-12 17:21:40 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 390/3300 [1:23:28<11:06:50, 13.75s/it]2026-04-12 17:21:54,244 | INFO | train_grpo_train | metrics_logged mode=train step=390

                                                       
{'loss': -0.05, 'grad_norm': 5.144541263580322, 'learning_rate': 8.821212121212121e-06, 'num_tokens': 730251.0, 'completions/mean_length': 102.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.85714340209961, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.4600223898887634, 'rewards/meter/std': 0.42484211921691895, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9141478538513184, 'rewards/lexical_plausibility/std': 0.17315734922885895, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.23868314921855927, 'rewards/repeat_penalty/mean': 0.9831186532974243, 'rewards/repeat_penalty/std': 0.023213526234030724, 'rewards/repeat_floor/mean': 0.9977360963821411, 'rewards/repeat_floor/std': 0.0028702779673039913, 'rewards/near_duplicate_penalty/mean': 0.9877917170524597, 'rewards/near_duplicate_penalty/std': 0.013979591429233551, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.995192289352417, 'rewards/distinct_2/std': 0.013598216697573662, 'rewards/total_composite/mean': 0.2282353937625885, 'rewards/total_composite/std': 0.214994877576828, 'reward': 0.2282353937625885, 'reward_std': 0.214994877576828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.198981374502182, 'sampling/sampling_logp_difference/max': 1.4062249660491943, 'sampling/importance_sampling_ratio/min': 0.24506668746471405, 'sampling/importance_sampling_ratio/mean': 1.0506387948989868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3851213604211807, 'clip_ratio/low_mean': 0.056114123202860355, 'clip_ratio/low_min': 0.056114123202860355, 'clip_ratio/high_mean': 0.09040492400527, 'clip_ratio/high_max': 0.09040492400527, 'clip_ratio/region_mean': 0.14651904720813036, 'reward_total_mean': 0.2282353937625885, 'reward_meter_mean': 0.4600223898887634, 'reward_meter_std': 0.42484211921691895, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9141478538513184, 'reward_lexical_plausibility_std': 0.17315734922885895, 'reward_repeat_penalty_mean': 0.9831186532974243, 'reward_repeat_penalty_std': 0.023213526234030724, 'reward_repeat_floor_mean': 0.9977360963821411, 'reward_repeat_floor_std': 0.0028702779673039913, 'reward_near_duplicate_penalty_mean': 0.9877917170524597, 'reward_near_duplicate_penalty_std': 0.013979591429233551, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.995192289352417, 'reward_distinct_2_std': 0.013598216697573662, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.23868314921855927, 'reward_total_composite_mean': 0.2282353937625885, 'reward_total_composite_std': 0.214994877576828, 'epoch': 0.02}

 12%|█▏        | 390/3300 [1:23:28<11:06:50, 13.75s/it]INFO 04-12 17:21:54 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 391/3300 [1:23:37<10:05:26, 12.49s/it]2026-04-12 17:22:03,789 | INFO | train_grpo_train | metrics_logged mode=train step=391

                                                       
{'loss': 0.0196, 'grad_norm': 9.996170043945312, 'learning_rate': 8.818181818181819e-06, 'num_tokens': 732125.0, 'completions/mean_length': 63.25, 'completions/min_length': 53.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.25, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.8383941650390625, 'rewards/meter/std': 0.1706671565771103, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9402313232421875, 'rewards/repeat_penalty/std': 0.1063796654343605, 'rewards/repeat_floor/mean': 0.9948868751525879, 'rewards/repeat_floor/std': 0.007647712714970112, 'rewards/near_duplicate_penalty/mean': 0.9843381643295288, 'rewards/near_duplicate_penalty/std': 0.017609547823667526, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9692903757095337, 'rewards/distinct_2/std': 0.057886600494384766, 'rewards/total_composite/mean': 0.32865601778030396, 'rewards/total_composite/std': 0.07581639289855957, 'reward': 0.32865601778030396, 'reward_std': 0.07581639289855957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16149884462356567, 'sampling/sampling_logp_difference/max': 1.7047510147094727, 'sampling/importance_sampling_ratio/min': 0.18181763589382172, 'sampling/importance_sampling_ratio/mean': 1.0544884204864502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4993118792772293, 'clip_ratio/low_mean': 0.061895521357655525, 'clip_ratio/low_min': 0.061895521357655525, 'clip_ratio/high_mean': 0.08875847700983286, 'clip_ratio/high_max': 0.08875847700983286, 'clip_ratio/region_mean': 0.15065399836748838, 'reward_total_mean': 0.32865601778030396, 'reward_meter_mean': 0.8383941650390625, 'reward_meter_std': 0.1706671565771103, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9402313232421875, 'reward_repeat_penalty_std': 0.1063796654343605, 'reward_repeat_floor_mean': 0.9948868751525879, 'reward_repeat_floor_std': 0.007647712714970112, 'reward_near_duplicate_penalty_mean': 0.9843381643295288, 'reward_near_duplicate_penalty_std': 0.017609547823667526, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9692903757095337, 'reward_distinct_2_std': 0.057886600494384766, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.32865601778030396, 'reward_total_composite_std': 0.07581639289855957, 'epoch': 0.02}

 12%|█▏        | 391/3300 [1:23:37<10:05:26, 12.49s/it]INFO 04-12 17:22:04 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 392/3300 [1:23:46<9:17:19, 11.50s/it] 2026-04-12 17:22:12,984 | INFO | train_grpo_train | metrics_logged mode=train step=392

                                                      
{'loss': 0.0233, 'grad_norm': 13.247349739074707, 'learning_rate': 8.815151515151516e-06, 'num_tokens': 733591.0, 'completions/mean_length': 40.25, 'completions/min_length': 33.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.47477659583091736, 'rewards/meter/std': 0.3723296523094177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.8869690895080566, 'rewards/repeat_penalty/std': 0.1098896786570549, 'rewards/repeat_floor/mean': 0.9859501719474792, 'rewards/repeat_floor/std': 0.01314405258744955, 'rewards/near_duplicate_penalty/mean': 0.9445058107376099, 'rewards/near_duplicate_penalty/std': 0.045414797961711884, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9363809823989868, 'rewards/distinct_2/std': 0.0766734927892685, 'rewards/total_composite/mean': 0.20537853240966797, 'rewards/total_composite/std': 0.14428956806659698, 'reward': 0.20537853240966797, 'reward_std': 0.14428956806659698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19985879957675934, 'sampling/sampling_logp_difference/max': 1.397996425628662, 'sampling/importance_sampling_ratio/min': 0.24709153175354004, 'sampling/importance_sampling_ratio/mean': 1.0509636402130127, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.148595839738846, 'clip_ratio/low_mean': 0.06792653910815716, 'clip_ratio/low_min': 0.06792653910815716, 'clip_ratio/high_mean': 0.12864204309880733, 'clip_ratio/high_max': 0.12864204309880733, 'clip_ratio/region_mean': 0.1965685822069645, 'reward_total_mean': 0.20537853240966797, 'reward_meter_mean': 0.47477659583091736, 'reward_meter_std': 0.3723296523094177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8869690895080566, 'reward_repeat_penalty_std': 0.1098896786570549, 'reward_repeat_floor_mean': 0.9859501719474792, 'reward_repeat_floor_std': 0.01314405258744955, 'reward_near_duplicate_penalty_mean': 0.9445058107376099, 'reward_near_duplicate_penalty_std': 0.045414797961711884, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9363809823989868, 'reward_distinct_2_std': 0.0766734927892685, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.20537853240966797, 'reward_total_composite_std': 0.14428956806659698, 'epoch': 0.02}

 12%|█▏        | 392/3300 [1:23:46<9:17:19, 11.50s/it]INFO 04-12 17:22:13 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 393/3300 [1:24:01<9:58:54, 12.36s/it]2026-04-12 17:22:27,354 | INFO | train_grpo_train | metrics_logged mode=train step=393

                                                      
{'loss': -0.0492, 'grad_norm': 4.1214375495910645, 'learning_rate': 8.812121212121213e-06, 'num_tokens': 735591.0, 'completions/mean_length': 136.0, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 82.28572082519531, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.7372268438339233, 'rewards/meter/std': 0.4070919156074524, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9264833927154541, 'rewards/lexical_plausibility/std': 0.1688588559627533, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.36286312341690063, 'rewards/repeat_penalty/mean': 0.9304853677749634, 'rewards/repeat_penalty/std': 0.0618341863155365, 'rewards/repeat_floor/mean': 0.991073727607727, 'rewards/repeat_floor/std': 0.008880944922566414, 'rewards/near_duplicate_penalty/mean': 0.9570054411888123, 'rewards/near_duplicate_penalty/std': 0.05864662677049637, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.972476601600647, 'rewards/distinct_2/std': 0.029476290568709373, 'rewards/total_composite/mean': 0.4474537968635559, 'rewards/total_composite/std': 0.3689633011817932, 'reward': 0.4474537968635559, 'reward_std': 0.3689633011817932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14681163430213928, 'sampling/sampling_logp_difference/max': 1.7314529418945312, 'sampling/importance_sampling_ratio/min': 0.17702700197696686, 'sampling/importance_sampling_ratio/mean': 1.010379672050476, 'sampling/importance_sampling_ratio/max': 1.965679407119751, 'entropy': 1.0856412202119827, 'clip_ratio/low_mean': 0.07229009084403515, 'clip_ratio/low_min': 0.07229009084403515, 'clip_ratio/high_mean': 0.058582477271556854, 'clip_ratio/high_max': 0.058582477271556854, 'clip_ratio/region_mean': 0.130872568115592, 'reward_total_mean': 0.4474537968635559, 'reward_meter_mean': 0.7372268438339233, 'reward_meter_std': 0.4070919156074524, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9264833927154541, 'reward_lexical_plausibility_std': 0.1688588559627533, 'reward_repeat_penalty_mean': 0.9304853677749634, 'reward_repeat_penalty_std': 0.0618341863155365, 'reward_repeat_floor_mean': 0.991073727607727, 'reward_repeat_floor_std': 0.008880944922566414, 'reward_near_duplicate_penalty_mean': 0.9570054411888123, 'reward_near_duplicate_penalty_std': 0.05864662677049637, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.972476601600647, 'reward_distinct_2_std': 0.029476290568709373, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.36286312341690063, 'reward_total_composite_mean': 0.4474537968635559, 'reward_total_composite_std': 0.3689633011817932, 'epoch': 0.02}

 12%|█▏        | 393/3300 [1:24:01<9:58:54, 12.36s/it]INFO 04-12 17:22:27 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 394/3300 [1:24:10<9:18:28, 11.53s/it]2026-04-12 17:22:36,949 | INFO | train_grpo_train | metrics_logged mode=train step=394

                                                      
{'loss': -0.0682, 'grad_norm': 12.230426788330078, 'learning_rate': 8.809090909090909e-06, 'num_tokens': 737169.0, 'completions/mean_length': 43.25, 'completions/min_length': 36.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8786708116531372, 'rewards/meter/std': 0.23519764840602875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.8454370498657227, 'rewards/repeat_penalty/std': 0.30686092376708984, 'rewards/repeat_floor/mean': 0.9786449074745178, 'rewards/repeat_floor/std': 0.044148582965135574, 'rewards/near_duplicate_penalty/mean': 0.9359985589981079, 'rewards/near_duplicate_penalty/std': 0.10977930575609207, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8902232646942139, 'rewards/distinct_2/std': 0.250618040561676, 'rewards/total_composite/mean': 0.42066481709480286, 'rewards/total_composite/std': 0.21841852366924286, 'reward': 0.42066481709480286, 'reward_std': 0.21841853857040405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15777400135993958, 'sampling/sampling_logp_difference/max': 1.178992748260498, 'sampling/importance_sampling_ratio/min': 0.3075883984565735, 'sampling/importance_sampling_ratio/mean': 1.0269209146499634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.279542215168476, 'clip_ratio/low_mean': 0.07671437226235867, 'clip_ratio/low_min': 0.07671437226235867, 'clip_ratio/high_mean': 0.06279671099036932, 'clip_ratio/high_max': 0.06279671099036932, 'clip_ratio/region_mean': 0.13951108325272799, 'reward_total_mean': 0.42066481709480286, 'reward_meter_mean': 0.8786708116531372, 'reward_meter_std': 0.23519764840602875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8454370498657227, 'reward_repeat_penalty_std': 0.30686092376708984, 'reward_repeat_floor_mean': 0.9786449074745178, 'reward_repeat_floor_std': 0.044148582965135574, 'reward_near_duplicate_penalty_mean': 0.9359985589981079, 'reward_near_duplicate_penalty_std': 0.10977930575609207, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8902232646942139, 'reward_distinct_2_std': 0.250618040561676, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.42066481709480286, 'reward_total_composite_std': 0.21841852366924286, 'epoch': 0.02}

 12%|█▏        | 394/3300 [1:24:10<9:18:28, 11.53s/it]INFO 04-12 17:22:37 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 395/3300 [1:24:19<8:30:54, 10.55s/it]2026-04-12 17:22:45,217 | INFO | train_grpo_train | metrics_logged mode=train step=395

                                                      
{'loss': 0.0401, 'grad_norm': 10.417993545532227, 'learning_rate': 8.806060606060608e-06, 'num_tokens': 738946.0, 'completions/mean_length': 40.125, 'completions/min_length': 34.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6346161365509033, 'rewards/meter/std': 0.3963431119918823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9247074723243713, 'rewards/repeat_penalty/std': 0.17210596799850464, 'rewards/repeat_floor/mean': 0.9911103844642639, 'rewards/repeat_floor/std': 0.01904297061264515, 'rewards/near_duplicate_penalty/mean': 0.970543622970581, 'rewards/near_duplicate_penalty/std': 0.04399280250072479, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9711538553237915, 'rewards/distinct_2/std': 0.0815892368555069, 'rewards/total_composite/mean': 0.37466081976890564, 'rewards/total_composite/std': 0.3277771472930908, 'reward': 0.37466081976890564, 'reward_std': 0.3277771472930908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18876314163208008, 'sampling/sampling_logp_difference/max': 3.047183036804199, 'sampling/importance_sampling_ratio/min': 0.04749252274632454, 'sampling/importance_sampling_ratio/mean': 1.014123797416687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0627145543694496, 'clip_ratio/low_mean': 0.12015937734395266, 'clip_ratio/low_min': 0.12015937734395266, 'clip_ratio/high_mean': 0.04669117787852883, 'clip_ratio/high_max': 0.04669117787852883, 'clip_ratio/region_mean': 0.1668505552224815, 'reward_total_mean': 0.37466081976890564, 'reward_meter_mean': 0.6346161365509033, 'reward_meter_std': 0.3963431119918823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9247074723243713, 'reward_repeat_penalty_std': 0.17210596799850464, 'reward_repeat_floor_mean': 0.9911103844642639, 'reward_repeat_floor_std': 0.01904297061264515, 'reward_near_duplicate_penalty_mean': 0.970543622970581, 'reward_near_duplicate_penalty_std': 0.04399280250072479, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9711538553237915, 'reward_distinct_2_std': 0.0815892368555069, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.37466081976890564, 'reward_total_composite_std': 0.3277771472930908, 'epoch': 0.02}

 12%|█▏        | 395/3300 [1:24:19<8:30:54, 10.55s/it]INFO 04-12 17:22:45 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 396/3300 [1:24:33<9:30:19, 11.78s/it]2026-04-12 17:22:59,875 | INFO | train_grpo_train | metrics_logged mode=train step=396

                                                      
{'loss': -0.0895, 'grad_norm': 4.484723091125488, 'learning_rate': 8.803030303030303e-06, 'num_tokens': 741130.0, 'completions/mean_length': 148.0, 'completions/min_length': 85.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 96.00000762939453, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.6980688571929932, 'rewards/meter/std': 0.37530940771102905, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.971875011920929, 'rewards/count_floor/std': 0.07954952120780945, 'rewards/lexical_plausibility/mean': 0.8836562037467957, 'rewards/lexical_plausibility/std': 0.2205982357263565, 'rewards/judge_quality/mean': 0.23749998211860657, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.7404608726501465, 'rewards/repeat_penalty/std': 0.27593088150024414, 'rewards/repeat_floor/mean': 0.9706962704658508, 'rewards/repeat_floor/std': 0.03453051298856735, 'rewards/near_duplicate_penalty/mean': 0.9310324192047119, 'rewards/near_duplicate_penalty/std': 0.08182311803102493, 'rewards/opening_diversity/mean': 0.9635416269302368, 'rewards/opening_diversity/std': 0.05970953777432442, 'rewards/distinct_2/mean': 0.8136544823646545, 'rewards/distinct_2/std': 0.21146509051322937, 'rewards/total_composite/mean': 0.17175447940826416, 'rewards/total_composite/std': 0.13983315229415894, 'reward': 0.17175447940826416, 'reward_std': 0.13983313739299774, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1565432995557785, 'sampling/sampling_logp_difference/max': 1.8109512329101562, 'sampling/importance_sampling_ratio/min': 0.19080238044261932, 'sampling/importance_sampling_ratio/mean': 1.0198642015457153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7546832039952278, 'clip_ratio/low_mean': 0.053342288825660944, 'clip_ratio/low_min': 0.053342288825660944, 'clip_ratio/high_mean': 0.07533847633749247, 'clip_ratio/high_max': 0.07533847633749247, 'clip_ratio/region_mean': 0.1286807651631534, 'reward_total_mean': 0.17175447940826416, 'reward_meter_mean': 0.6980688571929932, 'reward_meter_std': 0.37530940771102905, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.971875011920929, 'reward_count_floor_std': 0.07954952120780945, 'reward_lexical_plausibility_mean': 0.8836562037467957, 'reward_lexical_plausibility_std': 0.2205982357263565, 'reward_repeat_penalty_mean': 0.7404608726501465, 'reward_repeat_penalty_std': 0.27593088150024414, 'reward_repeat_floor_mean': 0.9706962704658508, 'reward_repeat_floor_std': 0.03453051298856735, 'reward_near_duplicate_penalty_mean': 0.9310324192047119, 'reward_near_duplicate_penalty_std': 0.08182311803102493, 'reward_opening_diversity_mean': 0.9635416269302368, 'reward_opening_diversity_std': 0.05970953777432442, 'reward_distinct_2_mean': 0.8136544823646545, 'reward_distinct_2_std': 0.21146509051322937, 'reward_judge_quality_mean': 0.23749998211860657, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.17175447940826416, 'reward_total_composite_std': 0.13983315229415894, 'epoch': 0.02}

 12%|█▏        | 396/3300 [1:24:33<9:30:19, 11.78s/it]INFO 04-12 17:23:00 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 397/3300 [1:24:48<10:15:52, 12.73s/it]2026-04-12 17:23:14,807 | INFO | train_grpo_train | metrics_logged mode=train step=397

                                                       
{'loss': -0.1715, 'grad_norm': 2.000181198120117, 'learning_rate': 8.8e-06, 'num_tokens': 743012.0, 'completions/mean_length': 312.25, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 112.5, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.7296103239059448, 'rewards/meter/std': 0.3876735270023346, 'rewards/count_adherence/mean': 0.6500000357627869, 'rewards/count_adherence/std': 0.3338091969490051, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8949999809265137, 'rewards/count_floor/std': 0.10014276206493378, 'rewards/lexical_plausibility/mean': 0.7854212522506714, 'rewards/lexical_plausibility/std': 0.23089616000652313, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9214908480644226, 'rewards/repeat_penalty/std': 0.16657783091068268, 'rewards/repeat_floor/mean': 0.9921917915344238, 'rewards/repeat_floor/std': 0.014982621185481548, 'rewards/near_duplicate_penalty/mean': 0.9822738766670227, 'rewards/near_duplicate_penalty/std': 0.017936529591679573, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9427855014801025, 'rewards/distinct_2/std': 0.14097727835178375, 'rewards/total_composite/mean': 0.18969127535820007, 'rewards/total_composite/std': 0.18752700090408325, 'reward': 0.18969127535820007, 'reward_std': 0.18752698600292206, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18511773645877838, 'sampling/sampling_logp_difference/max': 1.4762601852416992, 'sampling/importance_sampling_ratio/min': 0.22849060595035553, 'sampling/importance_sampling_ratio/mean': 1.0227693319320679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2498980313539505, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09984966926276684, 'clip_ratio/high_max': 0.09984966926276684, 'clip_ratio/region_mean': 0.09984966926276684, 'reward_total_mean': 0.18969127535820007, 'reward_meter_mean': 0.7296103239059448, 'reward_meter_std': 0.3876735270023346, 'reward_count_adherence_mean': 0.6500000357627869, 'reward_count_adherence_std': 0.3338091969490051, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8949999809265137, 'reward_count_floor_std': 0.10014276206493378, 'reward_lexical_plausibility_mean': 0.7854212522506714, 'reward_lexical_plausibility_std': 0.23089616000652313, 'reward_repeat_penalty_mean': 0.9214908480644226, 'reward_repeat_penalty_std': 0.16657783091068268, 'reward_repeat_floor_mean': 0.9921917915344238, 'reward_repeat_floor_std': 0.014982621185481548, 'reward_near_duplicate_penalty_mean': 0.9822738766670227, 'reward_near_duplicate_penalty_std': 0.017936529591679573, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9427855014801025, 'reward_distinct_2_std': 0.14097727835178375, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.18969127535820007, 'reward_total_composite_std': 0.18752700090408325, 'epoch': 0.02}

 12%|█▏        | 397/3300 [1:24:48<10:15:52, 12.73s/it]INFO 04-12 17:23:15 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 398/3300 [1:25:02<10:26:25, 12.95s/it]2026-04-12 17:23:28,279 | INFO | train_grpo_train | metrics_logged mode=train step=398

                                                       
{'loss': -0.0332, 'grad_norm': 4.894969940185547, 'learning_rate': 8.796969696969698e-06, 'num_tokens': 744860.0, 'completions/mean_length': 120.0, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.61540687084198, 'rewards/meter/std': 0.38487833738327026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.947494387626648, 'rewards/lexical_plausibility/std': 0.14850832521915436, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.9320471286773682, 'rewards/repeat_penalty/std': 0.07208981364965439, 'rewards/repeat_floor/mean': 0.9918498396873474, 'rewards/repeat_floor/std': 0.007867157459259033, 'rewards/near_duplicate_penalty/mean': 0.9689217805862427, 'rewards/near_duplicate_penalty/std': 0.023324813693761826, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.961239755153656, 'rewards/distinct_2/std': 0.05945970118045807, 'rewards/total_composite/mean': 0.3196439743041992, 'rewards/total_composite/std': 0.27725884318351746, 'reward': 0.3196439743041992, 'reward_std': 0.27725884318351746, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18957772850990295, 'sampling/sampling_logp_difference/max': 2.0010344982147217, 'sampling/importance_sampling_ratio/min': 0.13519535958766937, 'sampling/importance_sampling_ratio/mean': 1.0056103467941284, 'sampling/importance_sampling_ratio/max': 1.8509271144866943, 'entropy': 1.5113163739442825, 'clip_ratio/low_mean': 0.057552896440029144, 'clip_ratio/low_min': 0.057552896440029144, 'clip_ratio/high_mean': 0.12114899046719074, 'clip_ratio/high_max': 0.12114899046719074, 'clip_ratio/region_mean': 0.1787018869072199, 'reward_total_mean': 0.3196439743041992, 'reward_meter_mean': 0.61540687084198, 'reward_meter_std': 0.38487833738327026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.947494387626648, 'reward_lexical_plausibility_std': 0.14850832521915436, 'reward_repeat_penalty_mean': 0.9320471286773682, 'reward_repeat_penalty_std': 0.07208981364965439, 'reward_repeat_floor_mean': 0.9918498396873474, 'reward_repeat_floor_std': 0.007867157459259033, 'reward_near_duplicate_penalty_mean': 0.9689217805862427, 'reward_near_duplicate_penalty_std': 0.023324813693761826, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.961239755153656, 'reward_distinct_2_std': 0.05945970118045807, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.3196439743041992, 'reward_total_composite_std': 0.27725884318351746, 'epoch': 0.02}

 12%|█▏        | 398/3300 [1:25:02<10:26:25, 12.95s/it]INFO 04-12 17:23:28 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 399/3300 [1:25:15<10:31:22, 13.06s/it]2026-04-12 17:23:41,586 | INFO | train_grpo_train | metrics_logged mode=train step=399

                                                       
{'loss': -0.0415, 'grad_norm': 4.3880839347839355, 'learning_rate': 8.793939393939395e-06, 'num_tokens': 746519.0, 'completions/mean_length': 101.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.71428680419922, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7624013423919678, 'rewards/meter/std': 0.4070345461368561, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9524913430213928, 'rewards/lexical_plausibility/std': 0.12353385984897614, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121344089508057, 'rewards/repeat_penalty/mean': 0.9850335121154785, 'rewards/repeat_penalty/std': 0.029185784980654716, 'rewards/repeat_floor/mean': 0.9983240365982056, 'rewards/repeat_floor/std': 0.002946495311334729, 'rewards/near_duplicate_penalty/mean': 0.9945961833000183, 'rewards/near_duplicate_penalty/std': 0.0108893271535635, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.3531382977962494, 'rewards/total_composite/std': 0.2838740348815918, 'reward': 0.3531382977962494, 'reward_std': 0.2838740050792694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17453576624393463, 'sampling/sampling_logp_difference/max': 1.8483238220214844, 'sampling/importance_sampling_ratio/min': 0.15750093758106232, 'sampling/importance_sampling_ratio/mean': 1.0251365900039673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.410551592707634, 'clip_ratio/low_mean': 0.08065368328243494, 'clip_ratio/low_min': 0.08065368328243494, 'clip_ratio/high_mean': 0.06921791657805443, 'clip_ratio/high_max': 0.06921791657805443, 'clip_ratio/region_mean': 0.14987159986048937, 'reward_total_mean': 0.3531382977962494, 'reward_meter_mean': 0.7624013423919678, 'reward_meter_std': 0.4070345461368561, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9524913430213928, 'reward_lexical_plausibility_std': 0.12353385984897614, 'reward_repeat_penalty_mean': 0.9850335121154785, 'reward_repeat_penalty_std': 0.029185784980654716, 'reward_repeat_floor_mean': 0.9983240365982056, 'reward_repeat_floor_std': 0.002946495311334729, 'reward_near_duplicate_penalty_mean': 0.9945961833000183, 'reward_near_duplicate_penalty_std': 0.0108893271535635, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121344089508057, 'reward_total_composite_mean': 0.3531382977962494, 'reward_total_composite_std': 0.2838740348815918, 'epoch': 0.02}

 12%|█▏        | 399/3300 [1:25:15<10:31:22, 13.06s/it]INFO 04-12 17:23:41 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 400/3300 [1:25:23<9:22:37, 11.64s/it] 2026-04-12 17:23:49,918 | INFO | train_grpo_train | metrics_logged mode=train step=400

                                                      
{'loss': 0.0365, 'grad_norm': 13.021088600158691, 'learning_rate': 8.790909090909092e-06, 'num_tokens': 747958.0, 'completions/mean_length': 24.875, 'completions/min_length': 21.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.875, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8019142746925354, 'rewards/meter/std': 0.3328975737094879, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9475266933441162, 'rewards/lexical_plausibility/std': 0.0875638872385025, 'rewards/judge_quality/mean': 0.831250011920929, 'rewards/judge_quality/std': 0.27559998631477356, 'rewards/repeat_penalty/mean': 0.7228922843933105, 'rewards/repeat_penalty/std': 0.04000207781791687, 'rewards/repeat_floor/mean': 0.97957843542099, 'rewards/repeat_floor/std': 0.008000420406460762, 'rewards/near_duplicate_penalty/mean': 0.9638563394546509, 'rewards/near_duplicate_penalty/std': 0.05333612486720085, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7295103073120117, 'rewards/total_composite/std': 0.30076947808265686, 'reward': 0.7295103073120117, 'reward_std': 0.3007694482803345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16320884227752686, 'sampling/sampling_logp_difference/max': 1.4541816711425781, 'sampling/importance_sampling_ratio/min': 0.23359143733978271, 'sampling/importance_sampling_ratio/mean': 1.0245164632797241, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2223070114850998, 'clip_ratio/low_mean': 0.023148147389292717, 'clip_ratio/low_min': 0.023148147389292717, 'clip_ratio/high_mean': 0.1686567608267069, 'clip_ratio/high_max': 0.1686567608267069, 'clip_ratio/region_mean': 0.1918049082159996, 'reward_total_mean': 0.7295103073120117, 'reward_meter_mean': 0.8019142746925354, 'reward_meter_std': 0.3328975737094879, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9475266933441162, 'reward_lexical_plausibility_std': 0.0875638872385025, 'reward_repeat_penalty_mean': 0.7228922843933105, 'reward_repeat_penalty_std': 0.04000207781791687, 'reward_repeat_floor_mean': 0.97957843542099, 'reward_repeat_floor_std': 0.008000420406460762, 'reward_near_duplicate_penalty_mean': 0.9638563394546509, 'reward_near_duplicate_penalty_std': 0.05333612486720085, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.831250011920929, 'reward_judge_quality_std': 0.27559998631477356, 'reward_total_composite_mean': 0.7295103073120117, 'reward_total_composite_std': 0.30076947808265686, 'epoch': 0.02}

 12%|█▏        | 400/3300 [1:25:23<9:22:37, 11.64s/it]INFO 04-12 17:23:50 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:04,  5.83s/it]

 23%|██▎       | 3/13 [00:22<01:20,  8.02s/it]

 31%|███       | 4/13 [00:34<01:25,  9.50s/it]

 38%|███▊      | 5/13 [00:44<01:15,  9.43s/it]

 46%|████▌     | 6/13 [00:55<01:11, 10.15s/it]

 54%|█████▍    | 7/13 [01:07<01:03, 10.56s/it]

 62%|██████▏   | 8/13 [01:18<00:54, 10.87s/it]

 69%|██████▉   | 9/13 [01:30<00:45, 11.30s/it]

 77%|███████▋  | 10/13 [01:42<00:34, 11.45s/it]

 85%|████████▍ | 11/13 [01:54<00:23, 11.56s/it]

 92%|█████████▏| 12/13 [02:05<00:11, 11.35s/it]

100%|██████████| 13/13 [02:16<00:00, 11.31s/it]2026-04-12 17:26:18,486 | INFO | train_grpo_train | metrics_logged mode=eval step=400

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 148.5287, 'eval_samples_per_second': 0.7, 'eval_steps_per_second': 0.088, 'eval_num_tokens': 747958.0, 'eval_completions/mean_length': 212.75, 'eval_completions/min_length': 40.38461538461539, 'eval_completions/max_length': 493.15384615384613, 'eval_completions/clipped_ratio': 0.17307692307692307, 'eval_completions/mean_terminated_length': 149.4699677687425, 'eval_completions/min_terminated_length': 40.38461538461539, 'eval_completions/max_terminated_length': 333.46153846153845, 'eval_rewards/meter/mean': 0.6085947385201087, 'eval_rewards/meter/std': 0.3327841472167235, 'eval_rewards/count_adherence/mean': 0.8155458019329951, 'eval_rewards/count_adherence/std': 0.20745725299303347, 'eval_rewards/arabic_clean/mean': 0.7692307692307693, 'eval_rewards/arabic_clean/std': 0.3929927326165713, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9956730741720933, 'eval_rewards/arabic_floor/std': 0.009420730173587799, 'eval_rewards/count_floor/mean': 0.9446637447063739, 'eval_rewards/count_floor/std': 0.062237174894947275, 'eval_rewards/lexical_plausibility/mean': 0.9532026648521423, 'eval_rewards/lexical_plausibility/std': 0.08891198726800773, 'eval_rewards/judge_quality/mean': 0.3454807721651517, 'eval_rewards/judge_quality/std': 0.2318291744360557, 'eval_rewards/repeat_penalty/mean': 0.7250972390174866, 'eval_rewards/repeat_penalty/std': 0.2896957787183615, 'eval_rewards/repeat_floor/mean': 0.970545654113476, 'eval_rewards/repeat_floor/std': 0.03406601409929303, 'eval_rewards/near_duplicate_penalty/mean': 0.9396584767561692, 'eval_rewards/near_duplicate_penalty/std': 0.07124712576086704, 'eval_rewards/opening_diversity/mean': 0.9646484072391803, 'eval_rewards/opening_diversity/std': 0.07636884599924088, 'eval_rewards/distinct_2/mean': 0.7968663985912616, 'eval_rewards/distinct_2/std': 0.26321384425346667, 'eval_rewards/total_composite/mean': 0.21586975799157068, 'eval_rewards/total_composite/std': 0.21183620765805244, 'eval_reward': 0.21586975799157068, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.09321756718250421, 'eval_sampling/sampling_logp_difference/max': 1.1387993739201472, 'eval_sampling/importance_sampling_ratio/min': 0.32615469510738665, 'eval_sampling/importance_sampling_ratio/mean': 1.0274855632048387, 'eval_sampling/importance_sampling_ratio/max': 1.6013466945061317, 'eval_entropy': 1.2129662541242747, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.21586975799157068, 'eval_reward_meter_mean': 0.6085947385201087, 'eval_reward_meter_std': 0.3327841472167235, 'eval_reward_count_adherence_mean': 0.8155458019329951, 'eval_reward_count_adherence_std': 0.20745725299303347, 'eval_reward_arabic_clean_mean': 0.7692307692307693, 'eval_reward_arabic_clean_std': 0.3929927326165713, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9956730741720933, 'eval_reward_arabic_floor_std': 0.009420730173587799, 'eval_reward_count_floor_mean': 0.9446637447063739, 'eval_reward_count_floor_std': 0.062237174894947275, 'eval_reward_lexical_plausibility_mean': 0.9532026648521423, 'eval_reward_lexical_plausibility_std': 0.08891198726800773, 'eval_reward_repeat_penalty_mean': 0.7250972390174866, 'eval_reward_repeat_penalty_std': 0.2896957787183615, 'eval_reward_repeat_floor_mean': 0.970545654113476, 'eval_reward_repeat_floor_std': 0.03406601409929303, 'eval_reward_near_duplicate_penalty_mean': 0.9396584767561692, 'eval_reward_near_duplicate_penalty_std': 0.07124712576086704, 'eval_reward_opening_diversity_mean': 0.9646484072391803, 'eval_reward_opening_diversity_std': 0.07636884599924088, 'eval_reward_distinct_2_mean': 0.7968663985912616, 'eval_reward_distinct_2_std': 0.26321384425346667, 'eval_reward_judge_quality_mean': 0.3454807721651517, 'eval_reward_judge_quality_std': 0.2318291744360557, 'eval_reward_total_composite_mean': 0.21586975799157068, 'eval_reward_total_composite_std': 0.21183620765805244, 'epoch': 0.02}

 12%|█▏        | 400/3300 [1:27:52<9:22:37, 11.64s/it]

100%|██████████| 13/13 [02:16<00:00, 11.31s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 12%|█▏        | 401/3300 [1:28:05<45:36:28, 56.64s/it]2026-04-12 17:26:31,547 | INFO | train_grpo_train | metrics_logged mode=train step=401

                                                       
{'loss': 0.027, 'grad_norm': 14.887590408325195, 'learning_rate': 8.787878787878788e-06, 'num_tokens': 749831.0, 'completions/mean_length': 65.125, 'completions/min_length': 53.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.8512269854545593, 'rewards/meter/std': 0.3351706266403198, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.9267975091934204, 'rewards/repeat_penalty/std': 0.06511903554201126, 'rewards/repeat_floor/mean': 0.9922330975532532, 'rewards/repeat_floor/std': 0.006678598001599312, 'rewards/near_duplicate_penalty/mean': 0.9824880361557007, 'rewards/near_duplicate_penalty/std': 0.014683682471513748, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.942887544631958, 'rewards/distinct_2/std': 0.05745479464530945, 'rewards/total_composite/mean': 0.3106134533882141, 'rewards/total_composite/std': 0.15826581418514252, 'reward': 0.3106134533882141, 'reward_std': 0.15826581418514252, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14995956420898438, 'sampling/sampling_logp_difference/max': 1.8227968215942383, 'sampling/importance_sampling_ratio/min': 0.16157323122024536, 'sampling/importance_sampling_ratio/mean': 1.0287705659866333, 'sampling/importance_sampling_ratio/max': 1.9278740882873535, 'entropy': 1.2748309522867203, 'clip_ratio/low_mean': 0.02996794879436493, 'clip_ratio/low_min': 0.02996794879436493, 'clip_ratio/high_mean': 0.13120628520846367, 'clip_ratio/high_max': 0.13120628520846367, 'clip_ratio/region_mean': 0.1611742340028286, 'reward_total_mean': 0.3106134533882141, 'reward_meter_mean': 0.8512269854545593, 'reward_meter_std': 0.3351706266403198, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9267975091934204, 'reward_repeat_penalty_std': 0.06511903554201126, 'reward_repeat_floor_mean': 0.9922330975532532, 'reward_repeat_floor_std': 0.006678598001599312, 'reward_near_duplicate_penalty_mean': 0.9824880361557007, 'reward_near_duplicate_penalty_std': 0.014683682471513748, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.942887544631958, 'reward_distinct_2_std': 0.05745479464530945, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.3106134533882141, 'reward_total_composite_std': 0.15826581418514252, 'epoch': 0.02}

 12%|█▏        | 401/3300 [1:28:05<45:36:28, 56.64s/it]INFO 04-12 17:26:31 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 402/3300 [1:28:14<34:03:13, 42.30s/it]2026-04-12 17:26:40,407 | INFO | train_grpo_train | metrics_logged mode=train step=402

                                                       
{'loss': -0.0198, 'grad_norm': 14.945469856262207, 'learning_rate': 8.784848484848487e-06, 'num_tokens': 751533.0, 'completions/mean_length': 41.75, 'completions/min_length': 33.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6461836099624634, 'rewards/meter/std': 0.3091847598552704, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.2655990719795227, 'rewards/repeat_penalty/mean': 0.9799329042434692, 'rewards/repeat_penalty/std': 0.05391687527298927, 'rewards/repeat_floor/mean': 0.9978821277618408, 'rewards/repeat_floor/std': 0.005568044725805521, 'rewards/near_duplicate_penalty/mean': 0.9954961538314819, 'rewards/near_duplicate_penalty/std': 0.010154329240322113, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.497277170419693, 'rewards/total_composite/std': 0.33507657051086426, 'reward': 0.497277170419693, 'reward_std': 0.33507657051086426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14151373505592346, 'sampling/sampling_logp_difference/max': 1.328261375427246, 'sampling/importance_sampling_ratio/min': 0.26493749022483826, 'sampling/importance_sampling_ratio/mean': 1.0286755561828613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.880745142698288, 'clip_ratio/low_mean': 0.06724521797150373, 'clip_ratio/low_min': 0.06724521797150373, 'clip_ratio/high_mean': 0.03849832620471716, 'clip_ratio/high_max': 0.03849832620471716, 'clip_ratio/region_mean': 0.1057435441762209, 'reward_total_mean': 0.497277170419693, 'reward_meter_mean': 0.6461836099624634, 'reward_meter_std': 0.3091847598552704, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9799329042434692, 'reward_repeat_penalty_std': 0.05391687527298927, 'reward_repeat_floor_mean': 0.9978821277618408, 'reward_repeat_floor_std': 0.005568044725805521, 'reward_near_duplicate_penalty_mean': 0.9954961538314819, 'reward_near_duplicate_penalty_std': 0.010154329240322113, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.2655990719795227, 'reward_total_composite_mean': 0.497277170419693, 'reward_total_composite_std': 0.33507657051086426, 'epoch': 0.02}

 12%|█▏        | 402/3300 [1:28:14<34:03:13, 42.30s/it]INFO 04-12 17:26:40 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 403/3300 [1:28:22<25:49:31, 32.09s/it]2026-04-12 17:26:48,673 | INFO | train_grpo_train | metrics_logged mode=train step=403

                                                       
{'loss': 0.0335, 'grad_norm': 14.060990333557129, 'learning_rate': 8.781818181818182e-06, 'num_tokens': 753244.0, 'completions/mean_length': 40.875, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9835188388824463, 'rewards/meter/std': 0.013107487931847572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5587499737739563, 'rewards/judge_quality/std': 0.23491261899471283, 'rewards/repeat_penalty/mean': 0.9854729771614075, 'rewards/repeat_penalty/std': 0.0106391916051507, 'rewards/repeat_floor/mean': 0.9978209733963013, 'rewards/repeat_floor/std': 0.0015958774602040648, 'rewards/near_duplicate_penalty/mean': 0.9854729771614075, 'rewards/near_duplicate_penalty/std': 0.0106391916051507, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5484583973884583, 'rewards/total_composite/std': 0.23124751448631287, 'reward': 0.5484583973884583, 'reward_std': 0.23124751448631287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14659333229064941, 'sampling/sampling_logp_difference/max': 1.2318469285964966, 'sampling/importance_sampling_ratio/min': 0.29175323247909546, 'sampling/importance_sampling_ratio/mean': 1.0133779048919678, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1714520901441574, 'clip_ratio/low_mean': 0.12399742100387812, 'clip_ratio/low_min': 0.12399742100387812, 'clip_ratio/high_mean': 0.0324127902276814, 'clip_ratio/high_max': 0.0324127902276814, 'clip_ratio/region_mean': 0.15641021123155951, 'reward_total_mean': 0.5484583973884583, 'reward_meter_mean': 0.9835188388824463, 'reward_meter_std': 0.013107487931847572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9854729771614075, 'reward_repeat_penalty_std': 0.0106391916051507, 'reward_repeat_floor_mean': 0.9978209733963013, 'reward_repeat_floor_std': 0.0015958774602040648, 'reward_near_duplicate_penalty_mean': 0.9854729771614075, 'reward_near_duplicate_penalty_std': 0.0106391916051507, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5587499737739563, 'reward_judge_quality_std': 0.23491261899471283, 'reward_total_composite_mean': 0.5484583973884583, 'reward_total_composite_std': 0.23124751448631287, 'epoch': 0.02}

 12%|█▏        | 403/3300 [1:28:22<25:49:31, 32.09s/it]INFO 04-12 17:26:48 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 404/3300 [1:28:30<20:01:55, 24.90s/it]2026-04-12 17:26:56,796 | INFO | train_grpo_train | metrics_logged mode=train step=404

                                                       
{'loss': 0.1027, 'grad_norm': 17.967100143432617, 'learning_rate': 8.77878787878788e-06, 'num_tokens': 754673.0, 'completions/mean_length': 25.625, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9386039972305298, 'rewards/meter/std': 0.0921686589717865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.0289318785071373, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6715369820594788, 'rewards/total_composite/std': 0.24499766528606415, 'reward': 0.6715369820594788, 'reward_std': 0.24499766528606415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17219848930835724, 'sampling/sampling_logp_difference/max': 1.4513039588928223, 'sampling/importance_sampling_ratio/min': 0.23426462709903717, 'sampling/importance_sampling_ratio/mean': 1.0142366886138916, 'sampling/importance_sampling_ratio/max': 1.8152958154678345, 'entropy': 1.4703346639871597, 'clip_ratio/low_mean': 0.047901272773742676, 'clip_ratio/low_min': 0.047901272773742676, 'clip_ratio/high_mean': 0.06808359315618873, 'clip_ratio/high_max': 0.06808359315618873, 'clip_ratio/region_mean': 0.1159848659299314, 'reward_total_mean': 0.6715369820594788, 'reward_meter_mean': 0.9386039972305298, 'reward_meter_std': 0.0921686589717865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.0289318785071373, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.6715369820594788, 'reward_total_composite_std': 0.24499766528606415, 'epoch': 0.02}

 12%|█▏        | 404/3300 [1:28:30<20:01:55, 24.90s/it]INFO 04-12 17:26:57 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 405/3300 [1:28:39<16:07:02, 20.04s/it]2026-04-12 17:27:05,499 | INFO | train_grpo_train | metrics_logged mode=train step=405

                                                       
{'loss': 0.1602, 'grad_norm': 19.066726684570312, 'learning_rate': 8.775757575757577e-06, 'num_tokens': 756170.0, 'completions/mean_length': 30.125, 'completions/min_length': 21.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.125, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9536111354827881, 'rewards/meter/std': 0.10049214959144592, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.953125, 'rewards/lexical_plausibility/std': 0.13258251547813416, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3184336721897125, 'rewards/repeat_penalty/mean': 0.7586374282836914, 'rewards/repeat_penalty/std': 0.033012986183166504, 'rewards/repeat_floor/mean': 0.983735978603363, 'rewards/repeat_floor/std': 0.0024212803691625595, 'rewards/near_duplicate_penalty/mean': 0.9840898513793945, 'rewards/near_duplicate_penalty/std': 0.03568260744214058, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.5554136037826538, 'rewards/total_composite/std': 0.3026265799999237, 'reward': 0.5554136037826538, 'reward_std': 0.3026265799999237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15680178999900818, 'sampling/sampling_logp_difference/max': 1.4881324768066406, 'sampling/importance_sampling_ratio/min': 0.22579394280910492, 'sampling/importance_sampling_ratio/mean': 0.9988325238227844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0948723927140236, 'clip_ratio/low_mean': 0.11339311860501766, 'clip_ratio/low_min': 0.11339311860501766, 'clip_ratio/high_mean': 0.06383415637537837, 'clip_ratio/high_max': 0.06383415637537837, 'clip_ratio/region_mean': 0.17722727498039603, 'reward_total_mean': 0.5554136037826538, 'reward_meter_mean': 0.9536111354827881, 'reward_meter_std': 0.10049214959144592, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.953125, 'reward_lexical_plausibility_std': 0.13258251547813416, 'reward_repeat_penalty_mean': 0.7586374282836914, 'reward_repeat_penalty_std': 0.033012986183166504, 'reward_repeat_floor_mean': 0.983735978603363, 'reward_repeat_floor_std': 0.0024212803691625595, 'reward_near_duplicate_penalty_mean': 0.9840898513793945, 'reward_near_duplicate_penalty_std': 0.03568260744214058, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3184336721897125, 'reward_total_composite_mean': 0.5554136037826538, 'reward_total_composite_std': 0.3026265799999237, 'epoch': 0.02}

 12%|█▏        | 405/3300 [1:28:39<16:07:02, 20.04s/it]INFO 04-12 17:27:05 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 406/3300 [1:28:47<13:21:23, 16.61s/it]2026-04-12 17:27:14,116 | INFO | train_grpo_train | metrics_logged mode=train step=406

                                                       
{'loss': 0.0447, 'grad_norm': 15.60359001159668, 'learning_rate': 8.772727272727274e-06, 'num_tokens': 757688.0, 'completions/mean_length': 36.75, 'completions/min_length': 31.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7540900707244873, 'rewards/meter/std': 0.33582428097724915, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.26385873556137085, 'rewards/repeat_penalty/mean': 0.9719797372817993, 'rewards/repeat_penalty/std': 0.03192441165447235, 'rewards/repeat_floor/mean': 0.9957969784736633, 'rewards/repeat_floor/std': 0.004788653925061226, 'rewards/near_duplicate_penalty/mean': 0.9719797372817993, 'rewards/near_duplicate_penalty/std': 0.03192441165447235, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3961288034915924, 'rewards/total_composite/std': 0.3084753751754761, 'reward': 0.3961288034915924, 'reward_std': 0.3084753751754761, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14347271621227264, 'sampling/sampling_logp_difference/max': 1.327934741973877, 'sampling/importance_sampling_ratio/min': 0.2650240659713745, 'sampling/importance_sampling_ratio/mean': 1.0067261457443237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1793797463178635, 'clip_ratio/low_mean': 0.08616048749536276, 'clip_ratio/low_min': 0.08616048749536276, 'clip_ratio/high_mean': 0.03101503849029541, 'clip_ratio/high_max': 0.03101503849029541, 'clip_ratio/region_mean': 0.11717552598565817, 'reward_total_mean': 0.3961288034915924, 'reward_meter_mean': 0.7540900707244873, 'reward_meter_std': 0.33582428097724915, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9719797372817993, 'reward_repeat_penalty_std': 0.03192441165447235, 'reward_repeat_floor_mean': 0.9957969784736633, 'reward_repeat_floor_std': 0.004788653925061226, 'reward_near_duplicate_penalty_mean': 0.9719797372817993, 'reward_near_duplicate_penalty_std': 0.03192441165447235, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.26385873556137085, 'reward_total_composite_mean': 0.3961288034915924, 'reward_total_composite_std': 0.3084753751754761, 'epoch': 0.02}

 12%|█▏        | 406/3300 [1:28:47<13:21:23, 16.61s/it]INFO 04-12 17:27:14 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 407/3300 [1:29:02<12:53:27, 16.04s/it]2026-04-12 17:27:28,819 | INFO | train_grpo_train | metrics_logged mode=train step=407

                                                       
{'loss': -0.013, 'grad_norm': 5.35736608505249, 'learning_rate': 8.76969696969697e-06, 'num_tokens': 759146.0, 'completions/mean_length': 101.25, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.57143020629883, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.40924423933029175, 'rewards/meter/std': 0.47958657145500183, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9370801448822021, 'rewards/lexical_plausibility/std': 0.17796415090560913, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.39617589116096497, 'rewards/repeat_penalty/mean': 0.9908603429794312, 'rewards/repeat_penalty/std': 0.01109218504279852, 'rewards/repeat_floor/mean': 0.998629093170166, 'rewards/repeat_floor/std': 0.0016638237284496427, 'rewards/near_duplicate_penalty/mean': 0.9908603429794312, 'rewards/near_duplicate_penalty/std': 0.01109218504279852, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.37753111124038696, 'rewards/total_composite/std': 0.4559843838214874, 'reward': 0.37753111124038696, 'reward_std': 0.45598435401916504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15604357421398163, 'sampling/sampling_logp_difference/max': 1.76890230178833, 'sampling/importance_sampling_ratio/min': 0.17052006721496582, 'sampling/importance_sampling_ratio/mean': 1.0369576215744019, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9246109053492546, 'clip_ratio/low_mean': 0.09980427101254463, 'clip_ratio/low_min': 0.09980427101254463, 'clip_ratio/high_mean': 0.05201525054872036, 'clip_ratio/high_max': 0.05201525054872036, 'clip_ratio/region_mean': 0.151819521561265, 'reward_total_mean': 0.37753111124038696, 'reward_meter_mean': 0.40924423933029175, 'reward_meter_std': 0.47958657145500183, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9370801448822021, 'reward_lexical_plausibility_std': 0.17796415090560913, 'reward_repeat_penalty_mean': 0.9908603429794312, 'reward_repeat_penalty_std': 0.01109218504279852, 'reward_repeat_floor_mean': 0.998629093170166, 'reward_repeat_floor_std': 0.0016638237284496427, 'reward_near_duplicate_penalty_mean': 0.9908603429794312, 'reward_near_duplicate_penalty_std': 0.01109218504279852, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.39617589116096497, 'reward_total_composite_mean': 0.37753111124038696, 'reward_total_composite_std': 0.4559843838214874, 'epoch': 0.02}

 12%|█▏        | 407/3300 [1:29:02<12:53:27, 16.04s/it]INFO 04-12 17:27:29 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 408/3300 [1:29:13<11:39:10, 14.51s/it]2026-04-12 17:27:39,741 | INFO | train_grpo_train | metrics_logged mode=train step=408

                                                       
{'loss': -0.0267, 'grad_norm': 26.84180450439453, 'learning_rate': 8.766666666666669e-06, 'num_tokens': 760463.0, 'completions/mean_length': 20.625, 'completions/min_length': 16.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.6541098356246948, 'rewards/meter/std': 0.3562012016773224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6475000381469727, 'rewards/judge_quality/std': 0.30742016434669495, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4373846650123596, 'rewards/total_composite/std': 0.3502616882324219, 'reward': 0.4373846650123596, 'reward_std': 0.3502616882324219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18559736013412476, 'sampling/sampling_logp_difference/max': 1.6342525482177734, 'sampling/importance_sampling_ratio/min': 0.1950981318950653, 'sampling/importance_sampling_ratio/mean': 1.001253604888916, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.170207992196083, 'clip_ratio/low_mean': 0.08488005166873336, 'clip_ratio/low_min': 0.08488005166873336, 'clip_ratio/high_mean': 0.06979166716337204, 'clip_ratio/high_max': 0.06979166716337204, 'clip_ratio/region_mean': 0.1546717188321054, 'reward_total_mean': 0.4373846650123596, 'reward_meter_mean': 0.6541098356246948, 'reward_meter_std': 0.3562012016773224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6475000381469727, 'reward_judge_quality_std': 0.30742016434669495, 'reward_total_composite_mean': 0.4373846650123596, 'reward_total_composite_std': 0.3502616882324219, 'epoch': 0.02}

 12%|█▏        | 408/3300 [1:29:13<11:39:10, 14.51s/it]INFO 04-12 17:27:39 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 409/3300 [1:29:27<11:28:38, 14.29s/it]2026-04-12 17:27:53,543 | INFO | train_grpo_train | metrics_logged mode=train step=409

                                                       
{'loss': -0.2103, 'grad_norm': 2.0406625270843506, 'learning_rate': 8.763636363636364e-06, 'num_tokens': 764016.0, 'completions/mean_length': 281.125, 'completions/min_length': 222.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 248.1428680419922, 'completions/min_terminated_length': 222.0, 'completions/max_terminated_length': 281.0, 'rewards/meter/mean': 0.9281846284866333, 'rewards/meter/std': 0.1804521679878235, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.07120776921510696, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.02136232703924179, 'rewards/lexical_plausibility/mean': 0.9526408910751343, 'rewards/lexical_plausibility/std': 0.1339518129825592, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.1510518491268158, 'rewards/repeat_penalty/std': 0.27062907814979553, 'rewards/repeat_floor/mean': 0.8774338364601135, 'rewards/repeat_floor/std': 0.06200440600514412, 'rewards/near_duplicate_penalty/mean': 0.7272931337356567, 'rewards/near_duplicate_penalty/std': 0.1785890758037567, 'rewards/opening_diversity/mean': 0.8492288589477539, 'rewards/opening_diversity/std': 0.23756106197834015, 'rewards/distinct_2/mean': 0.19317904114723206, 'rewards/distinct_2/std': 0.3473024368286133, 'rewards/total_composite/mean': 0.09933464229106903, 'rewards/total_composite/std': 0.04377879574894905, 'reward': 0.09933464229106903, 'reward_std': 0.04377879202365875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07885255664587021, 'sampling/sampling_logp_difference/max': 1.6141021251678467, 'sampling/importance_sampling_ratio/min': 0.1990693360567093, 'sampling/importance_sampling_ratio/mean': 1.0149317979812622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5574203692376614, 'clip_ratio/low_mean': 0.004291845485568047, 'clip_ratio/low_min': 0.004291845485568047, 'clip_ratio/high_mean': 0.050099303014576435, 'clip_ratio/high_max': 0.050099303014576435, 'clip_ratio/region_mean': 0.05439114850014448, 'reward_total_mean': 0.09933464229106903, 'reward_meter_mean': 0.9281846284866333, 'reward_meter_std': 0.1804521679878235, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.07120776921510696, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.02136232703924179, 'reward_lexical_plausibility_mean': 0.9526408910751343, 'reward_lexical_plausibility_std': 0.1339518129825592, 'reward_repeat_penalty_mean': 0.1510518491268158, 'reward_repeat_penalty_std': 0.27062907814979553, 'reward_repeat_floor_mean': 0.8774338364601135, 'reward_repeat_floor_std': 0.06200440600514412, 'reward_near_duplicate_penalty_mean': 0.7272931337356567, 'reward_near_duplicate_penalty_std': 0.1785890758037567, 'reward_opening_diversity_mean': 0.8492288589477539, 'reward_opening_diversity_std': 0.23756106197834015, 'reward_distinct_2_mean': 0.19317904114723206, 'reward_distinct_2_std': 0.3473024368286133, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.09933464229106903, 'reward_total_composite_std': 0.04377879574894905, 'epoch': 0.02}

 12%|█▏        | 409/3300 [1:29:27<11:28:38, 14.29s/it]INFO 04-12 17:27:53 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 410/3300 [1:29:36<10:18:36, 12.84s/it]2026-04-12 17:28:02,998 | INFO | train_grpo_train | metrics_logged mode=train step=410

                                                       
{'loss': 0.0477, 'grad_norm': 12.290157318115234, 'learning_rate': 8.760606060606061e-06, 'num_tokens': 765918.0, 'completions/mean_length': 77.75, 'completions/min_length': 63.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.3320328891277313, 'rewards/meter/std': 0.33772704005241394, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4587499797344208, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9249914884567261, 'rewards/repeat_penalty/std': 0.07813272625207901, 'rewards/repeat_floor/mean': 0.9914798736572266, 'rewards/repeat_floor/std': 0.008529511280357838, 'rewards/near_duplicate_penalty/mean': 0.9740067720413208, 'rewards/near_duplicate_penalty/std': 0.022106532007455826, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9486543536186218, 'rewards/distinct_2/std': 0.06189310550689697, 'rewards/total_composite/mean': 0.12271323800086975, 'rewards/total_composite/std': 0.1129566878080368, 'reward': 0.12271323800086975, 'reward_std': 0.1129566952586174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19214822351932526, 'sampling/sampling_logp_difference/max': 1.9817166328430176, 'sampling/importance_sampling_ratio/min': 0.1378324329853058, 'sampling/importance_sampling_ratio/mean': 1.0120923519134521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9461139813065529, 'clip_ratio/low_mean': 0.131487344391644, 'clip_ratio/low_min': 0.131487344391644, 'clip_ratio/high_mean': 0.0534924641251564, 'clip_ratio/high_max': 0.0534924641251564, 'clip_ratio/region_mean': 0.1849798085168004, 'reward_total_mean': 0.12271323800086975, 'reward_meter_mean': 0.3320328891277313, 'reward_meter_std': 0.33772704005241394, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9249914884567261, 'reward_repeat_penalty_std': 0.07813272625207901, 'reward_repeat_floor_mean': 0.9914798736572266, 'reward_repeat_floor_std': 0.008529511280357838, 'reward_near_duplicate_penalty_mean': 0.9740067720413208, 'reward_near_duplicate_penalty_std': 0.022106532007455826, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9486543536186218, 'reward_distinct_2_std': 0.06189310550689697, 'reward_judge_quality_mean': 0.4587499797344208, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.12271323800086975, 'reward_total_composite_std': 0.1129566878080368, 'epoch': 0.02}

 12%|█▏        | 410/3300 [1:29:36<10:18:36, 12.84s/it]INFO 04-12 17:28:03 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 411/3300 [1:29:45<9:19:06, 11.61s/it] 2026-04-12 17:28:11,737 | INFO | train_grpo_train | metrics_logged mode=train step=411

                                                      
{'loss': 0.0942, 'grad_norm': 12.410826683044434, 'learning_rate': 8.757575757575759e-06, 'num_tokens': 767500.0, 'completions/mean_length': 39.75, 'completions/min_length': 33.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6078414916992188, 'rewards/meter/std': 0.43977683782577515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.2520735561847687, 'rewards/repeat_penalty/mean': 0.9345226287841797, 'rewards/repeat_penalty/std': 0.12379558384418488, 'rewards/repeat_floor/mean': 0.9934014678001404, 'rewards/repeat_floor/std': 0.011220952495932579, 'rewards/near_duplicate_penalty/mean': 0.9806556701660156, 'rewards/near_duplicate_penalty/std': 0.02532178722321987, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.03748259320855141, 'rewards/total_composite/mean': 0.22718289494514465, 'rewards/total_composite/std': 0.29405373334884644, 'reward': 0.22718289494514465, 'reward_std': 0.2940537631511688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21953687071800232, 'sampling/sampling_logp_difference/max': 2.6128580570220947, 'sampling/importance_sampling_ratio/min': 0.07332468032836914, 'sampling/importance_sampling_ratio/mean': 1.015641689300537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9310509115457535, 'clip_ratio/low_mean': 0.1193020697683096, 'clip_ratio/low_min': 0.1193020697683096, 'clip_ratio/high_mean': 0.06779939029365778, 'clip_ratio/high_max': 0.06779939029365778, 'clip_ratio/region_mean': 0.18710146006196737, 'reward_total_mean': 0.22718289494514465, 'reward_meter_mean': 0.6078414916992188, 'reward_meter_std': 0.43977683782577515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9345226287841797, 'reward_repeat_penalty_std': 0.12379558384418488, 'reward_repeat_floor_mean': 0.9934014678001404, 'reward_repeat_floor_std': 0.011220952495932579, 'reward_near_duplicate_penalty_mean': 0.9806556701660156, 'reward_near_duplicate_penalty_std': 0.02532178722321987, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.03748259320855141, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.2520735561847687, 'reward_total_composite_mean': 0.22718289494514465, 'reward_total_composite_std': 0.29405373334884644, 'epoch': 0.02}

 12%|█▏        | 411/3300 [1:29:45<9:19:06, 11.61s/it]INFO 04-12 17:28:11 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 412/3300 [1:29:55<8:57:30, 11.17s/it]2026-04-12 17:28:21,868 | INFO | train_grpo_train | metrics_logged mode=train step=412

                                                      
{'loss': 0.0502, 'grad_norm': 12.038230895996094, 'learning_rate': 8.754545454545456e-06, 'num_tokens': 770026.0, 'completions/mean_length': 133.75, 'completions/min_length': 120.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.75, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.6921815276145935, 'rewards/meter/std': 0.26543235778808594, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9357142448425293, 'rewards/count_floor/std': 0.02290808968245983, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.6986279487609863, 'rewards/repeat_penalty/std': 0.11810480058193207, 'rewards/repeat_floor/mean': 0.9692379236221313, 'rewards/repeat_floor/std': 0.00919030886143446, 'rewards/near_duplicate_penalty/mean': 0.9271507263183594, 'rewards/near_duplicate_penalty/std': 0.013372487388551235, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7796148061752319, 'rewards/distinct_2/std': 0.09739939123392105, 'rewards/total_composite/mean': 0.22802862524986267, 'rewards/total_composite/std': 0.12687252461910248, 'reward': 0.22802862524986267, 'reward_std': 0.12687252461910248, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19746123254299164, 'sampling/sampling_logp_difference/max': 3.173527479171753, 'sampling/importance_sampling_ratio/min': 0.04185568913817406, 'sampling/importance_sampling_ratio/mean': 1.0033645629882812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9051905423402786, 'clip_ratio/low_mean': 0.07040649559348822, 'clip_ratio/low_min': 0.07040649559348822, 'clip_ratio/high_mean': 0.07739872764796019, 'clip_ratio/high_max': 0.07739872764796019, 'clip_ratio/region_mean': 0.1478052232414484, 'reward_total_mean': 0.22802862524986267, 'reward_meter_mean': 0.6921815276145935, 'reward_meter_std': 0.26543235778808594, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9357142448425293, 'reward_count_floor_std': 0.02290808968245983, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6986279487609863, 'reward_repeat_penalty_std': 0.11810480058193207, 'reward_repeat_floor_mean': 0.9692379236221313, 'reward_repeat_floor_std': 0.00919030886143446, 'reward_near_duplicate_penalty_mean': 0.9271507263183594, 'reward_near_duplicate_penalty_std': 0.013372487388551235, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7796148061752319, 'reward_distinct_2_std': 0.09739939123392105, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.22802862524986267, 'reward_total_composite_std': 0.12687252461910248, 'epoch': 0.02}

 12%|█▏        | 412/3300 [1:29:55<8:57:30, 11.17s/it]INFO 04-12 17:28:22 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 413/3300 [1:30:04<8:23:30, 10.46s/it]2026-04-12 17:28:30,690 | INFO | train_grpo_train | metrics_logged mode=train step=413

                                                      
{'loss': 0.0989, 'grad_norm': 19.113256454467773, 'learning_rate': 8.751515151515151e-06, 'num_tokens': 771716.0, 'completions/mean_length': 46.25, 'completions/min_length': 40.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5463078618049622, 'rewards/meter/std': 0.3257794678211212, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5724999904632568, 'rewards/judge_quality/std': 0.38425251841545105, 'rewards/repeat_penalty/mean': 0.9594578742980957, 'rewards/repeat_penalty/std': 0.05256415903568268, 'rewards/repeat_floor/mean': 0.9948335289955139, 'rewards/repeat_floor/std': 0.006377866957336664, 'rewards/near_duplicate_penalty/mean': 0.9761340618133545, 'rewards/near_duplicate_penalty/std': 0.026294246315956116, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9823718070983887, 'rewards/distinct_2/std': 0.0328204445540905, 'rewards/total_composite/mean': 0.3291429281234741, 'rewards/total_composite/std': 0.31265631318092346, 'reward': 0.3291429281234741, 'reward_std': 0.31265631318092346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2219068855047226, 'sampling/sampling_logp_difference/max': 1.6267223358154297, 'sampling/importance_sampling_ratio/min': 0.19657281041145325, 'sampling/importance_sampling_ratio/mean': 0.9901831150054932, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2640323787927628, 'clip_ratio/low_mean': 0.10007031913846731, 'clip_ratio/low_min': 0.10007031913846731, 'clip_ratio/high_mean': 0.06926407013088465, 'clip_ratio/high_max': 0.06926407013088465, 'clip_ratio/region_mean': 0.16933438926935196, 'reward_total_mean': 0.3291429281234741, 'reward_meter_mean': 0.5463078618049622, 'reward_meter_std': 0.3257794678211212, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9594578742980957, 'reward_repeat_penalty_std': 0.05256415903568268, 'reward_repeat_floor_mean': 0.9948335289955139, 'reward_repeat_floor_std': 0.006377866957336664, 'reward_near_duplicate_penalty_mean': 0.9761340618133545, 'reward_near_duplicate_penalty_std': 0.026294246315956116, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9823718070983887, 'reward_distinct_2_std': 0.0328204445540905, 'reward_judge_quality_mean': 0.5724999904632568, 'reward_judge_quality_std': 0.38425251841545105, 'reward_total_composite_mean': 0.3291429281234741, 'reward_total_composite_std': 0.31265631318092346, 'epoch': 0.02}

 13%|█▎        | 413/3300 [1:30:04<8:23:30, 10.46s/it]INFO 04-12 17:28:30 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 414/3300 [1:30:17<9:04:33, 11.32s/it]2026-04-12 17:28:44,011 | INFO | train_grpo_train | metrics_logged mode=train step=414

                                                      
{'loss': -0.0454, 'grad_norm': 5.080606460571289, 'learning_rate': 8.748484848484849e-06, 'num_tokens': 773155.0, 'completions/mean_length': 101.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.28571701049805, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7053014039993286, 'rewards/meter/std': 0.3809899687767029, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9364877939224243, 'rewards/lexical_plausibility/std': 0.17963960766792297, 'rewards/judge_quality/mean': 0.6012500524520874, 'rewards/judge_quality/std': 0.37410610914230347, 'rewards/repeat_penalty/mean': 0.9379380345344543, 'rewards/repeat_penalty/std': 0.0914243832230568, 'rewards/repeat_floor/mean': 0.9940295815467834, 'rewards/repeat_floor/std': 0.007274527568370104, 'rewards/near_duplicate_penalty/mean': 0.9791073799133301, 'rewards/near_duplicate_penalty/std': 0.02847854606807232, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.5348511338233948, 'rewards/total_composite/std': 0.40322163701057434, 'reward': 0.5348511338233948, 'reward_std': 0.40322160720825195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1590871512889862, 'sampling/sampling_logp_difference/max': 1.3593673706054688, 'sampling/importance_sampling_ratio/min': 0.25682318210601807, 'sampling/importance_sampling_ratio/mean': 1.007158875465393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1926024556159973, 'clip_ratio/low_mean': 0.027985830791294575, 'clip_ratio/low_min': 0.027985830791294575, 'clip_ratio/high_mean': 0.07076023425906897, 'clip_ratio/high_max': 0.07076023425906897, 'clip_ratio/region_mean': 0.09874606505036354, 'reward_total_mean': 0.5348511338233948, 'reward_meter_mean': 0.7053014039993286, 'reward_meter_std': 0.3809899687767029, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9364877939224243, 'reward_lexical_plausibility_std': 0.17963960766792297, 'reward_repeat_penalty_mean': 0.9379380345344543, 'reward_repeat_penalty_std': 0.0914243832230568, 'reward_repeat_floor_mean': 0.9940295815467834, 'reward_repeat_floor_std': 0.007274527568370104, 'reward_near_duplicate_penalty_mean': 0.9791073799133301, 'reward_near_duplicate_penalty_std': 0.02847854606807232, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.6012500524520874, 'reward_judge_quality_std': 0.37410610914230347, 'reward_total_composite_mean': 0.5348511338233948, 'reward_total_composite_std': 0.40322163701057434, 'epoch': 0.02}

 13%|█▎        | 414/3300 [1:30:17<9:04:33, 11.32s/it]INFO 04-12 17:28:44 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 415/3300 [1:30:27<8:41:55, 10.85s/it]2026-04-12 17:28:53,779 | INFO | train_grpo_train | metrics_logged mode=train step=415

                                                      
{'loss': 0.0183, 'grad_norm': 7.434648036956787, 'learning_rate': 8.745454545454546e-06, 'num_tokens': 775845.0, 'completions/mean_length': 131.25, 'completions/min_length': 113.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.25, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.8679296970367432, 'rewards/meter/std': 0.1725633591413498, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.01767767407000065, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.537352442741394, 'rewards/repeat_penalty/std': 0.2563095986843109, 'rewards/repeat_floor/mean': 0.9467770457267761, 'rewards/repeat_floor/std': 0.03108583763241768, 'rewards/near_duplicate_penalty/mean': 0.8784315586090088, 'rewards/near_duplicate_penalty/std': 0.05615968629717827, 'rewards/opening_diversity/mean': 0.9776785373687744, 'rewards/opening_diversity/std': 0.05028164014220238, 'rewards/distinct_2/mean': 0.6261287331581116, 'rewards/distinct_2/std': 0.2367241531610489, 'rewards/total_composite/mean': 0.16756850481033325, 'rewards/total_composite/std': 0.07452666014432907, 'reward': 0.16756850481033325, 'reward_std': 0.07452665269374847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15547390282154083, 'sampling/sampling_logp_difference/max': 2.2485227584838867, 'sampling/importance_sampling_ratio/min': 0.10555503517389297, 'sampling/importance_sampling_ratio/mean': 1.0180752277374268, 'sampling/importance_sampling_ratio/max': 1.9504648447036743, 'entropy': 1.3454219177365303, 'clip_ratio/low_mean': 0.08070331392809749, 'clip_ratio/low_min': 0.08070331392809749, 'clip_ratio/high_mean': 0.0400482714176178, 'clip_ratio/high_max': 0.0400482714176178, 'clip_ratio/region_mean': 0.12075158534571528, 'reward_total_mean': 0.16756850481033325, 'reward_meter_mean': 0.8679296970367432, 'reward_meter_std': 0.1725633591413498, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.01767767407000065, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.537352442741394, 'reward_repeat_penalty_std': 0.2563095986843109, 'reward_repeat_floor_mean': 0.9467770457267761, 'reward_repeat_floor_std': 0.03108583763241768, 'reward_near_duplicate_penalty_mean': 0.8784315586090088, 'reward_near_duplicate_penalty_std': 0.05615968629717827, 'reward_opening_diversity_mean': 0.9776785373687744, 'reward_opening_diversity_std': 0.05028164014220238, 'reward_distinct_2_mean': 0.6261287331581116, 'reward_distinct_2_std': 0.2367241531610489, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.16756850481033325, 'reward_total_composite_std': 0.07452666014432907, 'epoch': 0.02}

 13%|█▎        | 415/3300 [1:30:27<8:41:55, 10.85s/it]INFO 04-12 17:28:54 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 416/3300 [1:30:37<8:22:33, 10.46s/it]2026-04-12 17:29:03,303 | INFO | train_grpo_train | metrics_logged mode=train step=416

                                                      
{'loss': 0.0445, 'grad_norm': 15.937089920043945, 'learning_rate': 8.742424242424243e-06, 'num_tokens': 777588.0, 'completions/mean_length': 40.875, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.3729636073112488, 'rewards/meter/std': 0.33335843682289124, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9930555820465088, 'rewards/lexical_plausibility/std': 0.01285861898213625, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2745385766029358, 'rewards/repeat_penalty/mean': 0.9592209458351135, 'rewards/repeat_penalty/std': 0.05305812135338783, 'rewards/repeat_floor/mean': 0.9949278235435486, 'rewards/repeat_floor/std': 0.006081957370042801, 'rewards/near_duplicate_penalty/mean': 0.9780904054641724, 'rewards/near_duplicate_penalty/std': 0.020435959100723267, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9801587462425232, 'rewards/distinct_2/std': 0.036883533000946045, 'rewards/total_composite/mean': 0.1650734692811966, 'rewards/total_composite/std': 0.1427987813949585, 'reward': 0.1650734692811966, 'reward_std': 0.1427987664937973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18382763862609863, 'sampling/sampling_logp_difference/max': 1.1045339107513428, 'sampling/importance_sampling_ratio/min': 0.3390491306781769, 'sampling/importance_sampling_ratio/mean': 1.0634757280349731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4608836770057678, 'clip_ratio/low_mean': 0.08704587444663048, 'clip_ratio/low_min': 0.08704587444663048, 'clip_ratio/high_mean': 0.07331375125795603, 'clip_ratio/high_max': 0.07331375125795603, 'clip_ratio/region_mean': 0.1603596257045865, 'reward_total_mean': 0.1650734692811966, 'reward_meter_mean': 0.3729636073112488, 'reward_meter_std': 0.33335843682289124, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9930555820465088, 'reward_lexical_plausibility_std': 0.01285861898213625, 'reward_repeat_penalty_mean': 0.9592209458351135, 'reward_repeat_penalty_std': 0.05305812135338783, 'reward_repeat_floor_mean': 0.9949278235435486, 'reward_repeat_floor_std': 0.006081957370042801, 'reward_near_duplicate_penalty_mean': 0.9780904054641724, 'reward_near_duplicate_penalty_std': 0.020435959100723267, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9801587462425232, 'reward_distinct_2_std': 0.036883533000946045, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2745385766029358, 'reward_total_composite_mean': 0.1650734692811966, 'reward_total_composite_std': 0.1427987813949585, 'epoch': 0.02}

 13%|█▎        | 416/3300 [1:30:37<8:22:33, 10.46s/it]INFO 04-12 17:29:03 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 417/3300 [1:30:45<7:46:01,  9.70s/it]2026-04-12 17:29:11,237 | INFO | train_grpo_train | metrics_logged mode=train step=417

                                                      
{'loss': 0.2634, 'grad_norm': 18.437881469726562, 'learning_rate': 8.73939393939394e-06, 'num_tokens': 779137.0, 'completions/mean_length': 31.625, 'completions/min_length': 20.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7448065876960754, 'rewards/meter/std': 0.35229867696762085, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.16035674512386322, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3046778440475464, 'rewards/repeat_penalty/mean': 0.8524281978607178, 'rewards/repeat_penalty/std': 0.11421186476945877, 'rewards/repeat_floor/mean': 0.989524245262146, 'rewards/repeat_floor/std': 0.0061520677991211414, 'rewards/near_duplicate_penalty/mean': 0.9849691390991211, 'rewards/near_duplicate_penalty/std': 0.020624451339244843, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.452991247177124, 'rewards/total_composite/std': 0.3771788477897644, 'reward': 0.452991247177124, 'reward_std': 0.377178817987442, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19702403247356415, 'sampling/sampling_logp_difference/max': 1.301346778869629, 'sampling/importance_sampling_ratio/min': 0.2721650004386902, 'sampling/importance_sampling_ratio/mean': 1.0572483539581299, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6861381083726883, 'clip_ratio/low_mean': 0.07361185643821955, 'clip_ratio/low_min': 0.07361185643821955, 'clip_ratio/high_mean': 0.05312500149011612, 'clip_ratio/high_max': 0.05312500149011612, 'clip_ratio/region_mean': 0.12673685792833567, 'reward_total_mean': 0.452991247177124, 'reward_meter_mean': 0.7448065876960754, 'reward_meter_std': 0.35229867696762085, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.16035674512386322, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8524281978607178, 'reward_repeat_penalty_std': 0.11421186476945877, 'reward_repeat_floor_mean': 0.989524245262146, 'reward_repeat_floor_std': 0.0061520677991211414, 'reward_near_duplicate_penalty_mean': 0.9849691390991211, 'reward_near_duplicate_penalty_std': 0.020624451339244843, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3046778440475464, 'reward_total_composite_mean': 0.452991247177124, 'reward_total_composite_std': 0.3771788477897644, 'epoch': 0.02}

 13%|█▎        | 417/3300 [1:30:45<7:46:01,  9.70s/it]INFO 04-12 17:29:11 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 418/3300 [1:30:59<8:55:22, 11.15s/it]2026-04-12 17:29:25,758 | INFO | train_grpo_train | metrics_logged mode=train step=418

                                                      
{'loss': -0.0278, 'grad_norm': 5.720499515533447, 'learning_rate': 8.736363636363638e-06, 'num_tokens': 781137.0, 'completions/mean_length': 134.0, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.2404271513223648, 'rewards/meter/std': 0.24963420629501343, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.8755707740783691, 'rewards/lexical_plausibility/std': 0.17385001480579376, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.5560752153396606, 'rewards/repeat_penalty/std': 0.31644949316978455, 'rewards/repeat_floor/mean': 0.9357200860977173, 'rewards/repeat_floor/std': 0.05104459077119827, 'rewards/near_duplicate_penalty/mean': 0.7400911450386047, 'rewards/near_duplicate_penalty/std': 0.2015363723039627, 'rewards/opening_diversity/mean': 0.8515625, 'rewards/opening_diversity/std': 0.3454705476760864, 'rewards/distinct_2/mean': 0.8179179430007935, 'rewards/distinct_2/std': 0.13774771988391876, 'rewards/total_composite/mean': 0.08427922427654266, 'rewards/total_composite/std': 0.09055059403181076, 'reward': 0.08427922427654266, 'reward_std': 0.09055058658123016, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14244744181632996, 'sampling/sampling_logp_difference/max': 2.121565580368042, 'sampling/importance_sampling_ratio/min': 0.11984385550022125, 'sampling/importance_sampling_ratio/mean': 1.0242865085601807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9750370159745216, 'clip_ratio/low_mean': 0.05169749353080988, 'clip_ratio/low_min': 0.05169749353080988, 'clip_ratio/high_mean': 0.06031937804073095, 'clip_ratio/high_max': 0.06031937804073095, 'clip_ratio/region_mean': 0.11201687157154083, 'reward_total_mean': 0.08427922427654266, 'reward_meter_mean': 0.2404271513223648, 'reward_meter_std': 0.24963420629501343, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.8755707740783691, 'reward_lexical_plausibility_std': 0.17385001480579376, 'reward_repeat_penalty_mean': 0.5560752153396606, 'reward_repeat_penalty_std': 0.31644949316978455, 'reward_repeat_floor_mean': 0.9357200860977173, 'reward_repeat_floor_std': 0.05104459077119827, 'reward_near_duplicate_penalty_mean': 0.7400911450386047, 'reward_near_duplicate_penalty_std': 0.2015363723039627, 'reward_opening_diversity_mean': 0.8515625, 'reward_opening_diversity_std': 0.3454705476760864, 'reward_distinct_2_mean': 0.8179179430007935, 'reward_distinct_2_std': 0.13774771988391876, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.08427922427654266, 'reward_total_composite_std': 0.09055059403181076, 'epoch': 0.02}

 13%|█▎        | 418/3300 [1:30:59<8:55:22, 11.15s/it]INFO 04-12 17:29:25 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 419/3300 [1:31:13<9:31:17, 11.90s/it]2026-04-12 17:29:39,410 | INFO | train_grpo_train | metrics_logged mode=train step=419

                                                      
{'loss': -0.06, 'grad_norm': 4.3258056640625, 'learning_rate': 8.733333333333333e-06, 'num_tokens': 783417.0, 'completions/mean_length': 137.0, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.42857360839844, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.7313331961631775, 'rewards/meter/std': 0.3466825485229492, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9925000071525574, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9555842280387878, 'rewards/lexical_plausibility/std': 0.12562675774097443, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.09425459802150726, 'rewards/repeat_penalty/mean': 0.47216612100601196, 'rewards/repeat_penalty/std': 0.28925347328186035, 'rewards/repeat_floor/mean': 0.9374450445175171, 'rewards/repeat_floor/std': 0.03818248212337494, 'rewards/near_duplicate_penalty/mean': 0.8511825203895569, 'rewards/near_duplicate_penalty/std': 0.08535610139369965, 'rewards/opening_diversity/mean': 0.9750000238418579, 'rewards/opening_diversity/std': 0.04629101976752281, 'rewards/distinct_2/mean': 0.569640576839447, 'rewards/distinct_2/std': 0.2676936686038971, 'rewards/total_composite/mean': 0.11668050289154053, 'rewards/total_composite/std': 0.09349052608013153, 'reward': 0.11668050289154053, 'reward_std': 0.09349052608013153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15329334139823914, 'sampling/sampling_logp_difference/max': 1.516845703125, 'sampling/importance_sampling_ratio/min': 0.21940286457538605, 'sampling/importance_sampling_ratio/mean': 1.0286436080932617, 'sampling/importance_sampling_ratio/max': 1.940280795097351, 'entropy': 1.1357912719249725, 'clip_ratio/low_mean': 0.037004128098487854, 'clip_ratio/low_min': 0.037004128098487854, 'clip_ratio/high_mean': 0.09222503937780857, 'clip_ratio/high_max': 0.09222503937780857, 'clip_ratio/region_mean': 0.12922916747629642, 'reward_total_mean': 0.11668050289154053, 'reward_meter_mean': 0.7313331961631775, 'reward_meter_std': 0.3466825485229492, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9925000071525574, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9555842280387878, 'reward_lexical_plausibility_std': 0.12562675774097443, 'reward_repeat_penalty_mean': 0.47216612100601196, 'reward_repeat_penalty_std': 0.28925347328186035, 'reward_repeat_floor_mean': 0.9374450445175171, 'reward_repeat_floor_std': 0.03818248212337494, 'reward_near_duplicate_penalty_mean': 0.8511825203895569, 'reward_near_duplicate_penalty_std': 0.08535610139369965, 'reward_opening_diversity_mean': 0.9750000238418579, 'reward_opening_diversity_std': 0.04629101976752281, 'reward_distinct_2_mean': 0.569640576839447, 'reward_distinct_2_std': 0.2676936686038971, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.09425459802150726, 'reward_total_composite_mean': 0.11668050289154053, 'reward_total_composite_std': 0.09349052608013153, 'epoch': 0.02}

 13%|█▎        | 419/3300 [1:31:13<9:31:17, 11.90s/it]INFO 04-12 17:29:39 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 420/3300 [1:31:26<9:53:24, 12.36s/it]2026-04-12 17:29:52,857 | INFO | train_grpo_train | metrics_logged mode=train step=420

                                                      
{'loss': -0.0663, 'grad_norm': 4.583530902862549, 'learning_rate': 8.73030303030303e-06, 'num_tokens': 785146.0, 'completions/mean_length': 104.125, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.85714340209961, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8950604200363159, 'rewards/meter/std': 0.10852807760238647, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9589052200317383, 'rewards/lexical_plausibility/std': 0.11623367667198181, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.8724360466003418, 'rewards/repeat_penalty/std': 0.10166767239570618, 'rewards/repeat_floor/mean': 0.9860841035842896, 'rewards/repeat_floor/std': 0.010874365456402302, 'rewards/near_duplicate_penalty/mean': 0.9505135416984558, 'rewards/near_duplicate_penalty/std': 0.04566945135593414, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9486898183822632, 'rewards/distinct_2/std': 0.06557372212409973, 'rewards/total_composite/mean': 0.3423081040382385, 'rewards/total_composite/std': 0.25631964206695557, 'reward': 0.3423081040382385, 'reward_std': 0.25631964206695557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16960261762142181, 'sampling/sampling_logp_difference/max': 1.1214208602905273, 'sampling/importance_sampling_ratio/min': 0.3258165121078491, 'sampling/importance_sampling_ratio/mean': 1.0269588232040405, 'sampling/importance_sampling_ratio/max': 1.7959256172180176, 'entropy': 1.2777791321277618, 'clip_ratio/low_mean': 0.07431362010538578, 'clip_ratio/low_min': 0.07431362010538578, 'clip_ratio/high_mean': 0.06017801910638809, 'clip_ratio/high_max': 0.06017801910638809, 'clip_ratio/region_mean': 0.13449163921177387, 'reward_total_mean': 0.3423081040382385, 'reward_meter_mean': 0.8950604200363159, 'reward_meter_std': 0.10852807760238647, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9589052200317383, 'reward_lexical_plausibility_std': 0.11623367667198181, 'reward_repeat_penalty_mean': 0.8724360466003418, 'reward_repeat_penalty_std': 0.10166767239570618, 'reward_repeat_floor_mean': 0.9860841035842896, 'reward_repeat_floor_std': 0.010874365456402302, 'reward_near_duplicate_penalty_mean': 0.9505135416984558, 'reward_near_duplicate_penalty_std': 0.04566945135593414, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9486898183822632, 'reward_distinct_2_std': 0.06557372212409973, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.3423081040382385, 'reward_total_composite_std': 0.25631964206695557, 'epoch': 0.02}

 13%|█▎        | 420/3300 [1:31:26<9:53:24, 12.36s/it]INFO 04-12 17:29:53 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 421/3300 [1:31:35<9:02:08, 11.30s/it]2026-04-12 17:30:01,673 | INFO | train_grpo_train | metrics_logged mode=train step=421

                                                      
{'loss': 0.0642, 'grad_norm': 11.137006759643555, 'learning_rate': 8.727272727272728e-06, 'num_tokens': 786939.0, 'completions/mean_length': 59.125, 'completions/min_length': 52.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7526957392692566, 'rewards/meter/std': 0.367506206035614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.8673490285873413, 'rewards/repeat_penalty/std': 0.08890773355960846, 'rewards/repeat_floor/mean': 0.9845092296600342, 'rewards/repeat_floor/std': 0.009965740144252777, 'rewards/near_duplicate_penalty/mean': 0.9495429992675781, 'rewards/near_duplicate_penalty/std': 0.024446433410048485, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9119751453399658, 'rewards/distinct_2/std': 0.07285679876804352, 'rewards/total_composite/mean': 0.31252261996269226, 'rewards/total_composite/std': 0.27877277135849, 'reward': 0.31252261996269226, 'reward_std': 0.2787728011608124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19893643260002136, 'sampling/sampling_logp_difference/max': 1.8573691844940186, 'sampling/importance_sampling_ratio/min': 0.15608270466327667, 'sampling/importance_sampling_ratio/mean': 1.010530710220337, 'sampling/importance_sampling_ratio/max': 1.9983532428741455, 'entropy': 1.5182730108499527, 'clip_ratio/low_mean': 0.08075428754091263, 'clip_ratio/low_min': 0.08075428754091263, 'clip_ratio/high_mean': 0.09795355051755905, 'clip_ratio/high_max': 0.09795355051755905, 'clip_ratio/region_mean': 0.17870783805847168, 'reward_total_mean': 0.31252261996269226, 'reward_meter_mean': 0.7526957392692566, 'reward_meter_std': 0.367506206035614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8673490285873413, 'reward_repeat_penalty_std': 0.08890773355960846, 'reward_repeat_floor_mean': 0.9845092296600342, 'reward_repeat_floor_std': 0.009965740144252777, 'reward_near_duplicate_penalty_mean': 0.9495429992675781, 'reward_near_duplicate_penalty_std': 0.024446433410048485, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9119751453399658, 'reward_distinct_2_std': 0.07285679876804352, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.31252261996269226, 'reward_total_composite_std': 0.27877277135849, 'epoch': 0.02}

 13%|█▎        | 421/3300 [1:31:35<9:02:08, 11.30s/it]INFO 04-12 17:30:01 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 422/3300 [1:31:45<8:38:38, 10.81s/it]2026-04-12 17:30:11,360 | INFO | train_grpo_train | metrics_logged mode=train step=422

                                                      
{'loss': 0.0672, 'grad_norm': 12.562152862548828, 'learning_rate': 8.724242424242425e-06, 'num_tokens': 788951.0, 'completions/mean_length': 66.5, 'completions/min_length': 56.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.5, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.7505141496658325, 'rewards/meter/std': 0.29966267943382263, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9925000071525574, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19905133545398712, 'rewards/repeat_penalty/mean': 0.8428101539611816, 'rewards/repeat_penalty/std': 0.15667498111724854, 'rewards/repeat_floor/mean': 0.9832167625427246, 'rewards/repeat_floor/std': 0.014885702170431614, 'rewards/near_duplicate_penalty/mean': 0.95533287525177, 'rewards/near_duplicate_penalty/std': 0.029687827453017235, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8879644870758057, 'rewards/distinct_2/std': 0.12172655016183853, 'rewards/total_composite/mean': 0.4289339780807495, 'rewards/total_composite/std': 0.26847225427627563, 'reward': 0.4289339780807495, 'reward_std': 0.26847225427627563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14342275261878967, 'sampling/sampling_logp_difference/max': 1.6937564611434937, 'sampling/importance_sampling_ratio/min': 0.1838276982307434, 'sampling/importance_sampling_ratio/mean': 1.010650396347046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7265492081642151, 'clip_ratio/low_mean': 0.0837098378688097, 'clip_ratio/low_min': 0.0837098378688097, 'clip_ratio/high_mean': 0.039722432382404804, 'clip_ratio/high_max': 0.039722432382404804, 'clip_ratio/region_mean': 0.1234322702512145, 'reward_total_mean': 0.4289339780807495, 'reward_meter_mean': 0.7505141496658325, 'reward_meter_std': 0.29966267943382263, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9925000071525574, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8428101539611816, 'reward_repeat_penalty_std': 0.15667498111724854, 'reward_repeat_floor_mean': 0.9832167625427246, 'reward_repeat_floor_std': 0.014885702170431614, 'reward_near_duplicate_penalty_mean': 0.95533287525177, 'reward_near_duplicate_penalty_std': 0.029687827453017235, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8879644870758057, 'reward_distinct_2_std': 0.12172655016183853, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19905133545398712, 'reward_total_composite_mean': 0.4289339780807495, 'reward_total_composite_std': 0.26847225427627563, 'epoch': 0.02}

 13%|█▎        | 422/3300 [1:31:45<8:38:38, 10.81s/it]INFO 04-12 17:30:11 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 423/3300 [1:31:53<8:07:17, 10.16s/it]2026-04-12 17:30:19,999 | INFO | train_grpo_train | metrics_logged mode=train step=423

                                                      
{'loss': 0.0614, 'grad_norm': 17.327491760253906, 'learning_rate': 8.72121212121212e-06, 'num_tokens': 790519.0, 'completions/mean_length': 40.0, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7543220520019531, 'rewards/meter/std': 0.33283770084381104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.219983771443367, 'rewards/repeat_penalty/mean': 0.9855372905731201, 'rewards/repeat_penalty/std': 0.017431318759918213, 'rewards/repeat_floor/mean': 0.9978306293487549, 'rewards/repeat_floor/std': 0.0026147011667490005, 'rewards/near_duplicate_penalty/mean': 0.9855372905731201, 'rewards/near_duplicate_penalty/std': 0.017431318759918213, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33214086294174194, 'rewards/total_composite/std': 0.2602840065956116, 'reward': 0.33214086294174194, 'reward_std': 0.2602840065956116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19631245732307434, 'sampling/sampling_logp_difference/max': 1.9757099151611328, 'sampling/importance_sampling_ratio/min': 0.13866284489631653, 'sampling/importance_sampling_ratio/mean': 1.0412321090698242, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7369295060634613, 'clip_ratio/low_mean': 0.09031642321497202, 'clip_ratio/low_min': 0.09031642321497202, 'clip_ratio/high_mean': 0.12331311963498592, 'clip_ratio/high_max': 0.12331311963498592, 'clip_ratio/region_mean': 0.21362954284995794, 'reward_total_mean': 0.33214086294174194, 'reward_meter_mean': 0.7543220520019531, 'reward_meter_std': 0.33283770084381104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9855372905731201, 'reward_repeat_penalty_std': 0.017431318759918213, 'reward_repeat_floor_mean': 0.9978306293487549, 'reward_repeat_floor_std': 0.0026147011667490005, 'reward_near_duplicate_penalty_mean': 0.9855372905731201, 'reward_near_duplicate_penalty_std': 0.017431318759918213, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.219983771443367, 'reward_total_composite_mean': 0.33214086294174194, 'reward_total_composite_std': 0.2602840065956116, 'epoch': 0.02}

 13%|█▎        | 423/3300 [1:31:53<8:07:17, 10.16s/it]INFO 04-12 17:30:20 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 424/3300 [1:32:03<8:05:40, 10.13s/it]2026-04-12 17:30:30,058 | INFO | train_grpo_train | metrics_logged mode=train step=424

                                                      
{'loss': 0.03, 'grad_norm': 10.294676780700684, 'learning_rate': 8.71818181818182e-06, 'num_tokens': 792719.0, 'completions/mean_length': 93.0, 'completions/min_length': 84.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.4554365873336792, 'rewards/meter/std': 0.26305270195007324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.8240758180618286, 'rewards/repeat_penalty/std': 0.17134004831314087, 'rewards/repeat_floor/mean': 0.9790796041488647, 'rewards/repeat_floor/std': 0.022784223780035973, 'rewards/near_duplicate_penalty/mean': 0.9403244256973267, 'rewards/near_duplicate_penalty/std': 0.06833906471729279, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8722193241119385, 'rewards/distinct_2/std': 0.1265467256307602, 'rewards/total_composite/mean': 0.17871516942977905, 'rewards/total_composite/std': 0.07699020206928253, 'reward': 0.17871516942977905, 'reward_std': 0.07699019461870193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1846231371164322, 'sampling/sampling_logp_difference/max': 1.644737720489502, 'sampling/importance_sampling_ratio/min': 0.1930631846189499, 'sampling/importance_sampling_ratio/mean': 1.0134590864181519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2414982467889786, 'clip_ratio/low_mean': 0.08379538543522358, 'clip_ratio/low_min': 0.08379538543522358, 'clip_ratio/high_mean': 0.07203113846480846, 'clip_ratio/high_max': 0.07203113846480846, 'clip_ratio/region_mean': 0.15582652390003204, 'reward_total_mean': 0.17871516942977905, 'reward_meter_mean': 0.4554365873336792, 'reward_meter_std': 0.26305270195007324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8240758180618286, 'reward_repeat_penalty_std': 0.17134004831314087, 'reward_repeat_floor_mean': 0.9790796041488647, 'reward_repeat_floor_std': 0.022784223780035973, 'reward_near_duplicate_penalty_mean': 0.9403244256973267, 'reward_near_duplicate_penalty_std': 0.06833906471729279, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8722193241119385, 'reward_distinct_2_std': 0.1265467256307602, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.17871516942977905, 'reward_total_composite_std': 0.07699020206928253, 'epoch': 0.02}

 13%|█▎        | 424/3300 [1:32:03<8:05:40, 10.13s/it]INFO 04-12 17:30:30 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 425/3300 [1:32:11<7:26:32,  9.32s/it]2026-04-12 17:30:37,481 | INFO | train_grpo_train | metrics_logged mode=train step=425

                                                      
{'loss': -0.0075, 'grad_norm': 38.55045700073242, 'learning_rate': 8.715151515151515e-06, 'num_tokens': 794072.0, 'completions/mean_length': 16.125, 'completions/min_length': 11.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.125, 'completions/min_terminated_length': 11.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.4318186640739441, 'rewards/meter/std': 0.36611291766166687, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9350000023841858, 'rewards/judge_quality/std': 0.0160356592386961, 'rewards/repeat_penalty/mean': 0.6771972179412842, 'rewards/repeat_penalty/std': 0.05999475345015526, 'rewards/repeat_floor/mean': 0.9688812494277954, 'rewards/repeat_floor/std': 0.01180236879736185, 'rewards/near_duplicate_penalty/mean': 0.894464910030365, 'rewards/near_duplicate_penalty/std': 0.07809729874134064, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9759615659713745, 'rewards/distinct_2/std': 0.06799104809761047, 'rewards/total_composite/mean': 0.3588297963142395, 'rewards/total_composite/std': 0.29621267318725586, 'reward': 0.3588297963142395, 'reward_std': 0.29621270298957825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20691446959972382, 'sampling/sampling_logp_difference/max': 1.2421596050262451, 'sampling/importance_sampling_ratio/min': 0.2887599468231201, 'sampling/importance_sampling_ratio/mean': 1.0438870191574097, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0148386992514133, 'clip_ratio/low_mean': 0.05116421636193991, 'clip_ratio/low_min': 0.05116421636193991, 'clip_ratio/high_mean': 0.12234322540462017, 'clip_ratio/high_max': 0.12234322540462017, 'clip_ratio/region_mean': 0.17350744176656008, 'reward_total_mean': 0.3588297963142395, 'reward_meter_mean': 0.4318186640739441, 'reward_meter_std': 0.36611291766166687, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6771972179412842, 'reward_repeat_penalty_std': 0.05999475345015526, 'reward_repeat_floor_mean': 0.9688812494277954, 'reward_repeat_floor_std': 0.01180236879736185, 'reward_near_duplicate_penalty_mean': 0.894464910030365, 'reward_near_duplicate_penalty_std': 0.07809729874134064, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9759615659713745, 'reward_distinct_2_std': 0.06799104809761047, 'reward_judge_quality_mean': 0.9350000023841858, 'reward_judge_quality_std': 0.0160356592386961, 'reward_total_composite_mean': 0.3588297963142395, 'reward_total_composite_std': 0.29621267318725586, 'epoch': 0.02}

 13%|█▎        | 425/3300 [1:32:11<7:26:32,  9.32s/it]INFO 04-12 17:30:37 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 426/3300 [1:32:21<7:42:15,  9.65s/it]2026-04-12 17:30:47,907 | INFO | train_grpo_train | metrics_logged mode=train step=426

                                                      
{'loss': 0.0984, 'grad_norm': 8.75532341003418, 'learning_rate': 8.712121212121212e-06, 'num_tokens': 796359.0, 'completions/mean_length': 95.875, 'completions/min_length': 75.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.8751838207244873, 'rewards/meter/std': 0.1776522397994995, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.6559178829193115, 'rewards/repeat_penalty/std': 0.31605398654937744, 'rewards/repeat_floor/mean': 0.9622302055358887, 'rewards/repeat_floor/std': 0.04364832490682602, 'rewards/near_duplicate_penalty/mean': 0.921332836151123, 'rewards/near_duplicate_penalty/std': 0.09565689414739609, 'rewards/opening_diversity/mean': 0.9508928656578064, 'rewards/opening_diversity/std': 0.09533105045557022, 'rewards/distinct_2/mean': 0.744186282157898, 'rewards/distinct_2/std': 0.2703051269054413, 'rewards/total_composite/mean': 0.24232810735702515, 'rewards/total_composite/std': 0.11076441407203674, 'reward': 0.24232810735702515, 'reward_std': 0.11076439917087555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15914596617221832, 'sampling/sampling_logp_difference/max': 1.3651857376098633, 'sampling/importance_sampling_ratio/min': 0.2553332448005676, 'sampling/importance_sampling_ratio/mean': 1.0339637994766235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.381018303334713, 'clip_ratio/low_mean': 0.0544103579595685, 'clip_ratio/low_min': 0.0544103579595685, 'clip_ratio/high_mean': 0.08059742767363787, 'clip_ratio/high_max': 0.08059742767363787, 'clip_ratio/region_mean': 0.13500778563320637, 'reward_total_mean': 0.24232810735702515, 'reward_meter_mean': 0.8751838207244873, 'reward_meter_std': 0.1776522397994995, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6559178829193115, 'reward_repeat_penalty_std': 0.31605398654937744, 'reward_repeat_floor_mean': 0.9622302055358887, 'reward_repeat_floor_std': 0.04364832490682602, 'reward_near_duplicate_penalty_mean': 0.921332836151123, 'reward_near_duplicate_penalty_std': 0.09565689414739609, 'reward_opening_diversity_mean': 0.9508928656578064, 'reward_opening_diversity_std': 0.09533105045557022, 'reward_distinct_2_mean': 0.744186282157898, 'reward_distinct_2_std': 0.2703051269054413, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.24232810735702515, 'reward_total_composite_std': 0.11076441407203674, 'epoch': 0.02}

 13%|█▎        | 426/3300 [1:32:21<7:42:15,  9.65s/it]INFO 04-12 17:30:48 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 427/3300 [1:32:30<7:30:14,  9.40s/it]2026-04-12 17:30:56,731 | INFO | train_grpo_train | metrics_logged mode=train step=427

                                                      
{'loss': 0.0511, 'grad_norm': 14.312180519104004, 'learning_rate': 8.70909090909091e-06, 'num_tokens': 797743.0, 'completions/mean_length': 38.0, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.49000275135040283, 'rewards/meter/std': 0.3732234537601471, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9409509897232056, 'rewards/repeat_penalty/std': 0.054482586681842804, 'rewards/repeat_floor/mean': 0.9925425052642822, 'rewards/repeat_floor/std': 0.006816257257014513, 'rewards/near_duplicate_penalty/mean': 0.9660857915878296, 'rewards/near_duplicate_penalty/std': 0.03407246991991997, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9736621975898743, 'rewards/distinct_2/std': 0.03648125380277634, 'rewards/total_composite/mean': 0.17703327536582947, 'rewards/total_composite/std': 0.15678516030311584, 'reward': 0.17703327536582947, 'reward_std': 0.15678516030311584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1936989724636078, 'sampling/sampling_logp_difference/max': 2.4302167892456055, 'sampling/importance_sampling_ratio/min': 0.08801774680614471, 'sampling/importance_sampling_ratio/mean': 1.0226032733917236, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7115192711353302, 'clip_ratio/low_mean': 0.11790913436561823, 'clip_ratio/low_min': 0.11790913436561823, 'clip_ratio/high_mean': 0.05874060280621052, 'clip_ratio/high_max': 0.05874060280621052, 'clip_ratio/region_mean': 0.17664973717182875, 'reward_total_mean': 0.17703327536582947, 'reward_meter_mean': 0.49000275135040283, 'reward_meter_std': 0.3732234537601471, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9409509897232056, 'reward_repeat_penalty_std': 0.054482586681842804, 'reward_repeat_floor_mean': 0.9925425052642822, 'reward_repeat_floor_std': 0.006816257257014513, 'reward_near_duplicate_penalty_mean': 0.9660857915878296, 'reward_near_duplicate_penalty_std': 0.03407246991991997, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9736621975898743, 'reward_distinct_2_std': 0.03648125380277634, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.17703327536582947, 'reward_total_composite_std': 0.15678516030311584, 'epoch': 0.02}

 13%|█▎        | 427/3300 [1:32:30<7:30:14,  9.40s/it]INFO 04-12 17:30:56 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 428/3300 [1:32:38<7:06:40,  8.91s/it]2026-04-12 17:31:04,528 | INFO | train_grpo_train | metrics_logged mode=train step=428

                                                      
{'loss': 0.1184, 'grad_norm': 21.285945892333984, 'learning_rate': 8.706060606060607e-06, 'num_tokens': 799048.0, 'completions/mean_length': 20.125, 'completions/min_length': 17.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8460110425949097, 'rewards/meter/std': 0.34095120429992676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.3620477616786957, 'rewards/repeat_penalty/mean': 0.73777174949646, 'rewards/repeat_penalty/std': 0.03458673879504204, 'rewards/repeat_floor/mean': 0.9825543761253357, 'rewards/repeat_floor/std': 0.006917351391166449, 'rewards/near_duplicate_penalty/mean': 0.9836956262588501, 'rewards/near_duplicate_penalty/std': 0.04611567035317421, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5261960625648499, 'rewards/total_composite/std': 0.3867631256580353, 'reward': 0.5261960625648499, 'reward_std': 0.38676315546035767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15875288844108582, 'sampling/sampling_logp_difference/max': 0.7255430221557617, 'sampling/importance_sampling_ratio/min': 0.48406165838241577, 'sampling/importance_sampling_ratio/mean': 1.0304712057113647, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5733484327793121, 'clip_ratio/low_mean': 0.05192877119407058, 'clip_ratio/low_min': 0.05192877119407058, 'clip_ratio/high_mean': 0.08513932023197412, 'clip_ratio/high_max': 0.08513932023197412, 'clip_ratio/region_mean': 0.1370680914260447, 'reward_total_mean': 0.5261960625648499, 'reward_meter_mean': 0.8460110425949097, 'reward_meter_std': 0.34095120429992676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.73777174949646, 'reward_repeat_penalty_std': 0.03458673879504204, 'reward_repeat_floor_mean': 0.9825543761253357, 'reward_repeat_floor_std': 0.006917351391166449, 'reward_near_duplicate_penalty_mean': 0.9836956262588501, 'reward_near_duplicate_penalty_std': 0.04611567035317421, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.3620477616786957, 'reward_total_composite_mean': 0.5261960625648499, 'reward_total_composite_std': 0.3867631256580353, 'epoch': 0.02}

 13%|█▎        | 428/3300 [1:32:38<7:06:40,  8.91s/it]INFO 04-12 17:31:04 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 429/3300 [1:32:52<8:28:28, 10.63s/it]2026-04-12 17:31:19,148 | INFO | train_grpo_train | metrics_logged mode=train step=429

                                                      
{'loss': -0.0691, 'grad_norm': 4.137322902679443, 'learning_rate': 8.703030303030304e-06, 'num_tokens': 800575.0, 'completions/mean_length': 99.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.763437032699585, 'rewards/meter/std': 0.40327855944633484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9443026781082153, 'rewards/lexical_plausibility/std': 0.15753573179244995, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.7007288336753845, 'rewards/repeat_penalty/std': 0.3058701753616333, 'rewards/repeat_floor/mean': 0.9595679640769958, 'rewards/repeat_floor/std': 0.05686198174953461, 'rewards/near_duplicate_penalty/mean': 0.8809818029403687, 'rewards/near_duplicate_penalty/std': 0.15458548069000244, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.8637024164199829, 'rewards/distinct_2/std': 0.2479252815246582, 'rewards/total_composite/mean': 0.3028221130371094, 'rewards/total_composite/std': 0.17363055050373077, 'reward': 0.3028221130371094, 'reward_std': 0.17363055050373077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1624930500984192, 'sampling/sampling_logp_difference/max': 2.1269805431365967, 'sampling/importance_sampling_ratio/min': 0.11919666081666946, 'sampling/importance_sampling_ratio/mean': 1.049901008605957, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0335667952895164, 'clip_ratio/low_mean': 0.04320652224123478, 'clip_ratio/low_min': 0.04320652224123478, 'clip_ratio/high_mean': 0.0928165651857853, 'clip_ratio/high_max': 0.0928165651857853, 'clip_ratio/region_mean': 0.13602308742702007, 'reward_total_mean': 0.3028221130371094, 'reward_meter_mean': 0.763437032699585, 'reward_meter_std': 0.40327855944633484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9443026781082153, 'reward_lexical_plausibility_std': 0.15753573179244995, 'reward_repeat_penalty_mean': 0.7007288336753845, 'reward_repeat_penalty_std': 0.3058701753616333, 'reward_repeat_floor_mean': 0.9595679640769958, 'reward_repeat_floor_std': 0.05686198174953461, 'reward_near_duplicate_penalty_mean': 0.8809818029403687, 'reward_near_duplicate_penalty_std': 0.15458548069000244, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.8637024164199829, 'reward_distinct_2_std': 0.2479252815246582, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.3028221130371094, 'reward_total_composite_std': 0.17363055050373077, 'epoch': 0.02}

 13%|█▎        | 429/3300 [1:32:53<8:28:28, 10.63s/it]INFO 04-12 17:31:19 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 430/3300 [1:33:02<8:07:22, 10.19s/it]2026-04-12 17:31:28,317 | INFO | train_grpo_train | metrics_logged mode=train step=430

                                                      
{'loss': 0.047, 'grad_norm': 14.473419189453125, 'learning_rate': 8.700000000000001e-06, 'num_tokens': 802063.0, 'completions/mean_length': 41.0, 'completions/min_length': 35.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7545582056045532, 'rewards/meter/std': 0.38726872205734253, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5962499976158142, 'rewards/judge_quality/std': 0.26065507531166077, 'rewards/repeat_penalty/mean': 0.914974570274353, 'rewards/repeat_penalty/std': 0.11135981976985931, 'rewards/repeat_floor/mean': 0.9899146556854248, 'rewards/repeat_floor/std': 0.0101882079616189, 'rewards/near_duplicate_penalty/mean': 0.9500719308853149, 'rewards/near_duplicate_penalty/std': 0.04354565590620041, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.38103383779525757, 'rewards/total_composite/std': 0.23693832755088806, 'reward': 0.38103383779525757, 'reward_std': 0.23693831264972687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.170084610581398, 'sampling/sampling_logp_difference/max': 1.1645030975341797, 'sampling/importance_sampling_ratio/min': 0.31207767128944397, 'sampling/importance_sampling_ratio/mean': 1.0287314653396606, 'sampling/importance_sampling_ratio/max': 1.9262477159500122, 'entropy': 1.3900617808103561, 'clip_ratio/low_mean': 0.0738748712465167, 'clip_ratio/low_min': 0.0738748712465167, 'clip_ratio/high_mean': 0.11073058936744928, 'clip_ratio/high_max': 0.11073058936744928, 'clip_ratio/region_mean': 0.184605460613966, 'reward_total_mean': 0.38103383779525757, 'reward_meter_mean': 0.7545582056045532, 'reward_meter_std': 0.38726872205734253, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.914974570274353, 'reward_repeat_penalty_std': 0.11135981976985931, 'reward_repeat_floor_mean': 0.9899146556854248, 'reward_repeat_floor_std': 0.0101882079616189, 'reward_near_duplicate_penalty_mean': 0.9500719308853149, 'reward_near_duplicate_penalty_std': 0.04354565590620041, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.5962499976158142, 'reward_judge_quality_std': 0.26065507531166077, 'reward_total_composite_mean': 0.38103383779525757, 'reward_total_composite_std': 0.23693832755088806, 'epoch': 0.02}

 13%|█▎        | 430/3300 [1:33:02<8:07:22, 10.19s/it]INFO 04-12 17:31:28 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 431/3300 [1:33:10<7:36:48,  9.55s/it]2026-04-12 17:31:36,386 | INFO | train_grpo_train | metrics_logged mode=train step=431

                                                      
{'loss': -0.0516, 'grad_norm': 20.342313766479492, 'learning_rate': 8.696969696969699e-06, 'num_tokens': 803343.0, 'completions/mean_length': 22.0, 'completions/min_length': 14.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.31310927867889404, 'rewards/meter/std': 0.40659424662590027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9321384429931641, 'rewards/lexical_plausibility/std': 0.10931902378797531, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.39445942640304565, 'rewards/repeat_penalty/mean': 0.8071969747543335, 'rewards/repeat_penalty/std': 0.10651341080665588, 'rewards/repeat_floor/mean': 0.9879545569419861, 'rewards/repeat_floor/std': 0.005729019176214933, 'rewards/near_duplicate_penalty/mean': 0.9946969747543335, 'rewards/near_duplicate_penalty/std': 0.01499924249947071, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1963495910167694, 'rewards/total_composite/std': 0.3161138594150543, 'reward': 0.1963495910167694, 'reward_std': 0.3161138594150543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18365509808063507, 'sampling/sampling_logp_difference/max': 1.4172420501708984, 'sampling/importance_sampling_ratio/min': 0.24238158762454987, 'sampling/importance_sampling_ratio/mean': 1.0191051959991455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2575480788946152, 'clip_ratio/low_mean': 0.09572916757315397, 'clip_ratio/low_min': 0.09572916757315397, 'clip_ratio/high_mean': 0.04442307632416487, 'clip_ratio/high_max': 0.04442307632416487, 'clip_ratio/region_mean': 0.14015224389731884, 'reward_total_mean': 0.1963495910167694, 'reward_meter_mean': 0.31310927867889404, 'reward_meter_std': 0.40659424662590027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9321384429931641, 'reward_lexical_plausibility_std': 0.10931902378797531, 'reward_repeat_penalty_mean': 0.8071969747543335, 'reward_repeat_penalty_std': 0.10651341080665588, 'reward_repeat_floor_mean': 0.9879545569419861, 'reward_repeat_floor_std': 0.005729019176214933, 'reward_near_duplicate_penalty_mean': 0.9946969747543335, 'reward_near_duplicate_penalty_std': 0.01499924249947071, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.39445942640304565, 'reward_total_composite_mean': 0.1963495910167694, 'reward_total_composite_std': 0.3161138594150543, 'epoch': 0.02}

 13%|█▎        | 431/3300 [1:33:10<7:36:48,  9.55s/it]INFO 04-12 17:31:36 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 432/3300 [1:33:18<7:21:20,  9.23s/it]2026-04-12 17:31:44,855 | INFO | train_grpo_train | metrics_logged mode=train step=432

                                                      
{'loss': 0.0122, 'grad_norm': 15.05434799194336, 'learning_rate': 8.693939393939394e-06, 'num_tokens': 804841.0, 'completions/mean_length': 36.25, 'completions/min_length': 32.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9286046028137207, 'rewards/meter/std': 0.10607271641492844, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.7789238691329956, 'rewards/repeat_penalty/std': 0.31931358575820923, 'rewards/repeat_floor/mean': 0.9670715928077698, 'rewards/repeat_floor/std': 0.05417557805776596, 'rewards/near_duplicate_penalty/mean': 0.8985941410064697, 'rewards/near_duplicate_penalty/std': 0.13835026323795319, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.8552219271659851, 'rewards/distinct_2/std': 0.2321971356868744, 'rewards/total_composite/mean': 0.33405745029449463, 'rewards/total_composite/std': 0.21716289222240448, 'reward': 0.33405745029449463, 'reward_std': 0.21716290712356567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1804109364748001, 'sampling/sampling_logp_difference/max': 1.2361602783203125, 'sampling/importance_sampling_ratio/min': 0.2904975116252899, 'sampling/importance_sampling_ratio/mean': 1.0170305967330933, 'sampling/importance_sampling_ratio/max': 1.888614535331726, 'entropy': 1.3487338498234749, 'clip_ratio/low_mean': 0.05966893583536148, 'clip_ratio/low_min': 0.05966893583536148, 'clip_ratio/high_mean': 0.10008982848376036, 'clip_ratio/high_max': 0.10008982848376036, 'clip_ratio/region_mean': 0.15975876431912184, 'reward_total_mean': 0.33405745029449463, 'reward_meter_mean': 0.9286046028137207, 'reward_meter_std': 0.10607271641492844, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7789238691329956, 'reward_repeat_penalty_std': 0.31931358575820923, 'reward_repeat_floor_mean': 0.9670715928077698, 'reward_repeat_floor_std': 0.05417557805776596, 'reward_near_duplicate_penalty_mean': 0.8985941410064697, 'reward_near_duplicate_penalty_std': 0.13835026323795319, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.8552219271659851, 'reward_distinct_2_std': 0.2321971356868744, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.33405745029449463, 'reward_total_composite_std': 0.21716289222240448, 'epoch': 0.02}

 13%|█▎        | 432/3300 [1:33:18<7:21:20,  9.23s/it]INFO 04-12 17:31:45 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 433/3300 [1:33:29<7:39:11,  9.61s/it]2026-04-12 17:31:55,347 | INFO | train_grpo_train | metrics_logged mode=train step=433

                                                      
{'loss': 0.0811, 'grad_norm': 21.103116989135742, 'learning_rate': 8.690909090909091e-06, 'num_tokens': 806260.0, 'completions/mean_length': 19.375, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.5606968998908997, 'rewards/meter/std': 0.42898276448249817, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.5866682529449463, 'rewards/repeat_penalty/std': 0.18245680630207062, 'rewards/repeat_floor/mean': 0.9532995820045471, 'rewards/repeat_floor/std': 0.037441376596689224, 'rewards/near_duplicate_penalty/mean': 0.8347302675247192, 'rewards/near_duplicate_penalty/std': 0.2094637155532837, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9232226014137268, 'rewards/distinct_2/std': 0.0840834379196167, 'rewards/total_composite/mean': 0.3905339241027832, 'rewards/total_composite/std': 0.3559979200363159, 'reward': 0.3905339241027832, 'reward_std': 0.3559979200363159, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14828713238239288, 'sampling/sampling_logp_difference/max': 1.1223297119140625, 'sampling/importance_sampling_ratio/min': 0.32552051544189453, 'sampling/importance_sampling_ratio/mean': 1.0559544563293457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2376205250620842, 'clip_ratio/low_mean': 0.10375180514529347, 'clip_ratio/low_min': 0.10375180514529347, 'clip_ratio/high_mean': 0.06840277835726738, 'clip_ratio/high_max': 0.06840277835726738, 'clip_ratio/region_mean': 0.17215458350256085, 'reward_total_mean': 0.3905339241027832, 'reward_meter_mean': 0.5606968998908997, 'reward_meter_std': 0.42898276448249817, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5866682529449463, 'reward_repeat_penalty_std': 0.18245680630207062, 'reward_repeat_floor_mean': 0.9532995820045471, 'reward_repeat_floor_std': 0.037441376596689224, 'reward_near_duplicate_penalty_mean': 0.8347302675247192, 'reward_near_duplicate_penalty_std': 0.2094637155532837, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9232226014137268, 'reward_distinct_2_std': 0.0840834379196167, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.3905339241027832, 'reward_total_composite_std': 0.3559979200363159, 'epoch': 0.02}

 13%|█▎        | 433/3300 [1:33:29<7:39:11,  9.61s/it]INFO 04-12 17:31:55 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 434/3300 [1:33:43<8:47:43, 11.05s/it]2026-04-12 17:32:09,743 | INFO | train_grpo_train | metrics_logged mode=train step=434

                                                      
{'loss': -0.0521, 'grad_norm': 4.7322797775268555, 'learning_rate': 8.687878787878789e-06, 'num_tokens': 807950.0, 'completions/mean_length': 98.25, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.142860412597656, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.4920887351036072, 'rewards/meter/std': 0.44183990359306335, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9341939687728882, 'rewards/lexical_plausibility/std': 0.18612757325172424, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9411939382553101, 'rewards/repeat_penalty/std': 0.08720644563436508, 'rewards/repeat_floor/mean': 0.9944504499435425, 'rewards/repeat_floor/std': 0.006299737375229597, 'rewards/near_duplicate_penalty/mean': 0.9807475805282593, 'rewards/near_duplicate_penalty/std': 0.024986431002616882, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.1768699586391449, 'rewards/total_composite/std': 0.15796123445034027, 'reward': 0.1768699586391449, 'reward_std': 0.15796124935150146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16028976440429688, 'sampling/sampling_logp_difference/max': 1.40220308303833, 'sampling/importance_sampling_ratio/min': 0.2460542768239975, 'sampling/importance_sampling_ratio/mean': 1.0458790063858032, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2747960537672043, 'clip_ratio/low_mean': 0.07279829680919647, 'clip_ratio/low_min': 0.07279829680919647, 'clip_ratio/high_mean': 0.07110042870044708, 'clip_ratio/high_max': 0.07110042870044708, 'clip_ratio/region_mean': 0.14389872550964355, 'reward_total_mean': 0.1768699586391449, 'reward_meter_mean': 0.4920887351036072, 'reward_meter_std': 0.44183990359306335, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9341939687728882, 'reward_lexical_plausibility_std': 0.18612757325172424, 'reward_repeat_penalty_mean': 0.9411939382553101, 'reward_repeat_penalty_std': 0.08720644563436508, 'reward_repeat_floor_mean': 0.9944504499435425, 'reward_repeat_floor_std': 0.006299737375229597, 'reward_near_duplicate_penalty_mean': 0.9807475805282593, 'reward_near_duplicate_penalty_std': 0.024986431002616882, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.1768699586391449, 'reward_total_composite_std': 0.15796123445034027, 'epoch': 0.02}

 13%|█▎        | 434/3300 [1:33:43<8:47:43, 11.05s/it]INFO 04-12 17:32:09 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 435/3300 [1:33:53<8:31:26, 10.71s/it]2026-04-12 17:32:19,670 | INFO | train_grpo_train | metrics_logged mode=train step=435

                                                      
{'loss': 0.1078, 'grad_norm': 11.790294647216797, 'learning_rate': 8.684848484848486e-06, 'num_tokens': 810514.0, 'completions/mean_length': 115.5, 'completions/min_length': 95.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.5, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.4564869999885559, 'rewards/meter/std': 0.3084644675254822, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.023145509883761406, 'rewards/lexical_plausibility/mean': 0.96875, 'rewards/lexical_plausibility/std': 0.0883883461356163, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.47967588901519775, 'rewards/repeat_penalty/std': 0.2529032826423645, 'rewards/repeat_floor/mean': 0.9452492594718933, 'rewards/repeat_floor/std': 0.034242741763591766, 'rewards/near_duplicate_penalty/mean': 0.9025230407714844, 'rewards/near_duplicate_penalty/std': 0.086373470723629, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5541201829910278, 'rewards/distinct_2/std': 0.25107184052467346, 'rewards/total_composite/mean': 0.11287913471460342, 'rewards/total_composite/std': 0.0824497640132904, 'reward': 0.11287913471460342, 'reward_std': 0.08244975656270981, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1585901826620102, 'sampling/sampling_logp_difference/max': 2.0680277347564697, 'sampling/importance_sampling_ratio/min': 0.12643489241600037, 'sampling/importance_sampling_ratio/mean': 1.0194590091705322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8978934735059738, 'clip_ratio/low_mean': 0.07746379356831312, 'clip_ratio/low_min': 0.07746379356831312, 'clip_ratio/high_mean': 0.08359695691615343, 'clip_ratio/high_max': 0.08359695691615343, 'clip_ratio/region_mean': 0.16106075048446655, 'reward_total_mean': 0.11287913471460342, 'reward_meter_mean': 0.4564869999885559, 'reward_meter_std': 0.3084644675254822, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.023145509883761406, 'reward_lexical_plausibility_mean': 0.96875, 'reward_lexical_plausibility_std': 0.0883883461356163, 'reward_repeat_penalty_mean': 0.47967588901519775, 'reward_repeat_penalty_std': 0.2529032826423645, 'reward_repeat_floor_mean': 0.9452492594718933, 'reward_repeat_floor_std': 0.034242741763591766, 'reward_near_duplicate_penalty_mean': 0.9025230407714844, 'reward_near_duplicate_penalty_std': 0.086373470723629, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5541201829910278, 'reward_distinct_2_std': 0.25107184052467346, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.11287913471460342, 'reward_total_composite_std': 0.0824497640132904, 'epoch': 0.02}

 13%|█▎        | 435/3300 [1:33:53<8:31:26, 10.71s/it]INFO 04-12 17:32:19 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 436/3300 [1:34:02<8:02:29, 10.11s/it]2026-04-12 17:32:28,370 | INFO | train_grpo_train | metrics_logged mode=train step=436

                                                      
{'loss': -0.0476, 'grad_norm': 15.734235763549805, 'learning_rate': 8.681818181818182e-06, 'num_tokens': 812143.0, 'completions/mean_length': 37.625, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9163669347763062, 'rewards/meter/std': 0.174087792634964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.1060660257935524, 'rewards/repeat_penalty/mean': 0.902525782585144, 'rewards/repeat_penalty/std': 0.12604406476020813, 'rewards/repeat_floor/mean': 0.9895317554473877, 'rewards/repeat_floor/std': 0.011998934671282768, 'rewards/near_duplicate_penalty/mean': 0.9656435251235962, 'rewards/near_duplicate_penalty/std': 0.031159186735749245, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9617804884910583, 'rewards/distinct_2/std': 0.05967911705374718, 'rewards/total_composite/mean': 0.44398781657218933, 'rewards/total_composite/std': 0.13645069301128387, 'reward': 0.44398781657218933, 'reward_std': 0.13645070791244507, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1768357753753662, 'sampling/sampling_logp_difference/max': 2.1466424465179443, 'sampling/importance_sampling_ratio/min': 0.11687591671943665, 'sampling/importance_sampling_ratio/mean': 1.0240319967269897, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3429600894451141, 'clip_ratio/low_mean': 0.14854550547897816, 'clip_ratio/low_min': 0.14854550547897816, 'clip_ratio/high_mean': 0.019886363297700882, 'clip_ratio/high_max': 0.019886363297700882, 'clip_ratio/region_mean': 0.16843186877667904, 'reward_total_mean': 0.44398781657218933, 'reward_meter_mean': 0.9163669347763062, 'reward_meter_std': 0.174087792634964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.902525782585144, 'reward_repeat_penalty_std': 0.12604406476020813, 'reward_repeat_floor_mean': 0.9895317554473877, 'reward_repeat_floor_std': 0.011998934671282768, 'reward_near_duplicate_penalty_mean': 0.9656435251235962, 'reward_near_duplicate_penalty_std': 0.031159186735749245, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9617804884910583, 'reward_distinct_2_std': 0.05967911705374718, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.1060660257935524, 'reward_total_composite_mean': 0.44398781657218933, 'reward_total_composite_std': 0.13645069301128387, 'epoch': 0.02}

 13%|█▎        | 436/3300 [1:34:02<8:02:29, 10.11s/it]INFO 04-12 17:32:28 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 437/3300 [1:34:10<7:43:05,  9.70s/it]2026-04-12 17:32:37,132 | INFO | train_grpo_train | metrics_logged mode=train step=437

                                                      
{'loss': -0.0727, 'grad_norm': 14.207025527954102, 'learning_rate': 8.67878787878788e-06, 'num_tokens': 813768.0, 'completions/mean_length': 44.125, 'completions/min_length': 20.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9627096056938171, 'rewards/meter/std': 0.07403799146413803, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.18137866258621216, 'rewards/repeat_penalty/mean': 0.9269751310348511, 'rewards/repeat_penalty/std': 0.0906093642115593, 'rewards/repeat_floor/mean': 0.9921700358390808, 'rewards/repeat_floor/std': 0.007709295954555273, 'rewards/near_duplicate_penalty/mean': 0.9649156928062439, 'rewards/near_duplicate_penalty/std': 0.03982192277908325, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.49118906259536743, 'rewards/total_composite/std': 0.18979567289352417, 'reward': 0.49118906259536743, 'reward_std': 0.18979567289352417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19425368309020996, 'sampling/sampling_logp_difference/max': 1.845548152923584, 'sampling/importance_sampling_ratio/min': 0.15793871879577637, 'sampling/importance_sampling_ratio/mean': 1.0171160697937012, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.537724032998085, 'clip_ratio/low_mean': 0.1682034321129322, 'clip_ratio/low_min': 0.1682034321129322, 'clip_ratio/high_mean': 0.04624304175376892, 'clip_ratio/high_max': 0.04624304175376892, 'clip_ratio/region_mean': 0.21444647386670113, 'reward_total_mean': 0.49118906259536743, 'reward_meter_mean': 0.9627096056938171, 'reward_meter_std': 0.07403799146413803, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9269751310348511, 'reward_repeat_penalty_std': 0.0906093642115593, 'reward_repeat_floor_mean': 0.9921700358390808, 'reward_repeat_floor_std': 0.007709295954555273, 'reward_near_duplicate_penalty_mean': 0.9649156928062439, 'reward_near_duplicate_penalty_std': 0.03982192277908325, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.18137866258621216, 'reward_total_composite_mean': 0.49118906259536743, 'reward_total_composite_std': 0.18979567289352417, 'epoch': 0.02}

 13%|█▎        | 437/3300 [1:34:10<7:43:05,  9.70s/it]INFO 04-12 17:32:37 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 438/3300 [1:34:24<8:33:34, 10.77s/it]2026-04-12 17:32:50,377 | INFO | train_grpo_train | metrics_logged mode=train step=438

                                                      
{'loss': -0.0364, 'grad_norm': 7.291502952575684, 'learning_rate': 8.675757575757576e-06, 'num_tokens': 815463.0, 'completions/mean_length': 105.875, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.85714340209961, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.4315311908721924, 'rewards/meter/std': 0.36646798253059387, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.3563483655452728, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9500000476837158, 'rewards/count_floor/std': 0.10690450668334961, 'rewards/lexical_plausibility/mean': 0.9189607501029968, 'rewards/lexical_plausibility/std': 0.2292136400938034, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.9554541110992432, 'rewards/repeat_penalty/std': 0.050407905131578445, 'rewards/repeat_floor/mean': 0.9946420788764954, 'rewards/repeat_floor/std': 0.005840022582560778, 'rewards/near_duplicate_penalty/mean': 0.9790205955505371, 'rewards/near_duplicate_penalty/std': 0.020779188722372055, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9754332304000854, 'rewards/distinct_2/std': 0.03412005305290222, 'rewards/total_composite/mean': 0.2373935580253601, 'rewards/total_composite/std': 0.29499682784080505, 'reward': 0.2373935580253601, 'reward_std': 0.29499682784080505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23944535851478577, 'sampling/sampling_logp_difference/max': 1.980813980102539, 'sampling/importance_sampling_ratio/min': 0.1379568874835968, 'sampling/importance_sampling_ratio/mean': 1.0251054763793945, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8470597416162491, 'clip_ratio/low_mean': 0.10668371617794037, 'clip_ratio/low_min': 0.10668371617794037, 'clip_ratio/high_mean': 0.048469388857483864, 'clip_ratio/high_max': 0.048469388857483864, 'clip_ratio/region_mean': 0.15515310503542423, 'reward_total_mean': 0.2373935580253601, 'reward_meter_mean': 0.4315311908721924, 'reward_meter_std': 0.36646798253059387, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.3563483655452728, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9500000476837158, 'reward_count_floor_std': 0.10690450668334961, 'reward_lexical_plausibility_mean': 0.9189607501029968, 'reward_lexical_plausibility_std': 0.2292136400938034, 'reward_repeat_penalty_mean': 0.9554541110992432, 'reward_repeat_penalty_std': 0.050407905131578445, 'reward_repeat_floor_mean': 0.9946420788764954, 'reward_repeat_floor_std': 0.005840022582560778, 'reward_near_duplicate_penalty_mean': 0.9790205955505371, 'reward_near_duplicate_penalty_std': 0.020779188722372055, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9754332304000854, 'reward_distinct_2_std': 0.03412005305290222, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.2373935580253601, 'reward_total_composite_std': 0.29499682784080505, 'epoch': 0.02}

 13%|█▎        | 438/3300 [1:34:24<8:33:34, 10.77s/it]INFO 04-12 17:32:50 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 439/3300 [1:34:32<7:58:31, 10.04s/it]2026-04-12 17:32:58,709 | INFO | train_grpo_train | metrics_logged mode=train step=439

                                                      
{'loss': -0.027, 'grad_norm': 14.869061470031738, 'learning_rate': 8.672727272727273e-06, 'num_tokens': 817137.0, 'completions/mean_length': 39.25, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8195381164550781, 'rewards/meter/std': 0.3328098952770233, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.8967796564102173, 'rewards/repeat_penalty/std': 0.1638745367527008, 'rewards/repeat_floor/mean': 0.9877615571022034, 'rewards/repeat_floor/std': 0.01785091497004032, 'rewards/near_duplicate_penalty/mean': 0.9547200202941895, 'rewards/near_duplicate_penalty/std': 0.05203406885266304, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9603174924850464, 'rewards/distinct_2/std': 0.07376706600189209, 'rewards/total_composite/mean': 0.376526415348053, 'rewards/total_composite/std': 0.2509756088256836, 'reward': 0.376526415348053, 'reward_std': 0.2509755790233612, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17136229574680328, 'sampling/sampling_logp_difference/max': 1.500025987625122, 'sampling/importance_sampling_ratio/min': 0.22312436997890472, 'sampling/importance_sampling_ratio/mean': 1.0411264896392822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.327915146946907, 'clip_ratio/low_mean': 0.09228302352130413, 'clip_ratio/low_min': 0.09228302352130413, 'clip_ratio/high_mean': 0.03553196042776108, 'clip_ratio/high_max': 0.03553196042776108, 'clip_ratio/region_mean': 0.1278149839490652, 'reward_total_mean': 0.376526415348053, 'reward_meter_mean': 0.8195381164550781, 'reward_meter_std': 0.3328098952770233, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8967796564102173, 'reward_repeat_penalty_std': 0.1638745367527008, 'reward_repeat_floor_mean': 0.9877615571022034, 'reward_repeat_floor_std': 0.01785091497004032, 'reward_near_duplicate_penalty_mean': 0.9547200202941895, 'reward_near_duplicate_penalty_std': 0.05203406885266304, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9603174924850464, 'reward_distinct_2_std': 0.07376706600189209, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.376526415348053, 'reward_total_composite_std': 0.2509756088256836, 'epoch': 0.02}

 13%|█▎        | 439/3300 [1:34:32<7:58:31, 10.04s/it]INFO 04-12 17:32:58 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 440/3300 [1:34:40<7:33:53,  9.52s/it]2026-04-12 17:33:07,030 | INFO | train_grpo_train | metrics_logged mode=train step=440

                                                      
{'loss': 0.0653, 'grad_norm': 16.776262283325195, 'learning_rate': 8.66969696969697e-06, 'num_tokens': 818602.0, 'completions/mean_length': 36.125, 'completions/min_length': 19.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7714211344718933, 'rewards/meter/std': 0.3801732361316681, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45374998450279236, 'rewards/judge_quality/std': 0.2404720038175583, 'rewards/repeat_penalty/mean': 0.9452999234199524, 'rewards/repeat_penalty/std': 0.08995723724365234, 'rewards/repeat_floor/mean': 0.9950763583183289, 'rewards/repeat_floor/std': 0.006854683626443148, 'rewards/near_duplicate_penalty/mean': 0.9849203824996948, 'rewards/near_duplicate_penalty/std': 0.029514934867620468, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.3650018572807312, 'rewards/total_composite/std': 0.279037743806839, 'reward': 0.3650018572807312, 'reward_std': 0.279037743806839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18914948403835297, 'sampling/sampling_logp_difference/max': 1.8713899850845337, 'sampling/importance_sampling_ratio/min': 0.1539095789194107, 'sampling/importance_sampling_ratio/mean': 1.0234311819076538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6712673604488373, 'clip_ratio/low_mean': 0.06190958060324192, 'clip_ratio/low_min': 0.06190958060324192, 'clip_ratio/high_mean': 0.1166908759623766, 'clip_ratio/high_max': 0.1166908759623766, 'clip_ratio/region_mean': 0.17860045656561852, 'reward_total_mean': 0.3650018572807312, 'reward_meter_mean': 0.7714211344718933, 'reward_meter_std': 0.3801732361316681, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9452999234199524, 'reward_repeat_penalty_std': 0.08995723724365234, 'reward_repeat_floor_mean': 0.9950763583183289, 'reward_repeat_floor_std': 0.006854683626443148, 'reward_near_duplicate_penalty_mean': 0.9849203824996948, 'reward_near_duplicate_penalty_std': 0.029514934867620468, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.45374998450279236, 'reward_judge_quality_std': 0.2404720038175583, 'reward_total_composite_mean': 0.3650018572807312, 'reward_total_composite_std': 0.279037743806839, 'epoch': 0.02}

 13%|█▎        | 440/3300 [1:34:40<7:33:53,  9.52s/it]INFO 04-12 17:33:07 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 441/3300 [1:34:49<7:17:25,  9.18s/it]2026-04-12 17:33:15,416 | INFO | train_grpo_train | metrics_logged mode=train step=441

                                                      
{'loss': 0.0323, 'grad_norm': 12.250515937805176, 'learning_rate': 8.666666666666668e-06, 'num_tokens': 820174.0, 'completions/mean_length': 40.5, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7683224678039551, 'rewards/meter/std': 0.40084394812583923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.680577278137207, 'rewards/repeat_penalty/std': 0.3836446702480316, 'rewards/repeat_floor/mean': 0.9568132162094116, 'rewards/repeat_floor/std': 0.06405427306890488, 'rewards/near_duplicate_penalty/mean': 0.890741229057312, 'rewards/near_duplicate_penalty/std': 0.18196484446525574, 'rewards/opening_diversity/mean': 0.828125, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.8168286681175232, 'rewards/distinct_2/std': 0.28185638785362244, 'rewards/total_composite/mean': 0.3095269799232483, 'rewards/total_composite/std': 0.2818174958229065, 'reward': 0.3095269799232483, 'reward_std': 0.2818174958229065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1876748502254486, 'sampling/sampling_logp_difference/max': 2.106761932373047, 'sampling/importance_sampling_ratio/min': 0.1216311827301979, 'sampling/importance_sampling_ratio/mean': 1.0105547904968262, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1488417834043503, 'clip_ratio/low_mean': 0.08663870207965374, 'clip_ratio/low_min': 0.08663870207965374, 'clip_ratio/high_mean': 0.0862985560670495, 'clip_ratio/high_max': 0.0862985560670495, 'clip_ratio/region_mean': 0.17293725814670324, 'reward_total_mean': 0.3095269799232483, 'reward_meter_mean': 0.7683224678039551, 'reward_meter_std': 0.40084394812583923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.680577278137207, 'reward_repeat_penalty_std': 0.3836446702480316, 'reward_repeat_floor_mean': 0.9568132162094116, 'reward_repeat_floor_std': 0.06405427306890488, 'reward_near_duplicate_penalty_mean': 0.890741229057312, 'reward_near_duplicate_penalty_std': 0.18196484446525574, 'reward_opening_diversity_mean': 0.828125, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.8168286681175232, 'reward_distinct_2_std': 0.28185638785362244, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.3095269799232483, 'reward_total_composite_std': 0.2818174958229065, 'epoch': 0.02}

 13%|█▎        | 441/3300 [1:34:49<7:17:25,  9.18s/it]INFO 04-12 17:33:15 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 442/3300 [1:34:57<7:09:57,  9.03s/it]2026-04-12 17:33:24,085 | INFO | train_grpo_train | metrics_logged mode=train step=442

                                                      
{'loss': 0.0547, 'grad_norm': 20.725318908691406, 'learning_rate': 8.663636363636363e-06, 'num_tokens': 821613.0, 'completions/mean_length': 33.875, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.4547012150287628, 'rewards/meter/std': 0.40436694025993347, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.268537700176239, 'rewards/repeat_penalty/mean': 0.8193002939224243, 'rewards/repeat_penalty/std': 0.27543652057647705, 'rewards/repeat_floor/mean': 0.9810307025909424, 'rewards/repeat_floor/std': 0.028231441974639893, 'rewards/near_duplicate_penalty/mean': 0.953036904335022, 'rewards/near_duplicate_penalty/std': 0.06711763888597488, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9091683030128479, 'rewards/distinct_2/std': 0.13809095323085785, 'rewards/total_composite/mean': 0.15017744898796082, 'rewards/total_composite/std': 0.1516982614994049, 'reward': 0.15017744898796082, 'reward_std': 0.1516982614994049, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2031320333480835, 'sampling/sampling_logp_difference/max': 2.945540428161621, 'sampling/importance_sampling_ratio/min': 0.052573639899492264, 'sampling/importance_sampling_ratio/mean': 1.0040764808654785, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2909222468733788, 'clip_ratio/low_mean': 0.09805652685463428, 'clip_ratio/low_min': 0.09805652685463428, 'clip_ratio/high_mean': 0.05971479695290327, 'clip_ratio/high_max': 0.05971479695290327, 'clip_ratio/region_mean': 0.15777132380753756, 'reward_total_mean': 0.15017744898796082, 'reward_meter_mean': 0.4547012150287628, 'reward_meter_std': 0.40436694025993347, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8193002939224243, 'reward_repeat_penalty_std': 0.27543652057647705, 'reward_repeat_floor_mean': 0.9810307025909424, 'reward_repeat_floor_std': 0.028231441974639893, 'reward_near_duplicate_penalty_mean': 0.953036904335022, 'reward_near_duplicate_penalty_std': 0.06711763888597488, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9091683030128479, 'reward_distinct_2_std': 0.13809095323085785, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.268537700176239, 'reward_total_composite_mean': 0.15017744898796082, 'reward_total_composite_std': 0.1516982614994049, 'epoch': 0.02}

 13%|█▎        | 442/3300 [1:34:57<7:09:57,  9.03s/it]INFO 04-12 17:33:24 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 443/3300 [1:35:07<7:17:08,  9.18s/it]2026-04-12 17:33:33,622 | INFO | train_grpo_train | metrics_logged mode=train step=443

                                                      
{'loss': 0.0462, 'grad_norm': 12.641680717468262, 'learning_rate': 8.660606060606062e-06, 'num_tokens': 823467.0, 'completions/mean_length': 53.75, 'completions/min_length': 38.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6973755955696106, 'rewards/meter/std': 0.4102723300457001, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.8206741809844971, 'rewards/repeat_penalty/std': 0.15012654662132263, 'rewards/repeat_floor/mean': 0.9779125452041626, 'rewards/repeat_floor/std': 0.017838358879089355, 'rewards/near_duplicate_penalty/mean': 0.9214118123054504, 'rewards/near_duplicate_penalty/std': 0.05443250760436058, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8855636119842529, 'rewards/distinct_2/std': 0.11458331346511841, 'rewards/total_composite/mean': 0.28731346130371094, 'rewards/total_composite/std': 0.27685046195983887, 'reward': 0.28731346130371094, 'reward_std': 0.27685049176216125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1657501608133316, 'sampling/sampling_logp_difference/max': 1.9893255233764648, 'sampling/importance_sampling_ratio/min': 0.13678765296936035, 'sampling/importance_sampling_ratio/mean': 1.0055123567581177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9584505185484886, 'clip_ratio/low_mean': 0.05728705367073417, 'clip_ratio/low_min': 0.05728705367073417, 'clip_ratio/high_mean': 0.09869378432631493, 'clip_ratio/high_max': 0.09869378432631493, 'clip_ratio/region_mean': 0.1559808379970491, 'reward_total_mean': 0.28731346130371094, 'reward_meter_mean': 0.6973755955696106, 'reward_meter_std': 0.4102723300457001, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8206741809844971, 'reward_repeat_penalty_std': 0.15012654662132263, 'reward_repeat_floor_mean': 0.9779125452041626, 'reward_repeat_floor_std': 0.017838358879089355, 'reward_near_duplicate_penalty_mean': 0.9214118123054504, 'reward_near_duplicate_penalty_std': 0.05443250760436058, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8855636119842529, 'reward_distinct_2_std': 0.11458331346511841, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.28731346130371094, 'reward_total_composite_std': 0.27685046195983887, 'epoch': 0.02}

 13%|█▎        | 443/3300 [1:35:07<7:17:08,  9.18s/it]INFO 04-12 17:33:33 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 444/3300 [1:35:21<8:27:59, 10.67s/it]2026-04-12 17:33:47,775 | INFO | train_grpo_train | metrics_logged mode=train step=444

                                                      
{'loss': -0.0532, 'grad_norm': 2.3879916667938232, 'learning_rate': 8.657575757575758e-06, 'num_tokens': 825054.0, 'completions/mean_length': 218.375, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 42.20000076293945, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.4307488203048706, 'rewards/meter/std': 0.45420384407043457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8206913471221924, 'rewards/lexical_plausibility/std': 0.23280160129070282, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.29235193133354187, 'rewards/repeat_penalty/mean': 0.9374526739120483, 'rewards/repeat_penalty/std': 0.08515334874391556, 'rewards/repeat_floor/mean': 0.9917943477630615, 'rewards/repeat_floor/std': 0.010660682804882526, 'rewards/near_duplicate_penalty/mean': 0.9609840512275696, 'rewards/near_duplicate_penalty/std': 0.044291090220212936, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9738528728485107, 'rewards/distinct_2/std': 0.050033703446388245, 'rewards/total_composite/mean': 0.13904240727424622, 'rewards/total_composite/std': 0.1691950410604477, 'reward': 0.13904240727424622, 'reward_std': 0.1691950410604477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20445603132247925, 'sampling/sampling_logp_difference/max': 1.2904787063598633, 'sampling/importance_sampling_ratio/min': 0.3813531994819641, 'sampling/importance_sampling_ratio/mean': 1.0190016031265259, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2257293909788132, 'clip_ratio/low_mean': 0.036382850259542465, 'clip_ratio/low_min': 0.036382850259542465, 'clip_ratio/high_mean': 0.0665881372988224, 'clip_ratio/high_max': 0.0665881372988224, 'clip_ratio/region_mean': 0.10297098755836487, 'reward_total_mean': 0.13904240727424622, 'reward_meter_mean': 0.4307488203048706, 'reward_meter_std': 0.45420384407043457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8206913471221924, 'reward_lexical_plausibility_std': 0.23280160129070282, 'reward_repeat_penalty_mean': 0.9374526739120483, 'reward_repeat_penalty_std': 0.08515334874391556, 'reward_repeat_floor_mean': 0.9917943477630615, 'reward_repeat_floor_std': 0.010660682804882526, 'reward_near_duplicate_penalty_mean': 0.9609840512275696, 'reward_near_duplicate_penalty_std': 0.044291090220212936, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9738528728485107, 'reward_distinct_2_std': 0.050033703446388245, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.29235193133354187, 'reward_total_composite_mean': 0.13904240727424622, 'reward_total_composite_std': 0.1691950410604477, 'epoch': 0.02}

 13%|█▎        | 444/3300 [1:35:21<8:27:59, 10.67s/it]INFO 04-12 17:33:48 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 445/3300 [1:35:30<7:57:52, 10.04s/it]2026-04-12 17:33:56,350 | INFO | train_grpo_train | metrics_logged mode=train step=445

                                                      
{'loss': 0.0521, 'grad_norm': 19.418930053710938, 'learning_rate': 8.654545454545455e-06, 'num_tokens': 826526.0, 'completions/mean_length': 35.0, 'completions/min_length': 32.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8228330612182617, 'rewards/meter/std': 0.3061090409755707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3184336721897125, 'rewards/repeat_penalty/mean': 0.9564417600631714, 'rewards/repeat_penalty/std': 0.08751532435417175, 'rewards/repeat_floor/mean': 0.9953893423080444, 'rewards/repeat_floor/std': 0.007846657186746597, 'rewards/near_duplicate_penalty/mean': 0.9884930849075317, 'rewards/near_duplicate_penalty/std': 0.016842078417539597, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9679487347602844, 'rewards/distinct_2/std': 0.09065470844507217, 'rewards/total_composite/mean': 0.5848497152328491, 'rewards/total_composite/std': 0.3616703748703003, 'reward': 0.5848497152328491, 'reward_std': 0.3616703450679779, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18440987169742584, 'sampling/sampling_logp_difference/max': 1.8155077695846558, 'sampling/importance_sampling_ratio/min': 0.16275525093078613, 'sampling/importance_sampling_ratio/mean': 1.028830885887146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.263876512646675, 'clip_ratio/low_mean': 0.11446377448737621, 'clip_ratio/low_min': 0.11446377448737621, 'clip_ratio/high_mean': 0.09691220428794622, 'clip_ratio/high_max': 0.09691220428794622, 'clip_ratio/region_mean': 0.21137597877532244, 'reward_total_mean': 0.5848497152328491, 'reward_meter_mean': 0.8228330612182617, 'reward_meter_std': 0.3061090409755707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9564417600631714, 'reward_repeat_penalty_std': 0.08751532435417175, 'reward_repeat_floor_mean': 0.9953893423080444, 'reward_repeat_floor_std': 0.007846657186746597, 'reward_near_duplicate_penalty_mean': 0.9884930849075317, 'reward_near_duplicate_penalty_std': 0.016842078417539597, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9679487347602844, 'reward_distinct_2_std': 0.09065470844507217, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3184336721897125, 'reward_total_composite_mean': 0.5848497152328491, 'reward_total_composite_std': 0.3616703748703003, 'epoch': 0.02}

 13%|█▎        | 445/3300 [1:35:30<7:57:52, 10.04s/it]INFO 04-12 17:33:56 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 446/3300 [1:35:43<8:48:33, 11.11s/it]2026-04-12 17:34:09,953 | INFO | train_grpo_train | metrics_logged mode=train step=446

                                                      
{'loss': -0.0795, 'grad_norm': 4.500837326049805, 'learning_rate': 8.651515151515152e-06, 'num_tokens': 828024.0, 'completions/mean_length': 101.25, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.57143020629883, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8515878319740295, 'rewards/meter/std': 0.3188558518886566, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9403458833694458, 'rewards/lexical_plausibility/std': 0.1687273234128952, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3160809278488159, 'rewards/repeat_penalty/mean': 0.6370641589164734, 'rewards/repeat_penalty/std': 0.31917449831962585, 'rewards/repeat_floor/mean': 0.951697051525116, 'rewards/repeat_floor/std': 0.05713529884815216, 'rewards/near_duplicate_penalty/mean': 0.8621770143508911, 'rewards/near_duplicate_penalty/std': 0.15992195904254913, 'rewards/opening_diversity/mean': 0.796875, 'rewards/opening_diversity/std': 0.2106272578239441, 'rewards/distinct_2/mean': 0.8284223079681396, 'rewards/distinct_2/std': 0.24711714684963226, 'rewards/total_composite/mean': 0.4267417788505554, 'rewards/total_composite/std': 0.3043561279773712, 'reward': 0.4267417788505554, 'reward_std': 0.3043561279773712, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13605427742004395, 'sampling/sampling_logp_difference/max': 1.4208288192749023, 'sampling/importance_sampling_ratio/min': 0.24151377379894257, 'sampling/importance_sampling_ratio/mean': 1.022839903831482, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8579638451337814, 'clip_ratio/low_mean': 0.05235937889665365, 'clip_ratio/low_min': 0.05235937889665365, 'clip_ratio/high_mean': 0.053034319542348385, 'clip_ratio/high_max': 0.053034319542348385, 'clip_ratio/region_mean': 0.10539369843900204, 'reward_total_mean': 0.4267417788505554, 'reward_meter_mean': 0.8515878319740295, 'reward_meter_std': 0.3188558518886566, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9403458833694458, 'reward_lexical_plausibility_std': 0.1687273234128952, 'reward_repeat_penalty_mean': 0.6370641589164734, 'reward_repeat_penalty_std': 0.31917449831962585, 'reward_repeat_floor_mean': 0.951697051525116, 'reward_repeat_floor_std': 0.05713529884815216, 'reward_near_duplicate_penalty_mean': 0.8621770143508911, 'reward_near_duplicate_penalty_std': 0.15992195904254913, 'reward_opening_diversity_mean': 0.796875, 'reward_opening_diversity_std': 0.2106272578239441, 'reward_distinct_2_mean': 0.8284223079681396, 'reward_distinct_2_std': 0.24711714684963226, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3160809278488159, 'reward_total_composite_mean': 0.4267417788505554, 'reward_total_composite_std': 0.3043561279773712, 'epoch': 0.02}

 14%|█▎        | 446/3300 [1:35:43<8:48:33, 11.11s/it]INFO 04-12 17:34:10 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 447/3300 [1:35:53<8:25:46, 10.64s/it]2026-04-12 17:34:19,482 | INFO | train_grpo_train | metrics_logged mode=train step=447

                                                      
{'loss': 0.0069, 'grad_norm': 11.352745056152344, 'learning_rate': 8.64848484848485e-06, 'num_tokens': 829983.0, 'completions/mean_length': 74.875, 'completions/min_length': 68.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.7570005655288696, 'rewards/meter/std': 0.37158462405204773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.892204761505127, 'rewards/repeat_penalty/std': 0.15515781939029694, 'rewards/repeat_floor/mean': 0.9886047840118408, 'rewards/repeat_floor/std': 0.013831005431711674, 'rewards/near_duplicate_penalty/mean': 0.9616425037384033, 'rewards/near_duplicate_penalty/std': 0.03131810575723648, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.942523717880249, 'rewards/distinct_2/std': 0.09394238889217377, 'rewards/total_composite/mean': 0.3357177972793579, 'rewards/total_composite/std': 0.19216515123844147, 'reward': 0.3357177972793579, 'reward_std': 0.19216513633728027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17933063209056854, 'sampling/sampling_logp_difference/max': 1.6456398963928223, 'sampling/importance_sampling_ratio/min': 0.19288909435272217, 'sampling/importance_sampling_ratio/mean': 1.0180919170379639, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.772596687078476, 'clip_ratio/low_mean': 0.06287023331969976, 'clip_ratio/low_min': 0.06287023331969976, 'clip_ratio/high_mean': 0.11908232234418392, 'clip_ratio/high_max': 0.11908232234418392, 'clip_ratio/region_mean': 0.18195255566388369, 'reward_total_mean': 0.3357177972793579, 'reward_meter_mean': 0.7570005655288696, 'reward_meter_std': 0.37158462405204773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.892204761505127, 'reward_repeat_penalty_std': 0.15515781939029694, 'reward_repeat_floor_mean': 0.9886047840118408, 'reward_repeat_floor_std': 0.013831005431711674, 'reward_near_duplicate_penalty_mean': 0.9616425037384033, 'reward_near_duplicate_penalty_std': 0.03131810575723648, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.942523717880249, 'reward_distinct_2_std': 0.09394238889217377, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.3357177972793579, 'reward_total_composite_std': 0.19216515123844147, 'epoch': 0.02}

 14%|█▎        | 447/3300 [1:35:53<8:25:46, 10.64s/it]INFO 04-12 17:34:19 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 448/3300 [1:36:01<7:50:30,  9.90s/it]2026-04-12 17:34:27,660 | INFO | train_grpo_train | metrics_logged mode=train step=448

                                                      
{'loss': 0.0722, 'grad_norm': 16.51186752319336, 'learning_rate': 8.645454545454545e-06, 'num_tokens': 831707.0, 'completions/mean_length': 39.5, 'completions/min_length': 30.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.07772711664438248, 'rewards/meter/std': 0.058112356811761856, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.8866095542907715, 'rewards/repeat_penalty/std': 0.13586875796318054, 'rewards/repeat_floor/mean': 0.9853266477584839, 'rewards/repeat_floor/std': 0.016353275626897812, 'rewards/near_duplicate_penalty/mean': 0.9361306428909302, 'rewards/near_duplicate_penalty/std': 0.056301020085811615, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9434118270874023, 'rewards/distinct_2/std': 0.10515422374010086, 'rewards/total_composite/mean': 0.051152303814888, 'rewards/total_composite/std': 0.05245295166969299, 'reward': 0.051152303814888, 'reward_std': 0.052452944219112396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20883063971996307, 'sampling/sampling_logp_difference/max': 4.22971248626709, 'sampling/importance_sampling_ratio/min': 0.014556575566530228, 'sampling/importance_sampling_ratio/mean': 0.99327552318573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9105018004775047, 'clip_ratio/low_mean': 0.07945175468921661, 'clip_ratio/low_min': 0.07945175468921661, 'clip_ratio/high_mean': 0.09045888856053352, 'clip_ratio/high_max': 0.09045888856053352, 'clip_ratio/region_mean': 0.16991064324975014, 'reward_total_mean': 0.051152303814888, 'reward_meter_mean': 0.07772711664438248, 'reward_meter_std': 0.058112356811761856, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8866095542907715, 'reward_repeat_penalty_std': 0.13586875796318054, 'reward_repeat_floor_mean': 0.9853266477584839, 'reward_repeat_floor_std': 0.016353275626897812, 'reward_near_duplicate_penalty_mean': 0.9361306428909302, 'reward_near_duplicate_penalty_std': 0.056301020085811615, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9434118270874023, 'reward_distinct_2_std': 0.10515422374010086, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.051152303814888, 'reward_total_composite_std': 0.05245295166969299, 'epoch': 0.02}

 14%|█▎        | 448/3300 [1:36:01<7:50:30,  9.90s/it]INFO 04-12 17:34:27 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 449/3300 [1:36:15<8:50:52, 11.17s/it]2026-04-12 17:34:41,804 | INFO | train_grpo_train | metrics_logged mode=train step=449

                                                      
{'loss': -0.272, 'grad_norm': 1.376067876815796, 'learning_rate': 8.642424242424242e-06, 'num_tokens': 834948.0, 'completions/mean_length': 270.125, 'completions/min_length': 188.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 235.57144165039062, 'completions/min_terminated_length': 188.0, 'completions/max_terminated_length': 279.0, 'rewards/meter/mean': 0.9818702936172485, 'rewards/meter/std': 0.03067575767636299, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.01781740039587021, 'rewards/lexical_plausibility/mean': 0.970257043838501, 'rewards/lexical_plausibility/std': 0.08412572741508484, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.05028318986296654, 'rewards/repeat_penalty/std': 0.11279795318841934, 'rewards/repeat_floor/mean': 0.8261778354644775, 'rewards/repeat_floor/std': 0.05766754597425461, 'rewards/near_duplicate_penalty/mean': 0.557167649269104, 'rewards/near_duplicate_penalty/std': 0.1810387521982193, 'rewards/opening_diversity/mean': 0.5578733682632446, 'rewards/opening_diversity/std': 0.2818715572357178, 'rewards/distinct_2/mean': 0.1014476865530014, 'rewards/distinct_2/std': 0.22239363193511963, 'rewards/total_composite/mean': 0.10441425442695618, 'rewards/total_composite/std': 0.024660808965563774, 'reward': 0.10441425442695618, 'reward_std': 0.024660807102918625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07688037306070328, 'sampling/sampling_logp_difference/max': 2.050701141357422, 'sampling/importance_sampling_ratio/min': 0.1286446750164032, 'sampling/importance_sampling_ratio/mean': 1.0193954706192017, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5089322850108147, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.04842671914957464, 'clip_ratio/high_max': 0.04842671914957464, 'clip_ratio/region_mean': 0.04842671914957464, 'reward_total_mean': 0.10441425442695618, 'reward_meter_mean': 0.9818702936172485, 'reward_meter_std': 0.03067575767636299, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.01781740039587021, 'reward_lexical_plausibility_mean': 0.970257043838501, 'reward_lexical_plausibility_std': 0.08412572741508484, 'reward_repeat_penalty_mean': 0.05028318986296654, 'reward_repeat_penalty_std': 0.11279795318841934, 'reward_repeat_floor_mean': 0.8261778354644775, 'reward_repeat_floor_std': 0.05766754597425461, 'reward_near_duplicate_penalty_mean': 0.557167649269104, 'reward_near_duplicate_penalty_std': 0.1810387521982193, 'reward_opening_diversity_mean': 0.5578733682632446, 'reward_opening_diversity_std': 0.2818715572357178, 'reward_distinct_2_mean': 0.1014476865530014, 'reward_distinct_2_std': 0.22239363193511963, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.10441425442695618, 'reward_total_composite_std': 0.024660808965563774, 'epoch': 0.02}

 14%|█▎        | 449/3300 [1:36:15<8:50:52, 11.17s/it]INFO 04-12 17:34:42 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 450/3300 [1:36:29<9:31:33, 12.03s/it]2026-04-12 17:34:55,845 | INFO | train_grpo_train | metrics_logged mode=train step=450

                                                      
{'loss': -0.0597, 'grad_norm': 2.5712053775787354, 'learning_rate': 8.63939393939394e-06, 'num_tokens': 836594.0, 'completions/mean_length': 156.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 38.333335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6731521487236023, 'rewards/meter/std': 0.29471877217292786, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8807671070098877, 'rewards/lexical_plausibility/std': 0.22261570394039154, 'rewards/judge_quality/mean': 0.4050000011920929, 'rewards/judge_quality/std': 0.3496120274066925, 'rewards/repeat_penalty/mean': 0.9917501211166382, 'rewards/repeat_penalty/std': 0.011602719314396381, 'rewards/repeat_floor/mean': 0.9987624883651733, 'rewards/repeat_floor/std': 0.001740400679409504, 'rewards/near_duplicate_penalty/mean': 0.9917501211166382, 'rewards/near_duplicate_penalty/std': 0.011602719314396381, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30958878993988037, 'rewards/total_composite/std': 0.3038085699081421, 'reward': 0.30958878993988037, 'reward_std': 0.3038085699081421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1802184134721756, 'sampling/sampling_logp_difference/max': 1.451115608215332, 'sampling/importance_sampling_ratio/min': 0.23430874943733215, 'sampling/importance_sampling_ratio/mean': 1.0301158428192139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9844834133982658, 'clip_ratio/low_mean': 0.031862745992839336, 'clip_ratio/low_min': 0.031862745992839336, 'clip_ratio/high_mean': 0.10054155811667442, 'clip_ratio/high_max': 0.10054155811667442, 'clip_ratio/region_mean': 0.13240430410951376, 'reward_total_mean': 0.30958878993988037, 'reward_meter_mean': 0.6731521487236023, 'reward_meter_std': 0.29471877217292786, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8807671070098877, 'reward_lexical_plausibility_std': 0.22261570394039154, 'reward_repeat_penalty_mean': 0.9917501211166382, 'reward_repeat_penalty_std': 0.011602719314396381, 'reward_repeat_floor_mean': 0.9987624883651733, 'reward_repeat_floor_std': 0.001740400679409504, 'reward_near_duplicate_penalty_mean': 0.9917501211166382, 'reward_near_duplicate_penalty_std': 0.011602719314396381, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4050000011920929, 'reward_judge_quality_std': 0.3496120274066925, 'reward_total_composite_mean': 0.30958878993988037, 'reward_total_composite_std': 0.3038085699081421, 'epoch': 0.02}

 14%|█▎        | 450/3300 [1:36:29<9:31:33, 12.03s/it]INFO 04-12 17:34:56 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:12<01:09,  6.28s/it]

 23%|██▎       | 3/13 [00:23<01:24,  8.41s/it]

 31%|███       | 4/13 [00:36<01:27,  9.78s/it]

 38%|███▊      | 5/13 [00:46<01:19,  9.90s/it]

 46%|████▌     | 6/13 [00:55<01:09,  9.87s/it]

 54%|█████▍    | 7/13 [01:07<01:02, 10.44s/it]

 62%|██████▏   | 8/13 [01:18<00:52, 10.59s/it]

 69%|██████▉   | 9/13 [01:30<00:44, 11.04s/it]

 77%|███████▋  | 10/13 [01:39<00:31, 10.34s/it]

 85%|████████▍ | 11/13 [01:51<00:21, 10.76s/it]

 92%|█████████▏| 12/13 [02:02<00:10, 10.90s/it]

100%|██████████| 13/13 [02:14<00:00, 11.21s/it]2026-04-12 17:37:20,096 | INFO | train_grpo_train | metrics_logged mode=eval step=450

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 144.2077, 'eval_samples_per_second': 0.721, 'eval_steps_per_second': 0.09, 'eval_num_tokens': 836594.0, 'eval_completions/mean_length': 197.1153846153846, 'eval_completions/min_length': 38.07692307692308, 'eval_completions/max_length': 460.9230769230769, 'eval_completions/clipped_ratio': 0.10576923076923077, 'eval_completions/mean_terminated_length': 157.49954575758713, 'eval_completions/min_terminated_length': 38.07692307692308, 'eval_completions/max_terminated_length': 350.0, 'eval_rewards/meter/mean': 0.6230765947928796, 'eval_rewards/meter/std': 0.39679821867209214, 'eval_rewards/count_adherence/mean': 0.8260979927503146, 'eval_rewards/count_adherence/std': 0.1801939681172371, 'eval_rewards/arabic_clean/mean': 0.8653846153846154, 'eval_rewards/arabic_clean/std': 0.26383458421780515, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9951923076923077, 'eval_rewards/arabic_floor/std': 0.013598207670908708, 'eval_rewards/count_floor/mean': 0.9478294115800124, 'eval_rewards/count_floor/std': 0.05405819014861034, 'eval_rewards/lexical_plausibility/mean': 0.9509462805894705, 'eval_rewards/lexical_plausibility/std': 0.09938178727260003, 'eval_rewards/judge_quality/mean': 0.3041346153387657, 'eval_rewards/judge_quality/std': 0.2317682011769368, 'eval_rewards/repeat_penalty/mean': 0.4748371495650365, 'eval_rewards/repeat_penalty/std': 0.4065972406130571, 'eval_rewards/repeat_floor/mean': 0.9208995837431687, 'eval_rewards/repeat_floor/std': 0.07328143744514538, 'eval_rewards/near_duplicate_penalty/mean': 0.8090284558442923, 'eval_rewards/near_duplicate_penalty/std': 0.19457622445546663, 'eval_rewards/opening_diversity/mean': 0.844254732131958, 'eval_rewards/opening_diversity/std': 0.22699065391833967, 'eval_rewards/distinct_2/mean': 0.543222583257235, 'eval_rewards/distinct_2/std': 0.40615633817819446, 'eval_rewards/total_composite/mean': 0.17142489953682974, 'eval_rewards/total_composite/std': 0.18232967589910215, 'eval_reward': 0.17142489953682974, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06136284969173945, 'eval_sampling/sampling_logp_difference/max': 1.2111104818490834, 'eval_sampling/importance_sampling_ratio/min': 0.3060183788721378, 'eval_sampling/importance_sampling_ratio/mean': 1.0175598584688628, 'eval_sampling/importance_sampling_ratio/max': 1.4855356491529024, 'eval_entropy': 0.7185706358689529, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.17142489953682974, 'eval_reward_meter_mean': 0.6230765947928796, 'eval_reward_meter_std': 0.39679821867209214, 'eval_reward_count_adherence_mean': 0.8260979927503146, 'eval_reward_count_adherence_std': 0.1801939681172371, 'eval_reward_arabic_clean_mean': 0.8653846153846154, 'eval_reward_arabic_clean_std': 0.26383458421780515, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9951923076923077, 'eval_reward_arabic_floor_std': 0.013598207670908708, 'eval_reward_count_floor_mean': 0.9478294115800124, 'eval_reward_count_floor_std': 0.05405819014861034, 'eval_reward_lexical_plausibility_mean': 0.9509462805894705, 'eval_reward_lexical_plausibility_std': 0.09938178727260003, 'eval_reward_repeat_penalty_mean': 0.4748371495650365, 'eval_reward_repeat_penalty_std': 0.4065972406130571, 'eval_reward_repeat_floor_mean': 0.9208995837431687, 'eval_reward_repeat_floor_std': 0.07328143744514538, 'eval_reward_near_duplicate_penalty_mean': 0.8090284558442923, 'eval_reward_near_duplicate_penalty_std': 0.19457622445546663, 'eval_reward_opening_diversity_mean': 0.844254732131958, 'eval_reward_opening_diversity_std': 0.22699065391833967, 'eval_reward_distinct_2_mean': 0.543222583257235, 'eval_reward_distinct_2_std': 0.40615633817819446, 'eval_reward_judge_quality_mean': 0.3041346153387657, 'eval_reward_judge_quality_std': 0.2317682011769368, 'eval_reward_total_composite_mean': 0.17142489953682974, 'eval_reward_total_composite_std': 0.18232967589910215, 'epoch': 0.02}

 14%|█▎        | 450/3300 [1:38:53<9:31:33, 12.03s/it]

100%|██████████| 13/13 [02:14<00:00, 11.21s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 14%|█▎        | 451/3300 [1:39:10<44:44:51, 56.54s/it]2026-04-12 17:37:36,247 | INFO | train_grpo_train | metrics_logged mode=train step=451

                                                       
{'loss': -0.114, 'grad_norm': 3.049464225769043, 'learning_rate': 8.636363636363637e-06, 'num_tokens': 838342.0, 'completions/mean_length': 180.5, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 70.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.826685905456543, 'rewards/meter/std': 0.34418532252311707, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 0.878539502620697, 'rewards/lexical_plausibility/std': 0.225218266248703, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.1898072361946106, 'rewards/repeat_penalty/mean': 0.7675365209579468, 'rewards/repeat_penalty/std': 0.32234644889831543, 'rewards/repeat_floor/mean': 0.9557523131370544, 'rewards/repeat_floor/std': 0.09061766415834427, 'rewards/near_duplicate_penalty/mean': 0.8593931198120117, 'rewards/near_duplicate_penalty/std': 0.3019898533821106, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.26726123690605164, 'rewards/distinct_2/mean': 0.8260376453399658, 'rewards/distinct_2/std': 0.3385562598705292, 'rewards/total_composite/mean': 0.22023338079452515, 'rewards/total_composite/std': 0.1806122362613678, 'reward': 0.22023338079452515, 'reward_std': 0.1806122213602066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.154558002948761, 'sampling/sampling_logp_difference/max': 1.3370943069458008, 'sampling/importance_sampling_ratio/min': 0.2626076340675354, 'sampling/importance_sampling_ratio/mean': 1.0516215562820435, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9712666273117065, 'clip_ratio/low_mean': 0.011194029822945595, 'clip_ratio/low_min': 0.011194029822945595, 'clip_ratio/high_mean': 0.0904910359531641, 'clip_ratio/high_max': 0.0904910359531641, 'clip_ratio/region_mean': 0.1016850657761097, 'reward_total_mean': 0.22023338079452515, 'reward_meter_mean': 0.826685905456543, 'reward_meter_std': 0.34418532252311707, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 0.878539502620697, 'reward_lexical_plausibility_std': 0.225218266248703, 'reward_repeat_penalty_mean': 0.7675365209579468, 'reward_repeat_penalty_std': 0.32234644889831543, 'reward_repeat_floor_mean': 0.9557523131370544, 'reward_repeat_floor_std': 0.09061766415834427, 'reward_near_duplicate_penalty_mean': 0.8593931198120117, 'reward_near_duplicate_penalty_std': 0.3019898533821106, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.26726123690605164, 'reward_distinct_2_mean': 0.8260376453399658, 'reward_distinct_2_std': 0.3385562598705292, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.1898072361946106, 'reward_total_composite_mean': 0.22023338079452515, 'reward_total_composite_std': 0.1806122362613678, 'epoch': 0.02}

 14%|█▎        | 451/3300 [1:39:10<44:44:51, 56.54s/it]INFO 04-12 17:37:36 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 452/3300 [1:39:19<33:35:12, 42.46s/it]2026-04-12 17:37:45,853 | INFO | train_grpo_train | metrics_logged mode=train step=452

                                                       
{'loss': 0.0923, 'grad_norm': 9.615572929382324, 'learning_rate': 8.633333333333334e-06, 'num_tokens': 840046.0, 'completions/mean_length': 48.0, 'completions/min_length': 41.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9875295758247375, 'rewards/meter/std': 0.005315815564244986, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.20701967179775238, 'rewards/repeat_penalty/mean': 0.47380179166793823, 'rewards/repeat_penalty/std': 0.3864913880825043, 'rewards/repeat_floor/mean': 0.9287439584732056, 'rewards/repeat_floor/std': 0.0697137787938118, 'rewards/near_duplicate_penalty/mean': 0.831121563911438, 'rewards/near_duplicate_penalty/std': 0.19502177834510803, 'rewards/opening_diversity/mean': 0.606249988079071, 'rewards/opening_diversity/std': 0.27894893288612366, 'rewards/distinct_2/mean': 0.7522299289703369, 'rewards/distinct_2/std': 0.30790141224861145, 'rewards/total_composite/mean': 0.41843971610069275, 'rewards/total_composite/std': 0.21579229831695557, 'reward': 0.41843971610069275, 'reward_std': 0.21579229831695557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1466285139322281, 'sampling/sampling_logp_difference/max': 1.9377379417419434, 'sampling/importance_sampling_ratio/min': 0.1440293788909912, 'sampling/importance_sampling_ratio/mean': 1.0440196990966797, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2133660018444061, 'clip_ratio/low_mean': 0.07121916394680738, 'clip_ratio/low_min': 0.07121916394680738, 'clip_ratio/high_mean': 0.05051717162132263, 'clip_ratio/high_max': 0.05051717162132263, 'clip_ratio/region_mean': 0.12173633556813002, 'reward_total_mean': 0.41843971610069275, 'reward_meter_mean': 0.9875295758247375, 'reward_meter_std': 0.005315815564244986, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.47380179166793823, 'reward_repeat_penalty_std': 0.3864913880825043, 'reward_repeat_floor_mean': 0.9287439584732056, 'reward_repeat_floor_std': 0.0697137787938118, 'reward_near_duplicate_penalty_mean': 0.831121563911438, 'reward_near_duplicate_penalty_std': 0.19502177834510803, 'reward_opening_diversity_mean': 0.606249988079071, 'reward_opening_diversity_std': 0.27894893288612366, 'reward_distinct_2_mean': 0.7522299289703369, 'reward_distinct_2_std': 0.30790141224861145, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.20701967179775238, 'reward_total_composite_mean': 0.41843971610069275, 'reward_total_composite_std': 0.21579229831695557, 'epoch': 0.02}

 14%|█▎        | 452/3300 [1:39:19<33:35:12, 42.46s/it]INFO 04-12 17:37:46 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 453/3300 [1:39:28<25:29:12, 32.23s/it]2026-04-12 17:37:54,216 | INFO | train_grpo_train | metrics_logged mode=train step=453

                                                       
{'loss': -0.0161, 'grad_norm': 24.25897789001465, 'learning_rate': 8.630303030303032e-06, 'num_tokens': 841440.0, 'completions/mean_length': 18.25, 'completions/min_length': 16.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.8243906497955322, 'rewards/meter/std': 0.30794912576675415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.8100000023841858, 'rewards/judge_quality/std': 0.22258225083351135, 'rewards/repeat_penalty/mean': 0.7126040458679199, 'rewards/repeat_penalty/std': 0.07829498499631882, 'rewards/repeat_floor/mean': 0.9775208234786987, 'rewards/repeat_floor/std': 0.01565900444984436, 'rewards/near_duplicate_penalty/mean': 0.9501386880874634, 'rewards/near_duplicate_penalty/std': 0.10439330339431763, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6381657719612122, 'rewards/total_composite/std': 0.2966378331184387, 'reward': 0.6381657719612122, 'reward_std': 0.29663780331611633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16541826725006104, 'sampling/sampling_logp_difference/max': 1.881052017211914, 'sampling/importance_sampling_ratio/min': 0.15242967009544373, 'sampling/importance_sampling_ratio/mean': 0.9908430576324463, 'sampling/importance_sampling_ratio/max': 1.7325114011764526, 'entropy': 0.9089064821600914, 'clip_ratio/low_mean': 0.04739583469927311, 'clip_ratio/low_min': 0.04739583469927311, 'clip_ratio/high_mean': 0.06712461402639747, 'clip_ratio/high_max': 0.06712461402639747, 'clip_ratio/region_mean': 0.11452044872567058, 'reward_total_mean': 0.6381657719612122, 'reward_meter_mean': 0.8243906497955322, 'reward_meter_std': 0.30794912576675415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.7126040458679199, 'reward_repeat_penalty_std': 0.07829498499631882, 'reward_repeat_floor_mean': 0.9775208234786987, 'reward_repeat_floor_std': 0.01565900444984436, 'reward_near_duplicate_penalty_mean': 0.9501386880874634, 'reward_near_duplicate_penalty_std': 0.10439330339431763, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8100000023841858, 'reward_judge_quality_std': 0.22258225083351135, 'reward_total_composite_mean': 0.6381657719612122, 'reward_total_composite_std': 0.2966378331184387, 'epoch': 0.02}

 14%|█▎        | 453/3300 [1:39:28<25:29:12, 32.23s/it]INFO 04-12 17:37:54 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 454/3300 [1:39:42<21:11:36, 26.81s/it]2026-04-12 17:38:08,356 | INFO | train_grpo_train | metrics_logged mode=train step=454

                                                       
{'loss': -0.0483, 'grad_norm': 3.276338577270508, 'learning_rate': 8.627272727272727e-06, 'num_tokens': 842988.0, 'completions/mean_length': 161.5, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.66666793823242, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6458823680877686, 'rewards/meter/std': 0.3947608172893524, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.8737679719924927, 'rewards/lexical_plausibility/std': 0.22537542879581451, 'rewards/judge_quality/mean': 0.21249999105930328, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.7688586115837097, 'rewards/repeat_penalty/std': 0.26979172229766846, 'rewards/repeat_floor/mean': 0.9733320474624634, 'rewards/repeat_floor/std': 0.03325796499848366, 'rewards/near_duplicate_penalty/mean': 0.9072303771972656, 'rewards/near_duplicate_penalty/std': 0.11587676405906677, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9208055734634399, 'rewards/distinct_2/std': 0.12130971252918243, 'rewards/total_composite/mean': 0.12177006155252457, 'rewards/total_composite/std': 0.1357794851064682, 'reward': 0.12177006155252457, 'reward_std': 0.135779470205307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15307927131652832, 'sampling/sampling_logp_difference/max': 1.7402987480163574, 'sampling/importance_sampling_ratio/min': 0.17546798288822174, 'sampling/importance_sampling_ratio/mean': 1.0444526672363281, 'sampling/importance_sampling_ratio/max': 1.9058141708374023, 'entropy': 1.0386455580592155, 'clip_ratio/low_mean': 0.01711956551298499, 'clip_ratio/low_min': 0.01711956551298499, 'clip_ratio/high_mean': 0.09181072749197483, 'clip_ratio/high_max': 0.09181072749197483, 'clip_ratio/region_mean': 0.10893029300495982, 'reward_total_mean': 0.12177006155252457, 'reward_meter_mean': 0.6458823680877686, 'reward_meter_std': 0.3947608172893524, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.8737679719924927, 'reward_lexical_plausibility_std': 0.22537542879581451, 'reward_repeat_penalty_mean': 0.7688586115837097, 'reward_repeat_penalty_std': 0.26979172229766846, 'reward_repeat_floor_mean': 0.9733320474624634, 'reward_repeat_floor_std': 0.03325796499848366, 'reward_near_duplicate_penalty_mean': 0.9072303771972656, 'reward_near_duplicate_penalty_std': 0.11587676405906677, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9208055734634399, 'reward_distinct_2_std': 0.12130971252918243, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.12177006155252457, 'reward_total_composite_std': 0.1357794851064682, 'epoch': 0.02}

 14%|█▍        | 454/3300 [1:39:42<21:11:36, 26.81s/it]INFO 04-12 17:38:08 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 455/3300 [1:39:51<17:00:01, 21.51s/it]2026-04-12 17:38:17,512 | INFO | train_grpo_train | metrics_logged mode=train step=455

                                                       
{'loss': -0.0052, 'grad_norm': 11.381220817565918, 'learning_rate': 8.624242424242424e-06, 'num_tokens': 844774.0, 'completions/mean_length': 52.25, 'completions/min_length': 44.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7359862327575684, 'rewards/meter/std': 0.35290268063545227, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.5788249969482422, 'rewards/repeat_penalty/std': 0.28753340244293213, 'rewards/repeat_floor/mean': 0.9511308073997498, 'rewards/repeat_floor/std': 0.03518301993608475, 'rewards/near_duplicate_penalty/mean': 0.8481342196464539, 'rewards/near_duplicate_penalty/std': 0.10036998987197876, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.7309515476226807, 'rewards/distinct_2/std': 0.2138870507478714, 'rewards/total_composite/mean': 0.18463751673698425, 'rewards/total_composite/std': 0.12065467983484268, 'reward': 0.18463751673698425, 'reward_std': 0.12065467238426208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15961751341819763, 'sampling/sampling_logp_difference/max': 1.2055935859680176, 'sampling/importance_sampling_ratio/min': 0.29951414465904236, 'sampling/importance_sampling_ratio/mean': 1.007577657699585, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3414967358112335, 'clip_ratio/low_mean': 0.08063552435487509, 'clip_ratio/low_min': 0.08063552435487509, 'clip_ratio/high_mean': 0.050244271755218506, 'clip_ratio/high_max': 0.050244271755218506, 'clip_ratio/region_mean': 0.1308797961100936, 'reward_total_mean': 0.18463751673698425, 'reward_meter_mean': 0.7359862327575684, 'reward_meter_std': 0.35290268063545227, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5788249969482422, 'reward_repeat_penalty_std': 0.28753340244293213, 'reward_repeat_floor_mean': 0.9511308073997498, 'reward_repeat_floor_std': 0.03518301993608475, 'reward_near_duplicate_penalty_mean': 0.8481342196464539, 'reward_near_duplicate_penalty_std': 0.10036998987197876, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.7309515476226807, 'reward_distinct_2_std': 0.2138870507478714, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.18463751673698425, 'reward_total_composite_std': 0.12065467983484268, 'epoch': 0.02}

 14%|█▍        | 455/3300 [1:39:51<17:00:01, 21.51s/it]INFO 04-12 17:38:17 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 456/3300 [1:40:04<15:05:14, 19.10s/it]2026-04-12 17:38:30,975 | INFO | train_grpo_train | metrics_logged mode=train step=456

                                                       
{'loss': -0.0678, 'grad_norm': 7.083357810974121, 'learning_rate': 8.621212121212122e-06, 'num_tokens': 846441.0, 'completions/mean_length': 103.375, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.000003814697266, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.37655889987945557, 'rewards/meter/std': 0.3353674113750458, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9286011457443237, 'rewards/lexical_plausibility/std': 0.20194637775421143, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.26692694425582886, 'rewards/repeat_penalty/mean': 0.755588173866272, 'rewards/repeat_penalty/std': 0.336740642786026, 'rewards/repeat_floor/mean': 0.9539693593978882, 'rewards/repeat_floor/std': 0.09076379239559174, 'rewards/near_duplicate_penalty/mean': 0.8471053838729858, 'rewards/near_duplicate_penalty/std': 0.3191220164299011, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.8162891268730164, 'rewards/distinct_2/std': 0.34389954805374146, 'rewards/total_composite/mean': 0.13373219966888428, 'rewards/total_composite/std': 0.11370719969272614, 'reward': 0.13373219966888428, 'reward_std': 0.11370719969272614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16646528244018555, 'sampling/sampling_logp_difference/max': 2.7328405380249023, 'sampling/importance_sampling_ratio/min': 0.06503429263830185, 'sampling/importance_sampling_ratio/mean': 1.0426028966903687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9241776242852211, 'clip_ratio/low_mean': 0.05442966986447573, 'clip_ratio/low_min': 0.05442966986447573, 'clip_ratio/high_mean': 0.040013296995311975, 'clip_ratio/high_max': 0.040013296995311975, 'clip_ratio/region_mean': 0.0944429668597877, 'reward_total_mean': 0.13373219966888428, 'reward_meter_mean': 0.37655889987945557, 'reward_meter_std': 0.3353674113750458, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9286011457443237, 'reward_lexical_plausibility_std': 0.20194637775421143, 'reward_repeat_penalty_mean': 0.755588173866272, 'reward_repeat_penalty_std': 0.336740642786026, 'reward_repeat_floor_mean': 0.9539693593978882, 'reward_repeat_floor_std': 0.09076379239559174, 'reward_near_duplicate_penalty_mean': 0.8471053838729858, 'reward_near_duplicate_penalty_std': 0.3191220164299011, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.8162891268730164, 'reward_distinct_2_std': 0.34389954805374146, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.26692694425582886, 'reward_total_composite_mean': 0.13373219966888428, 'reward_total_composite_std': 0.11370719969272614, 'epoch': 0.02}

 14%|█▍        | 456/3300 [1:40:04<15:05:14, 19.10s/it]INFO 04-12 17:38:31 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 457/3300 [1:40:13<12:42:00, 16.08s/it]2026-04-12 17:38:40,022 | INFO | train_grpo_train | metrics_logged mode=train step=457

                                                       
{'loss': 0.1114, 'grad_norm': 16.963098526000977, 'learning_rate': 8.618181818181819e-06, 'num_tokens': 848264.0, 'completions/mean_length': 50.875, 'completions/min_length': 41.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.45969995856285095, 'rewards/meter/std': 0.3252045512199402, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.9448733329772949, 'rewards/repeat_penalty/std': 0.06891842931509018, 'rewards/repeat_floor/mean': 0.9932032823562622, 'rewards/repeat_floor/std': 0.007673040963709354, 'rewards/near_duplicate_penalty/mean': 0.9714715480804443, 'rewards/near_duplicate_penalty/std': 0.026079701259732246, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9720279574394226, 'rewards/distinct_2/std': 0.05644134432077408, 'rewards/total_composite/mean': 0.21489249169826508, 'rewards/total_composite/std': 0.20255127549171448, 'reward': 0.21489249169826508, 'reward_std': 0.20255127549171448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20060986280441284, 'sampling/sampling_logp_difference/max': 1.8694686889648438, 'sampling/importance_sampling_ratio/min': 0.1542055755853653, 'sampling/importance_sampling_ratio/mean': 1.000983476638794, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2276152670383453, 'clip_ratio/low_mean': 0.07881452236324549, 'clip_ratio/low_min': 0.07881452236324549, 'clip_ratio/high_mean': 0.07786296680569649, 'clip_ratio/high_max': 0.07786296680569649, 'clip_ratio/region_mean': 0.15667748916894197, 'reward_total_mean': 0.21489249169826508, 'reward_meter_mean': 0.45969995856285095, 'reward_meter_std': 0.3252045512199402, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9448733329772949, 'reward_repeat_penalty_std': 0.06891842931509018, 'reward_repeat_floor_mean': 0.9932032823562622, 'reward_repeat_floor_std': 0.007673040963709354, 'reward_near_duplicate_penalty_mean': 0.9714715480804443, 'reward_near_duplicate_penalty_std': 0.026079701259732246, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9720279574394226, 'reward_distinct_2_std': 0.05644134432077408, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.21489249169826508, 'reward_total_composite_std': 0.20255127549171448, 'epoch': 0.02}

 14%|█▍        | 457/3300 [1:40:13<12:42:00, 16.08s/it]INFO 04-12 17:38:40 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 458/3300 [1:40:22<10:53:49, 13.80s/it]2026-04-12 17:38:48,508 | INFO | train_grpo_train | metrics_logged mode=train step=458

                                                       
{'loss': -0.0439, 'grad_norm': 13.403787612915039, 'learning_rate': 8.615151515151516e-06, 'num_tokens': 849850.0, 'completions/mean_length': 43.25, 'completions/min_length': 37.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9791386127471924, 'rewards/meter/std': 0.019205288961529732, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.8572620153427124, 'rewards/repeat_penalty/std': 0.14656752347946167, 'rewards/repeat_floor/mean': 0.9833243489265442, 'rewards/repeat_floor/std': 0.016158923506736755, 'rewards/near_duplicate_penalty/mean': 0.9427228569984436, 'rewards/near_duplicate_penalty/std': 0.048306528478860855, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9310103058815002, 'rewards/distinct_2/std': 0.07339107990264893, 'rewards/total_composite/mean': 0.402590274810791, 'rewards/total_composite/std': 0.2356213629245758, 'reward': 0.402590274810791, 'reward_std': 0.2356213480234146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14252756536006927, 'sampling/sampling_logp_difference/max': 1.5148115158081055, 'sampling/importance_sampling_ratio/min': 0.2198496162891388, 'sampling/importance_sampling_ratio/mean': 1.0008864402770996, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0320435538887978, 'clip_ratio/low_mean': 0.04681248031556606, 'clip_ratio/low_min': 0.04681248031556606, 'clip_ratio/high_mean': 0.07551885256543756, 'clip_ratio/high_max': 0.07551885256543756, 'clip_ratio/region_mean': 0.12233133288100362, 'reward_total_mean': 0.402590274810791, 'reward_meter_mean': 0.9791386127471924, 'reward_meter_std': 0.019205288961529732, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8572620153427124, 'reward_repeat_penalty_std': 0.14656752347946167, 'reward_repeat_floor_mean': 0.9833243489265442, 'reward_repeat_floor_std': 0.016158923506736755, 'reward_near_duplicate_penalty_mean': 0.9427228569984436, 'reward_near_duplicate_penalty_std': 0.048306528478860855, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9310103058815002, 'reward_distinct_2_std': 0.07339107990264893, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.402590274810791, 'reward_total_composite_std': 0.2356213629245758, 'epoch': 0.02}

 14%|█▍        | 458/3300 [1:40:22<10:53:49, 13.80s/it]INFO 04-12 17:38:48 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 459/3300 [1:40:33<10:14:12, 12.97s/it]2026-04-12 17:38:59,537 | INFO | train_grpo_train | metrics_logged mode=train step=459

                                                       
{'loss': 0.0458, 'grad_norm': 4.057854175567627, 'learning_rate': 8.612121212121213e-06, 'num_tokens': 853245.0, 'completions/mean_length': 220.375, 'completions/min_length': 196.0, 'completions/max_length': 253.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 220.375, 'completions/min_terminated_length': 196.0, 'completions/max_terminated_length': 253.0, 'rewards/meter/mean': 0.9919022917747498, 'rewards/meter/std': 0.0029621252324432135, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9142856597900391, 'rewards/count_floor/std': 0.022908104583621025, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.003524602623656392, 'rewards/repeat_penalty/std': 0.00656402762979269, 'rewards/repeat_floor/mean': 0.8029244542121887, 'rewards/repeat_floor/std': 0.030130090191960335, 'rewards/near_duplicate_penalty/mean': 0.5114600658416748, 'rewards/near_duplicate_penalty/std': 0.10638757795095444, 'rewards/opening_diversity/mean': 0.4069216251373291, 'rewards/opening_diversity/std': 0.2500130832195282, 'rewards/distinct_2/mean': 0.019890649244189262, 'rewards/distinct_2/std': 0.03774602338671684, 'rewards/total_composite/mean': 0.12793518602848053, 'rewards/total_composite/std': 0.054545022547245026, 'reward': 0.12793518602848053, 'reward_std': 0.05454501882195473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06892064213752747, 'sampling/sampling_logp_difference/max': 1.9090232849121094, 'sampling/importance_sampling_ratio/min': 0.1482250988483429, 'sampling/importance_sampling_ratio/mean': 1.0110036134719849, 'sampling/importance_sampling_ratio/max': 1.9041608572006226, 'entropy': 0.5043975673615932, 'clip_ratio/low_mean': 0.04216759325936437, 'clip_ratio/low_min': 0.04216759325936437, 'clip_ratio/high_mean': 0.006840796209871769, 'clip_ratio/high_max': 0.006840796209871769, 'clip_ratio/region_mean': 0.049008389469236135, 'reward_total_mean': 0.12793518602848053, 'reward_meter_mean': 0.9919022917747498, 'reward_meter_std': 0.0029621252324432135, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9142856597900391, 'reward_count_floor_std': 0.022908104583621025, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.003524602623656392, 'reward_repeat_penalty_std': 0.00656402762979269, 'reward_repeat_floor_mean': 0.8029244542121887, 'reward_repeat_floor_std': 0.030130090191960335, 'reward_near_duplicate_penalty_mean': 0.5114600658416748, 'reward_near_duplicate_penalty_std': 0.10638757795095444, 'reward_opening_diversity_mean': 0.4069216251373291, 'reward_opening_diversity_std': 0.2500130832195282, 'reward_distinct_2_mean': 0.019890649244189262, 'reward_distinct_2_std': 0.03774602338671684, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.12793518602848053, 'reward_total_composite_std': 0.054545022547245026, 'epoch': 0.02}

 14%|█▍        | 459/3300 [1:40:33<10:14:12, 12.97s/it]INFO 04-12 17:38:59 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 460/3300 [1:40:42<9:21:44, 11.87s/it] 2026-04-12 17:39:08,830 | INFO | train_grpo_train | metrics_logged mode=train step=460

                                                      
{'loss': -0.0389, 'grad_norm': 9.394753456115723, 'learning_rate': 8.60909090909091e-06, 'num_tokens': 855270.0, 'completions/mean_length': 69.125, 'completions/min_length': 41.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9830876588821411, 'rewards/meter/std': 0.0073325373232364655, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.6307567954063416, 'rewards/repeat_penalty/std': 0.290141224861145, 'rewards/repeat_floor/mean': 0.9599894285202026, 'rewards/repeat_floor/std': 0.0397547222673893, 'rewards/near_duplicate_penalty/mean': 0.8877747058868408, 'rewards/near_duplicate_penalty/std': 0.10667216032743454, 'rewards/opening_diversity/mean': 0.949999988079071, 'rewards/opening_diversity/std': 0.1414213478565216, 'rewards/distinct_2/mean': 0.7758134603500366, 'rewards/distinct_2/std': 0.18687871098518372, 'rewards/total_composite/mean': 0.3974243998527527, 'rewards/total_composite/std': 0.138163223862648, 'reward': 0.3974243998527527, 'reward_std': 0.138163223862648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13478593528270721, 'sampling/sampling_logp_difference/max': 1.2511844635009766, 'sampling/importance_sampling_ratio/min': 0.28616565465927124, 'sampling/importance_sampling_ratio/mean': 1.0205570459365845, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1178451552987099, 'clip_ratio/low_mean': 0.029906955547630787, 'clip_ratio/low_min': 0.029906955547630787, 'clip_ratio/high_mean': 0.11511932034045458, 'clip_ratio/high_max': 0.11511932034045458, 'clip_ratio/region_mean': 0.14502627588808537, 'reward_total_mean': 0.3974243998527527, 'reward_meter_mean': 0.9830876588821411, 'reward_meter_std': 0.0073325373232364655, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.6307567954063416, 'reward_repeat_penalty_std': 0.290141224861145, 'reward_repeat_floor_mean': 0.9599894285202026, 'reward_repeat_floor_std': 0.0397547222673893, 'reward_near_duplicate_penalty_mean': 0.8877747058868408, 'reward_near_duplicate_penalty_std': 0.10667216032743454, 'reward_opening_diversity_mean': 0.949999988079071, 'reward_opening_diversity_std': 0.1414213478565216, 'reward_distinct_2_mean': 0.7758134603500366, 'reward_distinct_2_std': 0.18687871098518372, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.3974243998527527, 'reward_total_composite_std': 0.138163223862648, 'epoch': 0.02}

 14%|█▍        | 460/3300 [1:40:42<9:21:44, 11.87s/it]INFO 04-12 17:39:09 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 461/3300 [1:40:51<8:42:10, 11.04s/it]2026-04-12 17:39:17,948 | INFO | train_grpo_train | metrics_logged mode=train step=461

                                                      
{'loss': -0.0315, 'grad_norm': 15.864015579223633, 'learning_rate': 8.606060606060606e-06, 'num_tokens': 856881.0, 'completions/mean_length': 45.375, 'completions/min_length': 40.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9021110534667969, 'rewards/meter/std': 0.22328464686870575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.992229700088501, 'rewards/repeat_penalty/std': 0.015372724272310734, 'rewards/repeat_floor/mean': 0.998834490776062, 'rewards/repeat_floor/std': 0.0023059083614498377, 'rewards/near_duplicate_penalty/mean': 0.992229700088501, 'rewards/near_duplicate_penalty/std': 0.015372724272310734, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4254631996154785, 'rewards/total_composite/std': 0.23813623189926147, 'reward': 0.4254631996154785, 'reward_std': 0.23813621699810028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1576824188232422, 'sampling/sampling_logp_difference/max': 1.5359315872192383, 'sampling/importance_sampling_ratio/min': 0.21525508165359497, 'sampling/importance_sampling_ratio/mean': 1.0277284383773804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.243036836385727, 'clip_ratio/low_mean': 0.05429601017385721, 'clip_ratio/low_min': 0.05429601017385721, 'clip_ratio/high_mean': 0.11604143399745226, 'clip_ratio/high_max': 0.11604143399745226, 'clip_ratio/region_mean': 0.17033744417130947, 'reward_total_mean': 0.4254631996154785, 'reward_meter_mean': 0.9021110534667969, 'reward_meter_std': 0.22328464686870575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.992229700088501, 'reward_repeat_penalty_std': 0.015372724272310734, 'reward_repeat_floor_mean': 0.998834490776062, 'reward_repeat_floor_std': 0.0023059083614498377, 'reward_near_duplicate_penalty_mean': 0.992229700088501, 'reward_near_duplicate_penalty_std': 0.015372724272310734, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.4254631996154785, 'reward_total_composite_std': 0.23813623189926147, 'epoch': 0.02}

 14%|█▍        | 461/3300 [1:40:51<8:42:10, 11.04s/it]INFO 04-12 17:39:18 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 462/3300 [1:41:01<8:19:13, 10.55s/it]2026-04-12 17:39:27,356 | INFO | train_grpo_train | metrics_logged mode=train step=462

                                                      
{'loss': 0.0783, 'grad_norm': 13.244956016540527, 'learning_rate': 8.603030303030303e-06, 'num_tokens': 858544.0, 'completions/mean_length': 42.875, 'completions/min_length': 39.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.661554217338562, 'rewards/meter/std': 0.45109081268310547, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.9347594976425171, 'rewards/repeat_penalty/std': 0.05908624082803726, 'rewards/repeat_floor/mean': 0.9911521673202515, 'rewards/repeat_floor/std': 0.008226786740124226, 'rewards/near_duplicate_penalty/mean': 0.9515254497528076, 'rewards/near_duplicate_penalty/std': 0.052539508789777756, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9824812412261963, 'rewards/distinct_2/std': 0.03248300403356552, 'rewards/total_composite/mean': 0.40324968099594116, 'rewards/total_composite/std': 0.3470838665962219, 'reward': 0.40324968099594116, 'reward_std': 0.3470838665962219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1557810753583908, 'sampling/sampling_logp_difference/max': 1.781308889389038, 'sampling/importance_sampling_ratio/min': 0.1684175729751587, 'sampling/importance_sampling_ratio/mean': 1.0169174671173096, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1622472777962685, 'clip_ratio/low_mean': 0.07083333376795053, 'clip_ratio/low_min': 0.07083333376795053, 'clip_ratio/high_mean': 0.07205445971339941, 'clip_ratio/high_max': 0.07205445971339941, 'clip_ratio/region_mean': 0.14288779348134995, 'reward_total_mean': 0.40324968099594116, 'reward_meter_mean': 0.661554217338562, 'reward_meter_std': 0.45109081268310547, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9347594976425171, 'reward_repeat_penalty_std': 0.05908624082803726, 'reward_repeat_floor_mean': 0.9911521673202515, 'reward_repeat_floor_std': 0.008226786740124226, 'reward_near_duplicate_penalty_mean': 0.9515254497528076, 'reward_near_duplicate_penalty_std': 0.052539508789777756, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9824812412261963, 'reward_distinct_2_std': 0.03248300403356552, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.40324968099594116, 'reward_total_composite_std': 0.3470838665962219, 'epoch': 0.02}

 14%|█▍        | 462/3300 [1:41:01<8:19:13, 10.55s/it]INFO 04-12 17:39:27 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 463/3300 [1:41:11<8:13:33, 10.44s/it]2026-04-12 17:39:37,522 | INFO | train_grpo_train | metrics_logged mode=train step=463

                                                      
{'loss': -0.0065, 'grad_norm': 16.66472625732422, 'learning_rate': 8.6e-06, 'num_tokens': 860257.0, 'completions/mean_length': 41.125, 'completions/min_length': 32.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8129892349243164, 'rewards/meter/std': 0.21980492770671844, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136146903038025, 'rewards/repeat_penalty/mean': 0.6516164541244507, 'rewards/repeat_penalty/std': 0.29724743962287903, 'rewards/repeat_floor/mean': 0.9582736492156982, 'rewards/repeat_floor/std': 0.03881257027387619, 'rewards/near_duplicate_penalty/mean': 0.8899881839752197, 'rewards/near_duplicate_penalty/std': 0.09444312751293182, 'rewards/opening_diversity/mean': 0.809374988079071, 'rewards/opening_diversity/std': 0.23449699580669403, 'rewards/distinct_2/mean': 0.8468108177185059, 'rewards/distinct_2/std': 0.14745701849460602, 'rewards/total_composite/mean': 0.2842075228691101, 'rewards/total_composite/std': 0.1423080712556839, 'reward': 0.2842075228691101, 'reward_std': 0.1423080712556839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15802541375160217, 'sampling/sampling_logp_difference/max': 1.492415428161621, 'sampling/importance_sampling_ratio/min': 0.22482894361019135, 'sampling/importance_sampling_ratio/mean': 1.012261986732483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2356071174144745, 'clip_ratio/low_mean': 0.07391707878559828, 'clip_ratio/low_min': 0.07391707878559828, 'clip_ratio/high_mean': 0.07254288718104362, 'clip_ratio/high_max': 0.07254288718104362, 'clip_ratio/region_mean': 0.1464599659666419, 'reward_total_mean': 0.2842075228691101, 'reward_meter_mean': 0.8129892349243164, 'reward_meter_std': 0.21980492770671844, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6516164541244507, 'reward_repeat_penalty_std': 0.29724743962287903, 'reward_repeat_floor_mean': 0.9582736492156982, 'reward_repeat_floor_std': 0.03881257027387619, 'reward_near_duplicate_penalty_mean': 0.8899881839752197, 'reward_near_duplicate_penalty_std': 0.09444312751293182, 'reward_opening_diversity_mean': 0.809374988079071, 'reward_opening_diversity_std': 0.23449699580669403, 'reward_distinct_2_mean': 0.8468108177185059, 'reward_distinct_2_std': 0.14745701849460602, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136146903038025, 'reward_total_composite_mean': 0.2842075228691101, 'reward_total_composite_std': 0.1423080712556839, 'epoch': 0.02}

 14%|█▍        | 463/3300 [1:41:11<8:13:33, 10.44s/it]INFO 04-12 17:39:37 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 464/3300 [1:41:19<7:42:34,  9.79s/it]2026-04-12 17:39:45,788 | INFO | train_grpo_train | metrics_logged mode=train step=464

                                                      
{'loss': 0.1111, 'grad_norm': 20.328224182128906, 'learning_rate': 8.596969696969698e-06, 'num_tokens': 861760.0, 'completions/mean_length': 35.875, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6358775496482849, 'rewards/meter/std': 0.4272748827934265, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45374998450279236, 'rewards/judge_quality/std': 0.20170257985591888, 'rewards/repeat_penalty/mean': 0.8019654750823975, 'rewards/repeat_penalty/std': 0.2824261486530304, 'rewards/repeat_floor/mean': 0.9779501557350159, 'rewards/repeat_floor/std': 0.030076604336500168, 'rewards/near_duplicate_penalty/mean': 0.9389964938163757, 'rewards/near_duplicate_penalty/std': 0.06664886325597763, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8983410000801086, 'rewards/distinct_2/std': 0.15062162280082703, 'rewards/total_composite/mean': 0.26632142066955566, 'rewards/total_composite/std': 0.17347563803195953, 'reward': 0.26632142066955566, 'reward_std': 0.17347563803195953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18181434273719788, 'sampling/sampling_logp_difference/max': 1.780573844909668, 'sampling/importance_sampling_ratio/min': 0.16854140162467957, 'sampling/importance_sampling_ratio/mean': 1.0317375659942627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3245523124933243, 'clip_ratio/low_mean': 0.049052730202674866, 'clip_ratio/low_min': 0.049052730202674866, 'clip_ratio/high_mean': 0.10146208759397268, 'clip_ratio/high_max': 0.10146208759397268, 'clip_ratio/region_mean': 0.15051481779664755, 'reward_total_mean': 0.26632142066955566, 'reward_meter_mean': 0.6358775496482849, 'reward_meter_std': 0.4272748827934265, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8019654750823975, 'reward_repeat_penalty_std': 0.2824261486530304, 'reward_repeat_floor_mean': 0.9779501557350159, 'reward_repeat_floor_std': 0.030076604336500168, 'reward_near_duplicate_penalty_mean': 0.9389964938163757, 'reward_near_duplicate_penalty_std': 0.06664886325597763, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8983410000801086, 'reward_distinct_2_std': 0.15062162280082703, 'reward_judge_quality_mean': 0.45374998450279236, 'reward_judge_quality_std': 0.20170257985591888, 'reward_total_composite_mean': 0.26632142066955566, 'reward_total_composite_std': 0.17347563803195953, 'epoch': 0.02}

 14%|█▍        | 464/3300 [1:41:19<7:42:34,  9.79s/it]INFO 04-12 17:39:46 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 465/3300 [1:41:28<7:29:22,  9.51s/it]2026-04-12 17:39:54,658 | INFO | train_grpo_train | metrics_logged mode=train step=465

                                                      
{'loss': -0.0292, 'grad_norm': 12.978865623474121, 'learning_rate': 8.593939393939395e-06, 'num_tokens': 863731.0, 'completions/mean_length': 63.375, 'completions/min_length': 56.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.8288000226020813, 'rewards/meter/std': 0.3157835006713867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9752435088157654, 'rewards/lexical_plausibility/std': 0.046044979244470596, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.2828427255153656, 'rewards/repeat_penalty/mean': 0.26078465580940247, 'rewards/repeat_penalty/std': 0.253579705953598, 'rewards/repeat_floor/mean': 0.885047197341919, 'rewards/repeat_floor/std': 0.0638018250465393, 'rewards/near_duplicate_penalty/mean': 0.7032175660133362, 'rewards/near_duplicate_penalty/std': 0.17859706282615662, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.32561755180358887, 'rewards/distinct_2/mean': 0.4465510845184326, 'rewards/distinct_2/std': 0.22540301084518433, 'rewards/total_composite/mean': 0.20578667521476746, 'rewards/total_composite/std': 0.28532278537750244, 'reward': 0.20578667521476746, 'reward_std': 0.28532278537750244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11823023110628128, 'sampling/sampling_logp_difference/max': 2.05720853805542, 'sampling/importance_sampling_ratio/min': 0.1278102546930313, 'sampling/importance_sampling_ratio/mean': 1.0175706148147583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9275213107466698, 'clip_ratio/low_mean': 0.08856649976223707, 'clip_ratio/low_min': 0.08856649976223707, 'clip_ratio/high_mean': 0.01584506966173649, 'clip_ratio/high_max': 0.01584506966173649, 'clip_ratio/region_mean': 0.10441156942397356, 'reward_total_mean': 0.20578667521476746, 'reward_meter_mean': 0.8288000226020813, 'reward_meter_std': 0.3157835006713867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9752435088157654, 'reward_lexical_plausibility_std': 0.046044979244470596, 'reward_repeat_penalty_mean': 0.26078465580940247, 'reward_repeat_penalty_std': 0.253579705953598, 'reward_repeat_floor_mean': 0.885047197341919, 'reward_repeat_floor_std': 0.0638018250465393, 'reward_near_duplicate_penalty_mean': 0.7032175660133362, 'reward_near_duplicate_penalty_std': 0.17859706282615662, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.32561755180358887, 'reward_distinct_2_mean': 0.4465510845184326, 'reward_distinct_2_std': 0.22540301084518433, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.2828427255153656, 'reward_total_composite_mean': 0.20578667521476746, 'reward_total_composite_std': 0.28532278537750244, 'epoch': 0.02}

 14%|█▍        | 465/3300 [1:41:28<7:29:22,  9.51s/it]INFO 04-12 17:39:54 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 466/3300 [1:41:37<7:18:28,  9.28s/it]2026-04-12 17:40:03,408 | INFO | train_grpo_train | metrics_logged mode=train step=466

                                                      
{'loss': 0.0338, 'grad_norm': 12.60587215423584, 'learning_rate': 8.590909090909092e-06, 'num_tokens': 865465.0, 'completions/mean_length': 47.75, 'completions/min_length': 43.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.846108615398407, 'rewards/meter/std': 0.24170660972595215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.5401484370231628, 'rewards/repeat_penalty/std': 0.40161025524139404, 'rewards/repeat_floor/mean': 0.9349104762077332, 'rewards/repeat_floor/std': 0.05784101411700249, 'rewards/near_duplicate_penalty/mean': 0.8248052597045898, 'rewards/near_duplicate_penalty/std': 0.1349523812532425, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.2734534442424774, 'rewards/distinct_2/mean': 0.7146074771881104, 'rewards/distinct_2/std': 0.2804373800754547, 'rewards/total_composite/mean': 0.2843344211578369, 'rewards/total_composite/std': 0.1477406769990921, 'reward': 0.2843344211578369, 'reward_std': 0.1477406919002533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15633729100227356, 'sampling/sampling_logp_difference/max': 1.929990291595459, 'sampling/importance_sampling_ratio/min': 0.1451496034860611, 'sampling/importance_sampling_ratio/mean': 1.035498023033142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.195248268544674, 'clip_ratio/low_mean': 0.040937500074505806, 'clip_ratio/low_min': 0.040937500074505806, 'clip_ratio/high_mean': 0.09840234834700823, 'clip_ratio/high_max': 0.09840234834700823, 'clip_ratio/region_mean': 0.13933984842151403, 'reward_total_mean': 0.2843344211578369, 'reward_meter_mean': 0.846108615398407, 'reward_meter_std': 0.24170660972595215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5401484370231628, 'reward_repeat_penalty_std': 0.40161025524139404, 'reward_repeat_floor_mean': 0.9349104762077332, 'reward_repeat_floor_std': 0.05784101411700249, 'reward_near_duplicate_penalty_mean': 0.8248052597045898, 'reward_near_duplicate_penalty_std': 0.1349523812532425, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.2734534442424774, 'reward_distinct_2_mean': 0.7146074771881104, 'reward_distinct_2_std': 0.2804373800754547, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.2843344211578369, 'reward_total_composite_std': 0.1477406769990921, 'epoch': 0.02}

 14%|█▍        | 466/3300 [1:41:37<7:18:28,  9.28s/it]INFO 04-12 17:40:03 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 467/3300 [1:41:51<8:25:33, 10.71s/it]2026-04-12 17:40:17,440 | INFO | train_grpo_train | metrics_logged mode=train step=467

                                                      
{'loss': -0.0011, 'grad_norm': 5.124169826507568, 'learning_rate': 8.587878787878788e-06, 'num_tokens': 867576.0, 'completions/mean_length': 141.875, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 89.00000762939453, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7691898345947266, 'rewards/meter/std': 0.37946081161499023, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.971875011920929, 'rewards/count_floor/std': 0.07954952120780945, 'rewards/lexical_plausibility/mean': 0.9341220259666443, 'rewards/lexical_plausibility/std': 0.1792246401309967, 'rewards/judge_quality/mean': 0.3162499964237213, 'rewards/judge_quality/std': 0.2650572657585144, 'rewards/repeat_penalty/mean': 0.7145575881004333, 'rewards/repeat_penalty/std': 0.29141151905059814, 'rewards/repeat_floor/mean': 0.9686734676361084, 'rewards/repeat_floor/std': 0.03696190565824509, 'rewards/near_duplicate_penalty/mean': 0.9290385246276855, 'rewards/near_duplicate_penalty/std': 0.08091890811920166, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.8326969146728516, 'rewards/distinct_2/std': 0.24042244255542755, 'rewards/total_composite/mean': 0.22213882207870483, 'rewards/total_composite/std': 0.27385786175727844, 'reward': 0.22213882207870483, 'reward_std': 0.27385786175727844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16399723291397095, 'sampling/sampling_logp_difference/max': 1.9604716300964355, 'sampling/importance_sampling_ratio/min': 0.14079201221466064, 'sampling/importance_sampling_ratio/mean': 1.0121349096298218, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.01725622266531, 'clip_ratio/low_mean': 0.08368723886087537, 'clip_ratio/low_min': 0.08368723886087537, 'clip_ratio/high_mean': 0.03216309007257223, 'clip_ratio/high_max': 0.03216309007257223, 'clip_ratio/region_mean': 0.1158503289334476, 'reward_total_mean': 0.22213882207870483, 'reward_meter_mean': 0.7691898345947266, 'reward_meter_std': 0.37946081161499023, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.971875011920929, 'reward_count_floor_std': 0.07954952120780945, 'reward_lexical_plausibility_mean': 0.9341220259666443, 'reward_lexical_plausibility_std': 0.1792246401309967, 'reward_repeat_penalty_mean': 0.7145575881004333, 'reward_repeat_penalty_std': 0.29141151905059814, 'reward_repeat_floor_mean': 0.9686734676361084, 'reward_repeat_floor_std': 0.03696190565824509, 'reward_near_duplicate_penalty_mean': 0.9290385246276855, 'reward_near_duplicate_penalty_std': 0.08091890811920166, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.8326969146728516, 'reward_distinct_2_std': 0.24042244255542755, 'reward_judge_quality_mean': 0.3162499964237213, 'reward_judge_quality_std': 0.2650572657585144, 'reward_total_composite_mean': 0.22213882207870483, 'reward_total_composite_std': 0.27385786175727844, 'epoch': 0.02}

 14%|█▍        | 467/3300 [1:41:51<8:25:33, 10.71s/it]INFO 04-12 17:40:17 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 468/3300 [1:42:00<8:00:18, 10.18s/it]2026-04-12 17:40:26,375 | INFO | train_grpo_train | metrics_logged mode=train step=468

                                                      
{'loss': 0.0493, 'grad_norm': 8.99153995513916, 'learning_rate': 8.584848484848485e-06, 'num_tokens': 869645.0, 'completions/mean_length': 81.625, 'completions/min_length': 76.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.625, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9223024249076843, 'rewards/meter/std': 0.09635310620069504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.12056658416986465, 'rewards/repeat_penalty/std': 0.1343386173248291, 'rewards/repeat_floor/mean': 0.8572448492050171, 'rewards/repeat_floor/std': 0.050898995250463486, 'rewards/near_duplicate_penalty/mean': 0.6321216821670532, 'rewards/near_duplicate_penalty/std': 0.15244291722774506, 'rewards/opening_diversity/mean': 0.5390625, 'rewards/opening_diversity/std': 0.2542880177497864, 'rewards/distinct_2/mean': 0.33425378799438477, 'rewards/distinct_2/std': 0.1574835479259491, 'rewards/total_composite/mean': 0.1184370294213295, 'rewards/total_composite/std': 0.01286934595555067, 'reward': 0.1184370294213295, 'reward_std': 0.01286934781819582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10954656451940536, 'sampling/sampling_logp_difference/max': 1.9089059829711914, 'sampling/importance_sampling_ratio/min': 0.14824248850345612, 'sampling/importance_sampling_ratio/mean': 1.0160108804702759, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8124596178531647, 'clip_ratio/low_mean': 0.03755123261362314, 'clip_ratio/low_min': 0.03755123261362314, 'clip_ratio/high_mean': 0.05008876696228981, 'clip_ratio/high_max': 0.05008876696228981, 'clip_ratio/region_mean': 0.08763999957591295, 'reward_total_mean': 0.1184370294213295, 'reward_meter_mean': 0.9223024249076843, 'reward_meter_std': 0.09635310620069504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.12056658416986465, 'reward_repeat_penalty_std': 0.1343386173248291, 'reward_repeat_floor_mean': 0.8572448492050171, 'reward_repeat_floor_std': 0.050898995250463486, 'reward_near_duplicate_penalty_mean': 0.6321216821670532, 'reward_near_duplicate_penalty_std': 0.15244291722774506, 'reward_opening_diversity_mean': 0.5390625, 'reward_opening_diversity_std': 0.2542880177497864, 'reward_distinct_2_mean': 0.33425378799438477, 'reward_distinct_2_std': 0.1574835479259491, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.1184370294213295, 'reward_total_composite_std': 0.01286934595555067, 'epoch': 0.02}

 14%|█▍        | 468/3300 [1:42:00<8:00:18, 10.18s/it]INFO 04-12 17:40:26 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 469/3300 [1:42:07<7:26:01,  9.45s/it]2026-04-12 17:40:34,141 | INFO | train_grpo_train | metrics_logged mode=train step=469

                                                      
{'loss': 0.0338, 'grad_norm': 14.935983657836914, 'learning_rate': 8.581818181818183e-06, 'num_tokens': 871213.0, 'completions/mean_length': 40.0, 'completions/min_length': 35.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.4505701959133148, 'rewards/meter/std': 0.4177137017250061, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.2655990719795227, 'rewards/repeat_penalty/mean': 0.8372052907943726, 'rewards/repeat_penalty/std': 0.24754536151885986, 'rewards/repeat_floor/mean': 0.9829343557357788, 'rewards/repeat_floor/std': 0.022863373160362244, 'rewards/near_duplicate_penalty/mean': 0.9355204105377197, 'rewards/near_duplicate_penalty/std': 0.06568436324596405, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9595141410827637, 'rewards/distinct_2/std': 0.08656234294176102, 'rewards/total_composite/mean': 0.38056474924087524, 'rewards/total_composite/std': 0.4107729196548462, 'reward': 0.38056474924087524, 'reward_std': 0.4107729196548462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13861869275569916, 'sampling/sampling_logp_difference/max': 2.2985057830810547, 'sampling/importance_sampling_ratio/min': 0.10040877014398575, 'sampling/importance_sampling_ratio/mean': 1.007846713066101, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8715974390506744, 'clip_ratio/low_mean': 0.07247200142592192, 'clip_ratio/low_min': 0.07247200142592192, 'clip_ratio/high_mean': 0.06966928578913212, 'clip_ratio/high_max': 0.06966928578913212, 'clip_ratio/region_mean': 0.14214128721505404, 'reward_total_mean': 0.38056474924087524, 'reward_meter_mean': 0.4505701959133148, 'reward_meter_std': 0.4177137017250061, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.8372052907943726, 'reward_repeat_penalty_std': 0.24754536151885986, 'reward_repeat_floor_mean': 0.9829343557357788, 'reward_repeat_floor_std': 0.022863373160362244, 'reward_near_duplicate_penalty_mean': 0.9355204105377197, 'reward_near_duplicate_penalty_std': 0.06568436324596405, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9595141410827637, 'reward_distinct_2_std': 0.08656234294176102, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.2655990719795227, 'reward_total_composite_mean': 0.38056474924087524, 'reward_total_composite_std': 0.4107729196548462, 'epoch': 0.02}

 14%|█▍        | 469/3300 [1:42:07<7:26:01,  9.45s/it]INFO 04-12 17:40:34 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 470/3300 [1:42:16<7:12:11,  9.16s/it]2026-04-12 17:40:42,627 | INFO | train_grpo_train | metrics_logged mode=train step=470

                                                      
{'loss': 0.0241, 'grad_norm': 13.975543022155762, 'learning_rate': 8.57878787878788e-06, 'num_tokens': 872829.0, 'completions/mean_length': 46.0, 'completions/min_length': 38.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.97574782371521, 'rewards/meter/std': 0.018756739795207977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.28749996423721313, 'rewards/judge_quality/std': 0.13024701178073883, 'rewards/repeat_penalty/mean': 0.7905802726745605, 'rewards/repeat_penalty/std': 0.24757647514343262, 'rewards/repeat_floor/mean': 0.9748726487159729, 'rewards/repeat_floor/std': 0.030745012685656548, 'rewards/near_duplicate_penalty/mean': 0.9211663007736206, 'rewards/near_duplicate_penalty/std': 0.08916570991277695, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8730300664901733, 'rewards/distinct_2/std': 0.1401810348033905, 'rewards/total_composite/mean': 0.27376270294189453, 'rewards/total_composite/std': 0.12548699975013733, 'reward': 0.27376270294189453, 'reward_std': 0.12548701465129852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14351414144039154, 'sampling/sampling_logp_difference/max': 1.2779419422149658, 'sampling/importance_sampling_ratio/min': 0.27861011028289795, 'sampling/importance_sampling_ratio/mean': 1.0189917087554932, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1294040009379387, 'clip_ratio/low_mean': 0.05070318467915058, 'clip_ratio/low_min': 0.05070318467915058, 'clip_ratio/high_mean': 0.07650966197252274, 'clip_ratio/high_max': 0.07650966197252274, 'clip_ratio/region_mean': 0.12721284665167332, 'reward_total_mean': 0.27376270294189453, 'reward_meter_mean': 0.97574782371521, 'reward_meter_std': 0.018756739795207977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7905802726745605, 'reward_repeat_penalty_std': 0.24757647514343262, 'reward_repeat_floor_mean': 0.9748726487159729, 'reward_repeat_floor_std': 0.030745012685656548, 'reward_near_duplicate_penalty_mean': 0.9211663007736206, 'reward_near_duplicate_penalty_std': 0.08916570991277695, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8730300664901733, 'reward_distinct_2_std': 0.1401810348033905, 'reward_judge_quality_mean': 0.28749996423721313, 'reward_judge_quality_std': 0.13024701178073883, 'reward_total_composite_mean': 0.27376270294189453, 'reward_total_composite_std': 0.12548699975013733, 'epoch': 0.02}

 14%|█▍        | 470/3300 [1:42:16<7:12:11,  9.16s/it]INFO 04-12 17:40:42 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 471/3300 [1:42:25<7:12:21,  9.17s/it]2026-04-12 17:40:51,816 | INFO | train_grpo_train | metrics_logged mode=train step=471

                                                      
{'loss': 0.0169, 'grad_norm': 8.806530952453613, 'learning_rate': 8.575757575757575e-06, 'num_tokens': 875072.0, 'completions/mean_length': 85.375, 'completions/min_length': 47.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.7614915370941162, 'rewards/meter/std': 0.3527231812477112, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5375000238418579, 'rewards/judge_quality/std': 0.2462722212076187, 'rewards/repeat_penalty/mean': 0.9216973781585693, 'rewards/repeat_penalty/std': 0.16981077194213867, 'rewards/repeat_floor/mean': 0.9922190308570862, 'rewards/repeat_floor/std': 0.01515916083008051, 'rewards/near_duplicate_penalty/mean': 0.9779188632965088, 'rewards/near_duplicate_penalty/std': 0.026434287428855896, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9503462910652161, 'rewards/distinct_2/std': 0.125265434384346, 'rewards/total_composite/mean': 0.40381038188934326, 'rewards/total_composite/std': 0.29029467701911926, 'reward': 0.40381038188934326, 'reward_std': 0.29029467701911926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15804153680801392, 'sampling/sampling_logp_difference/max': 1.5367183685302734, 'sampling/importance_sampling_ratio/min': 0.21508577466011047, 'sampling/importance_sampling_ratio/mean': 1.0180410146713257, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3140480741858482, 'clip_ratio/low_mean': 0.09662595391273499, 'clip_ratio/low_min': 0.09662595391273499, 'clip_ratio/high_mean': 0.06673349719494581, 'clip_ratio/high_max': 0.06673349719494581, 'clip_ratio/region_mean': 0.1633594511076808, 'reward_total_mean': 0.40381038188934326, 'reward_meter_mean': 0.7614915370941162, 'reward_meter_std': 0.3527231812477112, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9216973781585693, 'reward_repeat_penalty_std': 0.16981077194213867, 'reward_repeat_floor_mean': 0.9922190308570862, 'reward_repeat_floor_std': 0.01515916083008051, 'reward_near_duplicate_penalty_mean': 0.9779188632965088, 'reward_near_duplicate_penalty_std': 0.026434287428855896, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9503462910652161, 'reward_distinct_2_std': 0.125265434384346, 'reward_judge_quality_mean': 0.5375000238418579, 'reward_judge_quality_std': 0.2462722212076187, 'reward_total_composite_mean': 0.40381038188934326, 'reward_total_composite_std': 0.29029467701911926, 'epoch': 0.02}

 14%|█▍        | 471/3300 [1:42:25<7:12:21,  9.17s/it]INFO 04-12 17:40:52 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 472/3300 [1:42:39<8:17:10, 10.55s/it]2026-04-12 17:41:05,576 | INFO | train_grpo_train | metrics_logged mode=train step=472

                                                      
{'loss': -0.0261, 'grad_norm': 4.894377708435059, 'learning_rate': 8.572727272727274e-06, 'num_tokens': 876404.0, 'completions/mean_length': 81.5, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6767353415489197, 'rewards/meter/std': 0.4195828437805176, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.932405412197113, 'rewards/lexical_plausibility/std': 0.16736763715744019, 'rewards/judge_quality/mean': 0.5800000429153442, 'rewards/judge_quality/std': 0.3921734392642975, 'rewards/repeat_penalty/mean': 0.6140526533126831, 'rewards/repeat_penalty/std': 0.2528451979160309, 'rewards/repeat_floor/mean': 0.9682778120040894, 'rewards/repeat_floor/std': 0.027230918407440186, 'rewards/near_duplicate_penalty/mean': 0.9653545022010803, 'rewards/near_duplicate_penalty/std': 0.044731080532073975, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.9344406127929688, 'rewards/distinct_2/std': 0.09809565544128418, 'rewards/total_composite/mean': 0.41457605361938477, 'rewards/total_composite/std': 0.39520126581192017, 'reward': 0.41457605361938477, 'reward_std': 0.3952012062072754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19121161103248596, 'sampling/sampling_logp_difference/max': 1.447162389755249, 'sampling/importance_sampling_ratio/min': 0.23523686826229095, 'sampling/importance_sampling_ratio/mean': 1.0249830484390259, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1275591030716896, 'clip_ratio/low_mean': 0.08854811498895288, 'clip_ratio/low_min': 0.08854811498895288, 'clip_ratio/high_mean': 0.018750000279396772, 'clip_ratio/high_max': 0.018750000279396772, 'clip_ratio/region_mean': 0.10729811526834965, 'reward_total_mean': 0.41457605361938477, 'reward_meter_mean': 0.6767353415489197, 'reward_meter_std': 0.4195828437805176, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.932405412197113, 'reward_lexical_plausibility_std': 0.16736763715744019, 'reward_repeat_penalty_mean': 0.6140526533126831, 'reward_repeat_penalty_std': 0.2528451979160309, 'reward_repeat_floor_mean': 0.9682778120040894, 'reward_repeat_floor_std': 0.027230918407440186, 'reward_near_duplicate_penalty_mean': 0.9653545022010803, 'reward_near_duplicate_penalty_std': 0.044731080532073975, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.9344406127929688, 'reward_distinct_2_std': 0.09809565544128418, 'reward_judge_quality_mean': 0.5800000429153442, 'reward_judge_quality_std': 0.3921734392642975, 'reward_total_composite_mean': 0.41457605361938477, 'reward_total_composite_std': 0.39520126581192017, 'epoch': 0.02}

 14%|█▍        | 472/3300 [1:42:39<8:17:10, 10.55s/it]INFO 04-12 17:41:05 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 473/3300 [1:42:47<7:49:22,  9.96s/it]2026-04-12 17:41:14,173 | INFO | train_grpo_train | metrics_logged mode=train step=473

                                                      
{'loss': 0.0415, 'grad_norm': 15.320552825927734, 'learning_rate': 8.56969696969697e-06, 'num_tokens': 877964.0, 'completions/mean_length': 37.0, 'completions/min_length': 30.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.3447434604167938, 'rewards/meter/std': 0.31754711270332336, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.9638725519180298, 'rewards/repeat_penalty/std': 0.0356743298470974, 'rewards/repeat_floor/mean': 0.994580864906311, 'rewards/repeat_floor/std': 0.005351153668016195, 'rewards/near_duplicate_penalty/mean': 0.9638725519180298, 'rewards/near_duplicate_penalty/std': 0.0356743298470974, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21917980909347534, 'rewards/total_composite/std': 0.26487240195274353, 'reward': 0.21917980909347534, 'reward_std': 0.26487240195274353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1716521829366684, 'sampling/sampling_logp_difference/max': 2.157565116882324, 'sampling/importance_sampling_ratio/min': 0.11560626327991486, 'sampling/importance_sampling_ratio/mean': 1.0430837869644165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9588679075241089, 'clip_ratio/low_mean': 0.12824940914288163, 'clip_ratio/low_min': 0.12824940914288163, 'clip_ratio/high_mean': 0.022857625503093004, 'clip_ratio/high_max': 0.022857625503093004, 'clip_ratio/region_mean': 0.15110703464597464, 'reward_total_mean': 0.21917980909347534, 'reward_meter_mean': 0.3447434604167938, 'reward_meter_std': 0.31754711270332336, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9638725519180298, 'reward_repeat_penalty_std': 0.0356743298470974, 'reward_repeat_floor_mean': 0.994580864906311, 'reward_repeat_floor_std': 0.005351153668016195, 'reward_near_duplicate_penalty_mean': 0.9638725519180298, 'reward_near_duplicate_penalty_std': 0.0356743298470974, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.21917980909347534, 'reward_total_composite_std': 0.26487240195274353, 'epoch': 0.02}

 14%|█▍        | 473/3300 [1:42:48<7:49:22,  9.96s/it]INFO 04-12 17:41:14 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 474/3300 [1:43:02<8:48:48, 11.23s/it]2026-04-12 17:41:28,354 | INFO | train_grpo_train | metrics_logged mode=train step=474

                                                      
{'loss': -0.0492, 'grad_norm': 3.5581774711608887, 'learning_rate': 8.566666666666667e-06, 'num_tokens': 879673.0, 'completions/mean_length': 184.625, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.7618043422698975, 'rewards/meter/std': 0.351188063621521, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9318425059318542, 'rewards/lexical_plausibility/std': 0.1288023144006729, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.7190311551094055, 'rewards/repeat_penalty/std': 0.3231569230556488, 'rewards/repeat_floor/mean': 0.9632654786109924, 'rewards/repeat_floor/std': 0.05032624676823616, 'rewards/near_duplicate_penalty/mean': 0.8977975845336914, 'rewards/near_duplicate_penalty/std': 0.13175007700920105, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.18898223340511322, 'rewards/distinct_2/mean': 0.8455091118812561, 'rewards/distinct_2/std': 0.22642596065998077, 'rewards/total_composite/mean': 0.16438791155815125, 'rewards/total_composite/std': 0.11400491744279861, 'reward': 0.16438791155815125, 'reward_std': 0.11400491744279861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14609570801258087, 'sampling/sampling_logp_difference/max': 1.4534201622009277, 'sampling/importance_sampling_ratio/min': 0.23376940190792084, 'sampling/importance_sampling_ratio/mean': 1.0319956541061401, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9716388285160065, 'clip_ratio/low_mean': 0.0790879875421524, 'clip_ratio/low_min': 0.0790879875421524, 'clip_ratio/high_mean': 0.027110389433801174, 'clip_ratio/high_max': 0.027110389433801174, 'clip_ratio/region_mean': 0.10619837697595358, 'reward_total_mean': 0.16438791155815125, 'reward_meter_mean': 0.7618043422698975, 'reward_meter_std': 0.351188063621521, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9318425059318542, 'reward_lexical_plausibility_std': 0.1288023144006729, 'reward_repeat_penalty_mean': 0.7190311551094055, 'reward_repeat_penalty_std': 0.3231569230556488, 'reward_repeat_floor_mean': 0.9632654786109924, 'reward_repeat_floor_std': 0.05032624676823616, 'reward_near_duplicate_penalty_mean': 0.8977975845336914, 'reward_near_duplicate_penalty_std': 0.13175007700920105, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.18898223340511322, 'reward_distinct_2_mean': 0.8455091118812561, 'reward_distinct_2_std': 0.22642596065998077, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.16438791155815125, 'reward_total_composite_std': 0.11400491744279861, 'epoch': 0.02}

 14%|█▍        | 474/3300 [1:43:02<8:48:48, 11.23s/it]INFO 04-12 17:41:28 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 475/3300 [1:43:10<8:13:44, 10.49s/it]2026-04-12 17:41:37,110 | INFO | train_grpo_train | metrics_logged mode=train step=475

                                                      
{'loss': -0.0316, 'grad_norm': 20.224674224853516, 'learning_rate': 8.563636363636364e-06, 'num_tokens': 881236.0, 'completions/mean_length': 36.375, 'completions/min_length': 30.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9836364984512329, 'rewards/meter/std': 0.011494634672999382, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8443509936332703, 'rewards/repeat_penalty/std': 0.16900120675563812, 'rewards/repeat_floor/mean': 0.981834888458252, 'rewards/repeat_floor/std': 0.018487006425857544, 'rewards/near_duplicate_penalty/mean': 0.9215735197067261, 'rewards/near_duplicate_penalty/std': 0.08376346528530121, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9705433249473572, 'rewards/distinct_2/std': 0.04079176485538483, 'rewards/total_composite/mean': 0.3985815644264221, 'rewards/total_composite/std': 0.05256590247154236, 'reward': 0.3985815644264221, 'reward_std': 0.05256590247154236, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15195895731449127, 'sampling/sampling_logp_difference/max': 1.3622798919677734, 'sampling/importance_sampling_ratio/min': 0.25607630610466003, 'sampling/importance_sampling_ratio/mean': 1.0402600765228271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2950242161750793, 'clip_ratio/low_mean': 0.03435457614250481, 'clip_ratio/low_min': 0.03435457614250481, 'clip_ratio/high_mean': 0.08515602489933372, 'clip_ratio/high_max': 0.08515602489933372, 'clip_ratio/region_mean': 0.11951060104183853, 'reward_total_mean': 0.3985815644264221, 'reward_meter_mean': 0.9836364984512329, 'reward_meter_std': 0.011494634672999382, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8443509936332703, 'reward_repeat_penalty_std': 0.16900120675563812, 'reward_repeat_floor_mean': 0.981834888458252, 'reward_repeat_floor_std': 0.018487006425857544, 'reward_near_duplicate_penalty_mean': 0.9215735197067261, 'reward_near_duplicate_penalty_std': 0.08376346528530121, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9705433249473572, 'reward_distinct_2_std': 0.04079176485538483, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3985815644264221, 'reward_total_composite_std': 0.05256590247154236, 'epoch': 0.02}

 14%|█▍        | 475/3300 [1:43:10<8:13:44, 10.49s/it]INFO 04-12 17:41:37 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 476/3300 [1:43:24<8:58:27, 11.44s/it]2026-04-12 17:41:50,774 | INFO | train_grpo_train | metrics_logged mode=train step=476

                                                      
{'loss': 0.1432, 'grad_norm': 8.356103897094727, 'learning_rate': 8.560606060606062e-06, 'num_tokens': 884614.0, 'completions/mean_length': 215.25, 'completions/min_length': 158.0, 'completions/max_length': 366.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 215.25, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 366.0, 'rewards/meter/mean': 0.5926367044448853, 'rewards/meter/std': 0.2863155007362366, 'rewards/count_adherence/mean': 0.6666666269302368, 'rewards/count_adherence/std': 0.24487698078155518, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.073463074862957, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.6616848111152649, 'rewards/repeat_penalty/std': 0.28253722190856934, 'rewards/repeat_floor/mean': 0.9634898900985718, 'rewards/repeat_floor/std': 0.034645069390535355, 'rewards/near_duplicate_penalty/mean': 0.9370911717414856, 'rewards/near_duplicate_penalty/std': 0.06509105116128922, 'rewards/opening_diversity/mean': 0.9982142448425293, 'rewards/opening_diversity/std': 0.00505077326670289, 'rewards/distinct_2/mean': 0.7003699541091919, 'rewards/distinct_2/std': 0.27423763275146484, 'rewards/total_composite/mean': 0.17490671575069427, 'rewards/total_composite/std': 0.1152539923787117, 'reward': 0.17490671575069427, 'reward_std': 0.1152539849281311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16131575405597687, 'sampling/sampling_logp_difference/max': 2.7115721702575684, 'sampling/importance_sampling_ratio/min': 0.06643228232860565, 'sampling/importance_sampling_ratio/mean': 1.012266755104065, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9416213482618332, 'clip_ratio/low_mean': 0.07172723580151796, 'clip_ratio/low_min': 0.07172723580151796, 'clip_ratio/high_mean': 0.09089010022580624, 'clip_ratio/high_max': 0.09089010022580624, 'clip_ratio/region_mean': 0.1626173360273242, 'reward_total_mean': 0.17490671575069427, 'reward_meter_mean': 0.5926367044448853, 'reward_meter_std': 0.2863155007362366, 'reward_count_adherence_mean': 0.6666666269302368, 'reward_count_adherence_std': 0.24487698078155518, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.073463074862957, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6616848111152649, 'reward_repeat_penalty_std': 0.28253722190856934, 'reward_repeat_floor_mean': 0.9634898900985718, 'reward_repeat_floor_std': 0.034645069390535355, 'reward_near_duplicate_penalty_mean': 0.9370911717414856, 'reward_near_duplicate_penalty_std': 0.06509105116128922, 'reward_opening_diversity_mean': 0.9982142448425293, 'reward_opening_diversity_std': 0.00505077326670289, 'reward_distinct_2_mean': 0.7003699541091919, 'reward_distinct_2_std': 0.27423763275146484, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.17490671575069427, 'reward_total_composite_std': 0.1152539923787117, 'epoch': 0.03}

 14%|█▍        | 476/3300 [1:43:24<8:58:27, 11.44s/it]INFO 04-12 17:41:51 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 477/3300 [1:43:33<8:25:28, 10.74s/it]2026-04-12 17:41:59,891 | INFO | train_grpo_train | metrics_logged mode=train step=477

                                                      
{'loss': 0.1351, 'grad_norm': 18.758481979370117, 'learning_rate': 8.557575757575757e-06, 'num_tokens': 886249.0, 'completions/mean_length': 35.375, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5979759693145752, 'rewards/meter/std': 0.3991454541683197, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9858666062355042, 'rewards/repeat_penalty/std': 0.013685435988008976, 'rewards/repeat_floor/mean': 0.9978799819946289, 'rewards/repeat_floor/std': 0.002052822383120656, 'rewards/near_duplicate_penalty/mean': 0.9858666062355042, 'rewards/near_duplicate_penalty/std': 0.013685435988008976, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33318328857421875, 'rewards/total_composite/std': 0.3166455924510956, 'reward': 0.33318328857421875, 'reward_std': 0.3166455924510956, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1638694852590561, 'sampling/sampling_logp_difference/max': 1.8494634628295898, 'sampling/importance_sampling_ratio/min': 0.15732155740261078, 'sampling/importance_sampling_ratio/mean': 0.9977566003799438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9550353959202766, 'clip_ratio/low_mean': 0.0646360320970416, 'clip_ratio/low_min': 0.0646360320970416, 'clip_ratio/high_mean': 0.08017908968031406, 'clip_ratio/high_max': 0.08017908968031406, 'clip_ratio/region_mean': 0.14481512177735567, 'reward_total_mean': 0.33318328857421875, 'reward_meter_mean': 0.5979759693145752, 'reward_meter_std': 0.3991454541683197, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9858666062355042, 'reward_repeat_penalty_std': 0.013685435988008976, 'reward_repeat_floor_mean': 0.9978799819946289, 'reward_repeat_floor_std': 0.002052822383120656, 'reward_near_duplicate_penalty_mean': 0.9858666062355042, 'reward_near_duplicate_penalty_std': 0.013685435988008976, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.33318328857421875, 'reward_total_composite_std': 0.3166455924510956, 'epoch': 0.03}

 14%|█▍        | 477/3300 [1:43:33<8:25:28, 10.74s/it]INFO 04-12 17:42:00 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 478/3300 [1:43:43<8:14:46, 10.52s/it]2026-04-12 17:42:09,889 | INFO | train_grpo_train | metrics_logged mode=train step=478

                                                      
{'loss': 0.0528, 'grad_norm': 8.750658988952637, 'learning_rate': 8.554545454545456e-06, 'num_tokens': 888415.0, 'completions/mean_length': 87.75, 'completions/min_length': 79.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.75, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.8960630297660828, 'rewards/meter/std': 0.1679113656282425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9716117978096008, 'rewards/lexical_plausibility/std': 0.08029396086931229, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8216412663459778, 'rewards/repeat_penalty/std': 0.30016782879829407, 'rewards/repeat_floor/mean': 0.9763308167457581, 'rewards/repeat_floor/std': 0.0427468977868557, 'rewards/near_duplicate_penalty/mean': 0.9364006519317627, 'rewards/near_duplicate_penalty/std': 0.08619730919599533, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1530931144952774, 'rewards/distinct_2/mean': 0.8846746683120728, 'rewards/distinct_2/std': 0.23017317056655884, 'rewards/total_composite/mean': 0.3119262158870697, 'rewards/total_composite/std': 0.08966439217329025, 'reward': 0.3119262158870697, 'reward_std': 0.08966439217329025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13232532143592834, 'sampling/sampling_logp_difference/max': 1.586812973022461, 'sampling/importance_sampling_ratio/min': 0.20457656681537628, 'sampling/importance_sampling_ratio/mean': 1.024656057357788, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0769718065857887, 'clip_ratio/low_mean': 0.053361023776233196, 'clip_ratio/low_min': 0.053361023776233196, 'clip_ratio/high_mean': 0.10445518139749765, 'clip_ratio/high_max': 0.10445518139749765, 'clip_ratio/region_mean': 0.15781620517373085, 'reward_total_mean': 0.3119262158870697, 'reward_meter_mean': 0.8960630297660828, 'reward_meter_std': 0.1679113656282425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9716117978096008, 'reward_lexical_plausibility_std': 0.08029396086931229, 'reward_repeat_penalty_mean': 0.8216412663459778, 'reward_repeat_penalty_std': 0.30016782879829407, 'reward_repeat_floor_mean': 0.9763308167457581, 'reward_repeat_floor_std': 0.0427468977868557, 'reward_near_duplicate_penalty_mean': 0.9364006519317627, 'reward_near_duplicate_penalty_std': 0.08619730919599533, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1530931144952774, 'reward_distinct_2_mean': 0.8846746683120728, 'reward_distinct_2_std': 0.23017317056655884, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.3119262158870697, 'reward_total_composite_std': 0.08966439217329025, 'epoch': 0.03}

 14%|█▍        | 478/3300 [1:43:43<8:14:46, 10.52s/it]INFO 04-12 17:42:10 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 479/3300 [1:43:51<7:35:52,  9.70s/it]2026-04-12 17:42:17,672 | INFO | train_grpo_train | metrics_logged mode=train step=479

                                                      
{'loss': 0.0898, 'grad_norm': 29.00151252746582, 'learning_rate': 8.551515151515152e-06, 'num_tokens': 889883.0, 'completions/mean_length': 22.5, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9142677783966064, 'rewards/meter/std': 0.12872745096683502, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9505208134651184, 'rewards/lexical_plausibility/std': 0.07547593116760254, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.3620477616786957, 'rewards/repeat_penalty/mean': 0.7300546169281006, 'rewards/repeat_penalty/std': 0.04901907593011856, 'rewards/repeat_floor/mean': 0.9819168448448181, 'rewards/repeat_floor/std': 0.007275138981640339, 'rewards/near_duplicate_penalty/mean': 0.989935040473938, 'rewards/near_duplicate_penalty/std': 0.02014034427702427, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.5911568403244019, 'rewards/total_composite/std': 0.3244568705558777, 'reward': 0.5911568403244019, 'reward_std': 0.3244568705558777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16883940994739532, 'sampling/sampling_logp_difference/max': 1.9402134418487549, 'sampling/importance_sampling_ratio/min': 0.14367328584194183, 'sampling/importance_sampling_ratio/mean': 1.0014277696609497, 'sampling/importance_sampling_ratio/max': 1.8156558275222778, 'entropy': 1.0974134728312492, 'clip_ratio/low_mean': 0.03192640794441104, 'clip_ratio/low_min': 0.03192640794441104, 'clip_ratio/high_mean': 0.09963768254965544, 'clip_ratio/high_max': 0.09963768254965544, 'clip_ratio/region_mean': 0.13156409049406648, 'reward_total_mean': 0.5911568403244019, 'reward_meter_mean': 0.9142677783966064, 'reward_meter_std': 0.12872745096683502, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9505208134651184, 'reward_lexical_plausibility_std': 0.07547593116760254, 'reward_repeat_penalty_mean': 0.7300546169281006, 'reward_repeat_penalty_std': 0.04901907593011856, 'reward_repeat_floor_mean': 0.9819168448448181, 'reward_repeat_floor_std': 0.007275138981640339, 'reward_near_duplicate_penalty_mean': 0.989935040473938, 'reward_near_duplicate_penalty_std': 0.02014034427702427, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.3620477616786957, 'reward_total_composite_mean': 0.5911568403244019, 'reward_total_composite_std': 0.3244568705558777, 'epoch': 0.03}

 15%|█▍        | 479/3300 [1:43:51<7:35:52,  9.70s/it]INFO 04-12 17:42:17 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 480/3300 [1:43:59<7:11:49,  9.19s/it]2026-04-12 17:42:25,663 | INFO | train_grpo_train | metrics_logged mode=train step=480

                                                      
{'loss': 0.1218, 'grad_norm': 15.980237007141113, 'learning_rate': 8.548484848484849e-06, 'num_tokens': 891327.0, 'completions/mean_length': 26.5, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8697779178619385, 'rewards/meter/std': 0.2896170914173126, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9217138290405273, 'rewards/lexical_plausibility/std': 0.12752297520637512, 'rewards/judge_quality/mean': 0.7937500476837158, 'rewards/judge_quality/std': 0.2447119504213333, 'rewards/repeat_penalty/mean': 0.7800381183624268, 'rewards/repeat_penalty/std': 0.08749208599328995, 'rewards/repeat_floor/mean': 0.98665452003479, 'rewards/repeat_floor/std': 0.005198613274842501, 'rewards/near_duplicate_penalty/mean': 0.998529851436615, 'rewards/near_duplicate_penalty/std': 0.00300564244389534, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.658567488193512, 'rewards/total_composite/std': 0.3341985046863556, 'reward': 0.658567488193512, 'reward_std': 0.3341985046863556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1702299267053604, 'sampling/sampling_logp_difference/max': 2.050219774246216, 'sampling/importance_sampling_ratio/min': 0.12870661914348602, 'sampling/importance_sampling_ratio/mean': 1.0064083337783813, 'sampling/importance_sampling_ratio/max': 1.9248806238174438, 'entropy': 1.475433349609375, 'clip_ratio/low_mean': 0.08186069503426552, 'clip_ratio/low_min': 0.08186069503426552, 'clip_ratio/high_mean': 0.15122710820287466, 'clip_ratio/high_max': 0.15122710820287466, 'clip_ratio/region_mean': 0.23308780323714018, 'reward_total_mean': 0.658567488193512, 'reward_meter_mean': 0.8697779178619385, 'reward_meter_std': 0.2896170914173126, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9217138290405273, 'reward_lexical_plausibility_std': 0.12752297520637512, 'reward_repeat_penalty_mean': 0.7800381183624268, 'reward_repeat_penalty_std': 0.08749208599328995, 'reward_repeat_floor_mean': 0.98665452003479, 'reward_repeat_floor_std': 0.005198613274842501, 'reward_near_duplicate_penalty_mean': 0.998529851436615, 'reward_near_duplicate_penalty_std': 0.00300564244389534, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7937500476837158, 'reward_judge_quality_std': 0.2447119504213333, 'reward_total_composite_mean': 0.658567488193512, 'reward_total_composite_std': 0.3341985046863556, 'epoch': 0.03}

 15%|█▍        | 480/3300 [1:43:59<7:11:49,  9.19s/it]INFO 04-12 17:42:25 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 481/3300 [1:44:07<6:54:09,  8.82s/it]2026-04-12 17:42:33,610 | INFO | train_grpo_train | metrics_logged mode=train step=481

                                                      
{'loss': 0.0099, 'grad_norm': 10.96861457824707, 'learning_rate': 8.545454545454546e-06, 'num_tokens': 892900.0, 'completions/mean_length': 46.625, 'completions/min_length': 42.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.625, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8064861297607422, 'rewards/meter/std': 0.31927233934402466, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.8413097858428955, 'rewards/repeat_penalty/std': 0.14497357606887817, 'rewards/repeat_floor/mean': 0.9844622611999512, 'rewards/repeat_floor/std': 0.01077551580965519, 'rewards/near_duplicate_penalty/mean': 0.9492339491844177, 'rewards/near_duplicate_penalty/std': 0.02526203729212284, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.974468469619751, 'rewards/distinct_2/std': 0.035326093435287476, 'rewards/total_composite/mean': 0.43128859996795654, 'rewards/total_composite/std': 0.17588578164577484, 'reward': 0.43128859996795654, 'reward_std': 0.17588576674461365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1625947207212448, 'sampling/sampling_logp_difference/max': 1.3164739608764648, 'sampling/importance_sampling_ratio/min': 0.26807889342308044, 'sampling/importance_sampling_ratio/mean': 1.0100953578948975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3365557566285133, 'clip_ratio/low_mean': 0.044546227902173996, 'clip_ratio/low_min': 0.044546227902173996, 'clip_ratio/high_mean': 0.09171762876212597, 'clip_ratio/high_max': 0.09171762876212597, 'clip_ratio/region_mean': 0.13626385666429996, 'reward_total_mean': 0.43128859996795654, 'reward_meter_mean': 0.8064861297607422, 'reward_meter_std': 0.31927233934402466, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8413097858428955, 'reward_repeat_penalty_std': 0.14497357606887817, 'reward_repeat_floor_mean': 0.9844622611999512, 'reward_repeat_floor_std': 0.01077551580965519, 'reward_near_duplicate_penalty_mean': 0.9492339491844177, 'reward_near_duplicate_penalty_std': 0.02526203729212284, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.974468469619751, 'reward_distinct_2_std': 0.035326093435287476, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.43128859996795654, 'reward_total_composite_std': 0.17588578164577484, 'epoch': 0.03}

 15%|█▍        | 481/3300 [1:44:07<6:54:09,  8.82s/it]INFO 04-12 17:42:33 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 482/3300 [1:44:21<8:11:02, 10.46s/it]2026-04-12 17:42:47,895 | INFO | train_grpo_train | metrics_logged mode=train step=482

                                                      
{'loss': -0.0776, 'grad_norm': 4.801069259643555, 'learning_rate': 8.542424242424243e-06, 'num_tokens': 894426.0, 'completions/mean_length': 105.75, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.71428680419922, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7329697608947754, 'rewards/meter/std': 0.41056057810783386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9211018085479736, 'rewards/lexical_plausibility/std': 0.2231578826904297, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.2945577800273895, 'rewards/repeat_penalty/mean': 0.9583426713943481, 'rewards/repeat_penalty/std': 0.10578048974275589, 'rewards/repeat_floor/mean': 0.9967557191848755, 'rewards/repeat_floor/std': 0.007397820241749287, 'rewards/near_duplicate_penalty/mean': 0.9961162209510803, 'rewards/near_duplicate_penalty/std': 0.006006431765854359, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.4038626253604889, 'rewards/total_composite/std': 0.29983022809028625, 'reward': 0.4038626253604889, 'reward_std': 0.29983022809028625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15785160660743713, 'sampling/sampling_logp_difference/max': 1.3366641998291016, 'sampling/importance_sampling_ratio/min': 0.26272058486938477, 'sampling/importance_sampling_ratio/mean': 1.0346943140029907, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1752451807260513, 'clip_ratio/low_mean': 0.0785757452249527, 'clip_ratio/low_min': 0.0785757452249527, 'clip_ratio/high_mean': 0.07952513638883829, 'clip_ratio/high_max': 0.07952513638883829, 'clip_ratio/region_mean': 0.158100881613791, 'reward_total_mean': 0.4038626253604889, 'reward_meter_mean': 0.7329697608947754, 'reward_meter_std': 0.41056057810783386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9211018085479736, 'reward_lexical_plausibility_std': 0.2231578826904297, 'reward_repeat_penalty_mean': 0.9583426713943481, 'reward_repeat_penalty_std': 0.10578048974275589, 'reward_repeat_floor_mean': 0.9967557191848755, 'reward_repeat_floor_std': 0.007397820241749287, 'reward_near_duplicate_penalty_mean': 0.9961162209510803, 'reward_near_duplicate_penalty_std': 0.006006431765854359, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.2945577800273895, 'reward_total_composite_mean': 0.4038626253604889, 'reward_total_composite_std': 0.29983022809028625, 'epoch': 0.03}

 15%|█▍        | 482/3300 [1:44:21<8:11:02, 10.46s/it]INFO 04-12 17:42:48 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 483/3300 [1:44:29<7:33:27,  9.66s/it]2026-04-12 17:42:55,695 | INFO | train_grpo_train | metrics_logged mode=train step=483

                                                      
{'loss': 0.0472, 'grad_norm': 13.090096473693848, 'learning_rate': 8.539393939393939e-06, 'num_tokens': 896071.0, 'completions/mean_length': 42.625, 'completions/min_length': 41.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.484230637550354, 'rewards/meter/std': 0.32689040899276733, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390502154827118, 'rewards/repeat_penalty/mean': 0.8828092813491821, 'rewards/repeat_penalty/std': 0.24599142372608185, 'rewards/repeat_floor/mean': 0.9882823824882507, 'rewards/repeat_floor/std': 0.02364494651556015, 'rewards/near_duplicate_penalty/mean': 0.9690269231796265, 'rewards/near_duplicate_penalty/std': 0.05978485941886902, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9422594904899597, 'rewards/distinct_2/std': 0.10685522109270096, 'rewards/total_composite/mean': 0.22645294666290283, 'rewards/total_composite/std': 0.14072291553020477, 'reward': 0.22645294666290283, 'reward_std': 0.14072293043136597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17176489531993866, 'sampling/sampling_logp_difference/max': 1.9394941329956055, 'sampling/importance_sampling_ratio/min': 0.14377665519714355, 'sampling/importance_sampling_ratio/mean': 1.0345059633255005, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5112168416380882, 'clip_ratio/low_mean': 0.08229451440274715, 'clip_ratio/low_min': 0.08229451440274715, 'clip_ratio/high_mean': 0.08154318574815989, 'clip_ratio/high_max': 0.08154318574815989, 'clip_ratio/region_mean': 0.16383770015090704, 'reward_total_mean': 0.22645294666290283, 'reward_meter_mean': 0.484230637550354, 'reward_meter_std': 0.32689040899276733, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8828092813491821, 'reward_repeat_penalty_std': 0.24599142372608185, 'reward_repeat_floor_mean': 0.9882823824882507, 'reward_repeat_floor_std': 0.02364494651556015, 'reward_near_duplicate_penalty_mean': 0.9690269231796265, 'reward_near_duplicate_penalty_std': 0.05978485941886902, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9422594904899597, 'reward_distinct_2_std': 0.10685522109270096, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390502154827118, 'reward_total_composite_mean': 0.22645294666290283, 'reward_total_composite_std': 0.14072291553020477, 'epoch': 0.03}

 15%|█▍        | 483/3300 [1:44:29<7:33:27,  9.66s/it]INFO 04-12 17:42:55 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 484/3300 [1:44:37<7:06:25,  9.09s/it]2026-04-12 17:43:03,452 | INFO | train_grpo_train | metrics_logged mode=train step=484

                                                      
{'loss': 0.0005, 'grad_norm': 27.089628219604492, 'learning_rate': 8.536363636363636e-06, 'num_tokens': 897664.0, 'completions/mean_length': 18.125, 'completions/min_length': 15.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.125, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.9493619203567505, 'rewards/meter/std': 0.09226961433887482, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.831250011920929, 'rewards/judge_quality/std': 0.27559998631477356, 'rewards/repeat_penalty/mean': 0.7269262671470642, 'rewards/repeat_penalty/std': 0.06526235491037369, 'rewards/repeat_floor/mean': 0.9803853034973145, 'rewards/repeat_floor/std': 0.013052479363977909, 'rewards/near_duplicate_penalty/mean': 0.9692350625991821, 'rewards/near_duplicate_penalty/std': 0.08701647073030472, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7708688378334045, 'rewards/total_composite/std': 0.267892986536026, 'reward': 0.7708688378334045, 'reward_std': 0.267892986536026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16559292376041412, 'sampling/sampling_logp_difference/max': 1.049661636352539, 'sampling/importance_sampling_ratio/min': 0.35005617141723633, 'sampling/importance_sampling_ratio/mean': 1.027931809425354, 'sampling/importance_sampling_ratio/max': 1.7970988750457764, 'entropy': 1.1931760981678963, 'clip_ratio/low_mean': 0.034722222946584225, 'clip_ratio/low_min': 0.034722222946584225, 'clip_ratio/high_mean': 0.12408963870257139, 'clip_ratio/high_max': 0.12408963870257139, 'clip_ratio/region_mean': 0.15881186164915562, 'reward_total_mean': 0.7708688378334045, 'reward_meter_mean': 0.9493619203567505, 'reward_meter_std': 0.09226961433887482, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7269262671470642, 'reward_repeat_penalty_std': 0.06526235491037369, 'reward_repeat_floor_mean': 0.9803853034973145, 'reward_repeat_floor_std': 0.013052479363977909, 'reward_near_duplicate_penalty_mean': 0.9692350625991821, 'reward_near_duplicate_penalty_std': 0.08701647073030472, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.831250011920929, 'reward_judge_quality_std': 0.27559998631477356, 'reward_total_composite_mean': 0.7708688378334045, 'reward_total_composite_std': 0.267892986536026, 'epoch': 0.03}

 15%|█▍        | 484/3300 [1:44:37<7:06:25,  9.09s/it]INFO 04-12 17:43:03 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 485/3300 [1:44:45<6:54:37,  8.84s/it]2026-04-12 17:43:11,700 | INFO | train_grpo_train | metrics_logged mode=train step=485

                                                      
{'loss': 0.0282, 'grad_norm': 27.633771896362305, 'learning_rate': 8.533333333333335e-06, 'num_tokens': 899201.0, 'completions/mean_length': 33.125, 'completions/min_length': 30.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.5026096105575562, 'rewards/meter/std': 0.3751481771469116, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5462499856948853, 'rewards/judge_quality/std': 0.24401040375232697, 'rewards/repeat_penalty/mean': 0.9841090440750122, 'rewards/repeat_penalty/std': 0.016218602657318115, 'rewards/repeat_floor/mean': 0.9976163506507874, 'rewards/repeat_floor/std': 0.0024327824357897043, 'rewards/near_duplicate_penalty/mean': 0.9841090440750122, 'rewards/near_duplicate_penalty/std': 0.016218602657318115, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.26432886719703674, 'rewards/total_composite/std': 0.2809872031211853, 'reward': 0.26432886719703674, 'reward_std': 0.2809872031211853, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18319426476955414, 'sampling/sampling_logp_difference/max': 2.6407713890075684, 'sampling/importance_sampling_ratio/min': 0.0713062435388565, 'sampling/importance_sampling_ratio/mean': 1.0072911977767944, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0676851198077202, 'clip_ratio/low_mean': 0.10398916900157928, 'clip_ratio/low_min': 0.10398916900157928, 'clip_ratio/high_mean': 0.040719698183238506, 'clip_ratio/high_max': 0.040719698183238506, 'clip_ratio/region_mean': 0.1447088671848178, 'reward_total_mean': 0.26432886719703674, 'reward_meter_mean': 0.5026096105575562, 'reward_meter_std': 0.3751481771469116, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9841090440750122, 'reward_repeat_penalty_std': 0.016218602657318115, 'reward_repeat_floor_mean': 0.9976163506507874, 'reward_repeat_floor_std': 0.0024327824357897043, 'reward_near_duplicate_penalty_mean': 0.9841090440750122, 'reward_near_duplicate_penalty_std': 0.016218602657318115, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5462499856948853, 'reward_judge_quality_std': 0.24401040375232697, 'reward_total_composite_mean': 0.26432886719703674, 'reward_total_composite_std': 0.2809872031211853, 'epoch': 0.03}

 15%|█▍        | 485/3300 [1:44:45<6:54:37,  8.84s/it]INFO 04-12 17:43:11 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 486/3300 [1:44:55<7:11:26,  9.20s/it]2026-04-12 17:43:21,745 | INFO | train_grpo_train | metrics_logged mode=train step=486

                                                      
{'loss': -0.0285, 'grad_norm': 13.541166305541992, 'learning_rate': 8.53030303030303e-06, 'num_tokens': 901362.0, 'completions/mean_length': 100.125, 'completions/min_length': 80.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.41853219270706177, 'rewards/meter/std': 0.22888816893100739, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9850000143051147, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.7463910579681396, 'rewards/repeat_penalty/std': 0.16073068976402283, 'rewards/repeat_floor/mean': 0.9715670943260193, 'rewards/repeat_floor/std': 0.0172157883644104, 'rewards/near_duplicate_penalty/mean': 0.9110779166221619, 'rewards/near_duplicate_penalty/std': 0.03621843457221985, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8531154990196228, 'rewards/distinct_2/std': 0.08381636440753937, 'rewards/total_composite/mean': 0.13727928698062897, 'rewards/total_composite/std': 0.0861298069357872, 'reward': 0.13727928698062897, 'reward_std': 0.0861298143863678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21168260276317596, 'sampling/sampling_logp_difference/max': 2.3718762397766113, 'sampling/importance_sampling_ratio/min': 0.09330549836158752, 'sampling/importance_sampling_ratio/mean': 0.9802566170692444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9262030124664307, 'clip_ratio/low_mean': 0.09503144770860672, 'clip_ratio/low_min': 0.09503144770860672, 'clip_ratio/high_mean': 0.07580113597214222, 'clip_ratio/high_max': 0.07580113597214222, 'clip_ratio/region_mean': 0.17083258368074894, 'reward_total_mean': 0.13727928698062897, 'reward_meter_mean': 0.41853219270706177, 'reward_meter_std': 0.22888816893100739, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9850000143051147, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7463910579681396, 'reward_repeat_penalty_std': 0.16073068976402283, 'reward_repeat_floor_mean': 0.9715670943260193, 'reward_repeat_floor_std': 0.0172157883644104, 'reward_near_duplicate_penalty_mean': 0.9110779166221619, 'reward_near_duplicate_penalty_std': 0.03621843457221985, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8531154990196228, 'reward_distinct_2_std': 0.08381636440753937, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.13727928698062897, 'reward_total_composite_std': 0.0861298069357872, 'epoch': 0.03}

 15%|█▍        | 486/3300 [1:44:55<7:11:26,  9.20s/it]INFO 04-12 17:43:21 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 487/3300 [1:45:09<8:18:12, 10.63s/it]2026-04-12 17:43:35,702 | INFO | train_grpo_train | metrics_logged mode=train step=487

                                                      
{'loss': -0.0757, 'grad_norm': 3.5076582431793213, 'learning_rate': 8.527272727272728e-06, 'num_tokens': 904460.0, 'completions/mean_length': 248.25, 'completions/min_length': 193.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 210.57144165039062, 'completions/min_terminated_length': 193.0, 'completions/max_terminated_length': 236.0, 'rewards/meter/mean': 0.7735596895217896, 'rewards/meter/std': 0.3453555107116699, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.1079898476600647, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9571428298950195, 'rewards/count_floor/std': 0.03239695355296135, 'rewards/lexical_plausibility/mean': 0.9403283596038818, 'rewards/lexical_plausibility/std': 0.1687769591808319, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.08345229923725128, 'rewards/repeat_penalty/mean': 0.08856150507926941, 'rewards/repeat_penalty/std': 0.1794922798871994, 'rewards/repeat_floor/mean': 0.8054174184799194, 'rewards/repeat_floor/std': 0.07545901089906693, 'rewards/near_duplicate_penalty/mean': 0.45798158645629883, 'rewards/near_duplicate_penalty/std': 0.214552640914917, 'rewards/opening_diversity/mean': 0.37261903285980225, 'rewards/opening_diversity/std': 0.3214702010154724, 'rewards/distinct_2/mean': 0.15958885848522186, 'rewards/distinct_2/std': 0.28594765067100525, 'rewards/total_composite/mean': 0.10135243833065033, 'rewards/total_composite/std': 0.0846024826169014, 'reward': 0.10135243833065033, 'reward_std': 0.0846024826169014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05849408358335495, 'sampling/sampling_logp_difference/max': 1.6172943115234375, 'sampling/importance_sampling_ratio/min': 0.19843487441539764, 'sampling/importance_sampling_ratio/mean': 1.0120666027069092, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3639492578804493, 'clip_ratio/low_mean': 0.015279501210898161, 'clip_ratio/low_min': 0.015279501210898161, 'clip_ratio/high_mean': 0.034134830348193645, 'clip_ratio/high_max': 0.034134830348193645, 'clip_ratio/region_mean': 0.049414331559091806, 'reward_total_mean': 0.10135243833065033, 'reward_meter_mean': 0.7735596895217896, 'reward_meter_std': 0.3453555107116699, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.1079898476600647, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9571428298950195, 'reward_count_floor_std': 0.03239695355296135, 'reward_lexical_plausibility_mean': 0.9403283596038818, 'reward_lexical_plausibility_std': 0.1687769591808319, 'reward_repeat_penalty_mean': 0.08856150507926941, 'reward_repeat_penalty_std': 0.1794922798871994, 'reward_repeat_floor_mean': 0.8054174184799194, 'reward_repeat_floor_std': 0.07545901089906693, 'reward_near_duplicate_penalty_mean': 0.45798158645629883, 'reward_near_duplicate_penalty_std': 0.214552640914917, 'reward_opening_diversity_mean': 0.37261903285980225, 'reward_opening_diversity_std': 0.3214702010154724, 'reward_distinct_2_mean': 0.15958885848522186, 'reward_distinct_2_std': 0.28594765067100525, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.08345229923725128, 'reward_total_composite_mean': 0.10135243833065033, 'reward_total_composite_std': 0.0846024826169014, 'epoch': 0.03}

 15%|█▍        | 487/3300 [1:45:09<8:18:12, 10.63s/it]INFO 04-12 17:43:35 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 488/3300 [1:45:18<7:53:10, 10.10s/it]2026-04-12 17:43:44,560 | INFO | train_grpo_train | metrics_logged mode=train step=488

                                                      
{'loss': 0.0455, 'grad_norm': 11.663517951965332, 'learning_rate': 8.524242424242425e-06, 'num_tokens': 906185.0, 'completions/mean_length': 39.625, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5654488205909729, 'rewards/meter/std': 0.3592325747013092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.6431543827056885, 'rewards/repeat_penalty/std': 0.289379745721817, 'rewards/repeat_floor/mean': 0.9589046239852905, 'rewards/repeat_floor/std': 0.035241127014160156, 'rewards/near_duplicate_penalty/mean': 0.8935448527336121, 'rewards/near_duplicate_penalty/std': 0.09367447346448898, 'rewards/opening_diversity/mean': 0.796875, 'rewards/opening_diversity/std': 0.2106272578239441, 'rewards/distinct_2/mean': 0.8562270998954773, 'rewards/distinct_2/std': 0.12269478291273117, 'rewards/total_composite/mean': 0.17606037855148315, 'rewards/total_composite/std': 0.14399069547653198, 'reward': 0.17606037855148315, 'reward_std': 0.1439906805753708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1685486137866974, 'sampling/sampling_logp_difference/max': 1.5693578720092773, 'sampling/importance_sampling_ratio/min': 0.2081788182258606, 'sampling/importance_sampling_ratio/mean': 1.0289348363876343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.343580350279808, 'clip_ratio/low_mean': 0.05059830006211996, 'clip_ratio/low_min': 0.05059830006211996, 'clip_ratio/high_mean': 0.08685897663235664, 'clip_ratio/high_max': 0.08685897663235664, 'clip_ratio/region_mean': 0.1374572766944766, 'reward_total_mean': 0.17606037855148315, 'reward_meter_mean': 0.5654488205909729, 'reward_meter_std': 0.3592325747013092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6431543827056885, 'reward_repeat_penalty_std': 0.289379745721817, 'reward_repeat_floor_mean': 0.9589046239852905, 'reward_repeat_floor_std': 0.035241127014160156, 'reward_near_duplicate_penalty_mean': 0.8935448527336121, 'reward_near_duplicate_penalty_std': 0.09367447346448898, 'reward_opening_diversity_mean': 0.796875, 'reward_opening_diversity_std': 0.2106272578239441, 'reward_distinct_2_mean': 0.8562270998954773, 'reward_distinct_2_std': 0.12269478291273117, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.17606037855148315, 'reward_total_composite_std': 0.14399069547653198, 'epoch': 0.03}

 15%|█▍        | 488/3300 [1:45:18<7:53:10, 10.10s/it]INFO 04-12 17:43:44 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 489/3300 [1:45:27<7:37:50,  9.77s/it]2026-04-12 17:43:53,577 | INFO | train_grpo_train | metrics_logged mode=train step=489

                                                      
{'loss': 0.0849, 'grad_norm': 11.821260452270508, 'learning_rate': 8.521212121212123e-06, 'num_tokens': 907973.0, 'completions/mean_length': 43.5, 'completions/min_length': 39.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9157294631004333, 'rewards/meter/std': 0.10061603784561157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.28665250539779663, 'rewards/repeat_penalty/mean': 0.9101136326789856, 'rewards/repeat_penalty/std': 0.21360865235328674, 'rewards/repeat_floor/mean': 0.9878295660018921, 'rewards/repeat_floor/std': 0.028341015800833702, 'rewards/near_duplicate_penalty/mean': 0.9544407725334167, 'rewards/near_duplicate_penalty/std': 0.08910517394542694, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9615384340286255, 'rewards/distinct_2/std': 0.10878565162420273, 'rewards/total_composite/mean': 0.5179541110992432, 'rewards/total_composite/std': 0.2553533911705017, 'reward': 0.5179541110992432, 'reward_std': 0.2553533911705017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1325424611568451, 'sampling/sampling_logp_difference/max': 1.5777745246887207, 'sampling/importance_sampling_ratio/min': 0.20643401145935059, 'sampling/importance_sampling_ratio/mean': 1.0154997110366821, 'sampling/importance_sampling_ratio/max': 1.8528239727020264, 'entropy': 1.0343789383769035, 'clip_ratio/low_mean': 0.08047807216644287, 'clip_ratio/low_min': 0.08047807216644287, 'clip_ratio/high_mean': 0.05606740806251764, 'clip_ratio/high_max': 0.05606740806251764, 'clip_ratio/region_mean': 0.13654548022896051, 'reward_total_mean': 0.5179541110992432, 'reward_meter_mean': 0.9157294631004333, 'reward_meter_std': 0.10061603784561157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9101136326789856, 'reward_repeat_penalty_std': 0.21360865235328674, 'reward_repeat_floor_mean': 0.9878295660018921, 'reward_repeat_floor_std': 0.028341015800833702, 'reward_near_duplicate_penalty_mean': 0.9544407725334167, 'reward_near_duplicate_penalty_std': 0.08910517394542694, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9615384340286255, 'reward_distinct_2_std': 0.10878565162420273, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.28665250539779663, 'reward_total_composite_mean': 0.5179541110992432, 'reward_total_composite_std': 0.2553533911705017, 'epoch': 0.03}

 15%|█▍        | 489/3300 [1:45:27<7:37:50,  9.77s/it]INFO 04-12 17:43:53 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 490/3300 [1:45:36<7:29:05,  9.59s/it]2026-04-12 17:44:02,740 | INFO | train_grpo_train | metrics_logged mode=train step=490

                                                      
{'loss': 0.0485, 'grad_norm': 13.303430557250977, 'learning_rate': 8.518181818181818e-06, 'num_tokens': 909837.0, 'completions/mean_length': 61.0, 'completions/min_length': 54.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9696242213249207, 'rewards/meter/std': 0.028753528371453285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7327563762664795, 'rewards/repeat_penalty/std': 0.16056913137435913, 'rewards/repeat_floor/mean': 0.9657080173492432, 'rewards/repeat_floor/std': 0.020789634436368942, 'rewards/near_duplicate_penalty/mean': 0.8702481985092163, 'rewards/near_duplicate_penalty/std': 0.0724751353263855, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8352308869361877, 'rewards/distinct_2/std': 0.1276673525571823, 'rewards/total_composite/mean': 0.2598779499530792, 'rewards/total_composite/std': 0.13646939396858215, 'reward': 0.2598779499530792, 'reward_std': 0.13646937906742096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1456131488084793, 'sampling/sampling_logp_difference/max': 1.0862550735473633, 'sampling/importance_sampling_ratio/min': 0.33747798204421997, 'sampling/importance_sampling_ratio/mean': 1.037516474723816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1825586631894112, 'clip_ratio/low_mean': 0.06616924051195383, 'clip_ratio/low_min': 0.06616924051195383, 'clip_ratio/high_mean': 0.06505074258893728, 'clip_ratio/high_max': 0.06505074258893728, 'clip_ratio/region_mean': 0.1312199831008911, 'reward_total_mean': 0.2598779499530792, 'reward_meter_mean': 0.9696242213249207, 'reward_meter_std': 0.028753528371453285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7327563762664795, 'reward_repeat_penalty_std': 0.16056913137435913, 'reward_repeat_floor_mean': 0.9657080173492432, 'reward_repeat_floor_std': 0.020789634436368942, 'reward_near_duplicate_penalty_mean': 0.8702481985092163, 'reward_near_duplicate_penalty_std': 0.0724751353263855, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8352308869361877, 'reward_distinct_2_std': 0.1276673525571823, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.2598779499530792, 'reward_total_composite_std': 0.13646939396858215, 'epoch': 0.03}

 15%|█▍        | 490/3300 [1:45:36<7:29:05,  9.59s/it]INFO 04-12 17:44:02 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 491/3300 [1:45:44<7:05:00,  9.08s/it]2026-04-12 17:44:10,627 | INFO | train_grpo_train | metrics_logged mode=train step=491

                                                      
{'loss': 0.047, 'grad_norm': 24.3916015625, 'learning_rate': 8.515151515151517e-06, 'num_tokens': 911340.0, 'completions/mean_length': 19.875, 'completions/min_length': 19.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.8117765188217163, 'rewards/meter/std': 0.23110267519950867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.33683136105537415, 'rewards/repeat_penalty/mean': 0.7433712482452393, 'rewards/repeat_penalty/std': 0.018749047070741653, 'rewards/repeat_floor/mean': 0.9836742877960205, 'rewards/repeat_floor/std': 0.0037498052697628736, 'rewards/near_duplicate_penalty/mean': 0.991161584854126, 'rewards/near_duplicate_penalty/std': 0.024998728185892105, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6282688975334167, 'rewards/total_composite/std': 0.3177328109741211, 'reward': 0.6282688975334167, 'reward_std': 0.3177328109741211, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17767982184886932, 'sampling/sampling_logp_difference/max': 1.1688170433044434, 'sampling/importance_sampling_ratio/min': 0.3107343018054962, 'sampling/importance_sampling_ratio/mean': 1.010024905204773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.949689231812954, 'clip_ratio/low_mean': 0.02440476231276989, 'clip_ratio/low_min': 0.02440476231276989, 'clip_ratio/high_mean': 0.15888158092275262, 'clip_ratio/high_max': 0.15888158092275262, 'clip_ratio/region_mean': 0.1832863432355225, 'reward_total_mean': 0.6282688975334167, 'reward_meter_mean': 0.8117765188217163, 'reward_meter_std': 0.23110267519950867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7433712482452393, 'reward_repeat_penalty_std': 0.018749047070741653, 'reward_repeat_floor_mean': 0.9836742877960205, 'reward_repeat_floor_std': 0.0037498052697628736, 'reward_near_duplicate_penalty_mean': 0.991161584854126, 'reward_near_duplicate_penalty_std': 0.024998728185892105, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.33683136105537415, 'reward_total_composite_mean': 0.6282688975334167, 'reward_total_composite_std': 0.3177328109741211, 'epoch': 0.03}

 15%|█▍        | 491/3300 [1:45:44<7:05:00,  9.08s/it]INFO 04-12 17:44:10 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 492/3300 [1:45:53<7:01:38,  9.01s/it]2026-04-12 17:44:19,475 | INFO | train_grpo_train | metrics_logged mode=train step=492

                                                      
{'loss': -0.0398, 'grad_norm': 13.04485034942627, 'learning_rate': 8.512121212121213e-06, 'num_tokens': 912984.0, 'completions/mean_length': 47.5, 'completions/min_length': 34.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7978067398071289, 'rewards/meter/std': 0.29664960503578186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9685828685760498, 'rewards/lexical_plausibility/std': 0.08886100351810455, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.3215142786502838, 'rewards/repeat_penalty/mean': 0.9837372303009033, 'rewards/repeat_penalty/std': 0.028607802465558052, 'rewards/repeat_floor/mean': 0.9975606203079224, 'rewards/repeat_floor/std': 0.004291166085749865, 'rewards/near_duplicate_penalty/mean': 0.9837372303009033, 'rewards/near_duplicate_penalty/std': 0.028607802465558052, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4801263213157654, 'rewards/total_composite/std': 0.31657642126083374, 'reward': 0.4801263213157654, 'reward_std': 0.31657639145851135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1806916892528534, 'sampling/sampling_logp_difference/max': 2.440725326538086, 'sampling/importance_sampling_ratio/min': 0.0870976522564888, 'sampling/importance_sampling_ratio/mean': 1.026998519897461, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3460486084222794, 'clip_ratio/low_mean': 0.1184615008533001, 'clip_ratio/low_min': 0.1184615008533001, 'clip_ratio/high_mean': 0.06188082788139582, 'clip_ratio/high_max': 0.06188082788139582, 'clip_ratio/region_mean': 0.1803423287346959, 'reward_total_mean': 0.4801263213157654, 'reward_meter_mean': 0.7978067398071289, 'reward_meter_std': 0.29664960503578186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9685828685760498, 'reward_lexical_plausibility_std': 0.08886100351810455, 'reward_repeat_penalty_mean': 0.9837372303009033, 'reward_repeat_penalty_std': 0.028607802465558052, 'reward_repeat_floor_mean': 0.9975606203079224, 'reward_repeat_floor_std': 0.004291166085749865, 'reward_near_duplicate_penalty_mean': 0.9837372303009033, 'reward_near_duplicate_penalty_std': 0.028607802465558052, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.3215142786502838, 'reward_total_composite_mean': 0.4801263213157654, 'reward_total_composite_std': 0.31657642126083374, 'epoch': 0.03}

 15%|█▍        | 492/3300 [1:45:53<7:01:38,  9.01s/it]INFO 04-12 17:44:19 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 493/3300 [1:46:02<7:08:54,  9.17s/it]2026-04-12 17:44:29,013 | INFO | train_grpo_train | metrics_logged mode=train step=493

                                                      
{'loss': 0.0017, 'grad_norm': 9.214557647705078, 'learning_rate': 8.50909090909091e-06, 'num_tokens': 915314.0, 'completions/mean_length': 110.25, 'completions/min_length': 100.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.25, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.8839842677116394, 'rewards/meter/std': 0.20048923790454865, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.837070107460022, 'rewards/repeat_penalty/std': 0.15708816051483154, 'rewards/repeat_floor/mean': 0.9832630157470703, 'rewards/repeat_floor/std': 0.015097341500222683, 'rewards/near_duplicate_penalty/mean': 0.9602870941162109, 'rewards/near_duplicate_penalty/std': 0.03420589864253998, 'rewards/opening_diversity/mean': 0.979687511920929, 'rewards/opening_diversity/std': 0.03892387077212334, 'rewards/distinct_2/mean': 0.8937630653381348, 'rewards/distinct_2/std': 0.09784720838069916, 'rewards/total_composite/mean': 0.3213406205177307, 'rewards/total_composite/std': 0.06527985632419586, 'reward': 0.3213406205177307, 'reward_std': 0.06527985632419586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1578797847032547, 'sampling/sampling_logp_difference/max': 1.584580898284912, 'sampling/importance_sampling_ratio/min': 0.20503370463848114, 'sampling/importance_sampling_ratio/mean': 1.0144155025482178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3237784206867218, 'clip_ratio/low_mean': 0.04249957110732794, 'clip_ratio/low_min': 0.04249957110732794, 'clip_ratio/high_mean': 0.09139722678810358, 'clip_ratio/high_max': 0.09139722678810358, 'clip_ratio/region_mean': 0.13389679789543152, 'reward_total_mean': 0.3213406205177307, 'reward_meter_mean': 0.8839842677116394, 'reward_meter_std': 0.20048923790454865, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.837070107460022, 'reward_repeat_penalty_std': 0.15708816051483154, 'reward_repeat_floor_mean': 0.9832630157470703, 'reward_repeat_floor_std': 0.015097341500222683, 'reward_near_duplicate_penalty_mean': 0.9602870941162109, 'reward_near_duplicate_penalty_std': 0.03420589864253998, 'reward_opening_diversity_mean': 0.979687511920929, 'reward_opening_diversity_std': 0.03892387077212334, 'reward_distinct_2_mean': 0.8937630653381348, 'reward_distinct_2_std': 0.09784720838069916, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3213406205177307, 'reward_total_composite_std': 0.06527985632419586, 'epoch': 0.03}

 15%|█▍        | 493/3300 [1:46:02<7:08:54,  9.17s/it]INFO 04-12 17:44:29 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 494/3300 [1:46:10<6:53:49,  8.85s/it]2026-04-12 17:44:37,117 | INFO | train_grpo_train | metrics_logged mode=train step=494

                                                      
{'loss': 0.0805, 'grad_norm': 17.60483169555664, 'learning_rate': 8.506060606060607e-06, 'num_tokens': 917279.0, 'completions/mean_length': 59.625, 'completions/min_length': 42.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.46373629570007324, 'rewards/meter/std': 0.27098947763442993, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9890648722648621, 'rewards/lexical_plausibility/std': 0.03092925436794758, 'rewards/judge_quality/mean': 0.6499999761581421, 'rewards/judge_quality/std': 0.24951381981372833, 'rewards/repeat_penalty/mean': 0.9050288796424866, 'rewards/repeat_penalty/std': 0.12336964905261993, 'rewards/repeat_floor/mean': 0.9879235029220581, 'rewards/repeat_floor/std': 0.01584974303841591, 'rewards/near_duplicate_penalty/mean': 0.9508566856384277, 'rewards/near_duplicate_penalty/std': 0.05863592028617859, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9477227926254272, 'rewards/distinct_2/std': 0.08174040168523788, 'rewards/total_composite/mean': 0.29722103476524353, 'rewards/total_composite/std': 0.21231958270072937, 'reward': 0.29722103476524353, 'reward_std': 0.21231958270072937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2143169790506363, 'sampling/sampling_logp_difference/max': 3.088449478149414, 'sampling/importance_sampling_ratio/min': 0.045572564005851746, 'sampling/importance_sampling_ratio/mean': 1.012260913848877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8211357221007347, 'clip_ratio/low_mean': 0.12016293406486511, 'clip_ratio/low_min': 0.12016293406486511, 'clip_ratio/high_mean': 0.06825524382293224, 'clip_ratio/high_max': 0.06825524382293224, 'clip_ratio/region_mean': 0.18841817788779736, 'reward_total_mean': 0.29722103476524353, 'reward_meter_mean': 0.46373629570007324, 'reward_meter_std': 0.27098947763442993, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9890648722648621, 'reward_lexical_plausibility_std': 0.03092925436794758, 'reward_repeat_penalty_mean': 0.9050288796424866, 'reward_repeat_penalty_std': 0.12336964905261993, 'reward_repeat_floor_mean': 0.9879235029220581, 'reward_repeat_floor_std': 0.01584974303841591, 'reward_near_duplicate_penalty_mean': 0.9508566856384277, 'reward_near_duplicate_penalty_std': 0.05863592028617859, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9477227926254272, 'reward_distinct_2_std': 0.08174040168523788, 'reward_judge_quality_mean': 0.6499999761581421, 'reward_judge_quality_std': 0.24951381981372833, 'reward_total_composite_mean': 0.29722103476524353, 'reward_total_composite_std': 0.21231958270072937, 'epoch': 0.03}

 15%|█▍        | 494/3300 [1:46:10<6:53:49,  8.85s/it]INFO 04-12 17:44:37 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 495/3300 [1:46:25<8:08:19, 10.45s/it]2026-04-12 17:44:51,289 | INFO | train_grpo_train | metrics_logged mode=train step=495

                                                      
{'loss': -0.3182, 'grad_norm': 2.1097261905670166, 'learning_rate': 8.503030303030304e-06, 'num_tokens': 920445.0, 'completions/mean_length': 318.75, 'completions/min_length': 231.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 254.33334350585938, 'completions/min_terminated_length': 231.0, 'completions/max_terminated_length': 271.0, 'rewards/meter/mean': 0.9180511236190796, 'rewards/meter/std': 0.16932040452957153, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.25153848528862, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9124999642372131, 'rewards/count_floor/std': 0.07546153664588928, 'rewards/lexical_plausibility/mean': 0.8847241401672363, 'rewards/lexical_plausibility/std': 0.22568316757678986, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.4951223134994507, 'rewards/repeat_penalty/std': 0.4002896249294281, 'rewards/repeat_floor/mean': 0.9337047338485718, 'rewards/repeat_floor/std': 0.06018609553575516, 'rewards/near_duplicate_penalty/mean': 0.860082745552063, 'rewards/near_duplicate_penalty/std': 0.14075830578804016, 'rewards/opening_diversity/mean': 0.9040663838386536, 'rewards/opening_diversity/std': 0.15982766449451447, 'rewards/distinct_2/mean': 0.5605370402336121, 'rewards/distinct_2/std': 0.36832550168037415, 'rewards/total_composite/mean': 0.10159417241811752, 'rewards/total_composite/std': 0.045214757323265076, 'reward': 0.10159417241811752, 'reward_std': 0.045214757323265076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1317475438117981, 'sampling/sampling_logp_difference/max': 1.3974332809448242, 'sampling/importance_sampling_ratio/min': 0.24723073840141296, 'sampling/importance_sampling_ratio/mean': 1.0292994976043701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9910262078046799, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08706481289118528, 'clip_ratio/high_max': 0.08706481289118528, 'clip_ratio/region_mean': 0.08706481289118528, 'reward_total_mean': 0.10159417241811752, 'reward_meter_mean': 0.9180511236190796, 'reward_meter_std': 0.16932040452957153, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.25153848528862, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9124999642372131, 'reward_count_floor_std': 0.07546153664588928, 'reward_lexical_plausibility_mean': 0.8847241401672363, 'reward_lexical_plausibility_std': 0.22568316757678986, 'reward_repeat_penalty_mean': 0.4951223134994507, 'reward_repeat_penalty_std': 0.4002896249294281, 'reward_repeat_floor_mean': 0.9337047338485718, 'reward_repeat_floor_std': 0.06018609553575516, 'reward_near_duplicate_penalty_mean': 0.860082745552063, 'reward_near_duplicate_penalty_std': 0.14075830578804016, 'reward_opening_diversity_mean': 0.9040663838386536, 'reward_opening_diversity_std': 0.15982766449451447, 'reward_distinct_2_mean': 0.5605370402336121, 'reward_distinct_2_std': 0.36832550168037415, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.10159417241811752, 'reward_total_composite_std': 0.045214757323265076, 'epoch': 0.03}

 15%|█▌        | 495/3300 [1:46:25<8:08:19, 10.45s/it]INFO 04-12 17:44:51 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 496/3300 [1:46:39<9:00:48, 11.57s/it]2026-04-12 17:45:05,490 | INFO | train_grpo_train | metrics_logged mode=train step=496

                                                      
{'loss': -0.047, 'grad_norm': 0.4166216254234314, 'learning_rate': 8.5e-06, 'num_tokens': 922535.0, 'completions/mean_length': 398.25, 'completions/min_length': 200.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 208.6666717529297, 'completions/min_terminated_length': 200.0, 'completions/max_terminated_length': 215.0, 'rewards/meter/mean': 0.9173742532730103, 'rewards/meter/std': 0.09787047654390335, 'rewards/count_adherence/mean': 0.796875, 'rewards/count_adherence/std': 0.09300298243761063, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9390624761581421, 'rewards/count_floor/std': 0.027900882065296173, 'rewards/lexical_plausibility/mean': 0.8016838431358337, 'rewards/lexical_plausibility/std': 0.1831950843334198, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.5050085186958313, 'rewards/repeat_penalty/std': 0.4174470007419586, 'rewards/repeat_floor/mean': 0.9174864888191223, 'rewards/repeat_floor/std': 0.08609236776828766, 'rewards/near_duplicate_penalty/mean': 0.7942747473716736, 'rewards/near_duplicate_penalty/std': 0.19653551280498505, 'rewards/opening_diversity/mean': 0.7570332288742065, 'rewards/opening_diversity/std': 0.3361288905143738, 'rewards/distinct_2/mean': 0.588036060333252, 'rewards/distinct_2/std': 0.44063302874565125, 'rewards/total_composite/mean': 0.10686292499303818, 'rewards/total_composite/std': 0.027179429307579994, 'reward': 0.10686292499303818, 'reward_std': 0.027179429307579994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07973871380090714, 'sampling/sampling_logp_difference/max': 1.4409397840499878, 'sampling/importance_sampling_ratio/min': 0.23670519888401031, 'sampling/importance_sampling_ratio/mean': 1.0160298347473145, 'sampling/importance_sampling_ratio/max': 1.79944908618927, 'entropy': 0.2535622864961624, 'clip_ratio/low_mean': 0.004651162773370743, 'clip_ratio/low_min': 0.004651162773370743, 'clip_ratio/high_mean': 0.02072867378592491, 'clip_ratio/high_max': 0.02072867378592491, 'clip_ratio/region_mean': 0.025379836559295654, 'reward_total_mean': 0.10686292499303818, 'reward_meter_mean': 0.9173742532730103, 'reward_meter_std': 0.09787047654390335, 'reward_count_adherence_mean': 0.796875, 'reward_count_adherence_std': 0.09300298243761063, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9390624761581421, 'reward_count_floor_std': 0.027900882065296173, 'reward_lexical_plausibility_mean': 0.8016838431358337, 'reward_lexical_plausibility_std': 0.1831950843334198, 'reward_repeat_penalty_mean': 0.5050085186958313, 'reward_repeat_penalty_std': 0.4174470007419586, 'reward_repeat_floor_mean': 0.9174864888191223, 'reward_repeat_floor_std': 0.08609236776828766, 'reward_near_duplicate_penalty_mean': 0.7942747473716736, 'reward_near_duplicate_penalty_std': 0.19653551280498505, 'reward_opening_diversity_mean': 0.7570332288742065, 'reward_opening_diversity_std': 0.3361288905143738, 'reward_distinct_2_mean': 0.588036060333252, 'reward_distinct_2_std': 0.44063302874565125, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.10686292499303818, 'reward_total_composite_std': 0.027179429307579994, 'epoch': 0.03}

 15%|█▌        | 496/3300 [1:46:39<9:00:48, 11.57s/it]INFO 04-12 17:45:05 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 497/3300 [1:46:53<9:30:39, 12.22s/it]2026-04-12 17:45:19,207 | INFO | train_grpo_train | metrics_logged mode=train step=497

                                                      
{'loss': -0.0397, 'grad_norm': 5.391056537628174, 'learning_rate': 8.496969696969697e-06, 'num_tokens': 924183.0, 'completions/mean_length': 101.0, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.28571701049805, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7750760316848755, 'rewards/meter/std': 0.38514024019241333, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8815553188323975, 'rewards/lexical_plausibility/std': 0.2033458650112152, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.2133909910917282, 'rewards/repeat_penalty/mean': 0.9121878147125244, 'rewards/repeat_penalty/std': 0.08433179557323456, 'rewards/repeat_floor/mean': 0.9911338686943054, 'rewards/repeat_floor/std': 0.006743339356034994, 'rewards/near_duplicate_penalty/mean': 0.9704256057739258, 'rewards/near_duplicate_penalty/std': 0.027934445068240166, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.971611738204956, 'rewards/distinct_2/std': 0.03919585049152374, 'rewards/total_composite/mean': 0.35213810205459595, 'rewards/total_composite/std': 0.23824480175971985, 'reward': 0.35213810205459595, 'reward_std': 0.23824478685855865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17342451214790344, 'sampling/sampling_logp_difference/max': 1.6927714347839355, 'sampling/importance_sampling_ratio/min': 0.18400885164737701, 'sampling/importance_sampling_ratio/mean': 1.0089201927185059, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.298454836010933, 'clip_ratio/low_mean': 0.027465221472084522, 'clip_ratio/low_min': 0.027465221472084522, 'clip_ratio/high_mean': 0.10987714119255543, 'clip_ratio/high_max': 0.10987714119255543, 'clip_ratio/region_mean': 0.13734236266463995, 'reward_total_mean': 0.35213810205459595, 'reward_meter_mean': 0.7750760316848755, 'reward_meter_std': 0.38514024019241333, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8815553188323975, 'reward_lexical_plausibility_std': 0.2033458650112152, 'reward_repeat_penalty_mean': 0.9121878147125244, 'reward_repeat_penalty_std': 0.08433179557323456, 'reward_repeat_floor_mean': 0.9911338686943054, 'reward_repeat_floor_std': 0.006743339356034994, 'reward_near_duplicate_penalty_mean': 0.9704256057739258, 'reward_near_duplicate_penalty_std': 0.027934445068240166, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.971611738204956, 'reward_distinct_2_std': 0.03919585049152374, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.2133909910917282, 'reward_total_composite_mean': 0.35213810205459595, 'reward_total_composite_std': 0.23824480175971985, 'epoch': 0.03}

 15%|█▌        | 497/3300 [1:46:53<9:30:39, 12.22s/it]INFO 04-12 17:45:19 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 498/3300 [1:47:02<8:51:04, 11.37s/it]2026-04-12 17:45:28,612 | INFO | train_grpo_train | metrics_logged mode=train step=498

                                                      
{'loss': 0.0807, 'grad_norm': 13.27564525604248, 'learning_rate': 8.493939393939394e-06, 'num_tokens': 926069.0, 'completions/mean_length': 63.75, 'completions/min_length': 57.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6959583759307861, 'rewards/meter/std': 0.4100753664970398, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5837499499320984, 'rewards/judge_quality/std': 0.23231369256973267, 'rewards/repeat_penalty/mean': 0.9601210355758667, 'rewards/repeat_penalty/std': 0.05876023322343826, 'rewards/repeat_floor/mean': 0.9952081441879272, 'rewards/repeat_floor/std': 0.006436796393245459, 'rewards/near_duplicate_penalty/mean': 0.9813921451568604, 'rewards/near_duplicate_penalty/std': 0.018059436231851578, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9777709245681763, 'rewards/distinct_2/std': 0.046208228915929794, 'rewards/total_composite/mean': 0.38639330863952637, 'rewards/total_composite/std': 0.2618277966976166, 'reward': 0.38639330863952637, 'reward_std': 0.2618277668952942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15465795993804932, 'sampling/sampling_logp_difference/max': 1.8801250457763672, 'sampling/importance_sampling_ratio/min': 0.15257102251052856, 'sampling/importance_sampling_ratio/mean': 1.0143102407455444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1680407375097275, 'clip_ratio/low_mean': 0.07754085771739483, 'clip_ratio/low_min': 0.07754085771739483, 'clip_ratio/high_mean': 0.08074276801198721, 'clip_ratio/high_max': 0.08074276801198721, 'clip_ratio/region_mean': 0.15828362572938204, 'reward_total_mean': 0.38639330863952637, 'reward_meter_mean': 0.6959583759307861, 'reward_meter_std': 0.4100753664970398, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9601210355758667, 'reward_repeat_penalty_std': 0.05876023322343826, 'reward_repeat_floor_mean': 0.9952081441879272, 'reward_repeat_floor_std': 0.006436796393245459, 'reward_near_duplicate_penalty_mean': 0.9813921451568604, 'reward_near_duplicate_penalty_std': 0.018059436231851578, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9777709245681763, 'reward_distinct_2_std': 0.046208228915929794, 'reward_judge_quality_mean': 0.5837499499320984, 'reward_judge_quality_std': 0.23231369256973267, 'reward_total_composite_mean': 0.38639330863952637, 'reward_total_composite_std': 0.2618277966976166, 'epoch': 0.03}

 15%|█▌        | 498/3300 [1:47:02<8:51:04, 11.37s/it]INFO 04-12 17:45:28 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 499/3300 [1:47:12<8:25:55, 10.84s/it]2026-04-12 17:45:38,200 | INFO | train_grpo_train | metrics_logged mode=train step=499

                                                      
{'loss': -0.0282, 'grad_norm': 15.101868629455566, 'learning_rate': 8.490909090909092e-06, 'num_tokens': 928039.0, 'completions/mean_length': 72.25, 'completions/min_length': 61.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.10586683452129364, 'rewards/meter/std': 0.11134783923625946, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.7124732136726379, 'rewards/repeat_penalty/std': 0.25966528058052063, 'rewards/repeat_floor/mean': 0.9673922061920166, 'rewards/repeat_floor/std': 0.03315800055861473, 'rewards/near_duplicate_penalty/mean': 0.9282726645469666, 'rewards/near_duplicate_penalty/std': 0.07145749032497406, 'rewards/opening_diversity/mean': 0.8619791269302368, 'rewards/opening_diversity/std': 0.21791578829288483, 'rewards/distinct_2/mean': 0.849250078201294, 'rewards/distinct_2/std': 0.11509034037590027, 'rewards/total_composite/mean': 0.03694423288106918, 'rewards/total_composite/std': 0.037478912621736526, 'reward': 0.03694423288106918, 'reward_std': 0.037478912621736526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19849267601966858, 'sampling/sampling_logp_difference/max': 2.7425365447998047, 'sampling/importance_sampling_ratio/min': 0.06440676748752594, 'sampling/importance_sampling_ratio/mean': 1.0187902450561523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0103928744792938, 'clip_ratio/low_mean': 0.11610000394284725, 'clip_ratio/low_min': 0.11610000394284725, 'clip_ratio/high_mean': 0.06790966540575027, 'clip_ratio/high_max': 0.06790966540575027, 'clip_ratio/region_mean': 0.18400966934859753, 'reward_total_mean': 0.03694423288106918, 'reward_meter_mean': 0.10586683452129364, 'reward_meter_std': 0.11134783923625946, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7124732136726379, 'reward_repeat_penalty_std': 0.25966528058052063, 'reward_repeat_floor_mean': 0.9673922061920166, 'reward_repeat_floor_std': 0.03315800055861473, 'reward_near_duplicate_penalty_mean': 0.9282726645469666, 'reward_near_duplicate_penalty_std': 0.07145749032497406, 'reward_opening_diversity_mean': 0.8619791269302368, 'reward_opening_diversity_std': 0.21791578829288483, 'reward_distinct_2_mean': 0.849250078201294, 'reward_distinct_2_std': 0.11509034037590027, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.03694423288106918, 'reward_total_composite_std': 0.037478912621736526, 'epoch': 0.03}

 15%|█▌        | 499/3300 [1:47:12<8:25:55, 10.84s/it]INFO 04-12 17:45:38 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 500/3300 [1:47:20<7:55:15, 10.18s/it]2026-04-12 17:45:46,859 | INFO | train_grpo_train | metrics_logged mode=train step=500

                                                      
{'loss': 0.047, 'grad_norm': 13.783639907836914, 'learning_rate': 8.487878787878789e-06, 'num_tokens': 929661.0, 'completions/mean_length': 47.75, 'completions/min_length': 42.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8805423974990845, 'rewards/meter/std': 0.3042807877063751, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.9955461025238037, 'rewards/repeat_penalty/std': 0.008062711916863918, 'rewards/repeat_floor/mean': 0.9993319511413574, 'rewards/repeat_floor/std': 0.0012094092089682817, 'rewards/near_duplicate_penalty/mean': 0.9955461025238037, 'rewards/near_duplicate_penalty/std': 0.008062711916863918, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4048282504081726, 'rewards/total_composite/std': 0.2536325752735138, 'reward': 0.4048282504081726, 'reward_std': 0.2536325752735138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18422798812389374, 'sampling/sampling_logp_difference/max': 1.2962336540222168, 'sampling/importance_sampling_ratio/min': 0.2897098660469055, 'sampling/importance_sampling_ratio/mean': 1.0293875932693481, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6455870270729065, 'clip_ratio/low_mean': 0.039795113960281014, 'clip_ratio/low_min': 0.039795113960281014, 'clip_ratio/high_mean': 0.07668384909629822, 'clip_ratio/high_max': 0.07668384909629822, 'clip_ratio/region_mean': 0.11647896305657923, 'reward_total_mean': 0.4048282504081726, 'reward_meter_mean': 0.8805423974990845, 'reward_meter_std': 0.3042807877063751, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9955461025238037, 'reward_repeat_penalty_std': 0.008062711916863918, 'reward_repeat_floor_mean': 0.9993319511413574, 'reward_repeat_floor_std': 0.0012094092089682817, 'reward_near_duplicate_penalty_mean': 0.9955461025238037, 'reward_near_duplicate_penalty_std': 0.008062711916863918, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.4048282504081726, 'reward_total_composite_std': 0.2536325752735138, 'epoch': 0.03}

 15%|█▌        | 500/3300 [1:47:20<7:55:15, 10.18s/it]INFO 04-12 17:45:47 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:10<01:00,  5.47s/it]

 23%|██▎       | 3/13 [00:17<01:00,  6.10s/it]

 31%|███       | 4/13 [00:29<01:14,  8.23s/it]

 38%|███▊      | 5/13 [00:41<01:16,  9.57s/it]

 46%|████▌     | 6/13 [00:53<01:11, 10.19s/it]

 54%|█████▍    | 7/13 [01:02<00:59,  9.98s/it]

 62%|██████▏   | 8/13 [01:14<00:52, 10.47s/it]

 69%|██████▉   | 9/13 [01:25<00:43, 10.81s/it]

 77%|███████▋  | 10/13 [01:35<00:30, 10.29s/it]

 85%|████████▍ | 11/13 [01:46<00:21, 10.60s/it]

 92%|█████████▏| 12/13 [01:57<00:10, 10.84s/it]

100%|██████████| 13/13 [02:08<00:00, 10.70s/it]2026-04-12 17:48:06,937 | INFO | train_grpo_train | metrics_logged mode=eval step=500

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 140.0077, 'eval_samples_per_second': 0.743, 'eval_steps_per_second': 0.093, 'eval_num_tokens': 929661.0, 'eval_completions/mean_length': 207.94230769230768, 'eval_completions/min_length': 39.30769230769231, 'eval_completions/max_length': 445.15384615384613, 'eval_completions/clipped_ratio': 0.1346153846153846, 'eval_completions/mean_terminated_length': 161.69734896146335, 'eval_completions/min_terminated_length': 39.30769230769231, 'eval_completions/max_terminated_length': 353.38461538461536, 'eval_rewards/meter/mean': 0.6244613459477057, 'eval_rewards/meter/std': 0.3818479042786818, 'eval_rewards/count_adherence/mean': 0.8154045114150414, 'eval_rewards/count_adherence/std': 0.19789515779568598, 'eval_rewards/arabic_clean/mean': 0.8461538461538461, 'eval_rewards/arabic_clean/std': 0.2952340107697707, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9446213566339933, 'eval_rewards/count_floor/std': 0.05936854848494896, 'eval_rewards/lexical_plausibility/mean': 0.9487303587106558, 'eval_rewards/lexical_plausibility/std': 0.10835179056112583, 'eval_rewards/judge_quality/mean': 0.32451922733050126, 'eval_rewards/judge_quality/std': 0.1936850673877276, 'eval_rewards/repeat_penalty/mean': 0.6094352167386275, 'eval_rewards/repeat_penalty/std': 0.3754282702620213, 'eval_rewards/repeat_floor/mean': 0.9471771717071533, 'eval_rewards/repeat_floor/std': 0.06034939077038031, 'eval_rewards/near_duplicate_penalty/mean': 0.8858882188796997, 'eval_rewards/near_duplicate_penalty/std': 0.1401778683066368, 'eval_rewards/opening_diversity/mean': 0.904761268542363, 'eval_rewards/opening_diversity/std': 0.1724496314731928, 'eval_rewards/distinct_2/mean': 0.6667584410080543, 'eval_rewards/distinct_2/std': 0.3708963772425285, 'eval_rewards/total_composite/mean': 0.1807767548240148, 'eval_rewards/total_composite/std': 0.157410629093647, 'eval_reward': 0.1807767548240148, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07252437248826027, 'eval_sampling/sampling_logp_difference/max': 1.2025261658888597, 'eval_sampling/importance_sampling_ratio/min': 0.3037391614455443, 'eval_sampling/importance_sampling_ratio/mean': 1.0198851548708403, 'eval_sampling/importance_sampling_ratio/max': 1.5503454300073476, 'eval_entropy': 0.8679274183053237, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1807767548240148, 'eval_reward_meter_mean': 0.6244613459477057, 'eval_reward_meter_std': 0.3818479042786818, 'eval_reward_count_adherence_mean': 0.8154045114150414, 'eval_reward_count_adherence_std': 0.19789515779568598, 'eval_reward_arabic_clean_mean': 0.8461538461538461, 'eval_reward_arabic_clean_std': 0.2952340107697707, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9446213566339933, 'eval_reward_count_floor_std': 0.05936854848494896, 'eval_reward_lexical_plausibility_mean': 0.9487303587106558, 'eval_reward_lexical_plausibility_std': 0.10835179056112583, 'eval_reward_repeat_penalty_mean': 0.6094352167386275, 'eval_reward_repeat_penalty_std': 0.3754282702620213, 'eval_reward_repeat_floor_mean': 0.9471771717071533, 'eval_reward_repeat_floor_std': 0.06034939077038031, 'eval_reward_near_duplicate_penalty_mean': 0.8858882188796997, 'eval_reward_near_duplicate_penalty_std': 0.1401778683066368, 'eval_reward_opening_diversity_mean': 0.904761268542363, 'eval_reward_opening_diversity_std': 0.1724496314731928, 'eval_reward_distinct_2_mean': 0.6667584410080543, 'eval_reward_distinct_2_std': 0.3708963772425285, 'eval_reward_judge_quality_mean': 0.32451922733050126, 'eval_reward_judge_quality_std': 0.1936850673877276, 'eval_reward_total_composite_mean': 0.1807767548240148, 'eval_reward_total_composite_std': 0.157410629093647, 'epoch': 0.03}

 15%|█▌        | 500/3300 [1:49:40<7:55:15, 10.18s/it]

100%|██████████| 13/13 [02:08<00:00, 10.70s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 15%|█▌        | 501/3300 [1:49:58<42:16:10, 54.37s/it]2026-04-12 17:48:24,342 | INFO | train_grpo_train | metrics_logged mode=train step=501

                                                       
{'loss': -0.0266, 'grad_norm': 2.2380096912384033, 'learning_rate': 8.484848484848486e-06, 'num_tokens': 931082.0, 'completions/mean_length': 146.625, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 24.83333396911621, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6835510730743408, 'rewards/meter/std': 0.42393729090690613, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.619035542011261, 'rewards/lexical_plausibility/std': 0.19606757164001465, 'rewards/judge_quality/mean': 0.3175000250339508, 'rewards/judge_quality/std': 0.3744233548641205, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27723225951194763, 'rewards/total_composite/std': 0.3852379620075226, 'reward': 0.27723225951194763, 'reward_std': 0.3852379620075226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17146243155002594, 'sampling/sampling_logp_difference/max': 1.5724658966064453, 'sampling/importance_sampling_ratio/min': 0.20753279328346252, 'sampling/importance_sampling_ratio/mean': 1.0229686498641968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2139780223369598, 'clip_ratio/low_mean': 0.07788322027772665, 'clip_ratio/low_min': 0.07788322027772665, 'clip_ratio/high_mean': 0.04282407369464636, 'clip_ratio/high_max': 0.04282407369464636, 'clip_ratio/region_mean': 0.12070729397237301, 'reward_total_mean': 0.27723225951194763, 'reward_meter_mean': 0.6835510730743408, 'reward_meter_std': 0.42393729090690613, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.619035542011261, 'reward_lexical_plausibility_std': 0.19606757164001465, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3175000250339508, 'reward_judge_quality_std': 0.3744233548641205, 'reward_total_composite_mean': 0.27723225951194763, 'reward_total_composite_std': 0.3852379620075226, 'epoch': 0.03}

 15%|█▌        | 501/3300 [1:49:58<42:16:10, 54.37s/it]INFO 04-12 17:48:24 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 502/3300 [1:50:07<31:44:29, 40.84s/it]2026-04-12 17:48:33,618 | INFO | train_grpo_train | metrics_logged mode=train step=502

                                                       
{'loss': 0.0553, 'grad_norm': 18.66092872619629, 'learning_rate': 8.481818181818182e-06, 'num_tokens': 932642.0, 'completions/mean_length': 36.0, 'completions/min_length': 32.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8124307990074158, 'rewards/meter/std': 0.2910243570804596, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9420679211616516, 'rewards/repeat_penalty/std': 0.11843415349721909, 'rewards/repeat_floor/mean': 0.9940252900123596, 'rewards/repeat_floor/std': 0.010170619934797287, 'rewards/near_duplicate_penalty/mean': 0.9787415862083435, 'rewards/near_duplicate_penalty/std': 0.01927022822201252, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.23945167660713196, 'rewards/total_composite/std': 0.1405368149280548, 'reward': 0.23945167660713196, 'reward_std': 0.1405368149280548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1669476330280304, 'sampling/sampling_logp_difference/max': 1.7191133499145508, 'sampling/importance_sampling_ratio/min': 0.17922499775886536, 'sampling/importance_sampling_ratio/mean': 1.0241307020187378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2724262177944183, 'clip_ratio/low_mean': 0.09334236197173595, 'clip_ratio/low_min': 0.09334236197173595, 'clip_ratio/high_mean': 0.05911458469927311, 'clip_ratio/high_max': 0.05911458469927311, 'clip_ratio/region_mean': 0.15245694667100906, 'reward_total_mean': 0.23945167660713196, 'reward_meter_mean': 0.8124307990074158, 'reward_meter_std': 0.2910243570804596, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9420679211616516, 'reward_repeat_penalty_std': 0.11843415349721909, 'reward_repeat_floor_mean': 0.9940252900123596, 'reward_repeat_floor_std': 0.010170619934797287, 'reward_near_duplicate_penalty_mean': 0.9787415862083435, 'reward_near_duplicate_penalty_std': 0.01927022822201252, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.23945167660713196, 'reward_total_composite_std': 0.1405368149280548, 'epoch': 0.03}

 15%|█▌        | 502/3300 [1:50:07<31:44:29, 40.84s/it]INFO 04-12 17:48:33 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 503/3300 [1:50:16<24:20:22, 31.33s/it]2026-04-12 17:48:42,727 | INFO | train_grpo_train | metrics_logged mode=train step=503

                                                       
{'loss': 0.0894, 'grad_norm': 28.734704971313477, 'learning_rate': 8.478787878787879e-06, 'num_tokens': 934168.0, 'completions/mean_length': 32.75, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6736565828323364, 'rewards/meter/std': 0.42780157923698425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9625290632247925, 'rewards/repeat_penalty/std': 0.038888026028871536, 'rewards/repeat_floor/mean': 0.9943793416023254, 'rewards/repeat_floor/std': 0.005833211820572615, 'rewards/near_duplicate_penalty/mean': 0.9625290632247925, 'rewards/near_duplicate_penalty/std': 0.038888026028871536, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4264630377292633, 'rewards/total_composite/std': 0.31055524945259094, 'reward': 0.4264630377292633, 'reward_std': 0.31055521965026855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17910559475421906, 'sampling/sampling_logp_difference/max': 1.668534755706787, 'sampling/importance_sampling_ratio/min': 0.1885230988264084, 'sampling/importance_sampling_ratio/mean': 0.9994204044342041, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0970648378133774, 'clip_ratio/low_mean': 0.05626591667532921, 'clip_ratio/low_min': 0.05626591667532921, 'clip_ratio/high_mean': 0.08261030213907361, 'clip_ratio/high_max': 0.08261030213907361, 'clip_ratio/region_mean': 0.13887621881440282, 'reward_total_mean': 0.4264630377292633, 'reward_meter_mean': 0.6736565828323364, 'reward_meter_std': 0.42780157923698425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9625290632247925, 'reward_repeat_penalty_std': 0.038888026028871536, 'reward_repeat_floor_mean': 0.9943793416023254, 'reward_repeat_floor_std': 0.005833211820572615, 'reward_near_duplicate_penalty_mean': 0.9625290632247925, 'reward_near_duplicate_penalty_std': 0.038888026028871536, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.4264630377292633, 'reward_total_composite_std': 0.31055524945259094, 'epoch': 0.03}

 15%|█▌        | 503/3300 [1:50:16<24:20:22, 31.33s/it]INFO 04-12 17:48:43 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 504/3300 [1:50:24<18:53:50, 24.33s/it]2026-04-12 17:48:50,735 | INFO | train_grpo_train | metrics_logged mode=train step=504

                                                       
{'loss': 0.0829, 'grad_norm': 21.916610717773438, 'learning_rate': 8.475757575757576e-06, 'num_tokens': 935611.0, 'completions/mean_length': 20.375, 'completions/min_length': 18.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.779515266418457, 'rewards/meter/std': 0.30250662565231323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9703159332275391, 'rewards/lexical_plausibility/std': 0.08395931124687195, 'rewards/judge_quality/mean': 0.8237500190734863, 'rewards/judge_quality/std': 0.2722361385822296, 'rewards/repeat_penalty/mean': 0.7487824559211731, 'rewards/repeat_penalty/std': 0.003443692810833454, 'rewards/repeat_floor/mean': 0.9847564697265625, 'rewards/repeat_floor/std': 0.0006887427298352122, 'rewards/near_duplicate_penalty/mean': 0.9983766078948975, 'rewards/near_duplicate_penalty/std': 0.004591604229062796, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6573514342308044, 'rewards/total_composite/std': 0.3476335406303406, 'reward': 0.6573514342308044, 'reward_std': 0.3476335406303406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17827972769737244, 'sampling/sampling_logp_difference/max': 1.0622525215148926, 'sampling/importance_sampling_ratio/min': 0.3456763029098511, 'sampling/importance_sampling_ratio/mean': 1.0323456525802612, 'sampling/importance_sampling_ratio/max': 1.8639169931411743, 'entropy': 1.3325960487127304, 'clip_ratio/low_mean': 0.03958333469927311, 'clip_ratio/low_min': 0.03958333469927311, 'clip_ratio/high_mean': 0.07093254057690501, 'clip_ratio/high_max': 0.07093254057690501, 'clip_ratio/region_mean': 0.11051587527617812, 'reward_total_mean': 0.6573514342308044, 'reward_meter_mean': 0.779515266418457, 'reward_meter_std': 0.30250662565231323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9703159332275391, 'reward_lexical_plausibility_std': 0.08395931124687195, 'reward_repeat_penalty_mean': 0.7487824559211731, 'reward_repeat_penalty_std': 0.003443692810833454, 'reward_repeat_floor_mean': 0.9847564697265625, 'reward_repeat_floor_std': 0.0006887427298352122, 'reward_near_duplicate_penalty_mean': 0.9983766078948975, 'reward_near_duplicate_penalty_std': 0.004591604229062796, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8237500190734863, 'reward_judge_quality_std': 0.2722361385822296, 'reward_total_composite_mean': 0.6573514342308044, 'reward_total_composite_std': 0.3476335406303406, 'epoch': 0.03}

 15%|█▌        | 504/3300 [1:50:24<18:53:50, 24.33s/it]INFO 04-12 17:48:50 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 505/3300 [1:50:38<16:26:10, 21.17s/it]2026-04-12 17:49:04,529 | INFO | train_grpo_train | metrics_logged mode=train step=505

                                                       
{'loss': -0.1554, 'grad_norm': 3.5683786869049072, 'learning_rate': 8.472727272727274e-06, 'num_tokens': 937850.0, 'completions/mean_length': 216.875, 'completions/min_length': 103.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 118.5, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.34393617510795593, 'rewards/meter/std': 0.2458122968673706, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.19820624589920044, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.05946187674999237, 'rewards/lexical_plausibility/mean': 0.8972842693328857, 'rewards/lexical_plausibility/std': 0.21046949923038483, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.84952712059021, 'rewards/repeat_penalty/std': 0.22977428138256073, 'rewards/repeat_floor/mean': 0.9812049865722656, 'rewards/repeat_floor/std': 0.029502667486667633, 'rewards/near_duplicate_penalty/mean': 0.9400865435600281, 'rewards/near_duplicate_penalty/std': 0.07793454080820084, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9118558168411255, 'rewards/distinct_2/std': 0.14232002198696136, 'rewards/total_composite/mean': 0.11806898564100266, 'rewards/total_composite/std': 0.09317027777433395, 'reward': 0.11806898564100266, 'reward_std': 0.09317027777433395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18992212414741516, 'sampling/sampling_logp_difference/max': 2.0112175941467285, 'sampling/importance_sampling_ratio/min': 0.1338256299495697, 'sampling/importance_sampling_ratio/mean': 1.0228536128997803, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0625020116567612, 'clip_ratio/low_mean': 0.0348020913079381, 'clip_ratio/low_min': 0.0348020913079381, 'clip_ratio/high_mean': 0.08876976557075977, 'clip_ratio/high_max': 0.08876976557075977, 'clip_ratio/region_mean': 0.12357185687869787, 'reward_total_mean': 0.11806898564100266, 'reward_meter_mean': 0.34393617510795593, 'reward_meter_std': 0.2458122968673706, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.19820624589920044, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.05946187674999237, 'reward_lexical_plausibility_mean': 0.8972842693328857, 'reward_lexical_plausibility_std': 0.21046949923038483, 'reward_repeat_penalty_mean': 0.84952712059021, 'reward_repeat_penalty_std': 0.22977428138256073, 'reward_repeat_floor_mean': 0.9812049865722656, 'reward_repeat_floor_std': 0.029502667486667633, 'reward_near_duplicate_penalty_mean': 0.9400865435600281, 'reward_near_duplicate_penalty_std': 0.07793454080820084, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9118558168411255, 'reward_distinct_2_std': 0.14232002198696136, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.11806898564100266, 'reward_total_composite_std': 0.09317027777433395, 'epoch': 0.03}

 15%|█▌        | 505/3300 [1:50:38<16:26:10, 21.17s/it]INFO 04-12 17:49:04 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 506/3300 [1:50:52<14:41:36, 18.93s/it]2026-04-12 17:49:18,242 | INFO | train_grpo_train | metrics_logged mode=train step=506

                                                       
{'loss': -0.0499, 'grad_norm': 4.9672064781188965, 'learning_rate': 8.46969696969697e-06, 'num_tokens': 939513.0, 'completions/mean_length': 109.875, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.42857360839844, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7858970165252686, 'rewards/meter/std': 0.35001009702682495, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9274303913116455, 'rewards/lexical_plausibility/std': 0.19019381701946259, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.26384180784225464, 'rewards/repeat_penalty/mean': 0.9860967397689819, 'rewards/repeat_penalty/std': 0.02008829265832901, 'rewards/repeat_floor/mean': 0.9980258941650391, 'rewards/repeat_floor/std': 0.002791586797684431, 'rewards/near_duplicate_penalty/mean': 0.9880580306053162, 'rewards/near_duplicate_penalty/std': 0.016480281949043274, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9979685544967651, 'rewards/distinct_2/std': 0.005745711270719767, 'rewards/total_composite/mean': 0.3787696659564972, 'rewards/total_composite/std': 0.30826422572135925, 'reward': 0.3787696659564972, 'reward_std': 0.30826425552368164, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16874362528324127, 'sampling/sampling_logp_difference/max': 1.6193904876708984, 'sampling/importance_sampling_ratio/min': 0.1980193555355072, 'sampling/importance_sampling_ratio/mean': 1.0501161813735962, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.258274033665657, 'clip_ratio/low_mean': 0.06342687178403139, 'clip_ratio/low_min': 0.06342687178403139, 'clip_ratio/high_mean': 0.07070060167461634, 'clip_ratio/high_max': 0.07070060167461634, 'clip_ratio/region_mean': 0.13412747345864773, 'reward_total_mean': 0.3787696659564972, 'reward_meter_mean': 0.7858970165252686, 'reward_meter_std': 0.35001009702682495, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9274303913116455, 'reward_lexical_plausibility_std': 0.19019381701946259, 'reward_repeat_penalty_mean': 0.9860967397689819, 'reward_repeat_penalty_std': 0.02008829265832901, 'reward_repeat_floor_mean': 0.9980258941650391, 'reward_repeat_floor_std': 0.002791586797684431, 'reward_near_duplicate_penalty_mean': 0.9880580306053162, 'reward_near_duplicate_penalty_std': 0.016480281949043274, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9979685544967651, 'reward_distinct_2_std': 0.005745711270719767, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.26384180784225464, 'reward_total_composite_mean': 0.3787696659564972, 'reward_total_composite_std': 0.30826422572135925, 'epoch': 0.03}

 15%|█▌        | 506/3300 [1:50:52<14:41:36, 18.93s/it]INFO 04-12 17:49:18 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 507/3300 [1:51:06<13:36:38, 17.54s/it]2026-04-12 17:49:32,545 | INFO | train_grpo_train | metrics_logged mode=train step=507

                                                       
{'loss': -0.1184, 'grad_norm': 7.1238603591918945, 'learning_rate': 8.466666666666668e-06, 'num_tokens': 942101.0, 'completions/mean_length': 179.5, 'completions/min_length': 106.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 132.0, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.7116265296936035, 'rewards/meter/std': 0.21642476320266724, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9625000357627869, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 0.9538431763648987, 'rewards/lexical_plausibility/std': 0.13055117428302765, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9227674603462219, 'rewards/repeat_penalty/std': 0.15741252899169922, 'rewards/repeat_floor/mean': 0.9924988746643066, 'rewards/repeat_floor/std': 0.013243019580841064, 'rewards/near_duplicate_penalty/mean': 0.9739339351654053, 'rewards/near_duplicate_penalty/std': 0.033551253378391266, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9600974321365356, 'rewards/distinct_2/std': 0.0936034694314003, 'rewards/total_composite/mean': 0.22344699501991272, 'rewards/total_composite/std': 0.10760604590177536, 'reward': 0.22344699501991272, 'reward_std': 0.10760603845119476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23907411098480225, 'sampling/sampling_logp_difference/max': 2.4978647232055664, 'sampling/importance_sampling_ratio/min': 0.08226045966148376, 'sampling/importance_sampling_ratio/mean': 0.9915637969970703, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9913438484072685, 'clip_ratio/low_mean': 0.061237018555402756, 'clip_ratio/low_min': 0.061237018555402756, 'clip_ratio/high_mean': 0.10777921229600906, 'clip_ratio/high_max': 0.10777921229600906, 'clip_ratio/region_mean': 0.16901623085141182, 'reward_total_mean': 0.22344699501991272, 'reward_meter_mean': 0.7116265296936035, 'reward_meter_std': 0.21642476320266724, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9625000357627869, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 0.9538431763648987, 'reward_lexical_plausibility_std': 0.13055117428302765, 'reward_repeat_penalty_mean': 0.9227674603462219, 'reward_repeat_penalty_std': 0.15741252899169922, 'reward_repeat_floor_mean': 0.9924988746643066, 'reward_repeat_floor_std': 0.013243019580841064, 'reward_near_duplicate_penalty_mean': 0.9739339351654053, 'reward_near_duplicate_penalty_std': 0.033551253378391266, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9600974321365356, 'reward_distinct_2_std': 0.0936034694314003, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.22344699501991272, 'reward_total_composite_std': 0.10760604590177536, 'epoch': 0.03}

 15%|█▌        | 507/3300 [1:51:06<13:36:38, 17.54s/it]INFO 04-12 17:49:32 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 508/3300 [1:51:14<11:18:21, 14.58s/it]2026-04-12 17:49:40,200 | INFO | train_grpo_train | metrics_logged mode=train step=508

                                                       
{'loss': 0.0503, 'grad_norm': 16.859115600585938, 'learning_rate': 8.463636363636364e-06, 'num_tokens': 943665.0, 'completions/mean_length': 40.5, 'completions/min_length': 32.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.840571403503418, 'rewards/meter/std': 0.2462586909532547, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8600000143051147, 'rewards/judge_quality/std': 0.16673332452774048, 'rewards/repeat_penalty/mean': 0.9342851042747498, 'rewards/repeat_penalty/std': 0.11919376254081726, 'rewards/repeat_floor/mean': 0.9927542209625244, 'rewards/repeat_floor/std': 0.010998665355145931, 'rewards/near_duplicate_penalty/mean': 0.9692113399505615, 'rewards/near_duplicate_penalty/std': 0.03722553700208664, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.708553671836853, 'rewards/total_composite/std': 0.2429817020893097, 'reward': 0.708553671836853, 'reward_std': 0.2429817020893097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18467795848846436, 'sampling/sampling_logp_difference/max': 1.2500619888305664, 'sampling/importance_sampling_ratio/min': 0.28648704290390015, 'sampling/importance_sampling_ratio/mean': 1.0246526002883911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1481572911143303, 'clip_ratio/low_mean': 0.07120298407971859, 'clip_ratio/low_min': 0.07120298407971859, 'clip_ratio/high_mean': 0.10157212615013123, 'clip_ratio/high_max': 0.10157212615013123, 'clip_ratio/region_mean': 0.17277511022984982, 'reward_total_mean': 0.708553671836853, 'reward_meter_mean': 0.840571403503418, 'reward_meter_std': 0.2462586909532547, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9342851042747498, 'reward_repeat_penalty_std': 0.11919376254081726, 'reward_repeat_floor_mean': 0.9927542209625244, 'reward_repeat_floor_std': 0.010998665355145931, 'reward_near_duplicate_penalty_mean': 0.9692113399505615, 'reward_near_duplicate_penalty_std': 0.03722553700208664, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.8600000143051147, 'reward_judge_quality_std': 0.16673332452774048, 'reward_total_composite_mean': 0.708553671836853, 'reward_total_composite_std': 0.2429817020893097, 'epoch': 0.03}

 15%|█▌        | 508/3300 [1:51:14<11:18:21, 14.58s/it]INFO 04-12 17:49:40 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 509/3300 [1:51:22<9:48:46, 12.66s/it] 2026-04-12 17:49:48,377 | INFO | train_grpo_train | metrics_logged mode=train step=509

                                                      
{'loss': -0.0455, 'grad_norm': 15.400981903076172, 'learning_rate': 8.460606060606061e-06, 'num_tokens': 945077.0, 'completions/mean_length': 35.5, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8986762762069702, 'rewards/meter/std': 0.12459779530763626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9183171987533569, 'rewards/repeat_penalty/std': 0.10537010431289673, 'rewards/repeat_floor/mean': 0.9916671514511108, 'rewards/repeat_floor/std': 0.00818734522908926, 'rewards/near_duplicate_penalty/mean': 0.9737309217453003, 'rewards/near_duplicate_penalty/std': 0.020731303840875626, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9720279574394226, 'rewards/distinct_2/std': 0.03866533935070038, 'rewards/total_composite/mean': 0.30703824758529663, 'rewards/total_composite/std': 0.07253486663103104, 'reward': 0.30703824758529663, 'reward_std': 0.07253486663103104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17567403614521027, 'sampling/sampling_logp_difference/max': 2.484437942504883, 'sampling/importance_sampling_ratio/min': 0.08337239921092987, 'sampling/importance_sampling_ratio/mean': 1.0124120712280273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2451279684901237, 'clip_ratio/low_mean': 0.04838709533214569, 'clip_ratio/low_min': 0.04838709533214569, 'clip_ratio/high_mean': 0.130453669000417, 'clip_ratio/high_max': 0.130453669000417, 'clip_ratio/region_mean': 0.17884076433256269, 'reward_total_mean': 0.30703824758529663, 'reward_meter_mean': 0.8986762762069702, 'reward_meter_std': 0.12459779530763626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9183171987533569, 'reward_repeat_penalty_std': 0.10537010431289673, 'reward_repeat_floor_mean': 0.9916671514511108, 'reward_repeat_floor_std': 0.00818734522908926, 'reward_near_duplicate_penalty_mean': 0.9737309217453003, 'reward_near_duplicate_penalty_std': 0.020731303840875626, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9720279574394226, 'reward_distinct_2_std': 0.03866533935070038, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.30703824758529663, 'reward_total_composite_std': 0.07253486663103104, 'epoch': 0.03}

 15%|█▌        | 509/3300 [1:51:22<9:48:46, 12.66s/it]INFO 04-12 17:49:48 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 510/3300 [1:51:30<8:46:19, 11.32s/it]2026-04-12 17:49:56,573 | INFO | train_grpo_train | metrics_logged mode=train step=510

                                                      
{'loss': 0.1634, 'grad_norm': 18.07653045654297, 'learning_rate': 8.457575757575758e-06, 'num_tokens': 946654.0, 'completions/mean_length': 28.125, 'completions/min_length': 22.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9876656532287598, 'rewards/meter/std': 0.01162785105407238, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9527714848518372, 'rewards/lexical_plausibility/std': 0.13358239829540253, 'rewards/judge_quality/mean': 0.6549999713897705, 'rewards/judge_quality/std': 0.3013303875923157, 'rewards/repeat_penalty/mean': 0.7437795996665955, 'rewards/repeat_penalty/std': 0.07384127378463745, 'rewards/repeat_floor/mean': 0.980575680732727, 'rewards/repeat_floor/std': 0.009911520406603813, 'rewards/near_duplicate_penalty/mean': 0.9632494449615479, 'rewards/near_duplicate_penalty/std': 0.06476728618144989, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.6190110445022583, 'rewards/total_composite/std': 0.3101833462715149, 'reward': 0.6190110445022583, 'reward_std': 0.3101833164691925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1761094182729721, 'sampling/sampling_logp_difference/max': 1.0814104080200195, 'sampling/importance_sampling_ratio/min': 0.3391168713569641, 'sampling/importance_sampling_ratio/mean': 1.0074492692947388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1120131835341454, 'clip_ratio/low_mean': 0.06553030479699373, 'clip_ratio/low_min': 0.06553030479699373, 'clip_ratio/high_mean': 0.08131270948797464, 'clip_ratio/high_max': 0.08131270948797464, 'clip_ratio/region_mean': 0.14684301428496838, 'reward_total_mean': 0.6190110445022583, 'reward_meter_mean': 0.9876656532287598, 'reward_meter_std': 0.01162785105407238, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9527714848518372, 'reward_lexical_plausibility_std': 0.13358239829540253, 'reward_repeat_penalty_mean': 0.7437795996665955, 'reward_repeat_penalty_std': 0.07384127378463745, 'reward_repeat_floor_mean': 0.980575680732727, 'reward_repeat_floor_std': 0.009911520406603813, 'reward_near_duplicate_penalty_mean': 0.9632494449615479, 'reward_near_duplicate_penalty_std': 0.06476728618144989, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.6549999713897705, 'reward_judge_quality_std': 0.3013303875923157, 'reward_total_composite_mean': 0.6190110445022583, 'reward_total_composite_std': 0.3101833462715149, 'epoch': 0.03}

 15%|█▌        | 510/3300 [1:51:30<8:46:19, 11.32s/it]INFO 04-12 17:49:56 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 511/3300 [1:51:45<9:39:33, 12.47s/it]2026-04-12 17:50:11,725 | INFO | train_grpo_train | metrics_logged mode=train step=511

                                                      
{'loss': -0.0887, 'grad_norm': 5.195733070373535, 'learning_rate': 8.454545454545455e-06, 'num_tokens': 948639.0, 'completions/mean_length': 140.125, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 87.00000762939453, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7908278703689575, 'rewards/meter/std': 0.17973321676254272, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9310750365257263, 'rewards/lexical_plausibility/std': 0.16537347435951233, 'rewards/judge_quality/mean': 0.41624999046325684, 'rewards/judge_quality/std': 0.22859430313110352, 'rewards/repeat_penalty/mean': 0.7995262145996094, 'rewards/repeat_penalty/std': 0.23734745383262634, 'rewards/repeat_floor/mean': 0.9780957698822021, 'rewards/repeat_floor/std': 0.026306159794330597, 'rewards/near_duplicate_penalty/mean': 0.9527913928031921, 'rewards/near_duplicate_penalty/std': 0.05639262869954109, 'rewards/opening_diversity/mean': 0.9460227489471436, 'rewards/opening_diversity/std': 0.10159406810998917, 'rewards/distinct_2/mean': 0.8712852001190186, 'rewards/distinct_2/std': 0.14188195765018463, 'rewards/total_composite/mean': 0.31442081928253174, 'rewards/total_composite/std': 0.21579216420650482, 'reward': 0.31442081928253174, 'reward_std': 0.21579214930534363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16103772819042206, 'sampling/sampling_logp_difference/max': 1.7747416496276855, 'sampling/importance_sampling_ratio/min': 0.16952724754810333, 'sampling/importance_sampling_ratio/mean': 1.0292243957519531, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1683842241764069, 'clip_ratio/low_mean': 0.057608501985669136, 'clip_ratio/low_min': 0.057608501985669136, 'clip_ratio/high_mean': 0.06842339457944036, 'clip_ratio/high_max': 0.06842339457944036, 'clip_ratio/region_mean': 0.1260318965651095, 'reward_total_mean': 0.31442081928253174, 'reward_meter_mean': 0.7908278703689575, 'reward_meter_std': 0.17973321676254272, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9310750365257263, 'reward_lexical_plausibility_std': 0.16537347435951233, 'reward_repeat_penalty_mean': 0.7995262145996094, 'reward_repeat_penalty_std': 0.23734745383262634, 'reward_repeat_floor_mean': 0.9780957698822021, 'reward_repeat_floor_std': 0.026306159794330597, 'reward_near_duplicate_penalty_mean': 0.9527913928031921, 'reward_near_duplicate_penalty_std': 0.05639262869954109, 'reward_opening_diversity_mean': 0.9460227489471436, 'reward_opening_diversity_std': 0.10159406810998917, 'reward_distinct_2_mean': 0.8712852001190186, 'reward_distinct_2_std': 0.14188195765018463, 'reward_judge_quality_mean': 0.41624999046325684, 'reward_judge_quality_std': 0.22859430313110352, 'reward_total_composite_mean': 0.31442081928253174, 'reward_total_composite_std': 0.21579216420650482, 'epoch': 0.03}

 15%|█▌        | 511/3300 [1:51:45<9:39:33, 12.47s/it]INFO 04-12 17:50:11 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 512/3300 [1:52:01<10:26:36, 13.49s/it]2026-04-12 17:50:27,582 | INFO | train_grpo_train | metrics_logged mode=train step=512

                                                       
{'loss': -0.1608, 'grad_norm': 3.0728647708892822, 'learning_rate': 8.451515151515151e-06, 'num_tokens': 951667.0, 'completions/mean_length': 310.5, 'completions/min_length': 222.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 243.33334350585938, 'completions/min_terminated_length': 222.0, 'completions/max_terminated_length': 271.0, 'rewards/meter/mean': 0.7266141176223755, 'rewards/meter/std': 0.23444527387619019, 'rewards/count_adherence/mean': 0.6625000238418579, 'rewards/count_adherence/std': 0.1060660183429718, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8987499475479126, 'rewards/count_floor/std': 0.03181980550289154, 'rewards/lexical_plausibility/mean': 0.913486659526825, 'rewards/lexical_plausibility/std': 0.16599012911319733, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.15678924322128296, 'rewards/repeat_penalty/std': 0.3314603269100189, 'rewards/repeat_floor/mean': 0.8448523283004761, 'rewards/repeat_floor/std': 0.07679145038127899, 'rewards/near_duplicate_penalty/mean': 0.5846582651138306, 'rewards/near_duplicate_penalty/std': 0.24416083097457886, 'rewards/opening_diversity/mean': 0.5224862694740295, 'rewards/opening_diversity/std': 0.31657326221466064, 'rewards/distinct_2/mean': 0.2867155075073242, 'rewards/distinct_2/std': 0.34327632188796997, 'rewards/total_composite/mean': 0.11474845558404922, 'rewards/total_composite/std': 0.09143712371587753, 'reward': 0.11474845558404922, 'reward_std': 0.09143711626529694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0956878662109375, 'sampling/sampling_logp_difference/max': 1.985860824584961, 'sampling/importance_sampling_ratio/min': 0.13726240396499634, 'sampling/importance_sampling_ratio/mean': 1.031825065612793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6953889355063438, 'clip_ratio/low_mean': 0.030452340375632048, 'clip_ratio/low_min': 0.030452340375632048, 'clip_ratio/high_mean': 0.01627675397321582, 'clip_ratio/high_max': 0.01627675397321582, 'clip_ratio/region_mean': 0.046729094348847866, 'reward_total_mean': 0.11474845558404922, 'reward_meter_mean': 0.7266141176223755, 'reward_meter_std': 0.23444527387619019, 'reward_count_adherence_mean': 0.6625000238418579, 'reward_count_adherence_std': 0.1060660183429718, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8987499475479126, 'reward_count_floor_std': 0.03181980550289154, 'reward_lexical_plausibility_mean': 0.913486659526825, 'reward_lexical_plausibility_std': 0.16599012911319733, 'reward_repeat_penalty_mean': 0.15678924322128296, 'reward_repeat_penalty_std': 0.3314603269100189, 'reward_repeat_floor_mean': 0.8448523283004761, 'reward_repeat_floor_std': 0.07679145038127899, 'reward_near_duplicate_penalty_mean': 0.5846582651138306, 'reward_near_duplicate_penalty_std': 0.24416083097457886, 'reward_opening_diversity_mean': 0.5224862694740295, 'reward_opening_diversity_std': 0.31657326221466064, 'reward_distinct_2_mean': 0.2867155075073242, 'reward_distinct_2_std': 0.34327632188796997, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.11474845558404922, 'reward_total_composite_std': 0.09143712371587753, 'epoch': 0.03}

 16%|█▌        | 512/3300 [1:52:01<10:26:36, 13.49s/it]INFO 04-12 17:50:27 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 513/3300 [1:52:11<9:37:51, 12.44s/it] 2026-04-12 17:50:37,581 | INFO | train_grpo_train | metrics_logged mode=train step=513

                                                      
{'loss': 0.0787, 'grad_norm': 12.029963493347168, 'learning_rate': 8.44848484848485e-06, 'num_tokens': 954172.0, 'completions/mean_length': 138.125, 'completions/min_length': 122.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.125, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.7376955151557922, 'rewards/meter/std': 0.16614893078804016, 'rewards/count_adherence/mean': 0.660714328289032, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8982142806053162, 'rewards/count_floor/std': 0.022180693224072456, 'rewards/lexical_plausibility/mean': 0.9972826242446899, 'rewards/lexical_plausibility/std': 0.007685941644012928, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.6452541351318359, 'rewards/repeat_penalty/std': 0.25342443585395813, 'rewards/repeat_floor/mean': 0.9669176340103149, 'rewards/repeat_floor/std': 0.02324865385890007, 'rewards/near_duplicate_penalty/mean': 0.9520882368087769, 'rewards/near_duplicate_penalty/std': 0.03388504683971405, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7122707962989807, 'rewards/distinct_2/std': 0.21064119040966034, 'rewards/total_composite/mean': 0.24093817174434662, 'rewards/total_composite/std': 0.07627732306718826, 'reward': 0.24093817174434662, 'reward_std': 0.07627731561660767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19718462228775024, 'sampling/sampling_logp_difference/max': 2.534458875656128, 'sampling/importance_sampling_ratio/min': 0.07930462062358856, 'sampling/importance_sampling_ratio/mean': 1.0061147212982178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.074821561574936, 'clip_ratio/low_mean': 0.06576649006456137, 'clip_ratio/low_min': 0.06576649006456137, 'clip_ratio/high_mean': 0.10233640298247337, 'clip_ratio/high_max': 0.10233640298247337, 'clip_ratio/region_mean': 0.16810289304703474, 'reward_total_mean': 0.24093817174434662, 'reward_meter_mean': 0.7376955151557922, 'reward_meter_std': 0.16614893078804016, 'reward_count_adherence_mean': 0.660714328289032, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8982142806053162, 'reward_count_floor_std': 0.022180693224072456, 'reward_lexical_plausibility_mean': 0.9972826242446899, 'reward_lexical_plausibility_std': 0.007685941644012928, 'reward_repeat_penalty_mean': 0.6452541351318359, 'reward_repeat_penalty_std': 0.25342443585395813, 'reward_repeat_floor_mean': 0.9669176340103149, 'reward_repeat_floor_std': 0.02324865385890007, 'reward_near_duplicate_penalty_mean': 0.9520882368087769, 'reward_near_duplicate_penalty_std': 0.03388504683971405, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7122707962989807, 'reward_distinct_2_std': 0.21064119040966034, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.24093817174434662, 'reward_total_composite_std': 0.07627732306718826, 'epoch': 0.03}

 16%|█▌        | 513/3300 [1:52:11<9:37:51, 12.44s/it]INFO 04-12 17:50:37 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 514/3300 [1:52:25<10:03:09, 12.99s/it]2026-04-12 17:50:51,858 | INFO | train_grpo_train | metrics_logged mode=train step=514

                                                       
{'loss': -0.1644, 'grad_norm': 3.8267667293548584, 'learning_rate': 8.445454545454547e-06, 'num_tokens': 956592.0, 'completions/mean_length': 159.5, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 109.14286041259766, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.8357934355735779, 'rewards/meter/std': 0.21477803587913513, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9632163047790527, 'rewards/lexical_plausibility/std': 0.10404002666473389, 'rewards/judge_quality/mean': 0.4749999940395355, 'rewards/judge_quality/std': 0.29051679372787476, 'rewards/repeat_penalty/mean': 0.9527912139892578, 'rewards/repeat_penalty/std': 0.04206263646483421, 'rewards/repeat_floor/mean': 0.9949902892112732, 'rewards/repeat_floor/std': 0.004211359657347202, 'rewards/near_duplicate_penalty/mean': 0.9882081151008606, 'rewards/near_duplicate_penalty/std': 0.007910761050879955, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9639891386032104, 'rewards/distinct_2/std': 0.03702307492494583, 'rewards/total_composite/mean': 0.346296101808548, 'rewards/total_composite/std': 0.14345581829547882, 'reward': 0.346296101808548, 'reward_std': 0.14345581829547882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15752151608467102, 'sampling/sampling_logp_difference/max': 1.815791130065918, 'sampling/importance_sampling_ratio/min': 0.16270913183689117, 'sampling/importance_sampling_ratio/mean': 1.029537558555603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0509411171078682, 'clip_ratio/low_mean': 0.06316639482975006, 'clip_ratio/low_min': 0.06316639482975006, 'clip_ratio/high_mean': 0.06954278890043497, 'clip_ratio/high_max': 0.06954278890043497, 'clip_ratio/region_mean': 0.13270918373018503, 'reward_total_mean': 0.346296101808548, 'reward_meter_mean': 0.8357934355735779, 'reward_meter_std': 0.21477803587913513, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9632163047790527, 'reward_lexical_plausibility_std': 0.10404002666473389, 'reward_repeat_penalty_mean': 0.9527912139892578, 'reward_repeat_penalty_std': 0.04206263646483421, 'reward_repeat_floor_mean': 0.9949902892112732, 'reward_repeat_floor_std': 0.004211359657347202, 'reward_near_duplicate_penalty_mean': 0.9882081151008606, 'reward_near_duplicate_penalty_std': 0.007910761050879955, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9639891386032104, 'reward_distinct_2_std': 0.03702307492494583, 'reward_judge_quality_mean': 0.4749999940395355, 'reward_judge_quality_std': 0.29051679372787476, 'reward_total_composite_mean': 0.346296101808548, 'reward_total_composite_std': 0.14345581829547882, 'epoch': 0.03}

 16%|█▌        | 514/3300 [1:52:25<10:03:09, 12.99s/it]INFO 04-12 17:50:52 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 515/3300 [1:52:35<9:15:33, 11.97s/it] 2026-04-12 17:51:01,442 | INFO | train_grpo_train | metrics_logged mode=train step=515

                                                      
{'loss': 0.0591, 'grad_norm': 8.03563404083252, 'learning_rate': 8.442424242424243e-06, 'num_tokens': 958898.0, 'completions/mean_length': 105.25, 'completions/min_length': 96.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.25, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9241654872894287, 'rewards/meter/std': 0.051645684987306595, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9925000071525574, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.8615190386772156, 'rewards/repeat_penalty/std': 0.11042682081460953, 'rewards/repeat_floor/mean': 0.9857146739959717, 'rewards/repeat_floor/std': 0.010651892982423306, 'rewards/near_duplicate_penalty/mean': 0.969488263130188, 'rewards/near_duplicate_penalty/std': 0.017097249627113342, 'rewards/opening_diversity/mean': 0.971875011920929, 'rewards/opening_diversity/std': 0.05250425636768341, 'rewards/distinct_2/mean': 0.9108771085739136, 'rewards/distinct_2/std': 0.0654713362455368, 'rewards/total_composite/mean': 0.43101412057876587, 'rewards/total_composite/std': 0.08232199400663376, 'reward': 0.43101412057876587, 'reward_std': 0.08232198655605316, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.157629132270813, 'sampling/sampling_logp_difference/max': 5.239917278289795, 'sampling/importance_sampling_ratio/min': 0.005300695076584816, 'sampling/importance_sampling_ratio/mean': 1.0178728103637695, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9227953553199768, 'clip_ratio/low_mean': 0.09270178060978651, 'clip_ratio/low_min': 0.09270178060978651, 'clip_ratio/high_mean': 0.01417525764554739, 'clip_ratio/high_max': 0.01417525764554739, 'clip_ratio/region_mean': 0.1068770382553339, 'reward_total_mean': 0.43101412057876587, 'reward_meter_mean': 0.9241654872894287, 'reward_meter_std': 0.051645684987306595, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9925000071525574, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8615190386772156, 'reward_repeat_penalty_std': 0.11042682081460953, 'reward_repeat_floor_mean': 0.9857146739959717, 'reward_repeat_floor_std': 0.010651892982423306, 'reward_near_duplicate_penalty_mean': 0.969488263130188, 'reward_near_duplicate_penalty_std': 0.017097249627113342, 'reward_opening_diversity_mean': 0.971875011920929, 'reward_opening_diversity_std': 0.05250425636768341, 'reward_distinct_2_mean': 0.9108771085739136, 'reward_distinct_2_std': 0.0654713362455368, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.43101412057876587, 'reward_total_composite_std': 0.08232199400663376, 'epoch': 0.03}

 16%|█▌        | 515/3300 [1:52:35<9:15:33, 11.97s/it]INFO 04-12 17:51:01 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 516/3300 [1:52:45<8:49:52, 11.42s/it]2026-04-12 17:51:11,579 | INFO | train_grpo_train | metrics_logged mode=train step=516

                                                      
{'loss': 0.0255, 'grad_norm': 5.3206987380981445, 'learning_rate': 8.43939393939394e-06, 'num_tokens': 961458.0, 'completions/mean_length': 140.0, 'completions/min_length': 132.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.0, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.9921877384185791, 'rewards/meter/std': 0.0019624331034719944, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.331486314535141, 'rewards/repeat_penalty/std': 0.2879042327404022, 'rewards/repeat_floor/mean': 0.9014939069747925, 'rewards/repeat_floor/std': 0.06623192131519318, 'rewards/near_duplicate_penalty/mean': 0.7780974507331848, 'rewards/near_duplicate_penalty/std': 0.16105632483959198, 'rewards/opening_diversity/mean': 0.7230113744735718, 'rewards/opening_diversity/std': 0.3136232793331146, 'rewards/distinct_2/mean': 0.4599849581718445, 'rewards/distinct_2/std': 0.3084188997745514, 'rewards/total_composite/mean': 0.2763414978981018, 'rewards/total_composite/std': 0.0654524564743042, 'reward': 0.2763414978981018, 'reward_std': 0.0654524639248848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11591927707195282, 'sampling/sampling_logp_difference/max': 1.9515008926391602, 'sampling/importance_sampling_ratio/min': 0.14206069707870483, 'sampling/importance_sampling_ratio/mean': 1.0085686445236206, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8303218744695187, 'clip_ratio/low_mean': 0.014263114426285028, 'clip_ratio/low_min': 0.014263114426285028, 'clip_ratio/high_mean': 0.10018174815922976, 'clip_ratio/high_max': 0.10018174815922976, 'clip_ratio/region_mean': 0.11444486258551478, 'reward_total_mean': 0.2763414978981018, 'reward_meter_mean': 0.9921877384185791, 'reward_meter_std': 0.0019624331034719944, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.331486314535141, 'reward_repeat_penalty_std': 0.2879042327404022, 'reward_repeat_floor_mean': 0.9014939069747925, 'reward_repeat_floor_std': 0.06623192131519318, 'reward_near_duplicate_penalty_mean': 0.7780974507331848, 'reward_near_duplicate_penalty_std': 0.16105632483959198, 'reward_opening_diversity_mean': 0.7230113744735718, 'reward_opening_diversity_std': 0.3136232793331146, 'reward_distinct_2_mean': 0.4599849581718445, 'reward_distinct_2_std': 0.3084188997745514, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.2763414978981018, 'reward_total_composite_std': 0.0654524564743042, 'epoch': 0.03}

 16%|█▌        | 516/3300 [1:52:45<8:49:52, 11.42s/it]INFO 04-12 17:51:11 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 517/3300 [1:52:53<8:01:04, 10.37s/it]2026-04-12 17:51:19,509 | INFO | train_grpo_train | metrics_logged mode=train step=517

                                                      
{'loss': 0.2659, 'grad_norm': 30.32703971862793, 'learning_rate': 8.436363636363637e-06, 'num_tokens': 962874.0, 'completions/mean_length': 27.0, 'completions/min_length': 23.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.8983370661735535, 'rewards/meter/std': 0.22539734840393066, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7975000143051147, 'rewards/judge_quality/std': 0.2471408098936081, 'rewards/repeat_penalty/mean': 0.733729362487793, 'rewards/repeat_penalty/std': 0.12819167971611023, 'rewards/repeat_floor/mean': 0.9773709177970886, 'rewards/repeat_floor/std': 0.016670571640133858, 'rewards/near_duplicate_penalty/mean': 0.936639130115509, 'rewards/near_duplicate_penalty/std': 0.09639061987400055, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.727913498878479, 'rewards/total_composite/std': 0.30518433451652527, 'reward': 0.727913498878479, 'reward_std': 0.30518433451652527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15355220437049866, 'sampling/sampling_logp_difference/max': 1.2723546028137207, 'sampling/importance_sampling_ratio/min': 0.28017115592956543, 'sampling/importance_sampling_ratio/mean': 1.0076162815093994, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9819388464093208, 'clip_ratio/low_mean': 0.053321680054068565, 'clip_ratio/low_min': 0.053321680054068565, 'clip_ratio/high_mean': 0.11706173326820135, 'clip_ratio/high_max': 0.11706173326820135, 'clip_ratio/region_mean': 0.17038341332226992, 'reward_total_mean': 0.727913498878479, 'reward_meter_mean': 0.8983370661735535, 'reward_meter_std': 0.22539734840393066, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.733729362487793, 'reward_repeat_penalty_std': 0.12819167971611023, 'reward_repeat_floor_mean': 0.9773709177970886, 'reward_repeat_floor_std': 0.016670571640133858, 'reward_near_duplicate_penalty_mean': 0.936639130115509, 'reward_near_duplicate_penalty_std': 0.09639061987400055, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7975000143051147, 'reward_judge_quality_std': 0.2471408098936081, 'reward_total_composite_mean': 0.727913498878479, 'reward_total_composite_std': 0.30518433451652527, 'epoch': 0.03}

 16%|█▌        | 517/3300 [1:52:53<8:01:04, 10.37s/it]INFO 04-12 17:51:19 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 518/3300 [1:53:02<7:47:20, 10.08s/it]2026-04-12 17:51:28,904 | INFO | train_grpo_train | metrics_logged mode=train step=518

                                                      
{'loss': 0.1184, 'grad_norm': 13.314927101135254, 'learning_rate': 8.433333333333334e-06, 'num_tokens': 964540.0, 'completions/mean_length': 49.25, 'completions/min_length': 40.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.49174201488494873, 'rewards/meter/std': 0.48091262578964233, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9479242563247681, 'rewards/lexical_plausibility/std': 0.11055175960063934, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9025670289993286, 'rewards/repeat_penalty/std': 0.11271899193525314, 'rewards/repeat_floor/mean': 0.9901441931724548, 'rewards/repeat_floor/std': 0.009752021171152592, 'rewards/near_duplicate_penalty/mean': 0.9737732410430908, 'rewards/near_duplicate_penalty/std': 0.025190871208906174, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9550356864929199, 'rewards/distinct_2/std': 0.048800356686115265, 'rewards/total_composite/mean': 0.2097613513469696, 'rewards/total_composite/std': 0.20769169926643372, 'reward': 0.2097613513469696, 'reward_std': 0.20769169926643372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18047697842121124, 'sampling/sampling_logp_difference/max': 1.7908518314361572, 'sampling/importance_sampling_ratio/min': 0.2603965699672699, 'sampling/importance_sampling_ratio/mean': 1.0249801874160767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.371830441057682, 'clip_ratio/low_mean': 0.060898084193468094, 'clip_ratio/low_min': 0.060898084193468094, 'clip_ratio/high_mean': 0.0749156316742301, 'clip_ratio/high_max': 0.0749156316742301, 'clip_ratio/region_mean': 0.1358137158676982, 'reward_total_mean': 0.2097613513469696, 'reward_meter_mean': 0.49174201488494873, 'reward_meter_std': 0.48091262578964233, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9479242563247681, 'reward_lexical_plausibility_std': 0.11055175960063934, 'reward_repeat_penalty_mean': 0.9025670289993286, 'reward_repeat_penalty_std': 0.11271899193525314, 'reward_repeat_floor_mean': 0.9901441931724548, 'reward_repeat_floor_std': 0.009752021171152592, 'reward_near_duplicate_penalty_mean': 0.9737732410430908, 'reward_near_duplicate_penalty_std': 0.025190871208906174, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9550356864929199, 'reward_distinct_2_std': 0.048800356686115265, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.2097613513469696, 'reward_total_composite_std': 0.20769169926643372, 'epoch': 0.03}

 16%|█▌        | 518/3300 [1:53:02<7:47:20, 10.08s/it]INFO 04-12 17:51:29 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 519/3300 [1:53:13<7:53:19, 10.21s/it]2026-04-12 17:51:39,430 | INFO | train_grpo_train | metrics_logged mode=train step=519

                                                      
{'loss': 0.0359, 'grad_norm': 12.719137191772461, 'learning_rate': 8.43030303030303e-06, 'num_tokens': 966497.0, 'completions/mean_length': 73.625, 'completions/min_length': 55.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.4631587862968445, 'rewards/meter/std': 0.3079971373081207, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.7922379970550537, 'rewards/repeat_penalty/std': 0.16282393038272858, 'rewards/repeat_floor/mean': 0.9772928953170776, 'rewards/repeat_floor/std': 0.019161473959684372, 'rewards/near_duplicate_penalty/mean': 0.9444105625152588, 'rewards/near_duplicate_penalty/std': 0.05331508442759514, 'rewards/opening_diversity/mean': 0.9508928656578064, 'rewards/opening_diversity/std': 0.09533105045557022, 'rewards/distinct_2/mean': 0.8730860352516174, 'rewards/distinct_2/std': 0.0932348221540451, 'rewards/total_composite/mean': 0.13563811779022217, 'rewards/total_composite/std': 0.09495692700147629, 'reward': 0.13563811779022217, 'reward_std': 0.09495692700147629, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1856628954410553, 'sampling/sampling_logp_difference/max': 1.680562973022461, 'sampling/importance_sampling_ratio/min': 0.18626907467842102, 'sampling/importance_sampling_ratio/mean': 1.029004454612732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2884924337267876, 'clip_ratio/low_mean': 0.07619927451014519, 'clip_ratio/low_min': 0.07619927451014519, 'clip_ratio/high_mean': 0.12307463772594929, 'clip_ratio/high_max': 0.12307463772594929, 'clip_ratio/region_mean': 0.19927391223609447, 'reward_total_mean': 0.13563811779022217, 'reward_meter_mean': 0.4631587862968445, 'reward_meter_std': 0.3079971373081207, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7922379970550537, 'reward_repeat_penalty_std': 0.16282393038272858, 'reward_repeat_floor_mean': 0.9772928953170776, 'reward_repeat_floor_std': 0.019161473959684372, 'reward_near_duplicate_penalty_mean': 0.9444105625152588, 'reward_near_duplicate_penalty_std': 0.05331508442759514, 'reward_opening_diversity_mean': 0.9508928656578064, 'reward_opening_diversity_std': 0.09533105045557022, 'reward_distinct_2_mean': 0.8730860352516174, 'reward_distinct_2_std': 0.0932348221540451, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.13563811779022217, 'reward_total_composite_std': 0.09495692700147629, 'epoch': 0.03}

 16%|█▌        | 519/3300 [1:53:13<7:53:19, 10.21s/it]INFO 04-12 17:51:39 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 520/3300 [1:53:21<7:30:18,  9.72s/it]2026-04-12 17:51:47,993 | INFO | train_grpo_train | metrics_logged mode=train step=520

                                                      
{'loss': 0.1227, 'grad_norm': 15.349052429199219, 'learning_rate': 8.427272727272729e-06, 'num_tokens': 968065.0, 'completions/mean_length': 44.0, 'completions/min_length': 38.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.4164927303791046, 'rewards/meter/std': 0.3023623824119568, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9845391511917114, 'rewards/lexical_plausibility/std': 0.028630122542381287, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.2707628309726715, 'rewards/repeat_penalty/mean': 0.9862516522407532, 'rewards/repeat_penalty/std': 0.012122338637709618, 'rewards/repeat_floor/mean': 0.997937798500061, 'rewards/repeat_floor/std': 0.001818351331166923, 'rewards/near_duplicate_penalty/mean': 0.9862516522407532, 'rewards/near_duplicate_penalty/std': 0.012122338637709618, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3138221502304077, 'rewards/total_composite/std': 0.28780755400657654, 'reward': 0.3138221502304077, 'reward_std': 0.28780752420425415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2093144655227661, 'sampling/sampling_logp_difference/max': 2.6798653602600098, 'sampling/importance_sampling_ratio/min': 0.06857238709926605, 'sampling/importance_sampling_ratio/mean': 1.0018255710601807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4446628764271736, 'clip_ratio/low_mean': 0.12115802429616451, 'clip_ratio/low_min': 0.12115802429616451, 'clip_ratio/high_mean': 0.08064408600330353, 'clip_ratio/high_max': 0.08064408600330353, 'clip_ratio/region_mean': 0.20180211029946804, 'reward_total_mean': 0.3138221502304077, 'reward_meter_mean': 0.4164927303791046, 'reward_meter_std': 0.3023623824119568, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9845391511917114, 'reward_lexical_plausibility_std': 0.028630122542381287, 'reward_repeat_penalty_mean': 0.9862516522407532, 'reward_repeat_penalty_std': 0.012122338637709618, 'reward_repeat_floor_mean': 0.997937798500061, 'reward_repeat_floor_std': 0.001818351331166923, 'reward_near_duplicate_penalty_mean': 0.9862516522407532, 'reward_near_duplicate_penalty_std': 0.012122338637709618, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.2707628309726715, 'reward_total_composite_mean': 0.3138221502304077, 'reward_total_composite_std': 0.28780755400657654, 'epoch': 0.03}

 16%|█▌        | 520/3300 [1:53:21<7:30:18,  9.72s/it]INFO 04-12 17:51:48 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 521/3300 [1:53:30<7:13:38,  9.36s/it]2026-04-12 17:51:56,525 | INFO | train_grpo_train | metrics_logged mode=train step=521

                                                      
{'loss': 0.0643, 'grad_norm': 14.271038055419922, 'learning_rate': 8.424242424242425e-06, 'num_tokens': 969558.0, 'completions/mean_length': 40.625, 'completions/min_length': 38.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.852173924446106, 'rewards/meter/std': 0.30694180727005005, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.3285357356071472, 'rewards/repeat_penalty/mean': 0.8681364059448242, 'rewards/repeat_penalty/std': 0.1722218096256256, 'rewards/repeat_floor/mean': 0.9858099222183228, 'rewards/repeat_floor/std': 0.01767108216881752, 'rewards/near_duplicate_penalty/mean': 0.9533722996711731, 'rewards/near_duplicate_penalty/std': 0.054576653987169266, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9617122411727905, 'rewards/distinct_2/std': 0.05567097291350365, 'rewards/total_composite/mean': 0.4916231632232666, 'rewards/total_composite/std': 0.3457225561141968, 'reward': 0.4916231632232666, 'reward_std': 0.34572258591651917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17721252143383026, 'sampling/sampling_logp_difference/max': 1.1589140892028809, 'sampling/importance_sampling_ratio/min': 0.3138267993927002, 'sampling/importance_sampling_ratio/mean': 1.0233261585235596, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1198498606681824, 'clip_ratio/low_mean': 0.12171227484941483, 'clip_ratio/low_min': 0.12171227484941483, 'clip_ratio/high_mean': 0.060016026720404625, 'clip_ratio/high_max': 0.060016026720404625, 'clip_ratio/region_mean': 0.18172830156981945, 'reward_total_mean': 0.4916231632232666, 'reward_meter_mean': 0.852173924446106, 'reward_meter_std': 0.30694180727005005, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.8681364059448242, 'reward_repeat_penalty_std': 0.1722218096256256, 'reward_repeat_floor_mean': 0.9858099222183228, 'reward_repeat_floor_std': 0.01767108216881752, 'reward_near_duplicate_penalty_mean': 0.9533722996711731, 'reward_near_duplicate_penalty_std': 0.054576653987169266, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9617122411727905, 'reward_distinct_2_std': 0.05567097291350365, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.3285357356071472, 'reward_total_composite_mean': 0.4916231632232666, 'reward_total_composite_std': 0.3457225561141968, 'epoch': 0.03}

 16%|█▌        | 521/3300 [1:53:30<7:13:38,  9.36s/it]INFO 04-12 17:51:56 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 522/3300 [1:53:40<7:20:16,  9.51s/it]2026-04-12 17:52:06,375 | INFO | train_grpo_train | metrics_logged mode=train step=522

                                                      
{'loss': 0.0002, 'grad_norm': 12.01123046875, 'learning_rate': 8.421212121212122e-06, 'num_tokens': 971670.0, 'completions/mean_length': 96.0, 'completions/min_length': 85.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.9132922291755676, 'rewards/meter/std': 0.21457339823246002, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8785302639007568, 'rewards/repeat_penalty/std': 0.14265145361423492, 'rewards/repeat_floor/mean': 0.9867973327636719, 'rewards/repeat_floor/std': 0.014486290514469147, 'rewards/near_duplicate_penalty/mean': 0.9676364064216614, 'rewards/near_duplicate_penalty/std': 0.021895958110690117, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9124506711959839, 'rewards/distinct_2/std': 0.12714537978172302, 'rewards/total_composite/mean': 0.3384040296077728, 'rewards/total_composite/std': 0.08730526268482208, 'reward': 0.3384040296077728, 'reward_std': 0.08730526268482208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1593395620584488, 'sampling/sampling_logp_difference/max': 1.6513586044311523, 'sampling/importance_sampling_ratio/min': 0.1917891651391983, 'sampling/importance_sampling_ratio/mean': 1.0177624225616455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1946695148944855, 'clip_ratio/low_mean': 0.05445515923202038, 'clip_ratio/low_min': 0.05445515923202038, 'clip_ratio/high_mean': 0.10272705554962158, 'clip_ratio/high_max': 0.10272705554962158, 'clip_ratio/region_mean': 0.15718221478164196, 'reward_total_mean': 0.3384040296077728, 'reward_meter_mean': 0.9132922291755676, 'reward_meter_std': 0.21457339823246002, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8785302639007568, 'reward_repeat_penalty_std': 0.14265145361423492, 'reward_repeat_floor_mean': 0.9867973327636719, 'reward_repeat_floor_std': 0.014486290514469147, 'reward_near_duplicate_penalty_mean': 0.9676364064216614, 'reward_near_duplicate_penalty_std': 0.021895958110690117, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9124506711959839, 'reward_distinct_2_std': 0.12714537978172302, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3384040296077728, 'reward_total_composite_std': 0.08730526268482208, 'epoch': 0.03}

 16%|█▌        | 522/3300 [1:53:40<7:20:16,  9.51s/it]INFO 04-12 17:52:06 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 523/3300 [1:53:49<7:11:22,  9.32s/it]2026-04-12 17:52:15,258 | INFO | train_grpo_train | metrics_logged mode=train step=523

                                                      
{'loss': 0.062, 'grad_norm': 15.831817626953125, 'learning_rate': 8.418181818181819e-06, 'num_tokens': 973284.0, 'completions/mean_length': 35.75, 'completions/min_length': 32.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8376367092132568, 'rewards/meter/std': 0.2692849338054657, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.9702938795089722, 'rewards/repeat_penalty/std': 0.05283456668257713, 'rewards/repeat_floor/mean': 0.996000349521637, 'rewards/repeat_floor/std': 0.006708947941660881, 'rewards/near_duplicate_penalty/mean': 0.9791051149368286, 'rewards/near_duplicate_penalty/std': 0.030114414170384407, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.49114301800727844, 'rewards/total_composite/std': 0.28003180027008057, 'reward': 0.49114301800727844, 'reward_std': 0.2800317704677582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14815272390842438, 'sampling/sampling_logp_difference/max': 1.2352204322814941, 'sampling/importance_sampling_ratio/min': 0.29077064990997314, 'sampling/importance_sampling_ratio/mean': 1.0314936637878418, 'sampling/importance_sampling_ratio/max': 1.913636565208435, 'entropy': 1.0560617744922638, 'clip_ratio/low_mean': 0.11194766499102116, 'clip_ratio/low_min': 0.11194766499102116, 'clip_ratio/high_mean': 0.037913603708148, 'clip_ratio/high_max': 0.037913603708148, 'clip_ratio/region_mean': 0.14986126869916916, 'reward_total_mean': 0.49114301800727844, 'reward_meter_mean': 0.8376367092132568, 'reward_meter_std': 0.2692849338054657, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9702938795089722, 'reward_repeat_penalty_std': 0.05283456668257713, 'reward_repeat_floor_mean': 0.996000349521637, 'reward_repeat_floor_std': 0.006708947941660881, 'reward_near_duplicate_penalty_mean': 0.9791051149368286, 'reward_near_duplicate_penalty_std': 0.030114414170384407, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.49114301800727844, 'reward_total_composite_std': 0.28003180027008057, 'epoch': 0.03}

 16%|█▌        | 523/3300 [1:53:49<7:11:22,  9.32s/it]INFO 04-12 17:52:15 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 524/3300 [1:53:57<6:55:10,  8.97s/it]2026-04-12 17:52:23,421 | INFO | train_grpo_train | metrics_logged mode=train step=524

                                                      
{'loss': 0.0541, 'grad_norm': 16.983884811401367, 'learning_rate': 8.415151515151516e-06, 'num_tokens': 974885.0, 'completions/mean_length': 42.125, 'completions/min_length': 37.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.49647974967956543, 'rewards/meter/std': 0.40011730790138245, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.8025000095367432, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.9699174761772156, 'rewards/repeat_penalty/std': 0.03282570466399193, 'rewards/repeat_floor/mean': 0.9960370659828186, 'rewards/repeat_floor/std': 0.004304138477891684, 'rewards/near_duplicate_penalty/mean': 0.9790749549865723, 'rewards/near_duplicate_penalty/std': 0.029062652960419655, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.3848479390144348, 'rewards/total_composite/std': 0.34055766463279724, 'reward': 0.3848479390144348, 'reward_std': 0.34055769443511963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17654244601726532, 'sampling/sampling_logp_difference/max': 3.164750814437866, 'sampling/importance_sampling_ratio/min': 0.04222466051578522, 'sampling/importance_sampling_ratio/mean': 1.0286885499954224, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2037063613533974, 'clip_ratio/low_mean': 0.09554027020931244, 'clip_ratio/low_min': 0.09554027020931244, 'clip_ratio/high_mean': 0.06381778419017792, 'clip_ratio/high_max': 0.06381778419017792, 'clip_ratio/region_mean': 0.15935805439949036, 'reward_total_mean': 0.3848479390144348, 'reward_meter_mean': 0.49647974967956543, 'reward_meter_std': 0.40011730790138245, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9699174761772156, 'reward_repeat_penalty_std': 0.03282570466399193, 'reward_repeat_floor_mean': 0.9960370659828186, 'reward_repeat_floor_std': 0.004304138477891684, 'reward_near_duplicate_penalty_mean': 0.9790749549865723, 'reward_near_duplicate_penalty_std': 0.029062652960419655, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.8025000095367432, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.3848479390144348, 'reward_total_composite_std': 0.34055766463279724, 'epoch': 0.03}

 16%|█▌        | 524/3300 [1:53:57<6:55:10,  8.97s/it]INFO 04-12 17:52:23 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 525/3300 [1:54:05<6:47:28,  8.81s/it]2026-04-12 17:52:31,850 | INFO | train_grpo_train | metrics_logged mode=train step=525

                                                      
{'loss': 0.0999, 'grad_norm': 15.229125022888184, 'learning_rate': 8.412121212121212e-06, 'num_tokens': 976413.0, 'completions/mean_length': 37.0, 'completions/min_length': 34.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.5796794295310974, 'rewards/meter/std': 0.43055519461631775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.9137818813323975, 'rewards/repeat_penalty/std': 0.10843636095523834, 'rewards/repeat_floor/mean': 0.990498423576355, 'rewards/repeat_floor/std': 0.008728833869099617, 'rewards/near_duplicate_penalty/mean': 0.9619768261909485, 'rewards/near_duplicate_penalty/std': 0.025421274825930595, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.039564959704875946, 'rewards/total_composite/mean': 0.31257349252700806, 'rewards/total_composite/std': 0.37147992849349976, 'reward': 0.31257349252700806, 'reward_std': 0.37147992849349976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15412814915180206, 'sampling/sampling_logp_difference/max': 1.3310840129852295, 'sampling/importance_sampling_ratio/min': 0.2641907036304474, 'sampling/importance_sampling_ratio/mean': 1.0021681785583496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8568323999643326, 'clip_ratio/low_mean': 0.06261557387188077, 'clip_ratio/low_min': 0.06261557387188077, 'clip_ratio/high_mean': 0.07849761843681335, 'clip_ratio/high_max': 0.07849761843681335, 'clip_ratio/region_mean': 0.14111319230869412, 'reward_total_mean': 0.31257349252700806, 'reward_meter_mean': 0.5796794295310974, 'reward_meter_std': 0.43055519461631775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9137818813323975, 'reward_repeat_penalty_std': 0.10843636095523834, 'reward_repeat_floor_mean': 0.990498423576355, 'reward_repeat_floor_std': 0.008728833869099617, 'reward_near_duplicate_penalty_mean': 0.9619768261909485, 'reward_near_duplicate_penalty_std': 0.025421274825930595, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.039564959704875946, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.31257349252700806, 'reward_total_composite_std': 0.37147992849349976, 'epoch': 0.03}

 16%|█▌        | 525/3300 [1:54:05<6:47:28,  8.81s/it]INFO 04-12 17:52:32 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 526/3300 [1:54:15<7:02:37,  9.14s/it]2026-04-12 17:52:41,766 | INFO | train_grpo_train | metrics_logged mode=train step=526

                                                      
{'loss': 0.0467, 'grad_norm': 10.695037841796875, 'learning_rate': 8.40909090909091e-06, 'num_tokens': 978621.0, 'completions/mean_length': 96.0, 'completions/min_length': 80.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.4163585603237152, 'rewards/meter/std': 0.2186916470527649, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9775000214576721, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9287399053573608, 'rewards/repeat_penalty/std': 0.0394316203892231, 'rewards/repeat_floor/mean': 0.9922876954078674, 'rewards/repeat_floor/std': 0.004055831581354141, 'rewards/near_duplicate_penalty/mean': 0.9801652431488037, 'rewards/near_duplicate_penalty/std': 0.008899851702153683, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9473658204078674, 'rewards/distinct_2/std': 0.03438752517104149, 'rewards/total_composite/mean': 0.15805578231811523, 'rewards/total_composite/std': 0.0747164711356163, 'reward': 0.15805578231811523, 'reward_std': 0.0747164636850357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16614016890525818, 'sampling/sampling_logp_difference/max': 1.6688706874847412, 'sampling/importance_sampling_ratio/min': 0.18845976889133453, 'sampling/importance_sampling_ratio/mean': 1.003157615661621, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.111264280974865, 'clip_ratio/low_mean': 0.1024137856438756, 'clip_ratio/low_min': 0.1024137856438756, 'clip_ratio/high_mean': 0.07299947924911976, 'clip_ratio/high_max': 0.07299947924911976, 'clip_ratio/region_mean': 0.17541326489299536, 'reward_total_mean': 0.15805578231811523, 'reward_meter_mean': 0.4163585603237152, 'reward_meter_std': 0.2186916470527649, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9775000214576721, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9287399053573608, 'reward_repeat_penalty_std': 0.0394316203892231, 'reward_repeat_floor_mean': 0.9922876954078674, 'reward_repeat_floor_std': 0.004055831581354141, 'reward_near_duplicate_penalty_mean': 0.9801652431488037, 'reward_near_duplicate_penalty_std': 0.008899851702153683, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9473658204078674, 'reward_distinct_2_std': 0.03438752517104149, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.15805578231811523, 'reward_total_composite_std': 0.0747164711356163, 'epoch': 0.03}

 16%|█▌        | 526/3300 [1:54:15<7:02:37,  9.14s/it]INFO 04-12 17:52:42 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 527/3300 [1:54:24<7:00:55,  9.11s/it]2026-04-12 17:52:50,793 | INFO | train_grpo_train | metrics_logged mode=train step=527

                                                      
{'loss': 0.0454, 'grad_norm': 17.942092895507812, 'learning_rate': 8.406060606060606e-06, 'num_tokens': 980190.0, 'completions/mean_length': 35.125, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7994195222854614, 'rewards/meter/std': 0.33700355887413025, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9340277910232544, 'rewards/lexical_plausibility/std': 0.17564287781715393, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.8898806571960449, 'rewards/repeat_penalty/std': 0.17661099135875702, 'rewards/repeat_floor/mean': 0.9880690574645996, 'rewards/repeat_floor/std': 0.018163839355111122, 'rewards/near_duplicate_penalty/mean': 0.9567240476608276, 'rewards/near_duplicate_penalty/std': 0.05666833370923996, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9812270998954773, 'rewards/distinct_2/std': 0.034774426370859146, 'rewards/total_composite/mean': 0.36645856499671936, 'rewards/total_composite/std': 0.1453583687543869, 'reward': 0.36645856499671936, 'reward_std': 0.1453583836555481, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12800559401512146, 'sampling/sampling_logp_difference/max': 1.465423345565796, 'sampling/importance_sampling_ratio/min': 0.23098018765449524, 'sampling/importance_sampling_ratio/mean': 1.0225130319595337, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8144971951842308, 'clip_ratio/low_mean': 0.04513888992369175, 'clip_ratio/low_min': 0.04513888992369175, 'clip_ratio/high_mean': 0.08164018322713673, 'clip_ratio/high_max': 0.08164018322713673, 'clip_ratio/region_mean': 0.12677907315082848, 'reward_total_mean': 0.36645856499671936, 'reward_meter_mean': 0.7994195222854614, 'reward_meter_std': 0.33700355887413025, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9340277910232544, 'reward_lexical_plausibility_std': 0.17564287781715393, 'reward_repeat_penalty_mean': 0.8898806571960449, 'reward_repeat_penalty_std': 0.17661099135875702, 'reward_repeat_floor_mean': 0.9880690574645996, 'reward_repeat_floor_std': 0.018163839355111122, 'reward_near_duplicate_penalty_mean': 0.9567240476608276, 'reward_near_duplicate_penalty_std': 0.05666833370923996, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9812270998954773, 'reward_distinct_2_std': 0.034774426370859146, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.36645856499671936, 'reward_total_composite_std': 0.1453583687543869, 'epoch': 0.03}

 16%|█▌        | 527/3300 [1:54:24<7:00:55,  9.11s/it]INFO 04-12 17:52:51 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 528/3300 [1:54:38<8:05:03, 10.50s/it]2026-04-12 17:53:04,540 | INFO | train_grpo_train | metrics_logged mode=train step=528

                                                      
{'loss': -0.0509, 'grad_norm': 5.375563144683838, 'learning_rate': 8.403030303030304e-06, 'num_tokens': 981817.0, 'completions/mean_length': 110.375, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.000003814697266, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.3983202576637268, 'rewards/meter/std': 0.2766818106174469, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.945500373840332, 'rewards/lexical_plausibility/std': 0.14706578850746155, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8291400074958801, 'rewards/repeat_penalty/std': 0.12698650360107422, 'rewards/repeat_floor/mean': 0.978317379951477, 'rewards/repeat_floor/std': 0.01570338010787964, 'rewards/near_duplicate_penalty/mean': 0.9111659526824951, 'rewards/near_duplicate_penalty/std': 0.06265629827976227, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9071384072303772, 'rewards/distinct_2/std': 0.0984184741973877, 'rewards/total_composite/mean': 0.15317308902740479, 'rewards/total_composite/std': 0.10327749699354172, 'reward': 0.15317308902740479, 'reward_std': 0.10327748209238052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19312684237957, 'sampling/sampling_logp_difference/max': 1.3962161540985107, 'sampling/importance_sampling_ratio/min': 0.24753181636333466, 'sampling/importance_sampling_ratio/mean': 1.0329201221466064, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.378822848200798, 'clip_ratio/low_mean': 0.04297785647213459, 'clip_ratio/low_min': 0.04297785647213459, 'clip_ratio/high_mean': 0.11210514605045319, 'clip_ratio/high_max': 0.11210514605045319, 'clip_ratio/region_mean': 0.15508300252258778, 'reward_total_mean': 0.15317308902740479, 'reward_meter_mean': 0.3983202576637268, 'reward_meter_std': 0.2766818106174469, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.945500373840332, 'reward_lexical_plausibility_std': 0.14706578850746155, 'reward_repeat_penalty_mean': 0.8291400074958801, 'reward_repeat_penalty_std': 0.12698650360107422, 'reward_repeat_floor_mean': 0.978317379951477, 'reward_repeat_floor_std': 0.01570338010787964, 'reward_near_duplicate_penalty_mean': 0.9111659526824951, 'reward_near_duplicate_penalty_std': 0.06265629827976227, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9071384072303772, 'reward_distinct_2_std': 0.0984184741973877, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.15317308902740479, 'reward_total_composite_std': 0.10327749699354172, 'epoch': 0.03}

 16%|█▌        | 528/3300 [1:54:38<8:05:03, 10.50s/it]INFO 04-12 17:53:04 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 529/3300 [1:54:46<7:36:51,  9.89s/it]2026-04-12 17:53:13,015 | INFO | train_grpo_train | metrics_logged mode=train step=529

                                                      
{'loss': 0.0224, 'grad_norm': 17.020240783691406, 'learning_rate': 8.400000000000001e-06, 'num_tokens': 983604.0, 'completions/mean_length': 42.375, 'completions/min_length': 40.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6046037673950195, 'rewards/meter/std': 0.4366230070590973, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.9792095422744751, 'rewards/repeat_penalty/std': 0.024614235386252403, 'rewards/repeat_floor/mean': 0.9968814253807068, 'rewards/repeat_floor/std': 0.003692133817821741, 'rewards/near_duplicate_penalty/mean': 0.9792095422744751, 'rewards/near_duplicate_penalty/std': 0.024614235386252403, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5405032634735107, 'rewards/total_composite/std': 0.41472986340522766, 'reward': 0.5405032634735107, 'reward_std': 0.41472986340522766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13471050560474396, 'sampling/sampling_logp_difference/max': 1.173884391784668, 'sampling/importance_sampling_ratio/min': 0.3091636896133423, 'sampling/importance_sampling_ratio/mean': 1.0192580223083496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.02584907412529, 'clip_ratio/low_mean': 0.06833758763968945, 'clip_ratio/low_min': 0.06833758763968945, 'clip_ratio/high_mean': 0.0853796461597085, 'clip_ratio/high_max': 0.0853796461597085, 'clip_ratio/region_mean': 0.15371723379939795, 'reward_total_mean': 0.5405032634735107, 'reward_meter_mean': 0.6046037673950195, 'reward_meter_std': 0.4366230070590973, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9792095422744751, 'reward_repeat_penalty_std': 0.024614235386252403, 'reward_repeat_floor_mean': 0.9968814253807068, 'reward_repeat_floor_std': 0.003692133817821741, 'reward_near_duplicate_penalty_mean': 0.9792095422744751, 'reward_near_duplicate_penalty_std': 0.024614235386252403, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.5405032634735107, 'reward_total_composite_std': 0.41472986340522766, 'epoch': 0.03}

 16%|█▌        | 529/3300 [1:54:46<7:36:51,  9.89s/it]INFO 04-12 17:53:13 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 530/3300 [1:54:56<7:27:21,  9.69s/it]2026-04-12 17:53:22,233 | INFO | train_grpo_train | metrics_logged mode=train step=530

                                                      
{'loss': -0.0243, 'grad_norm': 12.753585815429688, 'learning_rate': 8.396969696969698e-06, 'num_tokens': 985536.0, 'completions/mean_length': 70.5, 'completions/min_length': 60.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.977649450302124, 'rewards/meter/std': 0.025170614942908287, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47874999046325684, 'rewards/judge_quality/std': 0.19208908081054688, 'rewards/repeat_penalty/mean': 0.6512997150421143, 'rewards/repeat_penalty/std': 0.29170456528663635, 'rewards/repeat_floor/mean': 0.9545592069625854, 'rewards/repeat_floor/std': 0.04167105257511139, 'rewards/near_duplicate_penalty/mean': 0.8683203458786011, 'rewards/near_duplicate_penalty/std': 0.09415994584560394, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.7770681381225586, 'rewards/distinct_2/std': 0.21163016557693481, 'rewards/total_composite/mean': 0.447043240070343, 'rewards/total_composite/std': 0.2018108069896698, 'reward': 0.447043240070343, 'reward_std': 0.201810821890831, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14128251373767853, 'sampling/sampling_logp_difference/max': 1.3856477737426758, 'sampling/importance_sampling_ratio/min': 0.29448401927948, 'sampling/importance_sampling_ratio/mean': 1.0213531255722046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0827714279294014, 'clip_ratio/low_mean': 0.08939542807638645, 'clip_ratio/low_min': 0.08939542807638645, 'clip_ratio/high_mean': 0.034137751907110214, 'clip_ratio/high_max': 0.034137751907110214, 'clip_ratio/region_mean': 0.12353317998349667, 'reward_total_mean': 0.447043240070343, 'reward_meter_mean': 0.977649450302124, 'reward_meter_std': 0.025170614942908287, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6512997150421143, 'reward_repeat_penalty_std': 0.29170456528663635, 'reward_repeat_floor_mean': 0.9545592069625854, 'reward_repeat_floor_std': 0.04167105257511139, 'reward_near_duplicate_penalty_mean': 0.8683203458786011, 'reward_near_duplicate_penalty_std': 0.09415994584560394, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.7770681381225586, 'reward_distinct_2_std': 0.21163016557693481, 'reward_judge_quality_mean': 0.47874999046325684, 'reward_judge_quality_std': 0.19208908081054688, 'reward_total_composite_mean': 0.447043240070343, 'reward_total_composite_std': 0.2018108069896698, 'epoch': 0.03}

 16%|█▌        | 530/3300 [1:54:56<7:27:21,  9.69s/it]INFO 04-12 17:53:22 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 531/3300 [1:55:10<8:29:47, 11.05s/it]2026-04-12 17:53:36,446 | INFO | train_grpo_train | metrics_logged mode=train step=531

                                                      
{'loss': -0.0403, 'grad_norm': 5.967904567718506, 'learning_rate': 8.393939393939394e-06, 'num_tokens': 987069.0, 'completions/mean_length': 103.625, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.28571701049805, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7417125701904297, 'rewards/meter/std': 0.41653189063072205, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.916822075843811, 'rewards/lexical_plausibility/std': 0.20004339516162872, 'rewards/judge_quality/mean': 0.3974999785423279, 'rewards/judge_quality/std': 0.24063903093338013, 'rewards/repeat_penalty/mean': 0.9902595281600952, 'rewards/repeat_penalty/std': 0.019846748560667038, 'rewards/repeat_floor/mean': 0.9985389709472656, 'rewards/repeat_floor/std': 0.0029770233668386936, 'rewards/near_duplicate_penalty/mean': 0.9902595281600952, 'rewards/near_duplicate_penalty/std': 0.019846748560667038, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33809584379196167, 'rewards/total_composite/std': 0.2639330327510834, 'reward': 0.33809584379196167, 'reward_std': 0.2639330327510834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18398135900497437, 'sampling/sampling_logp_difference/max': 1.5910964012145996, 'sampling/importance_sampling_ratio/min': 0.2037021517753601, 'sampling/importance_sampling_ratio/mean': 1.021674633026123, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.328366443514824, 'clip_ratio/low_mean': 0.043073542416095734, 'clip_ratio/low_min': 0.043073542416095734, 'clip_ratio/high_mean': 0.10310218390077353, 'clip_ratio/high_max': 0.10310218390077353, 'clip_ratio/region_mean': 0.14617572631686926, 'reward_total_mean': 0.33809584379196167, 'reward_meter_mean': 0.7417125701904297, 'reward_meter_std': 0.41653189063072205, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.916822075843811, 'reward_lexical_plausibility_std': 0.20004339516162872, 'reward_repeat_penalty_mean': 0.9902595281600952, 'reward_repeat_penalty_std': 0.019846748560667038, 'reward_repeat_floor_mean': 0.9985389709472656, 'reward_repeat_floor_std': 0.0029770233668386936, 'reward_near_duplicate_penalty_mean': 0.9902595281600952, 'reward_near_duplicate_penalty_std': 0.019846748560667038, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3974999785423279, 'reward_judge_quality_std': 0.24063903093338013, 'reward_total_composite_mean': 0.33809584379196167, 'reward_total_composite_std': 0.2639330327510834, 'epoch': 0.03}

 16%|█▌        | 531/3300 [1:55:10<8:29:47, 11.05s/it]INFO 04-12 17:53:36 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 532/3300 [1:55:17<7:41:15, 10.00s/it]2026-04-12 17:53:43,999 | INFO | train_grpo_train | metrics_logged mode=train step=532

                                                      
{'loss': -0.0458, 'grad_norm': 16.18859100341797, 'learning_rate': 8.390909090909091e-06, 'num_tokens': 988585.0, 'completions/mean_length': 21.5, 'completions/min_length': 18.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.990827202796936, 'rewards/meter/std': 0.004266964737325907, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9350000023841858, 'rewards/judge_quality/std': 0.0160356592386961, 'rewards/repeat_penalty/mean': 0.7407815456390381, 'rewards/repeat_penalty/std': 0.019436990842223167, 'rewards/repeat_floor/mean': 0.9831563234329224, 'rewards/repeat_floor/std': 0.003887410741299391, 'rewards/near_duplicate_penalty/mean': 0.9877086877822876, 'rewards/near_duplicate_penalty/std': 0.02591598778963089, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.9108440279960632, 'rewards/total_composite/std': 0.017982803285121918, 'reward': 0.9108440279960632, 'reward_std': 0.017982803285121918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12008897960186005, 'sampling/sampling_logp_difference/max': 1.1165573596954346, 'sampling/importance_sampling_ratio/min': 0.32740500569343567, 'sampling/importance_sampling_ratio/mean': 1.0318312644958496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8956025987863541, 'clip_ratio/low_mean': 0.10250000189989805, 'clip_ratio/low_min': 0.10250000189989805, 'clip_ratio/high_mean': 0.06027668062597513, 'clip_ratio/high_max': 0.06027668062597513, 'clip_ratio/region_mean': 0.16277668252587318, 'reward_total_mean': 0.9108440279960632, 'reward_meter_mean': 0.990827202796936, 'reward_meter_std': 0.004266964737325907, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7407815456390381, 'reward_repeat_penalty_std': 0.019436990842223167, 'reward_repeat_floor_mean': 0.9831563234329224, 'reward_repeat_floor_std': 0.003887410741299391, 'reward_near_duplicate_penalty_mean': 0.9877086877822876, 'reward_near_duplicate_penalty_std': 0.02591598778963089, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9350000023841858, 'reward_judge_quality_std': 0.0160356592386961, 'reward_total_composite_mean': 0.9108440279960632, 'reward_total_composite_std': 0.017982803285121918, 'epoch': 0.03}

 16%|█▌        | 532/3300 [1:55:17<7:41:15, 10.00s/it]INFO 04-12 17:53:44 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 533/3300 [1:55:27<7:34:48,  9.86s/it]2026-04-12 17:53:53,542 | INFO | train_grpo_train | metrics_logged mode=train step=533

                                                      
{'loss': 0.0013, 'grad_norm': 10.451583862304688, 'learning_rate': 8.387878787878788e-06, 'num_tokens': 990655.0, 'completions/mean_length': 85.75, 'completions/min_length': 69.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9029849767684937, 'rewards/meter/std': 0.16572394967079163, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9886363744735718, 'rewards/lexical_plausibility/std': 0.03214120864868164, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.7920286655426025, 'rewards/repeat_penalty/std': 0.1517161875963211, 'rewards/repeat_floor/mean': 0.9770760536193848, 'rewards/repeat_floor/std': 0.016692444682121277, 'rewards/near_duplicate_penalty/mean': 0.9443671703338623, 'rewards/near_duplicate_penalty/std': 0.036020245403051376, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.033407654613256454, 'rewards/distinct_2/mean': 0.8588440418243408, 'rewards/distinct_2/std': 0.10853662341833115, 'rewards/total_composite/mean': 0.3802032470703125, 'rewards/total_composite/std': 0.22947688400745392, 'reward': 0.3802032470703125, 'reward_std': 0.22947688400745392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17223356664180756, 'sampling/sampling_logp_difference/max': 1.8576600551605225, 'sampling/importance_sampling_ratio/min': 0.2245761901140213, 'sampling/importance_sampling_ratio/mean': 1.0213243961334229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1899310946464539, 'clip_ratio/low_mean': 0.11127193598076701, 'clip_ratio/low_min': 0.11127193598076701, 'clip_ratio/high_mean': 0.05669729225337505, 'clip_ratio/high_max': 0.05669729225337505, 'clip_ratio/region_mean': 0.16796922823414207, 'reward_total_mean': 0.3802032470703125, 'reward_meter_mean': 0.9029849767684937, 'reward_meter_std': 0.16572394967079163, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9886363744735718, 'reward_lexical_plausibility_std': 0.03214120864868164, 'reward_repeat_penalty_mean': 0.7920286655426025, 'reward_repeat_penalty_std': 0.1517161875963211, 'reward_repeat_floor_mean': 0.9770760536193848, 'reward_repeat_floor_std': 0.016692444682121277, 'reward_near_duplicate_penalty_mean': 0.9443671703338623, 'reward_near_duplicate_penalty_std': 0.036020245403051376, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.033407654613256454, 'reward_distinct_2_mean': 0.8588440418243408, 'reward_distinct_2_std': 0.10853662341833115, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.3802032470703125, 'reward_total_composite_std': 0.22947688400745392, 'epoch': 0.03}

 16%|█▌        | 533/3300 [1:55:27<7:34:48,  9.86s/it]INFO 04-12 17:53:53 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 534/3300 [1:55:35<7:15:04,  9.44s/it]2026-04-12 17:54:01,988 | INFO | train_grpo_train | metrics_logged mode=train step=534

                                                      
{'loss': -0.0169, 'grad_norm': 14.684247016906738, 'learning_rate': 8.384848484848485e-06, 'num_tokens': 992372.0, 'completions/mean_length': 45.625, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7595301270484924, 'rewards/meter/std': 0.35030755400657654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.2825110852718353, 'rewards/repeat_penalty/mean': 0.9128025770187378, 'rewards/repeat_penalty/std': 0.14674966037273407, 'rewards/repeat_floor/mean': 0.9907699823379517, 'rewards/repeat_floor/std': 0.01503970380872488, 'rewards/near_duplicate_penalty/mean': 0.9747364521026611, 'rewards/near_duplicate_penalty/std': 0.04412895813584328, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9603838324546814, 'rewards/distinct_2/std': 0.059518154710531235, 'rewards/total_composite/mean': 0.5552917718887329, 'rewards/total_composite/std': 0.3379598557949066, 'reward': 0.5552917718887329, 'reward_std': 0.3379598557949066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19227157533168793, 'sampling/sampling_logp_difference/max': 1.9252128601074219, 'sampling/importance_sampling_ratio/min': 0.1458446979522705, 'sampling/importance_sampling_ratio/mean': 1.0221185684204102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4830809086561203, 'clip_ratio/low_mean': 0.08226950652897358, 'clip_ratio/low_min': 0.08226950652897358, 'clip_ratio/high_mean': 0.12457510828971863, 'clip_ratio/high_max': 0.12457510828971863, 'clip_ratio/region_mean': 0.2068446148186922, 'reward_total_mean': 0.5552917718887329, 'reward_meter_mean': 0.7595301270484924, 'reward_meter_std': 0.35030755400657654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9128025770187378, 'reward_repeat_penalty_std': 0.14674966037273407, 'reward_repeat_floor_mean': 0.9907699823379517, 'reward_repeat_floor_std': 0.01503970380872488, 'reward_near_duplicate_penalty_mean': 0.9747364521026611, 'reward_near_duplicate_penalty_std': 0.04412895813584328, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9603838324546814, 'reward_distinct_2_std': 0.059518154710531235, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.2825110852718353, 'reward_total_composite_mean': 0.5552917718887329, 'reward_total_composite_std': 0.3379598557949066, 'epoch': 0.03}

 16%|█▌        | 534/3300 [1:55:35<7:15:04,  9.44s/it]INFO 04-12 17:54:02 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 535/3300 [1:55:44<7:02:05,  9.16s/it]2026-04-12 17:54:10,499 | INFO | train_grpo_train | metrics_logged mode=train step=535

                                                      
{'loss': 0.0063, 'grad_norm': 15.400646209716797, 'learning_rate': 8.381818181818183e-06, 'num_tokens': 994057.0, 'completions/mean_length': 36.625, 'completions/min_length': 28.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5755261778831482, 'rewards/meter/std': 0.33385905623435974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.23868314921855927, 'rewards/repeat_penalty/mean': 0.9926507472991943, 'rewards/repeat_penalty/std': 0.01203578058630228, 'rewards/repeat_floor/mean': 0.9988976120948792, 'rewards/repeat_floor/std': 0.0018053678795695305, 'rewards/near_duplicate_penalty/mean': 0.9926507472991943, 'rewards/near_duplicate_penalty/std': 0.01203578058630228, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27345919609069824, 'rewards/total_composite/std': 0.27108234167099, 'reward': 0.27345919609069824, 'reward_std': 0.27108234167099, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14011964201927185, 'sampling/sampling_logp_difference/max': 1.2333767414093018, 'sampling/importance_sampling_ratio/min': 0.3287924826145172, 'sampling/importance_sampling_ratio/mean': 1.0309313535690308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9671645909547806, 'clip_ratio/low_mean': 0.0746927666477859, 'clip_ratio/low_min': 0.0746927666477859, 'clip_ratio/high_mean': 0.05272368621081114, 'clip_ratio/high_max': 0.05272368621081114, 'clip_ratio/region_mean': 0.12741645285859704, 'reward_total_mean': 0.27345919609069824, 'reward_meter_mean': 0.5755261778831482, 'reward_meter_std': 0.33385905623435974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9926507472991943, 'reward_repeat_penalty_std': 0.01203578058630228, 'reward_repeat_floor_mean': 0.9988976120948792, 'reward_repeat_floor_std': 0.0018053678795695305, 'reward_near_duplicate_penalty_mean': 0.9926507472991943, 'reward_near_duplicate_penalty_std': 0.01203578058630228, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.23868314921855927, 'reward_total_composite_mean': 0.27345919609069824, 'reward_total_composite_std': 0.27108234167099, 'epoch': 0.03}

 16%|█▌        | 535/3300 [1:55:44<7:02:05,  9.16s/it]INFO 04-12 17:54:10 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 536/3300 [1:55:52<6:54:54,  9.01s/it]2026-04-12 17:54:19,149 | INFO | train_grpo_train | metrics_logged mode=train step=536

                                                      
{'loss': 0.0156, 'grad_norm': 14.113986015319824, 'learning_rate': 8.37878787878788e-06, 'num_tokens': 995744.0, 'completions/mean_length': 40.875, 'completions/min_length': 37.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9737036824226379, 'rewards/meter/std': 0.008200089447200298, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6924999952316284, 'rewards/judge_quality/std': 0.24702516198158264, 'rewards/repeat_penalty/mean': 0.9207477569580078, 'rewards/repeat_penalty/std': 0.16635946929454803, 'rewards/repeat_floor/mean': 0.9906090497970581, 'rewards/repeat_floor/std': 0.018031971529126167, 'rewards/near_duplicate_penalty/mean': 0.9663771390914917, 'rewards/near_duplicate_penalty/std': 0.04478232190012932, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9725275039672852, 'rewards/distinct_2/std': 0.0777040421962738, 'rewards/total_composite/mean': 0.6679074168205261, 'rewards/total_composite/std': 0.23968350887298584, 'reward': 0.6679074168205261, 'reward_std': 0.23968349397182465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1420191377401352, 'sampling/sampling_logp_difference/max': 1.8050583600997925, 'sampling/importance_sampling_ratio/min': 0.16446486115455627, 'sampling/importance_sampling_ratio/mean': 1.0114234685897827, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.072468377649784, 'clip_ratio/low_mean': 0.05411811778321862, 'clip_ratio/low_min': 0.05411811778321862, 'clip_ratio/high_mean': 0.07994132116436958, 'clip_ratio/high_max': 0.07994132116436958, 'clip_ratio/region_mean': 0.1340594389475882, 'reward_total_mean': 0.6679074168205261, 'reward_meter_mean': 0.9737036824226379, 'reward_meter_std': 0.008200089447200298, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9207477569580078, 'reward_repeat_penalty_std': 0.16635946929454803, 'reward_repeat_floor_mean': 0.9906090497970581, 'reward_repeat_floor_std': 0.018031971529126167, 'reward_near_duplicate_penalty_mean': 0.9663771390914917, 'reward_near_duplicate_penalty_std': 0.04478232190012932, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9725275039672852, 'reward_distinct_2_std': 0.0777040421962738, 'reward_judge_quality_mean': 0.6924999952316284, 'reward_judge_quality_std': 0.24702516198158264, 'reward_total_composite_mean': 0.6679074168205261, 'reward_total_composite_std': 0.23968350887298584, 'epoch': 0.03}

 16%|█▌        | 536/3300 [1:55:53<6:54:54,  9.01s/it]INFO 04-12 17:54:19 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 537/3300 [1:56:00<6:32:29,  8.52s/it]2026-04-12 17:54:26,576 | INFO | train_grpo_train | metrics_logged mode=train step=537

                                                      
{'loss': 0.0056, 'grad_norm': 27.017452239990234, 'learning_rate': 8.375757575757576e-06, 'num_tokens': 996955.0, 'completions/mean_length': 15.375, 'completions/min_length': 14.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.375, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.8183342218399048, 'rewards/meter/std': 0.22687792778015137, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.36358335614204407, 'rewards/repeat_penalty/mean': 0.5072780847549438, 'rewards/repeat_penalty/std': 0.33308693766593933, 'rewards/repeat_floor/mean': 0.9359098672866821, 'rewards/repeat_floor/std': 0.06790994107723236, 'rewards/near_duplicate_penalty/mean': 0.7990146279335022, 'rewards/near_duplicate_penalty/std': 0.2539031207561493, 'rewards/opening_diversity/mean': 0.5625, 'rewards/opening_diversity/std': 0.3471825420856476, 'rewards/distinct_2/mean': 0.9145299196243286, 'rewards/distinct_2/std': 0.12921862304210663, 'rewards/total_composite/mean': 0.41749513149261475, 'rewards/total_composite/std': 0.3125753104686737, 'reward': 0.41749513149261475, 'reward_std': 0.3125753104686737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1546861231327057, 'sampling/sampling_logp_difference/max': 0.8975248336791992, 'sampling/importance_sampling_ratio/min': 0.407577246427536, 'sampling/importance_sampling_ratio/mean': 1.0093579292297363, 'sampling/importance_sampling_ratio/max': 1.9865440130233765, 'entropy': 1.0206656903028488, 'clip_ratio/low_mean': 0.04010416753590107, 'clip_ratio/low_min': 0.04010416753590107, 'clip_ratio/high_mean': 0.10004814714193344, 'clip_ratio/high_max': 0.10004814714193344, 'clip_ratio/region_mean': 0.1401523146778345, 'reward_total_mean': 0.41749513149261475, 'reward_meter_mean': 0.8183342218399048, 'reward_meter_std': 0.22687792778015137, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5072780847549438, 'reward_repeat_penalty_std': 0.33308693766593933, 'reward_repeat_floor_mean': 0.9359098672866821, 'reward_repeat_floor_std': 0.06790994107723236, 'reward_near_duplicate_penalty_mean': 0.7990146279335022, 'reward_near_duplicate_penalty_std': 0.2539031207561493, 'reward_opening_diversity_mean': 0.5625, 'reward_opening_diversity_std': 0.3471825420856476, 'reward_distinct_2_mean': 0.9145299196243286, 'reward_distinct_2_std': 0.12921862304210663, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.36358335614204407, 'reward_total_composite_mean': 0.41749513149261475, 'reward_total_composite_std': 0.3125753104686737, 'epoch': 0.03}

 16%|█▋        | 537/3300 [1:56:00<6:32:29,  8.52s/it]INFO 04-12 17:54:26 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 538/3300 [1:56:08<6:30:13,  8.48s/it]2026-04-12 17:54:34,915 | INFO | train_grpo_train | metrics_logged mode=train step=538

                                                      
{'loss': -0.0252, 'grad_norm': 24.270811080932617, 'learning_rate': 8.372727272727273e-06, 'num_tokens': 998572.0, 'completions/mean_length': 23.125, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.5175116062164307, 'rewards/meter/std': 0.45019999146461487, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9537814855575562, 'rewards/lexical_plausibility/std': 0.13072560727596283, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.4051785469055176, 'rewards/repeat_penalty/mean': 0.7772529125213623, 'rewards/repeat_penalty/std': 0.09032660722732544, 'rewards/repeat_floor/mean': 0.9860755801200867, 'rewards/repeat_floor/std': 0.005829798057675362, 'rewards/near_duplicate_penalty/mean': 0.9946705102920532, 'rewards/near_duplicate_penalty/std': 0.010404122062027454, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3026256859302521, 'rewards/total_composite/std': 0.3392674922943115, 'reward': 0.3026256859302521, 'reward_std': 0.3392674922943115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15665315091609955, 'sampling/sampling_logp_difference/max': 1.524843692779541, 'sampling/importance_sampling_ratio/min': 0.21765509247779846, 'sampling/importance_sampling_ratio/mean': 1.0388096570968628, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1804536283016205, 'clip_ratio/low_mean': 0.10047619184479117, 'clip_ratio/low_min': 0.10047619184479117, 'clip_ratio/high_mean': 0.04678571503609419, 'clip_ratio/high_max': 0.04678571503609419, 'clip_ratio/region_mean': 0.14726190688088536, 'reward_total_mean': 0.3026256859302521, 'reward_meter_mean': 0.5175116062164307, 'reward_meter_std': 0.45019999146461487, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9537814855575562, 'reward_lexical_plausibility_std': 0.13072560727596283, 'reward_repeat_penalty_mean': 0.7772529125213623, 'reward_repeat_penalty_std': 0.09032660722732544, 'reward_repeat_floor_mean': 0.9860755801200867, 'reward_repeat_floor_std': 0.005829798057675362, 'reward_near_duplicate_penalty_mean': 0.9946705102920532, 'reward_near_duplicate_penalty_std': 0.010404122062027454, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.4051785469055176, 'reward_total_composite_mean': 0.3026256859302521, 'reward_total_composite_std': 0.3392674922943115, 'epoch': 0.03}

 16%|█▋        | 538/3300 [1:56:08<6:30:13,  8.48s/it]INFO 04-12 17:54:35 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 539/3300 [1:56:16<6:15:32,  8.16s/it]2026-04-12 17:54:42,367 | INFO | train_grpo_train | metrics_logged mode=train step=539

                                                      
{'loss': 0.0687, 'grad_norm': 28.37616539001465, 'learning_rate': 8.36969696969697e-06, 'num_tokens': 1000146.0, 'completions/mean_length': 20.75, 'completions/min_length': 16.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.836410403251648, 'rewards/meter/std': 0.34153905510902405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9712885022163391, 'rewards/lexical_plausibility/std': 0.08120834082365036, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.36130717396736145, 'rewards/repeat_penalty/mean': 0.7469173073768616, 'rewards/repeat_penalty/std': 0.00871915090829134, 'rewards/repeat_floor/mean': 0.9843834638595581, 'rewards/repeat_floor/std': 0.0017438469221815467, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6694763898849487, 'rewards/total_composite/std': 0.3791556656360626, 'reward': 0.6694763898849487, 'reward_std': 0.37915563583374023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.195589080452919, 'sampling/sampling_logp_difference/max': 2.2354326248168945, 'sampling/importance_sampling_ratio/min': 0.10694584995508194, 'sampling/importance_sampling_ratio/mean': 1.0180273056030273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5171124637126923, 'clip_ratio/low_mean': 0.03269675932824612, 'clip_ratio/low_min': 0.03269675932824612, 'clip_ratio/high_mean': 0.14932939317077398, 'clip_ratio/high_max': 0.14932939317077398, 'clip_ratio/region_mean': 0.1820261524990201, 'reward_total_mean': 0.6694763898849487, 'reward_meter_mean': 0.836410403251648, 'reward_meter_std': 0.34153905510902405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9712885022163391, 'reward_lexical_plausibility_std': 0.08120834082365036, 'reward_repeat_penalty_mean': 0.7469173073768616, 'reward_repeat_penalty_std': 0.00871915090829134, 'reward_repeat_floor_mean': 0.9843834638595581, 'reward_repeat_floor_std': 0.0017438469221815467, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.36130717396736145, 'reward_total_composite_mean': 0.6694763898849487, 'reward_total_composite_std': 0.3791556656360626, 'epoch': 0.03}

 16%|█▋        | 539/3300 [1:56:16<6:15:32,  8.16s/it]INFO 04-12 17:54:42 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 540/3300 [1:56:27<7:03:09,  9.20s/it]2026-04-12 17:54:53,989 | INFO | train_grpo_train | metrics_logged mode=train step=540

                                                      
{'loss': 0.0426, 'grad_norm': 5.609861373901367, 'learning_rate': 8.366666666666667e-06, 'num_tokens': 1003664.0, 'completions/mean_length': 221.75, 'completions/min_length': 200.0, 'completions/max_length': 237.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 221.75, 'completions/min_terminated_length': 200.0, 'completions/max_terminated_length': 237.0, 'rewards/meter/mean': 0.9899396896362305, 'rewards/meter/std': 0.006180691998451948, 'rewards/count_adherence/mean': 0.796875, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9390625357627869, 'rewards/count_floor/std': 0.019408077001571655, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.19504812359809875, 'rewards/repeat_penalty/std': 0.23113195598125458, 'rewards/repeat_floor/mean': 0.8998855352401733, 'rewards/repeat_floor/std': 0.03842740133404732, 'rewards/near_duplicate_penalty/mean': 0.804067850112915, 'rewards/near_duplicate_penalty/std': 0.10348296910524368, 'rewards/opening_diversity/mean': 0.9299342632293701, 'rewards/opening_diversity/std': 0.08241672813892365, 'rewards/distinct_2/mean': 0.26088157296180725, 'rewards/distinct_2/std': 0.2332412302494049, 'rewards/total_composite/mean': 0.21253344416618347, 'rewards/total_composite/std': 0.10148772597312927, 'reward': 0.21253344416618347, 'reward_std': 0.10148772597312927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10452376306056976, 'sampling/sampling_logp_difference/max': 2.5418431758880615, 'sampling/importance_sampling_ratio/min': 0.07872116565704346, 'sampling/importance_sampling_ratio/mean': 1.005106806755066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7273459509015083, 'clip_ratio/low_mean': 0.04295601695775986, 'clip_ratio/low_min': 0.04295601695775986, 'clip_ratio/high_mean': 0.04126647673547268, 'clip_ratio/high_max': 0.04126647673547268, 'clip_ratio/region_mean': 0.08422249369323254, 'reward_total_mean': 0.21253344416618347, 'reward_meter_mean': 0.9899396896362305, 'reward_meter_std': 0.006180691998451948, 'reward_count_adherence_mean': 0.796875, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9390625357627869, 'reward_count_floor_std': 0.019408077001571655, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.19504812359809875, 'reward_repeat_penalty_std': 0.23113195598125458, 'reward_repeat_floor_mean': 0.8998855352401733, 'reward_repeat_floor_std': 0.03842740133404732, 'reward_near_duplicate_penalty_mean': 0.804067850112915, 'reward_near_duplicate_penalty_std': 0.10348296910524368, 'reward_opening_diversity_mean': 0.9299342632293701, 'reward_opening_diversity_std': 0.08241672813892365, 'reward_distinct_2_mean': 0.26088157296180725, 'reward_distinct_2_std': 0.2332412302494049, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.21253344416618347, 'reward_total_composite_std': 0.10148772597312927, 'epoch': 0.03}

 16%|█▋        | 540/3300 [1:56:27<7:03:09,  9.20s/it]INFO 04-12 17:54:54 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 541/3300 [1:56:36<7:00:38,  9.15s/it]2026-04-12 17:55:02,987 | INFO | train_grpo_train | metrics_logged mode=train step=541

                                                      
{'loss': 0.0363, 'grad_norm': 10.28022289276123, 'learning_rate': 8.363636363636365e-06, 'num_tokens': 1005309.0, 'completions/mean_length': 43.625, 'completions/min_length': 39.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9291422963142395, 'rewards/meter/std': 0.1510709673166275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5337499976158142, 'rewards/judge_quality/std': 0.19449110329151154, 'rewards/repeat_penalty/mean': 0.8899205327033997, 'rewards/repeat_penalty/std': 0.1447400003671646, 'rewards/repeat_floor/mean': 0.9874789714813232, 'rewards/repeat_floor/std': 0.014373972080647945, 'rewards/near_duplicate_penalty/mean': 0.9584294557571411, 'rewards/near_duplicate_penalty/std': 0.02553483657538891, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.950995147228241, 'rewards/distinct_2/std': 0.07158949226140976, 'rewards/total_composite/mean': 0.4795469641685486, 'rewards/total_composite/std': 0.17618587613105774, 'reward': 0.4795469641685486, 'reward_std': 0.17618586122989655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17609557509422302, 'sampling/sampling_logp_difference/max': 3.096968650817871, 'sampling/importance_sampling_ratio/min': 0.04518597200512886, 'sampling/importance_sampling_ratio/mean': 1.0256534814834595, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.190860167145729, 'clip_ratio/low_mean': 0.1447890056297183, 'clip_ratio/low_min': 0.1447890056297183, 'clip_ratio/high_mean': 0.034090910106897354, 'clip_ratio/high_max': 0.034090910106897354, 'clip_ratio/region_mean': 0.17887991573661566, 'reward_total_mean': 0.4795469641685486, 'reward_meter_mean': 0.9291422963142395, 'reward_meter_std': 0.1510709673166275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8899205327033997, 'reward_repeat_penalty_std': 0.1447400003671646, 'reward_repeat_floor_mean': 0.9874789714813232, 'reward_repeat_floor_std': 0.014373972080647945, 'reward_near_duplicate_penalty_mean': 0.9584294557571411, 'reward_near_duplicate_penalty_std': 0.02553483657538891, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.950995147228241, 'reward_distinct_2_std': 0.07158949226140976, 'reward_judge_quality_mean': 0.5337499976158142, 'reward_judge_quality_std': 0.19449110329151154, 'reward_total_composite_mean': 0.4795469641685486, 'reward_total_composite_std': 0.17618587613105774, 'epoch': 0.03}

 16%|█▋        | 541/3300 [1:56:36<7:00:38,  9.15s/it]INFO 04-12 17:55:03 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 542/3300 [1:56:46<7:08:03,  9.31s/it]2026-04-12 17:55:12,686 | INFO | train_grpo_train | metrics_logged mode=train step=542

                                                      
{'loss': 0.038, 'grad_norm': 10.507553100585938, 'learning_rate': 8.360606060606062e-06, 'num_tokens': 1006834.0, 'completions/mean_length': 43.625, 'completions/min_length': 36.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5549032092094421, 'rewards/meter/std': 0.3387735188007355, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6924999952316284, 'rewards/judge_quality/std': 0.22403763234615326, 'rewards/repeat_penalty/mean': 0.9862676858901978, 'rewards/repeat_penalty/std': 0.017138002440333366, 'rewards/repeat_floor/mean': 0.9979401230812073, 'rewards/repeat_floor/std': 0.002570702228695154, 'rewards/near_duplicate_penalty/mean': 0.9862676858901978, 'rewards/near_duplicate_penalty/std': 0.017138002440333366, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3391241431236267, 'rewards/total_composite/std': 0.19354674220085144, 'reward': 0.3391241431236267, 'reward_std': 0.19354675710201263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1643466353416443, 'sampling/sampling_logp_difference/max': 1.9000043869018555, 'sampling/importance_sampling_ratio/min': 0.14956796169281006, 'sampling/importance_sampling_ratio/mean': 1.0039417743682861, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0861047953367233, 'clip_ratio/low_mean': 0.040495650842785835, 'clip_ratio/low_min': 0.040495650842785835, 'clip_ratio/high_mean': 0.08771135564893484, 'clip_ratio/high_max': 0.08771135564893484, 'clip_ratio/region_mean': 0.12820700649172068, 'reward_total_mean': 0.3391241431236267, 'reward_meter_mean': 0.5549032092094421, 'reward_meter_std': 0.3387735188007355, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9862676858901978, 'reward_repeat_penalty_std': 0.017138002440333366, 'reward_repeat_floor_mean': 0.9979401230812073, 'reward_repeat_floor_std': 0.002570702228695154, 'reward_near_duplicate_penalty_mean': 0.9862676858901978, 'reward_near_duplicate_penalty_std': 0.017138002440333366, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6924999952316284, 'reward_judge_quality_std': 0.22403763234615326, 'reward_total_composite_mean': 0.3391241431236267, 'reward_total_composite_std': 0.19354674220085144, 'epoch': 0.03}

 16%|█▋        | 542/3300 [1:56:46<7:08:03,  9.31s/it]INFO 04-12 17:55:12 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 543/3300 [1:57:00<8:09:46, 10.66s/it]2026-04-12 17:55:26,487 | INFO | train_grpo_train | metrics_logged mode=train step=543

                                                      
{'loss': -0.0588, 'grad_norm': 5.534773349761963, 'learning_rate': 8.357575757575759e-06, 'num_tokens': 1008491.0, 'completions/mean_length': 102.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.57143020629883, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7801710367202759, 'rewards/meter/std': 0.3322688639163971, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8668769598007202, 'rewards/lexical_plausibility/std': 0.2071887105703354, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.955765962600708, 'rewards/repeat_penalty/std': 0.0842575654387474, 'rewards/repeat_floor/mean': 0.9961774349212646, 'rewards/repeat_floor/std': 0.004959750920534134, 'rewards/near_duplicate_penalty/mean': 0.987015962600708, 'rewards/near_duplicate_penalty/std': 0.014638004824519157, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.34550759196281433, 'rewards/total_composite/std': 0.2685561776161194, 'reward': 0.34550759196281433, 'reward_std': 0.268556147813797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19492827355861664, 'sampling/sampling_logp_difference/max': 1.2642736434936523, 'sampling/importance_sampling_ratio/min': 0.2824443578720093, 'sampling/importance_sampling_ratio/mean': 1.0476534366607666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6444870680570602, 'clip_ratio/low_mean': 0.07559523917734623, 'clip_ratio/low_min': 0.07559523917734623, 'clip_ratio/high_mean': 0.0709772165864706, 'clip_ratio/high_max': 0.0709772165864706, 'clip_ratio/region_mean': 0.14657245576381683, 'reward_total_mean': 0.34550759196281433, 'reward_meter_mean': 0.7801710367202759, 'reward_meter_std': 0.3322688639163971, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8668769598007202, 'reward_lexical_plausibility_std': 0.2071887105703354, 'reward_repeat_penalty_mean': 0.955765962600708, 'reward_repeat_penalty_std': 0.0842575654387474, 'reward_repeat_floor_mean': 0.9961774349212646, 'reward_repeat_floor_std': 0.004959750920534134, 'reward_near_duplicate_penalty_mean': 0.987015962600708, 'reward_near_duplicate_penalty_std': 0.014638004824519157, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.34550759196281433, 'reward_total_composite_std': 0.2685561776161194, 'epoch': 0.03}

 16%|█▋        | 543/3300 [1:57:00<8:09:46, 10.66s/it]INFO 04-12 17:55:26 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 544/3300 [1:57:08<7:41:54, 10.06s/it]2026-04-12 17:55:35,140 | INFO | train_grpo_train | metrics_logged mode=train step=544

                                                      
{'loss': 0.0401, 'grad_norm': 13.619742393493652, 'learning_rate': 8.354545454545455e-06, 'num_tokens': 1010427.0, 'completions/mean_length': 55.0, 'completions/min_length': 45.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.610531210899353, 'rewards/meter/std': 0.3780951201915741, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.574999988079071, 'rewards/judge_quality/std': 0.23628070950508118, 'rewards/repeat_penalty/mean': 0.9364616870880127, 'rewards/repeat_penalty/std': 0.06729107350111008, 'rewards/repeat_floor/mean': 0.9928113222122192, 'rewards/repeat_floor/std': 0.007240903563797474, 'rewards/near_duplicate_penalty/mean': 0.9779541492462158, 'rewards/near_duplicate_penalty/std': 0.01755790412425995, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9568685293197632, 'rewards/distinct_2/std': 0.05472530052065849, 'rewards/total_composite/mean': 0.3467673659324646, 'rewards/total_composite/std': 0.26447150111198425, 'reward': 0.3467673659324646, 'reward_std': 0.26447150111198425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1923883855342865, 'sampling/sampling_logp_difference/max': 1.640580177307129, 'sampling/importance_sampling_ratio/min': 0.1938675343990326, 'sampling/importance_sampling_ratio/mean': 1.0281118154525757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5289423763751984, 'clip_ratio/low_mean': 0.1141222445294261, 'clip_ratio/low_min': 0.1141222445294261, 'clip_ratio/high_mean': 0.046415770426392555, 'clip_ratio/high_max': 0.046415770426392555, 'clip_ratio/region_mean': 0.16053801495581865, 'reward_total_mean': 0.3467673659324646, 'reward_meter_mean': 0.610531210899353, 'reward_meter_std': 0.3780951201915741, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9364616870880127, 'reward_repeat_penalty_std': 0.06729107350111008, 'reward_repeat_floor_mean': 0.9928113222122192, 'reward_repeat_floor_std': 0.007240903563797474, 'reward_near_duplicate_penalty_mean': 0.9779541492462158, 'reward_near_duplicate_penalty_std': 0.01755790412425995, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9568685293197632, 'reward_distinct_2_std': 0.05472530052065849, 'reward_judge_quality_mean': 0.574999988079071, 'reward_judge_quality_std': 0.23628070950508118, 'reward_total_composite_mean': 0.3467673659324646, 'reward_total_composite_std': 0.26447150111198425, 'epoch': 0.03}

 16%|█▋        | 544/3300 [1:57:08<7:41:54, 10.06s/it]INFO 04-12 17:55:35 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 545/3300 [1:57:17<7:19:06,  9.56s/it]2026-04-12 17:55:43,549 | INFO | train_grpo_train | metrics_logged mode=train step=545

                                                      
{'loss': -0.0339, 'grad_norm': 26.761737823486328, 'learning_rate': 8.351515151515152e-06, 'num_tokens': 1011888.0, 'completions/mean_length': 19.625, 'completions/min_length': 14.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.625, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8017735481262207, 'rewards/meter/std': 0.3023511469364166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7351142168045044, 'rewards/lexical_plausibility/std': 0.26900801062583923, 'rewards/judge_quality/mean': 0.6475000381469727, 'rewards/judge_quality/std': 0.30742016434669495, 'rewards/repeat_penalty/mean': 0.680923581123352, 'rewards/repeat_penalty/std': 0.11025054007768631, 'rewards/repeat_floor/mean': 0.9711847305297852, 'rewards/repeat_floor/std': 0.022050119936466217, 'rewards/near_duplicate_penalty/mean': 0.9078980684280396, 'rewards/near_duplicate_penalty/std': 0.14700071513652802, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5023888945579529, 'rewards/total_composite/std': 0.3166264593601227, 'reward': 0.5023888945579529, 'reward_std': 0.3166264593601227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18353956937789917, 'sampling/sampling_logp_difference/max': 1.4188008308410645, 'sampling/importance_sampling_ratio/min': 0.24200405180454254, 'sampling/importance_sampling_ratio/mean': 1.0561906099319458, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.584113746881485, 'clip_ratio/low_mean': 0.11533521302044392, 'clip_ratio/low_min': 0.11533521302044392, 'clip_ratio/high_mean': 0.0706196604296565, 'clip_ratio/high_max': 0.0706196604296565, 'clip_ratio/region_mean': 0.18595487345010042, 'reward_total_mean': 0.5023888945579529, 'reward_meter_mean': 0.8017735481262207, 'reward_meter_std': 0.3023511469364166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.7351142168045044, 'reward_lexical_plausibility_std': 0.26900801062583923, 'reward_repeat_penalty_mean': 0.680923581123352, 'reward_repeat_penalty_std': 0.11025054007768631, 'reward_repeat_floor_mean': 0.9711847305297852, 'reward_repeat_floor_std': 0.022050119936466217, 'reward_near_duplicate_penalty_mean': 0.9078980684280396, 'reward_near_duplicate_penalty_std': 0.14700071513652802, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6475000381469727, 'reward_judge_quality_std': 0.30742016434669495, 'reward_total_composite_mean': 0.5023888945579529, 'reward_total_composite_std': 0.3166264593601227, 'epoch': 0.03}

 17%|█▋        | 545/3300 [1:57:17<7:19:06,  9.56s/it]INFO 04-12 17:55:43 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 546/3300 [1:57:24<6:50:14,  8.94s/it]2026-04-12 17:55:51,027 | INFO | train_grpo_train | metrics_logged mode=train step=546

                                                      
{'loss': 0.039, 'grad_norm': 20.77408790588379, 'learning_rate': 8.348484848484849e-06, 'num_tokens': 1013367.0, 'completions/mean_length': 22.875, 'completions/min_length': 20.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9722200632095337, 'rewards/meter/std': 0.03387109562754631, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8274999856948853, 'rewards/judge_quality/std': 0.27395257353782654, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7910505533218384, 'rewards/total_composite/std': 0.2620963752269745, 'reward': 0.7910505533218384, 'reward_std': 0.2620963454246521, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16928167641162872, 'sampling/sampling_logp_difference/max': 1.0764293670654297, 'sampling/importance_sampling_ratio/min': 0.34081026911735535, 'sampling/importance_sampling_ratio/mean': 1.0310733318328857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4028927013278008, 'clip_ratio/low_mean': 0.02083333395421505, 'clip_ratio/low_min': 0.02083333395421505, 'clip_ratio/high_mean': 0.17040193639695644, 'clip_ratio/high_max': 0.17040193639695644, 'clip_ratio/region_mean': 0.1912352703511715, 'reward_total_mean': 0.7910505533218384, 'reward_meter_mean': 0.9722200632095337, 'reward_meter_std': 0.03387109562754631, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8274999856948853, 'reward_judge_quality_std': 0.27395257353782654, 'reward_total_composite_mean': 0.7910505533218384, 'reward_total_composite_std': 0.2620963752269745, 'epoch': 0.03}

 17%|█▋        | 546/3300 [1:57:24<6:50:14,  8.94s/it]INFO 04-12 17:55:51 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 547/3300 [1:57:33<6:53:00,  9.00s/it]2026-04-12 17:56:00,177 | INFO | train_grpo_train | metrics_logged mode=train step=547

                                                      
{'loss': 0.0912, 'grad_norm': 14.439604759216309, 'learning_rate': 8.345454545454546e-06, 'num_tokens': 1015122.0, 'completions/mean_length': 41.375, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.3898012340068817, 'rewards/meter/std': 0.43012189865112305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.9841147661209106, 'rewards/repeat_penalty/std': 0.020881956443190575, 'rewards/repeat_floor/mean': 0.9976171851158142, 'rewards/repeat_floor/std': 0.003132288111373782, 'rewards/near_duplicate_penalty/mean': 0.9841147661209106, 'rewards/near_duplicate_penalty/std': 0.020881956443190575, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21943525969982147, 'rewards/total_composite/std': 0.2665696144104004, 'reward': 0.21943525969982147, 'reward_std': 0.2665696144104004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19609205424785614, 'sampling/sampling_logp_difference/max': 1.6701555252075195, 'sampling/importance_sampling_ratio/min': 0.18821778893470764, 'sampling/importance_sampling_ratio/mean': 1.0267587900161743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7265300154685974, 'clip_ratio/low_mean': 0.09462734777480364, 'clip_ratio/low_min': 0.09462734777480364, 'clip_ratio/high_mean': 0.07235473114997149, 'clip_ratio/high_max': 0.07235473114997149, 'clip_ratio/region_mean': 0.16698207892477512, 'reward_total_mean': 0.21943525969982147, 'reward_meter_mean': 0.3898012340068817, 'reward_meter_std': 0.43012189865112305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9841147661209106, 'reward_repeat_penalty_std': 0.020881956443190575, 'reward_repeat_floor_mean': 0.9976171851158142, 'reward_repeat_floor_std': 0.003132288111373782, 'reward_near_duplicate_penalty_mean': 0.9841147661209106, 'reward_near_duplicate_penalty_std': 0.020881956443190575, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.21943525969982147, 'reward_total_composite_std': 0.2665696144104004, 'epoch': 0.03}

 17%|█▋        | 547/3300 [1:57:34<6:53:00,  9.00s/it]INFO 04-12 17:56:00 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 548/3300 [1:57:42<6:50:32,  8.95s/it]2026-04-12 17:56:09,007 | INFO | train_grpo_train | metrics_logged mode=train step=548

                                                      
{'loss': 0.0228, 'grad_norm': 14.816450119018555, 'learning_rate': 8.342424242424244e-06, 'num_tokens': 1016729.0, 'completions/mean_length': 42.875, 'completions/min_length': 39.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7090002298355103, 'rewards/meter/std': 0.42208850383758545, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9713541269302368, 'rewards/lexical_plausibility/std': 0.05990393087267876, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2975345253944397, 'rewards/repeat_penalty/mean': 0.9335046410560608, 'rewards/repeat_penalty/std': 0.12607936561107635, 'rewards/repeat_floor/mean': 0.9929938316345215, 'rewards/repeat_floor/std': 0.010820589028298855, 'rewards/near_duplicate_penalty/mean': 0.9758254885673523, 'rewards/near_duplicate_penalty/std': 0.025628648698329926, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.5091841220855713, 'rewards/total_composite/std': 0.37691253423690796, 'reward': 0.5091841220855713, 'reward_std': 0.37691250443458557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1511572003364563, 'sampling/sampling_logp_difference/max': 1.2685370445251465, 'sampling/importance_sampling_ratio/min': 0.2812427580356598, 'sampling/importance_sampling_ratio/mean': 1.0372276306152344, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0031752213835716, 'clip_ratio/low_mean': 0.07500832807272673, 'clip_ratio/low_min': 0.07500832807272673, 'clip_ratio/high_mean': 0.08761508017778397, 'clip_ratio/high_max': 0.08761508017778397, 'clip_ratio/region_mean': 0.1626234082505107, 'reward_total_mean': 0.5091841220855713, 'reward_meter_mean': 0.7090002298355103, 'reward_meter_std': 0.42208850383758545, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9713541269302368, 'reward_lexical_plausibility_std': 0.05990393087267876, 'reward_repeat_penalty_mean': 0.9335046410560608, 'reward_repeat_penalty_std': 0.12607936561107635, 'reward_repeat_floor_mean': 0.9929938316345215, 'reward_repeat_floor_std': 0.010820589028298855, 'reward_near_duplicate_penalty_mean': 0.9758254885673523, 'reward_near_duplicate_penalty_std': 0.025628648698329926, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2975345253944397, 'reward_total_composite_mean': 0.5091841220855713, 'reward_total_composite_std': 0.37691253423690796, 'epoch': 0.03}

 17%|█▋        | 548/3300 [1:57:42<6:50:32,  8.95s/it]INFO 04-12 17:56:09 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 549/3300 [1:57:52<7:07:11,  9.32s/it]2026-04-12 17:56:19,183 | INFO | train_grpo_train | metrics_logged mode=train step=549

                                                      
{'loss': 0.1447, 'grad_norm': 13.118990898132324, 'learning_rate': 8.339393939393941e-06, 'num_tokens': 1018643.0, 'completions/mean_length': 74.25, 'completions/min_length': 64.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.835888147354126, 'rewards/meter/std': 0.31511005759239197, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42875000834465027, 'rewards/judge_quality/std': 0.18566001951694489, 'rewards/repeat_penalty/mean': 0.9674917459487915, 'rewards/repeat_penalty/std': 0.03475857898592949, 'rewards/repeat_floor/mean': 0.995691180229187, 'rewards/repeat_floor/std': 0.004208729136735201, 'rewards/near_duplicate_penalty/mean': 0.9776894450187683, 'rewards/near_duplicate_penalty/std': 0.01718098670244217, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893079996109009, 'rewards/distinct_2/std': 0.019807932898402214, 'rewards/total_composite/mean': 0.3585785925388336, 'rewards/total_composite/std': 0.19560429453849792, 'reward': 0.3585785925388336, 'reward_std': 0.19560429453849792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22280728816986084, 'sampling/sampling_logp_difference/max': 4.220475196838379, 'sampling/importance_sampling_ratio/min': 0.014691662043333054, 'sampling/importance_sampling_ratio/mean': 1.0015207529067993, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.557713970541954, 'clip_ratio/low_mean': 0.1308778803795576, 'clip_ratio/low_min': 0.1308778803795576, 'clip_ratio/high_mean': 0.06722248159348965, 'clip_ratio/high_max': 0.06722248159348965, 'clip_ratio/region_mean': 0.19810036197304726, 'reward_total_mean': 0.3585785925388336, 'reward_meter_mean': 0.835888147354126, 'reward_meter_std': 0.31511005759239197, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9674917459487915, 'reward_repeat_penalty_std': 0.03475857898592949, 'reward_repeat_floor_mean': 0.995691180229187, 'reward_repeat_floor_std': 0.004208729136735201, 'reward_near_duplicate_penalty_mean': 0.9776894450187683, 'reward_near_duplicate_penalty_std': 0.01718098670244217, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893079996109009, 'reward_distinct_2_std': 0.019807932898402214, 'reward_judge_quality_mean': 0.42875000834465027, 'reward_judge_quality_std': 0.18566001951694489, 'reward_total_composite_mean': 0.3585785925388336, 'reward_total_composite_std': 0.19560429453849792, 'epoch': 0.03}

 17%|█▋        | 549/3300 [1:57:53<7:07:11,  9.32s/it]INFO 04-12 17:56:19 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 550/3300 [1:58:06<7:59:57, 10.47s/it]2026-04-12 17:56:32,346 | INFO | train_grpo_train | metrics_logged mode=train step=550

                                                      
{'loss': -0.0298, 'grad_norm': 5.064102649688721, 'learning_rate': 8.336363636363636e-06, 'num_tokens': 1020110.0, 'completions/mean_length': 92.375, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.42857360839844, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.43611469864845276, 'rewards/meter/std': 0.3893551230430603, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9332994818687439, 'rewards/lexical_plausibility/std': 0.18865752220153809, 'rewards/judge_quality/mean': 0.8187500238418579, 'rewards/judge_quality/std': 0.3109174370765686, 'rewards/repeat_penalty/mean': 0.9961467981338501, 'rewards/repeat_penalty/std': 0.010898458771407604, 'rewards/repeat_floor/mean': 0.999422013759613, 'rewards/repeat_floor/std': 0.001634771004319191, 'rewards/near_duplicate_penalty/mean': 0.9961467981338501, 'rewards/near_duplicate_penalty/std': 0.010898458771407604, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4066752791404724, 'rewards/total_composite/std': 0.3658042848110199, 'reward': 0.4066752791404724, 'reward_std': 0.3658042848110199, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12511000037193298, 'sampling/sampling_logp_difference/max': 1.336487889289856, 'sampling/importance_sampling_ratio/min': 0.26276692748069763, 'sampling/importance_sampling_ratio/mean': 1.0332398414611816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6308564990758896, 'clip_ratio/low_mean': 0.06636904831975698, 'clip_ratio/low_min': 0.06636904831975698, 'clip_ratio/high_mean': 0.04446512646973133, 'clip_ratio/high_max': 0.04446512646973133, 'clip_ratio/region_mean': 0.11083417478948832, 'reward_total_mean': 0.4066752791404724, 'reward_meter_mean': 0.43611469864845276, 'reward_meter_std': 0.3893551230430603, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9332994818687439, 'reward_lexical_plausibility_std': 0.18865752220153809, 'reward_repeat_penalty_mean': 0.9961467981338501, 'reward_repeat_penalty_std': 0.010898458771407604, 'reward_repeat_floor_mean': 0.999422013759613, 'reward_repeat_floor_std': 0.001634771004319191, 'reward_near_duplicate_penalty_mean': 0.9961467981338501, 'reward_near_duplicate_penalty_std': 0.010898458771407604, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8187500238418579, 'reward_judge_quality_std': 0.3109174370765686, 'reward_total_composite_mean': 0.4066752791404724, 'reward_total_composite_std': 0.3658042848110199, 'epoch': 0.03}

 17%|█▋        | 550/3300 [1:58:06<7:59:57, 10.47s/it]INFO 04-12 17:56:32 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:01,  5.62s/it]

 23%|██▎       | 3/13 [00:22<01:19,  7.93s/it]

 31%|███       | 4/13 [00:31<01:16,  8.53s/it]

 38%|███▊      | 5/13 [00:44<01:18,  9.80s/it]

 46%|████▌     | 6/13 [00:55<01:12, 10.41s/it]

 54%|█████▍    | 7/13 [01:06<01:03, 10.58s/it]

 62%|██████▏   | 8/13 [01:17<00:52, 10.56s/it]

 69%|██████▉   | 9/13 [01:29<00:44, 11.01s/it]

 77%|███████▋  | 10/13 [01:41<00:33, 11.30s/it]

 85%|████████▍ | 11/13 [01:52<00:22, 11.37s/it]

 92%|█████████▏| 12/13 [02:04<00:11, 11.56s/it]

100%|██████████| 13/13 [02:16<00:00, 11.62s/it]2026-04-12 17:59:00,752 | INFO | train_grpo_train | metrics_logged mode=eval step=550

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 148.3576, 'eval_samples_per_second': 0.701, 'eval_steps_per_second': 0.088, 'eval_num_tokens': 1020110.0, 'eval_completions/mean_length': 201.15384615384616, 'eval_completions/min_length': 39.0, 'eval_completions/max_length': 490.46153846153845, 'eval_completions/clipped_ratio': 0.14423076923076922, 'eval_completions/mean_terminated_length': 147.55146965613733, 'eval_completions/min_terminated_length': 39.0, 'eval_completions/max_terminated_length': 333.3076923076923, 'eval_rewards/meter/mean': 0.6186482149821061, 'eval_rewards/meter/std': 0.35500125472362226, 'eval_rewards/count_adherence/mean': 0.8089264081074641, 'eval_rewards/count_adherence/std': 0.22151099947782663, 'eval_rewards/arabic_clean/mean': 0.8173076923076923, 'eval_rewards/arabic_clean/std': 0.3768232533564934, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9426779105113103, 'eval_rewards/count_floor/std': 0.06645329855382442, 'eval_rewards/lexical_plausibility/mean': 0.9259520860818716, 'eval_rewards/lexical_plausibility/std': 0.14988471453006452, 'eval_rewards/judge_quality/mean': 0.3449038381759937, 'eval_rewards/judge_quality/std': 0.2187302645582419, 'eval_rewards/repeat_penalty/mean': 0.7233055692452651, 'eval_rewards/repeat_penalty/std': 0.318151416686865, 'eval_rewards/repeat_floor/mean': 0.9671928745049697, 'eval_rewards/repeat_floor/std': 0.04186510058263174, 'eval_rewards/near_duplicate_penalty/mean': 0.9314568822200482, 'eval_rewards/near_duplicate_penalty/std': 0.08834549498099548, 'eval_rewards/opening_diversity/mean': 0.9488580960493821, 'eval_rewards/opening_diversity/std': 0.10991556369341336, 'eval_rewards/distinct_2/mean': 0.7838094234466553, 'eval_rewards/distinct_2/std': 0.2979626308840055, 'eval_rewards/total_composite/mean': 0.20396145777060434, 'eval_rewards/total_composite/std': 0.18293480288523895, 'eval_reward': 0.20396145777060434, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0797056805055875, 'eval_sampling/sampling_logp_difference/max': 1.1188378700843225, 'eval_sampling/importance_sampling_ratio/min': 0.33932835780657256, 'eval_sampling/importance_sampling_ratio/mean': 1.019963759642381, 'eval_sampling/importance_sampling_ratio/max': 1.5410445745174701, 'eval_entropy': 1.1367097451136663, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.20396145777060434, 'eval_reward_meter_mean': 0.6186482149821061, 'eval_reward_meter_std': 0.35500125472362226, 'eval_reward_count_adherence_mean': 0.8089264081074641, 'eval_reward_count_adherence_std': 0.22151099947782663, 'eval_reward_arabic_clean_mean': 0.8173076923076923, 'eval_reward_arabic_clean_std': 0.3768232533564934, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9426779105113103, 'eval_reward_count_floor_std': 0.06645329855382442, 'eval_reward_lexical_plausibility_mean': 0.9259520860818716, 'eval_reward_lexical_plausibility_std': 0.14988471453006452, 'eval_reward_repeat_penalty_mean': 0.7233055692452651, 'eval_reward_repeat_penalty_std': 0.318151416686865, 'eval_reward_repeat_floor_mean': 0.9671928745049697, 'eval_reward_repeat_floor_std': 0.04186510058263174, 'eval_reward_near_duplicate_penalty_mean': 0.9314568822200482, 'eval_reward_near_duplicate_penalty_std': 0.08834549498099548, 'eval_reward_opening_diversity_mean': 0.9488580960493821, 'eval_reward_opening_diversity_std': 0.10991556369341336, 'eval_reward_distinct_2_mean': 0.7838094234466553, 'eval_reward_distinct_2_std': 0.2979626308840055, 'eval_reward_judge_quality_mean': 0.3449038381759937, 'eval_reward_judge_quality_std': 0.2187302645582419, 'eval_reward_total_composite_mean': 0.20396145777060434, 'eval_reward_total_composite_std': 0.18293480288523895, 'epoch': 0.03}

 17%|█▋        | 550/3300 [2:00:34<7:59:57, 10.47s/it]

100%|██████████| 13/13 [02:16<00:00, 11.62s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 17%|█▋        | 551/3300 [2:00:52<43:38:17, 57.15s/it]2026-04-12 17:59:18,406 | INFO | train_grpo_train | metrics_logged mode=train step=551

                                                       
{'loss': -0.0659, 'grad_norm': 5.255579948425293, 'learning_rate': 8.333333333333334e-06, 'num_tokens': 1021730.0, 'completions/mean_length': 102.5, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.000003814697266, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6737279891967773, 'rewards/meter/std': 0.38971206545829773, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9403427243232727, 'rewards/lexical_plausibility/std': 0.1687362641096115, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9403715133666992, 'rewards/repeat_penalty/std': 0.09566418081521988, 'rewards/repeat_floor/mean': 0.9947792291641235, 'rewards/repeat_floor/std': 0.007299729157239199, 'rewards/near_duplicate_penalty/mean': 0.9877285957336426, 'rewards/near_duplicate_penalty/std': 0.01969127357006073, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.23661760985851288, 'rewards/total_composite/std': 0.16402758657932281, 'reward': 0.23661760985851288, 'reward_std': 0.16402758657932281, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21333429217338562, 'sampling/sampling_logp_difference/max': 1.580709457397461, 'sampling/importance_sampling_ratio/min': 0.20582902431488037, 'sampling/importance_sampling_ratio/mean': 1.028855562210083, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7297458723187447, 'clip_ratio/low_mean': 0.03371212258934975, 'clip_ratio/low_min': 0.03371212258934975, 'clip_ratio/high_mean': 0.11647386383265257, 'clip_ratio/high_max': 0.11647386383265257, 'clip_ratio/region_mean': 0.15018598642200232, 'reward_total_mean': 0.23661760985851288, 'reward_meter_mean': 0.6737279891967773, 'reward_meter_std': 0.38971206545829773, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9403427243232727, 'reward_lexical_plausibility_std': 0.1687362641096115, 'reward_repeat_penalty_mean': 0.9403715133666992, 'reward_repeat_penalty_std': 0.09566418081521988, 'reward_repeat_floor_mean': 0.9947792291641235, 'reward_repeat_floor_std': 0.007299729157239199, 'reward_near_duplicate_penalty_mean': 0.9877285957336426, 'reward_near_duplicate_penalty_std': 0.01969127357006073, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.23661760985851288, 'reward_total_composite_std': 0.16402758657932281, 'epoch': 0.03}

 17%|█▋        | 551/3300 [2:00:52<43:38:17, 57.15s/it]INFO 04-12 17:59:18 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 552/3300 [2:01:07<34:02:19, 44.59s/it]2026-04-12 17:59:33,705 | INFO | train_grpo_train | metrics_logged mode=train step=552

                                                       
{'loss': -0.1397, 'grad_norm': 3.367072582244873, 'learning_rate': 8.330303030303031e-06, 'num_tokens': 1024008.0, 'completions/mean_length': 325.75, 'completions/min_length': 132.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 139.5, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.36627963185310364, 'rewards/meter/std': 0.2974469065666199, 'rewards/count_adherence/mean': 0.6785714626312256, 'rewards/count_adherence/std': 0.26175010204315186, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.903571367263794, 'rewards/count_floor/std': 0.07852503657341003, 'rewards/lexical_plausibility/mean': 0.7508395910263062, 'rewards/lexical_plausibility/std': 0.2670208215713501, 'rewards/judge_quality/mean': 0.19999998807907104, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.866409182548523, 'rewards/repeat_penalty/std': 0.23204495012760162, 'rewards/repeat_floor/mean': 0.9853261709213257, 'rewards/repeat_floor/std': 0.024082697927951813, 'rewards/near_duplicate_penalty/mean': 0.9658227562904358, 'rewards/near_duplicate_penalty/std': 0.031006718054413795, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8939195871353149, 'rewards/distinct_2/std': 0.22000761330127716, 'rewards/total_composite/mean': 0.08916445821523666, 'rewards/total_composite/std': 0.1075466051697731, 'reward': 0.08916445821523666, 'reward_std': 0.1075466051697731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18914061784744263, 'sampling/sampling_logp_difference/max': 1.97637939453125, 'sampling/importance_sampling_ratio/min': 0.13857004046440125, 'sampling/importance_sampling_ratio/mean': 1.01470947265625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5549485981464386, 'clip_ratio/low_mean': 0.023320894688367844, 'clip_ratio/low_min': 0.023320894688367844, 'clip_ratio/high_mean': 0.057010688818991184, 'clip_ratio/high_max': 0.057010688818991184, 'clip_ratio/region_mean': 0.08033158350735903, 'reward_total_mean': 0.08916445821523666, 'reward_meter_mean': 0.36627963185310364, 'reward_meter_std': 0.2974469065666199, 'reward_count_adherence_mean': 0.6785714626312256, 'reward_count_adherence_std': 0.26175010204315186, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.903571367263794, 'reward_count_floor_std': 0.07852503657341003, 'reward_lexical_plausibility_mean': 0.7508395910263062, 'reward_lexical_plausibility_std': 0.2670208215713501, 'reward_repeat_penalty_mean': 0.866409182548523, 'reward_repeat_penalty_std': 0.23204495012760162, 'reward_repeat_floor_mean': 0.9853261709213257, 'reward_repeat_floor_std': 0.024082697927951813, 'reward_near_duplicate_penalty_mean': 0.9658227562904358, 'reward_near_duplicate_penalty_std': 0.031006718054413795, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8939195871353149, 'reward_distinct_2_std': 0.22000761330127716, 'reward_judge_quality_mean': 0.19999998807907104, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.08916445821523666, 'reward_total_composite_std': 0.1075466051697731, 'epoch': 0.03}

 17%|█▋        | 552/3300 [2:01:07<34:02:19, 44.59s/it]INFO 04-12 17:59:33 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 553/3300 [2:01:16<25:50:15, 33.86s/it]2026-04-12 17:59:42,522 | INFO | train_grpo_train | metrics_logged mode=train step=553

                                                       
{'loss': -0.0342, 'grad_norm': 21.553974151611328, 'learning_rate': 8.327272727272728e-06, 'num_tokens': 1025561.0, 'completions/mean_length': 26.125, 'completions/min_length': 22.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.5366592407226562, 'rewards/meter/std': 0.2860164940357208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.981582522392273, 'rewards/repeat_penalty/std': 0.012491089291870594, 'rewards/repeat_floor/mean': 0.9972373247146606, 'rewards/repeat_floor/std': 0.0018736639758571982, 'rewards/near_duplicate_penalty/mean': 0.981582522392273, 'rewards/near_duplicate_penalty/std': 0.012491089291870594, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2340528666973114, 'rewards/total_composite/std': 0.13655918836593628, 'reward': 0.2340528666973114, 'reward_std': 0.13655918836593628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2177087813615799, 'sampling/sampling_logp_difference/max': 1.7831556797027588, 'sampling/importance_sampling_ratio/min': 0.1681068241596222, 'sampling/importance_sampling_ratio/mean': 0.9540435075759888, 'sampling/importance_sampling_ratio/max': 1.763105869293213, 'entropy': 0.966768778860569, 'clip_ratio/low_mean': 0.09590273816138506, 'clip_ratio/low_min': 0.09590273816138506, 'clip_ratio/high_mean': 0.13204094395041466, 'clip_ratio/high_max': 0.13204094395041466, 'clip_ratio/region_mean': 0.22794368211179972, 'reward_total_mean': 0.2340528666973114, 'reward_meter_mean': 0.5366592407226562, 'reward_meter_std': 0.2860164940357208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.981582522392273, 'reward_repeat_penalty_std': 0.012491089291870594, 'reward_repeat_floor_mean': 0.9972373247146606, 'reward_repeat_floor_std': 0.0018736639758571982, 'reward_near_duplicate_penalty_mean': 0.981582522392273, 'reward_near_duplicate_penalty_std': 0.012491089291870594, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.2340528666973114, 'reward_total_composite_std': 0.13655918836593628, 'epoch': 0.03}

 17%|█▋        | 553/3300 [2:01:16<25:50:15, 33.86s/it]INFO 04-12 17:59:42 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 554/3300 [2:01:25<20:07:58, 26.39s/it]2026-04-12 17:59:51,521 | INFO | train_grpo_train | metrics_logged mode=train step=554

                                                       
{'loss': 0.0324, 'grad_norm': 20.6953125, 'learning_rate': 8.324242424242425e-06, 'num_tokens': 1027146.0, 'completions/mean_length': 34.125, 'completions/min_length': 26.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7248470783233643, 'rewards/meter/std': 0.26182666420936584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9007190465927124, 'rewards/repeat_penalty/std': 0.20954850316047668, 'rewards/repeat_floor/mean': 0.9873193502426147, 'rewards/repeat_floor/std': 0.025214746594429016, 'rewards/near_duplicate_penalty/mean': 0.9568088054656982, 'rewards/near_duplicate_penalty/std': 0.05396223068237305, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9519230723381042, 'rewards/distinct_2/std': 0.13598206639289856, 'rewards/total_composite/mean': 0.4669719934463501, 'rewards/total_composite/std': 0.2306515872478485, 'reward': 0.4669719934463501, 'reward_std': 0.23065157234668732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17184989154338837, 'sampling/sampling_logp_difference/max': 3.478653907775879, 'sampling/importance_sampling_ratio/min': 0.22473815083503723, 'sampling/importance_sampling_ratio/mean': 1.046779990196228, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1770471036434174, 'clip_ratio/low_mean': 0.09403077885508537, 'clip_ratio/low_min': 0.09403077885508537, 'clip_ratio/high_mean': 0.031652189092710614, 'clip_ratio/high_max': 0.031652189092710614, 'clip_ratio/region_mean': 0.125682967947796, 'reward_total_mean': 0.4669719934463501, 'reward_meter_mean': 0.7248470783233643, 'reward_meter_std': 0.26182666420936584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9007190465927124, 'reward_repeat_penalty_std': 0.20954850316047668, 'reward_repeat_floor_mean': 0.9873193502426147, 'reward_repeat_floor_std': 0.025214746594429016, 'reward_near_duplicate_penalty_mean': 0.9568088054656982, 'reward_near_duplicate_penalty_std': 0.05396223068237305, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9519230723381042, 'reward_distinct_2_std': 0.13598206639289856, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.4669719934463501, 'reward_total_composite_std': 0.2306515872478485, 'epoch': 0.03}

 17%|█▋        | 554/3300 [2:01:25<20:07:58, 26.39s/it]INFO 04-12 17:59:51 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 555/3300 [2:01:34<16:11:37, 21.24s/it]2026-04-12 18:00:00,698 | INFO | train_grpo_train | metrics_logged mode=train step=555

                                                       
{'loss': -0.0116, 'grad_norm': 18.11533546447754, 'learning_rate': 8.321212121212123e-06, 'num_tokens': 1028430.0, 'completions/mean_length': 17.5, 'completions/min_length': 15.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.5, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.5945756435394287, 'rewards/meter/std': 0.4319617450237274, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.29624998569488525, 'rewards/judge_quality/std': 0.26773855090141296, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2279086858034134, 'rewards/total_composite/std': 0.2996399700641632, 'reward': 0.2279086858034134, 'reward_std': 0.2996399700641632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22744089365005493, 'sampling/sampling_logp_difference/max': 1.9784126281738281, 'sampling/importance_sampling_ratio/min': 0.13828857243061066, 'sampling/importance_sampling_ratio/mean': 1.0495731830596924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.804256483912468, 'clip_ratio/low_mean': 0.14473684411495924, 'clip_ratio/low_min': 0.14473684411495924, 'clip_ratio/high_mean': 0.07152777910232544, 'clip_ratio/high_max': 0.07152777910232544, 'clip_ratio/region_mean': 0.21626462321728468, 'reward_total_mean': 0.2279086858034134, 'reward_meter_mean': 0.5945756435394287, 'reward_meter_std': 0.4319617450237274, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.29624998569488525, 'reward_judge_quality_std': 0.26773855090141296, 'reward_total_composite_mean': 0.2279086858034134, 'reward_total_composite_std': 0.2996399700641632, 'epoch': 0.03}

 17%|█▋        | 555/3300 [2:01:34<16:11:37, 21.24s/it]INFO 04-12 18:00:01 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 556/3300 [2:01:43<13:25:59, 17.62s/it]2026-04-12 18:00:09,920 | INFO | train_grpo_train | metrics_logged mode=train step=556

                                                       
{'loss': 0.1307, 'grad_norm': 19.094688415527344, 'learning_rate': 8.318181818181818e-06, 'num_tokens': 1030183.0, 'completions/mean_length': 30.125, 'completions/min_length': 26.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.125, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.3635500967502594, 'rewards/meter/std': 0.27353185415267944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.492499977350235, 'rewards/judge_quality/std': 0.2796298861503601, 'rewards/repeat_penalty/mean': 0.9789423942565918, 'rewards/repeat_penalty/std': 0.03175196796655655, 'rewards/repeat_floor/mean': 0.9968413710594177, 'rewards/repeat_floor/std': 0.0047627887688577175, 'rewards/near_duplicate_penalty/mean': 0.9789423942565918, 'rewards/near_duplicate_penalty/std': 0.03175196796655655, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1594967097043991, 'rewards/total_composite/std': 0.14096620678901672, 'reward': 0.1594967097043991, 'reward_std': 0.14096622169017792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18325132131576538, 'sampling/sampling_logp_difference/max': 2.0480260848999023, 'sampling/importance_sampling_ratio/min': 0.12898926436901093, 'sampling/importance_sampling_ratio/mean': 0.9809970855712891, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8702652528882027, 'clip_ratio/low_mean': 0.09620174020528793, 'clip_ratio/low_min': 0.09620174020528793, 'clip_ratio/high_mean': 0.07234333548694849, 'clip_ratio/high_max': 0.07234333548694849, 'clip_ratio/region_mean': 0.16854507569223642, 'reward_total_mean': 0.1594967097043991, 'reward_meter_mean': 0.3635500967502594, 'reward_meter_std': 0.27353185415267944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9789423942565918, 'reward_repeat_penalty_std': 0.03175196796655655, 'reward_repeat_floor_mean': 0.9968413710594177, 'reward_repeat_floor_std': 0.0047627887688577175, 'reward_near_duplicate_penalty_mean': 0.9789423942565918, 'reward_near_duplicate_penalty_std': 0.03175196796655655, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.492499977350235, 'reward_judge_quality_std': 0.2796298861503601, 'reward_total_composite_mean': 0.1594967097043991, 'reward_total_composite_std': 0.14096620678901672, 'epoch': 0.03}

 17%|█▋        | 556/3300 [2:01:43<13:25:59, 17.62s/it]INFO 04-12 18:00:10 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 557/3300 [2:01:51<11:12:37, 14.71s/it]2026-04-12 18:00:17,842 | INFO | train_grpo_train | metrics_logged mode=train step=557

                                                       
{'loss': 0.0697, 'grad_norm': 33.31736373901367, 'learning_rate': 8.315151515151516e-06, 'num_tokens': 1031476.0, 'completions/mean_length': 17.625, 'completions/min_length': 16.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.4873313903808594, 'rewards/meter/std': 0.40829962491989136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8387500047683716, 'rewards/judge_quality/std': 0.27869275212287903, 'rewards/repeat_penalty/mean': 0.7358934283256531, 'rewards/repeat_penalty/std': 0.026865478605031967, 'rewards/repeat_floor/mean': 0.9821786880493164, 'rewards/repeat_floor/std': 0.0053730979561805725, 'rewards/near_duplicate_penalty/mean': 0.9811912178993225, 'rewards/near_duplicate_penalty/std': 0.03582063689827919, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4177014231681824, 'rewards/total_composite/std': 0.3959362804889679, 'reward': 0.4177014231681824, 'reward_std': 0.3959362506866455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14491981267929077, 'sampling/sampling_logp_difference/max': 1.1840513944625854, 'sampling/importance_sampling_ratio/min': 0.3060363531112671, 'sampling/importance_sampling_ratio/mean': 0.9935749173164368, 'sampling/importance_sampling_ratio/max': 1.8750287294387817, 'entropy': 0.8732364699244499, 'clip_ratio/low_mean': 0.05417956691235304, 'clip_ratio/low_min': 0.05417956691235304, 'clip_ratio/high_mean': 0.08652605768293142, 'clip_ratio/high_max': 0.08652605768293142, 'clip_ratio/region_mean': 0.14070562459528446, 'reward_total_mean': 0.4177014231681824, 'reward_meter_mean': 0.4873313903808594, 'reward_meter_std': 0.40829962491989136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7358934283256531, 'reward_repeat_penalty_std': 0.026865478605031967, 'reward_repeat_floor_mean': 0.9821786880493164, 'reward_repeat_floor_std': 0.0053730979561805725, 'reward_near_duplicate_penalty_mean': 0.9811912178993225, 'reward_near_duplicate_penalty_std': 0.03582063689827919, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8387500047683716, 'reward_judge_quality_std': 0.27869275212287903, 'reward_total_composite_mean': 0.4177014231681824, 'reward_total_composite_std': 0.3959362804889679, 'epoch': 0.03}

 17%|█▋        | 557/3300 [2:01:51<11:12:37, 14.71s/it]INFO 04-12 18:00:18 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 558/3300 [2:02:07<11:27:11, 15.04s/it]2026-04-12 18:00:33,607 | INFO | train_grpo_train | metrics_logged mode=train step=558

                                                       
{'loss': -0.164, 'grad_norm': 4.081277370452881, 'learning_rate': 8.312121212121213e-06, 'num_tokens': 1033472.0, 'completions/mean_length': 123.5, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.8523678779602051, 'rewards/meter/std': 0.33246737718582153, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.06943651288747787, 'rewards/lexical_plausibility/mean': 0.9511696100234985, 'rewards/lexical_plausibility/std': 0.13811323046684265, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.9099739193916321, 'rewards/repeat_penalty/std': 0.09572067111730576, 'rewards/repeat_floor/mean': 0.9904675483703613, 'rewards/repeat_floor/std': 0.008024848066270351, 'rewards/near_duplicate_penalty/mean': 0.9654771089553833, 'rewards/near_duplicate_penalty/std': 0.02118479460477829, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.972455620765686, 'rewards/distinct_2/std': 0.03426055610179901, 'rewards/total_composite/mean': 0.2987726032733917, 'rewards/total_composite/std': 0.1785968840122223, 'reward': 0.2987726032733917, 'reward_std': 0.1785968691110611, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.163870170712471, 'sampling/sampling_logp_difference/max': 1.990492582321167, 'sampling/importance_sampling_ratio/min': 0.13662810623645782, 'sampling/importance_sampling_ratio/mean': 1.0121395587921143, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0068789720535278, 'clip_ratio/low_mean': 0.023333333432674408, 'clip_ratio/low_min': 0.023333333432674408, 'clip_ratio/high_mean': 0.10875323042273521, 'clip_ratio/high_max': 0.10875323042273521, 'clip_ratio/region_mean': 0.13208656385540962, 'reward_total_mean': 0.2987726032733917, 'reward_meter_mean': 0.8523678779602051, 'reward_meter_std': 0.33246737718582153, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.06943651288747787, 'reward_lexical_plausibility_mean': 0.9511696100234985, 'reward_lexical_plausibility_std': 0.13811323046684265, 'reward_repeat_penalty_mean': 0.9099739193916321, 'reward_repeat_penalty_std': 0.09572067111730576, 'reward_repeat_floor_mean': 0.9904675483703613, 'reward_repeat_floor_std': 0.008024848066270351, 'reward_near_duplicate_penalty_mean': 0.9654771089553833, 'reward_near_duplicate_penalty_std': 0.02118479460477829, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.972455620765686, 'reward_distinct_2_std': 0.03426055610179901, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.2987726032733917, 'reward_total_composite_std': 0.1785968840122223, 'epoch': 0.03}

 17%|█▋        | 558/3300 [2:02:07<11:27:11, 15.04s/it]INFO 04-12 18:00:33 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 559/3300 [2:02:22<11:23:32, 14.96s/it]2026-04-12 18:00:48,395 | INFO | train_grpo_train | metrics_logged mode=train step=559

                                                       
{'loss': -0.0428, 'grad_norm': 1.2077611684799194, 'learning_rate': 8.30909090909091e-06, 'num_tokens': 1035074.0, 'completions/mean_length': 456.25, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7518603801727295, 'rewards/meter/std': 0.3013761043548584, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.2781743109226227, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9125000238418579, 'rewards/count_floor/std': 0.08345229923725128, 'rewards/lexical_plausibility/mean': 0.544519305229187, 'rewards/lexical_plausibility/std': 0.19723442196846008, 'rewards/judge_quality/mean': 0.1587500125169754, 'rewards/judge_quality/std': 0.3075914680957794, 'rewards/repeat_penalty/mean': 0.9920206665992737, 'rewards/repeat_penalty/std': 0.014792493544518948, 'rewards/repeat_floor/mean': 0.9988806247711182, 'rewards/repeat_floor/std': 0.002004166366532445, 'rewards/near_duplicate_penalty/mean': 0.9933830499649048, 'rewards/near_duplicate_penalty/std': 0.011038585565984249, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9985911846160889, 'rewards/distinct_2/std': 0.0039848159067332745, 'rewards/total_composite/mean': 0.11180394142866135, 'rewards/total_composite/std': 0.22277867794036865, 'reward': 0.11180394142866135, 'reward_std': 0.22277867794036865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21221083402633667, 'sampling/sampling_logp_difference/max': 1.367455244064331, 'sampling/importance_sampling_ratio/min': 0.2547544240951538, 'sampling/importance_sampling_ratio/mean': 1.0518114566802979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.1787753403186798, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.01515151560306549, 'clip_ratio/high_max': 0.01515151560306549, 'clip_ratio/region_mean': 0.01515151560306549, 'reward_total_mean': 0.11180394142866135, 'reward_meter_mean': 0.7518603801727295, 'reward_meter_std': 0.3013761043548584, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.2781743109226227, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9125000238418579, 'reward_count_floor_std': 0.08345229923725128, 'reward_lexical_plausibility_mean': 0.544519305229187, 'reward_lexical_plausibility_std': 0.19723442196846008, 'reward_repeat_penalty_mean': 0.9920206665992737, 'reward_repeat_penalty_std': 0.014792493544518948, 'reward_repeat_floor_mean': 0.9988806247711182, 'reward_repeat_floor_std': 0.002004166366532445, 'reward_near_duplicate_penalty_mean': 0.9933830499649048, 'reward_near_duplicate_penalty_std': 0.011038585565984249, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9985911846160889, 'reward_distinct_2_std': 0.0039848159067332745, 'reward_judge_quality_mean': 0.1587500125169754, 'reward_judge_quality_std': 0.3075914680957794, 'reward_total_composite_mean': 0.11180394142866135, 'reward_total_composite_std': 0.22277867794036865, 'epoch': 0.03}

 17%|█▋        | 559/3300 [2:02:22<11:23:32, 14.96s/it]INFO 04-12 18:00:48 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 560/3300 [2:02:37<11:22:05, 14.94s/it]2026-04-12 18:01:03,299 | INFO | train_grpo_train | metrics_logged mode=train step=560

                                                       
{'loss': -0.0636, 'grad_norm': 5.534683704376221, 'learning_rate': 8.306060606060606e-06, 'num_tokens': 1036620.0, 'completions/mean_length': 97.25, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5694347620010376, 'rewards/meter/std': 0.4263285994529724, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9341445565223694, 'rewards/lexical_plausibility/std': 0.18626733124256134, 'rewards/judge_quality/mean': 0.29625001549720764, 'rewards/judge_quality/std': 0.283293753862381, 'rewards/repeat_penalty/mean': 0.8989452123641968, 'rewards/repeat_penalty/std': 0.11513609439134598, 'rewards/repeat_floor/mean': 0.9894053339958191, 'rewards/repeat_floor/std': 0.011531993746757507, 'rewards/near_duplicate_penalty/mean': 0.9644253253936768, 'rewards/near_duplicate_penalty/std': 0.036088768392801285, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9624060392379761, 'rewards/distinct_2/std': 0.07887901365756989, 'rewards/total_composite/mean': 0.22839999198913574, 'rewards/total_composite/std': 0.2931581437587738, 'reward': 0.22839999198913574, 'reward_std': 0.2931581735610962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15873941779136658, 'sampling/sampling_logp_difference/max': 1.1122474670410156, 'sampling/importance_sampling_ratio/min': 0.3288191258907318, 'sampling/importance_sampling_ratio/mean': 1.0212668180465698, 'sampling/importance_sampling_ratio/max': 1.743553638458252, 'entropy': 1.5165867358446121, 'clip_ratio/low_mean': 0.05816003493964672, 'clip_ratio/low_min': 0.05816003493964672, 'clip_ratio/high_mean': 0.04723785724490881, 'clip_ratio/high_max': 0.04723785724490881, 'clip_ratio/region_mean': 0.10539789218455553, 'reward_total_mean': 0.22839999198913574, 'reward_meter_mean': 0.5694347620010376, 'reward_meter_std': 0.4263285994529724, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9341445565223694, 'reward_lexical_plausibility_std': 0.18626733124256134, 'reward_repeat_penalty_mean': 0.8989452123641968, 'reward_repeat_penalty_std': 0.11513609439134598, 'reward_repeat_floor_mean': 0.9894053339958191, 'reward_repeat_floor_std': 0.011531993746757507, 'reward_near_duplicate_penalty_mean': 0.9644253253936768, 'reward_near_duplicate_penalty_std': 0.036088768392801285, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9624060392379761, 'reward_distinct_2_std': 0.07887901365756989, 'reward_judge_quality_mean': 0.29625001549720764, 'reward_judge_quality_std': 0.283293753862381, 'reward_total_composite_mean': 0.22839999198913574, 'reward_total_composite_std': 0.2931581437587738, 'epoch': 0.03}

 17%|█▋        | 560/3300 [2:02:37<11:22:05, 14.94s/it]INFO 04-12 18:01:03 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 561/3300 [2:02:51<11:19:01, 14.87s/it]2026-04-12 18:01:18,001 | INFO | train_grpo_train | metrics_logged mode=train step=561

                                                       
{'loss': -0.1271, 'grad_norm': 2.5300896167755127, 'learning_rate': 8.303030303030305e-06, 'num_tokens': 1038453.0, 'completions/mean_length': 248.125, 'completions/min_length': 70.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 89.80000305175781, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9038985967636108, 'rewards/meter/std': 0.1426585167646408, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.28149792551994324, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.934374988079071, 'rewards/count_floor/std': 0.08444938063621521, 'rewards/lexical_plausibility/mean': 0.8190855979919434, 'rewards/lexical_plausibility/std': 0.25018277764320374, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.8170636892318726, 'rewards/repeat_penalty/std': 0.17387215793132782, 'rewards/repeat_floor/mean': 0.9830120801925659, 'rewards/repeat_floor/std': 0.01619563437998295, 'rewards/near_duplicate_penalty/mean': 0.9651481509208679, 'rewards/near_duplicate_penalty/std': 0.03389329835772514, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.9193319082260132, 'rewards/distinct_2/std': 0.08834436535835266, 'rewards/total_composite/mean': 0.20498758554458618, 'rewards/total_composite/std': 0.16108544170856476, 'reward': 0.20498758554458618, 'reward_std': 0.16108545660972595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17830517888069153, 'sampling/sampling_logp_difference/max': 1.7236967086791992, 'sampling/importance_sampling_ratio/min': 0.17840541899204254, 'sampling/importance_sampling_ratio/mean': 1.0018104314804077, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8101192861795425, 'clip_ratio/low_mean': 0.022836538031697273, 'clip_ratio/low_min': 0.022836538031697273, 'clip_ratio/high_mean': 0.08320409059524536, 'clip_ratio/high_max': 0.08320409059524536, 'clip_ratio/region_mean': 0.10604062862694263, 'reward_total_mean': 0.20498758554458618, 'reward_meter_mean': 0.9038985967636108, 'reward_meter_std': 0.1426585167646408, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.28149792551994324, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.934374988079071, 'reward_count_floor_std': 0.08444938063621521, 'reward_lexical_plausibility_mean': 0.8190855979919434, 'reward_lexical_plausibility_std': 0.25018277764320374, 'reward_repeat_penalty_mean': 0.8170636892318726, 'reward_repeat_penalty_std': 0.17387215793132782, 'reward_repeat_floor_mean': 0.9830120801925659, 'reward_repeat_floor_std': 0.01619563437998295, 'reward_near_duplicate_penalty_mean': 0.9651481509208679, 'reward_near_duplicate_penalty_std': 0.03389329835772514, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.9193319082260132, 'reward_distinct_2_std': 0.08834436535835266, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.20498758554458618, 'reward_total_composite_std': 0.16108544170856476, 'epoch': 0.03}

 17%|█▋        | 561/3300 [2:02:51<11:19:01, 14.87s/it]INFO 04-12 18:01:18 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 562/3300 [2:03:05<11:07:15, 14.62s/it]2026-04-12 18:01:32,035 | INFO | train_grpo_train | metrics_logged mode=train step=562

                                                       
{'loss': -0.066, 'grad_norm': 2.014894723892212, 'learning_rate': 8.3e-06, 'num_tokens': 1039999.0, 'completions/mean_length': 218.25, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.5807574391365051, 'rewards/meter/std': 0.44241222739219666, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.7765780687332153, 'rewards/lexical_plausibility/std': 0.28427284955978394, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.25599944591522217, 'rewards/repeat_penalty/mean': 0.8697566390037537, 'rewards/repeat_penalty/std': 0.1453157365322113, 'rewards/repeat_floor/mean': 0.9894654750823975, 'rewards/repeat_floor/std': 0.011609818786382675, 'rewards/near_duplicate_penalty/mean': 0.9827976226806641, 'rewards/near_duplicate_penalty/std': 0.02610170468688011, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9741200804710388, 'rewards/distinct_2/std': 0.050576355308294296, 'rewards/total_composite/mean': 0.2178936004638672, 'rewards/total_composite/std': 0.22457318007946014, 'reward': 0.2178936004638672, 'reward_std': 0.22457316517829895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18242985010147095, 'sampling/sampling_logp_difference/max': 1.454162836074829, 'sampling/importance_sampling_ratio/min': 0.2335958480834961, 'sampling/importance_sampling_ratio/mean': 1.0564383268356323, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2552347481250763, 'clip_ratio/low_mean': 0.02820512861944735, 'clip_ratio/low_min': 0.02820512861944735, 'clip_ratio/high_mean': 0.0937925186008215, 'clip_ratio/high_max': 0.0937925186008215, 'clip_ratio/region_mean': 0.12199764722026885, 'reward_total_mean': 0.2178936004638672, 'reward_meter_mean': 0.5807574391365051, 'reward_meter_std': 0.44241222739219666, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.7765780687332153, 'reward_lexical_plausibility_std': 0.28427284955978394, 'reward_repeat_penalty_mean': 0.8697566390037537, 'reward_repeat_penalty_std': 0.1453157365322113, 'reward_repeat_floor_mean': 0.9894654750823975, 'reward_repeat_floor_std': 0.011609818786382675, 'reward_near_duplicate_penalty_mean': 0.9827976226806641, 'reward_near_duplicate_penalty_std': 0.02610170468688011, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9741200804710388, 'reward_distinct_2_std': 0.050576355308294296, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.25599944591522217, 'reward_total_composite_mean': 0.2178936004638672, 'reward_total_composite_std': 0.22457318007946014, 'epoch': 0.03}

 17%|█▋        | 562/3300 [2:03:05<11:07:15, 14.62s/it]INFO 04-12 18:01:32 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 563/3300 [2:03:19<10:51:30, 14.28s/it]2026-04-12 18:01:45,524 | INFO | train_grpo_train | metrics_logged mode=train step=563

                                                       
{'loss': -0.0549, 'grad_norm': 4.267290115356445, 'learning_rate': 8.296969696969697e-06, 'num_tokens': 1041743.0, 'completions/mean_length': 103.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.57143020629883, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7931816577911377, 'rewards/meter/std': 0.34101036190986633, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9268641471862793, 'rewards/lexical_plausibility/std': 0.20685933530330658, 'rewards/judge_quality/mean': 0.5799999833106995, 'rewards/judge_quality/std': 0.31995537877082825, 'rewards/repeat_penalty/mean': 0.9798264503479004, 'rewards/repeat_penalty/std': 0.031223196536302567, 'rewards/repeat_floor/mean': 0.9969739317893982, 'rewards/repeat_floor/std': 0.004683476872742176, 'rewards/near_duplicate_penalty/mean': 0.9798264503479004, 'rewards/near_duplicate_penalty/std': 0.031223196536302567, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5329734086990356, 'rewards/total_composite/std': 0.3447689414024353, 'reward': 0.5329734086990356, 'reward_std': 0.3447689414024353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17494451999664307, 'sampling/sampling_logp_difference/max': 1.3205337524414062, 'sampling/importance_sampling_ratio/min': 0.2669927477836609, 'sampling/importance_sampling_ratio/mean': 1.0090452432632446, 'sampling/importance_sampling_ratio/max': 1.9617542028427124, 'entropy': 1.2997402250766754, 'clip_ratio/low_mean': 0.0697510838508606, 'clip_ratio/low_min': 0.0697510838508606, 'clip_ratio/high_mean': 0.057986111380159855, 'clip_ratio/high_max': 0.057986111380159855, 'clip_ratio/region_mean': 0.12773719523102045, 'reward_total_mean': 0.5329734086990356, 'reward_meter_mean': 0.7931816577911377, 'reward_meter_std': 0.34101036190986633, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9268641471862793, 'reward_lexical_plausibility_std': 0.20685933530330658, 'reward_repeat_penalty_mean': 0.9798264503479004, 'reward_repeat_penalty_std': 0.031223196536302567, 'reward_repeat_floor_mean': 0.9969739317893982, 'reward_repeat_floor_std': 0.004683476872742176, 'reward_near_duplicate_penalty_mean': 0.9798264503479004, 'reward_near_duplicate_penalty_std': 0.031223196536302567, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5799999833106995, 'reward_judge_quality_std': 0.31995537877082825, 'reward_total_composite_mean': 0.5329734086990356, 'reward_total_composite_std': 0.3447689414024353, 'epoch': 0.03}

 17%|█▋        | 563/3300 [2:03:19<10:51:30, 14.28s/it]INFO 04-12 18:01:45 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 564/3300 [2:03:33<10:46:40, 14.18s/it]2026-04-12 18:01:59,469 | INFO | train_grpo_train | metrics_logged mode=train step=564

                                                       
{'loss': -0.0951, 'grad_norm': 2.4203593730926514, 'learning_rate': 8.293939393939395e-06, 'num_tokens': 1043364.0, 'completions/mean_length': 162.625, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 46.16666793823242, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8536757826805115, 'rewards/meter/std': 0.30933067202568054, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8828497529029846, 'rewards/lexical_plausibility/std': 0.21698138117790222, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.30875036120414734, 'rewards/repeat_penalty/mean': 0.8874882459640503, 'rewards/repeat_penalty/std': 0.1500159651041031, 'rewards/repeat_floor/mean': 0.9917948246002197, 'rewards/repeat_floor/std': 0.01065937802195549, 'rewards/near_duplicate_penalty/mean': 0.9933937191963196, 'rewards/near_duplicate_penalty/std': 0.009810215793550014, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9823417663574219, 'rewards/distinct_2/std': 0.03137263283133507, 'rewards/total_composite/mean': 0.4434981644153595, 'rewards/total_composite/std': 0.305878609418869, 'reward': 0.4434981644153595, 'reward_std': 0.30587857961654663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15061204135417938, 'sampling/sampling_logp_difference/max': 1.1129411458969116, 'sampling/importance_sampling_ratio/min': 0.32859110832214355, 'sampling/importance_sampling_ratio/mean': 1.040624976158142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9117092192173004, 'clip_ratio/low_mean': 0.06774215400218964, 'clip_ratio/low_min': 0.06774215400218964, 'clip_ratio/high_mean': 0.04718712717294693, 'clip_ratio/high_max': 0.04718712717294693, 'clip_ratio/region_mean': 0.11492928117513657, 'reward_total_mean': 0.4434981644153595, 'reward_meter_mean': 0.8536757826805115, 'reward_meter_std': 0.30933067202568054, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8828497529029846, 'reward_lexical_plausibility_std': 0.21698138117790222, 'reward_repeat_penalty_mean': 0.8874882459640503, 'reward_repeat_penalty_std': 0.1500159651041031, 'reward_repeat_floor_mean': 0.9917948246002197, 'reward_repeat_floor_std': 0.01065937802195549, 'reward_near_duplicate_penalty_mean': 0.9933937191963196, 'reward_near_duplicate_penalty_std': 0.009810215793550014, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9823417663574219, 'reward_distinct_2_std': 0.03137263283133507, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.30875036120414734, 'reward_total_composite_mean': 0.4434981644153595, 'reward_total_composite_std': 0.305878609418869, 'epoch': 0.03}

 17%|█▋        | 564/3300 [2:03:33<10:46:40, 14.18s/it]INFO 04-12 18:01:59 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 565/3300 [2:03:47<10:42:20, 14.09s/it]2026-04-12 18:02:13,350 | INFO | train_grpo_train | metrics_logged mode=train step=565

                                                       
{'loss': -0.0916, 'grad_norm': 3.097459316253662, 'learning_rate': 8.290909090909092e-06, 'num_tokens': 1044987.0, 'completions/mean_length': 172.875, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 59.833335876464844, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.6703800559043884, 'rewards/meter/std': 0.4046465754508972, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.8771954774856567, 'rewards/lexical_plausibility/std': 0.22751611471176147, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.6906857490539551, 'rewards/repeat_penalty/std': 0.31215834617614746, 'rewards/repeat_floor/mean': 0.9587532877922058, 'rewards/repeat_floor/std': 0.04403448477387428, 'rewards/near_duplicate_penalty/mean': 0.8700329065322876, 'rewards/near_duplicate_penalty/std': 0.11799958348274231, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.7874820232391357, 'rewards/distinct_2/std': 0.24727487564086914, 'rewards/total_composite/mean': 0.212692528963089, 'rewards/total_composite/std': 0.18281516432762146, 'reward': 0.212692528963089, 'reward_std': 0.18281514942646027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12997393310070038, 'sampling/sampling_logp_difference/max': 1.0370960235595703, 'sampling/importance_sampling_ratio/min': 0.35448259115219116, 'sampling/importance_sampling_ratio/mean': 1.046034574508667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7987740859389305, 'clip_ratio/low_mean': 0.01456447970122099, 'clip_ratio/low_min': 0.01456447970122099, 'clip_ratio/high_mean': 0.06853776425123215, 'clip_ratio/high_max': 0.06853776425123215, 'clip_ratio/region_mean': 0.08310224395245314, 'reward_total_mean': 0.212692528963089, 'reward_meter_mean': 0.6703800559043884, 'reward_meter_std': 0.4046465754508972, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.8771954774856567, 'reward_lexical_plausibility_std': 0.22751611471176147, 'reward_repeat_penalty_mean': 0.6906857490539551, 'reward_repeat_penalty_std': 0.31215834617614746, 'reward_repeat_floor_mean': 0.9587532877922058, 'reward_repeat_floor_std': 0.04403448477387428, 'reward_near_duplicate_penalty_mean': 0.8700329065322876, 'reward_near_duplicate_penalty_std': 0.11799958348274231, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.7874820232391357, 'reward_distinct_2_std': 0.24727487564086914, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.212692528963089, 'reward_total_composite_std': 0.18281516432762146, 'epoch': 0.03}

 17%|█▋        | 565/3300 [2:03:47<10:42:20, 14.09s/it]INFO 04-12 18:02:13 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 566/3300 [2:04:01<10:48:03, 14.22s/it]2026-04-12 18:02:27,877 | INFO | train_grpo_train | metrics_logged mode=train step=566

                                                       
{'loss': -0.0738, 'grad_norm': 4.196310520172119, 'learning_rate': 8.287878787878787e-06, 'num_tokens': 1046916.0, 'completions/mean_length': 124.125, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 68.71428680419922, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9254423379898071, 'rewards/meter/std': 0.10226458311080933, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9500000476837158, 'rewards/count_floor/std': 0.07559289783239365, 'rewards/lexical_plausibility/mean': 0.9235543012619019, 'rewards/lexical_plausibility/std': 0.21622098982334137, 'rewards/judge_quality/mean': 0.4949999749660492, 'rewards/judge_quality/std': 0.2734958827495575, 'rewards/repeat_penalty/mean': 0.9449768662452698, 'rewards/repeat_penalty/std': 0.0846562534570694, 'rewards/repeat_floor/mean': 0.9951215982437134, 'rewards/repeat_floor/std': 0.0052273934707045555, 'rewards/near_duplicate_penalty/mean': 0.9858943223953247, 'rewards/near_duplicate_penalty/std': 0.01343545876443386, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.990138053894043, 'rewards/distinct_2/std': 0.027893755584955215, 'rewards/total_composite/mean': 0.4442390501499176, 'rewards/total_composite/std': 0.2518474757671356, 'reward': 0.4442390501499176, 'reward_std': 0.2518474757671356, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1668618768453598, 'sampling/sampling_logp_difference/max': 1.4340158700942993, 'sampling/importance_sampling_ratio/min': 0.2383498251438141, 'sampling/importance_sampling_ratio/mean': 1.028917908668518, 'sampling/importance_sampling_ratio/max': 1.8224741220474243, 'entropy': 1.2820668667554855, 'clip_ratio/low_mean': 0.059490740299224854, 'clip_ratio/low_min': 0.059490740299224854, 'clip_ratio/high_mean': 0.0811594258993864, 'clip_ratio/high_max': 0.0811594258993864, 'clip_ratio/region_mean': 0.14065016619861126, 'reward_total_mean': 0.4442390501499176, 'reward_meter_mean': 0.9254423379898071, 'reward_meter_std': 0.10226458311080933, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9500000476837158, 'reward_count_floor_std': 0.07559289783239365, 'reward_lexical_plausibility_mean': 0.9235543012619019, 'reward_lexical_plausibility_std': 0.21622098982334137, 'reward_repeat_penalty_mean': 0.9449768662452698, 'reward_repeat_penalty_std': 0.0846562534570694, 'reward_repeat_floor_mean': 0.9951215982437134, 'reward_repeat_floor_std': 0.0052273934707045555, 'reward_near_duplicate_penalty_mean': 0.9858943223953247, 'reward_near_duplicate_penalty_std': 0.01343545876443386, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.990138053894043, 'reward_distinct_2_std': 0.027893755584955215, 'reward_judge_quality_mean': 0.4949999749660492, 'reward_judge_quality_std': 0.2734958827495575, 'reward_total_composite_mean': 0.4442390501499176, 'reward_total_composite_std': 0.2518474757671356, 'epoch': 0.03}

 17%|█▋        | 566/3300 [2:04:01<10:48:03, 14.22s/it]INFO 04-12 18:02:28 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 567/3300 [2:04:16<10:54:27, 14.37s/it]2026-04-12 18:02:42,586 | INFO | train_grpo_train | metrics_logged mode=train step=567

                                                       
{'loss': -0.1443, 'grad_norm': 2.7742621898651123, 'learning_rate': 8.284848484848486e-06, 'num_tokens': 1048737.0, 'completions/mean_length': 106.625, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.71428680419922, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8627701997756958, 'rewards/meter/std': 0.2515907287597656, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9394711852073669, 'rewards/lexical_plausibility/std': 0.17120131850242615, 'rewards/judge_quality/mean': 0.39374998211860657, 'rewards/judge_quality/std': 0.1590990275144577, 'rewards/repeat_penalty/mean': 0.9668909311294556, 'rewards/repeat_penalty/std': 0.05515839904546738, 'rewards/repeat_floor/mean': 0.9957719445228577, 'rewards/repeat_floor/std': 0.006334956735372543, 'rewards/near_duplicate_penalty/mean': 0.9803599119186401, 'rewards/near_duplicate_penalty/std': 0.022773869335651398, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9857549667358398, 'rewards/distinct_2/std': 0.04029097780585289, 'rewards/total_composite/mean': 0.36731523275375366, 'rewards/total_composite/std': 0.14951860904693604, 'reward': 0.36731523275375366, 'reward_std': 0.14951860904693604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18495750427246094, 'sampling/sampling_logp_difference/max': 1.7382721900939941, 'sampling/importance_sampling_ratio/min': 0.23729169368743896, 'sampling/importance_sampling_ratio/mean': 1.0208741426467896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1288677677512169, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.17848812229931355, 'clip_ratio/high_max': 0.17848812229931355, 'clip_ratio/region_mean': 0.17848812229931355, 'reward_total_mean': 0.36731523275375366, 'reward_meter_mean': 0.8627701997756958, 'reward_meter_std': 0.2515907287597656, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9394711852073669, 'reward_lexical_plausibility_std': 0.17120131850242615, 'reward_repeat_penalty_mean': 0.9668909311294556, 'reward_repeat_penalty_std': 0.05515839904546738, 'reward_repeat_floor_mean': 0.9957719445228577, 'reward_repeat_floor_std': 0.006334956735372543, 'reward_near_duplicate_penalty_mean': 0.9803599119186401, 'reward_near_duplicate_penalty_std': 0.022773869335651398, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9857549667358398, 'reward_distinct_2_std': 0.04029097780585289, 'reward_judge_quality_mean': 0.39374998211860657, 'reward_judge_quality_std': 0.1590990275144577, 'reward_total_composite_mean': 0.36731523275375366, 'reward_total_composite_std': 0.14951860904693604, 'epoch': 0.03}

 17%|█▋        | 567/3300 [2:04:16<10:54:27, 14.37s/it]INFO 04-12 18:02:42 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 568/3300 [2:04:30<10:49:03, 14.25s/it]2026-04-12 18:02:56,579 | INFO | train_grpo_train | metrics_logged mode=train step=568

                                                       
{'loss': -0.1227, 'grad_norm': 5.495846748352051, 'learning_rate': 8.281818181818182e-06, 'num_tokens': 1051109.0, 'completions/mean_length': 165.5, 'completions/min_length': 98.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 116.00000762939453, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.7863644361495972, 'rewards/meter/std': 0.23657944798469543, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.28171807527542114, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9500000476837158, 'rewards/count_floor/std': 0.08451542258262634, 'rewards/lexical_plausibility/mean': 0.9286943674087524, 'rewards/lexical_plausibility/std': 0.20168279111385345, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.8176130652427673, 'rewards/repeat_penalty/std': 0.17285406589508057, 'rewards/repeat_floor/mean': 0.9803366661071777, 'rewards/repeat_floor/std': 0.018711449578404427, 'rewards/near_duplicate_penalty/mean': 0.9565730094909668, 'rewards/near_duplicate_penalty/std': 0.038160938769578934, 'rewards/opening_diversity/mean': 0.9955357313156128, 'rewards/opening_diversity/std': 0.012626901268959045, 'rewards/distinct_2/mean': 0.8568733930587769, 'rewards/distinct_2/std': 0.13849873840808868, 'rewards/total_composite/mean': 0.23523476719856262, 'rewards/total_composite/std': 0.15004345774650574, 'reward': 0.23523476719856262, 'reward_std': 0.15004344284534454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17758242785930634, 'sampling/sampling_logp_difference/max': 2.321695327758789, 'sampling/importance_sampling_ratio/min': 0.09810712188482285, 'sampling/importance_sampling_ratio/mean': 1.0052388906478882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9297135025262833, 'clip_ratio/low_mean': 0.060063484124839306, 'clip_ratio/low_min': 0.060063484124839306, 'clip_ratio/high_mean': 0.08620619587600231, 'clip_ratio/high_max': 0.08620619587600231, 'clip_ratio/region_mean': 0.14626968000084162, 'reward_total_mean': 0.23523476719856262, 'reward_meter_mean': 0.7863644361495972, 'reward_meter_std': 0.23657944798469543, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.28171807527542114, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9500000476837158, 'reward_count_floor_std': 0.08451542258262634, 'reward_lexical_plausibility_mean': 0.9286943674087524, 'reward_lexical_plausibility_std': 0.20168279111385345, 'reward_repeat_penalty_mean': 0.8176130652427673, 'reward_repeat_penalty_std': 0.17285406589508057, 'reward_repeat_floor_mean': 0.9803366661071777, 'reward_repeat_floor_std': 0.018711449578404427, 'reward_near_duplicate_penalty_mean': 0.9565730094909668, 'reward_near_duplicate_penalty_std': 0.038160938769578934, 'reward_opening_diversity_mean': 0.9955357313156128, 'reward_opening_diversity_std': 0.012626901268959045, 'reward_distinct_2_mean': 0.8568733930587769, 'reward_distinct_2_std': 0.13849873840808868, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.23523476719856262, 'reward_total_composite_std': 0.15004345774650574, 'epoch': 0.03}

 17%|█▋        | 568/3300 [2:04:30<10:49:03, 14.25s/it]INFO 04-12 18:02:56 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 569/3300 [2:04:43<10:29:01, 13.82s/it]2026-04-12 18:03:09,386 | INFO | train_grpo_train | metrics_logged mode=train step=569

                                                       
{'loss': -0.0502, 'grad_norm': 4.771590232849121, 'learning_rate': 8.27878787878788e-06, 'num_tokens': 1052496.0, 'completions/mean_length': 81.375, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.85714340209961, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7215157747268677, 'rewards/meter/std': 0.38841956853866577, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.928816020488739, 'rewards/lexical_plausibility/std': 0.2013387531042099, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3160809278488159, 'rewards/repeat_penalty/mean': 0.7777406573295593, 'rewards/repeat_penalty/std': 0.09034208953380585, 'rewards/repeat_floor/mean': 0.986173152923584, 'rewards/repeat_floor/std': 0.005922330543398857, 'rewards/near_duplicate_penalty/mean': 0.9953208565711975, 'rewards/near_duplicate_penalty/std': 0.013234616257250309, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.39093366265296936, 'rewards/total_composite/std': 0.3158460855484009, 'reward': 0.39093366265296936, 'reward_std': 0.3158460855484009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18371812999248505, 'sampling/sampling_logp_difference/max': 1.2501325607299805, 'sampling/importance_sampling_ratio/min': 0.2864668071269989, 'sampling/importance_sampling_ratio/mean': 1.0116584300994873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9686591550707817, 'clip_ratio/low_mean': 0.045833335258066654, 'clip_ratio/low_min': 0.045833335258066654, 'clip_ratio/high_mean': 0.12258522864431143, 'clip_ratio/high_max': 0.12258522864431143, 'clip_ratio/region_mean': 0.16841856390237808, 'reward_total_mean': 0.39093366265296936, 'reward_meter_mean': 0.7215157747268677, 'reward_meter_std': 0.38841956853866577, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.928816020488739, 'reward_lexical_plausibility_std': 0.2013387531042099, 'reward_repeat_penalty_mean': 0.7777406573295593, 'reward_repeat_penalty_std': 0.09034208953380585, 'reward_repeat_floor_mean': 0.986173152923584, 'reward_repeat_floor_std': 0.005922330543398857, 'reward_near_duplicate_penalty_mean': 0.9953208565711975, 'reward_near_duplicate_penalty_std': 0.013234616257250309, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3160809278488159, 'reward_total_composite_mean': 0.39093366265296936, 'reward_total_composite_std': 0.3158460855484009, 'epoch': 0.03}

 17%|█▋        | 569/3300 [2:04:43<10:29:01, 13.82s/it]INFO 04-12 18:03:09 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 570/3300 [2:04:51<9:13:54, 12.17s/it] 2026-04-12 18:03:17,714 | INFO | train_grpo_train | metrics_logged mode=train step=570

                                                      
{'loss': 0.0405, 'grad_norm': 16.06989288330078, 'learning_rate': 8.275757575757577e-06, 'num_tokens': 1053983.0, 'completions/mean_length': 36.875, 'completions/min_length': 29.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7739244103431702, 'rewards/meter/std': 0.34029388427734375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.2774340510368347, 'rewards/repeat_penalty/mean': 0.9469379186630249, 'rewards/repeat_penalty/std': 0.0600096620619297, 'rewards/repeat_floor/mean': 0.9929990768432617, 'rewards/repeat_floor/std': 0.007500375155359507, 'rewards/near_duplicate_penalty/mean': 0.9648656845092773, 'rewards/near_duplicate_penalty/std': 0.034399982541799545, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.03560846298933029, 'rewards/total_composite/mean': 0.2977449893951416, 'rewards/total_composite/std': 0.24070042371749878, 'reward': 0.2977449893951416, 'reward_std': 0.24070042371749878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16520251333713531, 'sampling/sampling_logp_difference/max': 2.277348041534424, 'sampling/importance_sampling_ratio/min': 0.10255581885576248, 'sampling/importance_sampling_ratio/mean': 1.0325301885604858, 'sampling/importance_sampling_ratio/max': 1.757184624671936, 'entropy': 1.1247583031654358, 'clip_ratio/low_mean': 0.062312956899404526, 'clip_ratio/low_min': 0.062312956899404526, 'clip_ratio/high_mean': 0.06243589986115694, 'clip_ratio/high_max': 0.06243589986115694, 'clip_ratio/region_mean': 0.12474885676056147, 'reward_total_mean': 0.2977449893951416, 'reward_meter_mean': 0.7739244103431702, 'reward_meter_std': 0.34029388427734375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9469379186630249, 'reward_repeat_penalty_std': 0.0600096620619297, 'reward_repeat_floor_mean': 0.9929990768432617, 'reward_repeat_floor_std': 0.007500375155359507, 'reward_near_duplicate_penalty_mean': 0.9648656845092773, 'reward_near_duplicate_penalty_std': 0.034399982541799545, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.03560846298933029, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.2774340510368347, 'reward_total_composite_mean': 0.2977449893951416, 'reward_total_composite_std': 0.24070042371749878, 'epoch': 0.03}

 17%|█▋        | 570/3300 [2:04:51<9:13:54, 12.17s/it]INFO 04-12 18:03:17 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 571/3300 [2:05:03<9:04:57, 11.98s/it]2026-04-12 18:03:29,251 | INFO | train_grpo_train | metrics_logged mode=train step=571

                                                      
{'loss': 0.1376, 'grad_norm': 7.757911205291748, 'learning_rate': 8.272727272727274e-06, 'num_tokens': 1057017.0, 'completions/mean_length': 190.25, 'completions/min_length': 139.0, 'completions/max_length': 239.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 190.25, 'completions/min_terminated_length': 139.0, 'completions/max_terminated_length': 239.0, 'rewards/meter/mean': 0.9191886186599731, 'rewards/meter/std': 0.1342712640762329, 'rewards/count_adherence/mean': 0.8392857313156128, 'rewards/count_adherence/std': 0.141575887799263, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.951785683631897, 'rewards/count_floor/std': 0.04247276857495308, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.4537621736526489, 'rewards/repeat_penalty/std': 0.2060713768005371, 'rewards/repeat_floor/mean': 0.9466707706451416, 'rewards/repeat_floor/std': 0.022419724613428116, 'rewards/near_duplicate_penalty/mean': 0.9332298040390015, 'rewards/near_duplicate_penalty/std': 0.02776387333869934, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5187369585037231, 'rewards/distinct_2/std': 0.21405723690986633, 'rewards/total_composite/mean': 0.21533070504665375, 'rewards/total_composite/std': 0.08425196260213852, 'reward': 0.21533070504665375, 'reward_std': 0.08425196260213852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13699856400489807, 'sampling/sampling_logp_difference/max': 2.8776745796203613, 'sampling/importance_sampling_ratio/min': 0.0562654510140419, 'sampling/importance_sampling_ratio/mean': 1.0142626762390137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.869613628834486, 'clip_ratio/low_mean': 0.06622637994587421, 'clip_ratio/low_min': 0.06622637994587421, 'clip_ratio/high_mean': 0.0607346361503005, 'clip_ratio/high_max': 0.0607346361503005, 'clip_ratio/region_mean': 0.12696101609617472, 'reward_total_mean': 0.21533070504665375, 'reward_meter_mean': 0.9191886186599731, 'reward_meter_std': 0.1342712640762329, 'reward_count_adherence_mean': 0.8392857313156128, 'reward_count_adherence_std': 0.141575887799263, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.951785683631897, 'reward_count_floor_std': 0.04247276857495308, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4537621736526489, 'reward_repeat_penalty_std': 0.2060713768005371, 'reward_repeat_floor_mean': 0.9466707706451416, 'reward_repeat_floor_std': 0.022419724613428116, 'reward_near_duplicate_penalty_mean': 0.9332298040390015, 'reward_near_duplicate_penalty_std': 0.02776387333869934, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5187369585037231, 'reward_distinct_2_std': 0.21405723690986633, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.21533070504665375, 'reward_total_composite_std': 0.08425196260213852, 'epoch': 0.03}

 17%|█▋        | 571/3300 [2:05:03<9:04:57, 11.98s/it]INFO 04-12 18:03:29 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 572/3300 [2:05:16<9:27:52, 12.49s/it]2026-04-12 18:03:42,923 | INFO | train_grpo_train | metrics_logged mode=train step=572

                                                      
{'loss': -0.0459, 'grad_norm': 2.5027718544006348, 'learning_rate': 8.269696969696971e-06, 'num_tokens': 1058380.0, 'completions/mean_length': 215.375, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 37.400001525878906, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.31860774755477905, 'rewards/meter/std': 0.3715546727180481, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8237892389297485, 'rewards/lexical_plausibility/std': 0.24426932632923126, 'rewards/judge_quality/mean': 0.2587500214576721, 'rewards/judge_quality/std': 0.2995920181274414, 'rewards/repeat_penalty/mean': 0.9666010141372681, 'rewards/repeat_penalty/std': 0.08760970085859299, 'rewards/repeat_floor/mean': 0.9978353381156921, 'rewards/repeat_floor/std': 0.005218965467065573, 'rewards/near_duplicate_penalty/mean': 0.9983958005905151, 'rewards/near_duplicate_penalty/std': 0.004009816329926252, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9994552135467529, 'rewards/distinct_2/std': 0.0015408677281811833, 'rewards/total_composite/mean': 0.09329602867364883, 'rewards/total_composite/std': 0.12459640204906464, 'reward': 0.09329602867364883, 'reward_std': 0.12459640204906464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17604339122772217, 'sampling/sampling_logp_difference/max': 2.6610355377197266, 'sampling/importance_sampling_ratio/min': 0.0698758214712143, 'sampling/importance_sampling_ratio/mean': 1.0377470254898071, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7948398143053055, 'clip_ratio/low_mean': 0.04713901598006487, 'clip_ratio/low_min': 0.04713901598006487, 'clip_ratio/high_mean': 0.04863256216049194, 'clip_ratio/high_max': 0.04863256216049194, 'clip_ratio/region_mean': 0.09577157814055681, 'reward_total_mean': 0.09329602867364883, 'reward_meter_mean': 0.31860774755477905, 'reward_meter_std': 0.3715546727180481, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8237892389297485, 'reward_lexical_plausibility_std': 0.24426932632923126, 'reward_repeat_penalty_mean': 0.9666010141372681, 'reward_repeat_penalty_std': 0.08760970085859299, 'reward_repeat_floor_mean': 0.9978353381156921, 'reward_repeat_floor_std': 0.005218965467065573, 'reward_near_duplicate_penalty_mean': 0.9983958005905151, 'reward_near_duplicate_penalty_std': 0.004009816329926252, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9994552135467529, 'reward_distinct_2_std': 0.0015408677281811833, 'reward_judge_quality_mean': 0.2587500214576721, 'reward_judge_quality_std': 0.2995920181274414, 'reward_total_composite_mean': 0.09329602867364883, 'reward_total_composite_std': 0.12459640204906464, 'epoch': 0.03}

 17%|█▋        | 572/3300 [2:05:16<9:27:52, 12.49s/it]INFO 04-12 18:03:43 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 573/3300 [2:05:26<8:51:13, 11.69s/it]2026-04-12 18:03:52,741 | INFO | train_grpo_train | metrics_logged mode=train step=573

                                                      
{'loss': 0.0385, 'grad_norm': 13.215439796447754, 'learning_rate': 8.266666666666667e-06, 'num_tokens': 1060255.0, 'completions/mean_length': 52.375, 'completions/min_length': 43.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.31971821188926697, 'rewards/meter/std': 0.26265063881874084, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6087499856948853, 'rewards/judge_quality/std': 0.24930977821350098, 'rewards/repeat_penalty/mean': 0.9568012952804565, 'rewards/repeat_penalty/std': 0.043867096304893494, 'rewards/repeat_floor/mean': 0.9946140050888062, 'rewards/repeat_floor/std': 0.00524016423150897, 'rewards/near_duplicate_penalty/mean': 0.9762766361236572, 'rewards/near_duplicate_penalty/std': 0.02132708951830864, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9796939492225647, 'rewards/distinct_2/std': 0.02823641709983349, 'rewards/total_composite/mean': 0.2179451435804367, 'rewards/total_composite/std': 0.24642625451087952, 'reward': 0.2179451435804367, 'reward_std': 0.24642623960971832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1934223175048828, 'sampling/sampling_logp_difference/max': 2.0406086444854736, 'sampling/importance_sampling_ratio/min': 0.12994959950447083, 'sampling/importance_sampling_ratio/mean': 0.9974057078361511, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7752466052770615, 'clip_ratio/low_mean': 0.09983043652027845, 'clip_ratio/low_min': 0.09983043652027845, 'clip_ratio/high_mean': 0.06216245796531439, 'clip_ratio/high_max': 0.06216245796531439, 'clip_ratio/region_mean': 0.16199289448559284, 'reward_total_mean': 0.2179451435804367, 'reward_meter_mean': 0.31971821188926697, 'reward_meter_std': 0.26265063881874084, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9568012952804565, 'reward_repeat_penalty_std': 0.043867096304893494, 'reward_repeat_floor_mean': 0.9946140050888062, 'reward_repeat_floor_std': 0.00524016423150897, 'reward_near_duplicate_penalty_mean': 0.9762766361236572, 'reward_near_duplicate_penalty_std': 0.02132708951830864, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9796939492225647, 'reward_distinct_2_std': 0.02823641709983349, 'reward_judge_quality_mean': 0.6087499856948853, 'reward_judge_quality_std': 0.24930977821350098, 'reward_total_composite_mean': 0.2179451435804367, 'reward_total_composite_std': 0.24642625451087952, 'epoch': 0.03}

 17%|█▋        | 573/3300 [2:05:26<8:51:13, 11.69s/it]INFO 04-12 18:03:52 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 574/3300 [2:05:40<9:16:01, 12.24s/it]2026-04-12 18:04:06,263 | INFO | train_grpo_train | metrics_logged mode=train step=574

                                                      
{'loss': -0.0429, 'grad_norm': 5.088376522064209, 'learning_rate': 8.263636363636366e-06, 'num_tokens': 1061807.0, 'completions/mean_length': 97.0, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.71428680419922, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8838696479797363, 'rewards/meter/std': 0.15098340809345245, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9406701326370239, 'rewards/lexical_plausibility/std': 0.13003993034362793, 'rewards/judge_quality/mean': 0.4449999928474426, 'rewards/judge_quality/std': 0.3022770583629608, 'rewards/repeat_penalty/mean': 0.9857231378555298, 'rewards/repeat_penalty/std': 0.01682031713426113, 'rewards/repeat_floor/mean': 0.9978585243225098, 'rewards/repeat_floor/std': 0.0025230494793504477, 'rewards/near_duplicate_penalty/mean': 0.9857231378555298, 'rewards/near_duplicate_penalty/std': 0.01682031713426113, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.41043856739997864, 'rewards/total_composite/std': 0.30440497398376465, 'reward': 0.41043856739997864, 'reward_std': 0.30440497398376465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18837955594062805, 'sampling/sampling_logp_difference/max': 1.6316981315612793, 'sampling/importance_sampling_ratio/min': 0.19559712707996368, 'sampling/importance_sampling_ratio/mean': 1.0205085277557373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1722734868526459, 'clip_ratio/low_mean': 0.07605504244565964, 'clip_ratio/low_min': 0.07605504244565964, 'clip_ratio/high_mean': 0.05871355161070824, 'clip_ratio/high_max': 0.05871355161070824, 'clip_ratio/region_mean': 0.13476859405636787, 'reward_total_mean': 0.41043856739997864, 'reward_meter_mean': 0.8838696479797363, 'reward_meter_std': 0.15098340809345245, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9406701326370239, 'reward_lexical_plausibility_std': 0.13003993034362793, 'reward_repeat_penalty_mean': 0.9857231378555298, 'reward_repeat_penalty_std': 0.01682031713426113, 'reward_repeat_floor_mean': 0.9978585243225098, 'reward_repeat_floor_std': 0.0025230494793504477, 'reward_near_duplicate_penalty_mean': 0.9857231378555298, 'reward_near_duplicate_penalty_std': 0.01682031713426113, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4449999928474426, 'reward_judge_quality_std': 0.3022770583629608, 'reward_total_composite_mean': 0.41043856739997864, 'reward_total_composite_std': 0.30440497398376465, 'epoch': 0.03}

 17%|█▋        | 574/3300 [2:05:40<9:16:01, 12.24s/it]INFO 04-12 18:04:06 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 575/3300 [2:05:48<8:23:00, 11.08s/it]2026-04-12 18:04:14,626 | INFO | train_grpo_train | metrics_logged mode=train step=575

                                                      
{'loss': -0.0124, 'grad_norm': 18.976665496826172, 'learning_rate': 8.260606060606061e-06, 'num_tokens': 1063184.0, 'completions/mean_length': 22.125, 'completions/min_length': 20.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.125, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9922141432762146, 'rewards/meter/std': 0.0037023602053523064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.34201088547706604, 'rewards/repeat_penalty/mean': 0.737289309501648, 'rewards/repeat_penalty/std': 0.034418314695358276, 'rewards/repeat_floor/mean': 0.9824578762054443, 'rewards/repeat_floor/std': 0.006883666850626469, 'rewards/near_duplicate_penalty/mean': 0.9830524325370789, 'rewards/near_duplicate_penalty/std': 0.045891109853982925, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5936582088470459, 'rewards/total_composite/std': 0.33149585127830505, 'reward': 0.5936582088470459, 'reward_std': 0.33149585127830505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19855232536792755, 'sampling/sampling_logp_difference/max': 1.9538545608520508, 'sampling/importance_sampling_ratio/min': 0.14172671735286713, 'sampling/importance_sampling_ratio/mean': 1.0286023616790771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5153897851705551, 'clip_ratio/low_mean': 0.09020563028752804, 'clip_ratio/low_min': 0.09020563028752804, 'clip_ratio/high_mean': 0.1141005763784051, 'clip_ratio/high_max': 0.1141005763784051, 'clip_ratio/region_mean': 0.20430620666593313, 'reward_total_mean': 0.5936582088470459, 'reward_meter_mean': 0.9922141432762146, 'reward_meter_std': 0.0037023602053523064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.737289309501648, 'reward_repeat_penalty_std': 0.034418314695358276, 'reward_repeat_floor_mean': 0.9824578762054443, 'reward_repeat_floor_std': 0.006883666850626469, 'reward_near_duplicate_penalty_mean': 0.9830524325370789, 'reward_near_duplicate_penalty_std': 0.045891109853982925, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.34201088547706604, 'reward_total_composite_mean': 0.5936582088470459, 'reward_total_composite_std': 0.33149585127830505, 'epoch': 0.03}

 17%|█▋        | 575/3300 [2:05:48<8:23:00, 11.08s/it]INFO 04-12 18:04:14 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 576/3300 [2:05:57<7:51:57, 10.40s/it]2026-04-12 18:04:23,436 | INFO | train_grpo_train | metrics_logged mode=train step=576

                                                      
{'loss': 0.0084, 'grad_norm': 14.198740005493164, 'learning_rate': 8.257575757575758e-06, 'num_tokens': 1065070.0, 'completions/mean_length': 61.75, 'completions/min_length': 45.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.796534538269043, 'rewards/meter/std': 0.22858215868473053, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.9449714422225952, 'rewards/repeat_penalty/std': 0.051366522908210754, 'rewards/repeat_floor/mean': 0.9934579133987427, 'rewards/repeat_floor/std': 0.005356706213206053, 'rewards/near_duplicate_penalty/mean': 0.9747199416160583, 'rewards/near_duplicate_penalty/std': 0.020100874826312065, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9694436192512512, 'rewards/distinct_2/std': 0.04780247062444687, 'rewards/total_composite/mean': 0.35789692401885986, 'rewards/total_composite/std': 0.17815133929252625, 'reward': 0.35789692401885986, 'reward_std': 0.17815133929252625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1710430234670639, 'sampling/sampling_logp_difference/max': 1.8441460132598877, 'sampling/importance_sampling_ratio/min': 0.15816032886505127, 'sampling/importance_sampling_ratio/mean': 1.0230984687805176, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1213613077998161, 'clip_ratio/low_mean': 0.07982475776225328, 'clip_ratio/low_min': 0.07982475776225328, 'clip_ratio/high_mean': 0.0809027049690485, 'clip_ratio/high_max': 0.0809027049690485, 'clip_ratio/region_mean': 0.16072746273130178, 'reward_total_mean': 0.35789692401885986, 'reward_meter_mean': 0.796534538269043, 'reward_meter_std': 0.22858215868473053, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9449714422225952, 'reward_repeat_penalty_std': 0.051366522908210754, 'reward_repeat_floor_mean': 0.9934579133987427, 'reward_repeat_floor_std': 0.005356706213206053, 'reward_near_duplicate_penalty_mean': 0.9747199416160583, 'reward_near_duplicate_penalty_std': 0.020100874826312065, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9694436192512512, 'reward_distinct_2_std': 0.04780247062444687, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.35789692401885986, 'reward_total_composite_std': 0.17815133929252625, 'epoch': 0.03}

 17%|█▋        | 576/3300 [2:05:57<7:51:57, 10.40s/it]INFO 04-12 18:04:23 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 577/3300 [2:06:04<7:10:45,  9.49s/it]2026-04-12 18:04:30,814 | INFO | train_grpo_train | metrics_logged mode=train step=577

                                                      
{'loss': 0.0412, 'grad_norm': 24.08905792236328, 'learning_rate': 8.254545454545456e-06, 'num_tokens': 1066505.0, 'completions/mean_length': 23.375, 'completions/min_length': 19.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.7529468536376953, 'rewards/meter/std': 0.33371686935424805, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8385416269302368, 'rewards/lexical_plausibility/std': 0.1855888068675995, 'rewards/judge_quality/mean': 0.6487500071525574, 'rewards/judge_quality/std': 0.413743793964386, 'rewards/repeat_penalty/mean': 0.7179175019264221, 'rewards/repeat_penalty/std': 0.0642748475074768, 'rewards/repeat_floor/mean': 0.9794076681137085, 'rewards/repeat_floor/std': 0.012317469343543053, 'rewards/near_duplicate_penalty/mean': 0.9709596037864685, 'rewards/near_duplicate_penalty/std': 0.08213866502046585, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.43734052777290344, 'rewards/total_composite/std': 0.3831902742385864, 'reward': 0.43734052777290344, 'reward_std': 0.3831902742385864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18308025598526, 'sampling/sampling_logp_difference/max': 0.9893598556518555, 'sampling/importance_sampling_ratio/min': 0.37181463837623596, 'sampling/importance_sampling_ratio/mean': 1.0063985586166382, 'sampling/importance_sampling_ratio/max': 1.9688750505447388, 'entropy': 1.371234193444252, 'clip_ratio/low_mean': 0.08781056059524417, 'clip_ratio/low_min': 0.08781056059524417, 'clip_ratio/high_mean': 0.07091317139565945, 'clip_ratio/high_max': 0.07091317139565945, 'clip_ratio/region_mean': 0.15872373199090362, 'reward_total_mean': 0.43734052777290344, 'reward_meter_mean': 0.7529468536376953, 'reward_meter_std': 0.33371686935424805, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8385416269302368, 'reward_lexical_plausibility_std': 0.1855888068675995, 'reward_repeat_penalty_mean': 0.7179175019264221, 'reward_repeat_penalty_std': 0.0642748475074768, 'reward_repeat_floor_mean': 0.9794076681137085, 'reward_repeat_floor_std': 0.012317469343543053, 'reward_near_duplicate_penalty_mean': 0.9709596037864685, 'reward_near_duplicate_penalty_std': 0.08213866502046585, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.6487500071525574, 'reward_judge_quality_std': 0.413743793964386, 'reward_total_composite_mean': 0.43734052777290344, 'reward_total_composite_std': 0.3831902742385864, 'epoch': 0.03}

 17%|█▋        | 577/3300 [2:06:04<7:10:45,  9.49s/it]INFO 04-12 18:04:31 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 578/3300 [2:06:14<7:20:51,  9.72s/it]2026-04-12 18:04:41,061 | INFO | train_grpo_train | metrics_logged mode=train step=578

                                                      
{'loss': 0.0593, 'grad_norm': 10.385157585144043, 'learning_rate': 8.251515151515153e-06, 'num_tokens': 1068572.0, 'completions/mean_length': 90.375, 'completions/min_length': 72.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9587092399597168, 'rewards/meter/std': 0.08925272524356842, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.8170046210289001, 'rewards/repeat_penalty/std': 0.1990501433610916, 'rewards/repeat_floor/mean': 0.9804220795631409, 'rewards/repeat_floor/std': 0.019898395985364914, 'rewards/near_duplicate_penalty/mean': 0.9495874643325806, 'rewards/near_duplicate_penalty/std': 0.0440165251493454, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8664884567260742, 'rewards/distinct_2/std': 0.15123692154884338, 'rewards/total_composite/mean': 0.43401795625686646, 'rewards/total_composite/std': 0.1833207905292511, 'reward': 0.43401795625686646, 'reward_std': 0.1833207607269287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15359997749328613, 'sampling/sampling_logp_difference/max': 2.1504414081573486, 'sampling/importance_sampling_ratio/min': 0.11643274873495102, 'sampling/importance_sampling_ratio/mean': 1.0121476650238037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8920102417469025, 'clip_ratio/low_mean': 0.04906655754894018, 'clip_ratio/low_min': 0.04906655754894018, 'clip_ratio/high_mean': 0.11039167083799839, 'clip_ratio/high_max': 0.11039167083799839, 'clip_ratio/region_mean': 0.15945822838693857, 'reward_total_mean': 0.43401795625686646, 'reward_meter_mean': 0.9587092399597168, 'reward_meter_std': 0.08925272524356842, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8170046210289001, 'reward_repeat_penalty_std': 0.1990501433610916, 'reward_repeat_floor_mean': 0.9804220795631409, 'reward_repeat_floor_std': 0.019898395985364914, 'reward_near_duplicate_penalty_mean': 0.9495874643325806, 'reward_near_duplicate_penalty_std': 0.0440165251493454, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8664884567260742, 'reward_distinct_2_std': 0.15123692154884338, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.43401795625686646, 'reward_total_composite_std': 0.1833207905292511, 'epoch': 0.03}

 18%|█▊        | 578/3300 [2:06:14<7:20:51,  9.72s/it]INFO 04-12 18:04:41 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 579/3300 [2:06:27<8:06:25, 10.73s/it]2026-04-12 18:04:54,141 | INFO | train_grpo_train | metrics_logged mode=train step=579

                                                      
{'loss': -0.0368, 'grad_norm': 7.788293361663818, 'learning_rate': 8.248484848484848e-06, 'num_tokens': 1070141.0, 'completions/mean_length': 92.125, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.142860412597656, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.40672704577445984, 'rewards/meter/std': 0.372258722782135, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9135165810585022, 'rewards/lexical_plausibility/std': 0.21889373660087585, 'rewards/judge_quality/mean': 0.6637499928474426, 'rewards/judge_quality/std': 0.33957695960998535, 'rewards/repeat_penalty/mean': 0.971304178237915, 'rewards/repeat_penalty/std': 0.05075687915086746, 'rewards/repeat_floor/mean': 0.9964439868927002, 'rewards/repeat_floor/std': 0.0057627735659480095, 'rewards/near_duplicate_penalty/mean': 0.9845349788665771, 'rewards/near_duplicate_penalty/std': 0.023037346079945564, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.2326650321483612, 'rewards/total_composite/std': 0.2507937252521515, 'reward': 0.2326650321483612, 'reward_std': 0.2507937550544739, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19097664952278137, 'sampling/sampling_logp_difference/max': 1.4531831741333008, 'sampling/importance_sampling_ratio/min': 0.23382480442523956, 'sampling/importance_sampling_ratio/mean': 1.0127637386322021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7697450146079063, 'clip_ratio/low_mean': 0.0793445548042655, 'clip_ratio/low_min': 0.0793445548042655, 'clip_ratio/high_mean': 0.08510783687233925, 'clip_ratio/high_max': 0.08510783687233925, 'clip_ratio/region_mean': 0.16445239167660475, 'reward_total_mean': 0.2326650321483612, 'reward_meter_mean': 0.40672704577445984, 'reward_meter_std': 0.372258722782135, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9135165810585022, 'reward_lexical_plausibility_std': 0.21889373660087585, 'reward_repeat_penalty_mean': 0.971304178237915, 'reward_repeat_penalty_std': 0.05075687915086746, 'reward_repeat_floor_mean': 0.9964439868927002, 'reward_repeat_floor_std': 0.0057627735659480095, 'reward_near_duplicate_penalty_mean': 0.9845349788665771, 'reward_near_duplicate_penalty_std': 0.023037346079945564, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.6637499928474426, 'reward_judge_quality_std': 0.33957695960998535, 'reward_total_composite_mean': 0.2326650321483612, 'reward_total_composite_std': 0.2507937252521515, 'epoch': 0.03}

 18%|█▊        | 579/3300 [2:06:27<8:06:25, 10.73s/it]INFO 04-12 18:04:54 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 580/3300 [2:06:42<9:03:01, 11.98s/it]2026-04-12 18:05:09,049 | INFO | train_grpo_train | metrics_logged mode=train step=580

                                                      
{'loss': -0.0992, 'grad_norm': 7.1847405433654785, 'learning_rate': 8.245454545454546e-06, 'num_tokens': 1072011.0, 'completions/mean_length': 120.75, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.85714721679688, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.705537736415863, 'rewards/meter/std': 0.2794915735721588, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.07954952120780945, 'rewards/lexical_plausibility/mean': 0.9416865110397339, 'rewards/lexical_plausibility/std': 0.16493549942970276, 'rewards/judge_quality/mean': 0.3187499940395355, 'rewards/judge_quality/std': 0.16243129968643188, 'rewards/repeat_penalty/mean': 0.572622537612915, 'rewards/repeat_penalty/std': 0.17941150069236755, 'rewards/repeat_floor/mean': 0.9546619653701782, 'rewards/repeat_floor/std': 0.02013588510453701, 'rewards/near_duplicate_penalty/mean': 0.8809958696365356, 'rewards/near_duplicate_penalty/std': 0.05666163191199303, 'rewards/opening_diversity/mean': 0.9296875, 'rewards/opening_diversity/std': 0.1129826009273529, 'rewards/distinct_2/mean': 0.7414588928222656, 'rewards/distinct_2/std': 0.15535785257816315, 'rewards/total_composite/mean': 0.20815955102443695, 'rewards/total_composite/std': 0.13052405416965485, 'reward': 0.20815955102443695, 'reward_std': 0.13052405416965485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15861372649669647, 'sampling/sampling_logp_difference/max': 1.8488059043884277, 'sampling/importance_sampling_ratio/min': 0.193994402885437, 'sampling/importance_sampling_ratio/mean': 1.0258125066757202, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7317588552832603, 'clip_ratio/low_mean': 0.03219697065651417, 'clip_ratio/low_min': 0.03219697065651417, 'clip_ratio/high_mean': 0.07353195920586586, 'clip_ratio/high_max': 0.07353195920586586, 'clip_ratio/region_mean': 0.10572892986238003, 'reward_total_mean': 0.20815955102443695, 'reward_meter_mean': 0.705537736415863, 'reward_meter_std': 0.2794915735721588, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.07954952120780945, 'reward_lexical_plausibility_mean': 0.9416865110397339, 'reward_lexical_plausibility_std': 0.16493549942970276, 'reward_repeat_penalty_mean': 0.572622537612915, 'reward_repeat_penalty_std': 0.17941150069236755, 'reward_repeat_floor_mean': 0.9546619653701782, 'reward_repeat_floor_std': 0.02013588510453701, 'reward_near_duplicate_penalty_mean': 0.8809958696365356, 'reward_near_duplicate_penalty_std': 0.05666163191199303, 'reward_opening_diversity_mean': 0.9296875, 'reward_opening_diversity_std': 0.1129826009273529, 'reward_distinct_2_mean': 0.7414588928222656, 'reward_distinct_2_std': 0.15535785257816315, 'reward_judge_quality_mean': 0.3187499940395355, 'reward_judge_quality_std': 0.16243129968643188, 'reward_total_composite_mean': 0.20815955102443695, 'reward_total_composite_std': 0.13052405416965485, 'epoch': 0.03}

 18%|█▊        | 580/3300 [2:06:42<9:03:01, 11.98s/it]INFO 04-12 18:05:09 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 581/3300 [2:06:53<8:50:28, 11.71s/it]2026-04-12 18:05:20,113 | INFO | train_grpo_train | metrics_logged mode=train step=581

                                                      
{'loss': -0.0358, 'grad_norm': 14.940107345581055, 'learning_rate': 8.242424242424243e-06, 'num_tokens': 1073881.0, 'completions/mean_length': 53.75, 'completions/min_length': 47.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.3694577217102051, 'rewards/meter/std': 0.301023930311203, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.28665250539779663, 'rewards/repeat_penalty/mean': 0.9378868341445923, 'rewards/repeat_penalty/std': 0.08922033756971359, 'rewards/repeat_floor/mean': 0.9928351640701294, 'rewards/repeat_floor/std': 0.009358384646475315, 'rewards/near_duplicate_penalty/mean': 0.9766912460327148, 'rewards/near_duplicate_penalty/std': 0.019322985783219337, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.95923912525177, 'rewards/distinct_2/std': 0.07707840204238892, 'rewards/total_composite/mean': 0.2371755689382553, 'rewards/total_composite/std': 0.2880220413208008, 'reward': 0.2371755689382553, 'reward_std': 0.2880220413208008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1628991961479187, 'sampling/sampling_logp_difference/max': 1.9592454433441162, 'sampling/importance_sampling_ratio/min': 0.14096474647521973, 'sampling/importance_sampling_ratio/mean': 1.0173221826553345, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8562130555510521, 'clip_ratio/low_mean': 0.12125014699995518, 'clip_ratio/low_min': 0.12125014699995518, 'clip_ratio/high_mean': 0.04394703730940819, 'clip_ratio/high_max': 0.04394703730940819, 'clip_ratio/region_mean': 0.16519718430936337, 'reward_total_mean': 0.2371755689382553, 'reward_meter_mean': 0.3694577217102051, 'reward_meter_std': 0.301023930311203, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9378868341445923, 'reward_repeat_penalty_std': 0.08922033756971359, 'reward_repeat_floor_mean': 0.9928351640701294, 'reward_repeat_floor_std': 0.009358384646475315, 'reward_near_duplicate_penalty_mean': 0.9766912460327148, 'reward_near_duplicate_penalty_std': 0.019322985783219337, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.95923912525177, 'reward_distinct_2_std': 0.07707840204238892, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.28665250539779663, 'reward_total_composite_mean': 0.2371755689382553, 'reward_total_composite_std': 0.2880220413208008, 'epoch': 0.03}

 18%|█▊        | 581/3300 [2:06:53<8:50:28, 11.71s/it]INFO 04-12 18:05:20 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 582/3300 [2:07:07<9:09:25, 12.13s/it]2026-04-12 18:05:33,234 | INFO | train_grpo_train | metrics_logged mode=train step=582

                                                      
{'loss': -0.0443, 'grad_norm': 5.245945453643799, 'learning_rate': 8.23939393939394e-06, 'num_tokens': 1075395.0, 'completions/mean_length': 99.25, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.28571701049805, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8541418313980103, 'rewards/meter/std': 0.28981056809425354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9368588924407959, 'rewards/lexical_plausibility/std': 0.17858998477458954, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.3003212809562683, 'rewards/repeat_penalty/mean': 0.945415735244751, 'rewards/repeat_penalty/std': 0.13051041960716248, 'rewards/repeat_floor/mean': 0.9943419098854065, 'rewards/repeat_floor/std': 0.012509886175394058, 'rewards/near_duplicate_penalty/mean': 0.9811896085739136, 'rewards/near_duplicate_penalty/std': 0.03266141563653946, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.44984495639801025, 'rewards/total_composite/std': 0.2928798794746399, 'reward': 0.44984495639801025, 'reward_std': 0.2928798794746399, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16057546436786652, 'sampling/sampling_logp_difference/max': 1.6267890930175781, 'sampling/importance_sampling_ratio/min': 0.19655968248844147, 'sampling/importance_sampling_ratio/mean': 1.0162990093231201, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0817658826708794, 'clip_ratio/low_mean': 0.06546593271195889, 'clip_ratio/low_min': 0.06546593271195889, 'clip_ratio/high_mean': 0.05600619316101074, 'clip_ratio/high_max': 0.05600619316101074, 'clip_ratio/region_mean': 0.12147212587296963, 'reward_total_mean': 0.44984495639801025, 'reward_meter_mean': 0.8541418313980103, 'reward_meter_std': 0.28981056809425354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9368588924407959, 'reward_lexical_plausibility_std': 0.17858998477458954, 'reward_repeat_penalty_mean': 0.945415735244751, 'reward_repeat_penalty_std': 0.13051041960716248, 'reward_repeat_floor_mean': 0.9943419098854065, 'reward_repeat_floor_std': 0.012509886175394058, 'reward_near_duplicate_penalty_mean': 0.9811896085739136, 'reward_near_duplicate_penalty_std': 0.03266141563653946, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.3003212809562683, 'reward_total_composite_mean': 0.44984495639801025, 'reward_total_composite_std': 0.2928798794746399, 'epoch': 0.03}

 18%|█▊        | 582/3300 [2:07:07<9:09:25, 12.13s/it]INFO 04-12 18:05:33 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 583/3300 [2:07:15<8:17:30, 10.99s/it]2026-04-12 18:05:41,549 | INFO | train_grpo_train | metrics_logged mode=train step=583

                                                      
{'loss': 0.0718, 'grad_norm': 18.93609046936035, 'learning_rate': 8.236363636363637e-06, 'num_tokens': 1076997.0, 'completions/mean_length': 38.25, 'completions/min_length': 33.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.6064172983169556, 'rewards/meter/std': 0.46976912021636963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.28222334384918213, 'rewards/repeat_penalty/mean': 0.9098682403564453, 'rewards/repeat_penalty/std': 0.09603958576917648, 'rewards/repeat_floor/mean': 0.9876708388328552, 'rewards/repeat_floor/std': 0.01218317449092865, 'rewards/near_duplicate_penalty/mean': 0.93613600730896, 'rewards/near_duplicate_penalty/std': 0.04395810514688492, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9694494009017944, 'rewards/distinct_2/std': 0.06382904946804047, 'rewards/total_composite/mean': 0.41691726446151733, 'rewards/total_composite/std': 0.4035947918891907, 'reward': 0.41691726446151733, 'reward_std': 0.4035947918891907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19777661561965942, 'sampling/sampling_logp_difference/max': 1.397841453552246, 'sampling/importance_sampling_ratio/min': 0.24712984263896942, 'sampling/importance_sampling_ratio/mean': 1.0057704448699951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.934145137667656, 'clip_ratio/low_mean': 0.07995073054917157, 'clip_ratio/low_min': 0.07995073054917157, 'clip_ratio/high_mean': 0.057914674282073975, 'clip_ratio/high_max': 0.057914674282073975, 'clip_ratio/region_mean': 0.13786540483124554, 'reward_total_mean': 0.41691726446151733, 'reward_meter_mean': 0.6064172983169556, 'reward_meter_std': 0.46976912021636963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9098682403564453, 'reward_repeat_penalty_std': 0.09603958576917648, 'reward_repeat_floor_mean': 0.9876708388328552, 'reward_repeat_floor_std': 0.01218317449092865, 'reward_near_duplicate_penalty_mean': 0.93613600730896, 'reward_near_duplicate_penalty_std': 0.04395810514688492, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9694494009017944, 'reward_distinct_2_std': 0.06382904946804047, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.28222334384918213, 'reward_total_composite_mean': 0.41691726446151733, 'reward_total_composite_std': 0.4035947918891907, 'epoch': 0.03}

 18%|█▊        | 583/3300 [2:07:15<8:17:30, 10.99s/it]INFO 04-12 18:05:41 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 584/3300 [2:07:29<9:00:21, 11.94s/it]2026-04-12 18:05:55,704 | INFO | train_grpo_train | metrics_logged mode=train step=584

                                                      
{'loss': -0.0751, 'grad_norm': 2.6814279556274414, 'learning_rate': 8.233333333333335e-06, 'num_tokens': 1078505.0, 'completions/mean_length': 161.5, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.66666793823242, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9531233906745911, 'rewards/meter/std': 0.07817742228507996, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8637686967849731, 'rewards/lexical_plausibility/std': 0.2527143359184265, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3366537094116211, 'rewards/repeat_penalty/mean': 0.9679023027420044, 'rewards/repeat_penalty/std': 0.05428579822182655, 'rewards/repeat_floor/mean': 0.9959695339202881, 'rewards/repeat_floor/std': 0.006082112900912762, 'rewards/near_duplicate_penalty/mean': 0.9820059537887573, 'rewards/near_duplicate_penalty/std': 0.020822148770093918, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.38559335470199585, 'rewards/total_composite/std': 0.29724568128585815, 'reward': 0.38559335470199585, 'reward_std': 0.29724568128585815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15141427516937256, 'sampling/sampling_logp_difference/max': 1.228102684020996, 'sampling/importance_sampling_ratio/min': 0.2928476631641388, 'sampling/importance_sampling_ratio/mean': 1.0269032716751099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.746183343231678, 'clip_ratio/low_mean': 0.08390332479029894, 'clip_ratio/low_min': 0.08390332479029894, 'clip_ratio/high_mean': 0.042247386649250984, 'clip_ratio/high_max': 0.042247386649250984, 'clip_ratio/region_mean': 0.12615071143954992, 'reward_total_mean': 0.38559335470199585, 'reward_meter_mean': 0.9531233906745911, 'reward_meter_std': 0.07817742228507996, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8637686967849731, 'reward_lexical_plausibility_std': 0.2527143359184265, 'reward_repeat_penalty_mean': 0.9679023027420044, 'reward_repeat_penalty_std': 0.05428579822182655, 'reward_repeat_floor_mean': 0.9959695339202881, 'reward_repeat_floor_std': 0.006082112900912762, 'reward_near_duplicate_penalty_mean': 0.9820059537887573, 'reward_near_duplicate_penalty_std': 0.020822148770093918, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3366537094116211, 'reward_total_composite_mean': 0.38559335470199585, 'reward_total_composite_std': 0.29724568128585815, 'epoch': 0.03}

 18%|█▊        | 584/3300 [2:07:29<9:00:21, 11.94s/it]INFO 04-12 18:05:55 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 585/3300 [2:07:38<8:20:29, 11.06s/it]2026-04-12 18:06:04,721 | INFO | train_grpo_train | metrics_logged mode=train step=585

                                                      
{'loss': 0.0964, 'grad_norm': 17.864479064941406, 'learning_rate': 8.23030303030303e-06, 'num_tokens': 1080205.0, 'completions/mean_length': 47.5, 'completions/min_length': 41.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6405714750289917, 'rewards/meter/std': 0.38932862877845764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9398886561393738, 'rewards/repeat_penalty/std': 0.11351761221885681, 'rewards/repeat_floor/mean': 0.99423748254776, 'rewards/repeat_floor/std': 0.00942007452249527, 'rewards/near_duplicate_penalty/mean': 0.9843447804450989, 'rewards/near_duplicate_penalty/std': 0.017178017646074295, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9828975200653076, 'rewards/distinct_2/std': 0.031678035855293274, 'rewards/total_composite/mean': 0.2329883575439453, 'rewards/total_composite/std': 0.15330007672309875, 'reward': 0.2329883575439453, 'reward_std': 0.15330006182193756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1745918095111847, 'sampling/sampling_logp_difference/max': 1.549882411956787, 'sampling/importance_sampling_ratio/min': 0.21227294206619263, 'sampling/importance_sampling_ratio/mean': 1.029954433441162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1156706884503365, 'clip_ratio/low_mean': 0.0724463053047657, 'clip_ratio/low_min': 0.0724463053047657, 'clip_ratio/high_mean': 0.08492917101830244, 'clip_ratio/high_max': 0.08492917101830244, 'clip_ratio/region_mean': 0.15737547632306814, 'reward_total_mean': 0.2329883575439453, 'reward_meter_mean': 0.6405714750289917, 'reward_meter_std': 0.38932862877845764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9398886561393738, 'reward_repeat_penalty_std': 0.11351761221885681, 'reward_repeat_floor_mean': 0.99423748254776, 'reward_repeat_floor_std': 0.00942007452249527, 'reward_near_duplicate_penalty_mean': 0.9843447804450989, 'reward_near_duplicate_penalty_std': 0.017178017646074295, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9828975200653076, 'reward_distinct_2_std': 0.031678035855293274, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.2329883575439453, 'reward_total_composite_std': 0.15330007672309875, 'epoch': 0.03}

 18%|█▊        | 585/3300 [2:07:38<8:20:29, 11.06s/it]INFO 04-12 18:06:04 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 586/3300 [2:07:52<8:54:04, 11.81s/it]2026-04-12 18:06:18,272 | INFO | train_grpo_train | metrics_logged mode=train step=586

                                                      
{'loss': -0.1275, 'grad_norm': 6.149367332458496, 'learning_rate': 8.227272727272728e-06, 'num_tokens': 1081982.0, 'completions/mean_length': 115.125, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.42857360839844, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.770018994808197, 'rewards/meter/std': 0.3668903112411499, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.07440237700939178, 'rewards/lexical_plausibility/mean': 0.9576707482337952, 'rewards/lexical_plausibility/std': 0.11972522735595703, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.16335216164588928, 'rewards/repeat_penalty/mean': 0.8881149291992188, 'rewards/repeat_penalty/std': 0.11621665954589844, 'rewards/repeat_floor/mean': 0.9900487661361694, 'rewards/repeat_floor/std': 0.007919838652014732, 'rewards/near_duplicate_penalty/mean': 0.9771410226821899, 'rewards/near_duplicate_penalty/std': 0.013170610181987286, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9691950678825378, 'rewards/distinct_2/std': 0.03478526324033737, 'rewards/total_composite/mean': 0.29879647493362427, 'rewards/total_composite/std': 0.17232559621334076, 'reward': 0.29879647493362427, 'reward_std': 0.17232558131217957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1661117970943451, 'sampling/sampling_logp_difference/max': 1.924523115158081, 'sampling/importance_sampling_ratio/min': 0.14594534039497375, 'sampling/importance_sampling_ratio/mean': 0.9916799664497375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9253994524478912, 'clip_ratio/low_mean': 0.017500000074505806, 'clip_ratio/low_min': 0.017500000074505806, 'clip_ratio/high_mean': 0.12571268435567617, 'clip_ratio/high_max': 0.12571268435567617, 'clip_ratio/region_mean': 0.14321268443018198, 'reward_total_mean': 0.29879647493362427, 'reward_meter_mean': 0.770018994808197, 'reward_meter_std': 0.3668903112411499, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.07440237700939178, 'reward_lexical_plausibility_mean': 0.9576707482337952, 'reward_lexical_plausibility_std': 0.11972522735595703, 'reward_repeat_penalty_mean': 0.8881149291992188, 'reward_repeat_penalty_std': 0.11621665954589844, 'reward_repeat_floor_mean': 0.9900487661361694, 'reward_repeat_floor_std': 0.007919838652014732, 'reward_near_duplicate_penalty_mean': 0.9771410226821899, 'reward_near_duplicate_penalty_std': 0.013170610181987286, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9691950678825378, 'reward_distinct_2_std': 0.03478526324033737, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.16335216164588928, 'reward_total_composite_mean': 0.29879647493362427, 'reward_total_composite_std': 0.17232559621334076, 'epoch': 0.03}

 18%|█▊        | 586/3300 [2:07:52<8:54:04, 11.81s/it]INFO 04-12 18:06:18 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 587/3300 [2:08:05<9:14:44, 12.27s/it]2026-04-12 18:06:31,614 | INFO | train_grpo_train | metrics_logged mode=train step=587

                                                      
{'loss': -0.0525, 'grad_norm': 6.476796627044678, 'learning_rate': 8.224242424242425e-06, 'num_tokens': 1083493.0, 'completions/mean_length': 101.875, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.28571701049805, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.7257331013679504, 'rewards/meter/std': 0.4257688820362091, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9304554462432861, 'rewards/lexical_plausibility/std': 0.196701779961586, 'rewards/judge_quality/mean': 0.543749988079071, 'rewards/judge_quality/std': 0.29952046275138855, 'rewards/repeat_penalty/mean': 0.959242045879364, 'rewards/repeat_penalty/std': 0.0620308481156826, 'rewards/repeat_floor/mean': 0.9947984218597412, 'rewards/repeat_floor/std': 0.006942821200937033, 'rewards/near_duplicate_penalty/mean': 0.9758121967315674, 'rewards/near_duplicate_penalty/std': 0.024305464699864388, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.4604256749153137, 'rewards/total_composite/std': 0.3339039981365204, 'reward': 0.4604256749153137, 'reward_std': 0.3339039981365204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15783986449241638, 'sampling/sampling_logp_difference/max': 1.6187973022460938, 'sampling/importance_sampling_ratio/min': 0.19813686609268188, 'sampling/importance_sampling_ratio/mean': 1.0214251279830933, 'sampling/importance_sampling_ratio/max': 1.9676741361618042, 'entropy': 1.1844745725393295, 'clip_ratio/low_mean': 0.04924048949033022, 'clip_ratio/low_min': 0.04924048949033022, 'clip_ratio/high_mean': 0.06058441661298275, 'clip_ratio/high_max': 0.06058441661298275, 'clip_ratio/region_mean': 0.10982490610331297, 'reward_total_mean': 0.4604256749153137, 'reward_meter_mean': 0.7257331013679504, 'reward_meter_std': 0.4257688820362091, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9304554462432861, 'reward_lexical_plausibility_std': 0.196701779961586, 'reward_repeat_penalty_mean': 0.959242045879364, 'reward_repeat_penalty_std': 0.0620308481156826, 'reward_repeat_floor_mean': 0.9947984218597412, 'reward_repeat_floor_std': 0.006942821200937033, 'reward_near_duplicate_penalty_mean': 0.9758121967315674, 'reward_near_duplicate_penalty_std': 0.024305464699864388, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.543749988079071, 'reward_judge_quality_std': 0.29952046275138855, 'reward_total_composite_mean': 0.4604256749153137, 'reward_total_composite_std': 0.3339039981365204, 'epoch': 0.03}

 18%|█▊        | 587/3300 [2:08:05<9:14:44, 12.27s/it]INFO 04-12 18:06:31 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 588/3300 [2:08:16<8:58:52, 11.92s/it]2026-04-12 18:06:42,727 | INFO | train_grpo_train | metrics_logged mode=train step=588

                                                      
{'loss': 0.1025, 'grad_norm': 5.921337127685547, 'learning_rate': 8.221212121212122e-06, 'num_tokens': 1086972.0, 'completions/mean_length': 217.875, 'completions/min_length': 184.0, 'completions/max_length': 270.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 217.875, 'completions/min_terminated_length': 184.0, 'completions/max_terminated_length': 270.0, 'rewards/meter/mean': 0.9886859655380249, 'rewards/meter/std': 0.003234405303373933, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.017359109595417976, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.13024701178073883, 'rewards/repeat_penalty/mean': 0.28278589248657227, 'rewards/repeat_penalty/std': 0.3578672409057617, 'rewards/repeat_floor/mean': 0.8869429230690002, 'rewards/repeat_floor/std': 0.07762374728918076, 'rewards/near_duplicate_penalty/mean': 0.7476008534431458, 'rewards/near_duplicate_penalty/std': 0.20883676409721375, 'rewards/opening_diversity/mean': 0.7758224010467529, 'rewards/opening_diversity/std': 0.2800602912902832, 'rewards/distinct_2/mean': 0.31392723321914673, 'rewards/distinct_2/std': 0.37888339161872864, 'rewards/total_composite/mean': 0.24719148874282837, 'rewards/total_composite/std': 0.1288624256849289, 'reward': 0.24719148874282837, 'reward_std': 0.1288624256849289, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10224106162786484, 'sampling/sampling_logp_difference/max': 2.688180446624756, 'sampling/importance_sampling_ratio/min': 0.06800457090139389, 'sampling/importance_sampling_ratio/mean': 1.006943941116333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5232706964015961, 'clip_ratio/low_mean': 0.0262136145029217, 'clip_ratio/low_min': 0.0262136145029217, 'clip_ratio/high_mean': 0.06160101946443319, 'clip_ratio/high_max': 0.06160101946443319, 'clip_ratio/region_mean': 0.0878146339673549, 'reward_total_mean': 0.24719148874282837, 'reward_meter_mean': 0.9886859655380249, 'reward_meter_std': 0.003234405303373933, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.017359109595417976, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.28278589248657227, 'reward_repeat_penalty_std': 0.3578672409057617, 'reward_repeat_floor_mean': 0.8869429230690002, 'reward_repeat_floor_std': 0.07762374728918076, 'reward_near_duplicate_penalty_mean': 0.7476008534431458, 'reward_near_duplicate_penalty_std': 0.20883676409721375, 'reward_opening_diversity_mean': 0.7758224010467529, 'reward_opening_diversity_std': 0.2800602912902832, 'reward_distinct_2_mean': 0.31392723321914673, 'reward_distinct_2_std': 0.37888339161872864, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.13024701178073883, 'reward_total_composite_mean': 0.24719148874282837, 'reward_total_composite_std': 0.1288624256849289, 'epoch': 0.03}

 18%|█▊        | 588/3300 [2:08:16<8:58:52, 11.92s/it]INFO 04-12 18:06:42 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 589/3300 [2:08:24<8:11:28, 10.88s/it]2026-04-12 18:06:51,169 | INFO | train_grpo_train | metrics_logged mode=train step=589

                                                      
{'loss': 0.0827, 'grad_norm': 18.29681968688965, 'learning_rate': 8.21818181818182e-06, 'num_tokens': 1088556.0, 'completions/mean_length': 37.0, 'completions/min_length': 32.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5002398490905762, 'rewards/meter/std': 0.4230419099330902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.20364098250865936, 'rewards/repeat_penalty/mean': 0.951219379901886, 'rewards/repeat_penalty/std': 0.08588853478431702, 'rewards/repeat_floor/mean': 0.9942461252212524, 'rewards/repeat_floor/std': 0.008486769162118435, 'rewards/near_duplicate_penalty/mean': 0.9789485931396484, 'rewards/near_duplicate_penalty/std': 0.010279958136379719, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9711538553237915, 'rewards/distinct_2/std': 0.0815892368555069, 'rewards/total_composite/mean': 0.26205557584762573, 'rewards/total_composite/std': 0.24468806385993958, 'reward': 0.26205557584762573, 'reward_std': 0.24468806385993958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.182035893201828, 'sampling/sampling_logp_difference/max': 2.0241386890411377, 'sampling/importance_sampling_ratio/min': 0.13210758566856384, 'sampling/importance_sampling_ratio/mean': 1.0191036462783813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1395717039704323, 'clip_ratio/low_mean': 0.09946038946509361, 'clip_ratio/low_min': 0.09946038946509361, 'clip_ratio/high_mean': 0.06316131725907326, 'clip_ratio/high_max': 0.06316131725907326, 'clip_ratio/region_mean': 0.16262170672416687, 'reward_total_mean': 0.26205557584762573, 'reward_meter_mean': 0.5002398490905762, 'reward_meter_std': 0.4230419099330902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.951219379901886, 'reward_repeat_penalty_std': 0.08588853478431702, 'reward_repeat_floor_mean': 0.9942461252212524, 'reward_repeat_floor_std': 0.008486769162118435, 'reward_near_duplicate_penalty_mean': 0.9789485931396484, 'reward_near_duplicate_penalty_std': 0.010279958136379719, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9711538553237915, 'reward_distinct_2_std': 0.0815892368555069, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.20364098250865936, 'reward_total_composite_mean': 0.26205557584762573, 'reward_total_composite_std': 0.24468806385993958, 'epoch': 0.03}

 18%|█▊        | 589/3300 [2:08:25<8:11:28, 10.88s/it]INFO 04-12 18:06:51 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 590/3300 [2:08:38<8:46:05, 11.65s/it]2026-04-12 18:07:04,614 | INFO | train_grpo_train | metrics_logged mode=train step=590

                                                      
{'loss': -0.0411, 'grad_norm': 5.579891204833984, 'learning_rate': 8.215151515151517e-06, 'num_tokens': 1090394.0, 'completions/mean_length': 121.75, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7304646968841553, 'rewards/meter/std': 0.34969362616539, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9395439624786377, 'rewards/lexical_plausibility/std': 0.15603625774383545, 'rewards/judge_quality/mean': 0.5762500166893005, 'rewards/judge_quality/std': 0.30123260617256165, 'rewards/repeat_penalty/mean': 0.9695484638214111, 'rewards/repeat_penalty/std': 0.037265244871377945, 'rewards/repeat_floor/mean': 0.996529221534729, 'rewards/repeat_floor/std': 0.00366105861030519, 'rewards/near_duplicate_penalty/mean': 0.988242506980896, 'rewards/near_duplicate_penalty/std': 0.009074272587895393, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9810318946838379, 'rewards/distinct_2/std': 0.035094089806079865, 'rewards/total_composite/mean': 0.47439393401145935, 'rewards/total_composite/std': 0.35635241866111755, 'reward': 0.47439393401145935, 'reward_std': 0.35635238885879517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1668364405632019, 'sampling/sampling_logp_difference/max': 2.9669766426086426, 'sampling/importance_sampling_ratio/min': 0.051458653062582016, 'sampling/importance_sampling_ratio/mean': 1.0063886642456055, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8602184131741524, 'clip_ratio/low_mean': 0.0969181340187788, 'clip_ratio/low_min': 0.0969181340187788, 'clip_ratio/high_mean': 0.05962243489921093, 'clip_ratio/high_max': 0.05962243489921093, 'clip_ratio/region_mean': 0.15654056891798973, 'reward_total_mean': 0.47439393401145935, 'reward_meter_mean': 0.7304646968841553, 'reward_meter_std': 0.34969362616539, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9395439624786377, 'reward_lexical_plausibility_std': 0.15603625774383545, 'reward_repeat_penalty_mean': 0.9695484638214111, 'reward_repeat_penalty_std': 0.037265244871377945, 'reward_repeat_floor_mean': 0.996529221534729, 'reward_repeat_floor_std': 0.00366105861030519, 'reward_near_duplicate_penalty_mean': 0.988242506980896, 'reward_near_duplicate_penalty_std': 0.009074272587895393, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9810318946838379, 'reward_distinct_2_std': 0.035094089806079865, 'reward_judge_quality_mean': 0.5762500166893005, 'reward_judge_quality_std': 0.30123260617256165, 'reward_total_composite_mean': 0.47439393401145935, 'reward_total_composite_std': 0.35635241866111755, 'epoch': 0.03}

 18%|█▊        | 590/3300 [2:08:38<8:46:05, 11.65s/it]INFO 04-12 18:07:04 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 591/3300 [2:08:47<8:07:05, 10.79s/it]2026-04-12 18:07:13,397 | INFO | train_grpo_train | metrics_logged mode=train step=591

                                                      
{'loss': 0.0622, 'grad_norm': 15.370757102966309, 'learning_rate': 8.212121212121212e-06, 'num_tokens': 1092228.0, 'completions/mean_length': 46.25, 'completions/min_length': 35.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.587675929069519, 'rewards/meter/std': 0.37471726536750793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9642857313156128, 'rewards/lexical_plausibility/std': 0.10101525485515594, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9389331340789795, 'rewards/repeat_penalty/std': 0.06536339968442917, 'rewards/repeat_floor/mean': 0.9920876622200012, 'rewards/repeat_floor/std': 0.008928068913519382, 'rewards/near_duplicate_penalty/mean': 0.9615490436553955, 'rewards/near_duplicate_penalty/std': 0.05235946550965309, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9761699438095093, 'rewards/distinct_2/std': 0.03302708640694618, 'rewards/total_composite/mean': 0.2845801115036011, 'rewards/total_composite/std': 0.20437119901180267, 'reward': 0.2845801115036011, 'reward_std': 0.20437119901180267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17702707648277283, 'sampling/sampling_logp_difference/max': 2.561309337615967, 'sampling/importance_sampling_ratio/min': 0.07848146557807922, 'sampling/importance_sampling_ratio/mean': 1.0149379968643188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.069802649319172, 'clip_ratio/low_mean': 0.07182778418064117, 'clip_ratio/low_min': 0.07182778418064117, 'clip_ratio/high_mean': 0.1115767490118742, 'clip_ratio/high_max': 0.1115767490118742, 'clip_ratio/region_mean': 0.18340453319251537, 'reward_total_mean': 0.2845801115036011, 'reward_meter_mean': 0.587675929069519, 'reward_meter_std': 0.37471726536750793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9642857313156128, 'reward_lexical_plausibility_std': 0.10101525485515594, 'reward_repeat_penalty_mean': 0.9389331340789795, 'reward_repeat_penalty_std': 0.06536339968442917, 'reward_repeat_floor_mean': 0.9920876622200012, 'reward_repeat_floor_std': 0.008928068913519382, 'reward_near_duplicate_penalty_mean': 0.9615490436553955, 'reward_near_duplicate_penalty_std': 0.05235946550965309, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9761699438095093, 'reward_distinct_2_std': 0.03302708640694618, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.2845801115036011, 'reward_total_composite_std': 0.20437119901180267, 'epoch': 0.03}

 18%|█▊        | 591/3300 [2:08:47<8:07:05, 10.79s/it]INFO 04-12 18:07:13 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 592/3300 [2:08:55<7:28:44,  9.94s/it]2026-04-12 18:07:21,399 | INFO | train_grpo_train | metrics_logged mode=train step=592

                                                      
{'loss': 0.021, 'grad_norm': 18.695114135742188, 'learning_rate': 8.20909090909091e-06, 'num_tokens': 1093785.0, 'completions/mean_length': 41.625, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.147118479013443, 'rewards/meter/std': 0.22690625488758087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.731249988079071, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.9846284985542297, 'rewards/repeat_penalty/std': 0.029142936691641808, 'rewards/repeat_floor/mean': 0.9981726408004761, 'rewards/repeat_floor/std': 0.0031048913951963186, 'rewards/near_duplicate_penalty/mean': 0.9928339719772339, 'rewards/near_duplicate_penalty/std': 0.009780322201550007, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.10930247604846954, 'rewards/total_composite/std': 0.20511575043201447, 'reward': 0.10930247604846954, 'reward_std': 0.20511573553085327, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16905662417411804, 'sampling/sampling_logp_difference/max': 1.3927001953125, 'sampling/importance_sampling_ratio/min': 0.2484036535024643, 'sampling/importance_sampling_ratio/mean': 1.0127214193344116, 'sampling/importance_sampling_ratio/max': 1.9360523223876953, 'entropy': 1.1897085085511208, 'clip_ratio/low_mean': 0.059014896396547556, 'clip_ratio/low_min': 0.059014896396547556, 'clip_ratio/high_mean': 0.03730158880352974, 'clip_ratio/high_max': 0.03730158880352974, 'clip_ratio/region_mean': 0.0963164852000773, 'reward_total_mean': 0.10930247604846954, 'reward_meter_mean': 0.147118479013443, 'reward_meter_std': 0.22690625488758087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9846284985542297, 'reward_repeat_penalty_std': 0.029142936691641808, 'reward_repeat_floor_mean': 0.9981726408004761, 'reward_repeat_floor_std': 0.0031048913951963186, 'reward_near_duplicate_penalty_mean': 0.9928339719772339, 'reward_near_duplicate_penalty_std': 0.009780322201550007, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.731249988079071, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.10930247604846954, 'reward_total_composite_std': 0.20511575043201447, 'epoch': 0.03}

 18%|█▊        | 592/3300 [2:08:55<7:28:44,  9.94s/it]INFO 04-12 18:07:21 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 593/3300 [2:09:03<7:01:26,  9.34s/it]2026-04-12 18:07:29,303 | INFO | train_grpo_train | metrics_logged mode=train step=593

                                                      
{'loss': 0.0997, 'grad_norm': 23.939882278442383, 'learning_rate': 8.206060606060607e-06, 'num_tokens': 1095159.0, 'completions/mean_length': 23.75, 'completions/min_length': 17.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.9602457284927368, 'rewards/meter/std': 0.0740094780921936, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9771242141723633, 'rewards/lexical_plausibility/std': 0.0647025778889656, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6635200381278992, 'rewards/total_composite/std': 0.2914673686027527, 'reward': 0.6635200381278992, 'reward_std': 0.2914673686027527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19499801099300385, 'sampling/sampling_logp_difference/max': 1.6875286102294922, 'sampling/importance_sampling_ratio/min': 0.30928120017051697, 'sampling/importance_sampling_ratio/mean': 1.0430655479431152, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4827243760228157, 'clip_ratio/low_mean': 0.05938390363007784, 'clip_ratio/low_min': 0.05938390363007784, 'clip_ratio/high_mean': 0.11629943223670125, 'clip_ratio/high_max': 0.11629943223670125, 'clip_ratio/region_mean': 0.1756833358667791, 'reward_total_mean': 0.6635200381278992, 'reward_meter_mean': 0.9602457284927368, 'reward_meter_std': 0.0740094780921936, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9771242141723633, 'reward_lexical_plausibility_std': 0.0647025778889656, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.6635200381278992, 'reward_total_composite_std': 0.2914673686027527, 'epoch': 0.03}

 18%|█▊        | 593/3300 [2:09:03<7:01:26,  9.34s/it]INFO 04-12 18:07:29 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 594/3300 [2:09:14<7:24:13,  9.85s/it]2026-04-12 18:07:40,338 | INFO | train_grpo_train | metrics_logged mode=train step=594

                                                      
{'loss': 0.0272, 'grad_norm': 5.911684989929199, 'learning_rate': 8.203030303030304e-06, 'num_tokens': 1098190.0, 'completions/mean_length': 207.875, 'completions/min_length': 183.0, 'completions/max_length': 228.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 207.875, 'completions/min_terminated_length': 183.0, 'completions/max_terminated_length': 228.0, 'rewards/meter/mean': 0.9815651178359985, 'rewards/meter/std': 0.020114127546548843, 'rewards/count_adherence/mean': 0.7678571343421936, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9303570985794067, 'rewards/count_floor/std': 0.02218066342175007, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.2147345095872879, 'rewards/repeat_penalty/std': 0.23416870832443237, 'rewards/repeat_floor/mean': 0.9049199819564819, 'rewards/repeat_floor/std': 0.03535691276192665, 'rewards/near_duplicate_penalty/mean': 0.822540283203125, 'rewards/near_duplicate_penalty/std': 0.0837874487042427, 'rewards/opening_diversity/mean': 0.9546568393707275, 'rewards/opening_diversity/std': 0.08771829307079315, 'rewards/distinct_2/mean': 0.26955029368400574, 'rewards/distinct_2/std': 0.24233220517635345, 'rewards/total_composite/mean': 0.27052873373031616, 'rewards/total_composite/std': 0.06528563797473907, 'reward': 0.27052873373031616, 'reward_std': 0.06528563052415848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10569387674331665, 'sampling/sampling_logp_difference/max': 2.49171781539917, 'sampling/importance_sampling_ratio/min': 0.08276766538619995, 'sampling/importance_sampling_ratio/mean': 0.9955204725265503, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6036545783281326, 'clip_ratio/low_mean': 0.008064515888690948, 'clip_ratio/low_min': 0.008064515888690948, 'clip_ratio/high_mean': 0.09638627106323838, 'clip_ratio/high_max': 0.09638627106323838, 'clip_ratio/region_mean': 0.10445078695192933, 'reward_total_mean': 0.27052873373031616, 'reward_meter_mean': 0.9815651178359985, 'reward_meter_std': 0.020114127546548843, 'reward_count_adherence_mean': 0.7678571343421936, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9303570985794067, 'reward_count_floor_std': 0.02218066342175007, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2147345095872879, 'reward_repeat_penalty_std': 0.23416870832443237, 'reward_repeat_floor_mean': 0.9049199819564819, 'reward_repeat_floor_std': 0.03535691276192665, 'reward_near_duplicate_penalty_mean': 0.822540283203125, 'reward_near_duplicate_penalty_std': 0.0837874487042427, 'reward_opening_diversity_mean': 0.9546568393707275, 'reward_opening_diversity_std': 0.08771829307079315, 'reward_distinct_2_mean': 0.26955029368400574, 'reward_distinct_2_std': 0.24233220517635345, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.27052873373031616, 'reward_total_composite_std': 0.06528563797473907, 'epoch': 0.03}

 18%|█▊        | 594/3300 [2:09:14<7:24:13,  9.85s/it]INFO 04-12 18:07:40 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 595/3300 [2:09:21<6:54:53,  9.20s/it]2026-04-12 18:07:48,032 | INFO | train_grpo_train | metrics_logged mode=train step=595

                                                      
{'loss': 0.1301, 'grad_norm': 16.424837112426758, 'learning_rate': 8.2e-06, 'num_tokens': 1099864.0, 'completions/mean_length': 39.25, 'completions/min_length': 34.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.4144309461116791, 'rewards/meter/std': 0.3266526162624359, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.9873858690261841, 'rewards/repeat_penalty/std': 0.014483274891972542, 'rewards/repeat_floor/mean': 0.99810791015625, 'rewards/repeat_floor/std': 0.0021724931430071592, 'rewards/near_duplicate_penalty/mean': 0.9873858690261841, 'rewards/near_duplicate_penalty/std': 0.014483274891972542, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31065911054611206, 'rewards/total_composite/std': 0.33523043990135193, 'reward': 0.31065911054611206, 'reward_std': 0.33523041009902954, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19718889892101288, 'sampling/sampling_logp_difference/max': 2.0223336219787598, 'sampling/importance_sampling_ratio/min': 0.1323462575674057, 'sampling/importance_sampling_ratio/mean': 1.0170778036117554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8910216242074966, 'clip_ratio/low_mean': 0.09677564026787877, 'clip_ratio/low_min': 0.09677564026787877, 'clip_ratio/high_mean': 0.06555268634110689, 'clip_ratio/high_max': 0.06555268634110689, 'clip_ratio/region_mean': 0.16232832660898566, 'reward_total_mean': 0.31065911054611206, 'reward_meter_mean': 0.4144309461116791, 'reward_meter_std': 0.3266526162624359, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9873858690261841, 'reward_repeat_penalty_std': 0.014483274891972542, 'reward_repeat_floor_mean': 0.99810791015625, 'reward_repeat_floor_std': 0.0021724931430071592, 'reward_near_duplicate_penalty_mean': 0.9873858690261841, 'reward_near_duplicate_penalty_std': 0.014483274891972542, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.31065911054611206, 'reward_total_composite_std': 0.33523043990135193, 'epoch': 0.03}

 18%|█▊        | 595/3300 [2:09:21<6:54:53,  9.20s/it]INFO 04-12 18:07:48 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 596/3300 [2:09:29<6:39:57,  8.87s/it]2026-04-12 18:07:56,142 | INFO | train_grpo_train | metrics_logged mode=train step=596

                                                      
{'loss': 0.0465, 'grad_norm': 16.912214279174805, 'learning_rate': 8.196969696969698e-06, 'num_tokens': 1101336.0, 'completions/mean_length': 44.0, 'completions/min_length': 40.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7226223349571228, 'rewards/meter/std': 0.32787397503852844, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.78125, 'rewards/judge_quality/std': 0.2126995027065277, 'rewards/repeat_penalty/mean': 0.9540501236915588, 'rewards/repeat_penalty/std': 0.08837031573057175, 'rewards/repeat_floor/mean': 0.9942610859870911, 'rewards/repeat_floor/std': 0.010143347084522247, 'rewards/near_duplicate_penalty/mean': 0.976163387298584, 'rewards/near_duplicate_penalty/std': 0.03246826305985451, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9759615659713745, 'rewards/distinct_2/std': 0.06799104809761047, 'rewards/total_composite/mean': 0.5714898705482483, 'rewards/total_composite/std': 0.30262914299964905, 'reward': 0.5714898705482483, 'reward_std': 0.30262914299964905, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1931528002023697, 'sampling/sampling_logp_difference/max': 1.415879249572754, 'sampling/importance_sampling_ratio/min': 0.2427121251821518, 'sampling/importance_sampling_ratio/mean': 0.9944076538085938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1498000472784042, 'clip_ratio/low_mean': 0.050173611380159855, 'clip_ratio/low_min': 0.050173611380159855, 'clip_ratio/high_mean': 0.12886305153369904, 'clip_ratio/high_max': 0.12886305153369904, 'clip_ratio/region_mean': 0.1790366629138589, 'reward_total_mean': 0.5714898705482483, 'reward_meter_mean': 0.7226223349571228, 'reward_meter_std': 0.32787397503852844, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9540501236915588, 'reward_repeat_penalty_std': 0.08837031573057175, 'reward_repeat_floor_mean': 0.9942610859870911, 'reward_repeat_floor_std': 0.010143347084522247, 'reward_near_duplicate_penalty_mean': 0.976163387298584, 'reward_near_duplicate_penalty_std': 0.03246826305985451, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9759615659713745, 'reward_distinct_2_std': 0.06799104809761047, 'reward_judge_quality_mean': 0.78125, 'reward_judge_quality_std': 0.2126995027065277, 'reward_total_composite_mean': 0.5714898705482483, 'reward_total_composite_std': 0.30262914299964905, 'epoch': 0.03}

 18%|█▊        | 596/3300 [2:09:29<6:39:57,  8.87s/it]INFO 04-12 18:07:56 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 597/3300 [2:09:44<7:53:43, 10.52s/it]2026-04-12 18:08:10,489 | INFO | train_grpo_train | metrics_logged mode=train step=597

                                                      
{'loss': -0.069, 'grad_norm': 6.055530071258545, 'learning_rate': 8.193939393939394e-06, 'num_tokens': 1103138.0, 'completions/mean_length': 128.25, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.42857360839844, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.6486878395080566, 'rewards/meter/std': 0.41599565744400024, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9223067760467529, 'rewards/lexical_plausibility/std': 0.21974965929985046, 'rewards/judge_quality/mean': 0.48375001549720764, 'rewards/judge_quality/std': 0.30514341592788696, 'rewards/repeat_penalty/mean': 0.9840086698532104, 'rewards/repeat_penalty/std': 0.024673406034708023, 'rewards/repeat_floor/mean': 0.9979274272918701, 'rewards/repeat_floor/std': 0.0028727182652801275, 'rewards/near_duplicate_penalty/mean': 0.9896791577339172, 'rewards/near_duplicate_penalty/std': 0.011669785715639591, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9941724538803101, 'rewards/distinct_2/std': 0.016482684761285782, 'rewards/total_composite/mean': 0.259575217962265, 'rewards/total_composite/std': 0.24509544670581818, 'reward': 0.259575217962265, 'reward_std': 0.24509546160697937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16639792919158936, 'sampling/sampling_logp_difference/max': 2.0793254375457764, 'sampling/importance_sampling_ratio/min': 0.1250145137310028, 'sampling/importance_sampling_ratio/mean': 1.0385816097259521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1286614760756493, 'clip_ratio/low_mean': 0.07283814251422882, 'clip_ratio/low_min': 0.07283814251422882, 'clip_ratio/high_mean': 0.08682319335639477, 'clip_ratio/high_max': 0.08682319335639477, 'clip_ratio/region_mean': 0.1596613358706236, 'reward_total_mean': 0.259575217962265, 'reward_meter_mean': 0.6486878395080566, 'reward_meter_std': 0.41599565744400024, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9223067760467529, 'reward_lexical_plausibility_std': 0.21974965929985046, 'reward_repeat_penalty_mean': 0.9840086698532104, 'reward_repeat_penalty_std': 0.024673406034708023, 'reward_repeat_floor_mean': 0.9979274272918701, 'reward_repeat_floor_std': 0.0028727182652801275, 'reward_near_duplicate_penalty_mean': 0.9896791577339172, 'reward_near_duplicate_penalty_std': 0.011669785715639591, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9941724538803101, 'reward_distinct_2_std': 0.016482684761285782, 'reward_judge_quality_mean': 0.48375001549720764, 'reward_judge_quality_std': 0.30514341592788696, 'reward_total_composite_mean': 0.259575217962265, 'reward_total_composite_std': 0.24509544670581818, 'epoch': 0.03}

 18%|█▊        | 597/3300 [2:09:44<7:53:43, 10.52s/it]INFO 04-12 18:08:10 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 598/3300 [2:09:59<8:51:49, 11.81s/it]2026-04-12 18:08:25,353 | INFO | train_grpo_train | metrics_logged mode=train step=598

                                                      
{'loss': -0.2691, 'grad_norm': 2.5904784202575684, 'learning_rate': 8.190909090909091e-06, 'num_tokens': 1105898.0, 'completions/mean_length': 271.0, 'completions/min_length': 172.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 190.6666717529297, 'completions/min_terminated_length': 172.0, 'completions/max_terminated_length': 203.0, 'rewards/meter/mean': 0.9814270734786987, 'rewards/meter/std': 0.015045571140944958, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.3148418068885803, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9357142448425293, 'rewards/count_floor/std': 0.09445255249738693, 'rewards/lexical_plausibility/mean': 0.9016356468200684, 'rewards/lexical_plausibility/std': 0.15866655111312866, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.486883282661438, 'rewards/repeat_penalty/std': 0.3581705093383789, 'rewards/repeat_floor/mean': 0.9371234774589539, 'rewards/repeat_floor/std': 0.05329469218850136, 'rewards/near_duplicate_penalty/mean': 0.8602719306945801, 'rewards/near_duplicate_penalty/std': 0.13260896503925323, 'rewards/opening_diversity/mean': 0.8804687261581421, 'rewards/opening_diversity/std': 0.15589021146297455, 'rewards/distinct_2/mean': 0.6139395236968994, 'rewards/distinct_2/std': 0.3230123817920685, 'rewards/total_composite/mean': 0.23201525211334229, 'rewards/total_composite/std': 0.12587951123714447, 'reward': 0.23201525211334229, 'reward_std': 0.12587951123714447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13244052231311798, 'sampling/sampling_logp_difference/max': 2.8504509925842285, 'sampling/importance_sampling_ratio/min': 0.05781824141740799, 'sampling/importance_sampling_ratio/mean': 1.0080112218856812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6020743548870087, 'clip_ratio/low_mean': 0.012354651466012001, 'clip_ratio/low_min': 0.012354651466012001, 'clip_ratio/high_mean': 0.091948670335114, 'clip_ratio/high_max': 0.091948670335114, 'clip_ratio/region_mean': 0.104303321801126, 'reward_total_mean': 0.23201525211334229, 'reward_meter_mean': 0.9814270734786987, 'reward_meter_std': 0.015045571140944958, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.3148418068885803, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9357142448425293, 'reward_count_floor_std': 0.09445255249738693, 'reward_lexical_plausibility_mean': 0.9016356468200684, 'reward_lexical_plausibility_std': 0.15866655111312866, 'reward_repeat_penalty_mean': 0.486883282661438, 'reward_repeat_penalty_std': 0.3581705093383789, 'reward_repeat_floor_mean': 0.9371234774589539, 'reward_repeat_floor_std': 0.05329469218850136, 'reward_near_duplicate_penalty_mean': 0.8602719306945801, 'reward_near_duplicate_penalty_std': 0.13260896503925323, 'reward_opening_diversity_mean': 0.8804687261581421, 'reward_opening_diversity_std': 0.15589021146297455, 'reward_distinct_2_mean': 0.6139395236968994, 'reward_distinct_2_std': 0.3230123817920685, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.23201525211334229, 'reward_total_composite_std': 0.12587951123714447, 'epoch': 0.03}

 18%|█▊        | 598/3300 [2:09:59<8:51:49, 11.81s/it]INFO 04-12 18:08:25 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 599/3300 [2:10:07<8:07:57, 10.84s/it]2026-04-12 18:08:33,889 | INFO | train_grpo_train | metrics_logged mode=train step=599

                                                      
{'loss': 0.1166, 'grad_norm': 23.549551010131836, 'learning_rate': 8.187878787878788e-06, 'num_tokens': 1107294.0, 'completions/mean_length': 21.5, 'completions/min_length': 17.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7523049116134644, 'rewards/meter/std': 0.31828737258911133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.2889884114265442, 'rewards/repeat_penalty/mean': 0.6145009994506836, 'rewards/repeat_penalty/std': 0.25493043661117554, 'rewards/repeat_floor/mean': 0.960392951965332, 'rewards/repeat_floor/std': 0.044148046523332596, 'rewards/near_duplicate_penalty/mean': 0.8849912285804749, 'rewards/near_duplicate_penalty/std': 0.16451109945774078, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.2985857129096985, 'rewards/total_composite/std': 0.20206375420093536, 'reward': 0.2985857129096985, 'reward_std': 0.20206372439861298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18337790668010712, 'sampling/sampling_logp_difference/max': 1.7136726379394531, 'sampling/importance_sampling_ratio/min': 0.18020275235176086, 'sampling/importance_sampling_ratio/mean': 1.0237829685211182, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1202788725495338, 'clip_ratio/low_mean': 0.027953165117651224, 'clip_ratio/low_min': 0.027953165117651224, 'clip_ratio/high_mean': 0.09070870745927095, 'clip_ratio/high_max': 0.09070870745927095, 'clip_ratio/region_mean': 0.11866187257692218, 'reward_total_mean': 0.2985857129096985, 'reward_meter_mean': 0.7523049116134644, 'reward_meter_std': 0.31828737258911133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6145009994506836, 'reward_repeat_penalty_std': 0.25493043661117554, 'reward_repeat_floor_mean': 0.960392951965332, 'reward_repeat_floor_std': 0.044148046523332596, 'reward_near_duplicate_penalty_mean': 0.8849912285804749, 'reward_near_duplicate_penalty_std': 0.16451109945774078, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.2889884114265442, 'reward_total_composite_mean': 0.2985857129096985, 'reward_total_composite_std': 0.20206375420093536, 'epoch': 0.03}

 18%|█▊        | 599/3300 [2:10:07<8:07:57, 10.84s/it]INFO 04-12 18:08:34 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 600/3300 [2:10:21<8:52:43, 11.84s/it]2026-04-12 18:08:48,057 | INFO | train_grpo_train | metrics_logged mode=train step=600

                                                      
{'loss': -0.035, 'grad_norm': 6.604052543640137, 'learning_rate': 8.184848484848486e-06, 'num_tokens': 1108936.0, 'completions/mean_length': 109.25, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.71428680419922, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.3260256052017212, 'rewards/meter/std': 0.3266305923461914, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.07559289783239365, 'rewards/lexical_plausibility/mean': 0.9401394724845886, 'rewards/lexical_plausibility/std': 0.16931118071079254, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.27614372968673706, 'rewards/repeat_penalty/mean': 0.9594191908836365, 'rewards/repeat_penalty/std': 0.08538877964019775, 'rewards/repeat_floor/mean': 0.99672532081604, 'rewards/repeat_floor/std': 0.005039273761212826, 'rewards/near_duplicate_penalty/mean': 0.9906691908836365, 'rewards/near_duplicate_penalty/std': 0.012051167897880077, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1640865057706833, 'rewards/total_composite/std': 0.19193319976329803, 'reward': 0.1640865057706833, 'reward_std': 0.19193319976329803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1753326654434204, 'sampling/sampling_logp_difference/max': 1.8509302139282227, 'sampling/importance_sampling_ratio/min': 0.15709097683429718, 'sampling/importance_sampling_ratio/mean': 1.0242491960525513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1086146757006645, 'clip_ratio/low_mean': 0.07389158569276333, 'clip_ratio/low_min': 0.07389158569276333, 'clip_ratio/high_mean': 0.0506847333163023, 'clip_ratio/high_max': 0.0506847333163023, 'clip_ratio/region_mean': 0.12457631900906563, 'reward_total_mean': 0.1640865057706833, 'reward_meter_mean': 0.3260256052017212, 'reward_meter_std': 0.3266305923461914, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.07559289783239365, 'reward_lexical_plausibility_mean': 0.9401394724845886, 'reward_lexical_plausibility_std': 0.16931118071079254, 'reward_repeat_penalty_mean': 0.9594191908836365, 'reward_repeat_penalty_std': 0.08538877964019775, 'reward_repeat_floor_mean': 0.99672532081604, 'reward_repeat_floor_std': 0.005039273761212826, 'reward_near_duplicate_penalty_mean': 0.9906691908836365, 'reward_near_duplicate_penalty_std': 0.012051167897880077, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.27614372968673706, 'reward_total_composite_mean': 0.1640865057706833, 'reward_total_composite_std': 0.19193319976329803, 'epoch': 0.03}

 18%|█▊        | 600/3300 [2:10:21<8:52:43, 11.84s/it]INFO 04-12 18:08:48 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:12<01:10,  6.40s/it]

 23%|██▎       | 3/13 [00:24<01:25,  8.53s/it]

 31%|███       | 4/13 [00:34<01:23,  9.27s/it]

 38%|███▊      | 5/13 [00:46<01:20, 10.08s/it]

 46%|████▌     | 6/13 [00:58<01:14, 10.69s/it]

 54%|█████▍    | 7/13 [01:10<01:06, 11.16s/it]

 62%|██████▏   | 8/13 [01:24<01:00, 12.04s/it]

 69%|██████▉   | 9/13 [01:35<00:47, 11.84s/it]

 77%|███████▋  | 10/13 [01:48<00:35, 11.96s/it]

 85%|████████▍ | 11/13 [01:59<00:23, 11.82s/it]

 92%|█████████▏| 12/13 [02:11<00:11, 11.79s/it]

100%|██████████| 13/13 [02:23<00:00, 11.86s/it]2026-04-12 18:11:21,902 | INFO | train_grpo_train | metrics_logged mode=eval step=600

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 153.7922, 'eval_samples_per_second': 0.676, 'eval_steps_per_second': 0.085, 'eval_num_tokens': 1108936.0, 'eval_completions/mean_length': 216.35576923076923, 'eval_completions/min_length': 51.0, 'eval_completions/max_length': 489.2307692307692, 'eval_completions/clipped_ratio': 0.15384615384615385, 'eval_completions/mean_terminated_length': 161.36337338961087, 'eval_completions/min_terminated_length': 51.0, 'eval_completions/max_terminated_length': 337.0, 'eval_rewards/meter/mean': 0.5323957594541403, 'eval_rewards/meter/std': 0.3750814680869763, 'eval_rewards/count_adherence/mean': 0.8576028759662921, 'eval_rewards/count_adherence/std': 0.19899147462386352, 'eval_rewards/arabic_clean/mean': 0.8076923076923077, 'eval_rewards/arabic_clean/std': 0.39691325105153596, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9971153827813956, 'eval_rewards/arabic_floor/std': 0.008158923341677738, 'eval_rewards/count_floor/mean': 0.957280851327456, 'eval_rewards/count_floor/std': 0.0596974457685764, 'eval_rewards/lexical_plausibility/mean': 0.944066753754249, 'eval_rewards/lexical_plausibility/std': 0.12990799660866076, 'eval_rewards/judge_quality/mean': 0.34480769015275514, 'eval_rewards/judge_quality/std': 0.2326679011950126, 'eval_rewards/repeat_penalty/mean': 0.6714496314525604, 'eval_rewards/repeat_penalty/std': 0.33662569580169827, 'eval_rewards/repeat_floor/mean': 0.9596992272597092, 'eval_rewards/repeat_floor/std': 0.04648636933416128, 'eval_rewards/near_duplicate_penalty/mean': 0.9150083982027494, 'eval_rewards/near_duplicate_penalty/std': 0.10893321180572876, 'eval_rewards/opening_diversity/mean': 0.96364092368346, 'eval_rewards/opening_diversity/std': 0.07593007643635456, 'eval_rewards/distinct_2/mean': 0.7181056050153879, 'eval_rewards/distinct_2/std': 0.3324068910800494, 'eval_rewards/total_composite/mean': 0.16826352018576402, 'eval_rewards/total_composite/std': 0.16927679972006723, 'eval_reward': 0.16826352018576402, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0662076072050975, 'eval_sampling/sampling_logp_difference/max': 1.1472235459547777, 'eval_sampling/importance_sampling_ratio/min': 0.3202722118451045, 'eval_sampling/importance_sampling_ratio/mean': 1.01552990766672, 'eval_sampling/importance_sampling_ratio/max': 1.5114481632526104, 'eval_entropy': 0.7519237811748798, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.16826352018576402, 'eval_reward_meter_mean': 0.5323957594541403, 'eval_reward_meter_std': 0.3750814680869763, 'eval_reward_count_adherence_mean': 0.8576028759662921, 'eval_reward_count_adherence_std': 0.19899147462386352, 'eval_reward_arabic_clean_mean': 0.8076923076923077, 'eval_reward_arabic_clean_std': 0.39691325105153596, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9971153827813956, 'eval_reward_arabic_floor_std': 0.008158923341677738, 'eval_reward_count_floor_mean': 0.957280851327456, 'eval_reward_count_floor_std': 0.0596974457685764, 'eval_reward_lexical_plausibility_mean': 0.944066753754249, 'eval_reward_lexical_plausibility_std': 0.12990799660866076, 'eval_reward_repeat_penalty_mean': 0.6714496314525604, 'eval_reward_repeat_penalty_std': 0.33662569580169827, 'eval_reward_repeat_floor_mean': 0.9596992272597092, 'eval_reward_repeat_floor_std': 0.04648636933416128, 'eval_reward_near_duplicate_penalty_mean': 0.9150083982027494, 'eval_reward_near_duplicate_penalty_std': 0.10893321180572876, 'eval_reward_opening_diversity_mean': 0.96364092368346, 'eval_reward_opening_diversity_std': 0.07593007643635456, 'eval_reward_distinct_2_mean': 0.7181056050153879, 'eval_reward_distinct_2_std': 0.3324068910800494, 'eval_reward_judge_quality_mean': 0.34480769015275514, 'eval_reward_judge_quality_std': 0.2326679011950126, 'eval_reward_total_composite_mean': 0.16826352018576402, 'eval_reward_total_composite_std': 0.16927679972006723, 'epoch': 0.03}

 18%|█▊        | 600/3300 [2:12:55<8:52:43, 11.84s/it]

100%|██████████| 13/13 [02:23<00:00, 11.86s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 18%|█▊        | 601/3300 [2:13:12<44:31:41, 59.39s/it]2026-04-12 18:11:38,417 | INFO | train_grpo_train | metrics_logged mode=train step=601

                                                       
{'loss': -0.1004, 'grad_norm': 4.635842323303223, 'learning_rate': 8.181818181818183e-06, 'num_tokens': 1110662.0, 'completions/mean_length': 110.75, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.42857360839844, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8872721791267395, 'rewards/meter/std': 0.20902809500694275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9427574872970581, 'rewards/lexical_plausibility/std': 0.16190630197525024, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9237461090087891, 'rewards/repeat_penalty/std': 0.16513754427433014, 'rewards/repeat_floor/mean': 0.9914565682411194, 'rewards/repeat_floor/std': 0.018150677904486656, 'rewards/near_duplicate_penalty/mean': 0.9758385419845581, 'rewards/near_duplicate_penalty/std': 0.04522961378097534, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9661751389503479, 'rewards/distinct_2/std': 0.07171893119812012, 'rewards/total_composite/mean': 0.30826202034950256, 'rewards/total_composite/std': 0.1486935317516327, 'reward': 0.30826202034950256, 'reward_std': 0.1486935168504715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15270750224590302, 'sampling/sampling_logp_difference/max': 1.7838517427444458, 'sampling/importance_sampling_ratio/min': 0.1679898500442505, 'sampling/importance_sampling_ratio/mean': 1.0060399770736694, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8088917434215546, 'clip_ratio/low_mean': 0.025216706097126007, 'clip_ratio/low_min': 0.025216706097126007, 'clip_ratio/high_mean': 0.11960294283926487, 'clip_ratio/high_max': 0.11960294283926487, 'clip_ratio/region_mean': 0.14481964893639088, 'reward_total_mean': 0.30826202034950256, 'reward_meter_mean': 0.8872721791267395, 'reward_meter_std': 0.20902809500694275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9427574872970581, 'reward_lexical_plausibility_std': 0.16190630197525024, 'reward_repeat_penalty_mean': 0.9237461090087891, 'reward_repeat_penalty_std': 0.16513754427433014, 'reward_repeat_floor_mean': 0.9914565682411194, 'reward_repeat_floor_std': 0.018150677904486656, 'reward_near_duplicate_penalty_mean': 0.9758385419845581, 'reward_near_duplicate_penalty_std': 0.04522961378097534, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9661751389503479, 'reward_distinct_2_std': 0.07171893119812012, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.30826202034950256, 'reward_total_composite_std': 0.1486935317516327, 'epoch': 0.03}

 18%|█▊        | 601/3300 [2:13:12<44:31:41, 59.39s/it]INFO 04-12 18:11:38 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 602/3300 [2:13:26<34:27:01, 45.97s/it]2026-04-12 18:11:53,056 | INFO | train_grpo_train | metrics_logged mode=train step=602

                                                       
{'loss': -0.0564, 'grad_norm': 4.765085220336914, 'learning_rate': 8.17878787878788e-06, 'num_tokens': 1113831.0, 'completions/mean_length': 243.125, 'completions/min_length': 176.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 204.71429443359375, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 236.0, 'rewards/meter/mean': 0.9532607197761536, 'rewards/meter/std': 0.0884724035859108, 'rewards/count_adherence/mean': 0.984375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.995312511920929, 'rewards/count_floor/std': 0.013258260674774647, 'rewards/lexical_plausibility/mean': 0.9606354236602783, 'rewards/lexical_plausibility/std': 0.0829610824584961, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.2547512948513031, 'rewards/repeat_penalty/mean': 0.5961198806762695, 'rewards/repeat_penalty/std': 0.35815364122390747, 'rewards/repeat_floor/mean': 0.950096845626831, 'rewards/repeat_floor/std': 0.05337326228618622, 'rewards/near_duplicate_penalty/mean': 0.9005410075187683, 'rewards/near_duplicate_penalty/std': 0.11840029805898666, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.6321191787719727, 'rewards/distinct_2/std': 0.35908475518226624, 'rewards/total_composite/mean': 0.31084227561950684, 'rewards/total_composite/std': 0.18292604386806488, 'reward': 0.31084227561950684, 'reward_std': 0.1829260289669037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13875505328178406, 'sampling/sampling_logp_difference/max': 1.8511505126953125, 'sampling/importance_sampling_ratio/min': 0.15705636143684387, 'sampling/importance_sampling_ratio/mean': 1.015142798423767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7933779545128345, 'clip_ratio/low_mean': 0.01934338128194213, 'clip_ratio/low_min': 0.01934338128194213, 'clip_ratio/high_mean': 0.08638909179717302, 'clip_ratio/high_max': 0.08638909179717302, 'clip_ratio/region_mean': 0.10573247307911515, 'reward_total_mean': 0.31084227561950684, 'reward_meter_mean': 0.9532607197761536, 'reward_meter_std': 0.0884724035859108, 'reward_count_adherence_mean': 0.984375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.995312511920929, 'reward_count_floor_std': 0.013258260674774647, 'reward_lexical_plausibility_mean': 0.9606354236602783, 'reward_lexical_plausibility_std': 0.0829610824584961, 'reward_repeat_penalty_mean': 0.5961198806762695, 'reward_repeat_penalty_std': 0.35815364122390747, 'reward_repeat_floor_mean': 0.950096845626831, 'reward_repeat_floor_std': 0.05337326228618622, 'reward_near_duplicate_penalty_mean': 0.9005410075187683, 'reward_near_duplicate_penalty_std': 0.11840029805898666, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.6321191787719727, 'reward_distinct_2_std': 0.35908475518226624, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.2547512948513031, 'reward_total_composite_mean': 0.31084227561950684, 'reward_total_composite_std': 0.18292604386806488, 'epoch': 0.03}

 18%|█▊        | 602/3300 [2:13:26<34:27:01, 45.97s/it]INFO 04-12 18:11:53 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 603/3300 [2:13:40<27:14:30, 36.36s/it]2026-04-12 18:12:07,009 | INFO | train_grpo_train | metrics_logged mode=train step=603

                                                       
{'loss': -0.067, 'grad_norm': 2.303831100463867, 'learning_rate': 8.175757575757577e-06, 'num_tokens': 1115309.0, 'completions/mean_length': 285.75, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 59.5, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.40588515996932983, 'rewards/meter/std': 0.4129186272621155, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 0.723495364189148, 'rewards/lexical_plausibility/std': 0.2994970977306366, 'rewards/judge_quality/mean': 0.24375000596046448, 'rewards/judge_quality/std': 0.22109711170196533, 'rewards/repeat_penalty/mean': 0.9916864633560181, 'rewards/repeat_penalty/std': 0.00821265671402216, 'rewards/repeat_floor/mean': 0.9987529516220093, 'rewards/repeat_floor/std': 0.0012319015804678202, 'rewards/near_duplicate_penalty/mean': 0.9916864633560181, 'rewards/near_duplicate_penalty/std': 0.00821265671402216, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1255316138267517, 'rewards/total_composite/std': 0.16291792690753937, 'reward': 0.1255316138267517, 'reward_std': 0.16291794180870056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19537453353405, 'sampling/sampling_logp_difference/max': 1.6980514526367188, 'sampling/importance_sampling_ratio/min': 0.1830398440361023, 'sampling/importance_sampling_ratio/mean': 1.0048465728759766, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7230313271284103, 'clip_ratio/low_mean': 0.027272727340459824, 'clip_ratio/low_min': 0.027272727340459824, 'clip_ratio/high_mean': 0.060100482776761055, 'clip_ratio/high_max': 0.060100482776761055, 'clip_ratio/region_mean': 0.08737321011722088, 'reward_total_mean': 0.1255316138267517, 'reward_meter_mean': 0.40588515996932983, 'reward_meter_std': 0.4129186272621155, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 0.723495364189148, 'reward_lexical_plausibility_std': 0.2994970977306366, 'reward_repeat_penalty_mean': 0.9916864633560181, 'reward_repeat_penalty_std': 0.00821265671402216, 'reward_repeat_floor_mean': 0.9987529516220093, 'reward_repeat_floor_std': 0.0012319015804678202, 'reward_near_duplicate_penalty_mean': 0.9916864633560181, 'reward_near_duplicate_penalty_std': 0.00821265671402216, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.24375000596046448, 'reward_judge_quality_std': 0.22109711170196533, 'reward_total_composite_mean': 0.1255316138267517, 'reward_total_composite_std': 0.16291792690753937, 'epoch': 0.03}

 18%|█▊        | 603/3300 [2:13:40<27:14:30, 36.36s/it]INFO 04-12 18:12:07 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 604/3300 [2:13:54<22:07:58, 29.55s/it]2026-04-12 18:12:20,676 | INFO | train_grpo_train | metrics_logged mode=train step=604

                                                       
{'loss': -0.1091, 'grad_norm': 2.0855977535247803, 'learning_rate': 8.172727272727273e-06, 'num_tokens': 1116861.0, 'completions/mean_length': 228.0, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 57.60000228881836, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.7371097207069397, 'rewards/meter/std': 0.44047844409942627, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8000662326812744, 'rewards/lexical_plausibility/std': 0.26414236426353455, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.40486329793930054, 'rewards/repeat_penalty/mean': 0.9048452377319336, 'rewards/repeat_penalty/std': 0.1098768338561058, 'rewards/repeat_floor/mean': 0.990536630153656, 'rewards/repeat_floor/std': 0.01118460576981306, 'rewards/near_duplicate_penalty/mean': 0.9740300178527832, 'rewards/near_duplicate_penalty/std': 0.04160391539335251, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9589678049087524, 'rewards/distinct_2/std': 0.058639682829380035, 'rewards/total_composite/mean': 0.4332438111305237, 'rewards/total_composite/std': 0.3999243378639221, 'reward': 0.4332438111305237, 'reward_std': 0.3999243378639221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1707833707332611, 'sampling/sampling_logp_difference/max': 2.1452465057373047, 'sampling/importance_sampling_ratio/min': 0.11703918129205704, 'sampling/importance_sampling_ratio/mean': 1.0064518451690674, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.672496497631073, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11224611662328243, 'clip_ratio/high_max': 0.11224611662328243, 'clip_ratio/region_mean': 0.11224611662328243, 'reward_total_mean': 0.4332438111305237, 'reward_meter_mean': 0.7371097207069397, 'reward_meter_std': 0.44047844409942627, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8000662326812744, 'reward_lexical_plausibility_std': 0.26414236426353455, 'reward_repeat_penalty_mean': 0.9048452377319336, 'reward_repeat_penalty_std': 0.1098768338561058, 'reward_repeat_floor_mean': 0.990536630153656, 'reward_repeat_floor_std': 0.01118460576981306, 'reward_near_duplicate_penalty_mean': 0.9740300178527832, 'reward_near_duplicate_penalty_std': 0.04160391539335251, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9589678049087524, 'reward_distinct_2_std': 0.058639682829380035, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.40486329793930054, 'reward_total_composite_mean': 0.4332438111305237, 'reward_total_composite_std': 0.3999243378639221, 'epoch': 0.03}

 18%|█▊        | 604/3300 [2:13:54<22:07:58, 29.55s/it]INFO 04-12 18:12:20 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 605/3300 [2:14:02<17:21:37, 23.19s/it]2026-04-12 18:12:29,018 | INFO | train_grpo_train | metrics_logged mode=train step=605

                                                       
{'loss': 0.0512, 'grad_norm': 15.98936939239502, 'learning_rate': 8.16969696969697e-06, 'num_tokens': 1118599.0, 'completions/mean_length': 42.25, 'completions/min_length': 32.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8217613697052002, 'rewards/meter/std': 0.25633203983306885, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6675000190734863, 'rewards/judge_quality/std': 0.26162952184677124, 'rewards/repeat_penalty/mean': 0.9919407367706299, 'rewards/repeat_penalty/std': 0.014203781262040138, 'rewards/repeat_floor/mean': 0.9987910985946655, 'rewards/repeat_floor/std': 0.002130561973899603, 'rewards/near_duplicate_penalty/mean': 0.9919407367706299, 'rewards/near_duplicate_penalty/std': 0.014203781262040138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5653389692306519, 'rewards/total_composite/std': 0.2998788058757782, 'reward': 0.5653389692306519, 'reward_std': 0.2998788058757782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1631499081850052, 'sampling/sampling_logp_difference/max': 1.499321460723877, 'sampling/importance_sampling_ratio/min': 0.2232816219329834, 'sampling/importance_sampling_ratio/mean': 1.0280369520187378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.960242934525013, 'clip_ratio/low_mean': 0.07336647715419531, 'clip_ratio/low_min': 0.07336647715419531, 'clip_ratio/high_mean': 0.0735264215618372, 'clip_ratio/high_max': 0.0735264215618372, 'clip_ratio/region_mean': 0.1468928987160325, 'reward_total_mean': 0.5653389692306519, 'reward_meter_mean': 0.8217613697052002, 'reward_meter_std': 0.25633203983306885, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9919407367706299, 'reward_repeat_penalty_std': 0.014203781262040138, 'reward_repeat_floor_mean': 0.9987910985946655, 'reward_repeat_floor_std': 0.002130561973899603, 'reward_near_duplicate_penalty_mean': 0.9919407367706299, 'reward_near_duplicate_penalty_std': 0.014203781262040138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6675000190734863, 'reward_judge_quality_std': 0.26162952184677124, 'reward_total_composite_mean': 0.5653389692306519, 'reward_total_composite_std': 0.2998788058757782, 'epoch': 0.03}

 18%|█▊        | 605/3300 [2:14:02<17:21:37, 23.19s/it]INFO 04-12 18:12:29 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 606/3300 [2:14:12<14:14:23, 19.03s/it]2026-04-12 18:12:38,346 | INFO | train_grpo_train | metrics_logged mode=train step=606

                                                       
{'loss': 0.0967, 'grad_norm': 15.18116283416748, 'learning_rate': 8.166666666666668e-06, 'num_tokens': 1120925.0, 'completions/mean_length': 92.75, 'completions/min_length': 80.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7198826670646667, 'rewards/meter/std': 0.34873974323272705, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.8551812767982483, 'rewards/repeat_penalty/std': 0.2196132391691208, 'rewards/repeat_floor/mean': 0.9843515753746033, 'rewards/repeat_floor/std': 0.024235205724835396, 'rewards/near_duplicate_penalty/mean': 0.9675413370132446, 'rewards/near_duplicate_penalty/std': 0.04756621643900871, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9010597467422485, 'rewards/distinct_2/std': 0.13354972004890442, 'rewards/total_composite/mean': 0.25184184312820435, 'rewards/total_composite/std': 0.14223897457122803, 'reward': 0.25184184312820435, 'reward_std': 0.14223895967006683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1485327035188675, 'sampling/sampling_logp_difference/max': 3.1874661445617676, 'sampling/importance_sampling_ratio/min': 0.04127632826566696, 'sampling/importance_sampling_ratio/mean': 1.0177041292190552, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7814310044050217, 'clip_ratio/low_mean': 0.077626409009099, 'clip_ratio/low_min': 0.077626409009099, 'clip_ratio/high_mean': 0.06050986424088478, 'clip_ratio/high_max': 0.06050986424088478, 'clip_ratio/region_mean': 0.1381362732499838, 'reward_total_mean': 0.25184184312820435, 'reward_meter_mean': 0.7198826670646667, 'reward_meter_std': 0.34873974323272705, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8551812767982483, 'reward_repeat_penalty_std': 0.2196132391691208, 'reward_repeat_floor_mean': 0.9843515753746033, 'reward_repeat_floor_std': 0.024235205724835396, 'reward_near_duplicate_penalty_mean': 0.9675413370132446, 'reward_near_duplicate_penalty_std': 0.04756621643900871, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9010597467422485, 'reward_distinct_2_std': 0.13354972004890442, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.25184184312820435, 'reward_total_composite_std': 0.14223897457122803, 'epoch': 0.03}

 18%|█▊        | 606/3300 [2:14:12<14:14:23, 19.03s/it]INFO 04-12 18:12:38 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 607/3300 [2:14:25<12:57:37, 17.33s/it]2026-04-12 18:12:51,689 | INFO | train_grpo_train | metrics_logged mode=train step=607

                                                       
{'loss': -0.0557, 'grad_norm': 2.7290549278259277, 'learning_rate': 8.163636363636365e-06, 'num_tokens': 1122338.0, 'completions/mean_length': 160.625, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8931992053985596, 'rewards/meter/std': 0.15338252484798431, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8617154359817505, 'rewards/lexical_plausibility/std': 0.21377067267894745, 'rewards/judge_quality/mean': 0.39249998331069946, 'rewards/judge_quality/std': 0.3495609760284424, 'rewards/repeat_penalty/mean': 0.922041118144989, 'rewards/repeat_penalty/std': 0.21444439888000488, 'rewards/repeat_floor/mean': 0.9909240007400513, 'rewards/repeat_floor/std': 0.0247652567923069, 'rewards/near_duplicate_penalty/mean': 0.9820936918258667, 'rewards/near_duplicate_penalty/std': 0.04483920335769653, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9498327970504761, 'rewards/distinct_2/std': 0.14189434051513672, 'rewards/total_composite/mean': 0.35684460401535034, 'rewards/total_composite/std': 0.32566359639167786, 'reward': 0.35684460401535034, 'reward_std': 0.32566359639167786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1367790251970291, 'sampling/sampling_logp_difference/max': 0.940155029296875, 'sampling/importance_sampling_ratio/min': 0.39056727290153503, 'sampling/importance_sampling_ratio/mean': 1.0394681692123413, 'sampling/importance_sampling_ratio/max': 1.8137564659118652, 'entropy': 0.8101457133889198, 'clip_ratio/low_mean': 0.06992868706583977, 'clip_ratio/low_min': 0.06992868706583977, 'clip_ratio/high_mean': 0.03372093103826046, 'clip_ratio/high_max': 0.03372093103826046, 'clip_ratio/region_mean': 0.10364961810410023, 'reward_total_mean': 0.35684460401535034, 'reward_meter_mean': 0.8931992053985596, 'reward_meter_std': 0.15338252484798431, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8617154359817505, 'reward_lexical_plausibility_std': 0.21377067267894745, 'reward_repeat_penalty_mean': 0.922041118144989, 'reward_repeat_penalty_std': 0.21444439888000488, 'reward_repeat_floor_mean': 0.9909240007400513, 'reward_repeat_floor_std': 0.0247652567923069, 'reward_near_duplicate_penalty_mean': 0.9820936918258667, 'reward_near_duplicate_penalty_std': 0.04483920335769653, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9498327970504761, 'reward_distinct_2_std': 0.14189434051513672, 'reward_judge_quality_mean': 0.39249998331069946, 'reward_judge_quality_std': 0.3495609760284424, 'reward_total_composite_mean': 0.35684460401535034, 'reward_total_composite_std': 0.32566359639167786, 'epoch': 0.03}

 18%|█▊        | 607/3300 [2:14:25<12:57:37, 17.33s/it]INFO 04-12 18:12:51 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 608/3300 [2:14:33<10:52:50, 14.55s/it]2026-04-12 18:12:59,762 | INFO | train_grpo_train | metrics_logged mode=train step=608

                                                       
{'loss': 0.0396, 'grad_norm': 24.5108699798584, 'learning_rate': 8.16060606060606e-06, 'num_tokens': 1123720.0, 'completions/mean_length': 20.75, 'completions/min_length': 15.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7343016862869263, 'rewards/meter/std': 0.3321674168109894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9093136787414551, 'rewards/lexical_plausibility/std': 0.12747789919376373, 'rewards/judge_quality/mean': 0.7425000071525574, 'rewards/judge_quality/std': 0.36593323945999146, 'rewards/repeat_penalty/mean': 0.7312376499176025, 'rewards/repeat_penalty/std': 0.034962061792612076, 'rewards/repeat_floor/mean': 0.9812475442886353, 'rewards/repeat_floor/std': 0.0069924150593578815, 'rewards/near_duplicate_penalty/mean': 0.9749835729598999, 'rewards/near_duplicate_penalty/std': 0.046616096049547195, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5027918815612793, 'rewards/total_composite/std': 0.37425485253334045, 'reward': 0.5027918815612793, 'reward_std': 0.37425482273101807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1717093288898468, 'sampling/sampling_logp_difference/max': 1.0389184951782227, 'sampling/importance_sampling_ratio/min': 0.37201645970344543, 'sampling/importance_sampling_ratio/mean': 1.0298277139663696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.262733832001686, 'clip_ratio/low_mean': 0.11354167014360428, 'clip_ratio/low_min': 0.11354167014360428, 'clip_ratio/high_mean': 0.09123792918398976, 'clip_ratio/high_max': 0.09123792918398976, 'clip_ratio/region_mean': 0.20477959932759404, 'reward_total_mean': 0.5027918815612793, 'reward_meter_mean': 0.7343016862869263, 'reward_meter_std': 0.3321674168109894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9093136787414551, 'reward_lexical_plausibility_std': 0.12747789919376373, 'reward_repeat_penalty_mean': 0.7312376499176025, 'reward_repeat_penalty_std': 0.034962061792612076, 'reward_repeat_floor_mean': 0.9812475442886353, 'reward_repeat_floor_std': 0.0069924150593578815, 'reward_near_duplicate_penalty_mean': 0.9749835729598999, 'reward_near_duplicate_penalty_std': 0.046616096049547195, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7425000071525574, 'reward_judge_quality_std': 0.36593323945999146, 'reward_total_composite_mean': 0.5027918815612793, 'reward_total_composite_std': 0.37425485253334045, 'epoch': 0.03}

 18%|█▊        | 608/3300 [2:14:33<10:52:50, 14.55s/it]INFO 04-12 18:12:59 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 609/3300 [2:14:47<10:45:31, 14.39s/it]2026-04-12 18:13:13,793 | INFO | train_grpo_train | metrics_logged mode=train step=609

                                                       
{'loss': -0.0634, 'grad_norm': 1.8359025716781616, 'learning_rate': 8.15757575757576e-06, 'num_tokens': 1125237.0, 'completions/mean_length': 220.625, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 45.79999923706055, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5188621282577515, 'rewards/meter/std': 0.38950106501579285, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8139591217041016, 'rewards/lexical_plausibility/std': 0.25860437750816345, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3612182140350342, 'rewards/repeat_penalty/mean': 0.940185546875, 'rewards/repeat_penalty/std': 0.08511374145746231, 'rewards/repeat_floor/mean': 0.9955215454101562, 'rewards/repeat_floor/std': 0.005360755138099194, 'rewards/near_duplicate_penalty/mean': 0.9994557499885559, 'rewards/near_duplicate_penalty/std': 0.001539413700811565, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9719797372817993, 'rewards/distinct_2/std': 0.038789909332990646, 'rewards/total_composite/mean': 0.27569475769996643, 'rewards/total_composite/std': 0.29816851019859314, 'reward': 0.27569475769996643, 'reward_std': 0.29816851019859314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12758906185626984, 'sampling/sampling_logp_difference/max': 1.1533756256103516, 'sampling/importance_sampling_ratio/min': 0.3155696988105774, 'sampling/importance_sampling_ratio/mean': 1.0322290658950806, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5662923827767372, 'clip_ratio/low_mean': 0.005681818351149559, 'clip_ratio/low_min': 0.005681818351149559, 'clip_ratio/high_mean': 0.05644345190376043, 'clip_ratio/high_max': 0.05644345190376043, 'clip_ratio/region_mean': 0.06212527025490999, 'reward_total_mean': 0.27569475769996643, 'reward_meter_mean': 0.5188621282577515, 'reward_meter_std': 0.38950106501579285, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8139591217041016, 'reward_lexical_plausibility_std': 0.25860437750816345, 'reward_repeat_penalty_mean': 0.940185546875, 'reward_repeat_penalty_std': 0.08511374145746231, 'reward_repeat_floor_mean': 0.9955215454101562, 'reward_repeat_floor_std': 0.005360755138099194, 'reward_near_duplicate_penalty_mean': 0.9994557499885559, 'reward_near_duplicate_penalty_std': 0.001539413700811565, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9719797372817993, 'reward_distinct_2_std': 0.038789909332990646, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3612182140350342, 'reward_total_composite_mean': 0.27569475769996643, 'reward_total_composite_std': 0.29816851019859314, 'epoch': 0.03}

 18%|█▊        | 609/3300 [2:14:47<10:45:31, 14.39s/it]INFO 04-12 18:13:14 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 610/3300 [2:14:56<9:35:28, 12.84s/it] 2026-04-12 18:13:22,990 | INFO | train_grpo_train | metrics_logged mode=train step=610

                                                      
{'loss': 0.0595, 'grad_norm': 15.591653823852539, 'learning_rate': 8.154545454545455e-06, 'num_tokens': 1126775.0, 'completions/mean_length': 39.25, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.2954864203929901, 'rewards/meter/std': 0.3326808214187622, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.7220776081085205, 'rewards/repeat_penalty/std': 0.2202475517988205, 'rewards/repeat_floor/mean': 0.9659847021102905, 'rewards/repeat_floor/std': 0.026132816448807716, 'rewards/near_duplicate_penalty/mean': 0.8857450485229492, 'rewards/near_duplicate_penalty/std': 0.06554435938596725, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8749772310256958, 'rewards/distinct_2/std': 0.10805743932723999, 'rewards/total_composite/mean': 0.12010689824819565, 'rewards/total_composite/std': 0.1389647275209427, 'reward': 0.12010689824819565, 'reward_std': 0.1389647126197815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.168714702129364, 'sampling/sampling_logp_difference/max': 1.9937443733215332, 'sampling/importance_sampling_ratio/min': 0.21070608496665955, 'sampling/importance_sampling_ratio/mean': 1.0192890167236328, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9132148772478104, 'clip_ratio/low_mean': 0.09924495033919811, 'clip_ratio/low_min': 0.09924495033919811, 'clip_ratio/high_mean': 0.04909370839595795, 'clip_ratio/high_max': 0.04909370839595795, 'clip_ratio/region_mean': 0.14833865873515606, 'reward_total_mean': 0.12010689824819565, 'reward_meter_mean': 0.2954864203929901, 'reward_meter_std': 0.3326808214187622, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7220776081085205, 'reward_repeat_penalty_std': 0.2202475517988205, 'reward_repeat_floor_mean': 0.9659847021102905, 'reward_repeat_floor_std': 0.026132816448807716, 'reward_near_duplicate_penalty_mean': 0.8857450485229492, 'reward_near_duplicate_penalty_std': 0.06554435938596725, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8749772310256958, 'reward_distinct_2_std': 0.10805743932723999, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.12010689824819565, 'reward_total_composite_std': 0.1389647275209427, 'epoch': 0.03}

 18%|█▊        | 610/3300 [2:14:56<9:35:28, 12.84s/it]INFO 04-12 18:13:23 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 611/3300 [2:15:10<9:43:54, 13.03s/it]2026-04-12 18:13:36,472 | INFO | train_grpo_train | metrics_logged mode=train step=611

                                                      
{'loss': -0.0121, 'grad_norm': 5.054934501647949, 'learning_rate': 8.151515151515152e-06, 'num_tokens': 1128256.0, 'completions/mean_length': 85.125, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.142858505249023, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.46992820501327515, 'rewards/meter/std': 0.5011417865753174, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9052891135215759, 'rewards/lexical_plausibility/std': 0.15319178998470306, 'rewards/judge_quality/mean': 0.8112499713897705, 'rewards/judge_quality/std': 0.3075914680957794, 'rewards/repeat_penalty/mean': 0.765625, 'rewards/repeat_penalty/std': 0.1043153703212738, 'rewards/repeat_floor/mean': 0.9837499856948853, 'rewards/repeat_floor/std': 0.010938146151602268, 'rewards/near_duplicate_penalty/mean': 0.9791666269302368, 'rewards/near_duplicate_penalty/std': 0.0589255727827549, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4230913519859314, 'rewards/total_composite/std': 0.4512261748313904, 'reward': 0.4230913519859314, 'reward_std': 0.4512261748313904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20187486708164215, 'sampling/sampling_logp_difference/max': 2.3587756156921387, 'sampling/importance_sampling_ratio/min': 0.09453590214252472, 'sampling/importance_sampling_ratio/mean': 1.0098243951797485, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1807630136609077, 'clip_ratio/low_mean': 0.08234926871955395, 'clip_ratio/low_min': 0.08234926871955395, 'clip_ratio/high_mean': 0.12944189086556435, 'clip_ratio/high_max': 0.12944189086556435, 'clip_ratio/region_mean': 0.2117911595851183, 'reward_total_mean': 0.4230913519859314, 'reward_meter_mean': 0.46992820501327515, 'reward_meter_std': 0.5011417865753174, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9052891135215759, 'reward_lexical_plausibility_std': 0.15319178998470306, 'reward_repeat_penalty_mean': 0.765625, 'reward_repeat_penalty_std': 0.1043153703212738, 'reward_repeat_floor_mean': 0.9837499856948853, 'reward_repeat_floor_std': 0.010938146151602268, 'reward_near_duplicate_penalty_mean': 0.9791666269302368, 'reward_near_duplicate_penalty_std': 0.0589255727827549, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8112499713897705, 'reward_judge_quality_std': 0.3075914680957794, 'reward_total_composite_mean': 0.4230913519859314, 'reward_total_composite_std': 0.4512261748313904, 'epoch': 0.03}

 19%|█▊        | 611/3300 [2:15:10<9:43:54, 13.03s/it]INFO 04-12 18:13:36 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 612/3300 [2:15:24<9:57:16, 13.33s/it]2026-04-12 18:13:50,513 | INFO | train_grpo_train | metrics_logged mode=train step=612

                                                      
{'loss': -0.0659, 'grad_norm': 5.352459907531738, 'learning_rate': 8.14848484848485e-06, 'num_tokens': 1129831.0, 'completions/mean_length': 102.875, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.42857360839844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6945694088935852, 'rewards/meter/std': 0.40696218609809875, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9336939454078674, 'rewards/lexical_plausibility/std': 0.1490423083305359, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.3358757197856903, 'rewards/repeat_penalty/mean': 0.9519752264022827, 'rewards/repeat_penalty/std': 0.08429176360368729, 'rewards/repeat_floor/mean': 0.9956088066101074, 'rewards/repeat_floor/std': 0.005327679682523012, 'rewards/near_duplicate_penalty/mean': 0.9832252264022827, 'rewards/near_duplicate_penalty/std': 0.022154025733470917, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.41193124651908875, 'rewards/total_composite/std': 0.3424032926559448, 'reward': 0.41193124651908875, 'reward_std': 0.34240326285362244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17750002443790436, 'sampling/sampling_logp_difference/max': 2.2086310386657715, 'sampling/importance_sampling_ratio/min': 0.1098509281873703, 'sampling/importance_sampling_ratio/mean': 1.017673134803772, 'sampling/importance_sampling_ratio/max': 1.9343492984771729, 'entropy': 1.1710313633084297, 'clip_ratio/low_mean': 0.0570170315913856, 'clip_ratio/low_min': 0.0570170315913856, 'clip_ratio/high_mean': 0.04812834318727255, 'clip_ratio/high_max': 0.04812834318727255, 'clip_ratio/region_mean': 0.10514537477865815, 'reward_total_mean': 0.41193124651908875, 'reward_meter_mean': 0.6945694088935852, 'reward_meter_std': 0.40696218609809875, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9336939454078674, 'reward_lexical_plausibility_std': 0.1490423083305359, 'reward_repeat_penalty_mean': 0.9519752264022827, 'reward_repeat_penalty_std': 0.08429176360368729, 'reward_repeat_floor_mean': 0.9956088066101074, 'reward_repeat_floor_std': 0.005327679682523012, 'reward_near_duplicate_penalty_mean': 0.9832252264022827, 'reward_near_duplicate_penalty_std': 0.022154025733470917, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.3358757197856903, 'reward_total_composite_mean': 0.41193124651908875, 'reward_total_composite_std': 0.3424032926559448, 'epoch': 0.03}

 19%|█▊        | 612/3300 [2:15:24<9:57:16, 13.33s/it]INFO 04-12 18:13:50 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 613/3300 [2:15:32<8:43:37, 11.69s/it]2026-04-12 18:13:58,377 | INFO | train_grpo_train | metrics_logged mode=train step=613

                                                      
{'loss': -0.0477, 'grad_norm': 18.587955474853516, 'learning_rate': 8.145454545454547e-06, 'num_tokens': 1131197.0, 'completions/mean_length': 25.75, 'completions/min_length': 17.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9855853319168091, 'rewards/meter/std': 0.018046841025352478, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6646794080734253, 'rewards/total_composite/std': 0.2433253824710846, 'reward': 0.6646794080734253, 'reward_std': 0.2433253824710846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1340295821428299, 'sampling/sampling_logp_difference/max': 1.0169172286987305, 'sampling/importance_sampling_ratio/min': 0.36170831322669983, 'sampling/importance_sampling_ratio/mean': 1.021571159362793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0363589599728584, 'clip_ratio/low_mean': 0.04903846140950918, 'clip_ratio/low_min': 0.04903846140950918, 'clip_ratio/high_mean': 0.08721532206982374, 'clip_ratio/high_max': 0.08721532206982374, 'clip_ratio/region_mean': 0.13625378347933292, 'reward_total_mean': 0.6646794080734253, 'reward_meter_mean': 0.9855853319168091, 'reward_meter_std': 0.018046841025352478, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.6646794080734253, 'reward_total_composite_std': 0.2433253824710846, 'epoch': 0.03}

 19%|█▊        | 613/3300 [2:15:32<8:43:37, 11.69s/it]INFO 04-12 18:13:58 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 614/3300 [2:15:45<9:08:17, 12.25s/it]2026-04-12 18:14:11,921 | INFO | train_grpo_train | metrics_logged mode=train step=614

                                                      
{'loss': 0.1047, 'grad_norm': 3.284402370452881, 'learning_rate': 8.142424242424242e-06, 'num_tokens': 1133087.0, 'completions/mean_length': 208.25, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 107.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 423.0, 'rewards/meter/mean': 0.23793752491474152, 'rewards/meter/std': 0.3073529601097107, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.11338934302330017, 'rewards/lexical_plausibility/mean': 0.778860330581665, 'rewards/lexical_plausibility/std': 0.30687153339385986, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.38975268602371216, 'rewards/repeat_penalty/mean': 0.9131630063056946, 'rewards/repeat_penalty/std': 0.10829947143793106, 'rewards/repeat_floor/mean': 0.9930588603019714, 'rewards/repeat_floor/std': 0.0069914464838802814, 'rewards/near_duplicate_penalty/mean': 0.9839702844619751, 'rewards/near_duplicate_penalty/std': 0.0249447263777256, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.10655052959918976, 'rewards/total_composite/std': 0.1975708156824112, 'reward': 0.10655052959918976, 'reward_std': 0.19757080078125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14779411256313324, 'sampling/sampling_logp_difference/max': 1.413313865661621, 'sampling/importance_sampling_ratio/min': 0.24333555996418, 'sampling/importance_sampling_ratio/mean': 1.0163108110427856, 'sampling/importance_sampling_ratio/max': 1.90984308719635, 'entropy': 1.836133986711502, 'clip_ratio/low_mean': 0.07186457002535462, 'clip_ratio/low_min': 0.07186457002535462, 'clip_ratio/high_mean': 0.054079560562968254, 'clip_ratio/high_max': 0.054079560562968254, 'clip_ratio/region_mean': 0.12594413058832288, 'reward_total_mean': 0.10655052959918976, 'reward_meter_mean': 0.23793752491474152, 'reward_meter_std': 0.3073529601097107, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.11338934302330017, 'reward_lexical_plausibility_mean': 0.778860330581665, 'reward_lexical_plausibility_std': 0.30687153339385986, 'reward_repeat_penalty_mean': 0.9131630063056946, 'reward_repeat_penalty_std': 0.10829947143793106, 'reward_repeat_floor_mean': 0.9930588603019714, 'reward_repeat_floor_std': 0.0069914464838802814, 'reward_near_duplicate_penalty_mean': 0.9839702844619751, 'reward_near_duplicate_penalty_std': 0.0249447263777256, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.38975268602371216, 'reward_total_composite_mean': 0.10655052959918976, 'reward_total_composite_std': 0.1975708156824112, 'epoch': 0.03}

 19%|█▊        | 614/3300 [2:15:45<9:08:17, 12.25s/it]INFO 04-12 18:14:12 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 615/3300 [2:15:59<9:34:35, 12.84s/it]2026-04-12 18:14:26,140 | INFO | train_grpo_train | metrics_logged mode=train step=615

                                                      
{'loss': -0.0769, 'grad_norm': 3.3002634048461914, 'learning_rate': 8.139393939393941e-06, 'num_tokens': 1134648.0, 'completions/mean_length': 162.125, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7952566742897034, 'rewards/meter/std': 0.30946919322013855, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8860474824905396, 'rewards/lexical_plausibility/std': 0.21209657192230225, 'rewards/judge_quality/mean': 0.3787500262260437, 'rewards/judge_quality/std': 0.2630283832550049, 'rewards/repeat_penalty/mean': 0.9653246402740479, 'rewards/repeat_penalty/std': 0.039832696318626404, 'rewards/repeat_floor/mean': 0.9947987198829651, 'rewards/repeat_floor/std': 0.0059749106876552105, 'rewards/near_duplicate_penalty/mean': 0.9653246402740479, 'rewards/near_duplicate_penalty/std': 0.039832696318626404, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3191142678260803, 'rewards/total_composite/std': 0.22737827897071838, 'reward': 0.3191142678260803, 'reward_std': 0.22737827897071838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17958132922649384, 'sampling/sampling_logp_difference/max': 2.4698095321655273, 'sampling/importance_sampling_ratio/min': 0.08460097014904022, 'sampling/importance_sampling_ratio/mean': 1.0286587476730347, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9334373250603676, 'clip_ratio/low_mean': 0.03500000014901161, 'clip_ratio/low_min': 0.03500000014901161, 'clip_ratio/high_mean': 0.11670032609254122, 'clip_ratio/high_max': 0.11670032609254122, 'clip_ratio/region_mean': 0.15170032624155283, 'reward_total_mean': 0.3191142678260803, 'reward_meter_mean': 0.7952566742897034, 'reward_meter_std': 0.30946919322013855, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8860474824905396, 'reward_lexical_plausibility_std': 0.21209657192230225, 'reward_repeat_penalty_mean': 0.9653246402740479, 'reward_repeat_penalty_std': 0.039832696318626404, 'reward_repeat_floor_mean': 0.9947987198829651, 'reward_repeat_floor_std': 0.0059749106876552105, 'reward_near_duplicate_penalty_mean': 0.9653246402740479, 'reward_near_duplicate_penalty_std': 0.039832696318626404, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3787500262260437, 'reward_judge_quality_std': 0.2630283832550049, 'reward_total_composite_mean': 0.3191142678260803, 'reward_total_composite_std': 0.22737827897071838, 'epoch': 0.03}

 19%|█▊        | 615/3300 [2:15:59<9:34:35, 12.84s/it]INFO 04-12 18:14:26 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 616/3300 [2:16:08<8:35:34, 11.53s/it]2026-04-12 18:14:34,601 | INFO | train_grpo_train | metrics_logged mode=train step=616

                                                      
{'loss': 0.06, 'grad_norm': 22.035568237304688, 'learning_rate': 8.136363636363637e-06, 'num_tokens': 1136371.0, 'completions/mean_length': 52.375, 'completions/min_length': 36.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.5446590185165405, 'rewards/meter/std': 0.32828229665756226, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5787500143051147, 'rewards/judge_quality/std': 0.18216457962989807, 'rewards/repeat_penalty/mean': 0.9390758872032166, 'rewards/repeat_penalty/std': 0.043568190187215805, 'rewards/repeat_floor/mean': 0.9933968186378479, 'rewards/repeat_floor/std': 0.004565213806927204, 'rewards/near_duplicate_penalty/mean': 0.982736349105835, 'rewards/near_duplicate_penalty/std': 0.013402133248746395, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9554044008255005, 'rewards/distinct_2/std': 0.03800548240542412, 'rewards/total_composite/mean': 0.3447574973106384, 'rewards/total_composite/std': 0.2765711545944214, 'reward': 0.3447574973106384, 'reward_std': 0.2765711545944214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19323676824569702, 'sampling/sampling_logp_difference/max': 2.2633309364318848, 'sampling/importance_sampling_ratio/min': 0.10400348156690598, 'sampling/importance_sampling_ratio/mean': 0.9937203526496887, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8703051880002022, 'clip_ratio/low_mean': 0.0836385521106422, 'clip_ratio/low_min': 0.0836385521106422, 'clip_ratio/high_mean': 0.05730654764920473, 'clip_ratio/high_max': 0.05730654764920473, 'clip_ratio/region_mean': 0.14094509975984693, 'reward_total_mean': 0.3447574973106384, 'reward_meter_mean': 0.5446590185165405, 'reward_meter_std': 0.32828229665756226, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9390758872032166, 'reward_repeat_penalty_std': 0.043568190187215805, 'reward_repeat_floor_mean': 0.9933968186378479, 'reward_repeat_floor_std': 0.004565213806927204, 'reward_near_duplicate_penalty_mean': 0.982736349105835, 'reward_near_duplicate_penalty_std': 0.013402133248746395, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9554044008255005, 'reward_distinct_2_std': 0.03800548240542412, 'reward_judge_quality_mean': 0.5787500143051147, 'reward_judge_quality_std': 0.18216457962989807, 'reward_total_composite_mean': 0.3447574973106384, 'reward_total_composite_std': 0.2765711545944214, 'epoch': 0.03}

 19%|█▊        | 616/3300 [2:16:08<8:35:34, 11.53s/it]INFO 04-12 18:14:34 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 617/3300 [2:16:22<9:03:33, 12.16s/it]2026-04-12 18:14:48,238 | INFO | train_grpo_train | metrics_logged mode=train step=617

                                                      
{'loss': -0.0628, 'grad_norm': 6.196832180023193, 'learning_rate': 8.133333333333334e-06, 'num_tokens': 1138060.0, 'completions/mean_length': 108.125, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.42857360839844, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.885595440864563, 'rewards/meter/std': 0.19156396389007568, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.940804660320282, 'rewards/lexical_plausibility/std': 0.14384816586971283, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.32275325059890747, 'rewards/repeat_penalty/mean': 0.8910260200500488, 'rewards/repeat_penalty/std': 0.13816334307193756, 'rewards/repeat_floor/mean': 0.9906647205352783, 'rewards/repeat_floor/std': 0.011830944567918777, 'rewards/near_duplicate_penalty/mean': 0.9871708154678345, 'rewards/near_duplicate_penalty/std': 0.02127412147819996, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9593233466148376, 'rewards/distinct_2/std': 0.046888317912817, 'rewards/total_composite/mean': 0.4802056550979614, 'rewards/total_composite/std': 0.304761677980423, 'reward': 0.4802056550979614, 'reward_std': 0.304761677980423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.175881490111351, 'sampling/sampling_logp_difference/max': 1.3144702911376953, 'sampling/importance_sampling_ratio/min': 0.26861658692359924, 'sampling/importance_sampling_ratio/mean': 0.9870852828025818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9020218998193741, 'clip_ratio/low_mean': 0.04474160494282842, 'clip_ratio/low_min': 0.04474160494282842, 'clip_ratio/high_mean': 0.06013353727757931, 'clip_ratio/high_max': 0.06013353727757931, 'clip_ratio/region_mean': 0.10487514222040772, 'reward_total_mean': 0.4802056550979614, 'reward_meter_mean': 0.885595440864563, 'reward_meter_std': 0.19156396389007568, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.940804660320282, 'reward_lexical_plausibility_std': 0.14384816586971283, 'reward_repeat_penalty_mean': 0.8910260200500488, 'reward_repeat_penalty_std': 0.13816334307193756, 'reward_repeat_floor_mean': 0.9906647205352783, 'reward_repeat_floor_std': 0.011830944567918777, 'reward_near_duplicate_penalty_mean': 0.9871708154678345, 'reward_near_duplicate_penalty_std': 0.02127412147819996, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9593233466148376, 'reward_distinct_2_std': 0.046888317912817, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.32275325059890747, 'reward_total_composite_mean': 0.4802056550979614, 'reward_total_composite_std': 0.304761677980423, 'epoch': 0.03}

 19%|█▊        | 617/3300 [2:16:22<9:03:33, 12.16s/it]INFO 04-12 18:14:48 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 618/3300 [2:16:35<9:19:43, 12.52s/it]2026-04-12 18:15:01,606 | INFO | train_grpo_train | metrics_logged mode=train step=618

                                                      
{'loss': -0.0683, 'grad_norm': 2.9688713550567627, 'learning_rate': 8.130303030303031e-06, 'num_tokens': 1139886.0, 'completions/mean_length': 228.25, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.597636878490448, 'rewards/meter/std': 0.42831969261169434, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8042219877243042, 'rewards/lexical_plausibility/std': 0.28561586141586304, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2881684899330139, 'rewards/repeat_penalty/mean': 0.9984383583068848, 'rewards/repeat_penalty/std': 0.004417074378579855, 'rewards/repeat_floor/mean': 0.9997657537460327, 'rewards/repeat_floor/std': 0.0006625695386901498, 'rewards/near_duplicate_penalty/mean': 0.9984383583068848, 'rewards/near_duplicate_penalty/std': 0.004417074378579855, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2403298169374466, 'rewards/total_composite/std': 0.2993147671222687, 'reward': 0.2403298169374466, 'reward_std': 0.2993147671222687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18364804983139038, 'sampling/sampling_logp_difference/max': 1.8630266189575195, 'sampling/importance_sampling_ratio/min': 0.15520218014717102, 'sampling/importance_sampling_ratio/mean': 1.0117346048355103, 'sampling/importance_sampling_ratio/max': 1.911372184753418, 'entropy': 0.9035596251487732, 'clip_ratio/low_mean': 0.04995519667863846, 'clip_ratio/low_min': 0.04995519667863846, 'clip_ratio/high_mean': 0.07724963873624802, 'clip_ratio/high_max': 0.07724963873624802, 'clip_ratio/region_mean': 0.12720483541488647, 'reward_total_mean': 0.2403298169374466, 'reward_meter_mean': 0.597636878490448, 'reward_meter_std': 0.42831969261169434, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8042219877243042, 'reward_lexical_plausibility_std': 0.28561586141586304, 'reward_repeat_penalty_mean': 0.9984383583068848, 'reward_repeat_penalty_std': 0.004417074378579855, 'reward_repeat_floor_mean': 0.9997657537460327, 'reward_repeat_floor_std': 0.0006625695386901498, 'reward_near_duplicate_penalty_mean': 0.9984383583068848, 'reward_near_duplicate_penalty_std': 0.004417074378579855, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2881684899330139, 'reward_total_composite_mean': 0.2403298169374466, 'reward_total_composite_std': 0.2993147671222687, 'epoch': 0.03}

 19%|█▊        | 618/3300 [2:16:35<9:19:43, 12.52s/it]INFO 04-12 18:15:01 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 619/3300 [2:16:44<8:36:51, 11.57s/it]2026-04-12 18:15:10,942 | INFO | train_grpo_train | metrics_logged mode=train step=619

                                                      
{'loss': 0.0874, 'grad_norm': 14.600446701049805, 'learning_rate': 8.127272727272728e-06, 'num_tokens': 1141776.0, 'completions/mean_length': 64.25, 'completions/min_length': 48.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.7336498498916626, 'rewards/meter/std': 0.3361385464668274, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9711818695068359, 'rewards/repeat_penalty/std': 0.05532743036746979, 'rewards/repeat_floor/mean': 0.9966529607772827, 'rewards/repeat_floor/std': 0.005594935268163681, 'rewards/near_duplicate_penalty/mean': 0.9881757497787476, 'rewards/near_duplicate_penalty/std': 0.011149524711072445, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.241560697555542, 'rewards/total_composite/std': 0.13178382813930511, 'reward': 0.241560697555542, 'reward_std': 0.1317838430404663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1809169054031372, 'sampling/sampling_logp_difference/max': 2.6580488681793213, 'sampling/importance_sampling_ratio/min': 0.0700848326086998, 'sampling/importance_sampling_ratio/mean': 1.0107845067977905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0946395546197891, 'clip_ratio/low_mean': 0.06394055904820561, 'clip_ratio/low_min': 0.06394055904820561, 'clip_ratio/high_mean': 0.06823549419641495, 'clip_ratio/high_max': 0.06823549419641495, 'clip_ratio/region_mean': 0.13217605324462056, 'reward_total_mean': 0.241560697555542, 'reward_meter_mean': 0.7336498498916626, 'reward_meter_std': 0.3361385464668274, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9711818695068359, 'reward_repeat_penalty_std': 0.05532743036746979, 'reward_repeat_floor_mean': 0.9966529607772827, 'reward_repeat_floor_std': 0.005594935268163681, 'reward_near_duplicate_penalty_mean': 0.9881757497787476, 'reward_near_duplicate_penalty_std': 0.011149524711072445, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.241560697555542, 'reward_total_composite_std': 0.13178382813930511, 'epoch': 0.03}

 19%|█▉        | 619/3300 [2:16:44<8:36:51, 11.57s/it]INFO 04-12 18:15:11 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 620/3300 [2:16:58<9:03:09, 12.16s/it]2026-04-12 18:15:24,488 | INFO | train_grpo_train | metrics_logged mode=train step=620

                                                      
{'loss': -0.0402, 'grad_norm': 5.1841020584106445, 'learning_rate': 8.124242424242424e-06, 'num_tokens': 1143734.0, 'completions/mean_length': 130.75, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 76.28572082519531, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7234073281288147, 'rewards/meter/std': 0.2889758050441742, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.053032997995615005, 'rewards/lexical_plausibility/mean': 0.9526209831237793, 'rewards/lexical_plausibility/std': 0.13400810956954956, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.7898324728012085, 'rewards/repeat_penalty/std': 0.22097037732601166, 'rewards/repeat_floor/mean': 0.9797949194908142, 'rewards/repeat_floor/std': 0.019347619265317917, 'rewards/near_duplicate_penalty/mean': 0.9350776672363281, 'rewards/near_duplicate_penalty/std': 0.045264650136232376, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.8941197991371155, 'rewards/distinct_2/std': 0.1336977779865265, 'rewards/total_composite/mean': 0.28049683570861816, 'rewards/total_composite/std': 0.19081680476665497, 'reward': 0.28049683570861816, 'reward_std': 0.19081680476665497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13807980716228485, 'sampling/sampling_logp_difference/max': 1.1613225936889648, 'sampling/importance_sampling_ratio/min': 0.3130718469619751, 'sampling/importance_sampling_ratio/mean': 1.0102351903915405, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.850733757019043, 'clip_ratio/low_mean': 0.0496748685836792, 'clip_ratio/low_min': 0.0496748685836792, 'clip_ratio/high_mean': 0.08766865264624357, 'clip_ratio/high_max': 0.08766865264624357, 'clip_ratio/region_mean': 0.13734352122992277, 'reward_total_mean': 0.28049683570861816, 'reward_meter_mean': 0.7234073281288147, 'reward_meter_std': 0.2889758050441742, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.053032997995615005, 'reward_lexical_plausibility_mean': 0.9526209831237793, 'reward_lexical_plausibility_std': 0.13400810956954956, 'reward_repeat_penalty_mean': 0.7898324728012085, 'reward_repeat_penalty_std': 0.22097037732601166, 'reward_repeat_floor_mean': 0.9797949194908142, 'reward_repeat_floor_std': 0.019347619265317917, 'reward_near_duplicate_penalty_mean': 0.9350776672363281, 'reward_near_duplicate_penalty_std': 0.045264650136232376, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.8941197991371155, 'reward_distinct_2_std': 0.1336977779865265, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.28049683570861816, 'reward_total_composite_std': 0.19081680476665497, 'epoch': 0.03}

 19%|█▉        | 620/3300 [2:16:58<9:03:09, 12.16s/it]INFO 04-12 18:15:24 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 621/3300 [2:17:13<9:38:14, 12.95s/it]2026-04-12 18:15:39,282 | INFO | train_grpo_train | metrics_logged mode=train step=621

                                                      
{'loss': -0.1467, 'grad_norm': 2.7168047428131104, 'learning_rate': 8.121212121212121e-06, 'num_tokens': 1146001.0, 'completions/mean_length': 183.375, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 73.83333587646484, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9033820629119873, 'rewards/meter/std': 0.22243911027908325, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8878495693206787, 'rewards/lexical_plausibility/std': 0.20941151678562164, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.8891706466674805, 'rewards/repeat_penalty/std': 0.06366286426782608, 'rewards/repeat_floor/mean': 0.9874754548072815, 'rewards/repeat_floor/std': 0.006605308502912521, 'rewards/near_duplicate_penalty/mean': 0.9622802734375, 'rewards/near_duplicate_penalty/std': 0.017249030992388725, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9237045049667358, 'rewards/distinct_2/std': 0.0581965409219265, 'rewards/total_composite/mean': 0.29970651865005493, 'rewards/total_composite/std': 0.14850907027721405, 'reward': 0.29970651865005493, 'reward_std': 0.14850907027721405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16845576465129852, 'sampling/sampling_logp_difference/max': 1.6724557876586914, 'sampling/importance_sampling_ratio/min': 0.187785342335701, 'sampling/importance_sampling_ratio/mean': 0.9902026653289795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6718973368406296, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.1276545450091362, 'clip_ratio/high_max': 0.1276545450091362, 'clip_ratio/region_mean': 0.1276545450091362, 'reward_total_mean': 0.29970651865005493, 'reward_meter_mean': 0.9033820629119873, 'reward_meter_std': 0.22243911027908325, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8878495693206787, 'reward_lexical_plausibility_std': 0.20941151678562164, 'reward_repeat_penalty_mean': 0.8891706466674805, 'reward_repeat_penalty_std': 0.06366286426782608, 'reward_repeat_floor_mean': 0.9874754548072815, 'reward_repeat_floor_std': 0.006605308502912521, 'reward_near_duplicate_penalty_mean': 0.9622802734375, 'reward_near_duplicate_penalty_std': 0.017249030992388725, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9237045049667358, 'reward_distinct_2_std': 0.0581965409219265, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.29970651865005493, 'reward_total_composite_std': 0.14850907027721405, 'epoch': 0.03}

 19%|█▉        | 621/3300 [2:17:13<9:38:14, 12.95s/it]INFO 04-12 18:15:39 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 622/3300 [2:17:21<8:34:23, 11.52s/it]2026-04-12 18:15:47,479 | INFO | train_grpo_train | metrics_logged mode=train step=622

                                                      
{'loss': 0.0119, 'grad_norm': 13.615230560302734, 'learning_rate': 8.118181818181819e-06, 'num_tokens': 1147420.0, 'completions/mean_length': 51.375, 'completions/min_length': 41.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.375, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6384586095809937, 'rewards/meter/std': 0.47671863436698914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5375000238418579, 'rewards/judge_quality/std': 0.2520062327384949, 'rewards/repeat_penalty/mean': 0.7313476800918579, 'rewards/repeat_penalty/std': 0.29345589876174927, 'rewards/repeat_floor/mean': 0.9674400091171265, 'rewards/repeat_floor/std': 0.0373542420566082, 'rewards/near_duplicate_penalty/mean': 0.9219187498092651, 'rewards/near_duplicate_penalty/std': 0.09316504001617432, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8308577537536621, 'rewards/distinct_2/std': 0.1842360943555832, 'rewards/total_composite/mean': 0.3183947205543518, 'rewards/total_composite/std': 0.31174230575561523, 'reward': 0.3183947205543518, 'reward_std': 0.3117423355579376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17254482209682465, 'sampling/sampling_logp_difference/max': 1.3535106182098389, 'sampling/importance_sampling_ratio/min': 0.2583317458629608, 'sampling/importance_sampling_ratio/mean': 1.0363050699234009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.272480845451355, 'clip_ratio/low_mean': 0.06573247816413641, 'clip_ratio/low_min': 0.06573247816413641, 'clip_ratio/high_mean': 0.08048379793763161, 'clip_ratio/high_max': 0.08048379793763161, 'clip_ratio/region_mean': 0.14621627610176802, 'reward_total_mean': 0.3183947205543518, 'reward_meter_mean': 0.6384586095809937, 'reward_meter_std': 0.47671863436698914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7313476800918579, 'reward_repeat_penalty_std': 0.29345589876174927, 'reward_repeat_floor_mean': 0.9674400091171265, 'reward_repeat_floor_std': 0.0373542420566082, 'reward_near_duplicate_penalty_mean': 0.9219187498092651, 'reward_near_duplicate_penalty_std': 0.09316504001617432, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8308577537536621, 'reward_distinct_2_std': 0.1842360943555832, 'reward_judge_quality_mean': 0.5375000238418579, 'reward_judge_quality_std': 0.2520062327384949, 'reward_total_composite_mean': 0.3183947205543518, 'reward_total_composite_std': 0.31174230575561523, 'epoch': 0.03}

 19%|█▉        | 622/3300 [2:17:21<8:34:23, 11.52s/it]INFO 04-12 18:15:47 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 623/3300 [2:17:34<8:58:40, 12.07s/it]2026-04-12 18:16:00,831 | INFO | train_grpo_train | metrics_logged mode=train step=623

                                                      
{'loss': -0.0487, 'grad_norm': 2.5123581886291504, 'learning_rate': 8.115151515151516e-06, 'num_tokens': 1148808.0, 'completions/mean_length': 153.5, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.6920024156570435, 'rewards/meter/std': 0.39770135283470154, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8688567876815796, 'rewards/lexical_plausibility/std': 0.2448626011610031, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.36570626497268677, 'rewards/repeat_penalty/mean': 0.9320791959762573, 'rewards/repeat_penalty/std': 0.061871372163295746, 'rewards/repeat_floor/mean': 0.9929609894752502, 'rewards/repeat_floor/std': 0.0061057982966303825, 'rewards/near_duplicate_penalty/mean': 0.9861485958099365, 'rewards/near_duplicate_penalty/std': 0.010002936236560345, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.94487464427948, 'rewards/distinct_2/std': 0.05666302144527435, 'rewards/total_composite/mean': 0.39848634600639343, 'rewards/total_composite/std': 0.3466055989265442, 'reward': 0.39848634600639343, 'reward_std': 0.3466055989265442, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1686234325170517, 'sampling/sampling_logp_difference/max': 2.383777141571045, 'sampling/importance_sampling_ratio/min': 0.09220165759325027, 'sampling/importance_sampling_ratio/mean': 1.0037939548492432, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5634059682488441, 'clip_ratio/low_mean': 0.02142857201397419, 'clip_ratio/low_min': 0.02142857201397419, 'clip_ratio/high_mean': 0.0881561990827322, 'clip_ratio/high_max': 0.0881561990827322, 'clip_ratio/region_mean': 0.10958477109670639, 'reward_total_mean': 0.39848634600639343, 'reward_meter_mean': 0.6920024156570435, 'reward_meter_std': 0.39770135283470154, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8688567876815796, 'reward_lexical_plausibility_std': 0.2448626011610031, 'reward_repeat_penalty_mean': 0.9320791959762573, 'reward_repeat_penalty_std': 0.061871372163295746, 'reward_repeat_floor_mean': 0.9929609894752502, 'reward_repeat_floor_std': 0.0061057982966303825, 'reward_near_duplicate_penalty_mean': 0.9861485958099365, 'reward_near_duplicate_penalty_std': 0.010002936236560345, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.94487464427948, 'reward_distinct_2_std': 0.05666302144527435, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.36570626497268677, 'reward_total_composite_mean': 0.39848634600639343, 'reward_total_composite_std': 0.3466055989265442, 'epoch': 0.03}

 19%|█▉        | 623/3300 [2:17:34<8:58:40, 12.07s/it]INFO 04-12 18:16:01 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 624/3300 [2:17:48<9:26:44, 12.71s/it]2026-04-12 18:16:15,018 | INFO | train_grpo_train | metrics_logged mode=train step=624

                                                      
{'loss': -0.1316, 'grad_norm': 2.875786781311035, 'learning_rate': 8.112121212121213e-06, 'num_tokens': 1150521.0, 'completions/mean_length': 173.125, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 60.16666793823242, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.6622650027275085, 'rewards/meter/std': 0.39949294924736023, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.07559289038181305, 'rewards/lexical_plausibility/mean': 0.8699882626533508, 'rewards/lexical_plausibility/std': 0.24216406047344208, 'rewards/judge_quality/mean': 0.4037500023841858, 'rewards/judge_quality/std': 0.26397714018821716, 'rewards/repeat_penalty/mean': 0.9251468181610107, 'rewards/repeat_penalty/std': 0.10896413773298264, 'rewards/repeat_floor/mean': 0.9943970441818237, 'rewards/repeat_floor/std': 0.006151876877993345, 'rewards/near_duplicate_penalty/mean': 0.9876468181610107, 'rewards/near_duplicate_penalty/std': 0.015654806047677994, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3148782551288605, 'rewards/total_composite/std': 0.2211579531431198, 'reward': 0.3148782551288605, 'reward_std': 0.22115793824195862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1675977110862732, 'sampling/sampling_logp_difference/max': 1.4057464599609375, 'sampling/importance_sampling_ratio/min': 0.24518397450447083, 'sampling/importance_sampling_ratio/mean': 1.0625616312026978, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8187977820634842, 'clip_ratio/low_mean': 0.04284227546304464, 'clip_ratio/low_min': 0.04284227546304464, 'clip_ratio/high_mean': 0.06946451030671597, 'clip_ratio/high_max': 0.06946451030671597, 'clip_ratio/region_mean': 0.11230678576976061, 'reward_total_mean': 0.3148782551288605, 'reward_meter_mean': 0.6622650027275085, 'reward_meter_std': 0.39949294924736023, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.07559289038181305, 'reward_lexical_plausibility_mean': 0.8699882626533508, 'reward_lexical_plausibility_std': 0.24216406047344208, 'reward_repeat_penalty_mean': 0.9251468181610107, 'reward_repeat_penalty_std': 0.10896413773298264, 'reward_repeat_floor_mean': 0.9943970441818237, 'reward_repeat_floor_std': 0.006151876877993345, 'reward_near_duplicate_penalty_mean': 0.9876468181610107, 'reward_near_duplicate_penalty_std': 0.015654806047677994, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4037500023841858, 'reward_judge_quality_std': 0.26397714018821716, 'reward_total_composite_mean': 0.3148782551288605, 'reward_total_composite_std': 0.2211579531431198, 'epoch': 0.03}

 19%|█▉        | 624/3300 [2:17:48<9:26:44, 12.71s/it]INFO 04-12 18:16:15 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 625/3300 [2:17:57<8:26:52, 11.37s/it]2026-04-12 18:16:23,266 | INFO | train_grpo_train | metrics_logged mode=train step=625

                                                      
{'loss': 0.0678, 'grad_norm': 16.772842407226562, 'learning_rate': 8.10909090909091e-06, 'num_tokens': 1152153.0, 'completions/mean_length': 44.0, 'completions/min_length': 33.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7899090051651001, 'rewards/meter/std': 0.25864115357398987, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.23259025812149048, 'rewards/repeat_penalty/mean': 0.9366745352745056, 'rewards/repeat_penalty/std': 0.14090968668460846, 'rewards/repeat_floor/mean': 0.9933446049690247, 'rewards/repeat_floor/std': 0.01317894458770752, 'rewards/near_duplicate_penalty/mean': 0.9802762866020203, 'rewards/near_duplicate_penalty/std': 0.023176228627562523, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.05725561082363129, 'rewards/total_composite/mean': 0.5422884225845337, 'rewards/total_composite/std': 0.31825241446495056, 'reward': 0.5422884225845337, 'reward_std': 0.3182523846626282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14097292721271515, 'sampling/sampling_logp_difference/max': 1.8590154647827148, 'sampling/importance_sampling_ratio/min': 0.15582597255706787, 'sampling/importance_sampling_ratio/mean': 1.0054686069488525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9811501502990723, 'clip_ratio/low_mean': 0.08203381858766079, 'clip_ratio/low_min': 0.08203381858766079, 'clip_ratio/high_mean': 0.048706404864788055, 'clip_ratio/high_max': 0.048706404864788055, 'clip_ratio/region_mean': 0.13074022345244884, 'reward_total_mean': 0.5422884225845337, 'reward_meter_mean': 0.7899090051651001, 'reward_meter_std': 0.25864115357398987, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9366745352745056, 'reward_repeat_penalty_std': 0.14090968668460846, 'reward_repeat_floor_mean': 0.9933446049690247, 'reward_repeat_floor_std': 0.01317894458770752, 'reward_near_duplicate_penalty_mean': 0.9802762866020203, 'reward_near_duplicate_penalty_std': 0.023176228627562523, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.05725561082363129, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.23259025812149048, 'reward_total_composite_mean': 0.5422884225845337, 'reward_total_composite_std': 0.31825241446495056, 'epoch': 0.03}

 19%|█▉        | 625/3300 [2:17:57<8:26:52, 11.37s/it]INFO 04-12 18:16:23 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 626/3300 [2:18:10<8:59:38, 12.11s/it]2026-04-12 18:16:37,098 | INFO | train_grpo_train | metrics_logged mode=train step=626

                                                      
{'loss': -0.0771, 'grad_norm': 2.380643844604492, 'learning_rate': 8.106060606060606e-06, 'num_tokens': 1153718.0, 'completions/mean_length': 158.625, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 40.833335876464844, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6858423948287964, 'rewards/meter/std': 0.3974907696247101, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8820273876190186, 'rewards/lexical_plausibility/std': 0.21894796192646027, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.30420562624931335, 'rewards/repeat_penalty/mean': 0.9202359318733215, 'rewards/repeat_penalty/std': 0.10980397462844849, 'rewards/repeat_floor/mean': 0.9942066669464111, 'rewards/repeat_floor/std': 0.006513233762234449, 'rewards/near_duplicate_penalty/mean': 0.9922199249267578, 'rewards/near_duplicate_penalty/std': 0.012621927075088024, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9902632236480713, 'rewards/distinct_2/std': 0.02571846917271614, 'rewards/total_composite/mean': 0.3047851622104645, 'rewards/total_composite/std': 0.22900506854057312, 'reward': 0.3047851622104645, 'reward_std': 0.22900508344173431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17130866646766663, 'sampling/sampling_logp_difference/max': 2.774261236190796, 'sampling/importance_sampling_ratio/min': 0.06239555776119232, 'sampling/importance_sampling_ratio/mean': 1.0127062797546387, 'sampling/importance_sampling_ratio/max': 1.9650647640228271, 'entropy': 0.7918702214956284, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.10613978188484907, 'clip_ratio/high_max': 0.10613978188484907, 'clip_ratio/region_mean': 0.12002867087721825, 'reward_total_mean': 0.3047851622104645, 'reward_meter_mean': 0.6858423948287964, 'reward_meter_std': 0.3974907696247101, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8820273876190186, 'reward_lexical_plausibility_std': 0.21894796192646027, 'reward_repeat_penalty_mean': 0.9202359318733215, 'reward_repeat_penalty_std': 0.10980397462844849, 'reward_repeat_floor_mean': 0.9942066669464111, 'reward_repeat_floor_std': 0.006513233762234449, 'reward_near_duplicate_penalty_mean': 0.9922199249267578, 'reward_near_duplicate_penalty_std': 0.012621927075088024, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9902632236480713, 'reward_distinct_2_std': 0.02571846917271614, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.30420562624931335, 'reward_total_composite_mean': 0.3047851622104645, 'reward_total_composite_std': 0.22900506854057312, 'epoch': 0.03}

 19%|█▉        | 626/3300 [2:18:10<8:59:38, 12.11s/it]INFO 04-12 18:16:37 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 627/3300 [2:18:24<9:18:58, 12.55s/it]2026-04-12 18:16:50,671 | INFO | train_grpo_train | metrics_logged mode=train step=627

                                                      
{'loss': -0.0458, 'grad_norm': 2.3469016551971436, 'learning_rate': 8.103030303030303e-06, 'num_tokens': 1155334.0, 'completions/mean_length': 283.0, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.193363219499588, 'rewards/meter/std': 0.16206784546375275, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.11338934302330017, 'rewards/lexical_plausibility/mean': 0.7676507234573364, 'rewards/lexical_plausibility/std': 0.25055408477783203, 'rewards/judge_quality/mean': 0.3424999713897705, 'rewards/judge_quality/std': 0.38975268602371216, 'rewards/repeat_penalty/mean': 0.9447952508926392, 'rewards/repeat_penalty/std': 0.09382601827383041, 'rewards/repeat_floor/mean': 0.9956594705581665, 'rewards/repeat_floor/std': 0.0064465817995369434, 'rewards/near_duplicate_penalty/mean': 0.994872510433197, 'rewards/near_duplicate_penalty/std': 0.014502751640975475, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9811508059501648, 'rewards/distinct_2/std': 0.05160851404070854, 'rewards/total_composite/mean': 0.1001095250248909, 'rewards/total_composite/std': 0.1601189821958542, 'reward': 0.1001095250248909, 'reward_std': 0.1601189821958542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1514362096786499, 'sampling/sampling_logp_difference/max': 1.0914926528930664, 'sampling/importance_sampling_ratio/min': 0.33571502566337585, 'sampling/importance_sampling_ratio/mean': 1.0199594497680664, 'sampling/importance_sampling_ratio/max': 1.8444223403930664, 'entropy': 0.4910387694835663, 'clip_ratio/low_mean': 0.038455989211797714, 'clip_ratio/low_min': 0.038455989211797714, 'clip_ratio/high_mean': 0.029325180687010288, 'clip_ratio/high_max': 0.029325180687010288, 'clip_ratio/region_mean': 0.067781169898808, 'reward_total_mean': 0.1001095250248909, 'reward_meter_mean': 0.193363219499588, 'reward_meter_std': 0.16206784546375275, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.11338934302330017, 'reward_lexical_plausibility_mean': 0.7676507234573364, 'reward_lexical_plausibility_std': 0.25055408477783203, 'reward_repeat_penalty_mean': 0.9447952508926392, 'reward_repeat_penalty_std': 0.09382601827383041, 'reward_repeat_floor_mean': 0.9956594705581665, 'reward_repeat_floor_std': 0.0064465817995369434, 'reward_near_duplicate_penalty_mean': 0.994872510433197, 'reward_near_duplicate_penalty_std': 0.014502751640975475, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9811508059501648, 'reward_distinct_2_std': 0.05160851404070854, 'reward_judge_quality_mean': 0.3424999713897705, 'reward_judge_quality_std': 0.38975268602371216, 'reward_total_composite_mean': 0.1001095250248909, 'reward_total_composite_std': 0.1601189821958542, 'epoch': 0.03}

 19%|█▉        | 627/3300 [2:18:24<9:18:58, 12.55s/it]INFO 04-12 18:16:50 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 628/3300 [2:18:34<8:38:40, 11.65s/it]2026-04-12 18:17:00,217 | INFO | train_grpo_train | metrics_logged mode=train step=628

                                                      
{'loss': -0.0437, 'grad_norm': 11.550919532775879, 'learning_rate': 8.1e-06, 'num_tokens': 1157482.0, 'completions/mean_length': 94.5, 'completions/min_length': 72.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.43770888447761536, 'rewards/meter/std': 0.17247872054576874, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9929601550102234, 'rewards/lexical_plausibility/std': 0.014150693081319332, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.9124069809913635, 'rewards/repeat_penalty/std': 0.13274580240249634, 'rewards/repeat_floor/mean': 0.9900774955749512, 'rewards/repeat_floor/std': 0.014343813993036747, 'rewards/near_duplicate_penalty/mean': 0.9725484848022461, 'rewards/near_duplicate_penalty/std': 0.027609558776021004, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9355016350746155, 'rewards/distinct_2/std': 0.11589191854000092, 'rewards/total_composite/mean': 0.1530650556087494, 'rewards/total_composite/std': 0.060808662325143814, 'reward': 0.1530650556087494, 'reward_std': 0.060808662325143814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16860195994377136, 'sampling/sampling_logp_difference/max': 3.2653942108154297, 'sampling/importance_sampling_ratio/min': 0.03818188235163689, 'sampling/importance_sampling_ratio/mean': 1.012155532836914, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8972285687923431, 'clip_ratio/low_mean': 0.06858719978481531, 'clip_ratio/low_min': 0.06858719978481531, 'clip_ratio/high_mean': 0.06843694765120745, 'clip_ratio/high_max': 0.06843694765120745, 'clip_ratio/region_mean': 0.13702414743602276, 'reward_total_mean': 0.1530650556087494, 'reward_meter_mean': 0.43770888447761536, 'reward_meter_std': 0.17247872054576874, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9929601550102234, 'reward_lexical_plausibility_std': 0.014150693081319332, 'reward_repeat_penalty_mean': 0.9124069809913635, 'reward_repeat_penalty_std': 0.13274580240249634, 'reward_repeat_floor_mean': 0.9900774955749512, 'reward_repeat_floor_std': 0.014343813993036747, 'reward_near_duplicate_penalty_mean': 0.9725484848022461, 'reward_near_duplicate_penalty_std': 0.027609558776021004, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9355016350746155, 'reward_distinct_2_std': 0.11589191854000092, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.1530650556087494, 'reward_total_composite_std': 0.060808662325143814, 'epoch': 0.03}

 19%|█▉        | 628/3300 [2:18:34<8:38:40, 11.65s/it]INFO 04-12 18:17:00 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 629/3300 [2:18:43<8:04:56, 10.89s/it]2026-04-12 18:17:09,352 | INFO | train_grpo_train | metrics_logged mode=train step=629

                                                      
{'loss': 0.0558, 'grad_norm': 14.109796524047852, 'learning_rate': 8.096969696969698e-06, 'num_tokens': 1159115.0, 'completions/mean_length': 52.125, 'completions/min_length': 40.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.696418285369873, 'rewards/meter/std': 0.40393397212028503, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6512500047683716, 'rewards/judge_quality/std': 0.27220991253852844, 'rewards/repeat_penalty/mean': 0.9879008531570435, 'rewards/repeat_penalty/std': 0.027250805869698524, 'rewards/repeat_floor/mean': 0.9986053705215454, 'rewards/repeat_floor/std': 0.002909204689785838, 'rewards/near_duplicate_penalty/mean': 0.9951404333114624, 'rewards/near_duplicate_penalty/std': 0.0072609977796673775, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.5278056859970093, 'rewards/total_composite/std': 0.40111443400382996, 'reward': 0.5278056859970093, 'reward_std': 0.40111443400382996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15522265434265137, 'sampling/sampling_logp_difference/max': 2.420238494873047, 'sampling/importance_sampling_ratio/min': 0.08890041708946228, 'sampling/importance_sampling_ratio/mean': 1.008543848991394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9298888742923737, 'clip_ratio/low_mean': 0.06552162393927574, 'clip_ratio/low_min': 0.06552162393927574, 'clip_ratio/high_mean': 0.07076923083513975, 'clip_ratio/high_max': 0.07076923083513975, 'clip_ratio/region_mean': 0.1362908547744155, 'reward_total_mean': 0.5278056859970093, 'reward_meter_mean': 0.696418285369873, 'reward_meter_std': 0.40393397212028503, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9879008531570435, 'reward_repeat_penalty_std': 0.027250805869698524, 'reward_repeat_floor_mean': 0.9986053705215454, 'reward_repeat_floor_std': 0.002909204689785838, 'reward_near_duplicate_penalty_mean': 0.9951404333114624, 'reward_near_duplicate_penalty_std': 0.0072609977796673775, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.6512500047683716, 'reward_judge_quality_std': 0.27220991253852844, 'reward_total_composite_mean': 0.5278056859970093, 'reward_total_composite_std': 0.40111443400382996, 'epoch': 0.03}

 19%|█▉        | 629/3300 [2:18:43<8:04:56, 10.89s/it]INFO 04-12 18:17:09 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 630/3300 [2:18:52<7:45:18, 10.46s/it]2026-04-12 18:17:18,795 | INFO | train_grpo_train | metrics_logged mode=train step=630

                                                      
{'loss': 0.0252, 'grad_norm': 12.736981391906738, 'learning_rate': 8.093939393939395e-06, 'num_tokens': 1160646.0, 'completions/mean_length': 41.375, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9313610196113586, 'rewards/meter/std': 0.09954942762851715, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.1642080545425415, 'rewards/repeat_penalty/mean': 0.8535189628601074, 'rewards/repeat_penalty/std': 0.24123379588127136, 'rewards/repeat_floor/mean': 0.983771562576294, 'rewards/repeat_floor/std': 0.023354189470410347, 'rewards/near_duplicate_penalty/mean': 0.9401407241821289, 'rewards/near_duplicate_penalty/std': 0.06618963181972504, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9402834177017212, 'rewards/distinct_2/std': 0.11225500702857971, 'rewards/total_composite/mean': 0.40136462450027466, 'rewards/total_composite/std': 0.16953253746032715, 'reward': 0.40136462450027466, 'reward_std': 0.16953253746032715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1317305564880371, 'sampling/sampling_logp_difference/max': 1.3465065956115723, 'sampling/importance_sampling_ratio/min': 0.26014748215675354, 'sampling/importance_sampling_ratio/mean': 1.0096341371536255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8016039058566093, 'clip_ratio/low_mean': 0.07051901984959841, 'clip_ratio/low_min': 0.07051901984959841, 'clip_ratio/high_mean': 0.08341802190989256, 'clip_ratio/high_max': 0.08341802190989256, 'clip_ratio/region_mean': 0.15393704175949097, 'reward_total_mean': 0.40136462450027466, 'reward_meter_mean': 0.9313610196113586, 'reward_meter_std': 0.09954942762851715, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8535189628601074, 'reward_repeat_penalty_std': 0.24123379588127136, 'reward_repeat_floor_mean': 0.983771562576294, 'reward_repeat_floor_std': 0.023354189470410347, 'reward_near_duplicate_penalty_mean': 0.9401407241821289, 'reward_near_duplicate_penalty_std': 0.06618963181972504, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9402834177017212, 'reward_distinct_2_std': 0.11225500702857971, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.1642080545425415, 'reward_total_composite_mean': 0.40136462450027466, 'reward_total_composite_std': 0.16953253746032715, 'epoch': 0.03}

 19%|█▉        | 630/3300 [2:18:52<7:45:18, 10.46s/it]INFO 04-12 18:17:19 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 631/3300 [2:19:06<8:36:06, 11.60s/it]2026-04-12 18:17:33,066 | INFO | train_grpo_train | metrics_logged mode=train step=631

                                                      
{'loss': -0.1075, 'grad_norm': 2.931724786758423, 'learning_rate': 8.090909090909092e-06, 'num_tokens': 1162553.0, 'completions/mean_length': 250.375, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 93.4000015258789, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.27989137172698975, 'rewards/meter/std': 0.3009420335292816, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.23299294710159302, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.06989788264036179, 'rewards/lexical_plausibility/mean': 0.8564306497573853, 'rewards/lexical_plausibility/std': 0.21001528203487396, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.8253175020217896, 'rewards/repeat_penalty/std': 0.2620405852794647, 'rewards/repeat_floor/mean': 0.9799808263778687, 'rewards/repeat_floor/std': 0.02932579256594181, 'rewards/near_duplicate_penalty/mean': 0.9481757283210754, 'rewards/near_duplicate_penalty/std': 0.056299153715372086, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8639386892318726, 'rewards/distinct_2/std': 0.23870481550693512, 'rewards/total_composite/mean': 0.08268521726131439, 'rewards/total_composite/std': 0.10065069049596786, 'reward': 0.08268521726131439, 'reward_std': 0.10065069049596786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16391897201538086, 'sampling/sampling_logp_difference/max': 1.7942862510681152, 'sampling/importance_sampling_ratio/min': 0.16624607145786285, 'sampling/importance_sampling_ratio/mean': 1.0135902166366577, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7055342271924019, 'clip_ratio/low_mean': 0.04161974973976612, 'clip_ratio/low_min': 0.04161974973976612, 'clip_ratio/high_mean': 0.07705851644277573, 'clip_ratio/high_max': 0.07705851644277573, 'clip_ratio/region_mean': 0.11867826618254185, 'reward_total_mean': 0.08268521726131439, 'reward_meter_mean': 0.27989137172698975, 'reward_meter_std': 0.3009420335292816, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.23299294710159302, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.06989788264036179, 'reward_lexical_plausibility_mean': 0.8564306497573853, 'reward_lexical_plausibility_std': 0.21001528203487396, 'reward_repeat_penalty_mean': 0.8253175020217896, 'reward_repeat_penalty_std': 0.2620405852794647, 'reward_repeat_floor_mean': 0.9799808263778687, 'reward_repeat_floor_std': 0.02932579256594181, 'reward_near_duplicate_penalty_mean': 0.9481757283210754, 'reward_near_duplicate_penalty_std': 0.056299153715372086, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8639386892318726, 'reward_distinct_2_std': 0.23870481550693512, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.08268521726131439, 'reward_total_composite_std': 0.10065069049596786, 'epoch': 0.03}

 19%|█▉        | 631/3300 [2:19:06<8:36:06, 11.60s/it]INFO 04-12 18:17:33 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 632/3300 [2:19:17<8:21:29, 11.28s/it]2026-04-12 18:17:43,587 | INFO | train_grpo_train | metrics_logged mode=train step=632

                                                      
{'loss': 0.0396, 'grad_norm': 10.458887100219727, 'learning_rate': 8.08787878787879e-06, 'num_tokens': 1164630.0, 'completions/mean_length': 76.625, 'completions/min_length': 67.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.984645426273346, 'rewards/meter/std': 0.02025916427373886, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9899553656578064, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9664424657821655, 'rewards/repeat_penalty/std': 0.04354164004325867, 'rewards/repeat_floor/mean': 0.995843768119812, 'rewards/repeat_floor/std': 0.00489113898947835, 'rewards/near_duplicate_penalty/mean': 0.9819217920303345, 'rewards/near_duplicate_penalty/std': 0.01741056516766548, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839502573013306, 'rewards/distinct_2/std': 0.03230852633714676, 'rewards/total_composite/mean': 0.34018394351005554, 'rewards/total_composite/std': 0.04311790689826012, 'reward': 0.34018394351005554, 'reward_std': 0.04311790317296982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16201089322566986, 'sampling/sampling_logp_difference/max': 2.309434175491333, 'sampling/importance_sampling_ratio/min': 0.09931743144989014, 'sampling/importance_sampling_ratio/mean': 1.0083796977996826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8590324893593788, 'clip_ratio/low_mean': 0.11193040199577808, 'clip_ratio/low_min': 0.11193040199577808, 'clip_ratio/high_mean': 0.03403242491185665, 'clip_ratio/high_max': 0.03403242491185665, 'clip_ratio/region_mean': 0.14596282690763474, 'reward_total_mean': 0.34018394351005554, 'reward_meter_mean': 0.984645426273346, 'reward_meter_std': 0.02025916427373886, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9899553656578064, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9664424657821655, 'reward_repeat_penalty_std': 0.04354164004325867, 'reward_repeat_floor_mean': 0.995843768119812, 'reward_repeat_floor_std': 0.00489113898947835, 'reward_near_duplicate_penalty_mean': 0.9819217920303345, 'reward_near_duplicate_penalty_std': 0.01741056516766548, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839502573013306, 'reward_distinct_2_std': 0.03230852633714676, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.34018394351005554, 'reward_total_composite_std': 0.04311790689826012, 'epoch': 0.03}

 19%|█▉        | 632/3300 [2:19:17<8:21:29, 11.28s/it]INFO 04-12 18:17:43 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 633/3300 [2:19:31<8:52:36, 11.98s/it]2026-04-12 18:17:57,209 | INFO | train_grpo_train | metrics_logged mode=train step=633

                                                      
{'loss': -0.0276, 'grad_norm': 1.9008045196533203, 'learning_rate': 8.084848484848485e-06, 'num_tokens': 1166135.0, 'completions/mean_length': 207.125, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 24.200000762939453, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.511986255645752, 'rewards/meter/std': 0.48121657967567444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7669378519058228, 'rewards/lexical_plausibility/std': 0.23855535686016083, 'rewards/judge_quality/mean': 0.38875001668930054, 'rewards/judge_quality/std': 0.4436355531215668, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24821455776691437, 'rewards/total_composite/std': 0.37103891372680664, 'reward': 0.24821455776691437, 'reward_std': 0.37103888392448425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1641748994588852, 'sampling/sampling_logp_difference/max': 1.1376252174377441, 'sampling/importance_sampling_ratio/min': 0.3205794394016266, 'sampling/importance_sampling_ratio/mean': 1.0472924709320068, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.165491834282875, 'clip_ratio/low_mean': 0.06007657293230295, 'clip_ratio/low_min': 0.06007657293230295, 'clip_ratio/high_mean': 0.04138370044529438, 'clip_ratio/high_max': 0.04138370044529438, 'clip_ratio/region_mean': 0.10146027337759733, 'reward_total_mean': 0.24821455776691437, 'reward_meter_mean': 0.511986255645752, 'reward_meter_std': 0.48121657967567444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.7669378519058228, 'reward_lexical_plausibility_std': 0.23855535686016083, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38875001668930054, 'reward_judge_quality_std': 0.4436355531215668, 'reward_total_composite_mean': 0.24821455776691437, 'reward_total_composite_std': 0.37103891372680664, 'epoch': 0.03}

 19%|█▉        | 633/3300 [2:19:31<8:52:36, 11.98s/it]INFO 04-12 18:17:57 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 634/3300 [2:19:39<8:09:50, 11.02s/it]2026-04-12 18:18:06,003 | INFO | train_grpo_train | metrics_logged mode=train step=634

                                                      
{'loss': 0.0847, 'grad_norm': 12.01213264465332, 'learning_rate': 8.081818181818182e-06, 'num_tokens': 1167876.0, 'completions/mean_length': 59.625, 'completions/min_length': 47.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.4651067554950714, 'rewards/meter/std': 0.36555662751197815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9450562000274658, 'rewards/repeat_penalty/std': 0.026517393067479134, 'rewards/repeat_floor/mean': 0.9947631359100342, 'rewards/repeat_floor/std': 0.0021158012095838785, 'rewards/near_duplicate_penalty/mean': 0.9952726364135742, 'rewards/near_duplicate_penalty/std': 0.006237310823053122, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9496915936470032, 'rewards/distinct_2/std': 0.03162623941898346, 'rewards/total_composite/mean': 0.2696032226085663, 'rewards/total_composite/std': 0.2899021804332733, 'reward': 0.2696032226085663, 'reward_std': 0.2899021506309509, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15532128512859344, 'sampling/sampling_logp_difference/max': 2.546969413757324, 'sampling/importance_sampling_ratio/min': 0.07831865549087524, 'sampling/importance_sampling_ratio/mean': 0.9956446886062622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.66331672295928, 'clip_ratio/low_mean': 0.06250449735671282, 'clip_ratio/low_min': 0.06250449735671282, 'clip_ratio/high_mean': 0.06522547174245119, 'clip_ratio/high_max': 0.06522547174245119, 'clip_ratio/region_mean': 0.127729969099164, 'reward_total_mean': 0.2696032226085663, 'reward_meter_mean': 0.4651067554950714, 'reward_meter_std': 0.36555662751197815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9450562000274658, 'reward_repeat_penalty_std': 0.026517393067479134, 'reward_repeat_floor_mean': 0.9947631359100342, 'reward_repeat_floor_std': 0.0021158012095838785, 'reward_near_duplicate_penalty_mean': 0.9952726364135742, 'reward_near_duplicate_penalty_std': 0.006237310823053122, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9496915936470032, 'reward_distinct_2_std': 0.03162623941898346, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.2696032226085663, 'reward_total_composite_std': 0.2899021804332733, 'epoch': 0.03}

 19%|█▉        | 634/3300 [2:19:39<8:09:50, 11.02s/it]INFO 04-12 18:18:06 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 635/3300 [2:19:48<7:36:35, 10.28s/it]2026-04-12 18:18:14,544 | INFO | train_grpo_train | metrics_logged mode=train step=635

                                                      
{'loss': 0.0591, 'grad_norm': 14.6242036819458, 'learning_rate': 8.07878787878788e-06, 'num_tokens': 1169377.0, 'completions/mean_length': 37.625, 'completions/min_length': 30.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.2919130027294159, 'rewards/meter/std': 0.39304572343826294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.2655990719795227, 'rewards/repeat_penalty/mean': 0.9970802068710327, 'rewards/repeat_penalty/std': 0.0040726871229708195, 'rewards/repeat_floor/mean': 0.9995620250701904, 'rewards/repeat_floor/std': 0.0006109100650064647, 'rewards/near_duplicate_penalty/mean': 0.9970802068710327, 'rewards/near_duplicate_penalty/std': 0.0040726871229708195, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18419796228408813, 'rewards/total_composite/std': 0.301580011844635, 'reward': 0.18419796228408813, 'reward_std': 0.301580011844635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1748860478401184, 'sampling/sampling_logp_difference/max': 1.8300681114196777, 'sampling/importance_sampling_ratio/min': 0.16040264070034027, 'sampling/importance_sampling_ratio/mean': 1.0145468711853027, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8902975171804428, 'clip_ratio/low_mean': 0.1224973239004612, 'clip_ratio/low_min': 0.1224973239004612, 'clip_ratio/high_mean': 0.0573300588876009, 'clip_ratio/high_max': 0.0573300588876009, 'clip_ratio/region_mean': 0.1798273827880621, 'reward_total_mean': 0.18419796228408813, 'reward_meter_mean': 0.2919130027294159, 'reward_meter_std': 0.39304572343826294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9970802068710327, 'reward_repeat_penalty_std': 0.0040726871229708195, 'reward_repeat_floor_mean': 0.9995620250701904, 'reward_repeat_floor_std': 0.0006109100650064647, 'reward_near_duplicate_penalty_mean': 0.9970802068710327, 'reward_near_duplicate_penalty_std': 0.0040726871229708195, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.2655990719795227, 'reward_total_composite_mean': 0.18419796228408813, 'reward_total_composite_std': 0.301580011844635, 'epoch': 0.03}

 19%|█▉        | 635/3300 [2:19:48<7:36:35, 10.28s/it]INFO 04-12 18:18:14 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 636/3300 [2:20:02<8:22:25, 11.32s/it]2026-04-12 18:18:28,280 | INFO | train_grpo_train | metrics_logged mode=train step=636

                                                      
{'loss': -0.093, 'grad_norm': 3.6681129932403564, 'learning_rate': 8.075757575757577e-06, 'num_tokens': 1171323.0, 'completions/mean_length': 173.25, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 60.333335876464844, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7711304426193237, 'rewards/meter/std': 0.32977500557899475, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8590909242630005, 'rewards/lexical_plausibility/std': 0.1982439011335373, 'rewards/judge_quality/mean': 0.5600000023841858, 'rewards/judge_quality/std': 0.4012124836444855, 'rewards/repeat_penalty/mean': 0.9764533042907715, 'rewards/repeat_penalty/std': 0.06421243399381638, 'rewards/repeat_floor/mean': 0.9976084232330322, 'rewards/repeat_floor/std': 0.006407818291336298, 'rewards/near_duplicate_penalty/mean': 0.996201753616333, 'rewards/near_duplicate_penalty/std': 0.008486887440085411, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.05725561082363129, 'rewards/total_composite/mean': 0.44851696491241455, 'rewards/total_composite/std': 0.35535600781440735, 'reward': 0.44851696491241455, 'reward_std': 0.35535597801208496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14211297035217285, 'sampling/sampling_logp_difference/max': 1.4528303146362305, 'sampling/importance_sampling_ratio/min': 0.2339073121547699, 'sampling/importance_sampling_ratio/mean': 1.0040576457977295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.872333899140358, 'clip_ratio/low_mean': 0.04713064804673195, 'clip_ratio/low_min': 0.04713064804673195, 'clip_ratio/high_mean': 0.04814477916806936, 'clip_ratio/high_max': 0.04814477916806936, 'clip_ratio/region_mean': 0.09527542721480131, 'reward_total_mean': 0.44851696491241455, 'reward_meter_mean': 0.7711304426193237, 'reward_meter_std': 0.32977500557899475, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8590909242630005, 'reward_lexical_plausibility_std': 0.1982439011335373, 'reward_repeat_penalty_mean': 0.9764533042907715, 'reward_repeat_penalty_std': 0.06421243399381638, 'reward_repeat_floor_mean': 0.9976084232330322, 'reward_repeat_floor_std': 0.006407818291336298, 'reward_near_duplicate_penalty_mean': 0.996201753616333, 'reward_near_duplicate_penalty_std': 0.008486887440085411, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.05725561082363129, 'reward_judge_quality_mean': 0.5600000023841858, 'reward_judge_quality_std': 0.4012124836444855, 'reward_total_composite_mean': 0.44851696491241455, 'reward_total_composite_std': 0.35535600781440735, 'epoch': 0.03}

 19%|█▉        | 636/3300 [2:20:02<8:22:25, 11.32s/it]INFO 04-12 18:18:28 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 637/3300 [2:20:11<7:58:31, 10.78s/it]2026-04-12 18:18:37,812 | INFO | train_grpo_train | metrics_logged mode=train step=637

                                                      
{'loss': 0.0649, 'grad_norm': 15.80407428741455, 'learning_rate': 8.072727272727274e-06, 'num_tokens': 1172911.0, 'completions/mean_length': 39.5, 'completions/min_length': 35.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6007897257804871, 'rewards/meter/std': 0.3716011345386505, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.1060660257935524, 'rewards/repeat_penalty/mean': 0.9790390133857727, 'rewards/repeat_penalty/std': 0.041826989501714706, 'rewards/repeat_floor/mean': 0.9973326921463013, 'rewards/repeat_floor/std': 0.004962644074112177, 'rewards/near_duplicate_penalty/mean': 0.9877128005027771, 'rewards/near_duplicate_penalty/std': 0.018546780571341515, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.3065003752708435, 'rewards/total_composite/std': 0.22941149771213531, 'reward': 0.3065003752708435, 'reward_std': 0.22941149771213531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1655583530664444, 'sampling/sampling_logp_difference/max': 1.7887907028198242, 'sampling/importance_sampling_ratio/min': 0.1671621948480606, 'sampling/importance_sampling_ratio/mean': 1.002976894378662, 'sampling/importance_sampling_ratio/max': 1.954919695854187, 'entropy': 0.9274764209985733, 'clip_ratio/low_mean': 0.06570768170058727, 'clip_ratio/low_min': 0.06570768170058727, 'clip_ratio/high_mean': 0.05936716962605715, 'clip_ratio/high_max': 0.05936716962605715, 'clip_ratio/region_mean': 0.12507485132664442, 'reward_total_mean': 0.3065003752708435, 'reward_meter_mean': 0.6007897257804871, 'reward_meter_std': 0.3716011345386505, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9790390133857727, 'reward_repeat_penalty_std': 0.041826989501714706, 'reward_repeat_floor_mean': 0.9973326921463013, 'reward_repeat_floor_std': 0.004962644074112177, 'reward_near_duplicate_penalty_mean': 0.9877128005027771, 'reward_near_duplicate_penalty_std': 0.018546780571341515, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.1060660257935524, 'reward_total_composite_mean': 0.3065003752708435, 'reward_total_composite_std': 0.22941149771213531, 'epoch': 0.03}

 19%|█▉        | 637/3300 [2:20:11<7:58:31, 10.78s/it]INFO 04-12 18:18:38 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 638/3300 [2:20:25<8:42:49, 11.78s/it]2026-04-12 18:18:51,936 | INFO | train_grpo_train | metrics_logged mode=train step=638

                                                      
{'loss': -0.1528, 'grad_norm': 2.7844018936157227, 'learning_rate': 8.069696969696971e-06, 'num_tokens': 1174738.0, 'completions/mean_length': 188.375, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.20511597394943237, 'rewards/meter/std': 0.12907274067401886, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.28149792551994324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.08444937318563461, 'rewards/lexical_plausibility/mean': 0.8517034649848938, 'rewards/lexical_plausibility/std': 0.2762708067893982, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.27105283737182617, 'rewards/repeat_penalty/mean': 0.9351958632469177, 'rewards/repeat_penalty/std': 0.08337276428937912, 'rewards/repeat_floor/mean': 0.9945733547210693, 'rewards/repeat_floor/std': 0.005232109688222408, 'rewards/near_duplicate_penalty/mean': 0.9915031790733337, 'rewards/near_duplicate_penalty/std': 0.010013229213654995, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9746991395950317, 'rewards/distinct_2/std': 0.037408240139484406, 'rewards/total_composite/mean': 0.07715211063623428, 'rewards/total_composite/std': 0.05077496916055679, 'reward': 0.07715211063623428, 'reward_std': 0.05077496916055679, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1477160006761551, 'sampling/sampling_logp_difference/max': 2.1589579582214355, 'sampling/importance_sampling_ratio/min': 0.11544536054134369, 'sampling/importance_sampling_ratio/mean': 0.9956884384155273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5546809062361717, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.09150512423366308, 'clip_ratio/high_max': 0.09150512423366308, 'clip_ratio/region_mean': 0.10539401322603226, 'reward_total_mean': 0.07715211063623428, 'reward_meter_mean': 0.20511597394943237, 'reward_meter_std': 0.12907274067401886, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.28149792551994324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.08444937318563461, 'reward_lexical_plausibility_mean': 0.8517034649848938, 'reward_lexical_plausibility_std': 0.2762708067893982, 'reward_repeat_penalty_mean': 0.9351958632469177, 'reward_repeat_penalty_std': 0.08337276428937912, 'reward_repeat_floor_mean': 0.9945733547210693, 'reward_repeat_floor_std': 0.005232109688222408, 'reward_near_duplicate_penalty_mean': 0.9915031790733337, 'reward_near_duplicate_penalty_std': 0.010013229213654995, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9746991395950317, 'reward_distinct_2_std': 0.037408240139484406, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.27105283737182617, 'reward_total_composite_mean': 0.07715211063623428, 'reward_total_composite_std': 0.05077496916055679, 'epoch': 0.03}

 19%|█▉        | 638/3300 [2:20:25<8:42:49, 11.78s/it]INFO 04-12 18:18:52 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 639/3300 [2:20:33<7:46:19, 10.51s/it]2026-04-12 18:18:59,488 | INFO | train_grpo_train | metrics_logged mode=train step=639

                                                      
{'loss': -0.0928, 'grad_norm': 12.352059364318848, 'learning_rate': 8.066666666666667e-06, 'num_tokens': 1176210.0, 'completions/mean_length': 37.0, 'completions/min_length': 18.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.557997465133667, 'rewards/meter/std': 0.4295482635498047, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9237500429153442, 'rewards/judge_quality/std': 0.01060659158974886, 'rewards/repeat_penalty/mean': 0.9146258234977722, 'rewards/repeat_penalty/std': 0.1343667358160019, 'rewards/repeat_floor/mean': 0.990487813949585, 'rewards/repeat_floor/std': 0.012669355608522892, 'rewards/near_duplicate_penalty/mean': 0.9659851789474487, 'rewards/near_duplicate_penalty/std': 0.02702588029205799, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9718337059020996, 'rewards/distinct_2/std': 0.05414172261953354, 'rewards/total_composite/mean': 0.5110194683074951, 'rewards/total_composite/std': 0.39803242683410645, 'reward': 0.5110194683074951, 'reward_std': 0.39803239703178406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16012847423553467, 'sampling/sampling_logp_difference/max': 1.1732993125915527, 'sampling/importance_sampling_ratio/min': 0.30934464931488037, 'sampling/importance_sampling_ratio/mean': 0.9790233373641968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8377412855625153, 'clip_ratio/low_mean': 0.04780982993543148, 'clip_ratio/low_min': 0.04780982993543148, 'clip_ratio/high_mean': 0.10969540849328041, 'clip_ratio/high_max': 0.10969540849328041, 'clip_ratio/region_mean': 0.1575052384287119, 'reward_total_mean': 0.5110194683074951, 'reward_meter_mean': 0.557997465133667, 'reward_meter_std': 0.4295482635498047, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9146258234977722, 'reward_repeat_penalty_std': 0.1343667358160019, 'reward_repeat_floor_mean': 0.990487813949585, 'reward_repeat_floor_std': 0.012669355608522892, 'reward_near_duplicate_penalty_mean': 0.9659851789474487, 'reward_near_duplicate_penalty_std': 0.02702588029205799, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9718337059020996, 'reward_distinct_2_std': 0.05414172261953354, 'reward_judge_quality_mean': 0.9237500429153442, 'reward_judge_quality_std': 0.01060659158974886, 'reward_total_composite_mean': 0.5110194683074951, 'reward_total_composite_std': 0.39803242683410645, 'epoch': 0.03}

 19%|█▉        | 639/3300 [2:20:33<7:46:19, 10.51s/it]INFO 04-12 18:18:59 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 640/3300 [2:20:42<7:25:50, 10.06s/it]2026-04-12 18:19:08,483 | INFO | train_grpo_train | metrics_logged mode=train step=640

                                                      
{'loss': 0.0464, 'grad_norm': 16.527528762817383, 'learning_rate': 8.063636363636364e-06, 'num_tokens': 1177927.0, 'completions/mean_length': 50.625, 'completions/min_length': 46.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.625, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.9800794124603271, 'rewards/meter/std': 0.011560414917767048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7475000023841858, 'rewards/judge_quality/std': 0.23468825221061707, 'rewards/repeat_penalty/mean': 0.9642617702484131, 'rewards/repeat_penalty/std': 0.042984336614608765, 'rewards/repeat_floor/mean': 0.9960994124412537, 'rewards/repeat_floor/std': 0.0045709069818258286, 'rewards/near_duplicate_penalty/mean': 0.9893150329589844, 'rewards/near_duplicate_penalty/std': 0.014057926833629608, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.974468469619751, 'rewards/distinct_2/std': 0.03532608970999718, 'rewards/total_composite/mean': 0.7300345301628113, 'rewards/total_composite/std': 0.23006996512413025, 'reward': 0.7300345301628113, 'reward_std': 0.23006996512413025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16267350316047668, 'sampling/sampling_logp_difference/max': 1.480438232421875, 'sampling/importance_sampling_ratio/min': 0.22753795981407166, 'sampling/importance_sampling_ratio/mean': 1.0314425230026245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9453489556908607, 'clip_ratio/low_mean': 0.04160125646740198, 'clip_ratio/low_min': 0.04160125646740198, 'clip_ratio/high_mean': 0.09325505793094635, 'clip_ratio/high_max': 0.09325505793094635, 'clip_ratio/region_mean': 0.13485631439834833, 'reward_total_mean': 0.7300345301628113, 'reward_meter_mean': 0.9800794124603271, 'reward_meter_std': 0.011560414917767048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9642617702484131, 'reward_repeat_penalty_std': 0.042984336614608765, 'reward_repeat_floor_mean': 0.9960994124412537, 'reward_repeat_floor_std': 0.0045709069818258286, 'reward_near_duplicate_penalty_mean': 0.9893150329589844, 'reward_near_duplicate_penalty_std': 0.014057926833629608, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.974468469619751, 'reward_distinct_2_std': 0.03532608970999718, 'reward_judge_quality_mean': 0.7475000023841858, 'reward_judge_quality_std': 0.23468825221061707, 'reward_total_composite_mean': 0.7300345301628113, 'reward_total_composite_std': 0.23006996512413025, 'epoch': 0.03}

 19%|█▉        | 640/3300 [2:20:42<7:25:50, 10.06s/it]INFO 04-12 18:19:08 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 641/3300 [2:20:50<7:04:12,  9.57s/it]2026-04-12 18:19:16,927 | INFO | train_grpo_train | metrics_logged mode=train step=641

                                                      
{'loss': 0.1177, 'grad_norm': 15.365981101989746, 'learning_rate': 8.060606060606061e-06, 'num_tokens': 1179713.0, 'completions/mean_length': 48.25, 'completions/min_length': 42.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4804297387599945, 'rewards/meter/std': 0.44888126850128174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9308338165283203, 'rewards/repeat_penalty/std': 0.07589451968669891, 'rewards/repeat_floor/mean': 0.9913982152938843, 'rewards/repeat_floor/std': 0.008873924612998962, 'rewards/near_duplicate_penalty/mean': 0.9639303684234619, 'rewards/near_duplicate_penalty/std': 0.031476281583309174, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9645408391952515, 'rewards/distinct_2/std': 0.05330968648195267, 'rewards/total_composite/mean': 0.19269824028015137, 'rewards/total_composite/std': 0.17805632948875427, 'reward': 0.19269824028015137, 'reward_std': 0.17805634438991547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17214347422122955, 'sampling/sampling_logp_difference/max': 1.5031428337097168, 'sampling/importance_sampling_ratio/min': 0.2224300056695938, 'sampling/importance_sampling_ratio/mean': 1.0379966497421265, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1672176271677017, 'clip_ratio/low_mean': 0.06790501344949007, 'clip_ratio/low_min': 0.06790501344949007, 'clip_ratio/high_mean': 0.07647515553981066, 'clip_ratio/high_max': 0.07647515553981066, 'clip_ratio/region_mean': 0.14438016898930073, 'reward_total_mean': 0.19269824028015137, 'reward_meter_mean': 0.4804297387599945, 'reward_meter_std': 0.44888126850128174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9308338165283203, 'reward_repeat_penalty_std': 0.07589451968669891, 'reward_repeat_floor_mean': 0.9913982152938843, 'reward_repeat_floor_std': 0.008873924612998962, 'reward_near_duplicate_penalty_mean': 0.9639303684234619, 'reward_near_duplicate_penalty_std': 0.031476281583309174, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9645408391952515, 'reward_distinct_2_std': 0.05330968648195267, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.19269824028015137, 'reward_total_composite_std': 0.17805632948875427, 'epoch': 0.03}

 19%|█▉        | 641/3300 [2:20:50<7:04:12,  9.57s/it]INFO 04-12 18:19:17 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 642/3300 [2:21:04<8:02:14, 10.89s/it]2026-04-12 18:19:30,870 | INFO | train_grpo_train | metrics_logged mode=train step=642

                                                      
{'loss': -0.0601, 'grad_norm': 5.796260356903076, 'learning_rate': 8.057575757575759e-06, 'num_tokens': 1181116.0, 'completions/mean_length': 89.375, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 29.000001907348633, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8147009611129761, 'rewards/meter/std': 0.3297954201698303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9335466027259827, 'rewards/lexical_plausibility/std': 0.18795861303806305, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.32802167534828186, 'rewards/repeat_penalty/mean': 0.9793533086776733, 'rewards/repeat_penalty/std': 0.03178567439317703, 'rewards/repeat_floor/mean': 0.9969362020492554, 'rewards/repeat_floor/std': 0.0047591682523489, 'rewards/near_duplicate_penalty/mean': 0.9799289703369141, 'rewards/near_duplicate_penalty/std': 0.03166106343269348, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.999410092830658, 'rewards/distinct_2/std': 0.001668488490395248, 'rewards/total_composite/mean': 0.624222993850708, 'rewards/total_composite/std': 0.32347583770751953, 'reward': 0.624222993850708, 'reward_std': 0.32347583770751953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16572195291519165, 'sampling/sampling_logp_difference/max': 1.5920886993408203, 'sampling/importance_sampling_ratio/min': 0.20350012183189392, 'sampling/importance_sampling_ratio/mean': 1.0216639041900635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7823265567421913, 'clip_ratio/low_mean': 0.045048702508211136, 'clip_ratio/low_min': 0.045048702508211136, 'clip_ratio/high_mean': 0.0937815411016345, 'clip_ratio/high_max': 0.0937815411016345, 'clip_ratio/region_mean': 0.13883024360984564, 'reward_total_mean': 0.624222993850708, 'reward_meter_mean': 0.8147009611129761, 'reward_meter_std': 0.3297954201698303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9335466027259827, 'reward_lexical_plausibility_std': 0.18795861303806305, 'reward_repeat_penalty_mean': 0.9793533086776733, 'reward_repeat_penalty_std': 0.03178567439317703, 'reward_repeat_floor_mean': 0.9969362020492554, 'reward_repeat_floor_std': 0.0047591682523489, 'reward_near_duplicate_penalty_mean': 0.9799289703369141, 'reward_near_duplicate_penalty_std': 0.03166106343269348, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.999410092830658, 'reward_distinct_2_std': 0.001668488490395248, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.32802167534828186, 'reward_total_composite_mean': 0.624222993850708, 'reward_total_composite_std': 0.32347583770751953, 'epoch': 0.03}

 19%|█▉        | 642/3300 [2:21:04<8:02:14, 10.89s/it]INFO 04-12 18:19:31 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 643/3300 [2:21:18<8:45:07, 11.86s/it]2026-04-12 18:19:44,999 | INFO | train_grpo_train | metrics_logged mode=train step=643

                                                      
{'loss': -0.0559, 'grad_norm': 6.062872409820557, 'learning_rate': 8.054545454545454e-06, 'num_tokens': 1183203.0, 'completions/mean_length': 134.875, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.8096573352813721, 'rewards/meter/std': 0.27412208914756775, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.07559289783239365, 'rewards/lexical_plausibility/mean': 0.9307242631912231, 'rewards/lexical_plausibility/std': 0.17208310961723328, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.3358757197856903, 'rewards/repeat_penalty/mean': 0.947460412979126, 'rewards/repeat_penalty/std': 0.08671032637357712, 'rewards/repeat_floor/mean': 0.995560884475708, 'rewards/repeat_floor/std': 0.005473115481436253, 'rewards/near_duplicate_penalty/mean': 0.989499568939209, 'rewards/near_duplicate_penalty/std': 0.010304414667189121, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9890109896659851, 'rewards/distinct_2/std': 0.031081615015864372, 'rewards/total_composite/mean': 0.4117696285247803, 'rewards/total_composite/std': 0.303749680519104, 'reward': 0.4117696285247803, 'reward_std': 0.303749680519104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13429448008537292, 'sampling/sampling_logp_difference/max': 1.6167144775390625, 'sampling/importance_sampling_ratio/min': 0.19854995608329773, 'sampling/importance_sampling_ratio/mean': 1.0122401714324951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8179492354393005, 'clip_ratio/low_mean': 0.055725112557411194, 'clip_ratio/low_min': 0.055725112557411194, 'clip_ratio/high_mean': 0.06733083166182041, 'clip_ratio/high_max': 0.06733083166182041, 'clip_ratio/region_mean': 0.1230559442192316, 'reward_total_mean': 0.4117696285247803, 'reward_meter_mean': 0.8096573352813721, 'reward_meter_std': 0.27412208914756775, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.07559289783239365, 'reward_lexical_plausibility_mean': 0.9307242631912231, 'reward_lexical_plausibility_std': 0.17208310961723328, 'reward_repeat_penalty_mean': 0.947460412979126, 'reward_repeat_penalty_std': 0.08671032637357712, 'reward_repeat_floor_mean': 0.995560884475708, 'reward_repeat_floor_std': 0.005473115481436253, 'reward_near_duplicate_penalty_mean': 0.989499568939209, 'reward_near_duplicate_penalty_std': 0.010304414667189121, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9890109896659851, 'reward_distinct_2_std': 0.031081615015864372, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.3358757197856903, 'reward_total_composite_mean': 0.4117696285247803, 'reward_total_composite_std': 0.303749680519104, 'epoch': 0.03}

 19%|█▉        | 643/3300 [2:21:18<8:45:07, 11.86s/it]INFO 04-12 18:19:45 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 644/3300 [2:21:32<9:04:14, 12.29s/it]2026-04-12 18:19:58,309 | INFO | train_grpo_train | metrics_logged mode=train step=644

                                                      
{'loss': -0.0233, 'grad_norm': 4.525539875030518, 'learning_rate': 8.051515151515153e-06, 'num_tokens': 1184503.0, 'completions/mean_length': 84.5, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.428571701049805, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.7014280557632446, 'rewards/meter/std': 0.4130121171474457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.933986246585846, 'rewards/lexical_plausibility/std': 0.1867150366306305, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.35419678688049316, 'rewards/repeat_penalty/mean': 0.7414772510528564, 'rewards/repeat_penalty/std': 0.02410591021180153, 'rewards/repeat_floor/mean': 0.9832954406738281, 'rewards/repeat_floor/std': 0.004821178037673235, 'rewards/near_duplicate_penalty/mean': 0.9886363744735718, 'rewards/near_duplicate_penalty/std': 0.03214120864868164, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.42818528413772583, 'rewards/total_composite/std': 0.4038875102996826, 'reward': 0.42818528413772583, 'reward_std': 0.40388745069503784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1365523338317871, 'sampling/sampling_logp_difference/max': 1.3165016174316406, 'sampling/importance_sampling_ratio/min': 0.2680715024471283, 'sampling/importance_sampling_ratio/mean': 1.0119614601135254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7186261564493179, 'clip_ratio/low_mean': 0.06812500022351742, 'clip_ratio/low_min': 0.06812500022351742, 'clip_ratio/high_mean': 0.0762648219242692, 'clip_ratio/high_max': 0.0762648219242692, 'clip_ratio/region_mean': 0.14438982214778662, 'reward_total_mean': 0.42818528413772583, 'reward_meter_mean': 0.7014280557632446, 'reward_meter_std': 0.4130121171474457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.933986246585846, 'reward_lexical_plausibility_std': 0.1867150366306305, 'reward_repeat_penalty_mean': 0.7414772510528564, 'reward_repeat_penalty_std': 0.02410591021180153, 'reward_repeat_floor_mean': 0.9832954406738281, 'reward_repeat_floor_std': 0.004821178037673235, 'reward_near_duplicate_penalty_mean': 0.9886363744735718, 'reward_near_duplicate_penalty_std': 0.03214120864868164, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.35419678688049316, 'reward_total_composite_mean': 0.42818528413772583, 'reward_total_composite_std': 0.4038875102996826, 'epoch': 0.03}

 20%|█▉        | 644/3300 [2:21:32<9:04:14, 12.29s/it]INFO 04-12 18:19:58 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 645/3300 [2:21:39<8:05:04, 10.96s/it]2026-04-12 18:20:06,166 | INFO | train_grpo_train | metrics_logged mode=train step=645

                                                      
{'loss': 0.0436, 'grad_norm': 30.82151985168457, 'learning_rate': 8.048484848484849e-06, 'num_tokens': 1185940.0, 'completions/mean_length': 18.625, 'completions/min_length': 15.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9655370712280273, 'rewards/meter/std': 0.02885575219988823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3366537094116211, 'rewards/repeat_penalty/mean': 0.7392919659614563, 'rewards/repeat_penalty/std': 0.030286915600299835, 'rewards/repeat_floor/mean': 0.982858419418335, 'rewards/repeat_floor/std': 0.006057387217879295, 'rewards/near_duplicate_penalty/mean': 0.9857226014137268, 'rewards/near_duplicate_penalty/std': 0.04038256034255028, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3927711844444275, 'rewards/total_composite/std': 0.31311947107315063, 'reward': 0.3927711844444275, 'reward_std': 0.31311944127082825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1656142771244049, 'sampling/sampling_logp_difference/max': 1.7993075847625732, 'sampling/importance_sampling_ratio/min': 0.16541339457035065, 'sampling/importance_sampling_ratio/mean': 0.9998873472213745, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9678597524762154, 'clip_ratio/low_mean': 0.08424375578761101, 'clip_ratio/low_min': 0.08424375578761101, 'clip_ratio/high_mean': 0.0791118424385786, 'clip_ratio/high_max': 0.0791118424385786, 'clip_ratio/region_mean': 0.1633555982261896, 'reward_total_mean': 0.3927711844444275, 'reward_meter_mean': 0.9655370712280273, 'reward_meter_std': 0.02885575219988823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7392919659614563, 'reward_repeat_penalty_std': 0.030286915600299835, 'reward_repeat_floor_mean': 0.982858419418335, 'reward_repeat_floor_std': 0.006057387217879295, 'reward_near_duplicate_penalty_mean': 0.9857226014137268, 'reward_near_duplicate_penalty_std': 0.04038256034255028, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3366537094116211, 'reward_total_composite_mean': 0.3927711844444275, 'reward_total_composite_std': 0.31311947107315063, 'epoch': 0.03}

 20%|█▉        | 645/3300 [2:21:40<8:05:04, 10.96s/it]INFO 04-12 18:20:06 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 646/3300 [2:21:48<7:29:03, 10.15s/it]2026-04-12 18:20:14,427 | INFO | train_grpo_train | metrics_logged mode=train step=646

                                                      
{'loss': 0.0835, 'grad_norm': 17.655353546142578, 'learning_rate': 8.045454545454546e-06, 'num_tokens': 1187454.0, 'completions/mean_length': 26.25, 'completions/min_length': 23.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8672947287559509, 'rewards/meter/std': 0.2532776892185211, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.7383207082748413, 'rewards/repeat_penalty/std': 0.03303402662277222, 'rewards/repeat_floor/mean': 0.982664167881012, 'rewards/repeat_floor/std': 0.0066068135201931, 'rewards/near_duplicate_penalty/mean': 0.9844275712966919, 'rewards/near_duplicate_penalty/std': 0.044045377522706985, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5407683849334717, 'rewards/total_composite/std': 0.30127987265586853, 'reward': 0.5407683849334717, 'reward_std': 0.30127984285354614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14098912477493286, 'sampling/sampling_logp_difference/max': 1.4134750366210938, 'sampling/importance_sampling_ratio/min': 0.24329635500907898, 'sampling/importance_sampling_ratio/mean': 1.015126347541809, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8547587618231773, 'clip_ratio/low_mean': 0.07694461708888412, 'clip_ratio/low_min': 0.07694461708888412, 'clip_ratio/high_mean': 0.04282794566825032, 'clip_ratio/high_max': 0.04282794566825032, 'clip_ratio/region_mean': 0.11977256275713444, 'reward_total_mean': 0.5407683849334717, 'reward_meter_mean': 0.8672947287559509, 'reward_meter_std': 0.2532776892185211, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7383207082748413, 'reward_repeat_penalty_std': 0.03303402662277222, 'reward_repeat_floor_mean': 0.982664167881012, 'reward_repeat_floor_std': 0.0066068135201931, 'reward_near_duplicate_penalty_mean': 0.9844275712966919, 'reward_near_duplicate_penalty_std': 0.044045377522706985, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.5407683849334717, 'reward_total_composite_std': 0.30127987265586853, 'epoch': 0.03}

 20%|█▉        | 646/3300 [2:21:48<7:29:03, 10.15s/it]INFO 04-12 18:20:14 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 647/3300 [2:21:58<7:33:10, 10.25s/it]2026-04-12 18:20:24,898 | INFO | train_grpo_train | metrics_logged mode=train step=647

                                                      
{'loss': -0.0292, 'grad_norm': 10.391853332519531, 'learning_rate': 8.042424242424243e-06, 'num_tokens': 1189251.0, 'completions/mean_length': 58.625, 'completions/min_length': 41.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6281118392944336, 'rewards/meter/std': 0.37566012144088745, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9382601976394653, 'rewards/repeat_penalty/std': 0.05668553337454796, 'rewards/repeat_floor/mean': 0.9934574365615845, 'rewards/repeat_floor/std': 0.005714621394872665, 'rewards/near_duplicate_penalty/mean': 0.9850168228149414, 'rewards/near_duplicate_penalty/std': 0.012420056387782097, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.952276349067688, 'rewards/distinct_2/std': 0.05141378566622734, 'rewards/total_composite/mean': 0.2798583507537842, 'rewards/total_composite/std': 0.16829027235507965, 'reward': 0.2798583507537842, 'reward_std': 0.16829027235507965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13313207030296326, 'sampling/sampling_logp_difference/max': 2.088466167449951, 'sampling/importance_sampling_ratio/min': 0.12387699633836746, 'sampling/importance_sampling_ratio/mean': 0.9937934279441833, 'sampling/importance_sampling_ratio/max': 1.9726005792617798, 'entropy': 0.7814536467194557, 'clip_ratio/low_mean': 0.04677249863743782, 'clip_ratio/low_min': 0.04677249863743782, 'clip_ratio/high_mean': 0.08222942613065243, 'clip_ratio/high_max': 0.08222942613065243, 'clip_ratio/region_mean': 0.12900192476809025, 'reward_total_mean': 0.2798583507537842, 'reward_meter_mean': 0.6281118392944336, 'reward_meter_std': 0.37566012144088745, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9382601976394653, 'reward_repeat_penalty_std': 0.05668553337454796, 'reward_repeat_floor_mean': 0.9934574365615845, 'reward_repeat_floor_std': 0.005714621394872665, 'reward_near_duplicate_penalty_mean': 0.9850168228149414, 'reward_near_duplicate_penalty_std': 0.012420056387782097, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.952276349067688, 'reward_distinct_2_std': 0.05141378566622734, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.2798583507537842, 'reward_total_composite_std': 0.16829027235507965, 'epoch': 0.03}

 20%|█▉        | 647/3300 [2:21:58<7:33:10, 10.25s/it]INFO 04-12 18:20:25 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 648/3300 [2:22:06<7:04:28,  9.60s/it]2026-04-12 18:20:33,002 | INFO | train_grpo_train | metrics_logged mode=train step=648

                                                      
{'loss': 0.0954, 'grad_norm': 21.5165958404541, 'learning_rate': 8.03939393939394e-06, 'num_tokens': 1190744.0, 'completions/mean_length': 28.625, 'completions/min_length': 26.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.28281521797180176, 'rewards/meter/std': 0.2929595410823822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8387500047683716, 'rewards/judge_quality/std': 0.27869275212287903, 'rewards/repeat_penalty/mean': 0.9794915914535522, 'rewards/repeat_penalty/std': 0.024691898375749588, 'rewards/repeat_floor/mean': 0.9969237446784973, 'rewards/repeat_floor/std': 0.0037037809379398823, 'rewards/near_duplicate_penalty/mean': 0.9794915914535522, 'rewards/near_duplicate_penalty/std': 0.024691898375749588, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2579765021800995, 'rewards/total_composite/std': 0.2762589752674103, 'reward': 0.2579765021800995, 'reward_std': 0.2762589752674103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1504424512386322, 'sampling/sampling_logp_difference/max': 1.6232399940490723, 'sampling/importance_sampling_ratio/min': 0.24931108951568604, 'sampling/importance_sampling_ratio/mean': 1.012108325958252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6194121949374676, 'clip_ratio/low_mean': 0.06246027257293463, 'clip_ratio/low_min': 0.06246027257293463, 'clip_ratio/high_mean': 0.04237891733646393, 'clip_ratio/high_max': 0.04237891733646393, 'clip_ratio/region_mean': 0.10483918990939856, 'reward_total_mean': 0.2579765021800995, 'reward_meter_mean': 0.28281521797180176, 'reward_meter_std': 0.2929595410823822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9794915914535522, 'reward_repeat_penalty_std': 0.024691898375749588, 'reward_repeat_floor_mean': 0.9969237446784973, 'reward_repeat_floor_std': 0.0037037809379398823, 'reward_near_duplicate_penalty_mean': 0.9794915914535522, 'reward_near_duplicate_penalty_std': 0.024691898375749588, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8387500047683716, 'reward_judge_quality_std': 0.27869275212287903, 'reward_total_composite_mean': 0.2579765021800995, 'reward_total_composite_std': 0.2762589752674103, 'epoch': 0.03}

 20%|█▉        | 648/3300 [2:22:06<7:04:28,  9.60s/it]INFO 04-12 18:20:33 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 649/3300 [2:22:16<7:10:28,  9.74s/it]2026-04-12 18:20:43,065 | INFO | train_grpo_train | metrics_logged mode=train step=649

                                                      
{'loss': 0.0498, 'grad_norm': 11.147249221801758, 'learning_rate': 8.036363636363636e-06, 'num_tokens': 1192930.0, 'completions/mean_length': 91.25, 'completions/min_length': 81.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.25, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.5871928334236145, 'rewards/meter/std': 0.3409288227558136, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9700000286102295, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5037499666213989, 'rewards/judge_quality/std': 0.15202796459197998, 'rewards/repeat_penalty/mean': 0.9305137991905212, 'rewards/repeat_penalty/std': 0.10091582685709, 'rewards/repeat_floor/mean': 0.9933023452758789, 'rewards/repeat_floor/std': 0.007848613895475864, 'rewards/near_duplicate_penalty/mean': 0.9761093258857727, 'rewards/near_duplicate_penalty/std': 0.02028193138539791, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9653991460800171, 'rewards/distinct_2/std': 0.05607825890183449, 'rewards/total_composite/mean': 0.27145183086395264, 'rewards/total_composite/std': 0.14429277181625366, 'reward': 0.27145183086395264, 'reward_std': 0.14429277181625366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16846567392349243, 'sampling/sampling_logp_difference/max': 2.271028518676758, 'sampling/importance_sampling_ratio/min': 0.10320597141981125, 'sampling/importance_sampling_ratio/mean': 0.992217481136322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8758502081036568, 'clip_ratio/low_mean': 0.041273620910942554, 'clip_ratio/low_min': 0.041273620910942554, 'clip_ratio/high_mean': 0.09511497523635626, 'clip_ratio/high_max': 0.09511497523635626, 'clip_ratio/region_mean': 0.1363885961472988, 'reward_total_mean': 0.27145183086395264, 'reward_meter_mean': 0.5871928334236145, 'reward_meter_std': 0.3409288227558136, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9700000286102295, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9305137991905212, 'reward_repeat_penalty_std': 0.10091582685709, 'reward_repeat_floor_mean': 0.9933023452758789, 'reward_repeat_floor_std': 0.007848613895475864, 'reward_near_duplicate_penalty_mean': 0.9761093258857727, 'reward_near_duplicate_penalty_std': 0.02028193138539791, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9653991460800171, 'reward_distinct_2_std': 0.05607825890183449, 'reward_judge_quality_mean': 0.5037499666213989, 'reward_judge_quality_std': 0.15202796459197998, 'reward_total_composite_mean': 0.27145183086395264, 'reward_total_composite_std': 0.14429277181625366, 'epoch': 0.03}

 20%|█▉        | 649/3300 [2:22:16<7:10:28,  9.74s/it]INFO 04-12 18:20:43 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 650/3300 [2:22:25<6:59:15,  9.49s/it]2026-04-12 18:20:51,973 | INFO | train_grpo_train | metrics_logged mode=train step=650

                                                      
{'loss': 0.0087, 'grad_norm': 23.668455123901367, 'learning_rate': 8.033333333333335e-06, 'num_tokens': 1194340.0, 'completions/mean_length': 19.25, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.8139858245849609, 'rewards/meter/std': 0.21032409369945526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6687500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.7229496240615845, 'rewards/repeat_penalty/std': 0.07651007920503616, 'rewards/repeat_floor/mean': 0.9795899391174316, 'rewards/repeat_floor/std': 0.015302024781703949, 'rewards/near_duplicate_penalty/mean': 0.9639328122138977, 'rewards/near_duplicate_penalty/std': 0.10201343148946762, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5177667140960693, 'rewards/total_composite/std': 0.3268375098705292, 'reward': 0.5177667140960693, 'reward_std': 0.3268374800682068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1575707495212555, 'sampling/sampling_logp_difference/max': 1.1000494956970215, 'sampling/importance_sampling_ratio/min': 0.33285459876060486, 'sampling/importance_sampling_ratio/mean': 1.0092663764953613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0452252626419067, 'clip_ratio/low_mean': 0.07959280535578728, 'clip_ratio/low_min': 0.07959280535578728, 'clip_ratio/high_mean': 0.08957219496369362, 'clip_ratio/high_max': 0.08957219496369362, 'clip_ratio/region_mean': 0.1691650003194809, 'reward_total_mean': 0.5177667140960693, 'reward_meter_mean': 0.8139858245849609, 'reward_meter_std': 0.21032409369945526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7229496240615845, 'reward_repeat_penalty_std': 0.07651007920503616, 'reward_repeat_floor_mean': 0.9795899391174316, 'reward_repeat_floor_std': 0.015302024781703949, 'reward_near_duplicate_penalty_mean': 0.9639328122138977, 'reward_near_duplicate_penalty_std': 0.10201343148946762, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6687500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.5177667140960693, 'reward_total_composite_std': 0.3268375098705292, 'epoch': 0.03}

 20%|█▉        | 650/3300 [2:22:25<6:59:15,  9.49s/it]INFO 04-12 18:20:52 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:21<02:00, 10.97s/it]

 23%|██▎       | 3/13 [00:28<01:31,  9.14s/it]

 31%|███       | 4/13 [00:40<01:31, 10.12s/it]

 38%|███▊      | 5/13 [00:50<01:21, 10.14s/it]

 46%|████▌     | 6/13 [01:00<01:10, 10.07s/it]

 54%|█████▍    | 7/13 [01:09<00:58,  9.83s/it]

 62%|██████▏   | 8/13 [01:20<00:49,  9.98s/it]

 69%|██████▉   | 9/13 [01:31<00:41, 10.45s/it]

 77%|███████▋  | 10/13 [01:40<00:30, 10.03s/it]

 85%|████████▍ | 11/13 [01:51<00:20, 10.23s/it]

 92%|█████████▏| 12/13 [02:02<00:10, 10.57s/it]

100%|██████████| 13/13 [02:11<00:00, 10.16s/it]2026-04-12 18:23:15,462 | INFO | train_grpo_train | metrics_logged mode=eval step=650

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 143.429, 'eval_samples_per_second': 0.725, 'eval_steps_per_second': 0.091, 'eval_num_tokens': 1194340.0, 'eval_completions/mean_length': 172.4903846153846, 'eval_completions/min_length': 42.23076923076923, 'eval_completions/max_length': 394.46153846153845, 'eval_completions/clipped_ratio': 0.07692307692307693, 'eval_completions/mean_terminated_length': 144.15045870267429, 'eval_completions/min_terminated_length': 42.23076923076923, 'eval_completions/max_terminated_length': 306.2307692307692, 'eval_rewards/meter/mean': 0.5530450894282415, 'eval_rewards/meter/std': 0.37604402808042675, 'eval_rewards/count_adherence/mean': 0.8929206407987155, 'eval_rewards/count_adherence/std': 0.16625133930490568, 'eval_rewards/arabic_clean/mean': 0.8653846153846154, 'eval_rewards/arabic_clean/std': 0.2337347658780905, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9971153827813956, 'eval_rewards/arabic_floor/std': 0.008158923341677738, 'eval_rewards/count_floor/mean': 0.9678761867376474, 'eval_rewards/count_floor/std': 0.049875402536529764, 'eval_rewards/lexical_plausibility/mean': 0.9647191029328567, 'eval_rewards/lexical_plausibility/std': 0.0759761416568206, 'eval_rewards/judge_quality/mean': 0.3601923103515918, 'eval_rewards/judge_quality/std': 0.2043610387123548, 'eval_rewards/repeat_penalty/mean': 0.6630720587877127, 'eval_rewards/repeat_penalty/std': 0.36234451601138484, 'eval_rewards/repeat_floor/mean': 0.9590880824969366, 'eval_rewards/repeat_floor/std': 0.049047087820676655, 'eval_rewards/near_duplicate_penalty/mean': 0.9166562694769639, 'eval_rewards/near_duplicate_penalty/std': 0.10195524646685673, 'eval_rewards/opening_diversity/mean': 0.960006805566641, 'eval_rewards/opening_diversity/std': 0.07581744271402176, 'eval_rewards/distinct_2/mean': 0.7109913780139043, 'eval_rewards/distinct_2/std': 0.3618535238962907, 'eval_rewards/total_composite/mean': 0.18821023404598236, 'eval_rewards/total_composite/std': 0.17175685069881952, 'eval_reward': 0.18821023404598236, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.057632267475128174, 'eval_sampling/sampling_logp_difference/max': 1.0790939331054688, 'eval_sampling/importance_sampling_ratio/min': 0.3468034748847668, 'eval_sampling/importance_sampling_ratio/mean': 1.0157257410196157, 'eval_sampling/importance_sampling_ratio/max': 1.6122803688049316, 'eval_entropy': 0.6549245073245122, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.18821023404598236, 'eval_reward_meter_mean': 0.5530450894282415, 'eval_reward_meter_std': 0.37604402808042675, 'eval_reward_count_adherence_mean': 0.8929206407987155, 'eval_reward_count_adherence_std': 0.16625133930490568, 'eval_reward_arabic_clean_mean': 0.8653846153846154, 'eval_reward_arabic_clean_std': 0.2337347658780905, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9971153827813956, 'eval_reward_arabic_floor_std': 0.008158923341677738, 'eval_reward_count_floor_mean': 0.9678761867376474, 'eval_reward_count_floor_std': 0.049875402536529764, 'eval_reward_lexical_plausibility_mean': 0.9647191029328567, 'eval_reward_lexical_plausibility_std': 0.0759761416568206, 'eval_reward_repeat_penalty_mean': 0.6630720587877127, 'eval_reward_repeat_penalty_std': 0.36234451601138484, 'eval_reward_repeat_floor_mean': 0.9590880824969366, 'eval_reward_repeat_floor_std': 0.049047087820676655, 'eval_reward_near_duplicate_penalty_mean': 0.9166562694769639, 'eval_reward_near_duplicate_penalty_std': 0.10195524646685673, 'eval_reward_opening_diversity_mean': 0.960006805566641, 'eval_reward_opening_diversity_std': 0.07581744271402176, 'eval_reward_distinct_2_mean': 0.7109913780139043, 'eval_reward_distinct_2_std': 0.3618535238962907, 'eval_reward_judge_quality_mean': 0.3601923103515918, 'eval_reward_judge_quality_std': 0.2043610387123548, 'eval_reward_total_composite_mean': 0.18821023404598236, 'eval_reward_total_composite_std': 0.17175685069881952, 'epoch': 0.03}

 20%|█▉        | 650/3300 [2:24:49<6:59:15,  9.49s/it]

100%|██████████| 13/13 [02:11<00:00, 10.16s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 20%|█▉        | 651/3300 [2:25:04<40:02:03, 54.41s/it]2026-04-12 18:23:31,183 | INFO | train_grpo_train | metrics_logged mode=train step=651

                                                       
{'loss': -0.0388, 'grad_norm': 2.5974252223968506, 'learning_rate': 8.03030303030303e-06, 'num_tokens': 1195733.0, 'completions/mean_length': 149.125, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 28.166667938232422, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.6173913478851318, 'rewards/meter/std': 0.4526126980781555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.887458324432373, 'rewards/lexical_plausibility/std': 0.20905417203903198, 'rewards/judge_quality/mean': 0.643750011920929, 'rewards/judge_quality/std': 0.40074706077575684, 'rewards/repeat_penalty/mean': 0.7300152778625488, 'rewards/repeat_penalty/std': 0.037389468401670456, 'rewards/repeat_floor/mean': 0.9810030460357666, 'rewards/repeat_floor/std': 0.007477888371795416, 'rewards/near_duplicate_penalty/mean': 0.9733536243438721, 'rewards/near_duplicate_penalty/std': 0.04985261708498001, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4706975817680359, 'rewards/total_composite/std': 0.4523340165615082, 'reward': 0.4706975817680359, 'reward_std': 0.4523339867591858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10816600173711777, 'sampling/sampling_logp_difference/max': 0.9233765602111816, 'sampling/importance_sampling_ratio/min': 0.39717569947242737, 'sampling/importance_sampling_ratio/mean': 1.0162279605865479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7349778488278389, 'clip_ratio/low_mean': 0.045770203694701195, 'clip_ratio/low_min': 0.045770203694701195, 'clip_ratio/high_mean': 0.0504464297555387, 'clip_ratio/high_max': 0.0504464297555387, 'clip_ratio/region_mean': 0.0962166334502399, 'reward_total_mean': 0.4706975817680359, 'reward_meter_mean': 0.6173913478851318, 'reward_meter_std': 0.4526126980781555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.887458324432373, 'reward_lexical_plausibility_std': 0.20905417203903198, 'reward_repeat_penalty_mean': 0.7300152778625488, 'reward_repeat_penalty_std': 0.037389468401670456, 'reward_repeat_floor_mean': 0.9810030460357666, 'reward_repeat_floor_std': 0.007477888371795416, 'reward_near_duplicate_penalty_mean': 0.9733536243438721, 'reward_near_duplicate_penalty_std': 0.04985261708498001, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.643750011920929, 'reward_judge_quality_std': 0.40074706077575684, 'reward_total_composite_mean': 0.4706975817680359, 'reward_total_composite_std': 0.4523340165615082, 'epoch': 0.03}

 20%|█▉        | 651/3300 [2:25:05<40:02:03, 54.41s/it]INFO 04-12 18:23:31 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 652/3300 [2:25:12<29:46:53, 40.49s/it]2026-04-12 18:23:39,195 | INFO | train_grpo_train | metrics_logged mode=train step=652

                                                       
{'loss': 0.0121, 'grad_norm': 17.63026237487793, 'learning_rate': 8.027272727272728e-06, 'num_tokens': 1197362.0, 'completions/mean_length': 43.625, 'completions/min_length': 39.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.13537968695163727, 'rewards/meter/std': 0.10082083940505981, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45500001311302185, 'rewards/judge_quality/std': 0.31883716583251953, 'rewards/repeat_penalty/mean': 0.9507007002830505, 'rewards/repeat_penalty/std': 0.054731372743844986, 'rewards/repeat_floor/mean': 0.9936259984970093, 'rewards/repeat_floor/std': 0.006548299919813871, 'rewards/near_duplicate_penalty/mean': 0.9693484306335449, 'rewards/near_duplicate_penalty/std': 0.02757902257144451, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9802631139755249, 'rewards/distinct_2/std': 0.036561544984579086, 'rewards/total_composite/mean': 0.07122999429702759, 'rewards/total_composite/std': 0.07895796746015549, 'reward': 0.07122999429702759, 'reward_std': 0.07895796746015549, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1790056824684143, 'sampling/sampling_logp_difference/max': 1.7705345153808594, 'sampling/importance_sampling_ratio/min': 0.17024196684360504, 'sampling/importance_sampling_ratio/mean': 1.005855917930603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9491819888353348, 'clip_ratio/low_mean': 0.10563959507271647, 'clip_ratio/low_min': 0.10563959507271647, 'clip_ratio/high_mean': 0.03306686133146286, 'clip_ratio/high_max': 0.03306686133146286, 'clip_ratio/region_mean': 0.13870645640417933, 'reward_total_mean': 0.07122999429702759, 'reward_meter_mean': 0.13537968695163727, 'reward_meter_std': 0.10082083940505981, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9507007002830505, 'reward_repeat_penalty_std': 0.054731372743844986, 'reward_repeat_floor_mean': 0.9936259984970093, 'reward_repeat_floor_std': 0.006548299919813871, 'reward_near_duplicate_penalty_mean': 0.9693484306335449, 'reward_near_duplicate_penalty_std': 0.02757902257144451, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9802631139755249, 'reward_distinct_2_std': 0.036561544984579086, 'reward_judge_quality_mean': 0.45500001311302185, 'reward_judge_quality_std': 0.31883716583251953, 'reward_total_composite_mean': 0.07122999429702759, 'reward_total_composite_std': 0.07895796746015549, 'epoch': 0.03}

 20%|█▉        | 652/3300 [2:25:13<29:46:53, 40.49s/it]INFO 04-12 18:23:39 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 653/3300 [2:25:21<22:39:58, 30.83s/it]2026-04-12 18:23:47,479 | INFO | train_grpo_train | metrics_logged mode=train step=653

                                                       
{'loss': 0.0321, 'grad_norm': 21.393150329589844, 'learning_rate': 8.024242424242425e-06, 'num_tokens': 1198911.0, 'completions/mean_length': 38.625, 'completions/min_length': 35.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8580807447433472, 'rewards/meter/std': 0.11206099390983582, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7900000214576721, 'rewards/judge_quality/std': 0.24219238758087158, 'rewards/repeat_penalty/mean': 0.8240015506744385, 'rewards/repeat_penalty/std': 0.1576167792081833, 'rewards/repeat_floor/mean': 0.9794058203697205, 'rewards/repeat_floor/std': 0.015769105404615402, 'rewards/near_duplicate_penalty/mean': 0.915531575679779, 'rewards/near_duplicate_penalty/std': 0.058078352361917496, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9536231160163879, 'rewards/distinct_2/std': 0.0537121407687664, 'rewards/total_composite/mean': 0.6526845097541809, 'rewards/total_composite/std': 0.1951036900281906, 'reward': 0.6526845097541809, 'reward_std': 0.1951036900281906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1395138055086136, 'sampling/sampling_logp_difference/max': 1.578580617904663, 'sampling/importance_sampling_ratio/min': 0.20626765489578247, 'sampling/importance_sampling_ratio/mean': 0.9674643278121948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7222080230712891, 'clip_ratio/low_mean': 0.03467680676840246, 'clip_ratio/low_min': 0.03467680676840246, 'clip_ratio/high_mean': 0.08485455485060811, 'clip_ratio/high_max': 0.08485455485060811, 'clip_ratio/region_mean': 0.11953136161901057, 'reward_total_mean': 0.6526845097541809, 'reward_meter_mean': 0.8580807447433472, 'reward_meter_std': 0.11206099390983582, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8240015506744385, 'reward_repeat_penalty_std': 0.1576167792081833, 'reward_repeat_floor_mean': 0.9794058203697205, 'reward_repeat_floor_std': 0.015769105404615402, 'reward_near_duplicate_penalty_mean': 0.915531575679779, 'reward_near_duplicate_penalty_std': 0.058078352361917496, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9536231160163879, 'reward_distinct_2_std': 0.0537121407687664, 'reward_judge_quality_mean': 0.7900000214576721, 'reward_judge_quality_std': 0.24219238758087158, 'reward_total_composite_mean': 0.6526845097541809, 'reward_total_composite_std': 0.1951036900281906, 'epoch': 0.03}

 20%|█▉        | 653/3300 [2:25:21<22:39:58, 30.83s/it]INFO 04-12 18:23:47 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 654/3300 [2:25:30<17:54:00, 24.35s/it]2026-04-12 18:23:56,729 | INFO | train_grpo_train | metrics_logged mode=train step=654

                                                       
{'loss': 0.0812, 'grad_norm': 16.97374153137207, 'learning_rate': 8.021212121212122e-06, 'num_tokens': 1200514.0, 'completions/mean_length': 44.375, 'completions/min_length': 39.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7214658260345459, 'rewards/meter/std': 0.38603147864341736, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.731249988079071, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.9975759387016296, 'rewards/repeat_penalty/std': 0.004586003255099058, 'rewards/repeat_floor/mean': 0.9996364116668701, 'rewards/repeat_floor/std': 0.0006879009888507426, 'rewards/near_duplicate_penalty/mean': 0.9975759387016296, 'rewards/near_duplicate_penalty/std': 0.004586003255099058, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5430784821510315, 'rewards/total_composite/std': 0.36221033334732056, 'reward': 0.5430784821510315, 'reward_std': 0.36221033334732056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15669618546962738, 'sampling/sampling_logp_difference/max': 2.4407830238342285, 'sampling/importance_sampling_ratio/min': 0.08709262311458588, 'sampling/importance_sampling_ratio/mean': 0.9790434837341309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8368480801582336, 'clip_ratio/low_mean': 0.0603094594553113, 'clip_ratio/low_min': 0.0603094594553113, 'clip_ratio/high_mean': 0.06950508616864681, 'clip_ratio/high_max': 0.06950508616864681, 'clip_ratio/region_mean': 0.1298145456239581, 'reward_total_mean': 0.5430784821510315, 'reward_meter_mean': 0.7214658260345459, 'reward_meter_std': 0.38603147864341736, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9975759387016296, 'reward_repeat_penalty_std': 0.004586003255099058, 'reward_repeat_floor_mean': 0.9996364116668701, 'reward_repeat_floor_std': 0.0006879009888507426, 'reward_near_duplicate_penalty_mean': 0.9975759387016296, 'reward_near_duplicate_penalty_std': 0.004586003255099058, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.731249988079071, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.5430784821510315, 'reward_total_composite_std': 0.36221033334732056, 'epoch': 0.03}

 20%|█▉        | 654/3300 [2:25:30<17:54:00, 24.35s/it]INFO 04-12 18:23:56 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 655/3300 [2:25:39<14:35:55, 19.87s/it]2026-04-12 18:24:06,132 | INFO | train_grpo_train | metrics_logged mode=train step=655

                                                       
{'loss': 0.0361, 'grad_norm': 15.985350608825684, 'learning_rate': 8.018181818181818e-06, 'num_tokens': 1202280.0, 'completions/mean_length': 49.75, 'completions/min_length': 43.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6333403587341309, 'rewards/meter/std': 0.47195398807525635, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.9566726684570312, 'rewards/repeat_penalty/std': 0.0907059758901596, 'rewards/repeat_floor/mean': 0.9943370819091797, 'rewards/repeat_floor/std': 0.012737623415887356, 'rewards/near_duplicate_penalty/mean': 0.9732609987258911, 'rewards/near_duplicate_penalty/std': 0.07100581377744675, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9816433191299438, 'rewards/distinct_2/std': 0.03404124453663826, 'rewards/total_composite/mean': 0.3853573203086853, 'rewards/total_composite/std': 0.35687434673309326, 'reward': 0.3853573203086853, 'reward_std': 0.35687434673309326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.174941286444664, 'sampling/sampling_logp_difference/max': 2.0234832763671875, 'sampling/importance_sampling_ratio/min': 0.13219419121742249, 'sampling/importance_sampling_ratio/mean': 1.038133144378662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9822410717606544, 'clip_ratio/low_mean': 0.08968568500131369, 'clip_ratio/low_min': 0.08968568500131369, 'clip_ratio/high_mean': 0.06355504505336285, 'clip_ratio/high_max': 0.06355504505336285, 'clip_ratio/region_mean': 0.15324073005467653, 'reward_total_mean': 0.3853573203086853, 'reward_meter_mean': 0.6333403587341309, 'reward_meter_std': 0.47195398807525635, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9566726684570312, 'reward_repeat_penalty_std': 0.0907059758901596, 'reward_repeat_floor_mean': 0.9943370819091797, 'reward_repeat_floor_std': 0.012737623415887356, 'reward_near_duplicate_penalty_mean': 0.9732609987258911, 'reward_near_duplicate_penalty_std': 0.07100581377744675, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9816433191299438, 'reward_distinct_2_std': 0.03404124453663826, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.3853573203086853, 'reward_total_composite_std': 0.35687434673309326, 'epoch': 0.03}

 20%|█▉        | 655/3300 [2:25:39<14:35:55, 19.87s/it]INFO 04-12 18:24:06 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 656/3300 [2:25:50<12:32:08, 17.07s/it]2026-04-12 18:24:16,669 | INFO | train_grpo_train | metrics_logged mode=train step=656

                                                       
{'loss': 0.0567, 'grad_norm': 19.898828506469727, 'learning_rate': 8.015151515151515e-06, 'num_tokens': 1204389.0, 'completions/mean_length': 75.625, 'completions/min_length': 62.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.7478793859481812, 'rewards/meter/std': 0.34890758991241455, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.04008918255567551, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9390250444412231, 'rewards/repeat_penalty/std': 0.08215352892875671, 'rewards/repeat_floor/mean': 0.9923896789550781, 'rewards/repeat_floor/std': 0.009498625062406063, 'rewards/near_duplicate_penalty/mean': 0.9681970477104187, 'rewards/near_duplicate_penalty/std': 0.029523510485887527, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9684461951255798, 'rewards/distinct_2/std': 0.058879781514406204, 'rewards/total_composite/mean': 0.28587985038757324, 'rewards/total_composite/std': 0.13996566832065582, 'reward': 0.28587985038757324, 'reward_std': 0.13996566832065582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12708786129951477, 'sampling/sampling_logp_difference/max': 2.4431800842285156, 'sampling/importance_sampling_ratio/min': 0.08688411116600037, 'sampling/importance_sampling_ratio/mean': 0.9986262321472168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6703781709074974, 'clip_ratio/low_mean': 0.03569597098976374, 'clip_ratio/low_min': 0.03569597098976374, 'clip_ratio/high_mean': 0.10775001719594002, 'clip_ratio/high_max': 0.10775001719594002, 'clip_ratio/region_mean': 0.14344598818570375, 'reward_total_mean': 0.28587985038757324, 'reward_meter_mean': 0.7478793859481812, 'reward_meter_std': 0.34890758991241455, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.04008918255567551, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9390250444412231, 'reward_repeat_penalty_std': 0.08215352892875671, 'reward_repeat_floor_mean': 0.9923896789550781, 'reward_repeat_floor_std': 0.009498625062406063, 'reward_near_duplicate_penalty_mean': 0.9681970477104187, 'reward_near_duplicate_penalty_std': 0.029523510485887527, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9684461951255798, 'reward_distinct_2_std': 0.058879781514406204, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.28587985038757324, 'reward_total_composite_std': 0.13996566832065582, 'epoch': 0.03}

 20%|█▉        | 656/3300 [2:25:50<12:32:08, 17.07s/it]INFO 04-12 18:24:16 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 657/3300 [2:26:02<11:27:07, 15.60s/it]2026-04-12 18:24:28,836 | INFO | train_grpo_train | metrics_logged mode=train step=657

                                                       
{'loss': 0.02, 'grad_norm': 7.60475492477417, 'learning_rate': 8.012121212121214e-06, 'num_tokens': 1207249.0, 'completions/mean_length': 156.5, 'completions/min_length': 142.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 156.5, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.9543006420135498, 'rewards/meter/std': 0.031204799190163612, 'rewards/count_adherence/mean': 0.8928571939468384, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9678571224212646, 'rewards/count_floor/std': 0.01983901485800743, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.5310271382331848, 'rewards/repeat_penalty/std': 0.2053688019514084, 'rewards/repeat_floor/mean': 0.9489916563034058, 'rewards/repeat_floor/std': 0.024536121636629105, 'rewards/near_duplicate_penalty/mean': 0.8843667507171631, 'rewards/near_duplicate_penalty/std': 0.052286289632320404, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.625962495803833, 'rewards/distinct_2/std': 0.19003194570541382, 'rewards/total_composite/mean': 0.3189767003059387, 'rewards/total_composite/std': 0.09191308915615082, 'reward': 0.3189767003059387, 'reward_std': 0.09191308170557022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12917010486125946, 'sampling/sampling_logp_difference/max': 3.227461814880371, 'sampling/importance_sampling_ratio/min': 0.03965802863240242, 'sampling/importance_sampling_ratio/mean': 0.9982758164405823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6592458188533783, 'clip_ratio/low_mean': 0.04193478776142001, 'clip_ratio/low_min': 0.04193478776142001, 'clip_ratio/high_mean': 0.06408673897385597, 'clip_ratio/high_max': 0.06408673897385597, 'clip_ratio/region_mean': 0.10602152673527598, 'reward_total_mean': 0.3189767003059387, 'reward_meter_mean': 0.9543006420135498, 'reward_meter_std': 0.031204799190163612, 'reward_count_adherence_mean': 0.8928571939468384, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9678571224212646, 'reward_count_floor_std': 0.01983901485800743, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5310271382331848, 'reward_repeat_penalty_std': 0.2053688019514084, 'reward_repeat_floor_mean': 0.9489916563034058, 'reward_repeat_floor_std': 0.024536121636629105, 'reward_near_duplicate_penalty_mean': 0.8843667507171631, 'reward_near_duplicate_penalty_std': 0.052286289632320404, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.625962495803833, 'reward_distinct_2_std': 0.19003194570541382, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.3189767003059387, 'reward_total_composite_std': 0.09191308915615082, 'epoch': 0.03}

 20%|█▉        | 657/3300 [2:26:02<11:27:07, 15.60s/it]INFO 04-12 18:24:29 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 658/3300 [2:26:11<9:57:10, 13.56s/it] 2026-04-12 18:24:37,644 | INFO | train_grpo_train | metrics_logged mode=train step=658

                                                      
{'loss': 0.0536, 'grad_norm': 11.497108459472656, 'learning_rate': 8.00909090909091e-06, 'num_tokens': 1209149.0, 'completions/mean_length': 69.5, 'completions/min_length': 46.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.26256340742111206, 'rewards/meter/std': 0.252474844455719, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.2522328794002533, 'rewards/repeat_penalty/mean': 0.8825502991676331, 'rewards/repeat_penalty/std': 0.11305183917284012, 'rewards/repeat_floor/mean': 0.9860538244247437, 'rewards/repeat_floor/std': 0.013264486566185951, 'rewards/near_duplicate_penalty/mean': 0.953132152557373, 'rewards/near_duplicate_penalty/std': 0.038152847439050674, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9231554865837097, 'rewards/distinct_2/std': 0.08683669567108154, 'rewards/total_composite/mean': 0.10729870200157166, 'rewards/total_composite/std': 0.07448439300060272, 'reward': 0.10729870200157166, 'reward_std': 0.07448439300060272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15363150835037231, 'sampling/sampling_logp_difference/max': 1.8525729179382324, 'sampling/importance_sampling_ratio/min': 0.15683312714099884, 'sampling/importance_sampling_ratio/mean': 1.017725944519043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9719163253903389, 'clip_ratio/low_mean': 0.0824795812368393, 'clip_ratio/low_min': 0.0824795812368393, 'clip_ratio/high_mean': 0.07663781940937042, 'clip_ratio/high_max': 0.07663781940937042, 'clip_ratio/region_mean': 0.15911740064620972, 'reward_total_mean': 0.10729870200157166, 'reward_meter_mean': 0.26256340742111206, 'reward_meter_std': 0.252474844455719, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8825502991676331, 'reward_repeat_penalty_std': 0.11305183917284012, 'reward_repeat_floor_mean': 0.9860538244247437, 'reward_repeat_floor_std': 0.013264486566185951, 'reward_near_duplicate_penalty_mean': 0.953132152557373, 'reward_near_duplicate_penalty_std': 0.038152847439050674, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9231554865837097, 'reward_distinct_2_std': 0.08683669567108154, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.2522328794002533, 'reward_total_composite_mean': 0.10729870200157166, 'reward_total_composite_std': 0.07448439300060272, 'epoch': 0.03}

 20%|█▉        | 658/3300 [2:26:11<9:57:10, 13.56s/it]INFO 04-12 18:24:37 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 659/3300 [2:26:20<8:54:51, 12.15s/it]2026-04-12 18:24:46,500 | INFO | train_grpo_train | metrics_logged mode=train step=659

                                                      
{'loss': 0.0404, 'grad_norm': 8.06682300567627, 'learning_rate': 8.006060606060607e-06, 'num_tokens': 1211519.0, 'completions/mean_length': 98.25, 'completions/min_length': 93.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.25, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.8991388082504272, 'rewards/meter/std': 0.20196320116519928, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.8384352326393127, 'rewards/repeat_penalty/std': 0.07862620055675507, 'rewards/repeat_floor/mean': 0.9821255803108215, 'rewards/repeat_floor/std': 0.008544066920876503, 'rewards/near_duplicate_penalty/mean': 0.9539580345153809, 'rewards/near_duplicate_penalty/std': 0.0216944869607687, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8781321048736572, 'rewards/distinct_2/std': 0.07006164640188217, 'rewards/total_composite/mean': 0.4534206986427307, 'rewards/total_composite/std': 0.2501770257949829, 'reward': 0.4534206986427307, 'reward_std': 0.2501770555973053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14626868069171906, 'sampling/sampling_logp_difference/max': 2.8476009368896484, 'sampling/importance_sampling_ratio/min': 0.05798326060175896, 'sampling/importance_sampling_ratio/mean': 1.0070701837539673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7126376032829285, 'clip_ratio/low_mean': 0.0903792018070817, 'clip_ratio/low_min': 0.0903792018070817, 'clip_ratio/high_mean': 0.03862307220697403, 'clip_ratio/high_max': 0.03862307220697403, 'clip_ratio/region_mean': 0.12900227401405573, 'reward_total_mean': 0.4534206986427307, 'reward_meter_mean': 0.8991388082504272, 'reward_meter_std': 0.20196320116519928, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8384352326393127, 'reward_repeat_penalty_std': 0.07862620055675507, 'reward_repeat_floor_mean': 0.9821255803108215, 'reward_repeat_floor_std': 0.008544066920876503, 'reward_near_duplicate_penalty_mean': 0.9539580345153809, 'reward_near_duplicate_penalty_std': 0.0216944869607687, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8781321048736572, 'reward_distinct_2_std': 0.07006164640188217, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.4534206986427307, 'reward_total_composite_std': 0.2501770257949829, 'epoch': 0.03}

 20%|█▉        | 659/3300 [2:26:20<8:54:51, 12.15s/it]INFO 04-12 18:24:46 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 660/3300 [2:26:33<9:11:15, 12.53s/it]2026-04-12 18:24:59,914 | INFO | train_grpo_train | metrics_logged mode=train step=660

                                                      
{'loss': -0.0491, 'grad_norm': 5.477097511291504, 'learning_rate': 8.003030303030304e-06, 'num_tokens': 1213052.0, 'completions/mean_length': 96.625, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.28571701049805, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5391556024551392, 'rewards/meter/std': 0.48435178399086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9576091766357422, 'rewards/lexical_plausibility/std': 0.11989942193031311, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.3486069440841675, 'rewards/repeat_penalty/mean': 0.9880307912826538, 'rewards/repeat_penalty/std': 0.018162276595830917, 'rewards/repeat_floor/mean': 0.9982045888900757, 'rewards/repeat_floor/std': 0.0027243425138294697, 'rewards/near_duplicate_penalty/mean': 0.9880307912826538, 'rewards/near_duplicate_penalty/std': 0.018162276595830917, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4118770956993103, 'rewards/total_composite/std': 0.4354664981365204, 'reward': 0.4118770956993103, 'reward_std': 0.4354664981365204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17071758210659027, 'sampling/sampling_logp_difference/max': 1.2555642127990723, 'sampling/importance_sampling_ratio/min': 0.2887059152126312, 'sampling/importance_sampling_ratio/mean': 1.0242204666137695, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.218810312449932, 'clip_ratio/low_mean': 0.0668847281485796, 'clip_ratio/low_min': 0.0668847281485796, 'clip_ratio/high_mean': 0.08562177047133446, 'clip_ratio/high_max': 0.08562177047133446, 'clip_ratio/region_mean': 0.15250649861991405, 'reward_total_mean': 0.4118770956993103, 'reward_meter_mean': 0.5391556024551392, 'reward_meter_std': 0.48435178399086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9576091766357422, 'reward_lexical_plausibility_std': 0.11989942193031311, 'reward_repeat_penalty_mean': 0.9880307912826538, 'reward_repeat_penalty_std': 0.018162276595830917, 'reward_repeat_floor_mean': 0.9982045888900757, 'reward_repeat_floor_std': 0.0027243425138294697, 'reward_near_duplicate_penalty_mean': 0.9880307912826538, 'reward_near_duplicate_penalty_std': 0.018162276595830917, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.3486069440841675, 'reward_total_composite_mean': 0.4118770956993103, 'reward_total_composite_std': 0.4354664981365204, 'epoch': 0.03}

 20%|██        | 660/3300 [2:26:33<9:11:15, 12.53s/it]INFO 04-12 18:25:00 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 661/3300 [2:26:43<8:30:23, 11.60s/it]2026-04-12 18:25:09,360 | INFO | train_grpo_train | metrics_logged mode=train step=661

                                                      
{'loss': 0.005, 'grad_norm': 12.345144271850586, 'learning_rate': 8.000000000000001e-06, 'num_tokens': 1214979.0, 'completions/mean_length': 66.875, 'completions/min_length': 58.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.37197738885879517, 'rewards/meter/std': 0.30547118186950684, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.11259915679693222, 'rewards/repeat_penalty/mean': 0.868878960609436, 'rewards/repeat_penalty/std': 0.20396563410758972, 'rewards/repeat_floor/mean': 0.9853951334953308, 'rewards/repeat_floor/std': 0.020186908543109894, 'rewards/near_duplicate_penalty/mean': 0.9496092200279236, 'rewards/near_duplicate_penalty/std': 0.04659753665328026, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9217078685760498, 'rewards/distinct_2/std': 0.14931339025497437, 'rewards/total_composite/mean': 0.14806479215621948, 'rewards/total_composite/std': 0.14574910700321198, 'reward': 0.14806479215621948, 'reward_std': 0.14574910700321198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1938203126192093, 'sampling/sampling_logp_difference/max': 2.2001328468322754, 'sampling/importance_sampling_ratio/min': 0.11078843474388123, 'sampling/importance_sampling_ratio/mean': 0.9971727728843689, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0106062814593315, 'clip_ratio/low_mean': 0.1024552658200264, 'clip_ratio/low_min': 0.1024552658200264, 'clip_ratio/high_mean': 0.07076810393482447, 'clip_ratio/high_max': 0.07076810393482447, 'clip_ratio/region_mean': 0.17322336975485086, 'reward_total_mean': 0.14806479215621948, 'reward_meter_mean': 0.37197738885879517, 'reward_meter_std': 0.30547118186950684, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.868878960609436, 'reward_repeat_penalty_std': 0.20396563410758972, 'reward_repeat_floor_mean': 0.9853951334953308, 'reward_repeat_floor_std': 0.020186908543109894, 'reward_near_duplicate_penalty_mean': 0.9496092200279236, 'reward_near_duplicate_penalty_std': 0.04659753665328026, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9217078685760498, 'reward_distinct_2_std': 0.14931339025497437, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.11259915679693222, 'reward_total_composite_mean': 0.14806479215621948, 'reward_total_composite_std': 0.14574910700321198, 'epoch': 0.03}

 20%|██        | 661/3300 [2:26:43<8:30:23, 11.60s/it]INFO 04-12 18:25:09 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 662/3300 [2:26:51<7:48:05, 10.65s/it]2026-04-12 18:25:17,772 | INFO | train_grpo_train | metrics_logged mode=train step=662

                                                      
{'loss': 0.0818, 'grad_norm': 19.352556228637695, 'learning_rate': 7.996969696969697e-06, 'num_tokens': 1216304.0, 'completions/mean_length': 19.625, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.7257665991783142, 'rewards/meter/std': 0.36524754762649536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.2737569510936737, 'rewards/repeat_penalty/mean': 0.7424530386924744, 'rewards/repeat_penalty/std': 0.01428307592868805, 'rewards/repeat_floor/mean': 0.9834905862808228, 'rewards/repeat_floor/std': 0.0028566198889166117, 'rewards/near_duplicate_penalty/mean': 0.9899373650550842, 'rewards/near_duplicate_penalty/std': 0.019044118002057076, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.36046797037124634, 'rewards/total_composite/std': 0.2924782633781433, 'reward': 0.36046797037124634, 'reward_std': 0.2924782633781433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1191195622086525, 'sampling/sampling_logp_difference/max': 1.4371376037597656, 'sampling/importance_sampling_ratio/min': 0.23760691285133362, 'sampling/importance_sampling_ratio/mean': 1.0074163675308228, 'sampling/importance_sampling_ratio/max': 1.9916117191314697, 'entropy': 0.7898770570755005, 'clip_ratio/low_mean': 0.0730695123784244, 'clip_ratio/low_min': 0.0730695123784244, 'clip_ratio/high_mean': 0.029389881063252687, 'clip_ratio/high_max': 0.029389881063252687, 'clip_ratio/region_mean': 0.1024593934416771, 'reward_total_mean': 0.36046797037124634, 'reward_meter_mean': 0.7257665991783142, 'reward_meter_std': 0.36524754762649536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7424530386924744, 'reward_repeat_penalty_std': 0.01428307592868805, 'reward_repeat_floor_mean': 0.9834905862808228, 'reward_repeat_floor_std': 0.0028566198889166117, 'reward_near_duplicate_penalty_mean': 0.9899373650550842, 'reward_near_duplicate_penalty_std': 0.019044118002057076, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.2737569510936737, 'reward_total_composite_mean': 0.36046797037124634, 'reward_total_composite_std': 0.2924782633781433, 'epoch': 0.03}

 20%|██        | 662/3300 [2:26:51<7:48:05, 10.65s/it]INFO 04-12 18:25:18 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 663/3300 [2:27:00<7:22:04, 10.06s/it]2026-04-12 18:25:26,456 | INFO | train_grpo_train | metrics_logged mode=train step=663

                                                      
{'loss': 0.1183, 'grad_norm': 11.3187894821167, 'learning_rate': 7.993939393939396e-06, 'num_tokens': 1218200.0, 'completions/mean_length': 71.0, 'completions/min_length': 46.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.7428855299949646, 'rewards/meter/std': 0.24329309165477753, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.05580177903175354, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390502154827118, 'rewards/repeat_penalty/mean': 0.7145324349403381, 'rewards/repeat_penalty/std': 0.2663004994392395, 'rewards/repeat_floor/mean': 0.9650921821594238, 'rewards/repeat_floor/std': 0.032483767718076706, 'rewards/near_duplicate_penalty/mean': 0.8942155241966248, 'rewards/near_duplicate_penalty/std': 0.07663911581039429, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.7936513423919678, 'rewards/distinct_2/std': 0.22877711057662964, 'rewards/total_composite/mean': 0.3045590817928314, 'rewards/total_composite/std': 0.16305504739284515, 'reward': 0.3045590817928314, 'reward_std': 0.16305504739284515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1332339197397232, 'sampling/sampling_logp_difference/max': 1.6979243755340576, 'sampling/importance_sampling_ratio/min': 0.1830631047487259, 'sampling/importance_sampling_ratio/mean': 1.0138534307479858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8701349049806595, 'clip_ratio/low_mean': 0.04047085391357541, 'clip_ratio/low_min': 0.04047085391357541, 'clip_ratio/high_mean': 0.09771819226443768, 'clip_ratio/high_max': 0.09771819226443768, 'clip_ratio/region_mean': 0.1381890461780131, 'reward_total_mean': 0.3045590817928314, 'reward_meter_mean': 0.7428855299949646, 'reward_meter_std': 0.24329309165477753, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.05580177903175354, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7145324349403381, 'reward_repeat_penalty_std': 0.2663004994392395, 'reward_repeat_floor_mean': 0.9650921821594238, 'reward_repeat_floor_std': 0.032483767718076706, 'reward_near_duplicate_penalty_mean': 0.8942155241966248, 'reward_near_duplicate_penalty_std': 0.07663911581039429, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.7936513423919678, 'reward_distinct_2_std': 0.22877711057662964, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390502154827118, 'reward_total_composite_mean': 0.3045590817928314, 'reward_total_composite_std': 0.16305504739284515, 'epoch': 0.03}

 20%|██        | 663/3300 [2:27:00<7:22:04, 10.06s/it]INFO 04-12 18:25:26 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 664/3300 [2:27:14<8:18:40, 11.35s/it]2026-04-12 18:25:40,825 | INFO | train_grpo_train | metrics_logged mode=train step=664

                                                      
{'loss': -0.0174, 'grad_norm': 6.1547770500183105, 'learning_rate': 7.990909090909091e-06, 'num_tokens': 1220048.0, 'completions/mean_length': 130.0, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 75.42857360839844, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.4293666481971741, 'rewards/meter/std': 0.3880077302455902, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.04008918255567551, 'rewards/lexical_plausibility/mean': 0.9557690620422363, 'rewards/lexical_plausibility/std': 0.1251039057970047, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1922609806060791, 'rewards/repeat_penalty/mean': 0.9673750400543213, 'rewards/repeat_penalty/std': 0.07846193760633469, 'rewards/repeat_floor/mean': 0.9960489273071289, 'rewards/repeat_floor/std': 0.009111514315009117, 'rewards/near_duplicate_penalty/mean': 0.9854937195777893, 'rewards/near_duplicate_penalty/std': 0.02757568471133709, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9802761077880859, 'rewards/distinct_2/std': 0.05578751116991043, 'rewards/total_composite/mean': 0.21620672941207886, 'rewards/total_composite/std': 0.2470053881406784, 'reward': 0.21620672941207886, 'reward_std': 0.2470053732395172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17725621163845062, 'sampling/sampling_logp_difference/max': 1.9277033805847168, 'sampling/importance_sampling_ratio/min': 0.14548194408416748, 'sampling/importance_sampling_ratio/mean': 0.9795347452163696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8164736106991768, 'clip_ratio/low_mean': 0.06796471774578094, 'clip_ratio/low_min': 0.06796471774578094, 'clip_ratio/high_mean': 0.05850048363208771, 'clip_ratio/high_max': 0.05850048363208771, 'clip_ratio/region_mean': 0.12646520137786865, 'reward_total_mean': 0.21620672941207886, 'reward_meter_mean': 0.4293666481971741, 'reward_meter_std': 0.3880077302455902, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.04008918255567551, 'reward_lexical_plausibility_mean': 0.9557690620422363, 'reward_lexical_plausibility_std': 0.1251039057970047, 'reward_repeat_penalty_mean': 0.9673750400543213, 'reward_repeat_penalty_std': 0.07846193760633469, 'reward_repeat_floor_mean': 0.9960489273071289, 'reward_repeat_floor_std': 0.009111514315009117, 'reward_near_duplicate_penalty_mean': 0.9854937195777893, 'reward_near_duplicate_penalty_std': 0.02757568471133709, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9802761077880859, 'reward_distinct_2_std': 0.05578751116991043, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1922609806060791, 'reward_total_composite_mean': 0.21620672941207886, 'reward_total_composite_std': 0.2470053881406784, 'epoch': 0.03}

 20%|██        | 664/3300 [2:27:14<8:18:40, 11.35s/it]INFO 04-12 18:25:41 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 665/3300 [2:27:29<9:00:43, 12.31s/it]2026-04-12 18:25:55,382 | INFO | train_grpo_train | metrics_logged mode=train step=665

                                                      
{'loss': -0.0639, 'grad_norm': 2.945300579071045, 'learning_rate': 7.987878787878789e-06, 'num_tokens': 1221582.0, 'completions/mean_length': 155.75, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.23700128495693207, 'rewards/meter/std': 0.32466304302215576, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9007634520530701, 'rewards/lexical_plausibility/std': 0.1845458298921585, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.3991396725177765, 'rewards/repeat_penalty/mean': 0.9618940353393555, 'rewards/repeat_penalty/std': 0.09047530591487885, 'rewards/repeat_floor/mean': 0.9970256090164185, 'rewards/repeat_floor/std': 0.005870083346962929, 'rewards/near_duplicate_penalty/mean': 0.9926705360412598, 'rewards/near_duplicate_penalty/std': 0.0076141818426549435, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.06932878494262695, 'rewards/total_composite/std': 0.06364665925502777, 'reward': 0.06932878494262695, 'reward_std': 0.06364665180444717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12925459444522858, 'sampling/sampling_logp_difference/max': 1.3852972984313965, 'sampling/importance_sampling_ratio/min': 0.25024938583374023, 'sampling/importance_sampling_ratio/mean': 1.0308729410171509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6449528858065605, 'clip_ratio/low_mean': 0.048448532819747925, 'clip_ratio/low_min': 0.048448532819747925, 'clip_ratio/high_mean': 0.058286977000534534, 'clip_ratio/high_max': 0.058286977000534534, 'clip_ratio/region_mean': 0.10673550982028246, 'reward_total_mean': 0.06932878494262695, 'reward_meter_mean': 0.23700128495693207, 'reward_meter_std': 0.32466304302215576, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9007634520530701, 'reward_lexical_plausibility_std': 0.1845458298921585, 'reward_repeat_penalty_mean': 0.9618940353393555, 'reward_repeat_penalty_std': 0.09047530591487885, 'reward_repeat_floor_mean': 0.9970256090164185, 'reward_repeat_floor_std': 0.005870083346962929, 'reward_near_duplicate_penalty_mean': 0.9926705360412598, 'reward_near_duplicate_penalty_std': 0.0076141818426549435, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.3991396725177765, 'reward_total_composite_mean': 0.06932878494262695, 'reward_total_composite_std': 0.06364665925502777, 'epoch': 0.03}

 20%|██        | 665/3300 [2:27:29<9:00:43, 12.31s/it]INFO 04-12 18:25:55 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 666/3300 [2:27:43<9:22:28, 12.81s/it]2026-04-12 18:26:09,363 | INFO | train_grpo_train | metrics_logged mode=train step=666

                                                      
{'loss': -0.0735, 'grad_norm': 6.352017879486084, 'learning_rate': 7.984848484848486e-06, 'num_tokens': 1223199.0, 'completions/mean_length': 105.125, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.000003814697266, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5255693197250366, 'rewards/meter/std': 0.35454121232032776, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9167780876159668, 'rewards/lexical_plausibility/std': 0.160339817404747, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.8807878494262695, 'rewards/repeat_penalty/std': 0.16998426616191864, 'rewards/repeat_floor/mean': 0.986284077167511, 'rewards/repeat_floor/std': 0.019304540008306503, 'rewards/near_duplicate_penalty/mean': 0.9594157934188843, 'rewards/near_duplicate_penalty/std': 0.049862559884786606, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9360747933387756, 'rewards/distinct_2/std': 0.08389385044574738, 'rewards/total_composite/mean': 0.20649249851703644, 'rewards/total_composite/std': 0.14781621098518372, 'reward': 0.20649249851703644, 'reward_std': 0.14781621098518372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15539716184139252, 'sampling/sampling_logp_difference/max': 1.267641544342041, 'sampling/importance_sampling_ratio/min': 0.28149473667144775, 'sampling/importance_sampling_ratio/mean': 1.0176290273666382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.932106114923954, 'clip_ratio/low_mean': 0.05587620520964265, 'clip_ratio/low_min': 0.05587620520964265, 'clip_ratio/high_mean': 0.054037779569625854, 'clip_ratio/high_max': 0.054037779569625854, 'clip_ratio/region_mean': 0.1099139847792685, 'reward_total_mean': 0.20649249851703644, 'reward_meter_mean': 0.5255693197250366, 'reward_meter_std': 0.35454121232032776, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9167780876159668, 'reward_lexical_plausibility_std': 0.160339817404747, 'reward_repeat_penalty_mean': 0.8807878494262695, 'reward_repeat_penalty_std': 0.16998426616191864, 'reward_repeat_floor_mean': 0.986284077167511, 'reward_repeat_floor_std': 0.019304540008306503, 'reward_near_duplicate_penalty_mean': 0.9594157934188843, 'reward_near_duplicate_penalty_std': 0.049862559884786606, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9360747933387756, 'reward_distinct_2_std': 0.08389385044574738, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.20649249851703644, 'reward_total_composite_std': 0.14781621098518372, 'epoch': 0.04}

 20%|██        | 666/3300 [2:27:43<9:22:28, 12.81s/it]INFO 04-12 18:26:09 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 667/3300 [2:27:51<8:24:56, 11.51s/it]2026-04-12 18:26:17,823 | INFO | train_grpo_train | metrics_logged mode=train step=667

                                                      
{'loss': 0.0458, 'grad_norm': 15.202601432800293, 'learning_rate': 7.981818181818183e-06, 'num_tokens': 1224678.0, 'completions/mean_length': 43.875, 'completions/min_length': 36.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.34886667132377625, 'rewards/meter/std': 0.3690391182899475, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8224999904632568, 'rewards/judge_quality/std': 0.2149917185306549, 'rewards/repeat_penalty/mean': 0.9940017461776733, 'rewards/repeat_penalty/std': 0.008982398547232151, 'rewards/repeat_floor/mean': 0.9991002082824707, 'rewards/repeat_floor/std': 0.0013473600847646594, 'rewards/near_duplicate_penalty/mean': 0.9940017461776733, 'rewards/near_duplicate_penalty/std': 0.008982398547232151, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23189206421375275, 'rewards/total_composite/std': 0.2210950255393982, 'reward': 0.23189206421375275, 'reward_std': 0.2210950255393982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15299059450626373, 'sampling/sampling_logp_difference/max': 2.9912333488464355, 'sampling/importance_sampling_ratio/min': 0.050225455313920975, 'sampling/importance_sampling_ratio/mean': 1.007691740989685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8653842359781265, 'clip_ratio/low_mean': 0.05927872471511364, 'clip_ratio/low_min': 0.05927872471511364, 'clip_ratio/high_mean': 0.0809075040742755, 'clip_ratio/high_max': 0.0809075040742755, 'clip_ratio/region_mean': 0.14018622878938913, 'reward_total_mean': 0.23189206421375275, 'reward_meter_mean': 0.34886667132377625, 'reward_meter_std': 0.3690391182899475, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9940017461776733, 'reward_repeat_penalty_std': 0.008982398547232151, 'reward_repeat_floor_mean': 0.9991002082824707, 'reward_repeat_floor_std': 0.0013473600847646594, 'reward_near_duplicate_penalty_mean': 0.9940017461776733, 'reward_near_duplicate_penalty_std': 0.008982398547232151, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8224999904632568, 'reward_judge_quality_std': 0.2149917185306549, 'reward_total_composite_mean': 0.23189206421375275, 'reward_total_composite_std': 0.2210950255393982, 'epoch': 0.04}

 20%|██        | 667/3300 [2:27:51<8:24:56, 11.51s/it]INFO 04-12 18:26:18 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 668/3300 [2:28:05<8:50:46, 12.10s/it]2026-04-12 18:26:31,310 | INFO | train_grpo_train | metrics_logged mode=train step=668

                                                      
{'loss': -0.0703, 'grad_norm': 4.15318489074707, 'learning_rate': 7.978787878787879e-06, 'num_tokens': 1226224.0, 'completions/mean_length': 101.25, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.57143020629883, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9759743213653564, 'rewards/meter/std': 0.017037374898791313, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9503446817398071, 'rewards/lexical_plausibility/std': 0.14044636487960815, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.30875036120414734, 'rewards/repeat_penalty/mean': 0.9090670347213745, 'rewards/repeat_penalty/std': 0.10325290262699127, 'rewards/repeat_floor/mean': 0.9900355935096741, 'rewards/repeat_floor/std': 0.010023010894656181, 'rewards/near_duplicate_penalty/mean': 0.9582163095474243, 'rewards/near_duplicate_penalty/std': 0.03900682181119919, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.05725561082363129, 'rewards/total_composite/mean': 0.43802177906036377, 'rewards/total_composite/std': 0.2764167785644531, 'reward': 0.43802177906036377, 'reward_std': 0.2764167785644531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13032785058021545, 'sampling/sampling_logp_difference/max': 1.1505379676818848, 'sampling/importance_sampling_ratio/min': 0.31646648049354553, 'sampling/importance_sampling_ratio/mean': 1.0114279985427856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7039266526699066, 'clip_ratio/low_mean': 0.0689329276792705, 'clip_ratio/low_min': 0.0689329276792705, 'clip_ratio/high_mean': 0.04873995203524828, 'clip_ratio/high_max': 0.04873995203524828, 'clip_ratio/region_mean': 0.11767287971451879, 'reward_total_mean': 0.43802177906036377, 'reward_meter_mean': 0.9759743213653564, 'reward_meter_std': 0.017037374898791313, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9503446817398071, 'reward_lexical_plausibility_std': 0.14044636487960815, 'reward_repeat_penalty_mean': 0.9090670347213745, 'reward_repeat_penalty_std': 0.10325290262699127, 'reward_repeat_floor_mean': 0.9900355935096741, 'reward_repeat_floor_std': 0.010023010894656181, 'reward_near_duplicate_penalty_mean': 0.9582163095474243, 'reward_near_duplicate_penalty_std': 0.03900682181119919, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.05725561082363129, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.30875036120414734, 'reward_total_composite_mean': 0.43802177906036377, 'reward_total_composite_std': 0.2764167785644531, 'epoch': 0.04}

 20%|██        | 668/3300 [2:28:05<8:50:46, 12.10s/it]INFO 04-12 18:26:31 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 669/3300 [2:28:14<8:12:43, 11.24s/it]2026-04-12 18:26:40,529 | INFO | train_grpo_train | metrics_logged mode=train step=669

                                                      
{'loss': 0.0085, 'grad_norm': 11.956597328186035, 'learning_rate': 7.975757575757576e-06, 'num_tokens': 1228293.0, 'completions/mean_length': 65.625, 'completions/min_length': 55.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.6452656984329224, 'rewards/meter/std': 0.3540770709514618, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9240579009056091, 'rewards/repeat_penalty/std': 0.1628832221031189, 'rewards/repeat_floor/mean': 0.990271270275116, 'rewards/repeat_floor/std': 0.0208668764680624, 'rewards/near_duplicate_penalty/mean': 0.964903712272644, 'rewards/near_duplicate_penalty/std': 0.06764344871044159, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9503968358039856, 'rewards/distinct_2/std': 0.11965266615152359, 'rewards/total_composite/mean': 0.3070964515209198, 'rewards/total_composite/std': 0.23158514499664307, 'reward': 0.3070964515209198, 'reward_std': 0.23158514499664307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1716596931219101, 'sampling/sampling_logp_difference/max': 1.6606016159057617, 'sampling/importance_sampling_ratio/min': 0.19002461433410645, 'sampling/importance_sampling_ratio/mean': 1.0029100179672241, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0608155205845833, 'clip_ratio/low_mean': 0.07844012696295977, 'clip_ratio/low_min': 0.07844012696295977, 'clip_ratio/high_mean': 0.09136004652827978, 'clip_ratio/high_max': 0.09136004652827978, 'clip_ratio/region_mean': 0.16980017349123955, 'reward_total_mean': 0.3070964515209198, 'reward_meter_mean': 0.6452656984329224, 'reward_meter_std': 0.3540770709514618, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9240579009056091, 'reward_repeat_penalty_std': 0.1628832221031189, 'reward_repeat_floor_mean': 0.990271270275116, 'reward_repeat_floor_std': 0.0208668764680624, 'reward_near_duplicate_penalty_mean': 0.964903712272644, 'reward_near_duplicate_penalty_std': 0.06764344871044159, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9503968358039856, 'reward_distinct_2_std': 0.11965266615152359, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.3070964515209198, 'reward_total_composite_std': 0.23158514499664307, 'epoch': 0.04}

 20%|██        | 669/3300 [2:28:14<8:12:43, 11.24s/it]INFO 04-12 18:26:40 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 670/3300 [2:28:22<7:37:06, 10.43s/it]2026-04-12 18:26:49,073 | INFO | train_grpo_train | metrics_logged mode=train step=670

                                                      
{'loss': 0.0393, 'grad_norm': 17.505361557006836, 'learning_rate': 7.972727272727273e-06, 'num_tokens': 1229799.0, 'completions/mean_length': 27.25, 'completions/min_length': 23.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.8154100179672241, 'rewards/meter/std': 0.3197248876094818, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.27969372272491455, 'rewards/repeat_penalty/mean': 0.862553596496582, 'rewards/repeat_penalty/std': 0.1551109403371811, 'rewards/repeat_floor/mean': 0.984084963798523, 'rewards/repeat_floor/std': 0.016064923256635666, 'rewards/near_duplicate_penalty/mean': 0.9442561864852905, 'rewards/near_duplicate_penalty/std': 0.04105212911963463, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9369060397148132, 'rewards/distinct_2/std': 0.09271392226219177, 'rewards/total_composite/mean': 0.4442387819290161, 'rewards/total_composite/std': 0.2854630947113037, 'reward': 0.4442387819290161, 'reward_std': 0.2854630649089813, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15387241542339325, 'sampling/sampling_logp_difference/max': 2.130800724029541, 'sampling/importance_sampling_ratio/min': 0.1851322054862976, 'sampling/importance_sampling_ratio/mean': 1.0311537981033325, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5915060341358185, 'clip_ratio/low_mean': 0.11391397751867771, 'clip_ratio/low_min': 0.11391397751867771, 'clip_ratio/high_mean': 0.03840579837560654, 'clip_ratio/high_max': 0.03840579837560654, 'clip_ratio/region_mean': 0.15231977589428425, 'reward_total_mean': 0.4442387819290161, 'reward_meter_mean': 0.8154100179672241, 'reward_meter_std': 0.3197248876094818, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.862553596496582, 'reward_repeat_penalty_std': 0.1551109403371811, 'reward_repeat_floor_mean': 0.984084963798523, 'reward_repeat_floor_std': 0.016064923256635666, 'reward_near_duplicate_penalty_mean': 0.9442561864852905, 'reward_near_duplicate_penalty_std': 0.04105212911963463, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9369060397148132, 'reward_distinct_2_std': 0.09271392226219177, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.27969372272491455, 'reward_total_composite_mean': 0.4442387819290161, 'reward_total_composite_std': 0.2854630947113037, 'epoch': 0.04}

 20%|██        | 670/3300 [2:28:22<7:37:06, 10.43s/it]INFO 04-12 18:26:49 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 671/3300 [2:28:36<8:18:58, 11.39s/it]2026-04-12 18:27:02,699 | INFO | train_grpo_train | metrics_logged mode=train step=671

                                                      
{'loss': -0.0711, 'grad_norm': 6.52949333190918, 'learning_rate': 7.96969696969697e-06, 'num_tokens': 1231529.0, 'completions/mean_length': 112.25, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.142860412597656, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6141881346702576, 'rewards/meter/std': 0.3707980513572693, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9398908019065857, 'rewards/lexical_plausibility/std': 0.17001453042030334, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.892742931842804, 'rewards/repeat_penalty/std': 0.1296996772289276, 'rewards/repeat_floor/mean': 0.9892590045928955, 'rewards/repeat_floor/std': 0.009078697301447392, 'rewards/near_duplicate_penalty/mean': 0.958815336227417, 'rewards/near_duplicate_penalty/std': 0.020125899463891983, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9492966532707214, 'rewards/distinct_2/std': 0.07004464417695999, 'rewards/total_composite/mean': 0.2518628239631653, 'rewards/total_composite/std': 0.15881061553955078, 'reward': 0.2518628239631653, 'reward_std': 0.15881061553955078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16067443788051605, 'sampling/sampling_logp_difference/max': 2.0222206115722656, 'sampling/importance_sampling_ratio/min': 0.13236121833324432, 'sampling/importance_sampling_ratio/mean': 1.0222755670547485, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9634770527482033, 'clip_ratio/low_mean': 0.0635985555127263, 'clip_ratio/low_min': 0.0635985555127263, 'clip_ratio/high_mean': 0.08093888964504004, 'clip_ratio/high_max': 0.08093888964504004, 'clip_ratio/region_mean': 0.14453744515776634, 'reward_total_mean': 0.2518628239631653, 'reward_meter_mean': 0.6141881346702576, 'reward_meter_std': 0.3707980513572693, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9398908019065857, 'reward_lexical_plausibility_std': 0.17001453042030334, 'reward_repeat_penalty_mean': 0.892742931842804, 'reward_repeat_penalty_std': 0.1296996772289276, 'reward_repeat_floor_mean': 0.9892590045928955, 'reward_repeat_floor_std': 0.009078697301447392, 'reward_near_duplicate_penalty_mean': 0.958815336227417, 'reward_near_duplicate_penalty_std': 0.020125899463891983, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9492966532707214, 'reward_distinct_2_std': 0.07004464417695999, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.2518628239631653, 'reward_total_composite_std': 0.15881061553955078, 'epoch': 0.04}

 20%|██        | 671/3300 [2:28:36<8:18:58, 11.39s/it]INFO 04-12 18:27:02 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 672/3300 [2:28:45<7:41:41, 10.54s/it]2026-04-12 18:27:11,264 | INFO | train_grpo_train | metrics_logged mode=train step=672

                                                      
{'loss': 0.0048, 'grad_norm': 14.852743148803711, 'learning_rate': 7.966666666666668e-06, 'num_tokens': 1233204.0, 'completions/mean_length': 50.375, 'completions/min_length': 45.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.708685040473938, 'rewards/meter/std': 0.2926980257034302, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.2794605791568756, 'rewards/repeat_penalty/mean': 0.9886763095855713, 'rewards/repeat_penalty/std': 0.008414451964199543, 'rewards/repeat_floor/mean': 0.9983013868331909, 'rewards/repeat_floor/std': 0.001262169098481536, 'rewards/near_duplicate_penalty/mean': 0.9886763095855713, 'rewards/near_duplicate_penalty/std': 0.008414451964199543, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4904243052005768, 'rewards/total_composite/std': 0.2901962101459503, 'reward': 0.4904243052005768, 'reward_std': 0.2901962399482727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14661242067813873, 'sampling/sampling_logp_difference/max': 1.7159264087677002, 'sampling/importance_sampling_ratio/min': 0.17979709804058075, 'sampling/importance_sampling_ratio/mean': 1.0275498628616333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7641573995351791, 'clip_ratio/low_mean': 0.10392376780509949, 'clip_ratio/low_min': 0.10392376780509949, 'clip_ratio/high_mean': 0.05837436392903328, 'clip_ratio/high_max': 0.05837436392903328, 'clip_ratio/region_mean': 0.16229813173413277, 'reward_total_mean': 0.4904243052005768, 'reward_meter_mean': 0.708685040473938, 'reward_meter_std': 0.2926980257034302, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9886763095855713, 'reward_repeat_penalty_std': 0.008414451964199543, 'reward_repeat_floor_mean': 0.9983013868331909, 'reward_repeat_floor_std': 0.001262169098481536, 'reward_near_duplicate_penalty_mean': 0.9886763095855713, 'reward_near_duplicate_penalty_std': 0.008414451964199543, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.2794605791568756, 'reward_total_composite_mean': 0.4904243052005768, 'reward_total_composite_std': 0.2901962101459503, 'epoch': 0.04}

 20%|██        | 672/3300 [2:28:45<7:41:41, 10.54s/it]INFO 04-12 18:27:11 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 673/3300 [2:28:52<7:04:59,  9.71s/it]2026-04-12 18:27:19,033 | INFO | train_grpo_train | metrics_logged mode=train step=673

                                                      
{'loss': -0.0381, 'grad_norm': 20.308509826660156, 'learning_rate': 7.963636363636365e-06, 'num_tokens': 1234615.0, 'completions/mean_length': 23.375, 'completions/min_length': 19.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6716791987419128, 'rewards/meter/std': 0.44118791818618774, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9635416269302368, 'rewards/lexical_plausibility/std': 0.05970953777432442, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3296508193016052, 'rewards/repeat_penalty/mean': 0.7041395902633667, 'rewards/repeat_penalty/std': 0.08856705576181412, 'rewards/repeat_floor/mean': 0.9758279323577881, 'rewards/repeat_floor/std': 0.01771341636776924, 'rewards/near_duplicate_penalty/mean': 0.9388527870178223, 'rewards/near_duplicate_penalty/std': 0.11808940768241882, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4570765197277069, 'rewards/total_composite/std': 0.3698100447654724, 'reward': 0.4570765197277069, 'reward_std': 0.36981001496315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16632592678070068, 'sampling/sampling_logp_difference/max': 1.8005461692810059, 'sampling/importance_sampling_ratio/min': 0.165208637714386, 'sampling/importance_sampling_ratio/mean': 1.0142191648483276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1466046571731567, 'clip_ratio/low_mean': 0.08267162414267659, 'clip_ratio/low_min': 0.08267162414267659, 'clip_ratio/high_mean': 0.04982286598533392, 'clip_ratio/high_max': 0.04982286598533392, 'clip_ratio/region_mean': 0.1324944901280105, 'reward_total_mean': 0.4570765197277069, 'reward_meter_mean': 0.6716791987419128, 'reward_meter_std': 0.44118791818618774, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9635416269302368, 'reward_lexical_plausibility_std': 0.05970953777432442, 'reward_repeat_penalty_mean': 0.7041395902633667, 'reward_repeat_penalty_std': 0.08856705576181412, 'reward_repeat_floor_mean': 0.9758279323577881, 'reward_repeat_floor_std': 0.01771341636776924, 'reward_near_duplicate_penalty_mean': 0.9388527870178223, 'reward_near_duplicate_penalty_std': 0.11808940768241882, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3296508193016052, 'reward_total_composite_mean': 0.4570765197277069, 'reward_total_composite_std': 0.3698100447654724, 'epoch': 0.04}

 20%|██        | 673/3300 [2:28:52<7:04:59,  9.71s/it]INFO 04-12 18:27:19 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 674/3300 [2:29:07<8:05:22, 11.09s/it]2026-04-12 18:27:33,342 | INFO | train_grpo_train | metrics_logged mode=train step=674

                                                      
{'loss': -0.0736, 'grad_norm': 5.6915693283081055, 'learning_rate': 7.96060606060606e-06, 'num_tokens': 1236181.0, 'completions/mean_length': 105.75, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.71428680419922, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8384084701538086, 'rewards/meter/std': 0.31080061197280884, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9239732027053833, 'rewards/lexical_plausibility/std': 0.2150362730026245, 'rewards/judge_quality/mean': 0.6424999833106995, 'rewards/judge_quality/std': 0.3556382656097412, 'rewards/repeat_penalty/mean': 0.9916764497756958, 'rewards/repeat_penalty/std': 0.011654396541416645, 'rewards/repeat_floor/mean': 0.9987514615058899, 'rewards/repeat_floor/std': 0.0017481568502262235, 'rewards/near_duplicate_penalty/mean': 0.9916764497756958, 'rewards/near_duplicate_penalty/std': 0.011654396541416645, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5048578977584839, 'rewards/total_composite/std': 0.369978129863739, 'reward': 0.5048578977584839, 'reward_std': 0.369978129863739, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12830685079097748, 'sampling/sampling_logp_difference/max': 1.4864661693572998, 'sampling/importance_sampling_ratio/min': 0.22617049515247345, 'sampling/importance_sampling_ratio/mean': 1.0071234703063965, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8284262157976627, 'clip_ratio/low_mean': 0.07619910221546888, 'clip_ratio/low_min': 0.07619910221546888, 'clip_ratio/high_mean': 0.05265436973422766, 'clip_ratio/high_max': 0.05265436973422766, 'clip_ratio/region_mean': 0.12885347194969654, 'reward_total_mean': 0.5048578977584839, 'reward_meter_mean': 0.8384084701538086, 'reward_meter_std': 0.31080061197280884, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9239732027053833, 'reward_lexical_plausibility_std': 0.2150362730026245, 'reward_repeat_penalty_mean': 0.9916764497756958, 'reward_repeat_penalty_std': 0.011654396541416645, 'reward_repeat_floor_mean': 0.9987514615058899, 'reward_repeat_floor_std': 0.0017481568502262235, 'reward_near_duplicate_penalty_mean': 0.9916764497756958, 'reward_near_duplicate_penalty_std': 0.011654396541416645, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6424999833106995, 'reward_judge_quality_std': 0.3556382656097412, 'reward_total_composite_mean': 0.5048578977584839, 'reward_total_composite_std': 0.369978129863739, 'epoch': 0.04}

 20%|██        | 674/3300 [2:29:07<8:05:22, 11.09s/it]INFO 04-12 18:27:33 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 675/3300 [2:29:21<8:51:17, 12.14s/it]2026-04-12 18:27:47,944 | INFO | train_grpo_train | metrics_logged mode=train step=675

                                                      
{'loss': -0.1414, 'grad_norm': 4.438897132873535, 'learning_rate': 7.957575757575758e-06, 'num_tokens': 1238134.0, 'completions/mean_length': 135.125, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.28572082519531, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7689068913459778, 'rewards/meter/std': 0.3179732859134674, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.2086307406425476, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.06258923560380936, 'rewards/lexical_plausibility/mean': 0.9577388763427734, 'rewards/lexical_plausibility/std': 0.11953240633010864, 'rewards/judge_quality/mean': 0.5012499690055847, 'rewards/judge_quality/std': 0.2806338369846344, 'rewards/repeat_penalty/mean': 0.8501274585723877, 'rewards/repeat_penalty/std': 0.11462368071079254, 'rewards/repeat_floor/mean': 0.9848185181617737, 'rewards/repeat_floor/std': 0.011386548168957233, 'rewards/near_duplicate_penalty/mean': 0.9630175232887268, 'rewards/near_duplicate_penalty/std': 0.030681267380714417, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.9189959764480591, 'rewards/distinct_2/std': 0.08283697813749313, 'rewards/total_composite/mean': 0.38563209772109985, 'rewards/total_composite/std': 0.23907926678657532, 'reward': 0.38563209772109985, 'reward_std': 0.23907926678657532, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13266560435295105, 'sampling/sampling_logp_difference/max': 1.4092607498168945, 'sampling/importance_sampling_ratio/min': 0.24432383477687836, 'sampling/importance_sampling_ratio/mean': 1.0034555196762085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7219468057155609, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.11606972105801105, 'clip_ratio/high_max': 0.11606972105801105, 'clip_ratio/region_mean': 0.13169472105801105, 'reward_total_mean': 0.38563209772109985, 'reward_meter_mean': 0.7689068913459778, 'reward_meter_std': 0.3179732859134674, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.2086307406425476, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.06258923560380936, 'reward_lexical_plausibility_mean': 0.9577388763427734, 'reward_lexical_plausibility_std': 0.11953240633010864, 'reward_repeat_penalty_mean': 0.8501274585723877, 'reward_repeat_penalty_std': 0.11462368071079254, 'reward_repeat_floor_mean': 0.9848185181617737, 'reward_repeat_floor_std': 0.011386548168957233, 'reward_near_duplicate_penalty_mean': 0.9630175232887268, 'reward_near_duplicate_penalty_std': 0.030681267380714417, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.9189959764480591, 'reward_distinct_2_std': 0.08283697813749313, 'reward_judge_quality_mean': 0.5012499690055847, 'reward_judge_quality_std': 0.2806338369846344, 'reward_total_composite_mean': 0.38563209772109985, 'reward_total_composite_std': 0.23907926678657532, 'epoch': 0.04}

 20%|██        | 675/3300 [2:29:21<8:51:17, 12.14s/it]INFO 04-12 18:27:48 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 676/3300 [2:29:32<8:33:29, 11.74s/it]2026-04-12 18:27:58,748 | INFO | train_grpo_train | metrics_logged mode=train step=676

                                                      
{'loss': 0.0206, 'grad_norm': 7.080842018127441, 'learning_rate': 7.954545454545455e-06, 'num_tokens': 1241062.0, 'completions/mean_length': 186.0, 'completions/min_length': 158.0, 'completions/max_length': 210.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.0, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 210.0, 'rewards/meter/mean': 0.3150366544723511, 'rewards/meter/std': 0.21612221002578735, 'rewards/count_adherence/mean': 0.930555522441864, 'rewards/count_adherence/std': 0.08266931027173996, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9791666865348816, 'rewards/count_floor/std': 0.024800799787044525, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.27125000953674316, 'rewards/judge_quality/std': 0.2713162302970886, 'rewards/repeat_penalty/mean': 0.1303415447473526, 'rewards/repeat_penalty/std': 0.2778764069080353, 'rewards/repeat_floor/mean': 0.8890479207038879, 'rewards/repeat_floor/std': 0.04458707943558693, 'rewards/near_duplicate_penalty/mean': 0.7908244132995605, 'rewards/near_duplicate_penalty/std': 0.13213153183460236, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.14499200880527496, 'rewards/distinct_2/std': 0.2861138582229614, 'rewards/total_composite/mean': 0.08844363689422607, 'rewards/total_composite/std': 0.11938827484846115, 'reward': 0.08844363689422607, 'reward_std': 0.11938827484846115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10392284393310547, 'sampling/sampling_logp_difference/max': 7.570887088775635, 'sampling/importance_sampling_ratio/min': 0.0005152351805008948, 'sampling/importance_sampling_ratio/mean': 1.0054212808609009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5497882030904293, 'clip_ratio/low_mean': 0.0376412901096046, 'clip_ratio/low_min': 0.0376412901096046, 'clip_ratio/high_mean': 0.0315225669182837, 'clip_ratio/high_max': 0.0315225669182837, 'clip_ratio/region_mean': 0.0691638570278883, 'reward_total_mean': 0.08844363689422607, 'reward_meter_mean': 0.3150366544723511, 'reward_meter_std': 0.21612221002578735, 'reward_count_adherence_mean': 0.930555522441864, 'reward_count_adherence_std': 0.08266931027173996, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9791666865348816, 'reward_count_floor_std': 0.024800799787044525, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.1303415447473526, 'reward_repeat_penalty_std': 0.2778764069080353, 'reward_repeat_floor_mean': 0.8890479207038879, 'reward_repeat_floor_std': 0.04458707943558693, 'reward_near_duplicate_penalty_mean': 0.7908244132995605, 'reward_near_duplicate_penalty_std': 0.13213153183460236, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.14499200880527496, 'reward_distinct_2_std': 0.2861138582229614, 'reward_judge_quality_mean': 0.27125000953674316, 'reward_judge_quality_std': 0.2713162302970886, 'reward_total_composite_mean': 0.08844363689422607, 'reward_total_composite_std': 0.11938827484846115, 'epoch': 0.04}

 20%|██        | 676/3300 [2:29:32<8:33:29, 11.74s/it]INFO 04-12 18:27:58 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 677/3300 [2:29:40<7:49:15, 10.73s/it]2026-04-12 18:28:07,132 | INFO | train_grpo_train | metrics_logged mode=train step=677

                                                      
{'loss': 0.051, 'grad_norm': 17.90215492248535, 'learning_rate': 7.951515151515152e-06, 'num_tokens': 1242564.0, 'completions/mean_length': 39.75, 'completions/min_length': 34.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8788691759109497, 'rewards/meter/std': 0.14421789348125458, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8062499761581421, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/repeat_penalty/mean': 0.9963901042938232, 'rewards/repeat_penalty/std': 0.005558798089623451, 'rewards/repeat_floor/mean': 0.9994585514068604, 'rewards/repeat_floor/std': 0.0008338202023878694, 'rewards/near_duplicate_penalty/mean': 0.9963901042938232, 'rewards/near_duplicate_penalty/std': 0.005558798089623451, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6976757049560547, 'rewards/total_composite/std': 0.2032840996980667, 'reward': 0.6976757049560547, 'reward_std': 0.2032841145992279, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15710601210594177, 'sampling/sampling_logp_difference/max': 1.5675959587097168, 'sampling/importance_sampling_ratio/min': 0.20854593813419342, 'sampling/importance_sampling_ratio/mean': 0.999754786491394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8704234212636948, 'clip_ratio/low_mean': 0.04006779193878174, 'clip_ratio/low_min': 0.04006779193878174, 'clip_ratio/high_mean': 0.10357406362891197, 'clip_ratio/high_max': 0.10357406362891197, 'clip_ratio/region_mean': 0.1436418555676937, 'reward_total_mean': 0.6976757049560547, 'reward_meter_mean': 0.8788691759109497, 'reward_meter_std': 0.14421789348125458, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9963901042938232, 'reward_repeat_penalty_std': 0.005558798089623451, 'reward_repeat_floor_mean': 0.9994585514068604, 'reward_repeat_floor_std': 0.0008338202023878694, 'reward_near_duplicate_penalty_mean': 0.9963901042938232, 'reward_near_duplicate_penalty_std': 0.005558798089623451, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8062499761581421, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.6976757049560547, 'reward_total_composite_std': 0.2032840996980667, 'epoch': 0.04}

 21%|██        | 677/3300 [2:29:40<7:49:15, 10.73s/it]INFO 04-12 18:28:07 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 678/3300 [2:29:51<7:40:30, 10.54s/it]2026-04-12 18:28:17,212 | INFO | train_grpo_train | metrics_logged mode=train step=678

                                                      
{'loss': -0.0058, 'grad_norm': 14.17061996459961, 'learning_rate': 7.948484848484848e-06, 'num_tokens': 1244675.0, 'completions/mean_length': 86.875, 'completions/min_length': 73.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.875, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.4399971663951874, 'rewards/meter/std': 0.31285810470581055, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.8435863256454468, 'rewards/repeat_penalty/std': 0.17496798932552338, 'rewards/repeat_floor/mean': 0.9819164276123047, 'rewards/repeat_floor/std': 0.021712088957428932, 'rewards/near_duplicate_penalty/mean': 0.9476750493049622, 'rewards/near_duplicate_penalty/std': 0.061347994953393936, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.9123215675354004, 'rewards/distinct_2/std': 0.09288833290338516, 'rewards/total_composite/mean': 0.14434631168842316, 'rewards/total_composite/std': 0.1120925322175026, 'reward': 0.14434631168842316, 'reward_std': 0.1120925322175026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19196632504463196, 'sampling/sampling_logp_difference/max': 2.207411289215088, 'sampling/importance_sampling_ratio/min': 0.10998500138521194, 'sampling/importance_sampling_ratio/mean': 1.0036708116531372, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2945523113012314, 'clip_ratio/low_mean': 0.07068119570612907, 'clip_ratio/low_min': 0.07068119570612907, 'clip_ratio/high_mean': 0.10044151172041893, 'clip_ratio/high_max': 0.10044151172041893, 'clip_ratio/region_mean': 0.171122707426548, 'reward_total_mean': 0.14434631168842316, 'reward_meter_mean': 0.4399971663951874, 'reward_meter_std': 0.31285810470581055, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.8435863256454468, 'reward_repeat_penalty_std': 0.17496798932552338, 'reward_repeat_floor_mean': 0.9819164276123047, 'reward_repeat_floor_std': 0.021712088957428932, 'reward_near_duplicate_penalty_mean': 0.9476750493049622, 'reward_near_duplicate_penalty_std': 0.061347994953393936, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.9123215675354004, 'reward_distinct_2_std': 0.09288833290338516, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.14434631168842316, 'reward_total_composite_std': 0.1120925322175026, 'epoch': 0.04}

 21%|██        | 678/3300 [2:29:51<7:40:30, 10.54s/it]INFO 04-12 18:28:17 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 679/3300 [2:30:00<7:26:23, 10.22s/it]2026-04-12 18:28:26,684 | INFO | train_grpo_train | metrics_logged mode=train step=679

                                                      
{'loss': 0.0172, 'grad_norm': 14.929986000061035, 'learning_rate': 7.945454545454547e-06, 'num_tokens': 1246508.0, 'completions/mean_length': 48.125, 'completions/min_length': 41.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7493664026260376, 'rewards/meter/std': 0.3446935713291168, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.23308108747005463, 'rewards/repeat_penalty/mean': 0.9263632893562317, 'rewards/repeat_penalty/std': 0.06771992146968842, 'rewards/repeat_floor/mean': 0.9913603067398071, 'rewards/repeat_floor/std': 0.00851268321275711, 'rewards/near_duplicate_penalty/mean': 0.9698535203933716, 'rewards/near_duplicate_penalty/std': 0.03847035765647888, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9542471170425415, 'rewards/distinct_2/std': 0.038344256579875946, 'rewards/total_composite/mean': 0.3571673631668091, 'rewards/total_composite/std': 0.27758970856666565, 'reward': 0.3571673631668091, 'reward_std': 0.27758967876434326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17775818705558777, 'sampling/sampling_logp_difference/max': 1.6910442113876343, 'sampling/importance_sampling_ratio/min': 0.18432694673538208, 'sampling/importance_sampling_ratio/mean': 1.0141254663467407, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2454250678420067, 'clip_ratio/low_mean': 0.14154711179435253, 'clip_ratio/low_min': 0.14154711179435253, 'clip_ratio/high_mean': 0.056379870511591434, 'clip_ratio/high_max': 0.056379870511591434, 'clip_ratio/region_mean': 0.19792698230594397, 'reward_total_mean': 0.3571673631668091, 'reward_meter_mean': 0.7493664026260376, 'reward_meter_std': 0.3446935713291168, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9263632893562317, 'reward_repeat_penalty_std': 0.06771992146968842, 'reward_repeat_floor_mean': 0.9913603067398071, 'reward_repeat_floor_std': 0.00851268321275711, 'reward_near_duplicate_penalty_mean': 0.9698535203933716, 'reward_near_duplicate_penalty_std': 0.03847035765647888, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9542471170425415, 'reward_distinct_2_std': 0.038344256579875946, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.23308108747005463, 'reward_total_composite_mean': 0.3571673631668091, 'reward_total_composite_std': 0.27758970856666565, 'epoch': 0.04}

 21%|██        | 679/3300 [2:30:00<7:26:23, 10.22s/it]INFO 04-12 18:28:26 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 680/3300 [2:30:09<7:12:32,  9.91s/it]2026-04-12 18:28:35,857 | INFO | train_grpo_train | metrics_logged mode=train step=680

                                                      
{'loss': 0.1259, 'grad_norm': 12.149139404296875, 'learning_rate': 7.942424242424242e-06, 'num_tokens': 1248589.0, 'completions/mean_length': 72.125, 'completions/min_length': 54.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9162817597389221, 'rewards/meter/std': 0.10908091813325882, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.04008918255567551, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6212499737739563, 'rewards/judge_quality/std': 0.2704724967479706, 'rewards/repeat_penalty/mean': 0.9279419183731079, 'rewards/repeat_penalty/std': 0.09104763716459274, 'rewards/repeat_floor/mean': 0.9915457963943481, 'rewards/repeat_floor/std': 0.01038378570228815, 'rewards/near_duplicate_penalty/mean': 0.9716588258743286, 'rewards/near_duplicate_penalty/std': 0.029558012261986732, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9532989859580994, 'rewards/distinct_2/std': 0.06751921027898788, 'rewards/total_composite/mean': 0.5090329647064209, 'rewards/total_composite/std': 0.2099083662033081, 'reward': 0.5090329647064209, 'reward_std': 0.2099083513021469, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1614697277545929, 'sampling/sampling_logp_difference/max': 1.658233642578125, 'sampling/importance_sampling_ratio/min': 0.19047513604164124, 'sampling/importance_sampling_ratio/mean': 1.0135349035263062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.034816287457943, 'clip_ratio/low_mean': 0.06443071644753218, 'clip_ratio/low_min': 0.06443071644753218, 'clip_ratio/high_mean': 0.09485700912773609, 'clip_ratio/high_max': 0.09485700912773609, 'clip_ratio/region_mean': 0.15928772557526827, 'reward_total_mean': 0.5090329647064209, 'reward_meter_mean': 0.9162817597389221, 'reward_meter_std': 0.10908091813325882, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.04008918255567551, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9279419183731079, 'reward_repeat_penalty_std': 0.09104763716459274, 'reward_repeat_floor_mean': 0.9915457963943481, 'reward_repeat_floor_std': 0.01038378570228815, 'reward_near_duplicate_penalty_mean': 0.9716588258743286, 'reward_near_duplicate_penalty_std': 0.029558012261986732, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9532989859580994, 'reward_distinct_2_std': 0.06751921027898788, 'reward_judge_quality_mean': 0.6212499737739563, 'reward_judge_quality_std': 0.2704724967479706, 'reward_total_composite_mean': 0.5090329647064209, 'reward_total_composite_std': 0.2099083662033081, 'epoch': 0.04}

 21%|██        | 680/3300 [2:30:09<7:12:32,  9.91s/it]INFO 04-12 18:28:36 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 681/3300 [2:30:18<6:52:30,  9.45s/it]2026-04-12 18:28:44,246 | INFO | train_grpo_train | metrics_logged mode=train step=681

                                                      
{'loss': 0.0554, 'grad_norm': 17.26517105102539, 'learning_rate': 7.93939393939394e-06, 'num_tokens': 1250063.0, 'completions/mean_length': 31.25, 'completions/min_length': 28.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.5412427186965942, 'rewards/meter/std': 0.4185452461242676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.231084406375885, 'rewards/repeat_penalty/mean': 0.9840534925460815, 'rewards/repeat_penalty/std': 0.02079745940864086, 'rewards/repeat_floor/mean': 0.9976080656051636, 'rewards/repeat_floor/std': 0.0031196200288832188, 'rewards/near_duplicate_penalty/mean': 0.9840534925460815, 'rewards/near_duplicate_penalty/std': 0.02079745940864086, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4303084909915924, 'rewards/total_composite/std': 0.351150780916214, 'reward': 0.4303084909915924, 'reward_std': 0.351150780916214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14856000244617462, 'sampling/sampling_logp_difference/max': 1.445095181465149, 'sampling/importance_sampling_ratio/min': 0.23572365939617157, 'sampling/importance_sampling_ratio/mean': 1.0016511678695679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6426580920815468, 'clip_ratio/low_mean': 0.09051120653748512, 'clip_ratio/low_min': 0.09051120653748512, 'clip_ratio/high_mean': 0.06935483776032925, 'clip_ratio/high_max': 0.06935483776032925, 'clip_ratio/region_mean': 0.15986604429781437, 'reward_total_mean': 0.4303084909915924, 'reward_meter_mean': 0.5412427186965942, 'reward_meter_std': 0.4185452461242676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9840534925460815, 'reward_repeat_penalty_std': 0.02079745940864086, 'reward_repeat_floor_mean': 0.9976080656051636, 'reward_repeat_floor_std': 0.0031196200288832188, 'reward_near_duplicate_penalty_mean': 0.9840534925460815, 'reward_near_duplicate_penalty_std': 0.02079745940864086, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.231084406375885, 'reward_total_composite_mean': 0.4303084909915924, 'reward_total_composite_std': 0.351150780916214, 'epoch': 0.04}

 21%|██        | 681/3300 [2:30:18<6:52:30,  9.45s/it]INFO 04-12 18:28:44 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 682/3300 [2:30:27<6:49:31,  9.39s/it]2026-04-12 18:28:53,478 | INFO | train_grpo_train | metrics_logged mode=train step=682

                                                      
{'loss': 0.1233, 'grad_norm': 13.69466781616211, 'learning_rate': 7.936363636363637e-06, 'num_tokens': 1251682.0, 'completions/mean_length': 42.375, 'completions/min_length': 34.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5714253783226013, 'rewards/meter/std': 0.40181785821914673, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9871336221694946, 'rewards/repeat_penalty/std': 0.013903828337788582, 'rewards/repeat_floor/mean': 0.9980700016021729, 'rewards/repeat_floor/std': 0.0020855737384408712, 'rewards/near_duplicate_penalty/mean': 0.9871336221694946, 'rewards/near_duplicate_penalty/std': 0.013903828337788582, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21997764706611633, 'rewards/total_composite/std': 0.15484604239463806, 'reward': 0.21997764706611633, 'reward_std': 0.15484604239463806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14972510933876038, 'sampling/sampling_logp_difference/max': 3.900174617767334, 'sampling/importance_sampling_ratio/min': 0.020238377153873444, 'sampling/importance_sampling_ratio/mean': 0.9987576603889465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6500987969338894, 'clip_ratio/low_mean': 0.04316326603293419, 'clip_ratio/low_min': 0.04316326603293419, 'clip_ratio/high_mean': 0.10029766196385026, 'clip_ratio/high_max': 0.10029766196385026, 'clip_ratio/region_mean': 0.14346092799678445, 'reward_total_mean': 0.21997764706611633, 'reward_meter_mean': 0.5714253783226013, 'reward_meter_std': 0.40181785821914673, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9871336221694946, 'reward_repeat_penalty_std': 0.013903828337788582, 'reward_repeat_floor_mean': 0.9980700016021729, 'reward_repeat_floor_std': 0.0020855737384408712, 'reward_near_duplicate_penalty_mean': 0.9871336221694946, 'reward_near_duplicate_penalty_std': 0.013903828337788582, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.21997764706611633, 'reward_total_composite_std': 0.15484604239463806, 'epoch': 0.04}

 21%|██        | 682/3300 [2:30:27<6:49:31,  9.39s/it]INFO 04-12 18:28:53 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 683/3300 [2:30:36<6:40:42,  9.19s/it]2026-04-12 18:29:02,206 | INFO | train_grpo_train | metrics_logged mode=train step=683

                                                      
{'loss': 0.0911, 'grad_norm': 15.283223152160645, 'learning_rate': 7.933333333333334e-06, 'num_tokens': 1253254.0, 'completions/mean_length': 43.5, 'completions/min_length': 39.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5645961165428162, 'rewards/meter/std': 0.45666199922561646, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.9908082485198975, 'rewards/repeat_penalty/std': 0.01736707240343094, 'rewards/repeat_floor/mean': 0.9986212253570557, 'rewards/repeat_floor/std': 0.002605058252811432, 'rewards/near_duplicate_penalty/mean': 0.9908082485198975, 'rewards/near_duplicate_penalty/std': 0.01736707240343094, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3609369397163391, 'rewards/total_composite/std': 0.37903764843940735, 'reward': 0.3609369397163391, 'reward_std': 0.37903761863708496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16934515535831451, 'sampling/sampling_logp_difference/max': 1.6285791397094727, 'sampling/importance_sampling_ratio/min': 0.1962081640958786, 'sampling/importance_sampling_ratio/mean': 1.0413309335708618, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.252444252371788, 'clip_ratio/low_mean': 0.07789652980864048, 'clip_ratio/low_min': 0.07789652980864048, 'clip_ratio/high_mean': 0.07041698787361383, 'clip_ratio/high_max': 0.07041698787361383, 'clip_ratio/region_mean': 0.14831351768225431, 'reward_total_mean': 0.3609369397163391, 'reward_meter_mean': 0.5645961165428162, 'reward_meter_std': 0.45666199922561646, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9908082485198975, 'reward_repeat_penalty_std': 0.01736707240343094, 'reward_repeat_floor_mean': 0.9986212253570557, 'reward_repeat_floor_std': 0.002605058252811432, 'reward_near_duplicate_penalty_mean': 0.9908082485198975, 'reward_near_duplicate_penalty_std': 0.01736707240343094, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.3609369397163391, 'reward_total_composite_std': 0.37903764843940735, 'epoch': 0.04}

 21%|██        | 683/3300 [2:30:36<6:40:42,  9.19s/it]INFO 04-12 18:29:02 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 684/3300 [2:30:49<7:42:19, 10.60s/it]2026-04-12 18:29:16,118 | INFO | train_grpo_train | metrics_logged mode=train step=684

                                                      
{'loss': -0.01, 'grad_norm': 5.522702217102051, 'learning_rate': 7.930303030303031e-06, 'num_tokens': 1254656.0, 'completions/mean_length': 99.25, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.28571701049805, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6928744316101074, 'rewards/meter/std': 0.4003216326236725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.929033637046814, 'rewards/lexical_plausibility/std': 0.20072321593761444, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.33342114090919495, 'rewards/repeat_penalty/mean': 0.9981769323348999, 'rewards/repeat_penalty/std': 0.004229435231536627, 'rewards/repeat_floor/mean': 0.9997265338897705, 'rewards/repeat_floor/std': 0.0006344058783724904, 'rewards/near_duplicate_penalty/mean': 0.9981769323348999, 'rewards/near_duplicate_penalty/std': 0.004229435231536627, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49300721287727356, 'rewards/total_composite/std': 0.365594744682312, 'reward': 0.49300721287727356, 'reward_std': 0.3655947148799896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16682495176792145, 'sampling/sampling_logp_difference/max': 1.26860511302948, 'sampling/importance_sampling_ratio/min': 0.2812236249446869, 'sampling/importance_sampling_ratio/mean': 1.0012701749801636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.004506193101406, 'clip_ratio/low_mean': 0.07665951643139124, 'clip_ratio/low_min': 0.07665951643139124, 'clip_ratio/high_mean': 0.09011470898985863, 'clip_ratio/high_max': 0.09011470898985863, 'clip_ratio/region_mean': 0.16677422542124987, 'reward_total_mean': 0.49300721287727356, 'reward_meter_mean': 0.6928744316101074, 'reward_meter_std': 0.4003216326236725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.929033637046814, 'reward_lexical_plausibility_std': 0.20072321593761444, 'reward_repeat_penalty_mean': 0.9981769323348999, 'reward_repeat_penalty_std': 0.004229435231536627, 'reward_repeat_floor_mean': 0.9997265338897705, 'reward_repeat_floor_std': 0.0006344058783724904, 'reward_near_duplicate_penalty_mean': 0.9981769323348999, 'reward_near_duplicate_penalty_std': 0.004229435231536627, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.33342114090919495, 'reward_total_composite_mean': 0.49300721287727356, 'reward_total_composite_std': 0.365594744682312, 'epoch': 0.04}

 21%|██        | 684/3300 [2:30:49<7:42:19, 10.60s/it]INFO 04-12 18:29:16 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 685/3300 [2:30:59<7:23:55, 10.19s/it]2026-04-12 18:29:25,327 | INFO | train_grpo_train | metrics_logged mode=train step=685

                                                      
{'loss': 0.1524, 'grad_norm': 14.242477416992188, 'learning_rate': 7.927272727272729e-06, 'num_tokens': 1256324.0, 'completions/mean_length': 50.5, 'completions/min_length': 33.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.48648005723953247, 'rewards/meter/std': 0.33642902970314026, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7012499570846558, 'rewards/judge_quality/std': 0.21963852643966675, 'rewards/repeat_penalty/mean': 0.972527027130127, 'rewards/repeat_penalty/std': 0.038944847881793976, 'rewards/repeat_floor/mean': 0.9969032406806946, 'rewards/repeat_floor/std': 0.004063354805111885, 'rewards/near_duplicate_penalty/mean': 0.9900432229042053, 'rewards/near_duplicate_penalty/std': 0.01294527668505907, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9821861982345581, 'rewards/distinct_2/std': 0.0333910807967186, 'rewards/total_composite/mean': 0.36581742763519287, 'rewards/total_composite/std': 0.2953268885612488, 'reward': 0.36581742763519287, 'reward_std': 0.2953268885612488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1397547721862793, 'sampling/sampling_logp_difference/max': 1.9408142566680908, 'sampling/importance_sampling_ratio/min': 0.14358699321746826, 'sampling/importance_sampling_ratio/mean': 0.9967378973960876, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.848524235188961, 'clip_ratio/low_mean': 0.0710381967946887, 'clip_ratio/low_min': 0.0710381967946887, 'clip_ratio/high_mean': 0.0545843867585063, 'clip_ratio/high_max': 0.0545843867585063, 'clip_ratio/region_mean': 0.125622583553195, 'reward_total_mean': 0.36581742763519287, 'reward_meter_mean': 0.48648005723953247, 'reward_meter_std': 0.33642902970314026, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.972527027130127, 'reward_repeat_penalty_std': 0.038944847881793976, 'reward_repeat_floor_mean': 0.9969032406806946, 'reward_repeat_floor_std': 0.004063354805111885, 'reward_near_duplicate_penalty_mean': 0.9900432229042053, 'reward_near_duplicate_penalty_std': 0.01294527668505907, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9821861982345581, 'reward_distinct_2_std': 0.0333910807967186, 'reward_judge_quality_mean': 0.7012499570846558, 'reward_judge_quality_std': 0.21963852643966675, 'reward_total_composite_mean': 0.36581742763519287, 'reward_total_composite_std': 0.2953268885612488, 'epoch': 0.04}

 21%|██        | 685/3300 [2:30:59<7:23:55, 10.19s/it]INFO 04-12 18:29:25 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 686/3300 [2:31:09<7:26:36, 10.25s/it]2026-04-12 18:29:35,745 | INFO | train_grpo_train | metrics_logged mode=train step=686

                                                      
{'loss': 0.0025, 'grad_norm': 7.851562023162842, 'learning_rate': 7.924242424242426e-06, 'num_tokens': 1259127.0, 'completions/mean_length': 151.375, 'completions/min_length': 124.0, 'completions/max_length': 169.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.375, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.703392744064331, 'rewards/meter/std': 0.3572615087032318, 'rewards/count_adherence/mean': 0.910714328289032, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9732142686843872, 'rewards/count_floor/std': 0.022180693224072456, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7438592910766602, 'rewards/repeat_penalty/std': 0.11683720350265503, 'rewards/repeat_floor/mean': 0.9738099575042725, 'rewards/repeat_floor/std': 0.012238663621246815, 'rewards/near_duplicate_penalty/mean': 0.9557079076766968, 'rewards/near_duplicate_penalty/std': 0.026552172377705574, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7828199863433838, 'rewards/distinct_2/std': 0.09453731775283813, 'rewards/total_composite/mean': 0.24712425470352173, 'rewards/total_composite/std': 0.1577533781528473, 'reward': 0.24712425470352173, 'reward_std': 0.1577533781528473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1356942504644394, 'sampling/sampling_logp_difference/max': 6.5828704833984375, 'sampling/importance_sampling_ratio/min': 0.0013838711893185973, 'sampling/importance_sampling_ratio/mean': 1.0152630805969238, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6535034105181694, 'clip_ratio/low_mean': 0.0425669401884079, 'clip_ratio/low_min': 0.0425669401884079, 'clip_ratio/high_mean': 0.07818795926868916, 'clip_ratio/high_max': 0.07818795926868916, 'clip_ratio/region_mean': 0.12075489945709705, 'reward_total_mean': 0.24712425470352173, 'reward_meter_mean': 0.703392744064331, 'reward_meter_std': 0.3572615087032318, 'reward_count_adherence_mean': 0.910714328289032, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9732142686843872, 'reward_count_floor_std': 0.022180693224072456, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7438592910766602, 'reward_repeat_penalty_std': 0.11683720350265503, 'reward_repeat_floor_mean': 0.9738099575042725, 'reward_repeat_floor_std': 0.012238663621246815, 'reward_near_duplicate_penalty_mean': 0.9557079076766968, 'reward_near_duplicate_penalty_std': 0.026552172377705574, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7828199863433838, 'reward_distinct_2_std': 0.09453731775283813, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.24712425470352173, 'reward_total_composite_std': 0.1577533781528473, 'epoch': 0.04}

 21%|██        | 686/3300 [2:31:09<7:26:36, 10.25s/it]INFO 04-12 18:29:35 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 687/3300 [2:31:19<7:18:07, 10.06s/it]2026-04-12 18:29:45,357 | INFO | train_grpo_train | metrics_logged mode=train step=687

                                                      
{'loss': 0.006, 'grad_norm': 10.999700546264648, 'learning_rate': 7.921212121212122e-06, 'num_tokens': 1261337.0, 'completions/mean_length': 90.25, 'completions/min_length': 77.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.25, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9758234620094299, 'rewards/meter/std': 0.018658699467778206, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.8926434516906738, 'rewards/repeat_penalty/std': 0.11182642728090286, 'rewards/repeat_floor/mean': 0.9894065260887146, 'rewards/repeat_floor/std': 0.009338720701634884, 'rewards/near_duplicate_penalty/mean': 0.976697564125061, 'rewards/near_duplicate_penalty/std': 0.01590649038553238, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9419655799865723, 'rewards/distinct_2/std': 0.058997996151447296, 'rewards/total_composite/mean': 0.5074703693389893, 'rewards/total_composite/std': 0.21573084592819214, 'reward': 0.5074703693389893, 'reward_std': 0.21573083102703094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1428309828042984, 'sampling/sampling_logp_difference/max': 1.6901178359985352, 'sampling/importance_sampling_ratio/min': 0.18449777364730835, 'sampling/importance_sampling_ratio/mean': 1.0279263257980347, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9295486807823181, 'clip_ratio/low_mean': 0.11103309225291014, 'clip_ratio/low_min': 0.11103309225291014, 'clip_ratio/high_mean': 0.03612903133034706, 'clip_ratio/high_max': 0.03612903133034706, 'clip_ratio/region_mean': 0.1471621235832572, 'reward_total_mean': 0.5074703693389893, 'reward_meter_mean': 0.9758234620094299, 'reward_meter_std': 0.018658699467778206, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8926434516906738, 'reward_repeat_penalty_std': 0.11182642728090286, 'reward_repeat_floor_mean': 0.9894065260887146, 'reward_repeat_floor_std': 0.009338720701634884, 'reward_near_duplicate_penalty_mean': 0.976697564125061, 'reward_near_duplicate_penalty_std': 0.01590649038553238, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9419655799865723, 'reward_distinct_2_std': 0.058997996151447296, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.5074703693389893, 'reward_total_composite_std': 0.21573084592819214, 'epoch': 0.04}

 21%|██        | 687/3300 [2:31:19<7:18:07, 10.06s/it]INFO 04-12 18:29:45 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 688/3300 [2:31:28<7:09:34,  9.87s/it]2026-04-12 18:29:54,770 | INFO | train_grpo_train | metrics_logged mode=train step=688

                                                      
{'loss': 0.1184, 'grad_norm': 12.7700777053833, 'learning_rate': 7.918181818181819e-06, 'num_tokens': 1263043.0, 'completions/mean_length': 53.25, 'completions/min_length': 35.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8128455877304077, 'rewards/meter/std': 0.2773248553276062, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9960116147994995, 'rewards/repeat_penalty/std': 0.005399797111749649, 'rewards/repeat_floor/mean': 0.9994017481803894, 'rewards/repeat_floor/std': 0.0008099714759737253, 'rewards/near_duplicate_penalty/mean': 0.9960116147994995, 'rewards/near_duplicate_penalty/std': 0.005399797111749649, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4248201251029968, 'rewards/total_composite/std': 0.1853087991476059, 'reward': 0.4248201251029968, 'reward_std': 0.1853087991476059, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14777840673923492, 'sampling/sampling_logp_difference/max': 1.8677616119384766, 'sampling/importance_sampling_ratio/min': 0.15446902811527252, 'sampling/importance_sampling_ratio/mean': 1.0275275707244873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0067977830767632, 'clip_ratio/low_mean': 0.1133290845900774, 'clip_ratio/low_min': 0.1133290845900774, 'clip_ratio/high_mean': 0.06153661757707596, 'clip_ratio/high_max': 0.06153661757707596, 'clip_ratio/region_mean': 0.17486570216715336, 'reward_total_mean': 0.4248201251029968, 'reward_meter_mean': 0.8128455877304077, 'reward_meter_std': 0.2773248553276062, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9960116147994995, 'reward_repeat_penalty_std': 0.005399797111749649, 'reward_repeat_floor_mean': 0.9994017481803894, 'reward_repeat_floor_std': 0.0008099714759737253, 'reward_near_duplicate_penalty_mean': 0.9960116147994995, 'reward_near_duplicate_penalty_std': 0.005399797111749649, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.4248201251029968, 'reward_total_composite_std': 0.1853087991476059, 'epoch': 0.04}

 21%|██        | 688/3300 [2:31:28<7:09:34,  9.87s/it]INFO 04-12 18:29:55 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 689/3300 [2:31:41<7:52:08, 10.85s/it]2026-04-12 18:30:07,902 | INFO | train_grpo_train | metrics_logged mode=train step=689

                                                      
{'loss': -0.0652, 'grad_norm': 3.553575038909912, 'learning_rate': 7.915151515151516e-06, 'num_tokens': 1264624.0, 'completions/mean_length': 162.625, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 46.16666793823242, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7992570996284485, 'rewards/meter/std': 0.33354029059410095, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8753854036331177, 'rewards/lexical_plausibility/std': 0.21299925446510315, 'rewards/judge_quality/mean': 0.5, 'rewards/judge_quality/std': 0.33717310428619385, 'rewards/repeat_penalty/mean': 0.9395312070846558, 'rewards/repeat_penalty/std': 0.08119502663612366, 'rewards/repeat_floor/mean': 0.9941484928131104, 'rewards/repeat_floor/std': 0.004828337579965591, 'rewards/near_duplicate_penalty/mean': 0.9779276847839355, 'rewards/near_duplicate_penalty/std': 0.016341637820005417, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.3845026195049286, 'rewards/total_composite/std': 0.3312537372112274, 'reward': 0.3845026195049286, 'reward_std': 0.33125370740890503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1796915978193283, 'sampling/sampling_logp_difference/max': 2.060620069503784, 'sampling/importance_sampling_ratio/min': 0.12737496197223663, 'sampling/importance_sampling_ratio/mean': 0.9975326061248779, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7718779966235161, 'clip_ratio/low_mean': 0.021276595070958138, 'clip_ratio/low_min': 0.021276595070958138, 'clip_ratio/high_mean': 0.10894730500876904, 'clip_ratio/high_max': 0.10894730500876904, 'clip_ratio/region_mean': 0.13022390007972717, 'reward_total_mean': 0.3845026195049286, 'reward_meter_mean': 0.7992570996284485, 'reward_meter_std': 0.33354029059410095, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8753854036331177, 'reward_lexical_plausibility_std': 0.21299925446510315, 'reward_repeat_penalty_mean': 0.9395312070846558, 'reward_repeat_penalty_std': 0.08119502663612366, 'reward_repeat_floor_mean': 0.9941484928131104, 'reward_repeat_floor_std': 0.004828337579965591, 'reward_near_duplicate_penalty_mean': 0.9779276847839355, 'reward_near_duplicate_penalty_std': 0.016341637820005417, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.5, 'reward_judge_quality_std': 0.33717310428619385, 'reward_total_composite_mean': 0.3845026195049286, 'reward_total_composite_std': 0.3312537372112274, 'epoch': 0.04}

 21%|██        | 689/3300 [2:31:41<7:52:08, 10.85s/it]INFO 04-12 18:30:08 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 690/3300 [2:31:49<7:17:00, 10.05s/it]2026-04-12 18:30:16,074 | INFO | train_grpo_train | metrics_logged mode=train step=690

                                                      
{'loss': 0.0178, 'grad_norm': 16.97161293029785, 'learning_rate': 7.912121212121213e-06, 'num_tokens': 1266125.0, 'completions/mean_length': 41.625, 'completions/min_length': 37.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8607565760612488, 'rewards/meter/std': 0.31258437037467957, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.731249988079071, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.9953497648239136, 'rewards/repeat_penalty/std': 0.009349735453724861, 'rewards/repeat_floor/mean': 0.9993025064468384, 'rewards/repeat_floor/std': 0.0014024607371538877, 'rewards/near_duplicate_penalty/mean': 0.9953497648239136, 'rewards/near_duplicate_penalty/std': 0.009349735453724861, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6114264726638794, 'rewards/total_composite/std': 0.3291594386100769, 'reward': 0.6114264726638794, 'reward_std': 0.3291594386100769, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15825603902339935, 'sampling/sampling_logp_difference/max': 2.149714469909668, 'sampling/importance_sampling_ratio/min': 0.11651741713285446, 'sampling/importance_sampling_ratio/mean': 1.015402913093567, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8685256093740463, 'clip_ratio/low_mean': 0.05555725609883666, 'clip_ratio/low_min': 0.05555725609883666, 'clip_ratio/high_mean': 0.0391608402132988, 'clip_ratio/high_max': 0.0391608402132988, 'clip_ratio/region_mean': 0.09471809631213546, 'reward_total_mean': 0.6114264726638794, 'reward_meter_mean': 0.8607565760612488, 'reward_meter_std': 0.31258437037467957, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9953497648239136, 'reward_repeat_penalty_std': 0.009349735453724861, 'reward_repeat_floor_mean': 0.9993025064468384, 'reward_repeat_floor_std': 0.0014024607371538877, 'reward_near_duplicate_penalty_mean': 0.9953497648239136, 'reward_near_duplicate_penalty_std': 0.009349735453724861, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.731249988079071, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.6114264726638794, 'reward_total_composite_std': 0.3291594386100769, 'epoch': 0.04}

 21%|██        | 690/3300 [2:31:49<7:17:00, 10.05s/it]INFO 04-12 18:30:16 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 691/3300 [2:31:58<6:57:26,  9.60s/it]2026-04-12 18:30:24,632 | INFO | train_grpo_train | metrics_logged mode=train step=691

                                                      
{'loss': 0.0312, 'grad_norm': 16.510276794433594, 'learning_rate': 7.909090909090909e-06, 'num_tokens': 1267799.0, 'completions/mean_length': 45.25, 'completions/min_length': 34.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.95619797706604, 'rewards/meter/std': 0.05634555593132973, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.1164964810013771, 'rewards/repeat_penalty/mean': 0.8895437717437744, 'rewards/repeat_penalty/std': 0.09774777293205261, 'rewards/repeat_floor/mean': 0.9860040545463562, 'rewards/repeat_floor/std': 0.01214191410690546, 'rewards/near_duplicate_penalty/mean': 0.9398403167724609, 'rewards/near_duplicate_penalty/std': 0.052910350263118744, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9447554349899292, 'rewards/distinct_2/std': 0.06508119404315948, 'rewards/total_composite/mean': 0.4247620701789856, 'rewards/total_composite/std': 0.10545188188552856, 'reward': 0.4247620701789856, 'reward_std': 0.10545188188552856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16715498268604279, 'sampling/sampling_logp_difference/max': 6.352803707122803, 'sampling/importance_sampling_ratio/min': 0.0017418566858395934, 'sampling/importance_sampling_ratio/mean': 1.036353588104248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0210946649312973, 'clip_ratio/low_mean': 0.11840536817908287, 'clip_ratio/low_min': 0.11840536817908287, 'clip_ratio/high_mean': 0.035920330323278904, 'clip_ratio/high_max': 0.035920330323278904, 'clip_ratio/region_mean': 0.15432569850236177, 'reward_total_mean': 0.4247620701789856, 'reward_meter_mean': 0.95619797706604, 'reward_meter_std': 0.05634555593132973, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8895437717437744, 'reward_repeat_penalty_std': 0.09774777293205261, 'reward_repeat_floor_mean': 0.9860040545463562, 'reward_repeat_floor_std': 0.01214191410690546, 'reward_near_duplicate_penalty_mean': 0.9398403167724609, 'reward_near_duplicate_penalty_std': 0.052910350263118744, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9447554349899292, 'reward_distinct_2_std': 0.06508119404315948, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.1164964810013771, 'reward_total_composite_mean': 0.4247620701789856, 'reward_total_composite_std': 0.10545188188552856, 'epoch': 0.04}

 21%|██        | 691/3300 [2:31:58<6:57:26,  9.60s/it]INFO 04-12 18:30:24 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 692/3300 [2:32:14<8:26:01, 11.64s/it]2026-04-12 18:30:41,038 | INFO | train_grpo_train | metrics_logged mode=train step=692

                                                      
{'loss': -0.0861, 'grad_norm': 4.097788333892822, 'learning_rate': 7.906060606060608e-06, 'num_tokens': 1269741.0, 'completions/mean_length': 185.75, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 77.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.8153953552246094, 'rewards/meter/std': 0.2948692739009857, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2357022613286972, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.0707106813788414, 'rewards/lexical_plausibility/mean': 0.8896616697311401, 'rewards/lexical_plausibility/std': 0.20752596855163574, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.2815771996974945, 'rewards/repeat_penalty/mean': 0.9806714653968811, 'rewards/repeat_penalty/std': 0.036163244396448135, 'rewards/repeat_floor/mean': 0.9976470470428467, 'rewards/repeat_floor/std': 0.00391932250931859, 'rewards/near_duplicate_penalty/mean': 0.9902307987213135, 'rewards/near_duplicate_penalty/std': 0.010942477732896805, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990138053894043, 'rewards/distinct_2/std': 0.027893755584955215, 'rewards/total_composite/mean': 0.3053979277610779, 'rewards/total_composite/std': 0.2750643789768219, 'reward': 0.3053979277610779, 'reward_std': 0.2750643491744995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16998818516731262, 'sampling/sampling_logp_difference/max': 1.905627727508545, 'sampling/importance_sampling_ratio/min': 0.14872926473617554, 'sampling/importance_sampling_ratio/mean': 1.0290447473526, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8471773117780685, 'clip_ratio/low_mean': 0.03693396970629692, 'clip_ratio/low_min': 0.03693396970629692, 'clip_ratio/high_mean': 0.07972583174705505, 'clip_ratio/high_max': 0.07972583174705505, 'clip_ratio/region_mean': 0.11665980145335197, 'reward_total_mean': 0.3053979277610779, 'reward_meter_mean': 0.8153953552246094, 'reward_meter_std': 0.2948692739009857, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2357022613286972, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.0707106813788414, 'reward_lexical_plausibility_mean': 0.8896616697311401, 'reward_lexical_plausibility_std': 0.20752596855163574, 'reward_repeat_penalty_mean': 0.9806714653968811, 'reward_repeat_penalty_std': 0.036163244396448135, 'reward_repeat_floor_mean': 0.9976470470428467, 'reward_repeat_floor_std': 0.00391932250931859, 'reward_near_duplicate_penalty_mean': 0.9902307987213135, 'reward_near_duplicate_penalty_std': 0.010942477732896805, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990138053894043, 'reward_distinct_2_std': 0.027893755584955215, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.2815771996974945, 'reward_total_composite_mean': 0.3053979277610779, 'reward_total_composite_std': 0.2750643789768219, 'epoch': 0.04}

 21%|██        | 692/3300 [2:32:14<8:26:01, 11.64s/it]INFO 04-12 18:30:41 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 693/3300 [2:32:28<8:48:51, 12.17s/it]2026-04-12 18:30:54,448 | INFO | train_grpo_train | metrics_logged mode=train step=693

                                                      
{'loss': -0.052, 'grad_norm': 5.776697158813477, 'learning_rate': 7.903030303030303e-06, 'num_tokens': 1271228.0, 'completions/mean_length': 95.875, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.42857360839844, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7285749316215515, 'rewards/meter/std': 0.323891282081604, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9372708797454834, 'rewards/lexical_plausibility/std': 0.1774248480796814, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.20310094952583313, 'rewards/repeat_penalty/mean': 0.9540948271751404, 'rewards/repeat_penalty/std': 0.08570332825183868, 'rewards/repeat_floor/mean': 0.9959267377853394, 'rewards/repeat_floor/std': 0.0056337621062994, 'rewards/near_duplicate_penalty/mean': 0.9853448271751404, 'rewards/near_duplicate_penalty/std': 0.024069856852293015, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24386470019817352, 'rewards/total_composite/std': 0.18078291416168213, 'reward': 0.24386470019817352, 'reward_std': 0.18078292906284332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1864989697933197, 'sampling/sampling_logp_difference/max': 2.0610222816467285, 'sampling/importance_sampling_ratio/min': 0.12732374668121338, 'sampling/importance_sampling_ratio/mean': 0.9933604598045349, 'sampling/importance_sampling_ratio/max': 1.884401559829712, 'entropy': 1.1467092335224152, 'clip_ratio/low_mean': 0.06282679922878742, 'clip_ratio/low_min': 0.06282679922878742, 'clip_ratio/high_mean': 0.08883412554860115, 'clip_ratio/high_max': 0.08883412554860115, 'clip_ratio/region_mean': 0.15166092477738857, 'reward_total_mean': 0.24386470019817352, 'reward_meter_mean': 0.7285749316215515, 'reward_meter_std': 0.323891282081604, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9372708797454834, 'reward_lexical_plausibility_std': 0.1774248480796814, 'reward_repeat_penalty_mean': 0.9540948271751404, 'reward_repeat_penalty_std': 0.08570332825183868, 'reward_repeat_floor_mean': 0.9959267377853394, 'reward_repeat_floor_std': 0.0056337621062994, 'reward_near_duplicate_penalty_mean': 0.9853448271751404, 'reward_near_duplicate_penalty_std': 0.024069856852293015, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.20310094952583313, 'reward_total_composite_mean': 0.24386470019817352, 'reward_total_composite_std': 0.18078291416168213, 'epoch': 0.04}

 21%|██        | 693/3300 [2:32:28<8:48:51, 12.17s/it]INFO 04-12 18:30:54 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 694/3300 [2:32:36<8:01:23, 11.08s/it]2026-04-12 18:31:02,991 | INFO | train_grpo_train | metrics_logged mode=train step=694

                                                      
{'loss': 0.05, 'grad_norm': 17.572412490844727, 'learning_rate': 7.9e-06, 'num_tokens': 1272788.0, 'completions/mean_length': 33.0, 'completions/min_length': 28.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5327678918838501, 'rewards/meter/std': 0.3950226306915283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9881653785705566, 'rewards/lexical_plausibility/std': 0.02336401864886284, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3296508193016052, 'rewards/repeat_penalty/mean': 0.9851109981536865, 'rewards/repeat_penalty/std': 0.02075997181236744, 'rewards/repeat_floor/mean': 0.9977666735649109, 'rewards/repeat_floor/std': 0.0031139894854277372, 'rewards/near_duplicate_penalty/mean': 0.9851109981536865, 'rewards/near_duplicate_penalty/std': 0.02075997181236744, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.28674715757369995, 'rewards/total_composite/std': 0.35322123765945435, 'reward': 0.28674715757369995, 'reward_std': 0.35322120785713196, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1541822850704193, 'sampling/sampling_logp_difference/max': 2.0245985984802246, 'sampling/importance_sampling_ratio/min': 0.13204683363437653, 'sampling/importance_sampling_ratio/mean': 1.004731297492981, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.892720527946949, 'clip_ratio/low_mean': 0.09476255066692829, 'clip_ratio/low_min': 0.09476255066692829, 'clip_ratio/high_mean': 0.023809524718672037, 'clip_ratio/high_max': 0.023809524718672037, 'clip_ratio/region_mean': 0.11857207538560033, 'reward_total_mean': 0.28674715757369995, 'reward_meter_mean': 0.5327678918838501, 'reward_meter_std': 0.3950226306915283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9881653785705566, 'reward_lexical_plausibility_std': 0.02336401864886284, 'reward_repeat_penalty_mean': 0.9851109981536865, 'reward_repeat_penalty_std': 0.02075997181236744, 'reward_repeat_floor_mean': 0.9977666735649109, 'reward_repeat_floor_std': 0.0031139894854277372, 'reward_near_duplicate_penalty_mean': 0.9851109981536865, 'reward_near_duplicate_penalty_std': 0.02075997181236744, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3296508193016052, 'reward_total_composite_mean': 0.28674715757369995, 'reward_total_composite_std': 0.35322123765945435, 'epoch': 0.04}

 21%|██        | 694/3300 [2:32:36<8:01:23, 11.08s/it]INFO 04-12 18:31:03 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 695/3300 [2:32:47<7:58:14, 11.02s/it]2026-04-12 18:31:13,846 | INFO | train_grpo_train | metrics_logged mode=train step=695

                                                      
{'loss': 0.1389, 'grad_norm': 16.48650550842285, 'learning_rate': 7.896969696969698e-06, 'num_tokens': 1274348.0, 'completions/mean_length': 28.0, 'completions/min_length': 22.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.8752923011779785, 'rewards/meter/std': 0.20829828083515167, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9713541269302368, 'rewards/lexical_plausibility/std': 0.05990393087267876, 'rewards/judge_quality/mean': 0.6050000190734863, 'rewards/judge_quality/std': 0.27244922518730164, 'rewards/repeat_penalty/mean': 0.7911126613616943, 'rewards/repeat_penalty/std': 0.088370680809021, 'rewards/repeat_floor/mean': 0.9864853620529175, 'rewards/repeat_floor/std': 0.005011714994907379, 'rewards/near_duplicate_penalty/mean': 0.9953917264938354, 'rewards/near_duplicate_penalty/std': 0.006611193995922804, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.4678093492984772, 'rewards/total_composite/std': 0.254795640707016, 'reward': 0.4678093492984772, 'reward_std': 0.2547956109046936, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1850403994321823, 'sampling/sampling_logp_difference/max': 1.6788057088851929, 'sampling/importance_sampling_ratio/min': 0.18659669160842896, 'sampling/importance_sampling_ratio/mean': 1.036389708518982, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3574261143803596, 'clip_ratio/low_mean': 0.13032166846096516, 'clip_ratio/low_min': 0.13032166846096516, 'clip_ratio/high_mean': 0.04796765837818384, 'clip_ratio/high_max': 0.04796765837818384, 'clip_ratio/region_mean': 0.178289326839149, 'reward_total_mean': 0.4678093492984772, 'reward_meter_mean': 0.8752923011779785, 'reward_meter_std': 0.20829828083515167, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9713541269302368, 'reward_lexical_plausibility_std': 0.05990393087267876, 'reward_repeat_penalty_mean': 0.7911126613616943, 'reward_repeat_penalty_std': 0.088370680809021, 'reward_repeat_floor_mean': 0.9864853620529175, 'reward_repeat_floor_std': 0.005011714994907379, 'reward_near_duplicate_penalty_mean': 0.9953917264938354, 'reward_near_duplicate_penalty_std': 0.006611193995922804, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.6050000190734863, 'reward_judge_quality_std': 0.27244922518730164, 'reward_total_composite_mean': 0.4678093492984772, 'reward_total_composite_std': 0.254795640707016, 'epoch': 0.04}

 21%|██        | 695/3300 [2:32:47<7:58:14, 11.02s/it]INFO 04-12 18:31:14 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 696/3300 [2:32:57<7:45:42, 10.73s/it]2026-04-12 18:31:23,916 | INFO | train_grpo_train | metrics_logged mode=train step=696

                                                      
{'loss': 0.0215, 'grad_norm': 10.675210952758789, 'learning_rate': 7.893939393939395e-06, 'num_tokens': 1276593.0, 'completions/mean_length': 79.625, 'completions/min_length': 46.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.625, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.8943508267402649, 'rewards/meter/std': 0.23351480066776276, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.22160132229328156, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.06648040562868118, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.9841098785400391, 'rewards/repeat_penalty/std': 0.020202450454235077, 'rewards/repeat_floor/mean': 0.9979227781295776, 'rewards/repeat_floor/std': 0.002310001291334629, 'rewards/near_duplicate_penalty/mean': 0.9893572330474854, 'rewards/near_duplicate_penalty/std': 0.010167546570301056, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9946581125259399, 'rewards/distinct_2/std': 0.01510911900550127, 'rewards/total_composite/mean': 0.3716614246368408, 'rewards/total_composite/std': 0.20161238312721252, 'reward': 0.3716614246368408, 'reward_std': 0.20161235332489014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1626281589269638, 'sampling/sampling_logp_difference/max': 1.7982239723205566, 'sampling/importance_sampling_ratio/min': 0.1655927300453186, 'sampling/importance_sampling_ratio/mean': 1.0247689485549927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.041448913514614, 'clip_ratio/low_mean': 0.060205587185919285, 'clip_ratio/low_min': 0.060205587185919285, 'clip_ratio/high_mean': 0.09158066054806113, 'clip_ratio/high_max': 0.09158066054806113, 'clip_ratio/region_mean': 0.15178624773398042, 'reward_total_mean': 0.3716614246368408, 'reward_meter_mean': 0.8943508267402649, 'reward_meter_std': 0.23351480066776276, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.22160132229328156, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.06648040562868118, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9841098785400391, 'reward_repeat_penalty_std': 0.020202450454235077, 'reward_repeat_floor_mean': 0.9979227781295776, 'reward_repeat_floor_std': 0.002310001291334629, 'reward_near_duplicate_penalty_mean': 0.9893572330474854, 'reward_near_duplicate_penalty_std': 0.010167546570301056, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9946581125259399, 'reward_distinct_2_std': 0.01510911900550127, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.3716614246368408, 'reward_total_composite_std': 0.20161238312721252, 'epoch': 0.04}

 21%|██        | 696/3300 [2:32:57<7:45:42, 10.73s/it]INFO 04-12 18:31:24 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 697/3300 [2:33:11<8:21:13, 11.55s/it]2026-04-12 18:31:37,387 | INFO | train_grpo_train | metrics_logged mode=train step=697

                                                      
{'loss': 0.011, 'grad_norm': 5.811776161193848, 'learning_rate': 7.89090909090909e-06, 'num_tokens': 1277995.0, 'completions/mean_length': 94.25, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.57143020629883, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5346740484237671, 'rewards/meter/std': 0.48755258321762085, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.9386634826660156, 'rewards/lexical_plausibility/std': 0.1734859049320221, 'rewards/judge_quality/mean': 0.6299999952316284, 'rewards/judge_quality/std': 0.3489985764026642, 'rewards/repeat_penalty/mean': 0.8246878385543823, 'rewards/repeat_penalty/std': 0.15275509655475616, 'rewards/repeat_floor/mean': 0.9861825704574585, 'rewards/repeat_floor/std': 0.016363805159926414, 'rewards/near_duplicate_penalty/mean': 0.9788626432418823, 'rewards/near_duplicate_penalty/std': 0.05838554725050926, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9650349617004395, 'rewards/distinct_2/std': 0.09889606386423111, 'rewards/total_composite/mean': 0.3222241997718811, 'rewards/total_composite/std': 0.35222432017326355, 'reward': 0.3222241997718811, 'reward_std': 0.35222432017326355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14557981491088867, 'sampling/sampling_logp_difference/max': 1.6572270393371582, 'sampling/importance_sampling_ratio/min': 0.19066695868968964, 'sampling/importance_sampling_ratio/mean': 0.9892773628234863, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.86921776086092, 'clip_ratio/low_mean': 0.027139748446643353, 'clip_ratio/low_min': 0.027139748446643353, 'clip_ratio/high_mean': 0.07627321360632777, 'clip_ratio/high_max': 0.07627321360632777, 'clip_ratio/region_mean': 0.10341296205297112, 'reward_total_mean': 0.3222241997718811, 'reward_meter_mean': 0.5346740484237671, 'reward_meter_std': 0.48755258321762085, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.9386634826660156, 'reward_lexical_plausibility_std': 0.1734859049320221, 'reward_repeat_penalty_mean': 0.8246878385543823, 'reward_repeat_penalty_std': 0.15275509655475616, 'reward_repeat_floor_mean': 0.9861825704574585, 'reward_repeat_floor_std': 0.016363805159926414, 'reward_near_duplicate_penalty_mean': 0.9788626432418823, 'reward_near_duplicate_penalty_std': 0.05838554725050926, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9650349617004395, 'reward_distinct_2_std': 0.09889606386423111, 'reward_judge_quality_mean': 0.6299999952316284, 'reward_judge_quality_std': 0.3489985764026642, 'reward_total_composite_mean': 0.3222241997718811, 'reward_total_composite_std': 0.35222432017326355, 'epoch': 0.04}

 21%|██        | 697/3300 [2:33:11<8:21:13, 11.55s/it]INFO 04-12 18:31:37 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 698/3300 [2:33:25<8:50:39, 12.24s/it]2026-04-12 18:31:51,231 | INFO | train_grpo_train | metrics_logged mode=train step=698

                                                      
{'loss': -0.0414, 'grad_norm': 1.8564060926437378, 'learning_rate': 7.88787878787879e-06, 'num_tokens': 1279502.0, 'completions/mean_length': 154.375, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.16666793823242, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6102007627487183, 'rewards/meter/std': 0.4394175410270691, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9030071496963501, 'rewards/lexical_plausibility/std': 0.1839004009962082, 'rewards/judge_quality/mean': 0.5074999928474426, 'rewards/judge_quality/std': 0.3796144723892212, 'rewards/repeat_penalty/mean': 0.8900975584983826, 'rewards/repeat_penalty/std': 0.12551118433475494, 'rewards/repeat_floor/mean': 0.9918233156204224, 'rewards/repeat_floor/std': 0.008244846947491169, 'rewards/near_duplicate_penalty/mean': 0.982988715171814, 'rewards/near_duplicate_penalty/std': 0.031867094337940216, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3795780837535858, 'rewards/total_composite/std': 0.3140793442726135, 'reward': 0.3795780837535858, 'reward_std': 0.3140793442726135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16782984137535095, 'sampling/sampling_logp_difference/max': 1.542236328125, 'sampling/importance_sampling_ratio/min': 0.21390222012996674, 'sampling/importance_sampling_ratio/mean': 1.003355860710144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7542218193411827, 'clip_ratio/low_mean': 0.027100469917058945, 'clip_ratio/low_min': 0.027100469917058945, 'clip_ratio/high_mean': 0.08988736011087894, 'clip_ratio/high_max': 0.08988736011087894, 'clip_ratio/region_mean': 0.11698783002793789, 'reward_total_mean': 0.3795780837535858, 'reward_meter_mean': 0.6102007627487183, 'reward_meter_std': 0.4394175410270691, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9030071496963501, 'reward_lexical_plausibility_std': 0.1839004009962082, 'reward_repeat_penalty_mean': 0.8900975584983826, 'reward_repeat_penalty_std': 0.12551118433475494, 'reward_repeat_floor_mean': 0.9918233156204224, 'reward_repeat_floor_std': 0.008244846947491169, 'reward_near_duplicate_penalty_mean': 0.982988715171814, 'reward_near_duplicate_penalty_std': 0.031867094337940216, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5074999928474426, 'reward_judge_quality_std': 0.3796144723892212, 'reward_total_composite_mean': 0.3795780837535858, 'reward_total_composite_std': 0.3140793442726135, 'epoch': 0.04}

 21%|██        | 698/3300 [2:33:25<8:50:39, 12.24s/it]INFO 04-12 18:31:51 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 699/3300 [2:33:38<9:07:26, 12.63s/it]2026-04-12 18:32:04,917 | INFO | train_grpo_train | metrics_logged mode=train step=699

                                                      
{'loss': -0.1166, 'grad_norm': 5.406973361968994, 'learning_rate': 7.884848484848485e-06, 'num_tokens': 1280998.0, 'completions/mean_length': 95.0, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.42857360839844, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8485106229782104, 'rewards/meter/std': 0.25493329763412476, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.947380542755127, 'rewards/lexical_plausibility/std': 0.148830384016037, 'rewards/judge_quality/mean': 0.5712499618530273, 'rewards/judge_quality/std': 0.3491392433643341, 'rewards/repeat_penalty/mean': 0.7777508497238159, 'rewards/repeat_penalty/std': 0.07377924025058746, 'rewards/repeat_floor/mean': 0.9824823141098022, 'rewards/repeat_floor/std': 0.0071553257293999195, 'rewards/near_duplicate_penalty/mean': 0.9701954126358032, 'rewards/near_duplicate_penalty/std': 0.0543275885283947, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9591997861862183, 'rewards/distinct_2/std': 0.060878071933984756, 'rewards/total_composite/mean': 0.49808424711227417, 'rewards/total_composite/std': 0.3303817808628082, 'reward': 0.49808424711227417, 'reward_std': 0.3303817808628082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16661261022090912, 'sampling/sampling_logp_difference/max': 0.992431640625, 'sampling/importance_sampling_ratio/min': 0.3706742525100708, 'sampling/importance_sampling_ratio/mean': 1.0287952423095703, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4277740269899368, 'clip_ratio/low_mean': 0.07378183864057064, 'clip_ratio/low_min': 0.07378183864057064, 'clip_ratio/high_mean': 0.09797375928610563, 'clip_ratio/high_max': 0.09797375928610563, 'clip_ratio/region_mean': 0.17175559792667627, 'reward_total_mean': 0.49808424711227417, 'reward_meter_mean': 0.8485106229782104, 'reward_meter_std': 0.25493329763412476, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.947380542755127, 'reward_lexical_plausibility_std': 0.148830384016037, 'reward_repeat_penalty_mean': 0.7777508497238159, 'reward_repeat_penalty_std': 0.07377924025058746, 'reward_repeat_floor_mean': 0.9824823141098022, 'reward_repeat_floor_std': 0.0071553257293999195, 'reward_near_duplicate_penalty_mean': 0.9701954126358032, 'reward_near_duplicate_penalty_std': 0.0543275885283947, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9591997861862183, 'reward_distinct_2_std': 0.060878071933984756, 'reward_judge_quality_mean': 0.5712499618530273, 'reward_judge_quality_std': 0.3491392433643341, 'reward_total_composite_mean': 0.49808424711227417, 'reward_total_composite_std': 0.3303817808628082, 'epoch': 0.04}

 21%|██        | 699/3300 [2:33:38<9:07:26, 12.63s/it]INFO 04-12 18:32:05 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 700/3300 [2:33:52<9:30:08, 13.16s/it]2026-04-12 18:32:19,162 | INFO | train_grpo_train | metrics_logged mode=train step=700

                                                      
{'loss': -0.1291, 'grad_norm': 3.2831649780273438, 'learning_rate': 7.881818181818182e-06, 'num_tokens': 1283048.0, 'completions/mean_length': 200.25, 'completions/min_length': 80.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 96.33333587646484, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.5204377174377441, 'rewards/meter/std': 0.41852569580078125, 'rewards/count_adherence/mean': 0.699999988079071, 'rewards/count_adherence/std': 0.32071349024772644, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9100000262260437, 'rewards/count_floor/std': 0.09621405601501465, 'rewards/lexical_plausibility/mean': 0.8878356218338013, 'rewards/lexical_plausibility/std': 0.20871137082576752, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.8468242883682251, 'rewards/repeat_penalty/std': 0.11950872093439102, 'rewards/repeat_floor/mean': 0.9871902465820312, 'rewards/repeat_floor/std': 0.008598395623266697, 'rewards/near_duplicate_penalty/mean': 0.9739341735839844, 'rewards/near_duplicate_penalty/std': 0.022811613976955414, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9427787065505981, 'rewards/distinct_2/std': 0.07691513746976852, 'rewards/total_composite/mean': 0.1796427071094513, 'rewards/total_composite/std': 0.14334683120250702, 'reward': 0.1796427071094513, 'reward_std': 0.14334683120250702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18485522270202637, 'sampling/sampling_logp_difference/max': 2.7764220237731934, 'sampling/importance_sampling_ratio/min': 0.06226087734103203, 'sampling/importance_sampling_ratio/mean': 0.9976781606674194, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7676830291748047, 'clip_ratio/low_mean': 0.04419191926717758, 'clip_ratio/low_min': 0.04419191926717758, 'clip_ratio/high_mean': 0.08549442142248154, 'clip_ratio/high_max': 0.08549442142248154, 'clip_ratio/region_mean': 0.12968634068965912, 'reward_total_mean': 0.1796427071094513, 'reward_meter_mean': 0.5204377174377441, 'reward_meter_std': 0.41852569580078125, 'reward_count_adherence_mean': 0.699999988079071, 'reward_count_adherence_std': 0.32071349024772644, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9100000262260437, 'reward_count_floor_std': 0.09621405601501465, 'reward_lexical_plausibility_mean': 0.8878356218338013, 'reward_lexical_plausibility_std': 0.20871137082576752, 'reward_repeat_penalty_mean': 0.8468242883682251, 'reward_repeat_penalty_std': 0.11950872093439102, 'reward_repeat_floor_mean': 0.9871902465820312, 'reward_repeat_floor_std': 0.008598395623266697, 'reward_near_duplicate_penalty_mean': 0.9739341735839844, 'reward_near_duplicate_penalty_std': 0.022811613976955414, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9427787065505981, 'reward_distinct_2_std': 0.07691513746976852, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.1796427071094513, 'reward_total_composite_std': 0.14334683120250702, 'epoch': 0.04}

 21%|██        | 700/3300 [2:33:53<9:30:08, 13.16s/it]INFO 04-12 18:32:19 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:13<01:14,  6.77s/it]

 23%|██▎       | 3/13 [00:24<01:24,  8.47s/it]

 31%|███       | 4/13 [00:33<01:19,  8.80s/it]

 38%|███▊      | 5/13 [00:44<01:16,  9.57s/it]

 46%|████▌     | 6/13 [00:56<01:12, 10.37s/it]

 54%|█████▍    | 7/13 [01:06<01:00, 10.04s/it]

 62%|██████▏   | 8/13 [01:21<00:58, 11.70s/it]

 69%|██████▉   | 9/13 [01:33<00:47, 11.86s/it]

 77%|███████▋  | 10/13 [01:45<00:35, 11.78s/it]

 85%|████████▍ | 11/13 [01:56<00:23, 11.72s/it]

 92%|█████████▏| 12/13 [02:06<00:11, 11.12s/it]

100%|██████████| 13/13 [02:15<00:00, 10.57s/it]2026-04-12 18:34:50,220 | INFO | train_grpo_train | metrics_logged mode=eval step=700

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 150.9885, 'eval_samples_per_second': 0.689, 'eval_steps_per_second': 0.086, 'eval_num_tokens': 1283048.0, 'eval_completions/mean_length': 155.39423076923077, 'eval_completions/min_length': 26.307692307692307, 'eval_completions/max_length': 420.84615384615387, 'eval_completions/clipped_ratio': 0.07692307692307693, 'eval_completions/mean_terminated_length': 126.42720002394456, 'eval_completions/min_terminated_length': 26.307692307692307, 'eval_completions/max_terminated_length': 297.46153846153845, 'eval_rewards/meter/mean': 0.5489744062607105, 'eval_rewards/meter/std': 0.3748012941617232, 'eval_rewards/count_adherence/mean': 0.7764834853319021, 'eval_rewards/count_adherence/std': 0.24915027274535254, 'eval_rewards/arabic_clean/mean': 0.875, 'eval_rewards/arabic_clean/std': 0.2972002969338344, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 1.0, 'eval_rewards/arabic_floor/std': 0.0, 'eval_rewards/count_floor/mean': 0.932945035971128, 'eval_rewards/count_floor/std': 0.07474507907262215, 'eval_rewards/lexical_plausibility/mean': 0.9671332882000849, 'eval_rewards/lexical_plausibility/std': 0.07702519601354232, 'eval_rewards/judge_quality/mean': 0.4059615364441505, 'eval_rewards/judge_quality/std': 0.2083253788833435, 'eval_rewards/repeat_penalty/mean': 0.7792832301213191, 'eval_rewards/repeat_penalty/std': 0.27835372720773405, 'eval_rewards/repeat_floor/mean': 0.9773236971635085, 'eval_rewards/repeat_floor/std': 0.0311901608052162, 'eval_rewards/near_duplicate_penalty/mean': 0.9611879816422095, 'eval_rewards/near_duplicate_penalty/std': 0.052638744218991354, 'eval_rewards/opening_diversity/mean': 0.9583390080011808, 'eval_rewards/opening_diversity/std': 0.08785006289298718, 'eval_rewards/distinct_2/mean': 0.8405017302586482, 'eval_rewards/distinct_2/std': 0.26118571483171904, 'eval_rewards/total_composite/mean': 0.214217674273711, 'eval_rewards/total_composite/std': 0.1935919368496308, 'eval_reward': 0.214217674273711, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07472726330161095, 'eval_sampling/sampling_logp_difference/max': 1.1051650047302246, 'eval_sampling/importance_sampling_ratio/min': 0.33584901919731724, 'eval_sampling/importance_sampling_ratio/mean': 1.0187156383807843, 'eval_sampling/importance_sampling_ratio/max': 1.5717311730751624, 'eval_entropy': 0.8667431015234727, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.214217674273711, 'eval_reward_meter_mean': 0.5489744062607105, 'eval_reward_meter_std': 0.3748012941617232, 'eval_reward_count_adherence_mean': 0.7764834853319021, 'eval_reward_count_adherence_std': 0.24915027274535254, 'eval_reward_arabic_clean_mean': 0.875, 'eval_reward_arabic_clean_std': 0.2972002969338344, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 1.0, 'eval_reward_arabic_floor_std': 0.0, 'eval_reward_count_floor_mean': 0.932945035971128, 'eval_reward_count_floor_std': 0.07474507907262215, 'eval_reward_lexical_plausibility_mean': 0.9671332882000849, 'eval_reward_lexical_plausibility_std': 0.07702519601354232, 'eval_reward_repeat_penalty_mean': 0.7792832301213191, 'eval_reward_repeat_penalty_std': 0.27835372720773405, 'eval_reward_repeat_floor_mean': 0.9773236971635085, 'eval_reward_repeat_floor_std': 0.0311901608052162, 'eval_reward_near_duplicate_penalty_mean': 0.9611879816422095, 'eval_reward_near_duplicate_penalty_std': 0.052638744218991354, 'eval_reward_opening_diversity_mean': 0.9583390080011808, 'eval_reward_opening_diversity_std': 0.08785006289298718, 'eval_reward_distinct_2_mean': 0.8405017302586482, 'eval_reward_distinct_2_std': 0.26118571483171904, 'eval_reward_judge_quality_mean': 0.4059615364441505, 'eval_reward_judge_quality_std': 0.2083253788833435, 'eval_reward_total_composite_mean': 0.214217674273711, 'eval_reward_total_composite_std': 0.1935919368496308, 'epoch': 0.04}

 21%|██        | 700/3300 [2:36:24<9:30:08, 13.16s/it]

100%|██████████| 13/13 [02:15<00:00, 10.57s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 21%|██        | 701/3300 [2:36:36<42:08:57, 58.38s/it]2026-04-12 18:35:03,069 | INFO | train_grpo_train | metrics_logged mode=train step=701

                                                       
{'loss': 0.027, 'grad_norm': 19.635427474975586, 'learning_rate': 7.87878787878788e-06, 'num_tokens': 1284493.0, 'completions/mean_length': 29.625, 'completions/min_length': 18.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6219130754470825, 'rewards/meter/std': 0.35015761852264404, 'rewards/count_adherence/mean': 0.5833333730697632, 'rewards/count_adherence/std': 0.15430335700511932, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.875, 'rewards/count_floor/std': 0.046290989965200424, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.3378080725669861, 'rewards/repeat_penalty/mean': 0.8384596109390259, 'rewards/repeat_penalty/std': 0.1622389703989029, 'rewards/repeat_floor/mean': 0.9866820573806763, 'rewards/repeat_floor/std': 0.013959908857941628, 'rewards/near_duplicate_penalty/mean': 0.9744111895561218, 'rewards/near_duplicate_penalty/std': 0.035747501999139786, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9780040383338928, 'rewards/distinct_2/std': 0.04075079783797264, 'rewards/total_composite/mean': 0.37776368856430054, 'rewards/total_composite/std': 0.30817466974258423, 'reward': 0.37776368856430054, 'reward_std': 0.30817466974258423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17166176438331604, 'sampling/sampling_logp_difference/max': 1.7431252002716064, 'sampling/importance_sampling_ratio/min': 0.17497271299362183, 'sampling/importance_sampling_ratio/mean': 1.0168657302856445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.26077351719141, 'clip_ratio/low_mean': 0.08873770013451576, 'clip_ratio/low_min': 0.08873770013451576, 'clip_ratio/high_mean': 0.07787176314741373, 'clip_ratio/high_max': 0.07787176314741373, 'clip_ratio/region_mean': 0.1666094632819295, 'reward_total_mean': 0.37776368856430054, 'reward_meter_mean': 0.6219130754470825, 'reward_meter_std': 0.35015761852264404, 'reward_count_adherence_mean': 0.5833333730697632, 'reward_count_adherence_std': 0.15430335700511932, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.875, 'reward_count_floor_std': 0.046290989965200424, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8384596109390259, 'reward_repeat_penalty_std': 0.1622389703989029, 'reward_repeat_floor_mean': 0.9866820573806763, 'reward_repeat_floor_std': 0.013959908857941628, 'reward_near_duplicate_penalty_mean': 0.9744111895561218, 'reward_near_duplicate_penalty_std': 0.035747501999139786, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9780040383338928, 'reward_distinct_2_std': 0.04075079783797264, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.3378080725669861, 'reward_total_composite_mean': 0.37776368856430054, 'reward_total_composite_std': 0.30817466974258423, 'epoch': 0.04}

 21%|██        | 701/3300 [2:36:36<42:08:57, 58.38s/it]INFO 04-12 18:35:03 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 702/3300 [2:36:45<31:27:00, 43.58s/it]2026-04-12 18:35:12,107 | INFO | train_grpo_train | metrics_logged mode=train step=702

                                                       
{'loss': 0.1869, 'grad_norm': 18.982019424438477, 'learning_rate': 7.875757575757577e-06, 'num_tokens': 1285975.0, 'completions/mean_length': 36.25, 'completions/min_length': 16.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7270604372024536, 'rewards/meter/std': 0.3584493100643158, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7225000262260437, 'rewards/judge_quality/std': 0.28272905945777893, 'rewards/repeat_penalty/mean': 0.9050331115722656, 'rewards/repeat_penalty/std': 0.12842176854610443, 'rewards/repeat_floor/mean': 0.9941924810409546, 'rewards/repeat_floor/std': 0.007628065068274736, 'rewards/near_duplicate_penalty/mean': 0.9987831115722656, 'rewards/near_duplicate_penalty/std': 0.0034419645089656115, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5188474655151367, 'rewards/total_composite/std': 0.35184794664382935, 'reward': 0.5188474655151367, 'reward_std': 0.35184794664382935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16477331519126892, 'sampling/sampling_logp_difference/max': 1.6497201919555664, 'sampling/importance_sampling_ratio/min': 0.19210365414619446, 'sampling/importance_sampling_ratio/mean': 1.0022728443145752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0120640695095062, 'clip_ratio/low_mean': 0.07160014472901821, 'clip_ratio/low_min': 0.07160014472901821, 'clip_ratio/high_mean': 0.10375485103577375, 'clip_ratio/high_max': 0.10375485103577375, 'clip_ratio/region_mean': 0.17535499576479197, 'reward_total_mean': 0.5188474655151367, 'reward_meter_mean': 0.7270604372024536, 'reward_meter_std': 0.3584493100643158, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9050331115722656, 'reward_repeat_penalty_std': 0.12842176854610443, 'reward_repeat_floor_mean': 0.9941924810409546, 'reward_repeat_floor_std': 0.007628065068274736, 'reward_near_duplicate_penalty_mean': 0.9987831115722656, 'reward_near_duplicate_penalty_std': 0.0034419645089656115, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7225000262260437, 'reward_judge_quality_std': 0.28272905945777893, 'reward_total_composite_mean': 0.5188474655151367, 'reward_total_composite_std': 0.35184794664382935, 'epoch': 0.04}

 21%|██▏       | 702/3300 [2:36:45<31:27:00, 43.58s/it]INFO 04-12 18:35:12 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 703/3300 [2:36:54<23:58:34, 33.24s/it]2026-04-12 18:35:21,206 | INFO | train_grpo_train | metrics_logged mode=train step=703

                                                       
{'loss': 0.368, 'grad_norm': 20.71617317199707, 'learning_rate': 7.872727272727273e-06, 'num_tokens': 1287520.0, 'completions/mean_length': 35.125, 'completions/min_length': 22.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.692795991897583, 'rewards/meter/std': 0.3991662263870239, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.8719308972358704, 'rewards/repeat_penalty/std': 0.13043223321437836, 'rewards/repeat_floor/mean': 0.992039680480957, 'rewards/repeat_floor/std': 0.0075578163377940655, 'rewards/near_duplicate_penalty/mean': 0.9969308972358704, 'rewards/near_duplicate_penalty/std': 0.005684043746441603, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45557886362075806, 'rewards/total_composite/std': 0.3125329315662384, 'reward': 0.45557886362075806, 'reward_std': 0.3125329315662384, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17352819442749023, 'sampling/sampling_logp_difference/max': 1.6558523178100586, 'sampling/importance_sampling_ratio/min': 0.19092924892902374, 'sampling/importance_sampling_ratio/mean': 1.009139895439148, 'sampling/importance_sampling_ratio/max': 1.898756980895996, 'entropy': 1.1124054417014122, 'clip_ratio/low_mean': 0.07493782881647348, 'clip_ratio/low_min': 0.07493782881647348, 'clip_ratio/high_mean': 0.08428030554205179, 'clip_ratio/high_max': 0.08428030554205179, 'clip_ratio/region_mean': 0.15921813435852528, 'reward_total_mean': 0.45557886362075806, 'reward_meter_mean': 0.692795991897583, 'reward_meter_std': 0.3991662263870239, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8719308972358704, 'reward_repeat_penalty_std': 0.13043223321437836, 'reward_repeat_floor_mean': 0.992039680480957, 'reward_repeat_floor_std': 0.0075578163377940655, 'reward_near_duplicate_penalty_mean': 0.9969308972358704, 'reward_near_duplicate_penalty_std': 0.005684043746441603, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.45557886362075806, 'reward_total_composite_std': 0.3125329315662384, 'epoch': 0.04}

 21%|██▏       | 703/3300 [2:36:55<23:58:34, 33.24s/it]INFO 04-12 18:35:21 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 704/3300 [2:37:05<19:01:55, 26.39s/it]2026-04-12 18:35:31,631 | INFO | train_grpo_train | metrics_logged mode=train step=704

                                                       
{'loss': 0.0259, 'grad_norm': 15.125969886779785, 'learning_rate': 7.86969696969697e-06, 'num_tokens': 1289401.0, 'completions/mean_length': 53.125, 'completions/min_length': 33.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.936540961265564, 'rewards/meter/std': 0.13762077689170837, 'rewards/count_adherence/mean': 0.5750000476837158, 'rewards/count_adherence/std': 0.16690459847450256, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8725000023841858, 'rewards/count_floor/std': 0.05007138103246689, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8017673492431641, 'rewards/repeat_penalty/std': 0.1994551420211792, 'rewards/repeat_floor/mean': 0.9753243923187256, 'rewards/repeat_floor/std': 0.02508958801627159, 'rewards/near_duplicate_penalty/mean': 0.9145094156265259, 'rewards/near_duplicate_penalty/std': 0.07198378443717957, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.894352912902832, 'rewards/distinct_2/std': 0.11988352984189987, 'rewards/total_composite/mean': 0.3104194402694702, 'rewards/total_composite/std': 0.07182146608829498, 'reward': 0.3104194402694702, 'reward_std': 0.07182146608829498, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1721804141998291, 'sampling/sampling_logp_difference/max': 2.0726966857910156, 'sampling/importance_sampling_ratio/min': 0.12584595382213593, 'sampling/importance_sampling_ratio/mean': 1.0132428407669067, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2515020370483398, 'clip_ratio/low_mean': 0.04043938894756138, 'clip_ratio/low_min': 0.04043938894756138, 'clip_ratio/high_mean': 0.07084248960018158, 'clip_ratio/high_max': 0.07084248960018158, 'clip_ratio/region_mean': 0.11128187854774296, 'reward_total_mean': 0.3104194402694702, 'reward_meter_mean': 0.936540961265564, 'reward_meter_std': 0.13762077689170837, 'reward_count_adherence_mean': 0.5750000476837158, 'reward_count_adherence_std': 0.16690459847450256, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8725000023841858, 'reward_count_floor_std': 0.05007138103246689, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8017673492431641, 'reward_repeat_penalty_std': 0.1994551420211792, 'reward_repeat_floor_mean': 0.9753243923187256, 'reward_repeat_floor_std': 0.02508958801627159, 'reward_near_duplicate_penalty_mean': 0.9145094156265259, 'reward_near_duplicate_penalty_std': 0.07198378443717957, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.894352912902832, 'reward_distinct_2_std': 0.11988352984189987, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3104194402694702, 'reward_total_composite_std': 0.07182146608829498, 'epoch': 0.04}

 21%|██▏       | 704/3300 [2:37:05<19:01:55, 26.39s/it]INFO 04-12 18:35:31 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 705/3300 [2:37:13<15:01:19, 20.84s/it]2026-04-12 18:35:39,521 | INFO | train_grpo_train | metrics_logged mode=train step=705

                                                       
{'loss': 0.0232, 'grad_norm': 20.257781982421875, 'learning_rate': 7.866666666666667e-06, 'num_tokens': 1290791.0, 'completions/mean_length': 24.75, 'completions/min_length': 17.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.8726639747619629, 'rewards/meter/std': 0.32054951786994934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9387500286102295, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.7448863983154297, 'rewards/repeat_penalty/std': 0.014463555067777634, 'rewards/repeat_floor/mean': 0.9839773178100586, 'rewards/repeat_floor/std': 0.002892707008868456, 'rewards/near_duplicate_penalty/mean': 0.9931818246841431, 'rewards/near_duplicate_penalty/std': 0.01928473263978958, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.8081232309341431, 'rewards/total_composite/std': 0.2979726493358612, 'reward': 0.8081232309341431, 'reward_std': 0.2979726195335388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10279171913862228, 'sampling/sampling_logp_difference/max': 0.6807451248168945, 'sampling/importance_sampling_ratio/min': 0.506239652633667, 'sampling/importance_sampling_ratio/mean': 1.0390857458114624, 'sampling/importance_sampling_ratio/max': 1.703858733177185, 'entropy': 0.9182189777493477, 'clip_ratio/low_mean': 0.027173912152647972, 'clip_ratio/low_min': 0.027173912152647972, 'clip_ratio/high_mean': 0.08591826166957617, 'clip_ratio/high_max': 0.08591826166957617, 'clip_ratio/region_mean': 0.11309217382222414, 'reward_total_mean': 0.8081232309341431, 'reward_meter_mean': 0.8726639747619629, 'reward_meter_std': 0.32054951786994934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7448863983154297, 'reward_repeat_penalty_std': 0.014463555067777634, 'reward_repeat_floor_mean': 0.9839773178100586, 'reward_repeat_floor_std': 0.002892707008868456, 'reward_near_duplicate_penalty_mean': 0.9931818246841431, 'reward_near_duplicate_penalty_std': 0.01928473263978958, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9387500286102295, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.8081232309341431, 'reward_total_composite_std': 0.2979726493358612, 'epoch': 0.04}

 21%|██▏       | 705/3300 [2:37:13<15:01:19, 20.84s/it]INFO 04-12 18:35:39 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 706/3300 [2:37:28<13:50:05, 19.20s/it]2026-04-12 18:35:54,888 | INFO | train_grpo_train | metrics_logged mode=train step=706

                                                       
{'loss': -0.1667, 'grad_norm': 3.175297498703003, 'learning_rate': 7.863636363636364e-06, 'num_tokens': 1295726.0, 'completions/mean_length': 387.875, 'completions/min_length': 332.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 370.14288330078125, 'completions/min_terminated_length': 332.0, 'completions/max_terminated_length': 421.0, 'rewards/meter/mean': 0.6071051359176636, 'rewards/meter/std': 0.2364199310541153, 'rewards/count_adherence/mean': 0.9097222089767456, 'rewards/count_adherence/std': 0.07821588963270187, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9729167222976685, 'rewards/count_floor/std': 0.02346477098762989, 'rewards/lexical_plausibility/mean': 0.9849501848220825, 'rewards/lexical_plausibility/std': 0.042567286640405655, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.48024994134902954, 'rewards/repeat_penalty/std': 0.33926162123680115, 'rewards/repeat_floor/mean': 0.9413257241249084, 'rewards/repeat_floor/std': 0.04095407575368881, 'rewards/near_duplicate_penalty/mean': 0.8847901821136475, 'rewards/near_duplicate_penalty/std': 0.0914272889494896, 'rewards/opening_diversity/mean': 0.9789685010910034, 'rewards/opening_diversity/std': 0.03901910036802292, 'rewards/distinct_2/mean': 0.5541009306907654, 'rewards/distinct_2/std': 0.3065713942050934, 'rewards/total_composite/mean': 0.17841105163097382, 'rewards/total_composite/std': 0.09289226680994034, 'reward': 0.17841105163097382, 'reward_std': 0.09289226680994034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12477786093950272, 'sampling/sampling_logp_difference/max': 3.1812477111816406, 'sampling/importance_sampling_ratio/min': 0.04153380170464516, 'sampling/importance_sampling_ratio/mean': 1.0115275382995605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6132855415344238, 'clip_ratio/low_mean': 0.022465783171355724, 'clip_ratio/low_min': 0.022465783171355724, 'clip_ratio/high_mean': 0.07483852189034224, 'clip_ratio/high_max': 0.07483852189034224, 'clip_ratio/region_mean': 0.09730430506169796, 'reward_total_mean': 0.17841105163097382, 'reward_meter_mean': 0.6071051359176636, 'reward_meter_std': 0.2364199310541153, 'reward_count_adherence_mean': 0.9097222089767456, 'reward_count_adherence_std': 0.07821588963270187, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9729167222976685, 'reward_count_floor_std': 0.02346477098762989, 'reward_lexical_plausibility_mean': 0.9849501848220825, 'reward_lexical_plausibility_std': 0.042567286640405655, 'reward_repeat_penalty_mean': 0.48024994134902954, 'reward_repeat_penalty_std': 0.33926162123680115, 'reward_repeat_floor_mean': 0.9413257241249084, 'reward_repeat_floor_std': 0.04095407575368881, 'reward_near_duplicate_penalty_mean': 0.8847901821136475, 'reward_near_duplicate_penalty_std': 0.0914272889494896, 'reward_opening_diversity_mean': 0.9789685010910034, 'reward_opening_diversity_std': 0.03901910036802292, 'reward_distinct_2_mean': 0.5541009306907654, 'reward_distinct_2_std': 0.3065713942050934, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.17841105163097382, 'reward_total_composite_std': 0.09289226680994034, 'epoch': 0.04}

 21%|██▏       | 706/3300 [2:37:28<13:50:05, 19.20s/it]INFO 04-12 18:35:55 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 707/3300 [2:37:42<12:35:09, 17.47s/it]2026-04-12 18:36:08,339 | INFO | train_grpo_train | metrics_logged mode=train step=707

                                                       
{'loss': 0.0803, 'grad_norm': 4.710929870605469, 'learning_rate': 7.860606060606062e-06, 'num_tokens': 1297194.0, 'completions/mean_length': 98.5, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.42857360839844, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.47339069843292236, 'rewards/meter/std': 0.47370660305023193, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.1060660183429718, 'rewards/lexical_plausibility/mean': 0.9365938901901245, 'rewards/lexical_plausibility/std': 0.17933957278728485, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.25753986835479736, 'rewards/repeat_penalty/mean': 0.8562992215156555, 'rewards/repeat_penalty/std': 0.12594392895698547, 'rewards/repeat_floor/mean': 0.9872283935546875, 'rewards/repeat_floor/std': 0.011465528979897499, 'rewards/near_duplicate_penalty/mean': 0.9612317681312561, 'rewards/near_duplicate_penalty/std': 0.05465776473283768, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.2304283231496811, 'rewards/total_composite/std': 0.27335724234580994, 'reward': 0.2304283231496811, 'reward_std': 0.27335724234580994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15772601962089539, 'sampling/sampling_logp_difference/max': 3.301652669906616, 'sampling/importance_sampling_ratio/min': 0.03682226315140724, 'sampling/importance_sampling_ratio/mean': 1.0255366563796997, 'sampling/importance_sampling_ratio/max': 1.865495204925537, 'entropy': 1.237963706254959, 'clip_ratio/low_mean': 0.05942331533879042, 'clip_ratio/low_min': 0.05942331533879042, 'clip_ratio/high_mean': 0.08035714458674192, 'clip_ratio/high_max': 0.08035714458674192, 'clip_ratio/region_mean': 0.13978045992553234, 'reward_total_mean': 0.2304283231496811, 'reward_meter_mean': 0.47339069843292236, 'reward_meter_std': 0.47370660305023193, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.1060660183429718, 'reward_lexical_plausibility_mean': 0.9365938901901245, 'reward_lexical_plausibility_std': 0.17933957278728485, 'reward_repeat_penalty_mean': 0.8562992215156555, 'reward_repeat_penalty_std': 0.12594392895698547, 'reward_repeat_floor_mean': 0.9872283935546875, 'reward_repeat_floor_std': 0.011465528979897499, 'reward_near_duplicate_penalty_mean': 0.9612317681312561, 'reward_near_duplicate_penalty_std': 0.05465776473283768, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.25753986835479736, 'reward_total_composite_mean': 0.2304283231496811, 'reward_total_composite_std': 0.27335724234580994, 'epoch': 0.04}

 21%|██▏       | 707/3300 [2:37:42<12:35:09, 17.47s/it]INFO 04-12 18:36:08 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 708/3300 [2:37:56<11:54:01, 16.53s/it]2026-04-12 18:36:22,659 | INFO | train_grpo_train | metrics_logged mode=train step=708

                                                       
{'loss': -0.007, 'grad_norm': 2.5022945404052734, 'learning_rate': 7.857575757575759e-06, 'num_tokens': 1298702.0, 'completions/mean_length': 208.5, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 26.399999618530273, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.28408458828926086, 'rewards/meter/std': 0.3408469259738922, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.8075497150421143, 'rewards/lexical_plausibility/std': 0.2512689530849457, 'rewards/judge_quality/mean': 0.23375001549720764, 'rewards/judge_quality/std': 0.29442134499549866, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.030608050525188446, 'rewards/total_composite/std': 0.042746976017951965, 'reward': 0.030608050525188446, 'reward_std': 0.04274697229266167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17112420499324799, 'sampling/sampling_logp_difference/max': 1.025547981262207, 'sampling/importance_sampling_ratio/min': 0.3585999310016632, 'sampling/importance_sampling_ratio/mean': 1.0705456733703613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.240393042564392, 'clip_ratio/low_mean': 0.06395432632416487, 'clip_ratio/low_min': 0.06395432632416487, 'clip_ratio/high_mean': 0.02654682332649827, 'clip_ratio/high_max': 0.02654682332649827, 'clip_ratio/region_mean': 0.09050114965066314, 'reward_total_mean': 0.030608050525188446, 'reward_meter_mean': 0.28408458828926086, 'reward_meter_std': 0.3408469259738922, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.8075497150421143, 'reward_lexical_plausibility_std': 0.2512689530849457, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.23375001549720764, 'reward_judge_quality_std': 0.29442134499549866, 'reward_total_composite_mean': 0.030608050525188446, 'reward_total_composite_std': 0.042746976017951965, 'epoch': 0.04}

 21%|██▏       | 708/3300 [2:37:56<11:54:01, 16.53s/it]INFO 04-12 18:36:22 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 709/3300 [2:38:10<11:15:37, 15.65s/it]2026-04-12 18:36:36,243 | INFO | train_grpo_train | metrics_logged mode=train step=709

                                                       
{'loss': -0.0131, 'grad_norm': 5.288625240325928, 'learning_rate': 7.854545454545454e-06, 'num_tokens': 1300120.0, 'completions/mean_length': 87.25, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.571430206298828, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5989590883255005, 'rewards/meter/std': 0.4722652733325958, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.08017837256193161, 'rewards/lexical_plausibility/mean': 0.935215950012207, 'rewards/lexical_plausibility/std': 0.18323691189289093, 'rewards/judge_quality/mean': 0.6974999904632568, 'rewards/judge_quality/std': 0.3388109505176544, 'rewards/repeat_penalty/mean': 0.7610735297203064, 'rewards/repeat_penalty/std': 0.11817479878664017, 'rewards/repeat_floor/mean': 0.9808283448219299, 'rewards/repeat_floor/std': 0.01071706973016262, 'rewards/near_duplicate_penalty/mean': 0.9629229307174683, 'rewards/near_duplicate_penalty/std': 0.052938204258680344, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9737762212753296, 'rewards/distinct_2/std': 0.05202964320778847, 'rewards/total_composite/mean': 0.41022107005119324, 'rewards/total_composite/std': 0.34409695863723755, 'reward': 0.41022107005119324, 'reward_std': 0.34409695863723755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16866233944892883, 'sampling/sampling_logp_difference/max': 1.2423224449157715, 'sampling/importance_sampling_ratio/min': 0.28871291875839233, 'sampling/importance_sampling_ratio/mean': 1.0020458698272705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9851214364171028, 'clip_ratio/low_mean': 0.06666666828095913, 'clip_ratio/low_min': 0.06666666828095913, 'clip_ratio/high_mean': 0.09111069655045867, 'clip_ratio/high_max': 0.09111069655045867, 'clip_ratio/region_mean': 0.1577773648314178, 'reward_total_mean': 0.41022107005119324, 'reward_meter_mean': 0.5989590883255005, 'reward_meter_std': 0.4722652733325958, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.08017837256193161, 'reward_lexical_plausibility_mean': 0.935215950012207, 'reward_lexical_plausibility_std': 0.18323691189289093, 'reward_repeat_penalty_mean': 0.7610735297203064, 'reward_repeat_penalty_std': 0.11817479878664017, 'reward_repeat_floor_mean': 0.9808283448219299, 'reward_repeat_floor_std': 0.01071706973016262, 'reward_near_duplicate_penalty_mean': 0.9629229307174683, 'reward_near_duplicate_penalty_std': 0.052938204258680344, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9737762212753296, 'reward_distinct_2_std': 0.05202964320778847, 'reward_judge_quality_mean': 0.6974999904632568, 'reward_judge_quality_std': 0.3388109505176544, 'reward_total_composite_mean': 0.41022107005119324, 'reward_total_composite_std': 0.34409695863723755, 'epoch': 0.04}

 21%|██▏       | 709/3300 [2:38:10<11:15:37, 15.65s/it]INFO 04-12 18:36:36 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 710/3300 [2:38:18<9:43:51, 13.53s/it] 2026-04-12 18:36:44,821 | INFO | train_grpo_train | metrics_logged mode=train step=710

                                                      
{'loss': 0.3582, 'grad_norm': 18.180816650390625, 'learning_rate': 7.851515151515152e-06, 'num_tokens': 1301607.0, 'completions/mean_length': 32.875, 'completions/min_length': 21.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5414481163024902, 'rewards/meter/std': 0.3958565890789032, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.8426564931869507, 'rewards/repeat_penalty/std': 0.12789233028888702, 'rewards/repeat_floor/mean': 0.990460991859436, 'rewards/repeat_floor/std': 0.007542266510426998, 'rewards/near_duplicate_penalty/mean': 0.9989064931869507, 'rewards/near_duplicate_penalty/std': 0.0024298341013491154, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1940850466489792, 'rewards/total_composite/std': 0.13349927961826324, 'reward': 0.1940850466489792, 'reward_std': 0.13349927961826324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1756649762392044, 'sampling/sampling_logp_difference/max': 1.2156119346618652, 'sampling/importance_sampling_ratio/min': 0.29652851819992065, 'sampling/importance_sampling_ratio/mean': 1.0077542066574097, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4264952093362808, 'clip_ratio/low_mean': 0.06860941834747791, 'clip_ratio/low_min': 0.06860941834747791, 'clip_ratio/high_mean': 0.08666666829958558, 'clip_ratio/high_max': 0.08666666829958558, 'clip_ratio/region_mean': 0.1552760866470635, 'reward_total_mean': 0.1940850466489792, 'reward_meter_mean': 0.5414481163024902, 'reward_meter_std': 0.3958565890789032, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8426564931869507, 'reward_repeat_penalty_std': 0.12789233028888702, 'reward_repeat_floor_mean': 0.990460991859436, 'reward_repeat_floor_std': 0.007542266510426998, 'reward_near_duplicate_penalty_mean': 0.9989064931869507, 'reward_near_duplicate_penalty_std': 0.0024298341013491154, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.1940850466489792, 'reward_total_composite_std': 0.13349927961826324, 'epoch': 0.04}

 22%|██▏       | 710/3300 [2:38:18<9:43:51, 13.53s/it]INFO 04-12 18:36:45 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 711/3300 [2:38:26<8:29:49, 11.82s/it]2026-04-12 18:36:52,647 | INFO | train_grpo_train | metrics_logged mode=train step=711

                                                      
{'loss': 0.0395, 'grad_norm': 26.305583953857422, 'learning_rate': 7.848484848484849e-06, 'num_tokens': 1303032.0, 'completions/mean_length': 19.125, 'completions/min_length': 17.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.9515354633331299, 'rewards/meter/std': 0.06079068034887314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45500001311302185, 'rewards/judge_quality/std': 0.31883716583251953, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.42562514543533325, 'rewards/total_composite/std': 0.30652469396591187, 'reward': 0.42562514543533325, 'reward_std': 0.30652469396591187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18814878165721893, 'sampling/sampling_logp_difference/max': 1.936065673828125, 'sampling/importance_sampling_ratio/min': 0.14427043497562408, 'sampling/importance_sampling_ratio/mean': 1.01978600025177, 'sampling/importance_sampling_ratio/max': 1.8287224769592285, 'entropy': 1.260237842798233, 'clip_ratio/low_mean': 0.10735294269397855, 'clip_ratio/low_min': 0.10735294269397855, 'clip_ratio/high_mean': 0.01315789483487606, 'clip_ratio/high_max': 0.01315789483487606, 'clip_ratio/region_mean': 0.12051083752885461, 'reward_total_mean': 0.42562514543533325, 'reward_meter_mean': 0.9515354633331299, 'reward_meter_std': 0.06079068034887314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45500001311302185, 'reward_judge_quality_std': 0.31883716583251953, 'reward_total_composite_mean': 0.42562514543533325, 'reward_total_composite_std': 0.30652469396591187, 'epoch': 0.04}

 22%|██▏       | 711/3300 [2:38:26<8:29:49, 11.82s/it]INFO 04-12 18:36:52 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 712/3300 [2:38:40<8:53:37, 12.37s/it]2026-04-12 18:37:06,316 | INFO | train_grpo_train | metrics_logged mode=train step=712

                                                      
{'loss': -0.0209, 'grad_norm': 9.332262992858887, 'learning_rate': 7.845454545454546e-06, 'num_tokens': 1304732.0, 'completions/mean_length': 93.5, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.71428680419922, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.39751294255256653, 'rewards/meter/std': 0.41686782240867615, 'rewards/count_adherence/mean': 0.3500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8049999475479126, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.9586276412010193, 'rewards/lexical_plausibility/std': 0.11701866239309311, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.8975834846496582, 'rewards/repeat_penalty/std': 0.09574146568775177, 'rewards/repeat_floor/mean': 0.9902625679969788, 'rewards/repeat_floor/std': 0.005486474838107824, 'rewards/near_duplicate_penalty/mean': 0.9600834846496582, 'rewards/near_duplicate_penalty/std': 0.0384177640080452, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.07124805450439453, 'rewards/total_composite/std': 0.09192438423633575, 'reward': 0.07124805450439453, 'reward_std': 0.09192438423633575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.205689936876297, 'sampling/sampling_logp_difference/max': 2.313621997833252, 'sampling/importance_sampling_ratio/min': 0.0989023745059967, 'sampling/importance_sampling_ratio/mean': 1.0081510543823242, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4003472030162811, 'clip_ratio/low_mean': 0.13853716012090445, 'clip_ratio/low_min': 0.13853716012090445, 'clip_ratio/high_mean': 0.057444853708148, 'clip_ratio/high_max': 0.057444853708148, 'clip_ratio/region_mean': 0.19598201382905245, 'reward_total_mean': 0.07124805450439453, 'reward_meter_mean': 0.39751294255256653, 'reward_meter_std': 0.41686782240867615, 'reward_count_adherence_mean': 0.3500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8049999475479126, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.9586276412010193, 'reward_lexical_plausibility_std': 0.11701866239309311, 'reward_repeat_penalty_mean': 0.8975834846496582, 'reward_repeat_penalty_std': 0.09574146568775177, 'reward_repeat_floor_mean': 0.9902625679969788, 'reward_repeat_floor_std': 0.005486474838107824, 'reward_near_duplicate_penalty_mean': 0.9600834846496582, 'reward_near_duplicate_penalty_std': 0.0384177640080452, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.07124805450439453, 'reward_total_composite_std': 0.09192438423633575, 'epoch': 0.04}

 22%|██▏       | 712/3300 [2:38:40<8:53:37, 12.37s/it]INFO 04-12 18:37:06 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 713/3300 [2:38:47<7:50:46, 10.92s/it]2026-04-12 18:37:13,839 | INFO | train_grpo_train | metrics_logged mode=train step=713

                                                      
{'loss': 0.084, 'grad_norm': 22.027294158935547, 'learning_rate': 7.842424242424243e-06, 'num_tokens': 1306327.0, 'completions/mean_length': 30.375, 'completions/min_length': 27.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.4532663822174072, 'rewards/meter/std': 0.35725316405296326, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.3660210967063904, 'rewards/repeat_penalty/mean': 0.9391168355941772, 'rewards/repeat_penalty/std': 0.15032248198986053, 'rewards/repeat_floor/mean': 0.9943631887435913, 'rewards/repeat_floor/std': 0.012695373967289925, 'rewards/near_duplicate_penalty/mean': 0.9931401610374451, 'rewards/near_duplicate_penalty/std': 0.009787586517632008, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9696356058120728, 'rewards/distinct_2/std': 0.08588341623544693, 'rewards/total_composite/mean': 0.3661108911037445, 'rewards/total_composite/std': 0.3673224449157715, 'reward': 0.3661108911037445, 'reward_std': 0.3673224151134491, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16539311408996582, 'sampling/sampling_logp_difference/max': 1.228032112121582, 'sampling/importance_sampling_ratio/min': 0.29286831617355347, 'sampling/importance_sampling_ratio/mean': 1.0022656917572021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1031476855278015, 'clip_ratio/low_mean': 0.0588333741761744, 'clip_ratio/low_min': 0.0588333741761744, 'clip_ratio/high_mean': 0.08416646718978882, 'clip_ratio/high_max': 0.08416646718978882, 'clip_ratio/region_mean': 0.14299984136596322, 'reward_total_mean': 0.3661108911037445, 'reward_meter_mean': 0.4532663822174072, 'reward_meter_std': 0.35725316405296326, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9391168355941772, 'reward_repeat_penalty_std': 0.15032248198986053, 'reward_repeat_floor_mean': 0.9943631887435913, 'reward_repeat_floor_std': 0.012695373967289925, 'reward_near_duplicate_penalty_mean': 0.9931401610374451, 'reward_near_duplicate_penalty_std': 0.009787586517632008, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9696356058120728, 'reward_distinct_2_std': 0.08588341623544693, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.3660210967063904, 'reward_total_composite_mean': 0.3661108911037445, 'reward_total_composite_std': 0.3673224449157715, 'epoch': 0.04}

 22%|██▏       | 713/3300 [2:38:47<7:50:46, 10.92s/it]INFO 04-12 18:37:14 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 714/3300 [2:39:01<8:26:54, 11.76s/it]2026-04-12 18:37:27,572 | INFO | train_grpo_train | metrics_logged mode=train step=714

                                                      
{'loss': -0.0551, 'grad_norm': 4.0953240394592285, 'learning_rate': 7.83939393939394e-06, 'num_tokens': 1307771.0, 'completions/mean_length': 84.5, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.428571701049805, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7046518325805664, 'rewards/meter/std': 0.3900322914123535, 'rewards/count_adherence/mean': 0.3333333432674408, 'rewards/count_adherence/std': 0.1781741827726364, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.800000011920929, 'rewards/count_floor/std': 0.05345224589109421, 'rewards/lexical_plausibility/mean': 0.9429162740707397, 'rewards/lexical_plausibility/std': 0.16145706176757812, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.2706572711467743, 'rewards/repeat_penalty/mean': 0.7986282110214233, 'rewards/repeat_penalty/std': 0.11976303160190582, 'rewards/repeat_floor/mean': 0.9861440062522888, 'rewards/repeat_floor/std': 0.009389491751790047, 'rewards/near_duplicate_penalty/mean': 0.9826266765594482, 'rewards/near_duplicate_penalty/std': 0.03939950838685036, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19854989647865295, 'rewards/total_composite/std': 0.18255853652954102, 'reward': 0.19854989647865295, 'reward_std': 0.18255853652954102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1470773071050644, 'sampling/sampling_logp_difference/max': 1.5912961959838867, 'sampling/importance_sampling_ratio/min': 0.2036614567041397, 'sampling/importance_sampling_ratio/mean': 0.9956262707710266, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9586853459477425, 'clip_ratio/low_mean': 0.08587185200303793, 'clip_ratio/low_min': 0.08587185200303793, 'clip_ratio/high_mean': 0.06704545579850674, 'clip_ratio/high_max': 0.06704545579850674, 'clip_ratio/region_mean': 0.15291730780154467, 'reward_total_mean': 0.19854989647865295, 'reward_meter_mean': 0.7046518325805664, 'reward_meter_std': 0.3900322914123535, 'reward_count_adherence_mean': 0.3333333432674408, 'reward_count_adherence_std': 0.1781741827726364, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.800000011920929, 'reward_count_floor_std': 0.05345224589109421, 'reward_lexical_plausibility_mean': 0.9429162740707397, 'reward_lexical_plausibility_std': 0.16145706176757812, 'reward_repeat_penalty_mean': 0.7986282110214233, 'reward_repeat_penalty_std': 0.11976303160190582, 'reward_repeat_floor_mean': 0.9861440062522888, 'reward_repeat_floor_std': 0.009389491751790047, 'reward_near_duplicate_penalty_mean': 0.9826266765594482, 'reward_near_duplicate_penalty_std': 0.03939950838685036, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.2706572711467743, 'reward_total_composite_mean': 0.19854989647865295, 'reward_total_composite_std': 0.18255853652954102, 'epoch': 0.04}

 22%|██▏       | 714/3300 [2:39:01<8:26:54, 11.76s/it]INFO 04-12 18:37:27 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 715/3300 [2:39:09<7:41:02, 10.70s/it]2026-04-12 18:37:35,829 | INFO | train_grpo_train | metrics_logged mode=train step=715

                                                      
{'loss': 0.2006, 'grad_norm': 18.80209732055664, 'learning_rate': 7.836363636363638e-06, 'num_tokens': 1309272.0, 'completions/mean_length': 26.625, 'completions/min_length': 20.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9644767045974731, 'rewards/meter/std': 0.07078598439693451, 'rewards/count_adherence/mean': 0.4166666865348816, 'rewards/count_adherence/std': 0.15430335700511932, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8250000476837158, 'rewards/count_floor/std': 0.046290989965200424, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.21738296747207642, 'rewards/repeat_penalty/mean': 0.7123791575431824, 'rewards/repeat_penalty/std': 0.1804841309785843, 'rewards/repeat_floor/mean': 0.974851131439209, 'rewards/repeat_floor/std': 0.02668357454240322, 'rewards/near_duplicate_penalty/mean': 0.9565541744232178, 'rewards/near_duplicate_penalty/std': 0.09183619916439056, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.938811182975769, 'rewards/distinct_2/std': 0.12641409039497375, 'rewards/total_composite/mean': 0.3323729634284973, 'rewards/total_composite/std': 0.12196414172649384, 'reward': 0.3323729634284973, 'reward_std': 0.12196414917707443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1371697336435318, 'sampling/sampling_logp_difference/max': 1.8691492080688477, 'sampling/importance_sampling_ratio/min': 0.15425483882427216, 'sampling/importance_sampling_ratio/mean': 1.0337226390838623, 'sampling/importance_sampling_ratio/max': 1.9927793741226196, 'entropy': 0.8754261136054993, 'clip_ratio/low_mean': 0.029415836557745934, 'clip_ratio/low_min': 0.029415836557745934, 'clip_ratio/high_mean': 0.06942287972196937, 'clip_ratio/high_max': 0.06942287972196937, 'clip_ratio/region_mean': 0.0988387162797153, 'reward_total_mean': 0.3323729634284973, 'reward_meter_mean': 0.9644767045974731, 'reward_meter_std': 0.07078598439693451, 'reward_count_adherence_mean': 0.4166666865348816, 'reward_count_adherence_std': 0.15430335700511932, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8250000476837158, 'reward_count_floor_std': 0.046290989965200424, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7123791575431824, 'reward_repeat_penalty_std': 0.1804841309785843, 'reward_repeat_floor_mean': 0.974851131439209, 'reward_repeat_floor_std': 0.02668357454240322, 'reward_near_duplicate_penalty_mean': 0.9565541744232178, 'reward_near_duplicate_penalty_std': 0.09183619916439056, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.938811182975769, 'reward_distinct_2_std': 0.12641409039497375, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.21738296747207642, 'reward_total_composite_mean': 0.3323729634284973, 'reward_total_composite_std': 0.12196414172649384, 'epoch': 0.04}

 22%|██▏       | 715/3300 [2:39:09<7:41:02, 10.70s/it]INFO 04-12 18:37:36 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 716/3300 [2:39:18<7:17:36, 10.16s/it]2026-04-12 18:37:44,701 | INFO | train_grpo_train | metrics_logged mode=train step=716

                                                      
{'loss': -0.0839, 'grad_norm': 20.68985939025879, 'learning_rate': 7.833333333333333e-06, 'num_tokens': 1310694.0, 'completions/mean_length': 27.75, 'completions/min_length': 21.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.75, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5627568960189819, 'rewards/meter/std': 0.4417295753955841, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.0289318785071373, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.7838326692581177, 'rewards/repeat_penalty/std': 0.15512166917324066, 'rewards/repeat_floor/mean': 0.9830164909362793, 'rewards/repeat_floor/std': 0.018982628360390663, 'rewards/near_duplicate_penalty/mean': 0.961776852607727, 'rewards/near_duplicate_penalty/std': 0.10811136662960052, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24523702263832092, 'rewards/total_composite/std': 0.2448812872171402, 'reward': 0.24523702263832092, 'reward_std': 0.2448812574148178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17926619946956635, 'sampling/sampling_logp_difference/max': 2.0898122787475586, 'sampling/importance_sampling_ratio/min': 0.12371035665273666, 'sampling/importance_sampling_ratio/mean': 1.0460915565490723, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4967299550771713, 'clip_ratio/low_mean': 0.05483974330127239, 'clip_ratio/low_min': 0.05483974330127239, 'clip_ratio/high_mean': 0.08721532393246889, 'clip_ratio/high_max': 0.08721532393246889, 'clip_ratio/region_mean': 0.14205506723374128, 'reward_total_mean': 0.24523702263832092, 'reward_meter_mean': 0.5627568960189819, 'reward_meter_std': 0.4417295753955841, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.0289318785071373, 'reward_repeat_penalty_mean': 0.7838326692581177, 'reward_repeat_penalty_std': 0.15512166917324066, 'reward_repeat_floor_mean': 0.9830164909362793, 'reward_repeat_floor_std': 0.018982628360390663, 'reward_near_duplicate_penalty_mean': 0.961776852607727, 'reward_near_duplicate_penalty_std': 0.10811136662960052, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.24523702263832092, 'reward_total_composite_std': 0.2448812872171402, 'epoch': 0.04}

 22%|██▏       | 716/3300 [2:39:18<7:17:36, 10.16s/it]INFO 04-12 18:37:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 717/3300 [2:39:33<8:17:17, 11.55s/it]2026-04-12 18:37:59,504 | INFO | train_grpo_train | metrics_logged mode=train step=717

                                                      
{'loss': -0.079, 'grad_norm': 10.555947303771973, 'learning_rate': 7.83030303030303e-06, 'num_tokens': 1312456.0, 'completions/mean_length': 168.25, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.66666793823242, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.503504753112793, 'rewards/meter/std': 0.3894895315170288, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8968750238418579, 'rewards/count_floor/std': 0.05580179765820503, 'rewards/lexical_plausibility/mean': 0.8695286512374878, 'rewards/lexical_plausibility/std': 0.19863541424274445, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.9882090091705322, 'rewards/repeat_penalty/std': 0.01315134484320879, 'rewards/repeat_floor/mean': 0.9982313513755798, 'rewards/repeat_floor/std': 0.001972706289961934, 'rewards/near_duplicate_penalty/mean': 0.9882090091705322, 'rewards/near_duplicate_penalty/std': 0.01315134484320879, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18480104207992554, 'rewards/total_composite/std': 0.1589871495962143, 'reward': 0.18480104207992554, 'reward_std': 0.1589871495962143, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16045673191547394, 'sampling/sampling_logp_difference/max': 1.2664761543273926, 'sampling/importance_sampling_ratio/min': 0.28182297945022583, 'sampling/importance_sampling_ratio/mean': 1.010556697845459, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7875117212533951, 'clip_ratio/low_mean': 0.038333334028720856, 'clip_ratio/low_min': 0.038333334028720856, 'clip_ratio/high_mean': 0.07271571457386017, 'clip_ratio/high_max': 0.07271571457386017, 'clip_ratio/region_mean': 0.11104904860258102, 'reward_total_mean': 0.18480104207992554, 'reward_meter_mean': 0.503504753112793, 'reward_meter_std': 0.3894895315170288, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8968750238418579, 'reward_count_floor_std': 0.05580179765820503, 'reward_lexical_plausibility_mean': 0.8695286512374878, 'reward_lexical_plausibility_std': 0.19863541424274445, 'reward_repeat_penalty_mean': 0.9882090091705322, 'reward_repeat_penalty_std': 0.01315134484320879, 'reward_repeat_floor_mean': 0.9982313513755798, 'reward_repeat_floor_std': 0.001972706289961934, 'reward_near_duplicate_penalty_mean': 0.9882090091705322, 'reward_near_duplicate_penalty_std': 0.01315134484320879, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.18480104207992554, 'reward_total_composite_std': 0.1589871495962143, 'epoch': 0.04}

 22%|██▏       | 717/3300 [2:39:33<8:17:17, 11.55s/it]INFO 04-12 18:37:59 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 718/3300 [2:39:47<8:53:04, 12.39s/it]2026-04-12 18:38:13,847 | INFO | train_grpo_train | metrics_logged mode=train step=718

                                                      
{'loss': -0.0327, 'grad_norm': 3.4864449501037598, 'learning_rate': 7.827272727272728e-06, 'num_tokens': 1314017.0, 'completions/mean_length': 144.125, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.6304418444633484, 'rewards/meter/std': 0.48267027735710144, 'rewards/count_adherence/mean': 0.28125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7843749523162842, 'rewards/count_floor/std': 0.026516521349549294, 'rewards/lexical_plausibility/mean': 0.8447605967521667, 'rewards/lexical_plausibility/std': 0.24316036701202393, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2881684899330139, 'rewards/repeat_penalty/mean': 0.7706804871559143, 'rewards/repeat_penalty/std': 0.0672103762626648, 'rewards/repeat_floor/mean': 0.9851903319358826, 'rewards/repeat_floor/std': 0.0026342247147113085, 'rewards/near_duplicate_penalty/mean': 0.9891163110733032, 'rewards/near_duplicate_penalty/std': 0.021902840584516525, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9994207620620728, 'rewards/distinct_2/std': 0.0016383323818445206, 'rewards/total_composite/mean': 0.135750412940979, 'rewards/total_composite/std': 0.1332475244998932, 'reward': 0.135750412940979, 'reward_std': 0.1332475244998932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15625827014446259, 'sampling/sampling_logp_difference/max': 1.0226421356201172, 'sampling/importance_sampling_ratio/min': 0.49665215611457825, 'sampling/importance_sampling_ratio/mean': 1.056526780128479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8905178010463715, 'clip_ratio/low_mean': 0.03348370920866728, 'clip_ratio/low_min': 0.03348370920866728, 'clip_ratio/high_mean': 0.060065790079534054, 'clip_ratio/high_max': 0.060065790079534054, 'clip_ratio/region_mean': 0.09354949928820133, 'reward_total_mean': 0.135750412940979, 'reward_meter_mean': 0.6304418444633484, 'reward_meter_std': 0.48267027735710144, 'reward_count_adherence_mean': 0.28125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7843749523162842, 'reward_count_floor_std': 0.026516521349549294, 'reward_lexical_plausibility_mean': 0.8447605967521667, 'reward_lexical_plausibility_std': 0.24316036701202393, 'reward_repeat_penalty_mean': 0.7706804871559143, 'reward_repeat_penalty_std': 0.0672103762626648, 'reward_repeat_floor_mean': 0.9851903319358826, 'reward_repeat_floor_std': 0.0026342247147113085, 'reward_near_duplicate_penalty_mean': 0.9891163110733032, 'reward_near_duplicate_penalty_std': 0.021902840584516525, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9994207620620728, 'reward_distinct_2_std': 0.0016383323818445206, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2881684899330139, 'reward_total_composite_mean': 0.135750412940979, 'reward_total_composite_std': 0.1332475244998932, 'epoch': 0.04}

 22%|██▏       | 718/3300 [2:39:47<8:53:04, 12.39s/it]INFO 04-12 18:38:14 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 719/3300 [2:40:01<9:10:25, 12.80s/it]2026-04-12 18:38:27,595 | INFO | train_grpo_train | metrics_logged mode=train step=719

                                                      
{'loss': -0.0403, 'grad_norm': 6.253236293792725, 'learning_rate': 7.824242424242425e-06, 'num_tokens': 1315463.0, 'completions/mean_length': 87.75, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 27.142858505249023, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.7696630954742432, 'rewards/meter/std': 0.3979122042655945, 'rewards/count_adherence/mean': 0.375, 'rewards/count_adherence/std': 0.1178511381149292, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8125, 'rewards/count_floor/std': 0.035355325788259506, 'rewards/lexical_plausibility/mean': 0.9454492330551147, 'rewards/lexical_plausibility/std': 0.1542929708957672, 'rewards/judge_quality/mean': 0.28999999165534973, 'rewards/judge_quality/std': 0.2899753749370575, 'rewards/repeat_penalty/mean': 0.7734279632568359, 'rewards/repeat_penalty/std': 0.06626439094543457, 'rewards/repeat_floor/mean': 0.9857016801834106, 'rewards/repeat_floor/std': 0.001984695205464959, 'rewards/near_duplicate_penalty/mean': 0.9921779632568359, 'rewards/near_duplicate_penalty/std': 0.02212395705282688, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12374009937047958, 'rewards/total_composite/std': 0.10643817484378815, 'reward': 0.12374009937047958, 'reward_std': 0.10643817484378815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17211350798606873, 'sampling/sampling_logp_difference/max': 1.4356813430786133, 'sampling/importance_sampling_ratio/min': 0.23795318603515625, 'sampling/importance_sampling_ratio/mean': 1.044828176498413, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.346255600452423, 'clip_ratio/low_mean': 0.05486994981765747, 'clip_ratio/low_min': 0.05486994981765747, 'clip_ratio/high_mean': 0.03685314720496535, 'clip_ratio/high_max': 0.03685314720496535, 'clip_ratio/region_mean': 0.09172309702262282, 'reward_total_mean': 0.12374009937047958, 'reward_meter_mean': 0.7696630954742432, 'reward_meter_std': 0.3979122042655945, 'reward_count_adherence_mean': 0.375, 'reward_count_adherence_std': 0.1178511381149292, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8125, 'reward_count_floor_std': 0.035355325788259506, 'reward_lexical_plausibility_mean': 0.9454492330551147, 'reward_lexical_plausibility_std': 0.1542929708957672, 'reward_repeat_penalty_mean': 0.7734279632568359, 'reward_repeat_penalty_std': 0.06626439094543457, 'reward_repeat_floor_mean': 0.9857016801834106, 'reward_repeat_floor_std': 0.001984695205464959, 'reward_near_duplicate_penalty_mean': 0.9921779632568359, 'reward_near_duplicate_penalty_std': 0.02212395705282688, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.28999999165534973, 'reward_judge_quality_std': 0.2899753749370575, 'reward_total_composite_mean': 0.12374009937047958, 'reward_total_composite_std': 0.10643817484378815, 'epoch': 0.04}

 22%|██▏       | 719/3300 [2:40:01<9:10:25, 12.80s/it]INFO 04-12 18:38:27 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 720/3300 [2:40:14<9:11:50, 12.83s/it]2026-04-12 18:38:40,524 | INFO | train_grpo_train | metrics_logged mode=train step=720

                                                      
{'loss': -0.0619, 'grad_norm': 2.590583324432373, 'learning_rate': 7.821212121212122e-06, 'num_tokens': 1316860.0, 'completions/mean_length': 148.625, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7158470153808594, 'rewards/meter/std': 0.4405026435852051, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.891234278678894, 'rewards/lexical_plausibility/std': 0.2013997733592987, 'rewards/judge_quality/mean': 0.6499999761581421, 'rewards/judge_quality/std': 0.4242640733718872, 'rewards/repeat_penalty/mean': 0.96875, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9981250166893005, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6014442443847656, 'rewards/total_composite/std': 0.42559149861335754, 'reward': 0.6014442443847656, 'reward_std': 0.42559149861335754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.095844566822052, 'sampling/sampling_logp_difference/max': 1.1192054748535156, 'sampling/importance_sampling_ratio/min': 0.3711637854576111, 'sampling/importance_sampling_ratio/mean': 1.0188664197921753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26757000386714935, 'clip_ratio/low_mean': 0.02777777798473835, 'clip_ratio/low_min': 0.02777777798473835, 'clip_ratio/high_mean': 0.04195601865649223, 'clip_ratio/high_max': 0.04195601865649223, 'clip_ratio/region_mean': 0.06973379664123058, 'reward_total_mean': 0.6014442443847656, 'reward_meter_mean': 0.7158470153808594, 'reward_meter_std': 0.4405026435852051, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.891234278678894, 'reward_lexical_plausibility_std': 0.2013997733592987, 'reward_repeat_penalty_mean': 0.96875, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9981250166893005, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6499999761581421, 'reward_judge_quality_std': 0.4242640733718872, 'reward_total_composite_mean': 0.6014442443847656, 'reward_total_composite_std': 0.42559149861335754, 'epoch': 0.04}

 22%|██▏       | 720/3300 [2:40:14<9:11:50, 12.83s/it]INFO 04-12 18:38:40 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 721/3300 [2:40:22<8:08:19, 11.36s/it]2026-04-12 18:38:48,463 | INFO | train_grpo_train | metrics_logged mode=train step=721

                                                      
{'loss': -0.0043, 'grad_norm': 18.8549747467041, 'learning_rate': 7.81818181818182e-06, 'num_tokens': 1318356.0, 'completions/mean_length': 32.0, 'completions/min_length': 16.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.12932734191417694, 'rewards/meter/std': 0.17947319149971008, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9312499761581421, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.9036004543304443, 'rewards/repeat_penalty/std': 0.10398723185062408, 'rewards/repeat_floor/mean': 0.9912065267562866, 'rewards/repeat_floor/std': 0.008883717469871044, 'rewards/near_duplicate_penalty/mean': 0.9849158525466919, 'rewards/near_duplicate_penalty/std': 0.016506239771842957, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9482681155204773, 'rewards/distinct_2/std': 0.07615123689174652, 'rewards/total_composite/mean': 0.11863408237695694, 'rewards/total_composite/std': 0.1647283434867859, 'reward': 0.11863408237695694, 'reward_std': 0.16472835838794708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13476267457008362, 'sampling/sampling_logp_difference/max': 1.7197775840759277, 'sampling/importance_sampling_ratio/min': 0.2881323993206024, 'sampling/importance_sampling_ratio/mean': 1.0299296379089355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5904507488012314, 'clip_ratio/low_mean': 0.06712105358019471, 'clip_ratio/low_min': 0.06712105358019471, 'clip_ratio/high_mean': 0.05870911106467247, 'clip_ratio/high_max': 0.05870911106467247, 'clip_ratio/region_mean': 0.12583016464486718, 'reward_total_mean': 0.11863408237695694, 'reward_meter_mean': 0.12932734191417694, 'reward_meter_std': 0.17947319149971008, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9036004543304443, 'reward_repeat_penalty_std': 0.10398723185062408, 'reward_repeat_floor_mean': 0.9912065267562866, 'reward_repeat_floor_std': 0.008883717469871044, 'reward_near_duplicate_penalty_mean': 0.9849158525466919, 'reward_near_duplicate_penalty_std': 0.016506239771842957, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9482681155204773, 'reward_distinct_2_std': 0.07615123689174652, 'reward_judge_quality_mean': 0.9312499761581421, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.11863408237695694, 'reward_total_composite_std': 0.1647283434867859, 'epoch': 0.04}

 22%|██▏       | 721/3300 [2:40:22<8:08:19, 11.36s/it]INFO 04-12 18:38:48 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 722/3300 [2:40:35<8:39:15, 12.09s/it]2026-04-12 18:39:02,536 | INFO | train_grpo_train | metrics_logged mode=train step=722

                                                      
{'loss': -0.0631, 'grad_norm': 3.870074987411499, 'learning_rate': 7.815151515151515e-06, 'num_tokens': 1319716.0, 'completions/mean_length': 82.0, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.571430206298828, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8439614176750183, 'rewards/meter/std': 0.3415564298629761, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.917374849319458, 'rewards/lexical_plausibility/std': 0.17628265917301178, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.38052597641944885, 'rewards/repeat_penalty/mean': 0.7113407850265503, 'rewards/repeat_penalty/std': 0.059995025396347046, 'rewards/repeat_floor/mean': 0.9780111312866211, 'rewards/repeat_floor/std': 0.010424705222249031, 'rewards/near_duplicate_penalty/mean': 0.9630230665206909, 'rewards/near_duplicate_penalty/std': 0.05389808118343353, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.56446373462677, 'rewards/total_composite/std': 0.31382542848587036, 'reward': 0.56446373462677, 'reward_std': 0.31382542848587036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13253246247768402, 'sampling/sampling_logp_difference/max': 1.212634801864624, 'sampling/importance_sampling_ratio/min': 0.297412633895874, 'sampling/importance_sampling_ratio/mean': 1.0298932790756226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8780987411737442, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09667454473674297, 'clip_ratio/high_max': 0.09667454473674297, 'clip_ratio/region_mean': 0.09667454473674297, 'reward_total_mean': 0.56446373462677, 'reward_meter_mean': 0.8439614176750183, 'reward_meter_std': 0.3415564298629761, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.917374849319458, 'reward_lexical_plausibility_std': 0.17628265917301178, 'reward_repeat_penalty_mean': 0.7113407850265503, 'reward_repeat_penalty_std': 0.059995025396347046, 'reward_repeat_floor_mean': 0.9780111312866211, 'reward_repeat_floor_std': 0.010424705222249031, 'reward_near_duplicate_penalty_mean': 0.9630230665206909, 'reward_near_duplicate_penalty_std': 0.05389808118343353, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.38052597641944885, 'reward_total_composite_mean': 0.56446373462677, 'reward_total_composite_std': 0.31382542848587036, 'epoch': 0.04}

 22%|██▏       | 722/3300 [2:40:36<8:39:15, 12.09s/it]INFO 04-12 18:39:02 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 723/3300 [2:40:52<9:32:28, 13.33s/it]2026-04-12 18:39:18,439 | INFO | train_grpo_train | metrics_logged mode=train step=723

                                                      
{'loss': -0.0304, 'grad_norm': 5.2299323081970215, 'learning_rate': 7.812121212121213e-06, 'num_tokens': 1321095.0, 'completions/mean_length': 84.375, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.285715103149414, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6755993366241455, 'rewards/meter/std': 0.43550217151641846, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9109703302383423, 'rewards/lexical_plausibility/std': 0.1917761266231537, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.34280356764793396, 'rewards/repeat_penalty/mean': 0.7712526321411133, 'rewards/repeat_penalty/std': 0.09453253448009491, 'rewards/repeat_floor/mean': 0.9848755598068237, 'rewards/repeat_floor/std': 0.007286030799150467, 'rewards/near_duplicate_penalty/mean': 0.986670196056366, 'rewards/near_duplicate_penalty/std': 0.026991460472345352, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6099381446838379, 'rewards/total_composite/std': 0.39566490054130554, 'reward': 0.6099381446838379, 'reward_std': 0.39566490054130554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13940078020095825, 'sampling/sampling_logp_difference/max': 0.8479976058006287, 'sampling/importance_sampling_ratio/min': 0.42827165126800537, 'sampling/importance_sampling_ratio/mean': 1.018211841583252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8917571529746056, 'clip_ratio/low_mean': 0.015224359463900328, 'clip_ratio/low_min': 0.015224359463900328, 'clip_ratio/high_mean': 0.09005434717983007, 'clip_ratio/high_max': 0.09005434717983007, 'clip_ratio/region_mean': 0.1052787066437304, 'reward_total_mean': 0.6099381446838379, 'reward_meter_mean': 0.6755993366241455, 'reward_meter_std': 0.43550217151641846, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9109703302383423, 'reward_lexical_plausibility_std': 0.1917761266231537, 'reward_repeat_penalty_mean': 0.7712526321411133, 'reward_repeat_penalty_std': 0.09453253448009491, 'reward_repeat_floor_mean': 0.9848755598068237, 'reward_repeat_floor_std': 0.007286030799150467, 'reward_near_duplicate_penalty_mean': 0.986670196056366, 'reward_near_duplicate_penalty_std': 0.026991460472345352, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.34280356764793396, 'reward_total_composite_mean': 0.6099381446838379, 'reward_total_composite_std': 0.39566490054130554, 'epoch': 0.04}

 22%|██▏       | 723/3300 [2:40:52<9:32:28, 13.33s/it]INFO 04-12 18:39:18 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 724/3300 [2:41:09<10:21:22, 14.47s/it]2026-04-12 18:39:35,597 | INFO | train_grpo_train | metrics_logged mode=train step=724

                                                       
{'loss': -0.0272, 'grad_norm': 5.167529106140137, 'learning_rate': 7.80909090909091e-06, 'num_tokens': 1322564.0, 'completions/mean_length': 84.625, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.571430206298828, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.7411857843399048, 'rewards/meter/std': 0.397968053817749, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9382473230361938, 'rewards/lexical_plausibility/std': 0.17466284334659576, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.3744877576828003, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3536131680011749, 'rewards/total_composite/std': 0.3555664122104645, 'reward': 0.3536131680011749, 'reward_std': 0.3555663824081421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15759819746017456, 'sampling/sampling_logp_difference/max': 1.3172483444213867, 'sampling/importance_sampling_ratio/min': 0.2678713798522949, 'sampling/importance_sampling_ratio/mean': 1.0373677015304565, 'sampling/importance_sampling_ratio/max': 1.9701427221298218, 'entropy': 1.4948654025793076, 'clip_ratio/low_mean': 0.06809163186699152, 'clip_ratio/low_min': 0.06809163186699152, 'clip_ratio/high_mean': 0.07676573470234871, 'clip_ratio/high_max': 0.07676573470234871, 'clip_ratio/region_mean': 0.14485736656934023, 'reward_total_mean': 0.3536131680011749, 'reward_meter_mean': 0.7411857843399048, 'reward_meter_std': 0.397968053817749, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9382473230361938, 'reward_lexical_plausibility_std': 0.17466284334659576, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.3744877576828003, 'reward_total_composite_mean': 0.3536131680011749, 'reward_total_composite_std': 0.3555664122104645, 'epoch': 0.04}

 22%|██▏       | 724/3300 [2:41:09<10:21:22, 14.47s/it]INFO 04-12 18:39:35 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 725/3300 [2:41:22<10:02:29, 14.04s/it]2026-04-12 18:39:48,606 | INFO | train_grpo_train | metrics_logged mode=train step=725

                                                       
{'loss': 0.0277, 'grad_norm': 4.60774040222168, 'learning_rate': 7.806060606060607e-06, 'num_tokens': 1323938.0, 'completions/mean_length': 81.75, 'completions/min_length': 13.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.285715103149414, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.07172045856714249, 'rewards/meter/std': 0.15121987462043762, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.938498854637146, 'rewards/lexical_plausibility/std': 0.17395159602165222, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.33961114287376404, 'rewards/repeat_penalty/mean': 0.7383946180343628, 'rewards/repeat_penalty/std': 0.02443493902683258, 'rewards/repeat_floor/mean': 0.9826788902282715, 'rewards/repeat_floor/std': 0.004886987619102001, 'rewards/near_duplicate_penalty/mean': 0.9845261573791504, 'rewards/near_duplicate_penalty/std': 0.03257991746068001, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.04277218505740166, 'rewards/total_composite/std': 0.09623362123966217, 'reward': 0.04277218505740166, 'reward_std': 0.09623362123966217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18291820585727692, 'sampling/sampling_logp_difference/max': 1.1260461807250977, 'sampling/importance_sampling_ratio/min': 0.3243129849433899, 'sampling/importance_sampling_ratio/mean': 1.0450856685638428, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4478702247142792, 'clip_ratio/low_mean': 0.12467185966670513, 'clip_ratio/low_min': 0.12467185966670513, 'clip_ratio/high_mean': 0.0475961547344923, 'clip_ratio/high_max': 0.0475961547344923, 'clip_ratio/region_mean': 0.17226801440119743, 'reward_total_mean': 0.04277218505740166, 'reward_meter_mean': 0.07172045856714249, 'reward_meter_std': 0.15121987462043762, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.938498854637146, 'reward_lexical_plausibility_std': 0.17395159602165222, 'reward_repeat_penalty_mean': 0.7383946180343628, 'reward_repeat_penalty_std': 0.02443493902683258, 'reward_repeat_floor_mean': 0.9826788902282715, 'reward_repeat_floor_std': 0.004886987619102001, 'reward_near_duplicate_penalty_mean': 0.9845261573791504, 'reward_near_duplicate_penalty_std': 0.03257991746068001, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.33961114287376404, 'reward_total_composite_mean': 0.04277218505740166, 'reward_total_composite_std': 0.09623362123966217, 'epoch': 0.04}

 22%|██▏       | 725/3300 [2:41:22<10:02:29, 14.04s/it]INFO 04-12 18:39:48 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 726/3300 [2:41:36<10:03:20, 14.06s/it]2026-04-12 18:40:02,752 | INFO | train_grpo_train | metrics_logged mode=train step=726

                                                       
{'loss': -0.0229, 'grad_norm': 1.2339202165603638, 'learning_rate': 7.803030303030303e-06, 'num_tokens': 1325287.0, 'completions/mean_length': 327.625, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 20.33333396911621, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.376447856426239, 'rewards/meter/std': 0.3692064583301544, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.6865270733833313, 'rewards/lexical_plausibility/std': 0.22714054584503174, 'rewards/judge_quality/mean': 0.10625000298023224, 'rewards/judge_quality/std': 0.11160357296466827, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.039985693991184235, 'rewards/total_composite/std': 0.05159784108400345, 'reward': 0.039985693991184235, 'reward_std': 0.05159783735871315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1866149753332138, 'sampling/sampling_logp_difference/max': 1.3450794219970703, 'sampling/importance_sampling_ratio/min': 0.26051902770996094, 'sampling/importance_sampling_ratio/mean': 1.0675848722457886, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5679557621479034, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08442942425608635, 'clip_ratio/high_max': 0.08442942425608635, 'clip_ratio/region_mean': 0.08442942425608635, 'reward_total_mean': 0.039985693991184235, 'reward_meter_mean': 0.376447856426239, 'reward_meter_std': 0.3692064583301544, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.6865270733833313, 'reward_lexical_plausibility_std': 0.22714054584503174, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.10625000298023224, 'reward_judge_quality_std': 0.11160357296466827, 'reward_total_composite_mean': 0.039985693991184235, 'reward_total_composite_std': 0.05159784108400345, 'epoch': 0.04}

 22%|██▏       | 726/3300 [2:41:36<10:03:20, 14.06s/it]INFO 04-12 18:40:03 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 727/3300 [2:41:52<10:22:00, 14.50s/it]2026-04-12 18:40:18,285 | INFO | train_grpo_train | metrics_logged mode=train step=727

                                                       
{'loss': -0.0122, 'grad_norm': 2.821991205215454, 'learning_rate': 7.800000000000002e-06, 'num_tokens': 1326843.0, 'completions/mean_length': 144.5, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.3461433947086334, 'rewards/meter/std': 0.40159371495246887, 'rewards/count_adherence/mean': 0.17500001192092896, 'rewards/count_adherence/std': 0.0707106813788414, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.7524999976158142, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.8556759357452393, 'rewards/lexical_plausibility/std': 0.2235196977853775, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.7688470482826233, 'rewards/repeat_penalty/std': 0.0959373190999031, 'rewards/repeat_floor/mean': 0.9850017428398132, 'rewards/repeat_floor/std': 0.006775653455406427, 'rewards/near_duplicate_penalty/mean': 0.993584156036377, 'rewards/near_duplicate_penalty/std': 0.015087136998772621, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.08100727945566177, 'rewards/total_composite/std': 0.10784495621919632, 'reward': 0.08100727945566177, 'reward_std': 0.10784494876861572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18328239023685455, 'sampling/sampling_logp_difference/max': 1.1372299194335938, 'sampling/importance_sampling_ratio/min': 0.32070618867874146, 'sampling/importance_sampling_ratio/mean': 1.0604618787765503, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4922990500926971, 'clip_ratio/low_mean': 0.07380952499806881, 'clip_ratio/low_min': 0.07380952499806881, 'clip_ratio/high_mean': 0.07006578985601664, 'clip_ratio/high_max': 0.07006578985601664, 'clip_ratio/region_mean': 0.14387531485408545, 'reward_total_mean': 0.08100727945566177, 'reward_meter_mean': 0.3461433947086334, 'reward_meter_std': 0.40159371495246887, 'reward_count_adherence_mean': 0.17500001192092896, 'reward_count_adherence_std': 0.0707106813788414, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.7524999976158142, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.8556759357452393, 'reward_lexical_plausibility_std': 0.2235196977853775, 'reward_repeat_penalty_mean': 0.7688470482826233, 'reward_repeat_penalty_std': 0.0959373190999031, 'reward_repeat_floor_mean': 0.9850017428398132, 'reward_repeat_floor_std': 0.006775653455406427, 'reward_near_duplicate_penalty_mean': 0.993584156036377, 'reward_near_duplicate_penalty_std': 0.015087136998772621, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.08100727945566177, 'reward_total_composite_std': 0.10784495621919632, 'epoch': 0.04}

 22%|██▏       | 727/3300 [2:41:52<10:22:00, 14.50s/it]INFO 04-12 18:40:18 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 728/3300 [2:42:05<10:06:13, 14.14s/it]2026-04-12 18:40:31,556 | INFO | train_grpo_train | metrics_logged mode=train step=728

                                                       
{'loss': -0.0376, 'grad_norm': 1.3681159019470215, 'learning_rate': 7.796969696969697e-06, 'num_tokens': 1328137.0, 'completions/mean_length': 208.75, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 26.80000114440918, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.5654644966125488, 'rewards/meter/std': 0.47302353382110596, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8394579887390137, 'rewards/lexical_plausibility/std': 0.22180584073066711, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.42570948600769043, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3893899917602539, 'rewards/total_composite/std': 0.43629246950149536, 'reward': 0.3893899917602539, 'reward_std': 0.43629246950149536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16032110154628754, 'sampling/sampling_logp_difference/max': 1.5410747528076172, 'sampling/importance_sampling_ratio/min': 0.2141508162021637, 'sampling/importance_sampling_ratio/mean': 1.0484509468078613, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8937837034463882, 'clip_ratio/low_mean': 0.044871795922517776, 'clip_ratio/low_min': 0.044871795922517776, 'clip_ratio/high_mean': 0.05648065451532602, 'clip_ratio/high_max': 0.05648065451532602, 'clip_ratio/region_mean': 0.1013524504378438, 'reward_total_mean': 0.3893899917602539, 'reward_meter_mean': 0.5654644966125488, 'reward_meter_std': 0.47302353382110596, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8394579887390137, 'reward_lexical_plausibility_std': 0.22180584073066711, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.42570948600769043, 'reward_total_composite_mean': 0.3893899917602539, 'reward_total_composite_std': 0.43629246950149536, 'epoch': 0.04}

 22%|██▏       | 728/3300 [2:42:05<10:06:13, 14.14s/it]INFO 04-12 18:40:31 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 729/3300 [2:42:21<10:36:20, 14.85s/it]2026-04-12 18:40:48,061 | INFO | train_grpo_train | metrics_logged mode=train step=729

                                                       
{'loss': -0.0243, 'grad_norm': 1.1098401546478271, 'learning_rate': 7.793939393939394e-06, 'num_tokens': 1329551.0, 'completions/mean_length': 329.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.33000853657722473, 'rewards/meter/std': 0.37807002663612366, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.7095398902893066, 'rewards/lexical_plausibility/std': 0.25487351417541504, 'rewards/judge_quality/mean': 0.2549999952316284, 'rewards/judge_quality/std': 0.3254886567592621, 'rewards/repeat_penalty/mean': 0.8006408214569092, 'rewards/repeat_penalty/std': 0.09557542949914932, 'rewards/repeat_floor/mean': 0.9869711399078369, 'rewards/repeat_floor/std': 0.004584609065204859, 'rewards/near_duplicate_penalty/mean': 0.9881408214569092, 'rewards/near_duplicate_penalty/std': 0.028710680082440376, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.06835195422172546, 'rewards/total_composite/std': 0.11151061952114105, 'reward': 0.06835195422172546, 'reward_std': 0.11151061207056046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1664460450410843, 'sampling/sampling_logp_difference/max': 1.539482593536377, 'sampling/importance_sampling_ratio/min': 0.214492067694664, 'sampling/importance_sampling_ratio/mean': 1.0515148639678955, 'sampling/importance_sampling_ratio/max': 1.8771865367889404, 'entropy': 0.5704511851072311, 'clip_ratio/low_mean': 0.02083333395421505, 'clip_ratio/low_min': 0.02083333395421505, 'clip_ratio/high_mean': 0.046875, 'clip_ratio/high_max': 0.046875, 'clip_ratio/region_mean': 0.06770833395421505, 'reward_total_mean': 0.06835195422172546, 'reward_meter_mean': 0.33000853657722473, 'reward_meter_std': 0.37807002663612366, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.7095398902893066, 'reward_lexical_plausibility_std': 0.25487351417541504, 'reward_repeat_penalty_mean': 0.8006408214569092, 'reward_repeat_penalty_std': 0.09557542949914932, 'reward_repeat_floor_mean': 0.9869711399078369, 'reward_repeat_floor_std': 0.004584609065204859, 'reward_near_duplicate_penalty_mean': 0.9881408214569092, 'reward_near_duplicate_penalty_std': 0.028710680082440376, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2549999952316284, 'reward_judge_quality_std': 0.3254886567592621, 'reward_total_composite_mean': 0.06835195422172546, 'reward_total_composite_std': 0.11151061952114105, 'epoch': 0.04}

 22%|██▏       | 729/3300 [2:42:21<10:36:20, 14.85s/it]INFO 04-12 18:40:48 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 730/3300 [2:42:29<9:08:37, 12.81s/it] 2026-04-12 18:40:56,119 | INFO | train_grpo_train | metrics_logged mode=train step=730

                                                      
{'loss': -0.0672, 'grad_norm': 21.157039642333984, 'learning_rate': 7.790909090909092e-06, 'num_tokens': 1330947.0, 'completions/mean_length': 19.5, 'completions/min_length': 12.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7525069117546082, 'rewards/meter/std': 0.3796283006668091, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9574281573295593, 'rewards/lexical_plausibility/std': 0.09763853251934052, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3236731290817261, 'rewards/repeat_penalty/mean': 0.7707452774047852, 'rewards/repeat_penalty/std': 0.09718868136405945, 'rewards/repeat_floor/mean': 0.9847740530967712, 'rewards/repeat_floor/std': 0.008511229418218136, 'rewards/near_duplicate_penalty/mean': 0.9859936237335205, 'rewards/near_duplicate_penalty/std': 0.03961591050028801, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33790332078933716, 'rewards/total_composite/std': 0.3171427845954895, 'reward': 0.33790332078933716, 'reward_std': 0.3171428143978119, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1961711347103119, 'sampling/sampling_logp_difference/max': 0.9623818397521973, 'sampling/importance_sampling_ratio/min': 0.3819819688796997, 'sampling/importance_sampling_ratio/mean': 1.0979080200195312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0567077547311783, 'clip_ratio/low_mean': 0.13083964958786964, 'clip_ratio/low_min': 0.13083964958786964, 'clip_ratio/high_mean': 0.08909605676308274, 'clip_ratio/high_max': 0.08909605676308274, 'clip_ratio/region_mean': 0.2199357063509524, 'reward_total_mean': 0.33790332078933716, 'reward_meter_mean': 0.7525069117546082, 'reward_meter_std': 0.3796283006668091, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9574281573295593, 'reward_lexical_plausibility_std': 0.09763853251934052, 'reward_repeat_penalty_mean': 0.7707452774047852, 'reward_repeat_penalty_std': 0.09718868136405945, 'reward_repeat_floor_mean': 0.9847740530967712, 'reward_repeat_floor_std': 0.008511229418218136, 'reward_near_duplicate_penalty_mean': 0.9859936237335205, 'reward_near_duplicate_penalty_std': 0.03961591050028801, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3236731290817261, 'reward_total_composite_mean': 0.33790332078933716, 'reward_total_composite_std': 0.3171427845954895, 'epoch': 0.04}

 22%|██▏       | 730/3300 [2:42:29<9:08:37, 12.81s/it]INFO 04-12 18:40:56 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 731/3300 [2:42:43<9:23:07, 13.15s/it]2026-04-12 18:41:10,081 | INFO | train_grpo_train | metrics_logged mode=train step=731

                                                      
{'loss': -0.118, 'grad_norm': 6.608311653137207, 'learning_rate': 7.787878787878789e-06, 'num_tokens': 1332451.0, 'completions/mean_length': 87.0, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.285715103149414, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.15218953788280487, 'rewards/meter/std': 0.333724707365036, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.08017837256193161, 'rewards/lexical_plausibility/mean': 0.9375545978546143, 'rewards/lexical_plausibility/std': 0.17662228643894196, 'rewards/judge_quality/mean': 0.5787500143051147, 'rewards/judge_quality/std': 0.38794100284576416, 'rewards/repeat_penalty/mean': 0.84375, 'rewards/repeat_penalty/std': 0.12938730418682098, 'rewards/repeat_floor/mean': 0.9906250238418579, 'rewards/repeat_floor/std': 0.00776322977617383, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11903752386569977, 'rewards/total_composite/std': 0.26284700632095337, 'reward': 0.11903752386569977, 'reward_std': 0.26284700632095337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15268048644065857, 'sampling/sampling_logp_difference/max': 1.066246509552002, 'sampling/importance_sampling_ratio/min': 0.34429842233657837, 'sampling/importance_sampling_ratio/mean': 1.070543646812439, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1604615151882172, 'clip_ratio/low_mean': 0.11805515456944704, 'clip_ratio/low_min': 0.11805515456944704, 'clip_ratio/high_mean': 0.050543476827442646, 'clip_ratio/high_max': 0.050543476827442646, 'clip_ratio/region_mean': 0.1685986313968897, 'reward_total_mean': 0.11903752386569977, 'reward_meter_mean': 0.15218953788280487, 'reward_meter_std': 0.333724707365036, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.08017837256193161, 'reward_lexical_plausibility_mean': 0.9375545978546143, 'reward_lexical_plausibility_std': 0.17662228643894196, 'reward_repeat_penalty_mean': 0.84375, 'reward_repeat_penalty_std': 0.12938730418682098, 'reward_repeat_floor_mean': 0.9906250238418579, 'reward_repeat_floor_std': 0.00776322977617383, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5787500143051147, 'reward_judge_quality_std': 0.38794100284576416, 'reward_total_composite_mean': 0.11903752386569977, 'reward_total_composite_std': 0.26284700632095337, 'epoch': 0.04}

 22%|██▏       | 731/3300 [2:42:43<9:23:07, 13.15s/it]INFO 04-12 18:41:10 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 732/3300 [2:42:59<9:57:04, 13.95s/it]2026-04-12 18:41:25,882 | INFO | train_grpo_train | metrics_logged mode=train step=732

                                                      
{'loss': -0.0242, 'grad_norm': 0.8254483342170715, 'learning_rate': 7.784848484848484e-06, 'num_tokens': 1333798.0, 'completions/mean_length': 327.375, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 19.666667938232422, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.26678141951560974, 'rewards/meter/std': 0.3807927370071411, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.6831883192062378, 'rewards/lexical_plausibility/std': 0.19720584154129028, 'rewards/judge_quality/mean': 0.11250000447034836, 'rewards/judge_quality/std': 0.1482035219669342, 'rewards/repeat_penalty/mean': 0.7455357313156128, 'rewards/repeat_penalty/std': 0.012626901268959045, 'rewards/repeat_floor/mean': 0.9841071367263794, 'rewards/repeat_floor/std': 0.002525375923141837, 'rewards/near_duplicate_penalty/mean': 0.9940476417541504, 'rewards/near_duplicate_penalty/std': 0.016835875809192657, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.043202705681324005, 'rewards/total_composite/std': 0.06016030162572861, 'reward': 0.043202705681324005, 'reward_std': 0.06016030162572861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22709383070468903, 'sampling/sampling_logp_difference/max': 1.682027816772461, 'sampling/importance_sampling_ratio/min': 0.1859964281320572, 'sampling/importance_sampling_ratio/mean': 1.0534539222717285, 'sampling/importance_sampling_ratio/max': 1.9848183393478394, 'entropy': 0.7943100333213806, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.05545587558299303, 'clip_ratio/high_max': 0.05545587558299303, 'clip_ratio/region_mean': 0.05545587558299303, 'reward_total_mean': 0.043202705681324005, 'reward_meter_mean': 0.26678141951560974, 'reward_meter_std': 0.3807927370071411, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.6831883192062378, 'reward_lexical_plausibility_std': 0.19720584154129028, 'reward_repeat_penalty_mean': 0.7455357313156128, 'reward_repeat_penalty_std': 0.012626901268959045, 'reward_repeat_floor_mean': 0.9841071367263794, 'reward_repeat_floor_std': 0.002525375923141837, 'reward_near_duplicate_penalty_mean': 0.9940476417541504, 'reward_near_duplicate_penalty_std': 0.016835875809192657, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.11250000447034836, 'reward_judge_quality_std': 0.1482035219669342, 'reward_total_composite_mean': 0.043202705681324005, 'reward_total_composite_std': 0.06016030162572861, 'epoch': 0.04}

 22%|██▏       | 732/3300 [2:42:59<9:57:04, 13.95s/it]INFO 04-12 18:41:26 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 733/3300 [2:43:13<9:56:36, 13.94s/it]2026-04-12 18:41:39,806 | INFO | train_grpo_train | metrics_logged mode=train step=733

                                                      
{'loss': -0.0337, 'grad_norm': 1.1059763431549072, 'learning_rate': 7.781818181818183e-06, 'num_tokens': 1335295.0, 'completions/mean_length': 145.125, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 22.83333396911621, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7005734443664551, 'rewards/meter/std': 0.420141339302063, 'rewards/count_adherence/mean': 0.3125, 'rewards/count_adherence/std': 0.22160132229328156, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.793749988079071, 'rewards/count_floor/std': 0.06648041307926178, 'rewards/lexical_plausibility/mean': 0.9005357027053833, 'rewards/lexical_plausibility/std': 0.18748725950717926, 'rewards/judge_quality/mean': 0.39250001311302185, 'rewards/judge_quality/std': 0.3454500436782837, 'rewards/repeat_penalty/mean': 0.8286024332046509, 'rewards/repeat_penalty/std': 0.14752699434757233, 'rewards/repeat_floor/mean': 0.9875989556312561, 'rewards/repeat_floor/std': 0.013096061535179615, 'rewards/near_duplicate_penalty/mean': 0.9798262119293213, 'rewards/near_duplicate_penalty/std': 0.05683772265911102, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2666128873825073, 'rewards/total_composite/std': 0.2470596432685852, 'reward': 0.2666128873825073, 'reward_std': 0.2470596432685852, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15543679893016815, 'sampling/sampling_logp_difference/max': 1.2707390785217285, 'sampling/importance_sampling_ratio/min': 0.2806241512298584, 'sampling/importance_sampling_ratio/mean': 0.9980720281600952, 'sampling/importance_sampling_ratio/max': 1.7403450012207031, 'entropy': 0.7436570003628731, 'clip_ratio/low_mean': 0.03405748726800084, 'clip_ratio/low_min': 0.03405748726800084, 'clip_ratio/high_mean': 0.06414473801851273, 'clip_ratio/high_max': 0.06414473801851273, 'clip_ratio/region_mean': 0.09820222528651357, 'reward_total_mean': 0.2666128873825073, 'reward_meter_mean': 0.7005734443664551, 'reward_meter_std': 0.420141339302063, 'reward_count_adherence_mean': 0.3125, 'reward_count_adherence_std': 0.22160132229328156, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.793749988079071, 'reward_count_floor_std': 0.06648041307926178, 'reward_lexical_plausibility_mean': 0.9005357027053833, 'reward_lexical_plausibility_std': 0.18748725950717926, 'reward_repeat_penalty_mean': 0.8286024332046509, 'reward_repeat_penalty_std': 0.14752699434757233, 'reward_repeat_floor_mean': 0.9875989556312561, 'reward_repeat_floor_std': 0.013096061535179615, 'reward_near_duplicate_penalty_mean': 0.9798262119293213, 'reward_near_duplicate_penalty_std': 0.05683772265911102, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.39250001311302185, 'reward_judge_quality_std': 0.3454500436782837, 'reward_total_composite_mean': 0.2666128873825073, 'reward_total_composite_std': 0.2470596432685852, 'epoch': 0.04}

 22%|██▏       | 733/3300 [2:43:13<9:56:36, 13.94s/it]INFO 04-12 18:41:40 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 734/3300 [2:43:29<10:16:29, 14.42s/it]2026-04-12 18:41:55,329 | INFO | train_grpo_train | metrics_logged mode=train step=734

                                                       
{'loss': -0.008, 'grad_norm': 0.5768758058547974, 'learning_rate': 7.778787878787879e-06, 'num_tokens': 1336934.0, 'completions/mean_length': 449.875, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 15.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 15.0, 'rewards/meter/mean': 0.143520325422287, 'rewards/meter/std': 0.24628643691539764, 'rewards/count_adherence/mean': 0.15000000596046448, 'rewards/count_adherence/std': 0.1414213627576828, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7450000047683716, 'rewards/count_floor/std': 0.04242641106247902, 'rewards/lexical_plausibility/mean': 0.5566787719726562, 'rewards/lexical_plausibility/std': 0.18475589156150818, 'rewards/judge_quality/mean': 0.05000000074505806, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/repeat_penalty/mean': 0.8745421171188354, 'rewards/repeat_penalty/std': 0.1341254860162735, 'rewards/repeat_floor/mean': 0.9924451112747192, 'rewards/repeat_floor/std': 0.008077853359282017, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9993895292282104, 'rewards/distinct_2/std': 0.0017267564544454217, 'rewards/total_composite/mean': 0.011730743572115898, 'rewards/total_composite/std': 0.02625987119972706, 'reward': 0.011730743572115898, 'reward_std': 0.026259873062372208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.29687464237213135, 'sampling/sampling_logp_difference/max': 1.0584502220153809, 'sampling/importance_sampling_ratio/min': 0.3469931483268738, 'sampling/importance_sampling_ratio/mean': 0.8634982109069824, 'sampling/importance_sampling_ratio/max': 1.2256780862808228, 'entropy': 0.17789536714553833, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.02500000037252903, 'clip_ratio/high_max': 0.02500000037252903, 'clip_ratio/region_mean': 0.02500000037252903, 'reward_total_mean': 0.011730743572115898, 'reward_meter_mean': 0.143520325422287, 'reward_meter_std': 0.24628643691539764, 'reward_count_adherence_mean': 0.15000000596046448, 'reward_count_adherence_std': 0.1414213627576828, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7450000047683716, 'reward_count_floor_std': 0.04242641106247902, 'reward_lexical_plausibility_mean': 0.5566787719726562, 'reward_lexical_plausibility_std': 0.18475589156150818, 'reward_repeat_penalty_mean': 0.8745421171188354, 'reward_repeat_penalty_std': 0.1341254860162735, 'reward_repeat_floor_mean': 0.9924451112747192, 'reward_repeat_floor_std': 0.008077853359282017, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9993895292282104, 'reward_distinct_2_std': 0.0017267564544454217, 'reward_judge_quality_mean': 0.05000000074505806, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.011730743572115898, 'reward_total_composite_std': 0.02625987119972706, 'epoch': 0.04}

 22%|██▏       | 734/3300 [2:43:29<10:16:29, 14.42s/it]INFO 04-12 18:41:55 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 735/3300 [2:43:43<10:17:54, 14.45s/it]2026-04-12 18:42:09,883 | INFO | train_grpo_train | metrics_logged mode=train step=735

                                                       
{'loss': 0.0181, 'grad_norm': 3.2832353115081787, 'learning_rate': 7.775757575757576e-06, 'num_tokens': 1338817.0, 'completions/mean_length': 97.375, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.142860412597656, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.7859209179878235, 'rewards/meter/std': 0.36400628089904785, 'rewards/count_adherence/mean': 0.30000001192092896, 'rewards/count_adherence/std': 0.21380899846553802, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7899999618530273, 'rewards/count_floor/std': 0.06414270401000977, 'rewards/lexical_plausibility/mean': 0.9369398355484009, 'rewards/lexical_plausibility/std': 0.15466144680976868, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.7620465159416199, 'rewards/repeat_penalty/std': 0.1048668697476387, 'rewards/repeat_floor/mean': 0.9820486307144165, 'rewards/repeat_floor/std': 0.011632430367171764, 'rewards/near_duplicate_penalty/mean': 0.9737856388092041, 'rewards/near_duplicate_penalty/std': 0.057021964341402054, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9692307710647583, 'rewards/distinct_2/std': 0.08702851831912994, 'rewards/total_composite/mean': 0.12267211079597473, 'rewards/total_composite/std': 0.07322044670581818, 'reward': 0.12267211079597473, 'reward_std': 0.07322043925523758, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20950034260749817, 'sampling/sampling_logp_difference/max': 3.6155519485473633, 'sampling/importance_sampling_ratio/min': 0.026902073994278908, 'sampling/importance_sampling_ratio/mean': 1.0456829071044922, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2990215495228767, 'clip_ratio/low_mean': 0.11206033732742071, 'clip_ratio/low_min': 0.11206033732742071, 'clip_ratio/high_mean': 0.06185300089418888, 'clip_ratio/high_max': 0.06185300089418888, 'clip_ratio/region_mean': 0.1739133382216096, 'reward_total_mean': 0.12267211079597473, 'reward_meter_mean': 0.7859209179878235, 'reward_meter_std': 0.36400628089904785, 'reward_count_adherence_mean': 0.30000001192092896, 'reward_count_adherence_std': 0.21380899846553802, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7899999618530273, 'reward_count_floor_std': 0.06414270401000977, 'reward_lexical_plausibility_mean': 0.9369398355484009, 'reward_lexical_plausibility_std': 0.15466144680976868, 'reward_repeat_penalty_mean': 0.7620465159416199, 'reward_repeat_penalty_std': 0.1048668697476387, 'reward_repeat_floor_mean': 0.9820486307144165, 'reward_repeat_floor_std': 0.011632430367171764, 'reward_near_duplicate_penalty_mean': 0.9737856388092041, 'reward_near_duplicate_penalty_std': 0.057021964341402054, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9692307710647583, 'reward_distinct_2_std': 0.08702851831912994, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.12267211079597473, 'reward_total_composite_std': 0.07322044670581818, 'epoch': 0.04}

 22%|██▏       | 735/3300 [2:43:43<10:17:54, 14.45s/it]INFO 04-12 18:42:10 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 736/3300 [2:43:57<10:04:55, 14.16s/it]2026-04-12 18:42:23,317 | INFO | train_grpo_train | metrics_logged mode=train step=736

                                                       
{'loss': -0.0214, 'grad_norm': 2.325249195098877, 'learning_rate': 7.772727272727273e-06, 'num_tokens': 1340452.0, 'completions/mean_length': 144.375, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 21.83333396911621, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.2917229235172272, 'rewards/meter/std': 0.3761116564273834, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.858216404914856, 'rewards/lexical_plausibility/std': 0.24734167754650116, 'rewards/judge_quality/mean': 0.27125000953674316, 'rewards/judge_quality/std': 0.28667742013931274, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12339712679386139, 'rewards/total_composite/std': 0.259798139333725, 'reward': 0.12339712679386139, 'reward_std': 0.2597981095314026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1476915031671524, 'sampling/sampling_logp_difference/max': 1.0003724098205566, 'sampling/importance_sampling_ratio/min': 0.36774247884750366, 'sampling/importance_sampling_ratio/mean': 1.018509864807129, 'sampling/importance_sampling_ratio/max': 1.9719990491867065, 'entropy': 0.9392887651920319, 'clip_ratio/low_mean': 0.08189863944426179, 'clip_ratio/low_min': 0.08189863944426179, 'clip_ratio/high_mean': 0.0416666679084301, 'clip_ratio/high_max': 0.0416666679084301, 'clip_ratio/region_mean': 0.12356530735269189, 'reward_total_mean': 0.12339712679386139, 'reward_meter_mean': 0.2917229235172272, 'reward_meter_std': 0.3761116564273834, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.858216404914856, 'reward_lexical_plausibility_std': 0.24734167754650116, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.27125000953674316, 'reward_judge_quality_std': 0.28667742013931274, 'reward_total_composite_mean': 0.12339712679386139, 'reward_total_composite_std': 0.259798139333725, 'epoch': 0.04}

 22%|██▏       | 736/3300 [2:43:57<10:04:55, 14.16s/it]INFO 04-12 18:42:23 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 737/3300 [2:44:10<9:58:52, 14.02s/it] 2026-04-12 18:42:37,021 | INFO | train_grpo_train | metrics_logged mode=train step=737

                                                      
{'loss': -0.008, 'grad_norm': 4.003470420837402, 'learning_rate': 7.76969696969697e-06, 'num_tokens': 1341807.0, 'completions/mean_length': 80.375, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 18.71428680419922, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.403460294008255, 'rewards/meter/std': 0.3681066930294037, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9372429847717285, 'rewards/lexical_plausibility/std': 0.1775035411119461, 'rewards/judge_quality/mean': 0.2837500274181366, 'rewards/judge_quality/std': 0.28278905153274536, 'rewards/repeat_penalty/mean': 0.7007063627243042, 'rewards/repeat_penalty/std': 0.09127402305603027, 'rewards/repeat_floor/mean': 0.9751412868499756, 'rewards/repeat_floor/std': 0.01825481653213501, 'rewards/near_duplicate_penalty/mean': 0.9342751502990723, 'rewards/near_duplicate_penalty/std': 0.12169872224330902, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.07287507504224777, 'rewards/total_composite/std': 0.07762982696294785, 'reward': 0.07287507504224777, 'reward_std': 0.07762983441352844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1422041952610016, 'sampling/sampling_logp_difference/max': 0.9437007904052734, 'sampling/importance_sampling_ratio/min': 0.3891848921775818, 'sampling/importance_sampling_ratio/mean': 1.0433428287506104, 'sampling/importance_sampling_ratio/max': 1.873073697090149, 'entropy': 1.1322224661707878, 'clip_ratio/low_mean': 0.05137844476848841, 'clip_ratio/low_min': 0.05137844476848841, 'clip_ratio/high_mean': 0.058654218912124634, 'clip_ratio/high_max': 0.058654218912124634, 'clip_ratio/region_mean': 0.11003266368061304, 'reward_total_mean': 0.07287507504224777, 'reward_meter_mean': 0.403460294008255, 'reward_meter_std': 0.3681066930294037, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9372429847717285, 'reward_lexical_plausibility_std': 0.1775035411119461, 'reward_repeat_penalty_mean': 0.7007063627243042, 'reward_repeat_penalty_std': 0.09127402305603027, 'reward_repeat_floor_mean': 0.9751412868499756, 'reward_repeat_floor_std': 0.01825481653213501, 'reward_near_duplicate_penalty_mean': 0.9342751502990723, 'reward_near_duplicate_penalty_std': 0.12169872224330902, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2837500274181366, 'reward_judge_quality_std': 0.28278905153274536, 'reward_total_composite_mean': 0.07287507504224777, 'reward_total_composite_std': 0.07762982696294785, 'epoch': 0.04}

 22%|██▏       | 737/3300 [2:44:10<9:58:52, 14.02s/it]INFO 04-12 18:42:37 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 738/3300 [2:44:24<9:50:50, 13.84s/it]2026-04-12 18:42:50,486 | INFO | train_grpo_train | metrics_logged mode=train step=738

                                                      
{'loss': -0.0494, 'grad_norm': 1.9637209177017212, 'learning_rate': 7.766666666666666e-06, 'num_tokens': 1343120.0, 'completions/mean_length': 205.125, 'completions/min_length': 14.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6143946647644043, 'rewards/meter/std': 0.46436938643455505, 'rewards/count_adherence/mean': 0.3333333432674408, 'rewards/count_adherence/std': 0.1781741827726364, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.800000011920929, 'rewards/count_floor/std': 0.05345224589109421, 'rewards/lexical_plausibility/mean': 0.7579483389854431, 'rewards/lexical_plausibility/std': 0.2553488612174988, 'rewards/judge_quality/mean': 0.18125000596046448, 'rewards/judge_quality/std': 0.14865468442440033, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.1157275140285492, 'rewards/repeat_floor/mean': 0.9887499809265137, 'rewards/repeat_floor/std': 0.0069436440244317055, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12995028495788574, 'rewards/total_composite/std': 0.12614144384860992, 'reward': 0.12995028495788574, 'reward_std': 0.12614144384860992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17581558227539062, 'sampling/sampling_logp_difference/max': 2.088022232055664, 'sampling/importance_sampling_ratio/min': 0.12393199652433395, 'sampling/importance_sampling_ratio/mean': 1.0554476976394653, 'sampling/importance_sampling_ratio/max': 1.8839465379714966, 'entropy': 1.0012018233537674, 'clip_ratio/low_mean': 0.019736841320991516, 'clip_ratio/low_min': 0.019736841320991516, 'clip_ratio/high_mean': 0.06529190763831139, 'clip_ratio/high_max': 0.06529190763831139, 'clip_ratio/region_mean': 0.0850287489593029, 'reward_total_mean': 0.12995028495788574, 'reward_meter_mean': 0.6143946647644043, 'reward_meter_std': 0.46436938643455505, 'reward_count_adherence_mean': 0.3333333432674408, 'reward_count_adherence_std': 0.1781741827726364, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.800000011920929, 'reward_count_floor_std': 0.05345224589109421, 'reward_lexical_plausibility_mean': 0.7579483389854431, 'reward_lexical_plausibility_std': 0.2553488612174988, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.1157275140285492, 'reward_repeat_floor_mean': 0.9887499809265137, 'reward_repeat_floor_std': 0.0069436440244317055, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.18125000596046448, 'reward_judge_quality_std': 0.14865468442440033, 'reward_total_composite_mean': 0.12995028495788574, 'reward_total_composite_std': 0.12614144384860992, 'epoch': 0.04}

 22%|██▏       | 738/3300 [2:44:24<9:50:50, 13.84s/it]INFO 04-12 18:42:50 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 739/3300 [2:44:38<9:53:36, 13.91s/it]2026-04-12 18:43:04,501 | INFO | train_grpo_train | metrics_logged mode=train step=739

                                                      
{'loss': -0.0193, 'grad_norm': 1.348751187324524, 'learning_rate': 7.763636363636364e-06, 'num_tokens': 1344432.0, 'completions/mean_length': 204.0, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 19.200000762939453, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.6815898418426514, 'rewards/meter/std': 0.4493662118911743, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8145034313201904, 'rewards/lexical_plausibility/std': 0.24308021366596222, 'rewards/judge_quality/mean': 0.23375001549720764, 'rewards/judge_quality/std': 0.31208688020706177, 'rewards/repeat_penalty/mean': 0.7448081970214844, 'rewards/repeat_penalty/std': 0.013514467515051365, 'rewards/repeat_floor/mean': 0.9839916229248047, 'rewards/repeat_floor/std': 0.002709784312173724, 'rewards/near_duplicate_penalty/mean': 0.9935770630836487, 'rewards/near_duplicate_penalty/std': 0.018166767433285713, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9995005130767822, 'rewards/distinct_2/std': 0.0014128044713288546, 'rewards/total_composite/mean': 0.18715158104896545, 'rewards/total_composite/std': 0.2601436674594879, 'reward': 0.18715158104896545, 'reward_std': 0.2601436972618103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1537979245185852, 'sampling/sampling_logp_difference/max': 0.7707515954971313, 'sampling/importance_sampling_ratio/min': 0.4626652002334595, 'sampling/importance_sampling_ratio/mean': 1.0036507844924927, 'sampling/importance_sampling_ratio/max': 1.7780812978744507, 'entropy': 0.7506260052323341, 'clip_ratio/low_mean': 0.04255952499806881, 'clip_ratio/low_min': 0.04255952499806881, 'clip_ratio/high_mean': 0.050245098769664764, 'clip_ratio/high_max': 0.050245098769664764, 'clip_ratio/region_mean': 0.09280462376773357, 'reward_total_mean': 0.18715158104896545, 'reward_meter_mean': 0.6815898418426514, 'reward_meter_std': 0.4493662118911743, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8145034313201904, 'reward_lexical_plausibility_std': 0.24308021366596222, 'reward_repeat_penalty_mean': 0.7448081970214844, 'reward_repeat_penalty_std': 0.013514467515051365, 'reward_repeat_floor_mean': 0.9839916229248047, 'reward_repeat_floor_std': 0.002709784312173724, 'reward_near_duplicate_penalty_mean': 0.9935770630836487, 'reward_near_duplicate_penalty_std': 0.018166767433285713, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9995005130767822, 'reward_distinct_2_std': 0.0014128044713288546, 'reward_judge_quality_mean': 0.23375001549720764, 'reward_judge_quality_std': 0.31208688020706177, 'reward_total_composite_mean': 0.18715158104896545, 'reward_total_composite_std': 0.2601436674594879, 'epoch': 0.04}

 22%|██▏       | 739/3300 [2:44:38<9:53:36, 13.91s/it]INFO 04-12 18:43:04 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 740/3300 [2:44:46<8:43:21, 12.27s/it]2026-04-12 18:43:12,938 | INFO | train_grpo_train | metrics_logged mode=train step=740

                                                      
{'loss': -0.0108, 'grad_norm': 19.96430206298828, 'learning_rate': 7.76060606060606e-06, 'num_tokens': 1345818.0, 'completions/mean_length': 20.25, 'completions/min_length': 15.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.20640191435813904, 'rewards/meter/std': 0.3374778926372528, 'rewards/count_adherence/mean': 0.3333333432674408, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.800000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.7469173073768616, 'rewards/repeat_penalty/std': 0.00871915090829134, 'rewards/repeat_floor/mean': 0.9843834638595581, 'rewards/repeat_floor/std': 0.0017438469221815467, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.028493516147136688, 'rewards/total_composite/std': 0.038897447288036346, 'reward': 0.028493516147136688, 'reward_std': 0.038897447288036346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1758432537317276, 'sampling/sampling_logp_difference/max': 1.4410896301269531, 'sampling/importance_sampling_ratio/min': 0.3202000856399536, 'sampling/importance_sampling_ratio/mean': 1.0307332277297974, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9976207837462425, 'clip_ratio/low_mean': 0.10916285635903478, 'clip_ratio/low_min': 0.10916285635903478, 'clip_ratio/high_mean': 0.0457251095212996, 'clip_ratio/high_max': 0.0457251095212996, 'clip_ratio/region_mean': 0.15488796588033438, 'reward_total_mean': 0.028493516147136688, 'reward_meter_mean': 0.20640191435813904, 'reward_meter_std': 0.3374778926372528, 'reward_count_adherence_mean': 0.3333333432674408, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.800000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7469173073768616, 'reward_repeat_penalty_std': 0.00871915090829134, 'reward_repeat_floor_mean': 0.9843834638595581, 'reward_repeat_floor_std': 0.0017438469221815467, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.028493516147136688, 'reward_total_composite_std': 0.038897447288036346, 'epoch': 0.04}

 22%|██▏       | 740/3300 [2:44:46<8:43:21, 12.27s/it]INFO 04-12 18:43:13 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 741/3300 [2:44:57<8:18:42, 11.69s/it]2026-04-12 18:43:23,294 | INFO | train_grpo_train | metrics_logged mode=train step=741

                                                      
{'loss': 0.0525, 'grad_norm': 23.208003997802734, 'learning_rate': 7.757575757575758e-06, 'num_tokens': 1347405.0, 'completions/mean_length': 30.375, 'completions/min_length': 27.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.4057590067386627, 'rewards/meter/std': 0.42456623911857605, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6599999666213989, 'rewards/judge_quality/std': 0.3724053204059601, 'rewards/repeat_penalty/mean': 0.9916204214096069, 'rewards/repeat_penalty/std': 0.01294863410294056, 'rewards/repeat_floor/mean': 0.9987430572509766, 'rewards/repeat_floor/std': 0.0019422972109168768, 'rewards/near_duplicate_penalty/mean': 0.9916204214096069, 'rewards/near_duplicate_penalty/std': 0.01294863410294056, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1763792186975479, 'rewards/total_composite/std': 0.23815672099590302, 'reward': 0.1763792186975479, 'reward_std': 0.23815672099590302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14278177917003632, 'sampling/sampling_logp_difference/max': 1.8892238140106201, 'sampling/importance_sampling_ratio/min': 0.15118910372257233, 'sampling/importance_sampling_ratio/mean': 1.0021116733551025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8032214045524597, 'clip_ratio/low_mean': 0.07851171866059303, 'clip_ratio/low_min': 0.07851171866059303, 'clip_ratio/high_mean': 0.030237268656492233, 'clip_ratio/high_max': 0.030237268656492233, 'clip_ratio/region_mean': 0.10874898731708527, 'reward_total_mean': 0.1763792186975479, 'reward_meter_mean': 0.4057590067386627, 'reward_meter_std': 0.42456623911857605, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9916204214096069, 'reward_repeat_penalty_std': 0.01294863410294056, 'reward_repeat_floor_mean': 0.9987430572509766, 'reward_repeat_floor_std': 0.0019422972109168768, 'reward_near_duplicate_penalty_mean': 0.9916204214096069, 'reward_near_duplicate_penalty_std': 0.01294863410294056, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6599999666213989, 'reward_judge_quality_std': 0.3724053204059601, 'reward_total_composite_mean': 0.1763792186975479, 'reward_total_composite_std': 0.23815672099590302, 'epoch': 0.04}

 22%|██▏       | 741/3300 [2:44:57<8:18:42, 11.69s/it]INFO 04-12 18:43:23 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 742/3300 [2:45:10<8:40:05, 12.20s/it]2026-04-12 18:43:36,675 | INFO | train_grpo_train | metrics_logged mode=train step=742

                                                      
{'loss': -0.0209, 'grad_norm': 2.7749855518341064, 'learning_rate': 7.754545454545455e-06, 'num_tokens': 1348740.0, 'completions/mean_length': 142.875, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 19.83333396911621, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.5375322103500366, 'rewards/meter/std': 0.45469042658805847, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.7862309217453003, 'rewards/lexical_plausibility/std': 0.2334330976009369, 'rewards/judge_quality/mean': 0.35500001907348633, 'rewards/judge_quality/std': 0.3702508807182312, 'rewards/repeat_penalty/mean': 0.777212917804718, 'rewards/repeat_penalty/std': 0.09072622656822205, 'rewards/repeat_floor/mean': 0.9860675930976868, 'rewards/repeat_floor/std': 0.006066360045224428, 'rewards/near_duplicate_penalty/mean': 0.994617223739624, 'rewards/near_duplicate_penalty/std': 0.015224790200591087, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2790070176124573, 'rewards/total_composite/std': 0.37286657094955444, 'reward': 0.2790070176124573, 'reward_std': 0.37286657094955444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21848545968532562, 'sampling/sampling_logp_difference/max': 1.837364673614502, 'sampling/importance_sampling_ratio/min': 0.15923650562763214, 'sampling/importance_sampling_ratio/mean': 1.0234222412109375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.452508419752121, 'clip_ratio/low_mean': 0.10812952741980553, 'clip_ratio/low_min': 0.10812952741980553, 'clip_ratio/high_mean': 0.034307065419852734, 'clip_ratio/high_max': 0.034307065419852734, 'clip_ratio/region_mean': 0.14243659283965826, 'reward_total_mean': 0.2790070176124573, 'reward_meter_mean': 0.5375322103500366, 'reward_meter_std': 0.45469042658805847, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.7862309217453003, 'reward_lexical_plausibility_std': 0.2334330976009369, 'reward_repeat_penalty_mean': 0.777212917804718, 'reward_repeat_penalty_std': 0.09072622656822205, 'reward_repeat_floor_mean': 0.9860675930976868, 'reward_repeat_floor_std': 0.006066360045224428, 'reward_near_duplicate_penalty_mean': 0.994617223739624, 'reward_near_duplicate_penalty_std': 0.015224790200591087, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.35500001907348633, 'reward_judge_quality_std': 0.3702508807182312, 'reward_total_composite_mean': 0.2790070176124573, 'reward_total_composite_std': 0.37286657094955444, 'epoch': 0.04}

 22%|██▏       | 742/3300 [2:45:10<8:40:05, 12.20s/it]INFO 04-12 18:43:36 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 743/3300 [2:45:26<9:30:29, 13.39s/it]2026-04-12 18:43:52,876 | INFO | train_grpo_train | metrics_logged mode=train step=743

                                                      
{'loss': 0.0012, 'grad_norm': 2.8965985774993896, 'learning_rate': 7.751515151515153e-06, 'num_tokens': 1350035.0, 'completions/mean_length': 148.875, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.83333396911621, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5172783732414246, 'rewards/meter/std': 0.41920652985572815, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.8744943141937256, 'rewards/lexical_plausibility/std': 0.23269939422607422, 'rewards/judge_quality/mean': 0.3799999952316284, 'rewards/judge_quality/std': 0.35306817293167114, 'rewards/repeat_penalty/mean': 0.8201873302459717, 'rewards/repeat_penalty/std': 0.16938941180706024, 'rewards/repeat_floor/mean': 0.984004020690918, 'rewards/repeat_floor/std': 0.015525962226092815, 'rewards/near_duplicate_penalty/mean': 0.9617943167686462, 'rewards/near_duplicate_penalty/std': 0.04469265043735504, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9692763090133667, 'rewards/distinct_2/std': 0.08328244090080261, 'rewards/total_composite/mean': 0.16020092368125916, 'rewards/total_composite/std': 0.24811586737632751, 'reward': 0.16020092368125916, 'reward_std': 0.24811585247516632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16965435445308685, 'sampling/sampling_logp_difference/max': 0.8831815719604492, 'sampling/importance_sampling_ratio/min': 0.41346535086631775, 'sampling/importance_sampling_ratio/mean': 1.0435234308242798, 'sampling/importance_sampling_ratio/max': 1.9052214622497559, 'entropy': 1.118041105568409, 'clip_ratio/low_mean': 0.060724363662302494, 'clip_ratio/low_min': 0.060724363662302494, 'clip_ratio/high_mean': 0.03281250037252903, 'clip_ratio/high_max': 0.03281250037252903, 'clip_ratio/region_mean': 0.09353686403483152, 'reward_total_mean': 0.16020092368125916, 'reward_meter_mean': 0.5172783732414246, 'reward_meter_std': 0.41920652985572815, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.8744943141937256, 'reward_lexical_plausibility_std': 0.23269939422607422, 'reward_repeat_penalty_mean': 0.8201873302459717, 'reward_repeat_penalty_std': 0.16938941180706024, 'reward_repeat_floor_mean': 0.984004020690918, 'reward_repeat_floor_std': 0.015525962226092815, 'reward_near_duplicate_penalty_mean': 0.9617943167686462, 'reward_near_duplicate_penalty_std': 0.04469265043735504, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9692763090133667, 'reward_distinct_2_std': 0.08328244090080261, 'reward_judge_quality_mean': 0.3799999952316284, 'reward_judge_quality_std': 0.35306817293167114, 'reward_total_composite_mean': 0.16020092368125916, 'reward_total_composite_std': 0.24811586737632751, 'epoch': 0.04}

 23%|██▎       | 743/3300 [2:45:26<9:30:29, 13.39s/it]INFO 04-12 18:43:53 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 744/3300 [2:45:43<10:13:51, 14.41s/it]2026-04-12 18:44:09,652 | INFO | train_grpo_train | metrics_logged mode=train step=744

                                                       
{'loss': -0.0973, 'grad_norm': 2.975322961807251, 'learning_rate': 7.74848484848485e-06, 'num_tokens': 1352126.0, 'completions/mean_length': 192.375, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 85.83333587646484, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.8800060153007507, 'rewards/meter/std': 0.259451687335968, 'rewards/count_adherence/mean': 0.4305555820465088, 'rewards/count_adherence/std': 0.32223591208457947, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8291666507720947, 'rewards/count_floor/std': 0.09667076170444489, 'rewards/lexical_plausibility/mean': 0.8968592286109924, 'rewards/lexical_plausibility/std': 0.19558745622634888, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.7106995582580566, 'rewards/repeat_penalty/std': 0.2822147309780121, 'rewards/repeat_floor/mean': 0.9704732894897461, 'rewards/repeat_floor/std': 0.032783858478069305, 'rewards/near_duplicate_penalty/mean': 0.9459105730056763, 'rewards/near_duplicate_penalty/std': 0.062300343066453934, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8037415146827698, 'rewards/distinct_2/std': 0.292168527841568, 'rewards/total_composite/mean': 0.1111559271812439, 'rewards/total_composite/std': 0.07378222048282623, 'reward': 0.1111559271812439, 'reward_std': 0.07378222048282623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15950749814510345, 'sampling/sampling_logp_difference/max': 1.5414862632751465, 'sampling/importance_sampling_ratio/min': 0.21406272053718567, 'sampling/importance_sampling_ratio/mean': 1.0336241722106934, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.192721113562584, 'clip_ratio/low_mean': 0.08326923102140427, 'clip_ratio/low_min': 0.08326923102140427, 'clip_ratio/high_mean': 0.08411479368805885, 'clip_ratio/high_max': 0.08411479368805885, 'clip_ratio/region_mean': 0.16738402470946312, 'reward_total_mean': 0.1111559271812439, 'reward_meter_mean': 0.8800060153007507, 'reward_meter_std': 0.259451687335968, 'reward_count_adherence_mean': 0.4305555820465088, 'reward_count_adherence_std': 0.32223591208457947, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8291666507720947, 'reward_count_floor_std': 0.09667076170444489, 'reward_lexical_plausibility_mean': 0.8968592286109924, 'reward_lexical_plausibility_std': 0.19558745622634888, 'reward_repeat_penalty_mean': 0.7106995582580566, 'reward_repeat_penalty_std': 0.2822147309780121, 'reward_repeat_floor_mean': 0.9704732894897461, 'reward_repeat_floor_std': 0.032783858478069305, 'reward_near_duplicate_penalty_mean': 0.9459105730056763, 'reward_near_duplicate_penalty_std': 0.062300343066453934, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8037415146827698, 'reward_distinct_2_std': 0.292168527841568, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.1111559271812439, 'reward_total_composite_std': 0.07378222048282623, 'epoch': 0.04}

 23%|██▎       | 744/3300 [2:45:43<10:13:51, 14.41s/it]INFO 04-12 18:44:09 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 745/3300 [2:45:50<8:44:16, 12.31s/it] 2026-04-12 18:44:17,043 | INFO | train_grpo_train | metrics_logged mode=train step=745

                                                      
{'loss': 0.013, 'grad_norm': 14.36792278289795, 'learning_rate': 7.745454545454545e-06, 'num_tokens': 1353491.0, 'completions/mean_length': 19.625, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.7133709788322449, 'rewards/meter/std': 0.4207753539085388, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.27125000953674316, 'rewards/judge_quality/std': 0.2713162302970886, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21824952960014343, 'rewards/total_composite/std': 0.2779955565929413, 'reward': 0.21824952960014343, 'reward_std': 0.2779955565929413, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09859257936477661, 'sampling/sampling_logp_difference/max': 1.220430612564087, 'sampling/importance_sampling_ratio/min': 0.2951030433177948, 'sampling/importance_sampling_ratio/mean': 0.9983386397361755, 'sampling/importance_sampling_ratio/max': 1.7585811614990234, 'entropy': 0.6698262840509415, 'clip_ratio/low_mean': 0.06917613837867975, 'clip_ratio/low_min': 0.06917613837867975, 'clip_ratio/high_mean': 0.013194444589316845, 'clip_ratio/high_max': 0.013194444589316845, 'clip_ratio/region_mean': 0.0823705829679966, 'reward_total_mean': 0.21824952960014343, 'reward_meter_mean': 0.7133709788322449, 'reward_meter_std': 0.4207753539085388, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.27125000953674316, 'reward_judge_quality_std': 0.2713162302970886, 'reward_total_composite_mean': 0.21824952960014343, 'reward_total_composite_std': 0.2779955565929413, 'epoch': 0.04}

 23%|██▎       | 745/3300 [2:45:50<8:44:16, 12.31s/it]INFO 04-12 18:44:17 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 746/3300 [2:46:03<8:52:30, 12.51s/it]2026-04-12 18:44:30,020 | INFO | train_grpo_train | metrics_logged mode=train step=746

                                                      
{'loss': 0.0138, 'grad_norm': 4.74420166015625, 'learning_rate': 7.742424242424244e-06, 'num_tokens': 1354978.0, 'completions/mean_length': 83.875, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.71428680419922, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5336765050888062, 'rewards/meter/std': 0.4166505038738251, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.08017837256193161, 'rewards/lexical_plausibility/mean': 0.9357050657272339, 'rewards/lexical_plausibility/std': 0.15811949968338013, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3366537392139435, 'rewards/repeat_penalty/mean': 0.7642590403556824, 'rewards/repeat_penalty/std': 0.13603028655052185, 'rewards/repeat_floor/mean': 0.9807535409927368, 'rewards/repeat_floor/std': 0.0151299349963665, 'rewards/near_duplicate_penalty/mean': 0.9571796655654907, 'rewards/near_duplicate_penalty/std': 0.08880757540464401, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.1489267647266388, 'rewards/total_composite/std': 0.12230773270130157, 'reward': 0.1489267647266388, 'reward_std': 0.12230773270130157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16524523496627808, 'sampling/sampling_logp_difference/max': 2.02655029296875, 'sampling/importance_sampling_ratio/min': 0.13178937137126923, 'sampling/importance_sampling_ratio/mean': 0.9896193742752075, 'sampling/importance_sampling_ratio/max': 1.637553334236145, 'entropy': 1.0790852010250092, 'clip_ratio/low_mean': 0.061038012616336346, 'clip_ratio/low_min': 0.061038012616336346, 'clip_ratio/high_mean': 0.05491946963593364, 'clip_ratio/high_max': 0.05491946963593364, 'clip_ratio/region_mean': 0.11595748225226998, 'reward_total_mean': 0.1489267647266388, 'reward_meter_mean': 0.5336765050888062, 'reward_meter_std': 0.4166505038738251, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.08017837256193161, 'reward_lexical_plausibility_mean': 0.9357050657272339, 'reward_lexical_plausibility_std': 0.15811949968338013, 'reward_repeat_penalty_mean': 0.7642590403556824, 'reward_repeat_penalty_std': 0.13603028655052185, 'reward_repeat_floor_mean': 0.9807535409927368, 'reward_repeat_floor_std': 0.0151299349963665, 'reward_near_duplicate_penalty_mean': 0.9571796655654907, 'reward_near_duplicate_penalty_std': 0.08880757540464401, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3366537392139435, 'reward_total_composite_mean': 0.1489267647266388, 'reward_total_composite_std': 0.12230773270130157, 'epoch': 0.04}

 23%|██▎       | 746/3300 [2:46:03<8:52:30, 12.51s/it]INFO 04-12 18:44:30 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 747/3300 [2:46:17<9:11:52, 12.97s/it]2026-04-12 18:44:44,074 | INFO | train_grpo_train | metrics_logged mode=train step=747

                                                      
{'loss': -0.0196, 'grad_norm': 1.8384373188018799, 'learning_rate': 7.73939393939394e-06, 'num_tokens': 1356303.0, 'completions/mean_length': 203.625, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 18.600000381469727, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.29425597190856934, 'rewards/meter/std': 0.35729554295539856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8731686472892761, 'rewards/lexical_plausibility/std': 0.18030144274234772, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.15811388194561005, 'rewards/repeat_penalty/mean': 0.7074088454246521, 'rewards/repeat_penalty/std': 0.10268846154212952, 'rewards/repeat_floor/mean': 0.9764817953109741, 'rewards/repeat_floor/std': 0.020537693053483963, 'rewards/near_duplicate_penalty/mean': 0.9432117938995361, 'rewards/near_duplicate_penalty/std': 0.13691793382167816, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.068361297249794, 'rewards/total_composite/std': 0.08780686557292938, 'reward': 0.068361297249794, 'reward_std': 0.08780687302350998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24706055223941803, 'sampling/sampling_logp_difference/max': 1.7024469375610352, 'sampling/importance_sampling_ratio/min': 0.18223705887794495, 'sampling/importance_sampling_ratio/mean': 1.0187609195709229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7845836654305458, 'clip_ratio/low_mean': 0.0416666679084301, 'clip_ratio/low_min': 0.0416666679084301, 'clip_ratio/high_mean': 0.051934524439275265, 'clip_ratio/high_max': 0.051934524439275265, 'clip_ratio/region_mean': 0.09360119234770536, 'reward_total_mean': 0.068361297249794, 'reward_meter_mean': 0.29425597190856934, 'reward_meter_std': 0.35729554295539856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8731686472892761, 'reward_lexical_plausibility_std': 0.18030144274234772, 'reward_repeat_penalty_mean': 0.7074088454246521, 'reward_repeat_penalty_std': 0.10268846154212952, 'reward_repeat_floor_mean': 0.9764817953109741, 'reward_repeat_floor_std': 0.020537693053483963, 'reward_near_duplicate_penalty_mean': 0.9432117938995361, 'reward_near_duplicate_penalty_std': 0.13691793382167816, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.15811388194561005, 'reward_total_composite_mean': 0.068361297249794, 'reward_total_composite_std': 0.08780686557292938, 'epoch': 0.04}

 23%|██▎       | 747/3300 [2:46:17<9:11:52, 12.97s/it]INFO 04-12 18:44:44 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 748/3300 [2:46:30<9:08:37, 12.90s/it]2026-04-12 18:44:56,806 | INFO | train_grpo_train | metrics_logged mode=train step=748

                                                      
{'loss': -0.0355, 'grad_norm': 4.142660617828369, 'learning_rate': 7.736363636363637e-06, 'num_tokens': 1357745.0, 'completions/mean_length': 84.25, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.142858505249023, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8712382316589355, 'rewards/meter/std': 0.3331470787525177, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9306751489639282, 'rewards/lexical_plausibility/std': 0.19608032703399658, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.377548485994339, 'rewards/repeat_penalty/mean': 0.7308650612831116, 'rewards/repeat_penalty/std': 0.021784374490380287, 'rewards/repeat_floor/mean': 0.981173038482666, 'rewards/repeat_floor/std': 0.004356873221695423, 'rewards/near_duplicate_penalty/mean': 0.974486768245697, 'rewards/near_duplicate_penalty/std': 0.0290458295494318, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4792823791503906, 'rewards/total_composite/std': 0.32279109954833984, 'reward': 0.4792823791503906, 'reward_std': 0.32279109954833984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14162302017211914, 'sampling/sampling_logp_difference/max': 0.8711061477661133, 'sampling/importance_sampling_ratio/min': 0.41848838329315186, 'sampling/importance_sampling_ratio/mean': 1.0485907793045044, 'sampling/importance_sampling_ratio/max': 1.6908886432647705, 'entropy': 1.1248897537589073, 'clip_ratio/low_mean': 0.062224079854786396, 'clip_ratio/low_min': 0.062224079854786396, 'clip_ratio/high_mean': 0.06815711595118046, 'clip_ratio/high_max': 0.06815711595118046, 'clip_ratio/region_mean': 0.13038119580596685, 'reward_total_mean': 0.4792823791503906, 'reward_meter_mean': 0.8712382316589355, 'reward_meter_std': 0.3331470787525177, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9306751489639282, 'reward_lexical_plausibility_std': 0.19608032703399658, 'reward_repeat_penalty_mean': 0.7308650612831116, 'reward_repeat_penalty_std': 0.021784374490380287, 'reward_repeat_floor_mean': 0.981173038482666, 'reward_repeat_floor_std': 0.004356873221695423, 'reward_near_duplicate_penalty_mean': 0.974486768245697, 'reward_near_duplicate_penalty_std': 0.0290458295494318, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.377548485994339, 'reward_total_composite_mean': 0.4792823791503906, 'reward_total_composite_std': 0.32279109954833984, 'epoch': 0.04}

 23%|██▎       | 748/3300 [2:46:30<9:08:37, 12.90s/it]INFO 04-12 18:44:57 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 749/3300 [2:46:44<9:18:21, 13.13s/it]2026-04-12 18:45:10,469 | INFO | train_grpo_train | metrics_logged mode=train step=749

                                                      
{'loss': -0.0524, 'grad_norm': 5.769345760345459, 'learning_rate': 7.733333333333334e-06, 'num_tokens': 1359366.0, 'completions/mean_length': 103.625, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.28571701049805, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5235595703125, 'rewards/meter/std': 0.4026947617530823, 'rewards/count_adherence/mean': 0.5250000357627869, 'rewards/count_adherence/std': 0.26049405336380005, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8574999570846558, 'rewards/count_floor/std': 0.07814821600914001, 'rewards/lexical_plausibility/mean': 0.9598478078842163, 'rewards/lexical_plausibility/std': 0.11356756091117859, 'rewards/judge_quality/mean': 0.3187499940395355, 'rewards/judge_quality/std': 0.16243129968643188, 'rewards/repeat_penalty/mean': 0.8233376741409302, 'rewards/repeat_penalty/std': 0.07907243818044662, 'rewards/repeat_floor/mean': 0.982809841632843, 'rewards/repeat_floor/std': 0.006920860148966312, 'rewards/near_duplicate_penalty/mean': 0.9534773826599121, 'rewards/near_duplicate_penalty/std': 0.040107861161231995, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9282026290893555, 'rewards/distinct_2/std': 0.0812455490231514, 'rewards/total_composite/mean': 0.1589653491973877, 'rewards/total_composite/std': 0.14139433205127716, 'reward': 0.1589653491973877, 'reward_std': 0.14139433205127716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18243052065372467, 'sampling/sampling_logp_difference/max': 1.4429073333740234, 'sampling/importance_sampling_ratio/min': 0.2362399399280548, 'sampling/importance_sampling_ratio/mean': 1.0197663307189941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2705709263682365, 'clip_ratio/low_mean': 0.06131018325686455, 'clip_ratio/low_min': 0.06131018325686455, 'clip_ratio/high_mean': 0.11268244124948978, 'clip_ratio/high_max': 0.11268244124948978, 'clip_ratio/region_mean': 0.17399262450635433, 'reward_total_mean': 0.1589653491973877, 'reward_meter_mean': 0.5235595703125, 'reward_meter_std': 0.4026947617530823, 'reward_count_adherence_mean': 0.5250000357627869, 'reward_count_adherence_std': 0.26049405336380005, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8574999570846558, 'reward_count_floor_std': 0.07814821600914001, 'reward_lexical_plausibility_mean': 0.9598478078842163, 'reward_lexical_plausibility_std': 0.11356756091117859, 'reward_repeat_penalty_mean': 0.8233376741409302, 'reward_repeat_penalty_std': 0.07907243818044662, 'reward_repeat_floor_mean': 0.982809841632843, 'reward_repeat_floor_std': 0.006920860148966312, 'reward_near_duplicate_penalty_mean': 0.9534773826599121, 'reward_near_duplicate_penalty_std': 0.040107861161231995, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9282026290893555, 'reward_distinct_2_std': 0.0812455490231514, 'reward_judge_quality_mean': 0.3187499940395355, 'reward_judge_quality_std': 0.16243129968643188, 'reward_total_composite_mean': 0.1589653491973877, 'reward_total_composite_std': 0.14139433205127716, 'epoch': 0.04}

 23%|██▎       | 749/3300 [2:46:44<9:18:21, 13.13s/it]INFO 04-12 18:45:10 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 750/3300 [2:46:58<9:28:25, 13.37s/it]2026-04-12 18:45:24,513 | INFO | train_grpo_train | metrics_logged mode=train step=750

                                                      
{'loss': -0.0335, 'grad_norm': 4.7941670417785645, 'learning_rate': 7.730303030303032e-06, 'num_tokens': 1360788.0, 'completions/mean_length': 84.75, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.71428680419922, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.47516119480133057, 'rewards/meter/std': 0.42735129594802856, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9183794856071472, 'rewards/lexical_plausibility/std': 0.15269480645656586, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.15221577882766724, 'rewards/repeat_penalty/mean': 0.7543042898178101, 'rewards/repeat_penalty/std': 0.26004454493522644, 'rewards/repeat_floor/mean': 0.9735600352287292, 'rewards/repeat_floor/std': 0.04295467212796211, 'rewards/near_duplicate_penalty/mean': 0.8862333297729492, 'rewards/near_duplicate_penalty/std': 0.27081748843193054, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15436387062072754, 'rewards/total_composite/std': 0.13202892243862152, 'reward': 0.15436387062072754, 'reward_std': 0.13202892243862152, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20861199498176575, 'sampling/sampling_logp_difference/max': 1.076934814453125, 'sampling/importance_sampling_ratio/min': 0.34063804149627686, 'sampling/importance_sampling_ratio/mean': 1.0120781660079956, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5639142394065857, 'clip_ratio/low_mean': 0.06005434785038233, 'clip_ratio/low_min': 0.06005434785038233, 'clip_ratio/high_mean': 0.11765479855239391, 'clip_ratio/high_max': 0.11765479855239391, 'clip_ratio/region_mean': 0.17770914640277624, 'reward_total_mean': 0.15436387062072754, 'reward_meter_mean': 0.47516119480133057, 'reward_meter_std': 0.42735129594802856, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9183794856071472, 'reward_lexical_plausibility_std': 0.15269480645656586, 'reward_repeat_penalty_mean': 0.7543042898178101, 'reward_repeat_penalty_std': 0.26004454493522644, 'reward_repeat_floor_mean': 0.9735600352287292, 'reward_repeat_floor_std': 0.04295467212796211, 'reward_near_duplicate_penalty_mean': 0.8862333297729492, 'reward_near_duplicate_penalty_std': 0.27081748843193054, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.15221577882766724, 'reward_total_composite_mean': 0.15436387062072754, 'reward_total_composite_std': 0.13202892243862152, 'epoch': 0.04}

 23%|██▎       | 750/3300 [2:46:58<9:28:25, 13.37s/it]INFO 04-12 18:45:24 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:04,  5.84s/it]

 23%|██▎       | 3/13 [00:17<00:57,  5.78s/it]

 31%|███       | 4/13 [00:29<01:12,  8.03s/it]

 38%|███▊      | 5/13 [00:40<01:13,  9.21s/it]

 46%|████▌     | 6/13 [00:48<01:00,  8.67s/it]

 54%|█████▍    | 7/13 [00:58<00:55,  9.30s/it]

 62%|██████▏   | 8/13 [01:10<00:49,  9.95s/it]

 69%|██████▉   | 9/13 [01:21<00:41, 10.45s/it]

 77%|███████▋  | 10/13 [01:35<00:34, 11.34s/it]

 85%|████████▍ | 11/13 [01:46<00:22, 11.24s/it]

 92%|█████████▏| 12/13 [01:57<00:11, 11.26s/it]

100%|██████████| 13/13 [02:09<00:00, 11.37s/it]2026-04-12 18:47:45,358 | INFO | train_grpo_train | metrics_logged mode=eval step=750

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 140.7759, 'eval_samples_per_second': 0.739, 'eval_steps_per_second': 0.092, 'eval_num_tokens': 1360788.0, 'eval_completions/mean_length': 161.79807692307693, 'eval_completions/min_length': 16.923076923076923, 'eval_completions/max_length': 457.0, 'eval_completions/clipped_ratio': 0.19230769230769232, 'eval_completions/mean_terminated_length': 79.48965366070087, 'eval_completions/min_terminated_length': 16.923076923076923, 'eval_completions/max_terminated_length': 217.30769230769232, 'eval_rewards/meter/mean': 0.4821709440304683, 'eval_rewards/meter/std': 0.37727194222120136, 'eval_rewards/count_adherence/mean': 0.5976877762721136, 'eval_rewards/count_adherence/std': 0.3097472580579611, 'eval_rewards/arabic_clean/mean': 0.7884615384615384, 'eval_rewards/arabic_clean/std': 0.39074395253108096, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9937499990830054, 'eval_rewards/arabic_floor/std': 0.017677669341747578, 'eval_rewards/count_floor/mean': 0.8793063393005958, 'eval_rewards/count_floor/std': 0.09292416990949558, 'eval_rewards/lexical_plausibility/mean': 0.9099988387181208, 'eval_rewards/lexical_plausibility/std': 0.1679023699118541, 'eval_rewards/judge_quality/mean': 0.3104807688639714, 'eval_rewards/judge_quality/std': 0.24153148440214303, 'eval_rewards/repeat_penalty/mean': 0.7412021710322454, 'eval_rewards/repeat_penalty/std': 0.29337632484160936, 'eval_rewards/repeat_floor/mean': 0.9711762620852544, 'eval_rewards/repeat_floor/std': 0.03903098776936531, 'eval_rewards/near_duplicate_penalty/mean': 0.9364464191290048, 'eval_rewards/near_duplicate_penalty/std': 0.0943520564872485, 'eval_rewards/opening_diversity/mean': 0.8834695265843318, 'eval_rewards/opening_diversity/std': 0.18799234181642532, 'eval_rewards/distinct_2/mean': 0.8633458797748272, 'eval_rewards/distinct_2/std': 0.24834957002447203, 'eval_rewards/total_composite/mean': 0.12990731230148903, 'eval_rewards/total_composite/std': 0.1406358927488327, 'eval_reward': 0.12990731230148903, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08168218141564956, 'eval_sampling/sampling_logp_difference/max': 1.0946788787841797, 'eval_sampling/importance_sampling_ratio/min': 0.3392829161423903, 'eval_sampling/importance_sampling_ratio/mean': 1.0242463533694928, 'eval_sampling/importance_sampling_ratio/max': 1.6500817353908832, 'eval_entropy': 0.9567869809957651, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.12990731230148903, 'eval_reward_meter_mean': 0.4821709440304683, 'eval_reward_meter_std': 0.37727194222120136, 'eval_reward_count_adherence_mean': 0.5976877762721136, 'eval_reward_count_adherence_std': 0.3097472580579611, 'eval_reward_arabic_clean_mean': 0.7884615384615384, 'eval_reward_arabic_clean_std': 0.39074395253108096, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9937499990830054, 'eval_reward_arabic_floor_std': 0.017677669341747578, 'eval_reward_count_floor_mean': 0.8793063393005958, 'eval_reward_count_floor_std': 0.09292416990949558, 'eval_reward_lexical_plausibility_mean': 0.9099988387181208, 'eval_reward_lexical_plausibility_std': 0.1679023699118541, 'eval_reward_repeat_penalty_mean': 0.7412021710322454, 'eval_reward_repeat_penalty_std': 0.29337632484160936, 'eval_reward_repeat_floor_mean': 0.9711762620852544, 'eval_reward_repeat_floor_std': 0.03903098776936531, 'eval_reward_near_duplicate_penalty_mean': 0.9364464191290048, 'eval_reward_near_duplicate_penalty_std': 0.0943520564872485, 'eval_reward_opening_diversity_mean': 0.8834695265843318, 'eval_reward_opening_diversity_std': 0.18799234181642532, 'eval_reward_distinct_2_mean': 0.8633458797748272, 'eval_reward_distinct_2_std': 0.24834957002447203, 'eval_reward_judge_quality_mean': 0.3104807688639714, 'eval_reward_judge_quality_std': 0.24153148440214303, 'eval_reward_total_composite_mean': 0.12990731230148903, 'eval_reward_total_composite_std': 0.1406358927488327, 'epoch': 0.04}

 23%|██▎       | 750/3300 [2:49:19<9:28:25, 13.37s/it]

100%|██████████| 13/13 [02:09<00:00, 11.37s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 23%|██▎       | 751/3300 [2:49:35<40:01:34, 56.53s/it]2026-04-12 18:48:01,638 | INFO | train_grpo_train | metrics_logged mode=train step=751

                                                       
{'loss': -0.0475, 'grad_norm': 2.932443380355835, 'learning_rate': 7.727272727272727e-06, 'num_tokens': 1362185.0, 'completions/mean_length': 143.625, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 20.83333396911621, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.3506571054458618, 'rewards/meter/std': 0.302674263715744, 'rewards/count_adherence/mean': 0.375, 'rewards/count_adherence/std': 0.2136233150959015, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8125, 'rewards/count_floor/std': 0.06408698856830597, 'rewards/lexical_plausibility/mean': 0.8737751841545105, 'rewards/lexical_plausibility/std': 0.219232439994812, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.8220474123954773, 'rewards/repeat_penalty/std': 0.10830463469028473, 'rewards/repeat_floor/mean': 0.9879381656646729, 'rewards/repeat_floor/std': 0.006480756681412458, 'rewards/near_duplicate_penalty/mean': 0.9888750910758972, 'rewards/near_duplicate_penalty/std': 0.02331627905368805, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.04655265808105469, 'rewards/total_composite/std': 0.04339646175503731, 'reward': 0.04655265808105469, 'reward_std': 0.043396465480327606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19946685433387756, 'sampling/sampling_logp_difference/max': 1.0557308197021484, 'sampling/importance_sampling_ratio/min': 0.34793806076049805, 'sampling/importance_sampling_ratio/mean': 1.045884370803833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4028751850128174, 'clip_ratio/low_mean': 0.040570175275206566, 'clip_ratio/low_min': 0.040570175275206566, 'clip_ratio/high_mean': 0.09315476473420858, 'clip_ratio/high_max': 0.09315476473420858, 'clip_ratio/region_mean': 0.13372494000941515, 'reward_total_mean': 0.04655265808105469, 'reward_meter_mean': 0.3506571054458618, 'reward_meter_std': 0.302674263715744, 'reward_count_adherence_mean': 0.375, 'reward_count_adherence_std': 0.2136233150959015, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8125, 'reward_count_floor_std': 0.06408698856830597, 'reward_lexical_plausibility_mean': 0.8737751841545105, 'reward_lexical_plausibility_std': 0.219232439994812, 'reward_repeat_penalty_mean': 0.8220474123954773, 'reward_repeat_penalty_std': 0.10830463469028473, 'reward_repeat_floor_mean': 0.9879381656646729, 'reward_repeat_floor_std': 0.006480756681412458, 'reward_near_duplicate_penalty_mean': 0.9888750910758972, 'reward_near_duplicate_penalty_std': 0.02331627905368805, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.04655265808105469, 'reward_total_composite_std': 0.04339646175503731, 'epoch': 0.04}

 23%|██▎       | 751/3300 [2:49:35<40:01:34, 56.53s/it]INFO 04-12 18:48:01 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 752/3300 [2:49:48<30:52:21, 43.62s/it]2026-04-12 18:48:15,132 | INFO | train_grpo_train | metrics_logged mode=train step=752

                                                       
{'loss': -0.0062, 'grad_norm': 5.7732062339782715, 'learning_rate': 7.724242424242424e-06, 'num_tokens': 1363560.0, 'completions/mean_length': 81.875, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.428571701049805, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.4490688741207123, 'rewards/meter/std': 0.430948406457901, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9362407326698303, 'rewards/lexical_plausibility/std': 0.15661919116973877, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.28101539611816406, 'rewards/repeat_penalty/mean': 0.7414772510528564, 'rewards/repeat_penalty/std': 0.02410591021180153, 'rewards/repeat_floor/mean': 0.9832954406738281, 'rewards/repeat_floor/std': 0.004821178037673235, 'rewards/near_duplicate_penalty/mean': 0.9886363744735718, 'rewards/near_duplicate_penalty/std': 0.03214120864868164, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18368887901306152, 'rewards/total_composite/std': 0.2430846393108368, 'reward': 0.18368887901306152, 'reward_std': 0.2430846393108368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18957577645778656, 'sampling/sampling_logp_difference/max': 1.344736099243164, 'sampling/importance_sampling_ratio/min': 0.2606084644794464, 'sampling/importance_sampling_ratio/mean': 1.065350890159607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.374651476740837, 'clip_ratio/low_mean': 0.050879035610705614, 'clip_ratio/low_min': 0.050879035610705614, 'clip_ratio/high_mean': 0.039227413944900036, 'clip_ratio/high_max': 0.039227413944900036, 'clip_ratio/region_mean': 0.09010644955560565, 'reward_total_mean': 0.18368887901306152, 'reward_meter_mean': 0.4490688741207123, 'reward_meter_std': 0.430948406457901, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9362407326698303, 'reward_lexical_plausibility_std': 0.15661919116973877, 'reward_repeat_penalty_mean': 0.7414772510528564, 'reward_repeat_penalty_std': 0.02410591021180153, 'reward_repeat_floor_mean': 0.9832954406738281, 'reward_repeat_floor_std': 0.004821178037673235, 'reward_near_duplicate_penalty_mean': 0.9886363744735718, 'reward_near_duplicate_penalty_std': 0.03214120864868164, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.28101539611816406, 'reward_total_composite_mean': 0.18368887901306152, 'reward_total_composite_std': 0.2430846393108368, 'epoch': 0.04}

 23%|██▎       | 752/3300 [2:49:48<30:52:21, 43.62s/it]INFO 04-12 18:48:15 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 753/3300 [2:49:56<23:15:04, 32.86s/it]2026-04-12 18:48:22,900 | INFO | train_grpo_train | metrics_logged mode=train step=753

                                                       
{'loss': -0.0098, 'grad_norm': 19.75664710998535, 'learning_rate': 7.721212121212122e-06, 'num_tokens': 1364939.0, 'completions/mean_length': 20.375, 'completions/min_length': 18.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.5459390878677368, 'rewards/meter/std': 0.43266281485557556, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.875, 'rewards/lexical_plausibility/std': 0.3535533845424652, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.28630343079566956, 'rewards/repeat_penalty/mean': 0.7368881106376648, 'rewards/repeat_penalty/std': 0.03708602488040924, 'rewards/repeat_floor/mean': 0.983426570892334, 'rewards/repeat_floor/std': 0.004450327716767788, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.24094471335411072, 'rewards/total_composite/std': 0.2565024495124817, 'reward': 0.24094471335411072, 'reward_std': 0.2565024495124817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1629667580127716, 'sampling/sampling_logp_difference/max': 1.4346981048583984, 'sampling/importance_sampling_ratio/min': 0.2381872683763504, 'sampling/importance_sampling_ratio/mean': 1.0214160680770874, 'sampling/importance_sampling_ratio/max': 1.8477272987365723, 'entropy': 1.4788138717412949, 'clip_ratio/low_mean': 0.07532467693090439, 'clip_ratio/low_min': 0.07532467693090439, 'clip_ratio/high_mean': 0.0680365702137351, 'clip_ratio/high_max': 0.0680365702137351, 'clip_ratio/region_mean': 0.1433612471446395, 'reward_total_mean': 0.24094471335411072, 'reward_meter_mean': 0.5459390878677368, 'reward_meter_std': 0.43266281485557556, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.875, 'reward_lexical_plausibility_std': 0.3535533845424652, 'reward_repeat_penalty_mean': 0.7368881106376648, 'reward_repeat_penalty_std': 0.03708602488040924, 'reward_repeat_floor_mean': 0.983426570892334, 'reward_repeat_floor_std': 0.004450327716767788, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.28630343079566956, 'reward_total_composite_mean': 0.24094471335411072, 'reward_total_composite_std': 0.2565024495124817, 'epoch': 0.04}

 23%|██▎       | 753/3300 [2:49:56<23:15:04, 32.86s/it]INFO 04-12 18:48:23 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 754/3300 [2:50:10<19:08:16, 27.06s/it]2026-04-12 18:48:36,424 | INFO | train_grpo_train | metrics_logged mode=train step=754

                                                       
{'loss': -0.0328, 'grad_norm': 5.844229221343994, 'learning_rate': 7.718181818181819e-06, 'num_tokens': 1366630.0, 'completions/mean_length': 95.375, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.85714340209961, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.6300411224365234, 'rewards/meter/std': 0.41531339287757874, 'rewards/count_adherence/mean': 0.5833333730697632, 'rewards/count_adherence/std': 0.15430335700511932, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.875, 'rewards/count_floor/std': 0.046290989965200424, 'rewards/lexical_plausibility/mean': 0.9491525888442993, 'rewards/lexical_plausibility/std': 0.14381824433803558, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.2517368495464325, 'rewards/repeat_penalty/mean': 0.9001684188842773, 'rewards/repeat_penalty/std': 0.12743665277957916, 'rewards/repeat_floor/mean': 0.9909023642539978, 'rewards/repeat_floor/std': 0.011028815992176533, 'rewards/near_duplicate_penalty/mean': 0.9729430675506592, 'rewards/near_duplicate_penalty/std': 0.039715174585580826, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9856764078140259, 'rewards/distinct_2/std': 0.02661900594830513, 'rewards/total_composite/mean': 0.27034127712249756, 'rewards/total_composite/std': 0.2623963952064514, 'reward': 0.27034127712249756, 'reward_std': 0.2623963952064514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15583452582359314, 'sampling/sampling_logp_difference/max': 1.686234951019287, 'sampling/importance_sampling_ratio/min': 0.185215562582016, 'sampling/importance_sampling_ratio/mean': 1.0056499242782593, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8357549533247948, 'clip_ratio/low_mean': 0.06000405736267567, 'clip_ratio/low_min': 0.06000405736267567, 'clip_ratio/high_mean': 0.03045112732797861, 'clip_ratio/high_max': 0.03045112732797861, 'clip_ratio/region_mean': 0.09045518469065428, 'reward_total_mean': 0.27034127712249756, 'reward_meter_mean': 0.6300411224365234, 'reward_meter_std': 0.41531339287757874, 'reward_count_adherence_mean': 0.5833333730697632, 'reward_count_adherence_std': 0.15430335700511932, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.875, 'reward_count_floor_std': 0.046290989965200424, 'reward_lexical_plausibility_mean': 0.9491525888442993, 'reward_lexical_plausibility_std': 0.14381824433803558, 'reward_repeat_penalty_mean': 0.9001684188842773, 'reward_repeat_penalty_std': 0.12743665277957916, 'reward_repeat_floor_mean': 0.9909023642539978, 'reward_repeat_floor_std': 0.011028815992176533, 'reward_near_duplicate_penalty_mean': 0.9729430675506592, 'reward_near_duplicate_penalty_std': 0.039715174585580826, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9856764078140259, 'reward_distinct_2_std': 0.02661900594830513, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.2517368495464325, 'reward_total_composite_mean': 0.27034127712249756, 'reward_total_composite_std': 0.2623963952064514, 'epoch': 0.04}

 23%|██▎       | 754/3300 [2:50:10<19:08:16, 27.06s/it]INFO 04-12 18:48:36 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 755/3300 [2:50:23<16:09:22, 22.85s/it]2026-04-12 18:48:49,457 | INFO | train_grpo_train | metrics_logged mode=train step=755

                                                       
{'loss': -0.0229, 'grad_norm': 3.789351224899292, 'learning_rate': 7.715151515151516e-06, 'num_tokens': 1368035.0, 'completions/mean_length': 81.625, 'completions/min_length': 14.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.142858505249023, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.3168658912181854, 'rewards/meter/std': 0.3613809645175934, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9468423128128052, 'rewards/lexical_plausibility/std': 0.1503526270389557, 'rewards/judge_quality/mean': 0.5475000143051147, 'rewards/judge_quality/std': 0.4066500961780548, 'rewards/repeat_penalty/mean': 0.7781673073768616, 'rewards/repeat_penalty/std': 0.0900484099984169, 'rewards/repeat_floor/mean': 0.9862584471702576, 'rewards/repeat_floor/std': 0.005814488511532545, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16401417553424835, 'rewards/total_composite/std': 0.27484244108200073, 'reward': 0.16401417553424835, 'reward_std': 0.27484241127967834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1589149534702301, 'sampling/sampling_logp_difference/max': 1.23565673828125, 'sampling/importance_sampling_ratio/min': 0.2906438410282135, 'sampling/importance_sampling_ratio/mean': 1.0463230609893799, 'sampling/importance_sampling_ratio/max': 1.8723198175430298, 'entropy': 1.4279182776808739, 'clip_ratio/low_mean': 0.08604692062363029, 'clip_ratio/low_min': 0.08604692062363029, 'clip_ratio/high_mean': 0.02291666716337204, 'clip_ratio/high_max': 0.02291666716337204, 'clip_ratio/region_mean': 0.10896358778700233, 'reward_total_mean': 0.16401417553424835, 'reward_meter_mean': 0.3168658912181854, 'reward_meter_std': 0.3613809645175934, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9468423128128052, 'reward_lexical_plausibility_std': 0.1503526270389557, 'reward_repeat_penalty_mean': 0.7781673073768616, 'reward_repeat_penalty_std': 0.0900484099984169, 'reward_repeat_floor_mean': 0.9862584471702576, 'reward_repeat_floor_std': 0.005814488511532545, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5475000143051147, 'reward_judge_quality_std': 0.4066500961780548, 'reward_total_composite_mean': 0.16401417553424835, 'reward_total_composite_std': 0.27484244108200073, 'epoch': 0.04}

 23%|██▎       | 755/3300 [2:50:23<16:09:22, 22.85s/it]INFO 04-12 18:48:49 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 756/3300 [2:50:36<14:04:34, 19.92s/it]2026-04-12 18:49:02,527 | INFO | train_grpo_train | metrics_logged mode=train step=756

                                                       
{'loss': -0.0174, 'grad_norm': 5.291820049285889, 'learning_rate': 7.712121212121213e-06, 'num_tokens': 1369236.0, 'completions/mean_length': 83.125, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.85714340209961, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.46067625284194946, 'rewards/meter/std': 0.46334439516067505, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9267593622207642, 'rewards/lexical_plausibility/std': 0.20715588331222534, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.2840491831302643, 'rewards/repeat_penalty/mean': 0.735744059085846, 'rewards/repeat_penalty/std': 0.03279595449566841, 'rewards/repeat_floor/mean': 0.9821488261222839, 'rewards/repeat_floor/std': 0.006559199187904596, 'rewards/near_duplicate_penalty/mean': 0.9809920787811279, 'rewards/near_duplicate_penalty/std': 0.043727945536375046, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14458605647087097, 'rewards/total_composite/std': 0.18077252805233002, 'reward': 0.14458605647087097, 'reward_std': 0.18077251315116882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21747484803199768, 'sampling/sampling_logp_difference/max': 1.2292728424072266, 'sampling/importance_sampling_ratio/min': 0.2925052046775818, 'sampling/importance_sampling_ratio/mean': 1.0790289640426636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.625254675745964, 'clip_ratio/low_mean': 0.10710093099623919, 'clip_ratio/low_min': 0.10710093099623919, 'clip_ratio/high_mean': 0.04680735990405083, 'clip_ratio/high_max': 0.04680735990405083, 'clip_ratio/region_mean': 0.15390829090029, 'reward_total_mean': 0.14458605647087097, 'reward_meter_mean': 0.46067625284194946, 'reward_meter_std': 0.46334439516067505, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9267593622207642, 'reward_lexical_plausibility_std': 0.20715588331222534, 'reward_repeat_penalty_mean': 0.735744059085846, 'reward_repeat_penalty_std': 0.03279595449566841, 'reward_repeat_floor_mean': 0.9821488261222839, 'reward_repeat_floor_std': 0.006559199187904596, 'reward_near_duplicate_penalty_mean': 0.9809920787811279, 'reward_near_duplicate_penalty_std': 0.043727945536375046, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.2840491831302643, 'reward_total_composite_mean': 0.14458605647087097, 'reward_total_composite_std': 0.18077252805233002, 'epoch': 0.04}

 23%|██▎       | 756/3300 [2:50:36<14:04:34, 19.92s/it]INFO 04-12 18:49:02 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 757/3300 [2:50:45<11:43:33, 16.60s/it]2026-04-12 18:49:11,383 | INFO | train_grpo_train | metrics_logged mode=train step=757

                                                       
{'loss': -0.1391, 'grad_norm': 18.59693717956543, 'learning_rate': 7.709090909090909e-06, 'num_tokens': 1370688.0, 'completions/mean_length': 24.5, 'completions/min_length': 18.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.4819759726524353, 'rewards/meter/std': 0.31004247069358826, 'rewards/count_adherence/mean': 0.4166666865348816, 'rewards/count_adherence/std': 0.15430335700511932, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8250000476837158, 'rewards/count_floor/std': 0.046290989965200424, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.222898930311203, 'rewards/repeat_penalty/mean': 0.72732013463974, 'rewards/repeat_penalty/std': 0.053876038640737534, 'rewards/repeat_floor/mean': 0.9797546863555908, 'rewards/repeat_floor/std': 0.006581379100680351, 'rewards/near_duplicate_penalty/mean': 0.9795132279396057, 'rewards/near_duplicate_penalty/std': 0.03453369438648224, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9550296068191528, 'rewards/distinct_2/std': 0.06208225339651108, 'rewards/total_composite/mean': 0.1905805915594101, 'rewards/total_composite/std': 0.2514793574810028, 'reward': 0.1905805915594101, 'reward_std': 0.2514793574810028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18915100395679474, 'sampling/sampling_logp_difference/max': 1.3864779472351074, 'sampling/importance_sampling_ratio/min': 0.24995411932468414, 'sampling/importance_sampling_ratio/mean': 1.0098398923873901, 'sampling/importance_sampling_ratio/max': 1.930884838104248, 'entropy': 1.3853571638464928, 'clip_ratio/low_mean': 0.11050179274752736, 'clip_ratio/low_min': 0.11050179274752736, 'clip_ratio/high_mean': 0.053520623594522476, 'clip_ratio/high_max': 0.053520623594522476, 'clip_ratio/region_mean': 0.16402241634204984, 'reward_total_mean': 0.1905805915594101, 'reward_meter_mean': 0.4819759726524353, 'reward_meter_std': 0.31004247069358826, 'reward_count_adherence_mean': 0.4166666865348816, 'reward_count_adherence_std': 0.15430335700511932, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8250000476837158, 'reward_count_floor_std': 0.046290989965200424, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.72732013463974, 'reward_repeat_penalty_std': 0.053876038640737534, 'reward_repeat_floor_mean': 0.9797546863555908, 'reward_repeat_floor_std': 0.006581379100680351, 'reward_near_duplicate_penalty_mean': 0.9795132279396057, 'reward_near_duplicate_penalty_std': 0.03453369438648224, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9550296068191528, 'reward_distinct_2_std': 0.06208225339651108, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.222898930311203, 'reward_total_composite_mean': 0.1905805915594101, 'reward_total_composite_std': 0.2514793574810028, 'epoch': 0.04}

 23%|██▎       | 757/3300 [2:50:45<11:43:33, 16.60s/it]INFO 04-12 18:49:11 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 758/3300 [2:50:53<9:52:24, 13.98s/it] 2026-04-12 18:49:19,263 | INFO | train_grpo_train | metrics_logged mode=train step=758

                                                      
{'loss': 0.0694, 'grad_norm': 19.346956253051758, 'learning_rate': 7.706060606060606e-06, 'num_tokens': 1372126.0, 'completions/mean_length': 20.75, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.4748724699020386, 'rewards/meter/std': 0.4451950490474701, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.6617925763130188, 'rewards/lexical_plausibility/std': 0.2446226179599762, 'rewards/judge_quality/mean': 0.48875001072883606, 'rewards/judge_quality/std': 0.3665841519832611, 'rewards/repeat_penalty/mean': 0.7433712482452393, 'rewards/repeat_penalty/std': 0.018749047070741653, 'rewards/repeat_floor/mean': 0.9836742877960205, 'rewards/repeat_floor/std': 0.0037498052697628736, 'rewards/near_duplicate_penalty/mean': 0.991161584854126, 'rewards/near_duplicate_penalty/std': 0.024998728185892105, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14455866813659668, 'rewards/total_composite/std': 0.17668859660625458, 'reward': 0.14455866813659668, 'reward_std': 0.17668859660625458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18759851157665253, 'sampling/sampling_logp_difference/max': 1.6440258026123047, 'sampling/importance_sampling_ratio/min': 0.19320067763328552, 'sampling/importance_sampling_ratio/mean': 1.0177215337753296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5919382125139236, 'clip_ratio/low_mean': 0.09926338586956263, 'clip_ratio/low_min': 0.09926338586956263, 'clip_ratio/high_mean': 0.0575396828353405, 'clip_ratio/high_max': 0.0575396828353405, 'clip_ratio/region_mean': 0.15680306870490313, 'reward_total_mean': 0.14455866813659668, 'reward_meter_mean': 0.4748724699020386, 'reward_meter_std': 0.4451950490474701, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.6617925763130188, 'reward_lexical_plausibility_std': 0.2446226179599762, 'reward_repeat_penalty_mean': 0.7433712482452393, 'reward_repeat_penalty_std': 0.018749047070741653, 'reward_repeat_floor_mean': 0.9836742877960205, 'reward_repeat_floor_std': 0.0037498052697628736, 'reward_near_duplicate_penalty_mean': 0.991161584854126, 'reward_near_duplicate_penalty_std': 0.024998728185892105, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48875001072883606, 'reward_judge_quality_std': 0.3665841519832611, 'reward_total_composite_mean': 0.14455866813659668, 'reward_total_composite_std': 0.17668859660625458, 'epoch': 0.04}

 23%|██▎       | 758/3300 [2:50:53<9:52:24, 13.98s/it]INFO 04-12 18:49:19 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 759/3300 [2:51:08<10:11:47, 14.45s/it]2026-04-12 18:49:34,946 | INFO | train_grpo_train | metrics_logged mode=train step=759

                                                       
{'loss': -0.104, 'grad_norm': 4.506317138671875, 'learning_rate': 7.703030303030304e-06, 'num_tokens': 1374570.0, 'completions/mean_length': 184.5, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 137.71429443359375, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 190.0, 'rewards/meter/mean': 0.8124877214431763, 'rewards/meter/std': 0.16841427981853485, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.16690459847450256, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.05007137358188629, 'rewards/lexical_plausibility/mean': 0.9502505660057068, 'rewards/lexical_plausibility/std': 0.1407126933336258, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.7897309064865112, 'rewards/repeat_penalty/std': 0.18313255906105042, 'rewards/repeat_floor/mean': 0.9783972501754761, 'rewards/repeat_floor/std': 0.01762981154024601, 'rewards/near_duplicate_penalty/mean': 0.9580878019332886, 'rewards/near_duplicate_penalty/std': 0.018783865496516228, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.829822838306427, 'rewards/distinct_2/std': 0.1714695394039154, 'rewards/total_composite/mean': 0.24015222489833832, 'rewards/total_composite/std': 0.11441846936941147, 'reward': 0.24015222489833832, 'reward_std': 0.11441846936941147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1576281487941742, 'sampling/sampling_logp_difference/max': 2.4853200912475586, 'sampling/importance_sampling_ratio/min': 0.08329888433218002, 'sampling/importance_sampling_ratio/mean': 1.0077157020568848, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7918769866228104, 'clip_ratio/low_mean': 0.04094129614531994, 'clip_ratio/low_min': 0.04094129614531994, 'clip_ratio/high_mean': 0.10996764153242111, 'clip_ratio/high_max': 0.10996764153242111, 'clip_ratio/region_mean': 0.15090893767774105, 'reward_total_mean': 0.24015222489833832, 'reward_meter_mean': 0.8124877214431763, 'reward_meter_std': 0.16841427981853485, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.16690459847450256, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.05007137358188629, 'reward_lexical_plausibility_mean': 0.9502505660057068, 'reward_lexical_plausibility_std': 0.1407126933336258, 'reward_repeat_penalty_mean': 0.7897309064865112, 'reward_repeat_penalty_std': 0.18313255906105042, 'reward_repeat_floor_mean': 0.9783972501754761, 'reward_repeat_floor_std': 0.01762981154024601, 'reward_near_duplicate_penalty_mean': 0.9580878019332886, 'reward_near_duplicate_penalty_std': 0.018783865496516228, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.829822838306427, 'reward_distinct_2_std': 0.1714695394039154, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.24015222489833832, 'reward_total_composite_std': 0.11441846936941147, 'epoch': 0.04}

 23%|██▎       | 759/3300 [2:51:08<10:11:47, 14.45s/it]INFO 04-12 18:49:35 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 760/3300 [2:51:17<9:02:32, 12.82s/it] 2026-04-12 18:49:43,806 | INFO | train_grpo_train | metrics_logged mode=train step=760

                                                      
{'loss': 0.0818, 'grad_norm': 19.084365844726562, 'learning_rate': 7.7e-06, 'num_tokens': 1375952.0, 'completions/mean_length': 18.75, 'completions/min_length': 16.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7186318635940552, 'rewards/meter/std': 0.3999112546443939, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.36402514576911926, 'rewards/repeat_penalty/mean': 0.5965873003005981, 'rewards/repeat_penalty/std': 0.18547387421131134, 'rewards/repeat_floor/mean': 0.9547529220581055, 'rewards/repeat_floor/std': 0.03867047652602196, 'rewards/near_duplicate_penalty/mean': 0.8332011699676514, 'rewards/near_duplicate_penalty/std': 0.22255395352840424, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.941919207572937, 'rewards/distinct_2/std': 0.08058978617191315, 'rewards/total_composite/mean': 0.3209429979324341, 'rewards/total_composite/std': 0.3452916443347931, 'reward': 0.3209429979324341, 'reward_std': 0.3452916145324707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16951622068881989, 'sampling/sampling_logp_difference/max': 1.3270177841186523, 'sampling/importance_sampling_ratio/min': 0.26526719331741333, 'sampling/importance_sampling_ratio/mean': 1.0366358757019043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1174676269292831, 'clip_ratio/low_mean': 0.11512605333700776, 'clip_ratio/low_min': 0.11512605333700776, 'clip_ratio/high_mean': 0.045036764815449715, 'clip_ratio/high_max': 0.045036764815449715, 'clip_ratio/region_mean': 0.16016281815245748, 'reward_total_mean': 0.3209429979324341, 'reward_meter_mean': 0.7186318635940552, 'reward_meter_std': 0.3999112546443939, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5965873003005981, 'reward_repeat_penalty_std': 0.18547387421131134, 'reward_repeat_floor_mean': 0.9547529220581055, 'reward_repeat_floor_std': 0.03867047652602196, 'reward_near_duplicate_penalty_mean': 0.8332011699676514, 'reward_near_duplicate_penalty_std': 0.22255395352840424, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.941919207572937, 'reward_distinct_2_std': 0.08058978617191315, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.36402514576911926, 'reward_total_composite_mean': 0.3209429979324341, 'reward_total_composite_std': 0.3452916443347931, 'epoch': 0.04}

 23%|██▎       | 760/3300 [2:51:17<9:02:32, 12.82s/it]INFO 04-12 18:49:44 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 761/3300 [2:51:31<9:13:01, 13.07s/it]2026-04-12 18:49:57,461 | INFO | train_grpo_train | metrics_logged mode=train step=761

                                                      
{'loss': -0.0231, 'grad_norm': 4.444335460662842, 'learning_rate': 7.696969696969696e-06, 'num_tokens': 1377264.0, 'completions/mean_length': 83.0, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.71428680419922, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6338531970977783, 'rewards/meter/std': 0.4407992959022522, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9379624128341675, 'rewards/lexical_plausibility/std': 0.17546886205673218, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3537932336330414, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.38226625323295593, 'rewards/total_composite/std': 0.3537594974040985, 'reward': 0.38226625323295593, 'reward_std': 0.3537594676017761, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18164275586605072, 'sampling/sampling_logp_difference/max': 1.1295795440673828, 'sampling/importance_sampling_ratio/min': 0.3735601305961609, 'sampling/importance_sampling_ratio/mean': 1.0501426458358765, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2641790956258774, 'clip_ratio/low_mean': 0.0896728141233325, 'clip_ratio/low_min': 0.0896728141233325, 'clip_ratio/high_mean': 0.0623641312122345, 'clip_ratio/high_max': 0.0623641312122345, 'clip_ratio/region_mean': 0.152036945335567, 'reward_total_mean': 0.38226625323295593, 'reward_meter_mean': 0.6338531970977783, 'reward_meter_std': 0.4407992959022522, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9379624128341675, 'reward_lexical_plausibility_std': 0.17546886205673218, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3537932336330414, 'reward_total_composite_mean': 0.38226625323295593, 'reward_total_composite_std': 0.3537594974040985, 'epoch': 0.04}

 23%|██▎       | 761/3300 [2:51:31<9:13:01, 13.07s/it]INFO 04-12 18:49:57 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 762/3300 [2:51:45<9:22:46, 13.30s/it]2026-04-12 18:50:11,312 | INFO | train_grpo_train | metrics_logged mode=train step=762

                                                      
{'loss': -0.0231, 'grad_norm': 2.7324411869049072, 'learning_rate': 7.693939393939395e-06, 'num_tokens': 1378702.0, 'completions/mean_length': 143.75, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.331742525100708, 'rewards/meter/std': 0.3675425350666046, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.871656060218811, 'rewards/lexical_plausibility/std': 0.17920765280723572, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.2070196568965912, 'rewards/repeat_penalty/mean': 0.760830819606781, 'rewards/repeat_penalty/std': 0.10899937152862549, 'rewards/repeat_floor/mean': 0.9828301668167114, 'rewards/repeat_floor/std': 0.012803645804524422, 'rewards/near_duplicate_penalty/mean': 0.9730343818664551, 'rewards/near_duplicate_penalty/std': 0.07378175109624863, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1149199828505516, 'rewards/total_composite/std': 0.1412688046693802, 'reward': 0.1149199828505516, 'reward_std': 0.141268789768219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14124970138072968, 'sampling/sampling_logp_difference/max': 1.1507205963134766, 'sampling/importance_sampling_ratio/min': 0.4002891778945923, 'sampling/importance_sampling_ratio/mean': 1.0113756656646729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0592435225844383, 'clip_ratio/low_mean': 0.041125542018562555, 'clip_ratio/low_min': 0.041125542018562555, 'clip_ratio/high_mean': 0.04743392672389746, 'clip_ratio/high_max': 0.04743392672389746, 'clip_ratio/region_mean': 0.08855946874246001, 'reward_total_mean': 0.1149199828505516, 'reward_meter_mean': 0.331742525100708, 'reward_meter_std': 0.3675425350666046, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.871656060218811, 'reward_lexical_plausibility_std': 0.17920765280723572, 'reward_repeat_penalty_mean': 0.760830819606781, 'reward_repeat_penalty_std': 0.10899937152862549, 'reward_repeat_floor_mean': 0.9828301668167114, 'reward_repeat_floor_std': 0.012803645804524422, 'reward_near_duplicate_penalty_mean': 0.9730343818664551, 'reward_near_duplicate_penalty_std': 0.07378175109624863, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.2070196568965912, 'reward_total_composite_mean': 0.1149199828505516, 'reward_total_composite_std': 0.1412688046693802, 'epoch': 0.04}

 23%|██▎       | 762/3300 [2:51:45<9:22:46, 13.30s/it]INFO 04-12 18:50:11 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 763/3300 [2:51:54<8:27:58, 12.01s/it]2026-04-12 18:50:20,312 | INFO | train_grpo_train | metrics_logged mode=train step=763

                                                      
{'loss': -0.1787, 'grad_norm': 18.535858154296875, 'learning_rate': 7.690909090909091e-06, 'num_tokens': 1380205.0, 'completions/mean_length': 24.875, 'completions/min_length': 16.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7672863006591797, 'rewards/meter/std': 0.3053992688655853, 'rewards/count_adherence/mean': 0.4583333432674408, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8375000357627869, 'rewards/count_floor/std': 0.0517548993229866, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.8299659490585327, 'rewards/repeat_penalty/std': 0.13155357539653778, 'rewards/repeat_floor/mean': 0.9881312847137451, 'rewards/repeat_floor/std': 0.009411572478711605, 'rewards/near_duplicate_penalty/mean': 0.983375072479248, 'rewards/near_duplicate_penalty/std': 0.03218412771821022, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1267128735780716, 'rewards/total_composite/std': 0.05846923589706421, 'reward': 0.1267128735780716, 'reward_std': 0.05846923217177391, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15891475975513458, 'sampling/sampling_logp_difference/max': 1.1123703718185425, 'sampling/importance_sampling_ratio/min': 0.33009979128837585, 'sampling/importance_sampling_ratio/mean': 1.0086604356765747, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0872657150030136, 'clip_ratio/low_mean': 0.10213108593598008, 'clip_ratio/low_min': 0.10213108593598008, 'clip_ratio/high_mean': 0.03632478695362806, 'clip_ratio/high_max': 0.03632478695362806, 'clip_ratio/region_mean': 0.13845587288960814, 'reward_total_mean': 0.1267128735780716, 'reward_meter_mean': 0.7672863006591797, 'reward_meter_std': 0.3053992688655853, 'reward_count_adherence_mean': 0.4583333432674408, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8375000357627869, 'reward_count_floor_std': 0.0517548993229866, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8299659490585327, 'reward_repeat_penalty_std': 0.13155357539653778, 'reward_repeat_floor_mean': 0.9881312847137451, 'reward_repeat_floor_std': 0.009411572478711605, 'reward_near_duplicate_penalty_mean': 0.983375072479248, 'reward_near_duplicate_penalty_std': 0.03218412771821022, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1267128735780716, 'reward_total_composite_std': 0.05846923589706421, 'epoch': 0.04}

 23%|██▎       | 763/3300 [2:51:54<8:27:58, 12.01s/it]INFO 04-12 18:50:20 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 764/3300 [2:52:03<7:50:17, 11.13s/it]2026-04-12 18:50:29,372 | INFO | train_grpo_train | metrics_logged mode=train step=764

                                                      
{'loss': 0.2131, 'grad_norm': 22.37726593017578, 'learning_rate': 7.687878787878788e-06, 'num_tokens': 1381584.0, 'completions/mean_length': 22.375, 'completions/min_length': 14.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.375, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5623931884765625, 'rewards/meter/std': 0.3798246681690216, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.3184336721897125, 'rewards/repeat_penalty/mean': 0.8179469704627991, 'rewards/repeat_penalty/std': 0.12802553176879883, 'rewards/repeat_floor/mean': 0.9861041307449341, 'rewards/repeat_floor/std': 0.01027802936732769, 'rewards/near_duplicate_penalty/mean': 0.9698607921600342, 'rewards/near_duplicate_penalty/std': 0.0485549233853817, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30107706785202026, 'rewards/total_composite/std': 0.2895428240299225, 'reward': 0.30107706785202026, 'reward_std': 0.2895428240299225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15181639790534973, 'sampling/sampling_logp_difference/max': 1.1574530601501465, 'sampling/importance_sampling_ratio/min': 0.3174467384815216, 'sampling/importance_sampling_ratio/mean': 0.9998915195465088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8709793724119663, 'clip_ratio/low_mean': 0.09685446508228779, 'clip_ratio/low_min': 0.09685446508228779, 'clip_ratio/high_mean': 0.05131579004228115, 'clip_ratio/high_max': 0.05131579004228115, 'clip_ratio/region_mean': 0.14817025512456894, 'reward_total_mean': 0.30107706785202026, 'reward_meter_mean': 0.5623931884765625, 'reward_meter_std': 0.3798246681690216, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8179469704627991, 'reward_repeat_penalty_std': 0.12802553176879883, 'reward_repeat_floor_mean': 0.9861041307449341, 'reward_repeat_floor_std': 0.01027802936732769, 'reward_near_duplicate_penalty_mean': 0.9698607921600342, 'reward_near_duplicate_penalty_std': 0.0485549233853817, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.3184336721897125, 'reward_total_composite_mean': 0.30107706785202026, 'reward_total_composite_std': 0.2895428240299225, 'epoch': 0.04}

 23%|██▎       | 764/3300 [2:52:03<7:50:17, 11.13s/it]INFO 04-12 18:50:29 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 765/3300 [2:52:11<7:16:42, 10.34s/it]2026-04-12 18:50:37,924 | INFO | train_grpo_train | metrics_logged mode=train step=765

                                                      
{'loss': 0.0753, 'grad_norm': 15.800399780273438, 'learning_rate': 7.684848484848485e-06, 'num_tokens': 1383197.0, 'completions/mean_length': 22.625, 'completions/min_length': 18.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9605438709259033, 'rewards/meter/std': 0.032537493854761124, 'rewards/count_adherence/mean': 0.375, 'rewards/count_adherence/std': 0.1178511381149292, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8125, 'rewards/count_floor/std': 0.035355325788259506, 'rewards/lexical_plausibility/mean': 0.953125, 'rewards/lexical_plausibility/std': 0.13258251547813416, 'rewards/judge_quality/mean': 0.35875001549720764, 'rewards/judge_quality/std': 0.27105283737182617, 'rewards/repeat_penalty/mean': 0.564224362373352, 'rewards/repeat_penalty/std': 0.2216183990240097, 'rewards/repeat_floor/mean': 0.9506906270980835, 'rewards/repeat_floor/std': 0.03978775069117546, 'rewards/near_duplicate_penalty/mean': 0.8543699979782104, 'rewards/near_duplicate_penalty/std': 0.1575518399477005, 'rewards/opening_diversity/mean': 0.703125, 'rewards/opening_diversity/std': 0.13258251547813416, 'rewards/distinct_2/mean': 0.8927515149116516, 'rewards/distinct_2/std': 0.1609673798084259, 'rewards/total_composite/mean': 0.2627718448638916, 'rewards/total_composite/std': 0.19434994459152222, 'reward': 0.2627718448638916, 'reward_std': 0.1943499594926834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13608522713184357, 'sampling/sampling_logp_difference/max': 1.119798183441162, 'sampling/importance_sampling_ratio/min': 0.32634562253952026, 'sampling/importance_sampling_ratio/mean': 1.0080320835113525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9093529507517815, 'clip_ratio/low_mean': 0.06417606631293893, 'clip_ratio/low_min': 0.06417606631293893, 'clip_ratio/high_mean': 0.07747114170342684, 'clip_ratio/high_max': 0.07747114170342684, 'clip_ratio/region_mean': 0.14164720801636577, 'reward_total_mean': 0.2627718448638916, 'reward_meter_mean': 0.9605438709259033, 'reward_meter_std': 0.032537493854761124, 'reward_count_adherence_mean': 0.375, 'reward_count_adherence_std': 0.1178511381149292, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8125, 'reward_count_floor_std': 0.035355325788259506, 'reward_lexical_plausibility_mean': 0.953125, 'reward_lexical_plausibility_std': 0.13258251547813416, 'reward_repeat_penalty_mean': 0.564224362373352, 'reward_repeat_penalty_std': 0.2216183990240097, 'reward_repeat_floor_mean': 0.9506906270980835, 'reward_repeat_floor_std': 0.03978775069117546, 'reward_near_duplicate_penalty_mean': 0.8543699979782104, 'reward_near_duplicate_penalty_std': 0.1575518399477005, 'reward_opening_diversity_mean': 0.703125, 'reward_opening_diversity_std': 0.13258251547813416, 'reward_distinct_2_mean': 0.8927515149116516, 'reward_distinct_2_std': 0.1609673798084259, 'reward_judge_quality_mean': 0.35875001549720764, 'reward_judge_quality_std': 0.27105283737182617, 'reward_total_composite_mean': 0.2627718448638916, 'reward_total_composite_std': 0.19434994459152222, 'epoch': 0.04}

 23%|██▎       | 765/3300 [2:52:11<7:16:42, 10.34s/it]INFO 04-12 18:50:38 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 766/3300 [2:52:20<6:54:46,  9.82s/it]2026-04-12 18:50:46,483 | INFO | train_grpo_train | metrics_logged mode=train step=766

                                                      
{'loss': 0.0216, 'grad_norm': 19.032678604125977, 'learning_rate': 7.681818181818183e-06, 'num_tokens': 1384716.0, 'completions/mean_length': 21.875, 'completions/min_length': 19.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.5013790726661682, 'rewards/meter/std': 0.30632445216178894, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9652406573295593, 'rewards/lexical_plausibility/std': 0.0983143076300621, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.32275325059890747, 'rewards/repeat_penalty/mean': 0.7008553743362427, 'rewards/repeat_penalty/std': 0.0752108246088028, 'rewards/repeat_floor/mean': 0.9760586619377136, 'rewards/repeat_floor/std': 0.01471381913870573, 'rewards/near_duplicate_penalty/mean': 0.949266791343689, 'rewards/near_duplicate_penalty/std': 0.10010528564453125, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.3125600814819336, 'rewards/total_composite/std': 0.27480924129486084, 'reward': 0.3125600814819336, 'reward_std': 0.27480921149253845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15522697567939758, 'sampling/sampling_logp_difference/max': 1.051997184753418, 'sampling/importance_sampling_ratio/min': 0.3492395579814911, 'sampling/importance_sampling_ratio/mean': 1.0159152746200562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.231025904417038, 'clip_ratio/low_mean': 0.08325826842337847, 'clip_ratio/low_min': 0.08325826842337847, 'clip_ratio/high_mean': 0.04471343941986561, 'clip_ratio/high_max': 0.04471343941986561, 'clip_ratio/region_mean': 0.12797170784324408, 'reward_total_mean': 0.3125600814819336, 'reward_meter_mean': 0.5013790726661682, 'reward_meter_std': 0.30632445216178894, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9652406573295593, 'reward_lexical_plausibility_std': 0.0983143076300621, 'reward_repeat_penalty_mean': 0.7008553743362427, 'reward_repeat_penalty_std': 0.0752108246088028, 'reward_repeat_floor_mean': 0.9760586619377136, 'reward_repeat_floor_std': 0.01471381913870573, 'reward_near_duplicate_penalty_mean': 0.949266791343689, 'reward_near_duplicate_penalty_std': 0.10010528564453125, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.6937500238418579, 'reward_judge_quality_std': 0.32275325059890747, 'reward_total_composite_mean': 0.3125600814819336, 'reward_total_composite_std': 0.27480924129486084, 'epoch': 0.04}

 23%|██▎       | 766/3300 [2:52:20<6:54:46,  9.82s/it]INFO 04-12 18:50:46 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 767/3300 [2:52:28<6:31:08,  9.27s/it]2026-04-12 18:50:54,450 | INFO | train_grpo_train | metrics_logged mode=train step=767

                                                      
{'loss': 0.047, 'grad_norm': 21.347537994384766, 'learning_rate': 7.678787878787878e-06, 'num_tokens': 1386086.0, 'completions/mean_length': 19.25, 'completions/min_length': 17.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9754035472869873, 'rewards/meter/std': 0.016338838264346123, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.476250022649765, 'rewards/judge_quality/std': 0.38119879364967346, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3886607885360718, 'rewards/total_composite/std': 0.3102341890335083, 'reward': 0.3886607885360718, 'reward_std': 0.3102341592311859, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16929809749126434, 'sampling/sampling_logp_difference/max': 1.5765259265899658, 'sampling/importance_sampling_ratio/min': 0.2066919058561325, 'sampling/importance_sampling_ratio/mean': 0.9983431100845337, 'sampling/importance_sampling_ratio/max': 1.9500958919525146, 'entropy': 0.8574133589863777, 'clip_ratio/low_mean': 0.10224089957773685, 'clip_ratio/low_min': 0.10224089957773685, 'clip_ratio/high_mean': 0.05807072902098298, 'clip_ratio/high_max': 0.05807072902098298, 'clip_ratio/region_mean': 0.16031162859871984, 'reward_total_mean': 0.3886607885360718, 'reward_meter_mean': 0.9754035472869873, 'reward_meter_std': 0.016338838264346123, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.476250022649765, 'reward_judge_quality_std': 0.38119879364967346, 'reward_total_composite_mean': 0.3886607885360718, 'reward_total_composite_std': 0.3102341890335083, 'epoch': 0.04}

 23%|██▎       | 767/3300 [2:52:28<6:31:08,  9.27s/it]INFO 04-12 18:50:54 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 768/3300 [2:52:36<6:18:39,  8.97s/it]2026-04-12 18:51:02,742 | INFO | train_grpo_train | metrics_logged mode=train step=768

                                                      
{'loss': -0.0195, 'grad_norm': 20.816057205200195, 'learning_rate': 7.675757575757577e-06, 'num_tokens': 1387476.0, 'completions/mean_length': 24.75, 'completions/min_length': 17.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.41209256649017334, 'rewards/meter/std': 0.38532668352127075, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9887152910232544, 'rewards/lexical_plausibility/std': 0.022863244637846947, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.33457863330841064, 'rewards/repeat_penalty/mean': 0.8101656436920166, 'rewards/repeat_penalty/std': 0.1115168109536171, 'rewards/repeat_floor/mean': 0.988399863243103, 'rewards/repeat_floor/std': 0.006339644081890583, 'rewards/near_duplicate_penalty/mean': 0.9976656436920166, 'rewards/near_duplicate_penalty/std': 0.0066026621498167515, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2490522861480713, 'rewards/total_composite/std': 0.2924487590789795, 'reward': 0.2490522861480713, 'reward_std': 0.2924487888813019, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1720062792301178, 'sampling/sampling_logp_difference/max': 1.9223260879516602, 'sampling/importance_sampling_ratio/min': 0.14626634120941162, 'sampling/importance_sampling_ratio/mean': 1.0118714570999146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.984496034681797, 'clip_ratio/low_mean': 0.08240398345515132, 'clip_ratio/low_min': 0.08240398345515132, 'clip_ratio/high_mean': 0.052736480720341206, 'clip_ratio/high_max': 0.052736480720341206, 'clip_ratio/region_mean': 0.13514046417549253, 'reward_total_mean': 0.2490522861480713, 'reward_meter_mean': 0.41209256649017334, 'reward_meter_std': 0.38532668352127075, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9887152910232544, 'reward_lexical_plausibility_std': 0.022863244637846947, 'reward_repeat_penalty_mean': 0.8101656436920166, 'reward_repeat_penalty_std': 0.1115168109536171, 'reward_repeat_floor_mean': 0.988399863243103, 'reward_repeat_floor_std': 0.006339644081890583, 'reward_near_duplicate_penalty_mean': 0.9976656436920166, 'reward_near_duplicate_penalty_std': 0.0066026621498167515, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.33457863330841064, 'reward_total_composite_mean': 0.2490522861480713, 'reward_total_composite_std': 0.2924487590789795, 'epoch': 0.04}

 23%|██▎       | 768/3300 [2:52:36<6:18:39,  8.97s/it]INFO 04-12 18:51:02 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 769/3300 [2:52:44<6:04:28,  8.64s/it]2026-04-12 18:51:10,611 | INFO | train_grpo_train | metrics_logged mode=train step=769

                                                      
{'loss': 0.0236, 'grad_norm': 25.497764587402344, 'learning_rate': 7.672727272727273e-06, 'num_tokens': 1389023.0, 'completions/mean_length': 19.375, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.9459700584411621, 'rewards/meter/std': 0.06000656634569168, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8487499952316284, 'rewards/judge_quality/std': 0.20152544975280762, 'rewards/repeat_penalty/mean': 0.7207263112068176, 'rewards/repeat_penalty/std': 0.055174242705106735, 'rewards/repeat_floor/mean': 0.9791452884674072, 'rewards/repeat_floor/std': 0.011034860275685787, 'rewards/near_duplicate_penalty/mean': 0.9609683752059937, 'rewards/near_duplicate_penalty/std': 0.07356566190719604, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6651339530944824, 'rewards/total_composite/std': 0.15791328251361847, 'reward': 0.6651339530944824, 'reward_std': 0.15791326761245728, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15176261961460114, 'sampling/sampling_logp_difference/max': 1.8464980125427246, 'sampling/importance_sampling_ratio/min': 0.15778878331184387, 'sampling/importance_sampling_ratio/mean': 0.9938704967498779, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8959499150514603, 'clip_ratio/low_mean': 0.01315789483487606, 'clip_ratio/low_min': 0.01315789483487606, 'clip_ratio/high_mean': 0.12202381249517202, 'clip_ratio/high_max': 0.12202381249517202, 'clip_ratio/region_mean': 0.13518170733004808, 'reward_total_mean': 0.6651339530944824, 'reward_meter_mean': 0.9459700584411621, 'reward_meter_std': 0.06000656634569168, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7207263112068176, 'reward_repeat_penalty_std': 0.055174242705106735, 'reward_repeat_floor_mean': 0.9791452884674072, 'reward_repeat_floor_std': 0.011034860275685787, 'reward_near_duplicate_penalty_mean': 0.9609683752059937, 'reward_near_duplicate_penalty_std': 0.07356566190719604, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8487499952316284, 'reward_judge_quality_std': 0.20152544975280762, 'reward_total_composite_mean': 0.6651339530944824, 'reward_total_composite_std': 0.15791328251361847, 'epoch': 0.04}

 23%|██▎       | 769/3300 [2:52:44<6:04:28,  8.64s/it]INFO 04-12 18:51:10 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 770/3300 [2:52:53<6:12:57,  8.84s/it]2026-04-12 18:51:19,931 | INFO | train_grpo_train | metrics_logged mode=train step=770

                                                      
{'loss': -0.0293, 'grad_norm': 13.898656845092773, 'learning_rate': 7.66969696969697e-06, 'num_tokens': 1390773.0, 'completions/mean_length': 48.75, 'completions/min_length': 37.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5010685920715332, 'rewards/meter/std': 0.4373815357685089, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.7499796152114868, 'rewards/repeat_penalty/std': 0.14986592531204224, 'rewards/repeat_floor/mean': 0.9721471071243286, 'rewards/repeat_floor/std': 0.016216160729527473, 'rewards/near_duplicate_penalty/mean': 0.9233490824699402, 'rewards/near_duplicate_penalty/std': 0.030025022104382515, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8599413633346558, 'rewards/distinct_2/std': 0.09072302281856537, 'rewards/total_composite/mean': 0.14605870842933655, 'rewards/total_composite/std': 0.14427883923053741, 'reward': 0.14605870842933655, 'reward_std': 0.14427883923053741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.154762402176857, 'sampling/sampling_logp_difference/max': 2.8827919960021973, 'sampling/importance_sampling_ratio/min': 0.05597825348377228, 'sampling/importance_sampling_ratio/mean': 0.9965939521789551, 'sampling/importance_sampling_ratio/max': 1.865016222000122, 'entropy': 0.8376650288701057, 'clip_ratio/low_mean': 0.05851264717057347, 'clip_ratio/low_min': 0.05851264717057347, 'clip_ratio/high_mean': 0.05808691866695881, 'clip_ratio/high_max': 0.05808691866695881, 'clip_ratio/region_mean': 0.11659956583753228, 'reward_total_mean': 0.14605870842933655, 'reward_meter_mean': 0.5010685920715332, 'reward_meter_std': 0.4373815357685089, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7499796152114868, 'reward_repeat_penalty_std': 0.14986592531204224, 'reward_repeat_floor_mean': 0.9721471071243286, 'reward_repeat_floor_std': 0.016216160729527473, 'reward_near_duplicate_penalty_mean': 0.9233490824699402, 'reward_near_duplicate_penalty_std': 0.030025022104382515, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8599413633346558, 'reward_distinct_2_std': 0.09072302281856537, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.14605870842933655, 'reward_total_composite_std': 0.14427883923053741, 'epoch': 0.04}

 23%|██▎       | 770/3300 [2:52:53<6:12:57,  8.84s/it]INFO 04-12 18:51:20 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 771/3300 [2:53:06<7:06:18, 10.11s/it]2026-04-12 18:51:33,012 | INFO | train_grpo_train | metrics_logged mode=train step=771

                                                      
{'loss': 0.0163, 'grad_norm': 5.95081901550293, 'learning_rate': 7.666666666666667e-06, 'num_tokens': 1394565.0, 'completions/mean_length': 230.0, 'completions/min_length': 46.0, 'completions/max_length': 321.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 230.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 321.0, 'rewards/meter/mean': 0.5900455713272095, 'rewards/meter/std': 0.28857189416885376, 'rewards/count_adherence/mean': 0.7578125, 'rewards/count_adherence/std': 0.2681731879711151, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9273437261581421, 'rewards/count_floor/std': 0.08045195043087006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.4184042811393738, 'rewards/repeat_penalty/std': 0.2967893183231354, 'rewards/repeat_floor/mean': 0.9340304732322693, 'rewards/repeat_floor/std': 0.04270326346158981, 'rewards/near_duplicate_penalty/mean': 0.8616848587989807, 'rewards/near_duplicate_penalty/std': 0.11876453459262848, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.49753016233444214, 'rewards/distinct_2/std': 0.29442593455314636, 'rewards/total_composite/mean': 0.12593936920166016, 'rewards/total_composite/std': 0.06832549721002579, 'reward': 0.12593936920166016, 'reward_std': 0.06832549721002579, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13308840990066528, 'sampling/sampling_logp_difference/max': 2.553238868713379, 'sampling/importance_sampling_ratio/min': 0.07782917469739914, 'sampling/importance_sampling_ratio/mean': 1.0099968910217285, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8225430846214294, 'clip_ratio/low_mean': 0.07484854385256767, 'clip_ratio/low_min': 0.07484854385256767, 'clip_ratio/high_mean': 0.05350201576948166, 'clip_ratio/high_max': 0.05350201576948166, 'clip_ratio/region_mean': 0.12835055962204933, 'reward_total_mean': 0.12593936920166016, 'reward_meter_mean': 0.5900455713272095, 'reward_meter_std': 0.28857189416885376, 'reward_count_adherence_mean': 0.7578125, 'reward_count_adherence_std': 0.2681731879711151, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9273437261581421, 'reward_count_floor_std': 0.08045195043087006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4184042811393738, 'reward_repeat_penalty_std': 0.2967893183231354, 'reward_repeat_floor_mean': 0.9340304732322693, 'reward_repeat_floor_std': 0.04270326346158981, 'reward_near_duplicate_penalty_mean': 0.8616848587989807, 'reward_near_duplicate_penalty_std': 0.11876453459262848, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.49753016233444214, 'reward_distinct_2_std': 0.29442593455314636, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.12593936920166016, 'reward_total_composite_std': 0.06832549721002579, 'epoch': 0.04}

 23%|██▎       | 771/3300 [2:53:06<7:06:18, 10.11s/it]INFO 04-12 18:51:33 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 772/3300 [2:53:20<7:50:02, 11.16s/it]2026-04-12 18:51:46,591 | INFO | train_grpo_train | metrics_logged mode=train step=772

                                                      
{'loss': -0.0801, 'grad_norm': 6.436338901519775, 'learning_rate': 7.663636363636364e-06, 'num_tokens': 1396581.0, 'completions/mean_length': 129.0, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.28572082519531, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.19514545798301697, 'rewards/meter/std': 0.19754625856876373, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.27368009090423584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.08210402727127075, 'rewards/lexical_plausibility/mean': 0.9324172735214233, 'rewards/lexical_plausibility/std': 0.19115281105041504, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8844287991523743, 'rewards/repeat_penalty/std': 0.07957295328378677, 'rewards/repeat_floor/mean': 0.9859362840652466, 'rewards/repeat_floor/std': 0.008351316675543785, 'rewards/near_duplicate_penalty/mean': 0.9252746105194092, 'rewards/near_duplicate_penalty/std': 0.05803005024790764, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9891117811203003, 'rewards/distinct_2/std': 0.020227652043104172, 'rewards/total_composite/mean': 0.07932163774967194, 'rewards/total_composite/std': 0.0918048620223999, 'reward': 0.07932163774967194, 'reward_std': 0.0918048620223999, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15314671397209167, 'sampling/sampling_logp_difference/max': 2.5624470710754395, 'sampling/importance_sampling_ratio/min': 0.07711579650640488, 'sampling/importance_sampling_ratio/mean': 1.0120174884796143, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5992117673158646, 'clip_ratio/low_mean': 0.10750545747578144, 'clip_ratio/low_min': 0.10750545747578144, 'clip_ratio/high_mean': 0.040029761381447315, 'clip_ratio/high_max': 0.040029761381447315, 'clip_ratio/region_mean': 0.14753521885722876, 'reward_total_mean': 0.07932163774967194, 'reward_meter_mean': 0.19514545798301697, 'reward_meter_std': 0.19754625856876373, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.27368009090423584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.08210402727127075, 'reward_lexical_plausibility_mean': 0.9324172735214233, 'reward_lexical_plausibility_std': 0.19115281105041504, 'reward_repeat_penalty_mean': 0.8844287991523743, 'reward_repeat_penalty_std': 0.07957295328378677, 'reward_repeat_floor_mean': 0.9859362840652466, 'reward_repeat_floor_std': 0.008351316675543785, 'reward_near_duplicate_penalty_mean': 0.9252746105194092, 'reward_near_duplicate_penalty_std': 0.05803005024790764, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9891117811203003, 'reward_distinct_2_std': 0.020227652043104172, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.07932163774967194, 'reward_total_composite_std': 0.0918048620223999, 'epoch': 0.04}

 23%|██▎       | 772/3300 [2:53:20<7:50:02, 11.16s/it]INFO 04-12 18:51:46 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 773/3300 [2:53:28<7:16:23, 10.36s/it]2026-04-12 18:51:55,103 | INFO | train_grpo_train | metrics_logged mode=train step=773

                                                      
{'loss': 0.0417, 'grad_norm': 23.878934860229492, 'learning_rate': 7.660606060606062e-06, 'num_tokens': 1397985.0, 'completions/mean_length': 26.5, 'completions/min_length': 16.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7131302952766418, 'rewards/meter/std': 0.2686402499675751, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.5762500166893005, 'rewards/judge_quality/std': 0.3882538676261902, 'rewards/repeat_penalty/mean': 0.8704565763473511, 'rewards/repeat_penalty/std': 0.1289837807416916, 'rewards/repeat_floor/mean': 0.9918185472488403, 'rewards/repeat_floor/std': 0.0073724305257201195, 'rewards/near_duplicate_penalty/mean': 0.9954565763473511, 'rewards/near_duplicate_penalty/std': 0.008819373324513435, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3349722623825073, 'rewards/total_composite/std': 0.24952907860279083, 'reward': 0.3349722623825073, 'reward_std': 0.24952907860279083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16240321099758148, 'sampling/sampling_logp_difference/max': 1.4973726272583008, 'sampling/importance_sampling_ratio/min': 0.22371718287467957, 'sampling/importance_sampling_ratio/mean': 1.0314677953720093, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9310154989361763, 'clip_ratio/low_mean': 0.07899436168372631, 'clip_ratio/low_min': 0.07899436168372631, 'clip_ratio/high_mean': 0.08602390484884381, 'clip_ratio/high_max': 0.08602390484884381, 'clip_ratio/region_mean': 0.16501826653257012, 'reward_total_mean': 0.3349722623825073, 'reward_meter_mean': 0.7131302952766418, 'reward_meter_std': 0.2686402499675751, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.8704565763473511, 'reward_repeat_penalty_std': 0.1289837807416916, 'reward_repeat_floor_mean': 0.9918185472488403, 'reward_repeat_floor_std': 0.0073724305257201195, 'reward_near_duplicate_penalty_mean': 0.9954565763473511, 'reward_near_duplicate_penalty_std': 0.008819373324513435, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5762500166893005, 'reward_judge_quality_std': 0.3882538676261902, 'reward_total_composite_mean': 0.3349722623825073, 'reward_total_composite_std': 0.24952907860279083, 'epoch': 0.04}

 23%|██▎       | 773/3300 [2:53:28<7:16:23, 10.36s/it]INFO 04-12 18:51:55 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 774/3300 [2:53:37<6:52:34,  9.80s/it]2026-04-12 18:52:03,589 | INFO | train_grpo_train | metrics_logged mode=train step=774

                                                      
{'loss': 0.0013, 'grad_norm': 17.045730590820312, 'learning_rate': 7.657575757575757e-06, 'num_tokens': 1399485.0, 'completions/mean_length': 23.5, 'completions/min_length': 19.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8890976905822754, 'rewards/meter/std': 0.2827149033546448, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.27523693442344666, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.39331406354904175, 'rewards/total_composite/std': 0.2601402699947357, 'reward': 0.39331406354904175, 'reward_std': 0.2601402699947357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12740592658519745, 'sampling/sampling_logp_difference/max': 1.1320905685424805, 'sampling/importance_sampling_ratio/min': 0.3223586678504944, 'sampling/importance_sampling_ratio/mean': 1.0031788349151611, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7947101294994354, 'clip_ratio/low_mean': 0.06576450075954199, 'clip_ratio/low_min': 0.06576450075954199, 'clip_ratio/high_mean': 0.04166666604578495, 'clip_ratio/high_max': 0.04166666604578495, 'clip_ratio/region_mean': 0.10743116680532694, 'reward_total_mean': 0.39331406354904175, 'reward_meter_mean': 0.8890976905822754, 'reward_meter_std': 0.2827149033546448, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.27523693442344666, 'reward_total_composite_mean': 0.39331406354904175, 'reward_total_composite_std': 0.2601402699947357, 'epoch': 0.04}

 23%|██▎       | 774/3300 [2:53:37<6:52:34,  9.80s/it]INFO 04-12 18:52:03 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 775/3300 [2:53:46<6:38:00,  9.46s/it]2026-04-12 18:52:12,274 | INFO | train_grpo_train | metrics_logged mode=train step=775

                                                      
{'loss': 0.2762, 'grad_norm': 19.8206729888916, 'learning_rate': 7.654545454545456e-06, 'num_tokens': 1400995.0, 'completions/mean_length': 28.75, 'completions/min_length': 12.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6935251951217651, 'rewards/meter/std': 0.40879955887794495, 'rewards/count_adherence/mean': 0.5416666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.862500011920929, 'rewards/count_floor/std': 0.0517548993229866, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5387500524520874, 'rewards/judge_quality/std': 0.3228638768196106, 'rewards/repeat_penalty/mean': 0.8410468697547913, 'rewards/repeat_penalty/std': 0.1193373054265976, 'rewards/repeat_floor/mean': 0.986666202545166, 'rewards/repeat_floor/std': 0.010483510792255402, 'rewards/near_duplicate_penalty/mean': 0.9743757843971252, 'rewards/near_duplicate_penalty/std': 0.03457370027899742, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9570537805557251, 'rewards/distinct_2/std': 0.08207473903894424, 'rewards/total_composite/mean': 0.3474574089050293, 'rewards/total_composite/std': 0.31205835938453674, 'reward': 0.3474574089050293, 'reward_std': 0.31205835938453674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1438266634941101, 'sampling/sampling_logp_difference/max': 1.1797747611999512, 'sampling/importance_sampling_ratio/min': 0.30734795331954956, 'sampling/importance_sampling_ratio/mean': 1.004528284072876, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8478525876998901, 'clip_ratio/low_mean': 0.08346232445910573, 'clip_ratio/low_min': 0.08346232445910573, 'clip_ratio/high_mean': 0.04856601823121309, 'clip_ratio/high_max': 0.04856601823121309, 'clip_ratio/region_mean': 0.13202834269031882, 'reward_total_mean': 0.3474574089050293, 'reward_meter_mean': 0.6935251951217651, 'reward_meter_std': 0.40879955887794495, 'reward_count_adherence_mean': 0.5416666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.862500011920929, 'reward_count_floor_std': 0.0517548993229866, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8410468697547913, 'reward_repeat_penalty_std': 0.1193373054265976, 'reward_repeat_floor_mean': 0.986666202545166, 'reward_repeat_floor_std': 0.010483510792255402, 'reward_near_duplicate_penalty_mean': 0.9743757843971252, 'reward_near_duplicate_penalty_std': 0.03457370027899742, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9570537805557251, 'reward_distinct_2_std': 0.08207473903894424, 'reward_judge_quality_mean': 0.5387500524520874, 'reward_judge_quality_std': 0.3228638768196106, 'reward_total_composite_mean': 0.3474574089050293, 'reward_total_composite_std': 0.31205835938453674, 'epoch': 0.04}

 23%|██▎       | 775/3300 [2:53:46<6:38:00,  9.46s/it]INFO 04-12 18:52:12 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 776/3300 [2:53:54<6:25:53,  9.17s/it]2026-04-12 18:52:20,759 | INFO | train_grpo_train | metrics_logged mode=train step=776

                                                      
{'loss': 0.0504, 'grad_norm': 13.736774444580078, 'learning_rate': 7.651515151515152e-06, 'num_tokens': 1402390.0, 'completions/mean_length': 18.375, 'completions/min_length': 18.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.892931342124939, 'rewards/meter/std': 0.2214132845401764, 'rewards/count_adherence/mean': 0.25, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7749999761581421, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.644318163394928, 'rewards/repeat_penalty/std': 0.2624778747558594, 'rewards/repeat_floor/mean': 0.9733480215072632, 'rewards/repeat_floor/std': 0.026111982762813568, 'rewards/near_duplicate_penalty/mean': 0.9703094959259033, 'rewards/near_duplicate_penalty/std': 0.05516405403614044, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.23547068238258362, 'rewards/total_composite/std': 0.05812937766313553, 'reward': 0.23547068238258362, 'reward_std': 0.05812938138842583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1799391359090805, 'sampling/sampling_logp_difference/max': 1.858041763305664, 'sampling/importance_sampling_ratio/min': 0.1559777706861496, 'sampling/importance_sampling_ratio/mean': 0.9716464281082153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8302898481488228, 'clip_ratio/low_mean': 0.032894738018512726, 'clip_ratio/low_min': 0.032894738018512726, 'clip_ratio/high_mean': 0.13633040990680456, 'clip_ratio/high_max': 0.13633040990680456, 'clip_ratio/region_mean': 0.1692251479253173, 'reward_total_mean': 0.23547068238258362, 'reward_meter_mean': 0.892931342124939, 'reward_meter_std': 0.2214132845401764, 'reward_count_adherence_mean': 0.25, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7749999761581421, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.644318163394928, 'reward_repeat_penalty_std': 0.2624778747558594, 'reward_repeat_floor_mean': 0.9733480215072632, 'reward_repeat_floor_std': 0.026111982762813568, 'reward_near_duplicate_penalty_mean': 0.9703094959259033, 'reward_near_duplicate_penalty_std': 0.05516405403614044, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.23547068238258362, 'reward_total_composite_std': 0.05812937766313553, 'epoch': 0.04}

 24%|██▎       | 776/3300 [2:53:54<6:25:53,  9.17s/it]INFO 04-12 18:52:21 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 777/3300 [2:54:03<6:22:51,  9.10s/it]2026-04-12 18:52:29,702 | INFO | train_grpo_train | metrics_logged mode=train step=777

                                                      
{'loss': 0.0782, 'grad_norm': 18.51751708984375, 'learning_rate': 7.648484848484849e-06, 'num_tokens': 1403824.0, 'completions/mean_length': 20.25, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8164852261543274, 'rewards/meter/std': 0.3480514883995056, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.7368881106376648, 'rewards/repeat_penalty/std': 0.03708602488040924, 'rewards/repeat_floor/mean': 0.983426570892334, 'rewards/repeat_floor/std': 0.004450327716767788, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.25499558448791504, 'rewards/total_composite/std': 0.1301426887512207, 'reward': 0.25499558448791504, 'reward_std': 0.1301426887512207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14567610621452332, 'sampling/sampling_logp_difference/max': 1.052895188331604, 'sampling/importance_sampling_ratio/min': 0.3489260673522949, 'sampling/importance_sampling_ratio/mean': 1.0312671661376953, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1905914917588234, 'clip_ratio/low_mean': 0.051037549041211605, 'clip_ratio/low_min': 0.051037549041211605, 'clip_ratio/high_mean': 0.08946805074810982, 'clip_ratio/high_max': 0.08946805074810982, 'clip_ratio/region_mean': 0.14050559978932142, 'reward_total_mean': 0.25499558448791504, 'reward_meter_mean': 0.8164852261543274, 'reward_meter_std': 0.3480514883995056, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.7368881106376648, 'reward_repeat_penalty_std': 0.03708602488040924, 'reward_repeat_floor_mean': 0.983426570892334, 'reward_repeat_floor_std': 0.004450327716767788, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.25499558448791504, 'reward_total_composite_std': 0.1301426887512207, 'epoch': 0.04}

 24%|██▎       | 777/3300 [2:54:03<6:22:51,  9.10s/it]INFO 04-12 18:52:29 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 778/3300 [2:54:10<5:59:40,  8.56s/it]2026-04-12 18:52:36,981 | INFO | train_grpo_train | metrics_logged mode=train step=778

                                                      
{'loss': -0.1217, 'grad_norm': 20.82761573791504, 'learning_rate': 7.645454545454546e-06, 'num_tokens': 1405206.0, 'completions/mean_length': 19.75, 'completions/min_length': 12.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.6180265545845032, 'rewards/meter/std': 0.4441920816898346, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9312499761581421, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.7890859246253967, 'rewards/repeat_penalty/std': 0.1365775465965271, 'rewards/repeat_floor/mean': 0.9859402775764465, 'rewards/repeat_floor/std': 0.00999416969716549, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9687812328338623, 'rewards/distinct_2/std': 0.05835825204849243, 'rewards/total_composite/mean': 0.564579963684082, 'rewards/total_composite/std': 0.40636146068573, 'reward': 0.564579963684082, 'reward_std': 0.40636146068573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1289128065109253, 'sampling/sampling_logp_difference/max': 1.2647600173950195, 'sampling/importance_sampling_ratio/min': 0.2823070287704468, 'sampling/importance_sampling_ratio/mean': 0.982999324798584, 'sampling/importance_sampling_ratio/max': 1.4815729856491089, 'entropy': 0.6345707513391972, 'clip_ratio/low_mean': 0.05520979035645723, 'clip_ratio/low_min': 0.05520979035645723, 'clip_ratio/high_mean': 0.05383971333503723, 'clip_ratio/high_max': 0.05383971333503723, 'clip_ratio/region_mean': 0.10904950369149446, 'reward_total_mean': 0.564579963684082, 'reward_meter_mean': 0.6180265545845032, 'reward_meter_std': 0.4441920816898346, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7890859246253967, 'reward_repeat_penalty_std': 0.1365775465965271, 'reward_repeat_floor_mean': 0.9859402775764465, 'reward_repeat_floor_std': 0.00999416969716549, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9687812328338623, 'reward_distinct_2_std': 0.05835825204849243, 'reward_judge_quality_mean': 0.9312499761581421, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.564579963684082, 'reward_total_composite_std': 0.40636146068573, 'epoch': 0.04}

 24%|██▎       | 778/3300 [2:54:10<5:59:40,  8.56s/it]INFO 04-12 18:52:37 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 779/3300 [2:54:18<5:50:22,  8.34s/it]2026-04-12 18:52:44,850 | INFO | train_grpo_train | metrics_logged mode=train step=779

                                                      
{'loss': 0.0559, 'grad_norm': 22.683591842651367, 'learning_rate': 7.642424242424244e-06, 'num_tokens': 1406452.0, 'completions/mean_length': 26.75, 'completions/min_length': 17.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.855337381362915, 'rewards/meter/std': 0.3440517783164978, 'rewards/count_adherence/mean': 0.34375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8031250238418579, 'rewards/count_floor/std': 0.038816213607788086, 'rewards/lexical_plausibility/mean': 0.9713541269302368, 'rewards/lexical_plausibility/std': 0.05990393087267876, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101404666900635, 'rewards/repeat_penalty/mean': 0.809446394443512, 'rewards/repeat_penalty/std': 0.10299328714609146, 'rewards/repeat_floor/mean': 0.9855901002883911, 'rewards/repeat_floor/std': 0.006884741596877575, 'rewards/near_duplicate_penalty/mean': 0.9712417125701904, 'rewards/near_duplicate_penalty/std': 0.04128456115722656, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.29597342014312744, 'rewards/total_composite/std': 0.2679290473461151, 'reward': 0.29597342014312744, 'reward_std': 0.2679290473461151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17572063207626343, 'sampling/sampling_logp_difference/max': 1.550114631652832, 'sampling/importance_sampling_ratio/min': 0.21222363412380219, 'sampling/importance_sampling_ratio/mean': 1.0125097036361694, 'sampling/importance_sampling_ratio/max': 1.7293217182159424, 'entropy': 1.4376767873764038, 'clip_ratio/low_mean': 0.07957887835800648, 'clip_ratio/low_min': 0.07957887835800648, 'clip_ratio/high_mean': 0.08406385313719511, 'clip_ratio/high_max': 0.08406385313719511, 'clip_ratio/region_mean': 0.1636427314952016, 'reward_total_mean': 0.29597342014312744, 'reward_meter_mean': 0.855337381362915, 'reward_meter_std': 0.3440517783164978, 'reward_count_adherence_mean': 0.34375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8031250238418579, 'reward_count_floor_std': 0.038816213607788086, 'reward_lexical_plausibility_mean': 0.9713541269302368, 'reward_lexical_plausibility_std': 0.05990393087267876, 'reward_repeat_penalty_mean': 0.809446394443512, 'reward_repeat_penalty_std': 0.10299328714609146, 'reward_repeat_floor_mean': 0.9855901002883911, 'reward_repeat_floor_std': 0.006884741596877575, 'reward_near_duplicate_penalty_mean': 0.9712417125701904, 'reward_near_duplicate_penalty_std': 0.04128456115722656, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101404666900635, 'reward_total_composite_mean': 0.29597342014312744, 'reward_total_composite_std': 0.2679290473461151, 'epoch': 0.04}

 24%|██▎       | 779/3300 [2:54:18<5:50:22,  8.34s/it]INFO 04-12 18:52:45 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 780/3300 [2:54:26<5:47:10,  8.27s/it]2026-04-12 18:52:52,918 | INFO | train_grpo_train | metrics_logged mode=train step=780

                                                      
{'loss': 0.0862, 'grad_norm': 21.682519912719727, 'learning_rate': 7.639393939393939e-06, 'num_tokens': 1407900.0, 'completions/mean_length': 20.0, 'completions/min_length': 18.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9203027486801147, 'rewards/meter/std': 0.16414465010166168, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3890189230442047, 'rewards/repeat_penalty/mean': 0.7219443321228027, 'rewards/repeat_penalty/std': 0.06725989282131195, 'rewards/repeat_floor/mean': 0.9793888926506042, 'rewards/repeat_floor/std': 0.013451973907649517, 'rewards/near_duplicate_penalty/mean': 0.9625924229621887, 'rewards/near_duplicate_penalty/std': 0.08967985957860947, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3738017678260803, 'rewards/total_composite/std': 0.32943981885910034, 'reward': 0.3738017678260803, 'reward_std': 0.32943978905677795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15233810245990753, 'sampling/sampling_logp_difference/max': 0.9184441566467285, 'sampling/importance_sampling_ratio/min': 0.39913955330848694, 'sampling/importance_sampling_ratio/mean': 1.0313408374786377, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.019579291343689, 'clip_ratio/low_mean': 0.09335442213341594, 'clip_ratio/low_min': 0.09335442213341594, 'clip_ratio/high_mean': 0.020467836409807205, 'clip_ratio/high_max': 0.020467836409807205, 'clip_ratio/region_mean': 0.11382225854322314, 'reward_total_mean': 0.3738017678260803, 'reward_meter_mean': 0.9203027486801147, 'reward_meter_std': 0.16414465010166168, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7219443321228027, 'reward_repeat_penalty_std': 0.06725989282131195, 'reward_repeat_floor_mean': 0.9793888926506042, 'reward_repeat_floor_std': 0.013451973907649517, 'reward_near_duplicate_penalty_mean': 0.9625924229621887, 'reward_near_duplicate_penalty_std': 0.08967985957860947, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3890189230442047, 'reward_total_composite_mean': 0.3738017678260803, 'reward_total_composite_std': 0.32943981885910034, 'epoch': 0.04}

 24%|██▎       | 780/3300 [2:54:26<5:47:10,  8.27s/it]INFO 04-12 18:52:53 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 781/3300 [2:54:39<6:50:00,  9.77s/it]2026-04-12 18:53:06,174 | INFO | train_grpo_train | metrics_logged mode=train step=781

                                                      
{'loss': -0.0465, 'grad_norm': 1.260072112083435, 'learning_rate': 7.636363636363638e-06, 'num_tokens': 1409366.0, 'completions/mean_length': 149.25, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 28.33333396911621, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7019569277763367, 'rewards/meter/std': 0.4358839988708496, 'rewards/count_adherence/mean': 0.28125, 'rewards/count_adherence/std': 0.2086307406425476, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7843749523162842, 'rewards/count_floor/std': 0.06258922815322876, 'rewards/lexical_plausibility/mean': 0.8568053841590881, 'rewards/lexical_plausibility/std': 0.23534192144870758, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.2841905951499939, 'rewards/repeat_penalty/mean': 0.8117424249649048, 'rewards/repeat_penalty/std': 0.11433623731136322, 'rewards/repeat_floor/mean': 0.9886363744735718, 'rewards/repeat_floor/std': 0.006737613119184971, 'rewards/near_duplicate_penalty/mean': 0.9992424249649048, 'rewards/near_duplicate_penalty/std': 0.0021427457686513662, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2913835048675537, 'rewards/total_composite/std': 0.21905705332756042, 'reward': 0.2913835048675537, 'reward_std': 0.21905705332756042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15715575218200684, 'sampling/sampling_logp_difference/max': 1.2289297580718994, 'sampling/importance_sampling_ratio/min': 0.34187573194503784, 'sampling/importance_sampling_ratio/mean': 1.0295498371124268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7227039933204651, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11323255952447653, 'clip_ratio/high_max': 0.11323255952447653, 'clip_ratio/region_mean': 0.11323255952447653, 'reward_total_mean': 0.2913835048675537, 'reward_meter_mean': 0.7019569277763367, 'reward_meter_std': 0.4358839988708496, 'reward_count_adherence_mean': 0.28125, 'reward_count_adherence_std': 0.2086307406425476, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7843749523162842, 'reward_count_floor_std': 0.06258922815322876, 'reward_lexical_plausibility_mean': 0.8568053841590881, 'reward_lexical_plausibility_std': 0.23534192144870758, 'reward_repeat_penalty_mean': 0.8117424249649048, 'reward_repeat_penalty_std': 0.11433623731136322, 'reward_repeat_floor_mean': 0.9886363744735718, 'reward_repeat_floor_std': 0.006737613119184971, 'reward_near_duplicate_penalty_mean': 0.9992424249649048, 'reward_near_duplicate_penalty_std': 0.0021427457686513662, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.2841905951499939, 'reward_total_composite_mean': 0.2913835048675537, 'reward_total_composite_std': 0.21905705332756042, 'epoch': 0.04}

 24%|██▎       | 781/3300 [2:54:40<6:50:00,  9.77s/it]INFO 04-12 18:53:06 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 782/3300 [2:54:47<6:26:47,  9.22s/it]2026-04-12 18:53:14,109 | INFO | train_grpo_train | metrics_logged mode=train step=782

                                                      
{'loss': -0.0101, 'grad_norm': 18.58893585205078, 'learning_rate': 7.633333333333334e-06, 'num_tokens': 1410900.0, 'completions/mean_length': 21.75, 'completions/min_length': 19.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9149412512779236, 'rewards/meter/std': 0.18534734845161438, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.22369863092899323, 'rewards/repeat_penalty/mean': 0.7145075798034668, 'rewards/repeat_penalty/std': 0.054901186376810074, 'rewards/repeat_floor/mean': 0.9779015183448792, 'rewards/repeat_floor/std': 0.010980247519910336, 'rewards/near_duplicate_penalty/mean': 0.9526767730712891, 'rewards/near_duplicate_penalty/std': 0.07320158183574677, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.39992910623550415, 'rewards/total_composite/std': 0.20364829897880554, 'reward': 0.39992910623550415, 'reward_std': 0.20364831387996674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15350231528282166, 'sampling/sampling_logp_difference/max': 1.6557645797729492, 'sampling/importance_sampling_ratio/min': 0.19094599783420563, 'sampling/importance_sampling_ratio/mean': 1.0164743661880493, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.043894425034523, 'clip_ratio/low_mean': 0.09295454667881131, 'clip_ratio/low_min': 0.09295454667881131, 'clip_ratio/high_mean': 0.024027460254728794, 'clip_ratio/high_max': 0.024027460254728794, 'clip_ratio/region_mean': 0.1169820069335401, 'reward_total_mean': 0.39992910623550415, 'reward_meter_mean': 0.9149412512779236, 'reward_meter_std': 0.18534734845161438, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7145075798034668, 'reward_repeat_penalty_std': 0.054901186376810074, 'reward_repeat_floor_mean': 0.9779015183448792, 'reward_repeat_floor_std': 0.010980247519910336, 'reward_near_duplicate_penalty_mean': 0.9526767730712891, 'reward_near_duplicate_penalty_std': 0.07320158183574677, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.22369863092899323, 'reward_total_composite_mean': 0.39992910623550415, 'reward_total_composite_std': 0.20364829897880554, 'epoch': 0.04}

 24%|██▎       | 782/3300 [2:54:47<6:26:47,  9.22s/it]INFO 04-12 18:53:14 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 783/3300 [2:54:56<6:12:43,  8.89s/it]2026-04-12 18:53:22,221 | INFO | train_grpo_train | metrics_logged mode=train step=783

                                                      
{'loss': -0.0432, 'grad_norm': 18.332656860351562, 'learning_rate': 7.630303030303031e-06, 'num_tokens': 1412408.0, 'completions/mean_length': 15.5, 'completions/min_length': 13.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.5, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.9762332439422607, 'rewards/meter/std': 0.024528803303837776, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7063689827919006, 'rewards/total_composite/std': 0.13651399314403534, 'reward': 0.7063689827919006, 'reward_std': 0.13651399314403534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13365532457828522, 'sampling/sampling_logp_difference/max': 1.4031152725219727, 'sampling/importance_sampling_ratio/min': 0.24582993984222412, 'sampling/importance_sampling_ratio/mean': 1.006777286529541, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7066499292850494, 'clip_ratio/low_mean': 0.008928571827709675, 'clip_ratio/low_min': 0.008928571827709675, 'clip_ratio/high_mean': 0.05147058889269829, 'clip_ratio/high_max': 0.05147058889269829, 'clip_ratio/region_mean': 0.06039916072040796, 'reward_total_mean': 0.7063689827919006, 'reward_meter_mean': 0.9762332439422607, 'reward_meter_std': 0.024528803303837776, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.7063689827919006, 'reward_total_composite_std': 0.13651399314403534, 'epoch': 0.04}

 24%|██▎       | 783/3300 [2:54:56<6:12:43,  8.89s/it]INFO 04-12 18:53:22 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 784/3300 [2:55:04<6:02:06,  8.64s/it]2026-04-12 18:53:30,285 | INFO | train_grpo_train | metrics_logged mode=train step=784

                                                      
{'loss': -0.074, 'grad_norm': 19.76032829284668, 'learning_rate': 7.627272727272727e-06, 'num_tokens': 1414020.0, 'completions/mean_length': 25.5, 'completions/min_length': 17.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6522606611251831, 'rewards/meter/std': 0.2949579358100891, 'rewards/count_adherence/mean': 0.40625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8218749761581421, 'rewards/count_floor/std': 0.038816213607788086, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.865331768989563, 'rewards/repeat_penalty/std': 0.1077810749411583, 'rewards/repeat_floor/mean': 0.9888465404510498, 'rewards/repeat_floor/std': 0.006757983472198248, 'rewards/near_duplicate_penalty/mean': 0.9708355665206909, 'rewards/near_duplicate_penalty/std': 0.032967522740364075, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.2589820623397827, 'rewards/total_composite/std': 0.2225312888622284, 'reward': 0.2589820623397827, 'reward_std': 0.2225312888622284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1912866234779358, 'sampling/sampling_logp_difference/max': 1.6955208778381348, 'sampling/importance_sampling_ratio/min': 0.18350361287593842, 'sampling/importance_sampling_ratio/mean': 0.9580504298210144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.648014560341835, 'clip_ratio/low_mean': 0.08849715208634734, 'clip_ratio/low_min': 0.08849715208634734, 'clip_ratio/high_mean': 0.0570512842386961, 'clip_ratio/high_max': 0.0570512842386961, 'clip_ratio/region_mean': 0.14554843632504344, 'reward_total_mean': 0.2589820623397827, 'reward_meter_mean': 0.6522606611251831, 'reward_meter_std': 0.2949579358100891, 'reward_count_adherence_mean': 0.40625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8218749761581421, 'reward_count_floor_std': 0.038816213607788086, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.865331768989563, 'reward_repeat_penalty_std': 0.1077810749411583, 'reward_repeat_floor_mean': 0.9888465404510498, 'reward_repeat_floor_std': 0.006757983472198248, 'reward_near_duplicate_penalty_mean': 0.9708355665206909, 'reward_near_duplicate_penalty_std': 0.032967522740364075, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.2589820623397827, 'reward_total_composite_std': 0.2225312888622284, 'epoch': 0.04}

 24%|██▍       | 784/3300 [2:55:04<6:02:06,  8.64s/it]INFO 04-12 18:53:30 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 785/3300 [2:55:12<5:59:23,  8.57s/it]2026-04-12 18:53:38,714 | INFO | train_grpo_train | metrics_logged mode=train step=785

                                                      
{'loss': -0.1302, 'grad_norm': 17.985218048095703, 'learning_rate': 7.6242424242424254e-06, 'num_tokens': 1415606.0, 'completions/mean_length': 30.25, 'completions/min_length': 17.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9774325489997864, 'rewards/meter/std': 0.028761059045791626, 'rewards/count_adherence/mean': 0.34375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8031250238418579, 'rewards/count_floor/std': 0.038816213607788086, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.28400203585624695, 'rewards/repeat_penalty/mean': 0.7474256753921509, 'rewards/repeat_penalty/std': 0.13144512474536896, 'rewards/repeat_floor/mean': 0.9782500267028809, 'rewards/repeat_floor/std': 0.015048199333250523, 'rewards/near_duplicate_penalty/mean': 0.9494141340255737, 'rewards/near_duplicate_penalty/std': 0.066337950527668, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9676434993743896, 'rewards/distinct_2/std': 0.0638899877667427, 'rewards/total_composite/mean': 0.37300124764442444, 'rewards/total_composite/std': 0.23304860293865204, 'reward': 0.37300124764442444, 'reward_std': 0.23304857313632965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15827888250350952, 'sampling/sampling_logp_difference/max': 1.247124433517456, 'sampling/importance_sampling_ratio/min': 0.28732985258102417, 'sampling/importance_sampling_ratio/mean': 1.0281550884246826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.024580493569374, 'clip_ratio/low_mean': 0.1004318417981267, 'clip_ratio/low_min': 0.1004318417981267, 'clip_ratio/high_mean': 0.0513392873108387, 'clip_ratio/high_max': 0.0513392873108387, 'clip_ratio/region_mean': 0.1517711291089654, 'reward_total_mean': 0.37300124764442444, 'reward_meter_mean': 0.9774325489997864, 'reward_meter_std': 0.028761059045791626, 'reward_count_adherence_mean': 0.34375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8031250238418579, 'reward_count_floor_std': 0.038816213607788086, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7474256753921509, 'reward_repeat_penalty_std': 0.13144512474536896, 'reward_repeat_floor_mean': 0.9782500267028809, 'reward_repeat_floor_std': 0.015048199333250523, 'reward_near_duplicate_penalty_mean': 0.9494141340255737, 'reward_near_duplicate_penalty_std': 0.066337950527668, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9676434993743896, 'reward_distinct_2_std': 0.0638899877667427, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.28400203585624695, 'reward_total_composite_mean': 0.37300124764442444, 'reward_total_composite_std': 0.23304860293865204, 'epoch': 0.04}

 24%|██▍       | 785/3300 [2:55:12<5:59:23,  8.57s/it]INFO 04-12 18:53:38 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 786/3300 [2:55:21<6:09:55,  8.83s/it]2026-04-12 18:53:48,147 | INFO | train_grpo_train | metrics_logged mode=train step=786

                                                      
{'loss': -0.0504, 'grad_norm': 11.129945755004883, 'learning_rate': 7.621212121212122e-06, 'num_tokens': 1417780.0, 'completions/mean_length': 87.75, 'completions/min_length': 64.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.7454812526702881, 'rewards/meter/std': 0.27467748522758484, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.1259881556034088, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.03779645636677742, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.6524744033813477, 'rewards/repeat_penalty/std': 0.23615077137947083, 'rewards/repeat_floor/mean': 0.9642062187194824, 'rewards/repeat_floor/std': 0.025166330859065056, 'rewards/near_duplicate_penalty/mean': 0.917383074760437, 'rewards/near_duplicate_penalty/std': 0.046802930533885956, 'rewards/opening_diversity/mean': 0.9955357313156128, 'rewards/opening_diversity/std': 0.012626901268959045, 'rewards/distinct_2/mean': 0.7429624795913696, 'rewards/distinct_2/std': 0.21188749372959137, 'rewards/total_composite/mean': 0.27600806951522827, 'rewards/total_composite/std': 0.10370464622974396, 'reward': 0.27600806951522827, 'reward_std': 0.10370464622974396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16440506279468536, 'sampling/sampling_logp_difference/max': 3.912323236465454, 'sampling/importance_sampling_ratio/min': 0.019993996247649193, 'sampling/importance_sampling_ratio/mean': 1.0130337476730347, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7387968674302101, 'clip_ratio/low_mean': 0.05429098196327686, 'clip_ratio/low_min': 0.05429098196327686, 'clip_ratio/high_mean': 0.06683196406811476, 'clip_ratio/high_max': 0.06683196406811476, 'clip_ratio/region_mean': 0.12112294603139162, 'reward_total_mean': 0.27600806951522827, 'reward_meter_mean': 0.7454812526702881, 'reward_meter_std': 0.27467748522758484, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.1259881556034088, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.03779645636677742, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6524744033813477, 'reward_repeat_penalty_std': 0.23615077137947083, 'reward_repeat_floor_mean': 0.9642062187194824, 'reward_repeat_floor_std': 0.025166330859065056, 'reward_near_duplicate_penalty_mean': 0.917383074760437, 'reward_near_duplicate_penalty_std': 0.046802930533885956, 'reward_opening_diversity_mean': 0.9955357313156128, 'reward_opening_diversity_std': 0.012626901268959045, 'reward_distinct_2_mean': 0.7429624795913696, 'reward_distinct_2_std': 0.21188749372959137, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.27600806951522827, 'reward_total_composite_std': 0.10370464622974396, 'epoch': 0.04}

 24%|██▍       | 786/3300 [2:55:22<6:09:55,  8.83s/it]INFO 04-12 18:53:48 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 787/3300 [2:55:30<6:05:06,  8.72s/it]2026-04-12 18:53:56,587 | INFO | train_grpo_train | metrics_logged mode=train step=787

                                                      
{'loss': 0.0608, 'grad_norm': 25.511476516723633, 'learning_rate': 7.618181818181819e-06, 'num_tokens': 1419155.0, 'completions/mean_length': 20.875, 'completions/min_length': 17.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.6830810308456421, 'rewards/meter/std': 0.36726513504981995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7599999904632568, 'rewards/judge_quality/std': 0.31955328583717346, 'rewards/repeat_penalty/mean': 0.7362592816352844, 'rewards/repeat_penalty/std': 0.03886464238166809, 'rewards/repeat_floor/mean': 0.9822518825531006, 'rewards/repeat_floor/std': 0.007772932760417461, 'rewards/near_duplicate_penalty/mean': 0.9816790223121643, 'rewards/near_duplicate_penalty/std': 0.05181950703263283, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5857986211776733, 'rewards/total_composite/std': 0.3785436153411865, 'reward': 0.5857986211776733, 'reward_std': 0.3785436153411865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18615415692329407, 'sampling/sampling_logp_difference/max': 1.6064338684082031, 'sampling/importance_sampling_ratio/min': 0.2006017118692398, 'sampling/importance_sampling_ratio/mean': 1.0401734113693237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2724601104855537, 'clip_ratio/low_mean': 0.07842246256768703, 'clip_ratio/low_min': 0.07842246256768703, 'clip_ratio/high_mean': 0.11835214169695973, 'clip_ratio/high_max': 0.11835214169695973, 'clip_ratio/region_mean': 0.19677460426464677, 'reward_total_mean': 0.5857986211776733, 'reward_meter_mean': 0.6830810308456421, 'reward_meter_std': 0.36726513504981995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7362592816352844, 'reward_repeat_penalty_std': 0.03886464238166809, 'reward_repeat_floor_mean': 0.9822518825531006, 'reward_repeat_floor_std': 0.007772932760417461, 'reward_near_duplicate_penalty_mean': 0.9816790223121643, 'reward_near_duplicate_penalty_std': 0.05181950703263283, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7599999904632568, 'reward_judge_quality_std': 0.31955328583717346, 'reward_total_composite_mean': 0.5857986211776733, 'reward_total_composite_std': 0.3785436153411865, 'epoch': 0.04}

 24%|██▍       | 787/3300 [2:55:30<6:05:06,  8.72s/it]INFO 04-12 18:53:56 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 788/3300 [2:55:38<5:58:08,  8.55s/it]2026-04-12 18:54:04,760 | INFO | train_grpo_train | metrics_logged mode=train step=788

                                                      
{'loss': 0.044, 'grad_norm': 18.48018455505371, 'learning_rate': 7.6151515151515155e-06, 'num_tokens': 1420563.0, 'completions/mean_length': 25.0, 'completions/min_length': 23.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7448182106018066, 'rewards/meter/std': 0.41385018825531006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9514399766921997, 'rewards/lexical_plausibility/std': 0.09064352512359619, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.3384418189525604, 'rewards/repeat_penalty/mean': 0.7196282148361206, 'rewards/repeat_penalty/std': 0.06807498633861542, 'rewards/repeat_floor/mean': 0.9789255857467651, 'rewards/repeat_floor/std': 0.013615014031529427, 'rewards/near_duplicate_penalty/mean': 0.959504246711731, 'rewards/near_duplicate_penalty/std': 0.09076667577028275, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5100557804107666, 'rewards/total_composite/std': 0.3952203392982483, 'reward': 0.5100557804107666, 'reward_std': 0.3952203392982483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14302374422550201, 'sampling/sampling_logp_difference/max': 1.2075798511505127, 'sampling/importance_sampling_ratio/min': 0.32901689410209656, 'sampling/importance_sampling_ratio/mean': 0.9992201924324036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7803836241364479, 'clip_ratio/low_mean': 0.05947115458548069, 'clip_ratio/low_min': 0.05947115458548069, 'clip_ratio/high_mean': 0.055292642675340176, 'clip_ratio/high_max': 0.055292642675340176, 'clip_ratio/region_mean': 0.11476379726082087, 'reward_total_mean': 0.5100557804107666, 'reward_meter_mean': 0.7448182106018066, 'reward_meter_std': 0.41385018825531006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9514399766921997, 'reward_lexical_plausibility_std': 0.09064352512359619, 'reward_repeat_penalty_mean': 0.7196282148361206, 'reward_repeat_penalty_std': 0.06807498633861542, 'reward_repeat_floor_mean': 0.9789255857467651, 'reward_repeat_floor_std': 0.013615014031529427, 'reward_near_duplicate_penalty_mean': 0.959504246711731, 'reward_near_duplicate_penalty_std': 0.09076667577028275, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.3384418189525604, 'reward_total_composite_mean': 0.5100557804107666, 'reward_total_composite_std': 0.3952203392982483, 'epoch': 0.04}

 24%|██▍       | 788/3300 [2:55:38<5:58:08,  8.55s/it]INFO 04-12 18:54:05 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 789/3300 [2:55:47<6:02:11,  8.65s/it]2026-04-12 18:54:13,647 | INFO | train_grpo_train | metrics_logged mode=train step=789

                                                      
{'loss': -0.0437, 'grad_norm': 20.2589168548584, 'learning_rate': 7.612121212121213e-06, 'num_tokens': 1422221.0, 'completions/mean_length': 41.25, 'completions/min_length': 25.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7006809711456299, 'rewards/meter/std': 0.378921777009964, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9298311471939087, 'rewards/repeat_penalty/std': 0.09482555836439133, 'rewards/repeat_floor/mean': 0.9919579029083252, 'rewards/repeat_floor/std': 0.00876479409635067, 'rewards/near_duplicate_penalty/mean': 0.9686388373374939, 'rewards/near_duplicate_penalty/std': 0.01773165911436081, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.9748168587684631, 'rewards/distinct_2/std': 0.048887718468904495, 'rewards/total_composite/mean': 0.28199052810668945, 'rewards/total_composite/std': 0.15307407081127167, 'reward': 0.28199052810668945, 'reward_std': 0.15307407081127167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19265636801719666, 'sampling/sampling_logp_difference/max': 1.7889890670776367, 'sampling/importance_sampling_ratio/min': 0.1671290546655655, 'sampling/importance_sampling_ratio/mean': 0.9828406572341919, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6921148151159286, 'clip_ratio/low_mean': 0.07544191926717758, 'clip_ratio/low_min': 0.07544191926717758, 'clip_ratio/high_mean': 0.12784320954233408, 'clip_ratio/high_max': 0.12784320954233408, 'clip_ratio/region_mean': 0.20328512880951166, 'reward_total_mean': 0.28199052810668945, 'reward_meter_mean': 0.7006809711456299, 'reward_meter_std': 0.378921777009964, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9298311471939087, 'reward_repeat_penalty_std': 0.09482555836439133, 'reward_repeat_floor_mean': 0.9919579029083252, 'reward_repeat_floor_std': 0.00876479409635067, 'reward_near_duplicate_penalty_mean': 0.9686388373374939, 'reward_near_duplicate_penalty_std': 0.01773165911436081, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.9748168587684631, 'reward_distinct_2_std': 0.048887718468904495, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.28199052810668945, 'reward_total_composite_std': 0.15307407081127167, 'epoch': 0.04}

 24%|██▍       | 789/3300 [2:55:47<6:02:11,  8.65s/it]INFO 04-12 18:54:13 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 790/3300 [2:55:58<6:28:29,  9.29s/it]2026-04-12 18:54:24,410 | INFO | train_grpo_train | metrics_logged mode=train step=790

                                                      
{'loss': 0.3816, 'grad_norm': 7.537034034729004, 'learning_rate': 7.609090909090909e-06, 'num_tokens': 1424783.0, 'completions/mean_length': 138.25, 'completions/min_length': 24.0, 'completions/max_length': 226.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.25, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 226.0, 'rewards/meter/mean': 0.9772459864616394, 'rewards/meter/std': 0.024871040135622025, 'rewards/count_adherence/mean': 0.6805555820465088, 'rewards/count_adherence/std': 0.3868996798992157, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9041666388511658, 'rewards/count_floor/std': 0.11606989800930023, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.3546792268753052, 'rewards/repeat_penalty/std': 0.39633709192276, 'rewards/repeat_floor/mean': 0.9023541808128357, 'rewards/repeat_floor/std': 0.07556961476802826, 'rewards/near_duplicate_penalty/mean': 0.7567660212516785, 'rewards/near_duplicate_penalty/std': 0.20138370990753174, 'rewards/opening_diversity/mean': 0.8033853769302368, 'rewards/opening_diversity/std': 0.27091559767723083, 'rewards/distinct_2/mean': 0.45151287317276, 'rewards/distinct_2/std': 0.41189688444137573, 'rewards/total_composite/mean': 0.23552927374839783, 'rewards/total_composite/std': 0.10092758387327194, 'reward': 0.23552927374839783, 'reward_std': 0.10092758387327194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11563397943973541, 'sampling/sampling_logp_difference/max': 1.9845504760742188, 'sampling/importance_sampling_ratio/min': 0.13744238018989563, 'sampling/importance_sampling_ratio/mean': 1.0082074403762817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6662660092115402, 'clip_ratio/low_mean': 0.025670488364994526, 'clip_ratio/low_min': 0.025670488364994526, 'clip_ratio/high_mean': 0.09135061781853437, 'clip_ratio/high_max': 0.09135061781853437, 'clip_ratio/region_mean': 0.1170211061835289, 'reward_total_mean': 0.23552927374839783, 'reward_meter_mean': 0.9772459864616394, 'reward_meter_std': 0.024871040135622025, 'reward_count_adherence_mean': 0.6805555820465088, 'reward_count_adherence_std': 0.3868996798992157, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9041666388511658, 'reward_count_floor_std': 0.11606989800930023, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3546792268753052, 'reward_repeat_penalty_std': 0.39633709192276, 'reward_repeat_floor_mean': 0.9023541808128357, 'reward_repeat_floor_std': 0.07556961476802826, 'reward_near_duplicate_penalty_mean': 0.7567660212516785, 'reward_near_duplicate_penalty_std': 0.20138370990753174, 'reward_opening_diversity_mean': 0.8033853769302368, 'reward_opening_diversity_std': 0.27091559767723083, 'reward_distinct_2_mean': 0.45151287317276, 'reward_distinct_2_std': 0.41189688444137573, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.23552927374839783, 'reward_total_composite_std': 0.10092758387327194, 'epoch': 0.04}

 24%|██▍       | 790/3300 [2:55:58<6:28:29,  9.29s/it]INFO 04-12 18:54:24 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 791/3300 [2:56:06<6:19:56,  9.09s/it]2026-04-12 18:54:33,025 | INFO | train_grpo_train | metrics_logged mode=train step=791

                                                      
{'loss': -0.0084, 'grad_norm': 15.763554573059082, 'learning_rate': 7.606060606060606e-06, 'num_tokens': 1426446.0, 'completions/mean_length': 31.875, 'completions/min_length': 18.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9808659553527832, 'rewards/meter/std': 0.01319903414696455, 'rewards/count_adherence/mean': 0.2750000059604645, 'rewards/count_adherence/std': 0.14880476891994476, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7824999690055847, 'rewards/count_floor/std': 0.04464143142104149, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136148393154144, 'rewards/repeat_penalty/mean': 0.6188188791275024, 'rewards/repeat_penalty/std': 0.19888631999492645, 'rewards/repeat_floor/mean': 0.9593135714530945, 'rewards/repeat_floor/std': 0.03195086866617203, 'rewards/near_duplicate_penalty/mean': 0.8738785982131958, 'rewards/near_duplicate_penalty/std': 0.14913766086101532, 'rewards/opening_diversity/mean': 0.7946428656578064, 'rewards/opening_diversity/std': 0.09105392545461655, 'rewards/distinct_2/mean': 0.8950357437133789, 'rewards/distinct_2/std': 0.14770591259002686, 'rewards/total_composite/mean': 0.29286256432533264, 'rewards/total_composite/std': 0.17971602082252502, 'reward': 0.29286256432533264, 'reward_std': 0.17971602082252502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16245712339878082, 'sampling/sampling_logp_difference/max': 1.6938426494598389, 'sampling/importance_sampling_ratio/min': 0.18381184339523315, 'sampling/importance_sampling_ratio/mean': 1.0143431425094604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9917872846126556, 'clip_ratio/low_mean': 0.06146375695243478, 'clip_ratio/low_min': 0.06146375695243478, 'clip_ratio/high_mean': 0.04757246468216181, 'clip_ratio/high_max': 0.04757246468216181, 'clip_ratio/region_mean': 0.10903622163459659, 'reward_total_mean': 0.29286256432533264, 'reward_meter_mean': 0.9808659553527832, 'reward_meter_std': 0.01319903414696455, 'reward_count_adherence_mean': 0.2750000059604645, 'reward_count_adherence_std': 0.14880476891994476, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7824999690055847, 'reward_count_floor_std': 0.04464143142104149, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.6188188791275024, 'reward_repeat_penalty_std': 0.19888631999492645, 'reward_repeat_floor_mean': 0.9593135714530945, 'reward_repeat_floor_std': 0.03195086866617203, 'reward_near_duplicate_penalty_mean': 0.8738785982131958, 'reward_near_duplicate_penalty_std': 0.14913766086101532, 'reward_opening_diversity_mean': 0.7946428656578064, 'reward_opening_diversity_std': 0.09105392545461655, 'reward_distinct_2_mean': 0.8950357437133789, 'reward_distinct_2_std': 0.14770591259002686, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136148393154144, 'reward_total_composite_mean': 0.29286256432533264, 'reward_total_composite_std': 0.17971602082252502, 'epoch': 0.04}

 24%|██▍       | 791/3300 [2:56:06<6:19:56,  9.09s/it]INFO 04-12 18:54:33 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 792/3300 [2:56:15<6:19:06,  9.07s/it]2026-04-12 18:54:42,057 | INFO | train_grpo_train | metrics_logged mode=train step=792

                                                      
{'loss': 0.1101, 'grad_norm': 19.253833770751953, 'learning_rate': 7.603030303030303e-06, 'num_tokens': 1428138.0, 'completions/mean_length': 32.5, 'completions/min_length': 21.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.9065608978271484, 'rewards/meter/std': 0.24304506182670593, 'rewards/count_adherence/mean': 0.34375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8031250238418579, 'rewards/count_floor/std': 0.038816213607788086, 'rewards/lexical_plausibility/mean': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.8103132843971252, 'rewards/repeat_penalty/std': 0.1503400355577469, 'rewards/repeat_floor/mean': 0.9843347072601318, 'rewards/repeat_floor/std': 0.015232930891215801, 'rewards/near_duplicate_penalty/mean': 0.9580647945404053, 'rewards/near_duplicate_penalty/std': 0.0765187069773674, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24058473110198975, 'rewards/total_composite/std': 0.10412845760583878, 'reward': 0.24058473110198975, 'reward_std': 0.10412845760583878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18579772114753723, 'sampling/sampling_logp_difference/max': 1.3998298645019531, 'sampling/importance_sampling_ratio/min': 0.2466389387845993, 'sampling/importance_sampling_ratio/mean': 1.0183417797088623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1508845165371895, 'clip_ratio/low_mean': 0.061420829966664314, 'clip_ratio/low_min': 0.061420829966664314, 'clip_ratio/high_mean': 0.11376503109931946, 'clip_ratio/high_max': 0.11376503109931946, 'clip_ratio/region_mean': 0.17518586106598377, 'reward_total_mean': 0.24058473110198975, 'reward_meter_mean': 0.9065608978271484, 'reward_meter_std': 0.24304506182670593, 'reward_count_adherence_mean': 0.34375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8031250238418579, 'reward_count_floor_std': 0.038816213607788086, 'reward_lexical_plausibility_mean': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.8103132843971252, 'reward_repeat_penalty_std': 0.1503400355577469, 'reward_repeat_floor_mean': 0.9843347072601318, 'reward_repeat_floor_std': 0.015232930891215801, 'reward_near_duplicate_penalty_mean': 0.9580647945404053, 'reward_near_duplicate_penalty_std': 0.0765187069773674, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.24058473110198975, 'reward_total_composite_std': 0.10412845760583878, 'epoch': 0.04}

 24%|██▍       | 792/3300 [2:56:15<6:19:06,  9.07s/it]INFO 04-12 18:54:42 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 793/3300 [2:56:23<6:02:36,  8.68s/it]2026-04-12 18:54:49,823 | INFO | train_grpo_train | metrics_logged mode=train step=793

                                                      
{'loss': -0.0045, 'grad_norm': 33.363319396972656, 'learning_rate': 7.600000000000001e-06, 'num_tokens': 1429496.0, 'completions/mean_length': 17.75, 'completions/min_length': 13.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.75, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5651530027389526, 'rewards/meter/std': 0.41453009843826294, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.953125, 'rewards/lexical_plausibility/std': 0.13258251547813416, 'rewards/judge_quality/mean': 0.8762500286102295, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/repeat_penalty/mean': 0.7601220607757568, 'rewards/repeat_penalty/std': 0.028629617765545845, 'rewards/repeat_floor/mean': 0.9842380285263062, 'rewards/repeat_floor/std': 0.0021551367826759815, 'rewards/near_duplicate_penalty/mean': 0.9892076253890991, 'rewards/near_duplicate_penalty/std': 0.03052542544901371, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.4184347987174988, 'rewards/total_composite/std': 0.3332328498363495, 'reward': 0.4184347987174988, 'reward_std': 0.3332328498363495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15353116393089294, 'sampling/sampling_logp_difference/max': 1.4581780433654785, 'sampling/importance_sampling_ratio/min': 0.3867521584033966, 'sampling/importance_sampling_ratio/mean': 1.0651296377182007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7655248232185841, 'clip_ratio/low_mean': 0.10860514268279076, 'clip_ratio/low_min': 0.10860514268279076, 'clip_ratio/high_mean': 0.04757834831252694, 'clip_ratio/high_max': 0.04757834831252694, 'clip_ratio/region_mean': 0.1561834909953177, 'reward_total_mean': 0.4184347987174988, 'reward_meter_mean': 0.5651530027389526, 'reward_meter_std': 0.41453009843826294, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.953125, 'reward_lexical_plausibility_std': 0.13258251547813416, 'reward_repeat_penalty_mean': 0.7601220607757568, 'reward_repeat_penalty_std': 0.028629617765545845, 'reward_repeat_floor_mean': 0.9842380285263062, 'reward_repeat_floor_std': 0.0021551367826759815, 'reward_near_duplicate_penalty_mean': 0.9892076253890991, 'reward_near_duplicate_penalty_std': 0.03052542544901371, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.8762500286102295, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.4184347987174988, 'reward_total_composite_std': 0.3332328498363495, 'epoch': 0.04}

 24%|██▍       | 793/3300 [2:56:23<6:02:36,  8.68s/it]INFO 04-12 18:54:50 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 794/3300 [2:56:32<5:59:03,  8.60s/it]2026-04-12 18:54:58,229 | INFO | train_grpo_train | metrics_logged mode=train step=794

                                                      
{'loss': -0.0253, 'grad_norm': 23.46124267578125, 'learning_rate': 7.596969696969697e-06, 'num_tokens': 1430999.0, 'completions/mean_length': 19.875, 'completions/min_length': 17.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9850809574127197, 'rewards/meter/std': 0.005409306846559048, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.699721097946167, 'rewards/repeat_penalty/std': 0.09741540998220444, 'rewards/repeat_floor/mean': 0.9753839373588562, 'rewards/repeat_floor/std': 0.018439989537000656, 'rewards/near_duplicate_penalty/mean': 0.9463824033737183, 'rewards/near_duplicate_penalty/std': 0.09973803162574768, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.3930644989013672, 'rewards/total_composite/std': 0.24613673985004425, 'reward': 0.3930644989013672, 'reward_std': 0.24613673985004425, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1470496654510498, 'sampling/sampling_logp_difference/max': 1.793107509613037, 'sampling/importance_sampling_ratio/min': 0.1664421409368515, 'sampling/importance_sampling_ratio/mean': 0.9975727796554565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.788835734128952, 'clip_ratio/low_mean': 0.0747162033803761, 'clip_ratio/low_min': 0.0747162033803761, 'clip_ratio/high_mean': 0.05014822259545326, 'clip_ratio/high_max': 0.05014822259545326, 'clip_ratio/region_mean': 0.12486442597582936, 'reward_total_mean': 0.3930644989013672, 'reward_meter_mean': 0.9850809574127197, 'reward_meter_std': 0.005409306846559048, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.699721097946167, 'reward_repeat_penalty_std': 0.09741540998220444, 'reward_repeat_floor_mean': 0.9753839373588562, 'reward_repeat_floor_std': 0.018439989537000656, 'reward_near_duplicate_penalty_mean': 0.9463824033737183, 'reward_near_duplicate_penalty_std': 0.09973803162574768, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.3930644989013672, 'reward_total_composite_std': 0.24613673985004425, 'epoch': 0.04}

 24%|██▍       | 794/3300 [2:56:32<5:59:03,  8.60s/it]INFO 04-12 18:54:58 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 795/3300 [2:56:39<5:43:43,  8.23s/it]2026-04-12 18:55:05,613 | INFO | train_grpo_train | metrics_logged mode=train step=795

                                                      
{'loss': -0.013, 'grad_norm': 18.708932876586914, 'learning_rate': 7.593939393939395e-06, 'num_tokens': 1432507.0, 'completions/mean_length': 24.5, 'completions/min_length': 18.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.5848500728607178, 'rewards/meter/std': 0.29298242926597595, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9312499761581421, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.8162202835083008, 'rewards/repeat_penalty/std': 0.10509595274925232, 'rewards/repeat_floor/mean': 0.9873518943786621, 'rewards/repeat_floor/std': 0.006554624531418085, 'rewards/near_duplicate_penalty/mean': 0.988213062286377, 'rewards/near_duplicate_penalty/std': 0.021667474880814552, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.46388813853263855, 'rewards/total_composite/std': 0.22653377056121826, 'reward': 0.46388813853263855, 'reward_std': 0.22653377056121826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15578028559684753, 'sampling/sampling_logp_difference/max': 2.1698503494262695, 'sampling/importance_sampling_ratio/min': 0.11419470608234406, 'sampling/importance_sampling_ratio/mean': 1.0178667306900024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8114653304219246, 'clip_ratio/low_mean': 0.06194380950182676, 'clip_ratio/low_min': 0.06194380950182676, 'clip_ratio/high_mean': 0.05900621181353927, 'clip_ratio/high_max': 0.05900621181353927, 'clip_ratio/region_mean': 0.12095002131536603, 'reward_total_mean': 0.46388813853263855, 'reward_meter_mean': 0.5848500728607178, 'reward_meter_std': 0.29298242926597595, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8162202835083008, 'reward_repeat_penalty_std': 0.10509595274925232, 'reward_repeat_floor_mean': 0.9873518943786621, 'reward_repeat_floor_std': 0.006554624531418085, 'reward_near_duplicate_penalty_mean': 0.988213062286377, 'reward_near_duplicate_penalty_std': 0.021667474880814552, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.9312499761581421, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.46388813853263855, 'reward_total_composite_std': 0.22653377056121826, 'epoch': 0.04}

 24%|██▍       | 795/3300 [2:56:39<5:43:43,  8.23s/it]INFO 04-12 18:55:05 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 796/3300 [2:56:48<5:53:17,  8.47s/it]2026-04-12 18:55:14,625 | INFO | train_grpo_train | metrics_logged mode=train step=796

                                                      
{'loss': 0.0494, 'grad_norm': 16.400909423828125, 'learning_rate': 7.590909090909091e-06, 'num_tokens': 1434292.0, 'completions/mean_length': 36.125, 'completions/min_length': 33.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7702640891075134, 'rewards/meter/std': 0.3129320740699768, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9930901527404785, 'rewards/repeat_penalty/std': 0.008361689746379852, 'rewards/repeat_floor/mean': 0.9989635348320007, 'rewards/repeat_floor/std': 0.0012542513431981206, 'rewards/near_duplicate_penalty/mean': 0.9930901527404785, 'rewards/near_duplicate_penalty/std': 0.008361689746379852, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27638447284698486, 'rewards/total_composite/std': 0.1023259311914444, 'reward': 0.27638447284698486, 'reward_std': 0.102325938642025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2024848759174347, 'sampling/sampling_logp_difference/max': 2.381497621536255, 'sampling/importance_sampling_ratio/min': 0.0924120768904686, 'sampling/importance_sampling_ratio/mean': 0.9887649416923523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2036852315068245, 'clip_ratio/low_mean': 0.04448198340833187, 'clip_ratio/low_min': 0.04448198340833187, 'clip_ratio/high_mean': 0.1774127334356308, 'clip_ratio/high_max': 0.1774127334356308, 'clip_ratio/region_mean': 0.22189471684396267, 'reward_total_mean': 0.27638447284698486, 'reward_meter_mean': 0.7702640891075134, 'reward_meter_std': 0.3129320740699768, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9930901527404785, 'reward_repeat_penalty_std': 0.008361689746379852, 'reward_repeat_floor_mean': 0.9989635348320007, 'reward_repeat_floor_std': 0.0012542513431981206, 'reward_near_duplicate_penalty_mean': 0.9930901527404785, 'reward_near_duplicate_penalty_std': 0.008361689746379852, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.27638447284698486, 'reward_total_composite_std': 0.1023259311914444, 'epoch': 0.04}

 24%|██▍       | 796/3300 [2:56:48<5:53:17,  8.47s/it]INFO 04-12 18:55:14 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 797/3300 [2:56:57<5:55:03,  8.51s/it]2026-04-12 18:55:23,239 | INFO | train_grpo_train | metrics_logged mode=train step=797

                                                      
{'loss': -0.1661, 'grad_norm': 18.46236228942871, 'learning_rate': 7.587878787878788e-06, 'num_tokens': 1436009.0, 'completions/mean_length': 42.625, 'completions/min_length': 19.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.6396142840385437, 'rewards/meter/std': 0.24471114575862885, 'rewards/count_adherence/mean': 0.3958333432674408, 'rewards/count_adherence/std': 0.19795581698417664, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8187500238418579, 'rewards/count_floor/std': 0.05938674136996269, 'rewards/lexical_plausibility/mean': 0.9418402910232544, 'rewards/lexical_plausibility/std': 0.12996836006641388, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2728912830352783, 'rewards/repeat_penalty/mean': 0.8997710943222046, 'rewards/repeat_penalty/std': 0.12221883237361908, 'rewards/repeat_floor/mean': 0.9906902313232422, 'rewards/repeat_floor/std': 0.009867928922176361, 'rewards/near_duplicate_penalty/mean': 0.9690076112747192, 'rewards/near_duplicate_penalty/std': 0.03308709338307381, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.18128123879432678, 'rewards/total_composite/std': 0.19653424620628357, 'reward': 0.18128123879432678, 'reward_std': 0.19653426110744476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17326223850250244, 'sampling/sampling_logp_difference/max': 1.9223346710205078, 'sampling/importance_sampling_ratio/min': 0.14626507461071014, 'sampling/importance_sampling_ratio/mean': 1.0231902599334717, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7868901789188385, 'clip_ratio/low_mean': 0.13473488576710224, 'clip_ratio/low_min': 0.13473488576710224, 'clip_ratio/high_mean': 0.04135572165250778, 'clip_ratio/high_max': 0.04135572165250778, 'clip_ratio/region_mean': 0.17609060741961002, 'reward_total_mean': 0.18128123879432678, 'reward_meter_mean': 0.6396142840385437, 'reward_meter_std': 0.24471114575862885, 'reward_count_adherence_mean': 0.3958333432674408, 'reward_count_adherence_std': 0.19795581698417664, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8187500238418579, 'reward_count_floor_std': 0.05938674136996269, 'reward_lexical_plausibility_mean': 0.9418402910232544, 'reward_lexical_plausibility_std': 0.12996836006641388, 'reward_repeat_penalty_mean': 0.8997710943222046, 'reward_repeat_penalty_std': 0.12221883237361908, 'reward_repeat_floor_mean': 0.9906902313232422, 'reward_repeat_floor_std': 0.009867928922176361, 'reward_near_duplicate_penalty_mean': 0.9690076112747192, 'reward_near_duplicate_penalty_std': 0.03308709338307381, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2728912830352783, 'reward_total_composite_mean': 0.18128123879432678, 'reward_total_composite_std': 0.19653424620628357, 'epoch': 0.04}

 24%|██▍       | 797/3300 [2:56:57<5:55:03,  8.51s/it]INFO 04-12 18:55:23 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 798/3300 [2:57:05<5:48:18,  8.35s/it]2026-04-12 18:55:31,223 | INFO | train_grpo_train | metrics_logged mode=train step=798

                                                      
{'loss': -0.0038, 'grad_norm': 20.410425186157227, 'learning_rate': 7.584848484848486e-06, 'num_tokens': 1437513.0, 'completions/mean_length': 25.0, 'completions/min_length': 21.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7125061750411987, 'rewards/meter/std': 0.40914562344551086, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.749517560005188, 'rewards/repeat_penalty/std': 0.001364483148790896, 'rewards/repeat_floor/mean': 0.9849035143852234, 'rewards/repeat_floor/std': 0.00027290082653053105, 'rewards/near_duplicate_penalty/mean': 0.999356746673584, 'rewards/near_duplicate_penalty/std': 0.0018193108262494206, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4167323112487793, 'rewards/total_composite/std': 0.2952266037464142, 'reward': 0.4167323112487793, 'reward_std': 0.2952266037464142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15916723012924194, 'sampling/sampling_logp_difference/max': 1.1265678405761719, 'sampling/importance_sampling_ratio/min': 0.3241438567638397, 'sampling/importance_sampling_ratio/mean': 1.0218888521194458, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0149695873260498, 'clip_ratio/low_mean': 0.0910462555475533, 'clip_ratio/low_min': 0.0910462555475533, 'clip_ratio/high_mean': 0.03993055643513799, 'clip_ratio/high_max': 0.03993055643513799, 'clip_ratio/region_mean': 0.1309768119826913, 'reward_total_mean': 0.4167323112487793, 'reward_meter_mean': 0.7125061750411987, 'reward_meter_std': 0.40914562344551086, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.749517560005188, 'reward_repeat_penalty_std': 0.001364483148790896, 'reward_repeat_floor_mean': 0.9849035143852234, 'reward_repeat_floor_std': 0.00027290082653053105, 'reward_near_duplicate_penalty_mean': 0.999356746673584, 'reward_near_duplicate_penalty_std': 0.0018193108262494206, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.4167323112487793, 'reward_total_composite_std': 0.2952266037464142, 'epoch': 0.04}

 24%|██▍       | 798/3300 [2:57:05<5:48:18,  8.35s/it]INFO 04-12 18:55:31 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 799/3300 [2:57:12<5:38:20,  8.12s/it]2026-04-12 18:55:38,792 | INFO | train_grpo_train | metrics_logged mode=train step=799

                                                      
{'loss': 0.109, 'grad_norm': 17.887121200561523, 'learning_rate': 7.581818181818183e-06, 'num_tokens': 1438926.0, 'completions/mean_length': 20.625, 'completions/min_length': 17.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8786775469779968, 'rewards/meter/std': 0.27019941806793213, 'rewards/count_adherence/mean': 0.3333333432674408, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.800000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3236731290817261, 'rewards/repeat_penalty/mean': 0.7480332255363464, 'rewards/repeat_penalty/std': 0.005562899634242058, 'rewards/repeat_floor/mean': 0.9846066236495972, 'rewards/repeat_floor/std': 0.0011125925229862332, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25092214345932007, 'rewards/total_composite/std': 0.19736409187316895, 'reward': 0.25092214345932007, 'reward_std': 0.19736410677433014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14094756543636322, 'sampling/sampling_logp_difference/max': 1.2471961975097656, 'sampling/importance_sampling_ratio/min': 0.2873092293739319, 'sampling/importance_sampling_ratio/mean': 1.0189201831817627, 'sampling/importance_sampling_ratio/max': 1.57198166847229, 'entropy': 1.0235275998711586, 'clip_ratio/low_mean': 0.10895504243671894, 'clip_ratio/low_min': 0.10895504243671894, 'clip_ratio/high_mean': 0.07022058963775635, 'clip_ratio/high_max': 0.07022058963775635, 'clip_ratio/region_mean': 0.1791756320744753, 'reward_total_mean': 0.25092214345932007, 'reward_meter_mean': 0.8786775469779968, 'reward_meter_std': 0.27019941806793213, 'reward_count_adherence_mean': 0.3333333432674408, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.800000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7480332255363464, 'reward_repeat_penalty_std': 0.005562899634242058, 'reward_repeat_floor_mean': 0.9846066236495972, 'reward_repeat_floor_std': 0.0011125925229862332, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3236731290817261, 'reward_total_composite_mean': 0.25092214345932007, 'reward_total_composite_std': 0.19736409187316895, 'epoch': 0.04}

 24%|██▍       | 799/3300 [2:57:12<5:38:20,  8.12s/it]INFO 04-12 18:55:39 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 800/3300 [2:57:20<5:41:30,  8.20s/it]2026-04-12 18:55:47,170 | INFO | train_grpo_train | metrics_logged mode=train step=800

                                                      
{'loss': 0.0704, 'grad_norm': 24.220542907714844, 'learning_rate': 7.57878787878788e-06, 'num_tokens': 1440451.0, 'completions/mean_length': 23.625, 'completions/min_length': 20.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9771468639373779, 'rewards/meter/std': 0.034696560353040695, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.7439630627632141, 'rewards/repeat_penalty/std': 0.011203979142010212, 'rewards/repeat_floor/mean': 0.9837926626205444, 'rewards/repeat_floor/std': 0.0022408016957342625, 'rewards/near_duplicate_penalty/mean': 0.9919507503509521, 'rewards/near_duplicate_penalty/std': 0.014938644133508205, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29844123125076294, 'rewards/total_composite/std': 0.0860796570777893, 'reward': 0.29844123125076294, 'reward_std': 0.0860796645283699, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16077981889247894, 'sampling/sampling_logp_difference/max': 1.5124053955078125, 'sampling/importance_sampling_ratio/min': 0.22037924826145172, 'sampling/importance_sampling_ratio/mean': 1.0170494318008423, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3279575258493423, 'clip_ratio/low_mean': 0.07468993868678808, 'clip_ratio/low_min': 0.07468993868678808, 'clip_ratio/high_mean': 0.046363635919988155, 'clip_ratio/high_max': 0.046363635919988155, 'clip_ratio/region_mean': 0.12105357460677624, 'reward_total_mean': 0.29844123125076294, 'reward_meter_mean': 0.9771468639373779, 'reward_meter_std': 0.034696560353040695, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.7439630627632141, 'reward_repeat_penalty_std': 0.011203979142010212, 'reward_repeat_floor_mean': 0.9837926626205444, 'reward_repeat_floor_std': 0.0022408016957342625, 'reward_near_duplicate_penalty_mean': 0.9919507503509521, 'reward_near_duplicate_penalty_std': 0.014938644133508205, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.29844123125076294, 'reward_total_composite_std': 0.0860796570777893, 'epoch': 0.04}

 24%|██▍       | 800/3300 [2:57:21<5:41:30,  8.20s/it]INFO 04-12 18:55:47 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:12<01:06,  6.01s/it]

 23%|██▎       | 3/13 [00:17<00:58,  5.85s/it]

 31%|███       | 4/13 [00:24<00:56,  6.23s/it]

 38%|███▊      | 5/13 [00:32<00:55,  6.93s/it]

 46%|████▌     | 6/13 [00:39<00:48,  6.90s/it]

 54%|█████▍    | 7/13 [00:47<00:43,  7.25s/it]

 62%|██████▏   | 8/13 [00:59<00:42,  8.56s/it]

 69%|██████▉   | 9/13 [01:08<00:34,  8.74s/it]

 77%|███████▋  | 10/13 [01:15<00:24,  8.30s/it]

 85%|████████▍ | 11/13 [01:24<00:16,  8.46s/it]

 92%|█████████▏| 12/13 [01:32<00:08,  8.36s/it]

100%|██████████| 13/13 [01:41<00:00,  8.45s/it]2026-04-12 18:57:37,936 | INFO | train_grpo_train | metrics_logged mode=eval step=800

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 110.6364, 'eval_samples_per_second': 0.94, 'eval_steps_per_second': 0.118, 'eval_num_tokens': 1440451.0, 'eval_completions/mean_length': 100.07692307692308, 'eval_completions/min_length': 18.46153846153846, 'eval_completions/max_length': 280.0769230769231, 'eval_completions/clipped_ratio': 0.019230769230769232, 'eval_completions/mean_terminated_length': 92.75137446476863, 'eval_completions/min_terminated_length': 18.46153846153846, 'eval_completions/max_terminated_length': 252.0, 'eval_rewards/meter/mean': 0.6896416499064519, 'eval_rewards/meter/std': 0.3716986041802626, 'eval_rewards/count_adherence/mean': 0.6200550198554993, 'eval_rewards/count_adherence/std': 0.30198927452931035, 'eval_rewards/arabic_clean/mean': 0.9423076923076923, 'eval_rewards/arabic_clean/std': 0.12140072079805228, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 1.0, 'eval_rewards/arabic_floor/std': 0.0, 'eval_rewards/count_floor/mean': 0.886016492660229, 'eval_rewards/count_floor/std': 0.09059677788844475, 'eval_rewards/lexical_plausibility/mean': 0.989690349652217, 'eval_rewards/lexical_plausibility/std': 0.02916010025028999, 'eval_rewards/judge_quality/mean': 0.3920192260008592, 'eval_rewards/judge_quality/std': 0.19813704719910255, 'eval_rewards/repeat_penalty/mean': 0.6619293185380789, 'eval_rewards/repeat_penalty/std': 0.25210404797242236, 'eval_rewards/repeat_floor/mean': 0.9658236366051894, 'eval_rewards/repeat_floor/std': 0.031848148406984716, 'eval_rewards/near_duplicate_penalty/mean': 0.9338104174687312, 'eval_rewards/near_duplicate_penalty/std': 0.07629959844052792, 'eval_rewards/opening_diversity/mean': 0.883915364742279, 'eval_rewards/opening_diversity/std': 0.13033672307546323, 'eval_rewards/distinct_2/mean': 0.8079682359328637, 'eval_rewards/distinct_2/std': 0.2624016593282039, 'eval_rewards/total_composite/mean': 0.2375198000898728, 'eval_rewards/total_composite/std': 0.19253497742689574, 'eval_reward': 0.2375198000898728, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07209557724686769, 'eval_sampling/sampling_logp_difference/max': 1.008132164294903, 'eval_sampling/importance_sampling_ratio/min': 0.37383420077654034, 'eval_sampling/importance_sampling_ratio/mean': 1.0177982128583467, 'eval_sampling/importance_sampling_ratio/max': 1.4980343763644879, 'eval_entropy': 0.8217866466595576, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.2375198000898728, 'eval_reward_meter_mean': 0.6896416499064519, 'eval_reward_meter_std': 0.3716986041802626, 'eval_reward_count_adherence_mean': 0.6200550198554993, 'eval_reward_count_adherence_std': 0.30198927452931035, 'eval_reward_arabic_clean_mean': 0.9423076923076923, 'eval_reward_arabic_clean_std': 0.12140072079805228, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 1.0, 'eval_reward_arabic_floor_std': 0.0, 'eval_reward_count_floor_mean': 0.886016492660229, 'eval_reward_count_floor_std': 0.09059677788844475, 'eval_reward_lexical_plausibility_mean': 0.989690349652217, 'eval_reward_lexical_plausibility_std': 0.02916010025028999, 'eval_reward_repeat_penalty_mean': 0.6619293185380789, 'eval_reward_repeat_penalty_std': 0.25210404797242236, 'eval_reward_repeat_floor_mean': 0.9658236366051894, 'eval_reward_repeat_floor_std': 0.031848148406984716, 'eval_reward_near_duplicate_penalty_mean': 0.9338104174687312, 'eval_reward_near_duplicate_penalty_std': 0.07629959844052792, 'eval_reward_opening_diversity_mean': 0.883915364742279, 'eval_reward_opening_diversity_std': 0.13033672307546323, 'eval_reward_distinct_2_mean': 0.8079682359328637, 'eval_reward_distinct_2_std': 0.2624016593282039, 'eval_reward_judge_quality_mean': 0.3920192260008592, 'eval_reward_judge_quality_std': 0.19813704719910255, 'eval_reward_total_composite_mean': 0.2375198000898728, 'eval_reward_total_composite_std': 0.19253497742689574, 'epoch': 0.04}

 24%|██▍       | 800/3300 [2:59:11<5:41:30,  8.20s/it]

100%|██████████| 13/13 [01:41<00:00,  8.45s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 24%|██▍       | 801/3300 [2:59:24<29:40:30, 42.75s/it]2026-04-12 18:57:50,544 | INFO | train_grpo_train | metrics_logged mode=train step=801

                                                       
{'loss': 0.0564, 'grad_norm': 11.878886222839355, 'learning_rate': 7.5757575757575764e-06, 'num_tokens': 1442159.0, 'completions/mean_length': 36.5, 'completions/min_length': 22.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9927767515182495, 'rewards/meter/std': 0.005641669034957886, 'rewards/count_adherence/mean': 0.34375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8031250238418579, 'rewards/count_floor/std': 0.038816213607788086, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.7536890506744385, 'rewards/repeat_penalty/std': 0.16974732279777527, 'rewards/repeat_floor/mean': 0.9775967597961426, 'rewards/repeat_floor/std': 0.021494954824447632, 'rewards/near_duplicate_penalty/mean': 0.9495554566383362, 'rewards/near_duplicate_penalty/std': 0.07761404663324356, 'rewards/opening_diversity/mean': 0.831250011920929, 'rewards/opening_diversity/std': 0.11630471050739288, 'rewards/distinct_2/mean': 0.9476495981216431, 'rewards/distinct_2/std': 0.09696105867624283, 'rewards/total_composite/mean': 0.29813772439956665, 'rewards/total_composite/std': 0.18174539506435394, 'reward': 0.29813772439956665, 'reward_std': 0.18174538016319275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1707857996225357, 'sampling/sampling_logp_difference/max': 1.5292596817016602, 'sampling/importance_sampling_ratio/min': 0.21669605374336243, 'sampling/importance_sampling_ratio/mean': 1.0212535858154297, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1316742971539497, 'clip_ratio/low_mean': 0.15517114102840424, 'clip_ratio/low_min': 0.15517114102840424, 'clip_ratio/high_mean': 0.052759742364287376, 'clip_ratio/high_max': 0.052759742364287376, 'clip_ratio/region_mean': 0.2079308833926916, 'reward_total_mean': 0.29813772439956665, 'reward_meter_mean': 0.9927767515182495, 'reward_meter_std': 0.005641669034957886, 'reward_count_adherence_mean': 0.34375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8031250238418579, 'reward_count_floor_std': 0.038816213607788086, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7536890506744385, 'reward_repeat_penalty_std': 0.16974732279777527, 'reward_repeat_floor_mean': 0.9775967597961426, 'reward_repeat_floor_std': 0.021494954824447632, 'reward_near_duplicate_penalty_mean': 0.9495554566383362, 'reward_near_duplicate_penalty_std': 0.07761404663324356, 'reward_opening_diversity_mean': 0.831250011920929, 'reward_opening_diversity_std': 0.11630471050739288, 'reward_distinct_2_mean': 0.9476495981216431, 'reward_distinct_2_std': 0.09696105867624283, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.29813772439956665, 'reward_total_composite_std': 0.18174539506435394, 'epoch': 0.04}

 24%|██▍       | 801/3300 [2:59:24<29:40:30, 42.75s/it]INFO 04-12 18:57:50 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 802/3300 [2:59:31<22:20:51, 32.21s/it]2026-04-12 18:57:58,151 | INFO | train_grpo_train | metrics_logged mode=train step=802

                                                       
{'loss': -0.0634, 'grad_norm': 23.30038070678711, 'learning_rate': 7.572727272727274e-06, 'num_tokens': 1443757.0, 'completions/mean_length': 20.75, 'completions/min_length': 11.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 11.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.3598850965499878, 'rewards/meter/std': 0.4232175648212433, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.7764204740524292, 'rewards/repeat_penalty/std': 0.14579880237579346, 'rewards/repeat_floor/mean': 0.9815341234207153, 'rewards/repeat_floor/std': 0.014780850149691105, 'rewards/near_duplicate_penalty/mean': 0.9518939256668091, 'rewards/near_duplicate_penalty/std': 0.06941288709640503, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10499314963817596, 'rewards/total_composite/std': 0.12517425417900085, 'reward': 0.10499314963817596, 'reward_std': 0.12517425417900085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16747726500034332, 'sampling/sampling_logp_difference/max': 1.394655704498291, 'sampling/importance_sampling_ratio/min': 0.24791836738586426, 'sampling/importance_sampling_ratio/mean': 1.0148138999938965, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0280520394444466, 'clip_ratio/low_mean': 0.0721969697624445, 'clip_ratio/low_min': 0.0721969697624445, 'clip_ratio/high_mean': 0.06514550279825926, 'clip_ratio/high_max': 0.06514550279825926, 'clip_ratio/region_mean': 0.13734247256070375, 'reward_total_mean': 0.10499314963817596, 'reward_meter_mean': 0.3598850965499878, 'reward_meter_std': 0.4232175648212433, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7764204740524292, 'reward_repeat_penalty_std': 0.14579880237579346, 'reward_repeat_floor_mean': 0.9815341234207153, 'reward_repeat_floor_std': 0.014780850149691105, 'reward_near_duplicate_penalty_mean': 0.9518939256668091, 'reward_near_duplicate_penalty_std': 0.06941288709640503, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.10499314963817596, 'reward_total_composite_std': 0.12517425417900085, 'epoch': 0.04}

 24%|██▍       | 802/3300 [2:59:32<22:20:51, 32.21s/it]INFO 04-12 18:57:58 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 803/3300 [2:59:39<17:11:14, 24.78s/it]2026-04-12 18:58:05,600 | INFO | train_grpo_train | metrics_logged mode=train step=803

                                                       
{'loss': 0.1123, 'grad_norm': 23.425832748413086, 'learning_rate': 7.56969696969697e-06, 'num_tokens': 1445090.0, 'completions/mean_length': 21.625, 'completions/min_length': 18.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.8531711101531982, 'rewards/meter/std': 0.342794805765152, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7975000143051147, 'rewards/judge_quality/std': 0.2471408098936081, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6548610329627991, 'rewards/total_composite/std': 0.34972184896469116, 'reward': 0.6548610329627991, 'reward_std': 0.3497218191623688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12653936445713043, 'sampling/sampling_logp_difference/max': 0.9467382431030273, 'sampling/importance_sampling_ratio/min': 0.3880045413970947, 'sampling/importance_sampling_ratio/mean': 1.0381673574447632, 'sampling/importance_sampling_ratio/max': 1.8500672578811646, 'entropy': 1.0275596976280212, 'clip_ratio/low_mean': 0.04130434803664684, 'clip_ratio/low_min': 0.04130434803664684, 'clip_ratio/high_mean': 0.05765319149941206, 'clip_ratio/high_max': 0.05765319149941206, 'clip_ratio/region_mean': 0.0989575395360589, 'reward_total_mean': 0.6548610329627991, 'reward_meter_mean': 0.8531711101531982, 'reward_meter_std': 0.342794805765152, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7975000143051147, 'reward_judge_quality_std': 0.2471408098936081, 'reward_total_composite_mean': 0.6548610329627991, 'reward_total_composite_std': 0.34972184896469116, 'epoch': 0.04}

 24%|██▍       | 803/3300 [2:59:39<17:11:14, 24.78s/it]INFO 04-12 18:58:05 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 804/3300 [2:59:52<14:45:57, 21.30s/it]2026-04-12 18:58:18,775 | INFO | train_grpo_train | metrics_logged mode=train step=804

                                                       
{'loss': -0.014, 'grad_norm': 4.55588436126709, 'learning_rate': 7.566666666666667e-06, 'num_tokens': 1446354.0, 'completions/mean_length': 79.0, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 17.142858505249023, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.6146837472915649, 'rewards/meter/std': 0.45543384552001953, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.951554000377655, 'rewards/lexical_plausibility/std': 0.13702602684497833, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.35379326343536377, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32480186223983765, 'rewards/total_composite/std': 0.3363499343395233, 'reward': 0.32480186223983765, 'reward_std': 0.3363499343395233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17231829464435577, 'sampling/sampling_logp_difference/max': 0.9110507965087891, 'sampling/importance_sampling_ratio/min': 0.4021014869213104, 'sampling/importance_sampling_ratio/mean': 1.05978524684906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9506350383162498, 'clip_ratio/low_mean': 0.08088235463947058, 'clip_ratio/low_min': 0.08088235463947058, 'clip_ratio/high_mean': 0.07962461328133941, 'clip_ratio/high_max': 0.07962461328133941, 'clip_ratio/region_mean': 0.16050696792080998, 'reward_total_mean': 0.32480186223983765, 'reward_meter_mean': 0.6146837472915649, 'reward_meter_std': 0.45543384552001953, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.951554000377655, 'reward_lexical_plausibility_std': 0.13702602684497833, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.35379326343536377, 'reward_total_composite_mean': 0.32480186223983765, 'reward_total_composite_std': 0.3363499343395233, 'epoch': 0.04}

 24%|██▍       | 804/3300 [2:59:52<14:45:57, 21.30s/it]INFO 04-12 18:58:19 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 805/3300 [3:00:00<11:57:01, 17.24s/it]2026-04-12 18:58:26,554 | INFO | train_grpo_train | metrics_logged mode=train step=805

                                                       
{'loss': 0.003, 'grad_norm': 19.495641708374023, 'learning_rate': 7.563636363636364e-06, 'num_tokens': 1447637.0, 'completions/mean_length': 19.375, 'completions/min_length': 15.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.375, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9700566530227661, 'rewards/meter/std': 0.04708840325474739, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5387500524520874, 'rewards/judge_quality/std': 0.3358757197856903, 'rewards/repeat_penalty/mean': 0.7480332255363464, 'rewards/repeat_penalty/std': 0.005562899634242058, 'rewards/repeat_floor/mean': 0.9846066236495972, 'rewards/repeat_floor/std': 0.0011125925229862332, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4391022324562073, 'rewards/total_composite/std': 0.2897334098815918, 'reward': 0.4391022324562073, 'reward_std': 0.2897334098815918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11404381692409515, 'sampling/sampling_logp_difference/max': 0.9312906265258789, 'sampling/importance_sampling_ratio/min': 0.3940448462963104, 'sampling/importance_sampling_ratio/mean': 1.0381765365600586, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7706129997968674, 'clip_ratio/low_mean': 0.09775410033762455, 'clip_ratio/low_min': 0.09775410033762455, 'clip_ratio/high_mean': 0.05393518693745136, 'clip_ratio/high_max': 0.05393518693745136, 'clip_ratio/region_mean': 0.1516892872750759, 'reward_total_mean': 0.4391022324562073, 'reward_meter_mean': 0.9700566530227661, 'reward_meter_std': 0.04708840325474739, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7480332255363464, 'reward_repeat_penalty_std': 0.005562899634242058, 'reward_repeat_floor_mean': 0.9846066236495972, 'reward_repeat_floor_std': 0.0011125925229862332, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5387500524520874, 'reward_judge_quality_std': 0.3358757197856903, 'reward_total_composite_mean': 0.4391022324562073, 'reward_total_composite_std': 0.2897334098815918, 'epoch': 0.04}

 24%|██▍       | 805/3300 [3:00:00<11:57:01, 17.24s/it]INFO 04-12 18:58:26 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 806/3300 [3:00:08<10:02:37, 14.50s/it]2026-04-12 18:58:34,699 | INFO | train_grpo_train | metrics_logged mode=train step=806

                                                       
{'loss': -0.0149, 'grad_norm': 18.91968536376953, 'learning_rate': 7.560606060606062e-06, 'num_tokens': 1449163.0, 'completions/mean_length': 26.75, 'completions/min_length': 23.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.545417070388794, 'rewards/meter/std': 0.29303255677223206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9387500286102295, 'rewards/judge_quality/std': 0.015526460483670235, 'rewards/repeat_penalty/mean': 0.9581937789916992, 'rewards/repeat_penalty/std': 0.025177184492349625, 'rewards/repeat_floor/mean': 0.9937291145324707, 'rewards/repeat_floor/std': 0.003776577301323414, 'rewards/near_duplicate_penalty/mean': 0.9581937789916992, 'rewards/near_duplicate_penalty/std': 0.025177184492349625, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5101175308227539, 'rewards/total_composite/std': 0.27494683861732483, 'reward': 0.5101175308227539, 'reward_std': 0.27494683861732483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12949466705322266, 'sampling/sampling_logp_difference/max': 1.5344762802124023, 'sampling/importance_sampling_ratio/min': 0.21556855738162994, 'sampling/importance_sampling_ratio/mean': 0.985395610332489, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5876073054969311, 'clip_ratio/low_mean': 0.04502249648794532, 'clip_ratio/low_min': 0.04502249648794532, 'clip_ratio/high_mean': 0.04750000033527613, 'clip_ratio/high_max': 0.04750000033527613, 'clip_ratio/region_mean': 0.09252249682322145, 'reward_total_mean': 0.5101175308227539, 'reward_meter_mean': 0.545417070388794, 'reward_meter_std': 0.29303255677223206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9581937789916992, 'reward_repeat_penalty_std': 0.025177184492349625, 'reward_repeat_floor_mean': 0.9937291145324707, 'reward_repeat_floor_std': 0.003776577301323414, 'reward_near_duplicate_penalty_mean': 0.9581937789916992, 'reward_near_duplicate_penalty_std': 0.025177184492349625, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9387500286102295, 'reward_judge_quality_std': 0.015526460483670235, 'reward_total_composite_mean': 0.5101175308227539, 'reward_total_composite_std': 0.27494683861732483, 'epoch': 0.04}

 24%|██▍       | 806/3300 [3:00:08<10:02:37, 14.50s/it]INFO 04-12 18:58:35 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 807/3300 [3:00:17<8:57:37, 12.94s/it] 2026-04-12 18:58:43,951 | INFO | train_grpo_train | metrics_logged mode=train step=807

                                                      
{'loss': 0.3515, 'grad_norm': 19.447965621948242, 'learning_rate': 7.557575757575758e-06, 'num_tokens': 1451024.0, 'completions/mean_length': 34.625, 'completions/min_length': 17.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5784199833869934, 'rewards/meter/std': 0.4045107960700989, 'rewards/count_adherence/mean': 0.4375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.831250011920929, 'rewards/count_floor/std': 0.05303302779793739, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.2707628309726715, 'rewards/repeat_penalty/mean': 0.877953290939331, 'rewards/repeat_penalty/std': 0.14206625521183014, 'rewards/repeat_floor/mean': 0.9894434809684753, 'rewards/repeat_floor/std': 0.012300090864300728, 'rewards/near_duplicate_penalty/mean': 0.9671230316162109, 'rewards/near_duplicate_penalty/std': 0.051729150116443634, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3764670491218567, 'rewards/total_composite/std': 0.3131535053253174, 'reward': 0.3764670491218567, 'reward_std': 0.31315353512763977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17660307884216309, 'sampling/sampling_logp_difference/max': 1.5156049728393555, 'sampling/importance_sampling_ratio/min': 0.21967525780200958, 'sampling/importance_sampling_ratio/mean': 1.0280940532684326, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.951075442135334, 'clip_ratio/low_mean': 0.10332360491156578, 'clip_ratio/low_min': 0.10332360491156578, 'clip_ratio/high_mean': 0.06471755588427186, 'clip_ratio/high_max': 0.06471755588427186, 'clip_ratio/region_mean': 0.16804116079583764, 'reward_total_mean': 0.3764670491218567, 'reward_meter_mean': 0.5784199833869934, 'reward_meter_std': 0.4045107960700989, 'reward_count_adherence_mean': 0.4375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.831250011920929, 'reward_count_floor_std': 0.05303302779793739, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.877953290939331, 'reward_repeat_penalty_std': 0.14206625521183014, 'reward_repeat_floor_mean': 0.9894434809684753, 'reward_repeat_floor_std': 0.012300090864300728, 'reward_near_duplicate_penalty_mean': 0.9671230316162109, 'reward_near_duplicate_penalty_std': 0.051729150116443634, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.2707628309726715, 'reward_total_composite_mean': 0.3764670491218567, 'reward_total_composite_std': 0.3131535053253174, 'epoch': 0.04}

 24%|██▍       | 807/3300 [3:00:17<8:57:37, 12.94s/it]INFO 04-12 18:58:44 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 808/3300 [3:00:25<7:49:54, 11.31s/it]2026-04-12 18:58:51,472 | INFO | train_grpo_train | metrics_logged mode=train step=808

                                                      
{'loss': 0.0823, 'grad_norm': 19.021160125732422, 'learning_rate': 7.5545454545454555e-06, 'num_tokens': 1452329.0, 'completions/mean_length': 22.125, 'completions/min_length': 16.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.125, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.4860062897205353, 'rewards/meter/std': 0.4817962050437927, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8350000381469727, 'rewards/judge_quality/std': 0.27717968821525574, 'rewards/repeat_penalty/mean': 0.6842336654663086, 'rewards/repeat_penalty/std': 0.09754367172718048, 'rewards/repeat_floor/mean': 0.9721978902816772, 'rewards/repeat_floor/std': 0.01882293075323105, 'rewards/near_duplicate_penalty/mean': 0.9228942394256592, 'rewards/near_duplicate_penalty/std': 0.11104470491409302, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.42997312545776367, 'rewards/total_composite/std': 0.44086191058158875, 'reward': 0.42997312545776367, 'reward_std': 0.44086188077926636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17718452215194702, 'sampling/sampling_logp_difference/max': 2.492250442504883, 'sampling/importance_sampling_ratio/min': 0.08272358775138855, 'sampling/importance_sampling_ratio/mean': 1.0032590627670288, 'sampling/importance_sampling_ratio/max': 1.846886396408081, 'entropy': 1.052128829061985, 'clip_ratio/low_mean': 0.07777014747262001, 'clip_ratio/low_min': 0.07777014747262001, 'clip_ratio/high_mean': 0.07617374509572983, 'clip_ratio/high_max': 0.07617374509572983, 'clip_ratio/region_mean': 0.15394389256834984, 'reward_total_mean': 0.42997312545776367, 'reward_meter_mean': 0.4860062897205353, 'reward_meter_std': 0.4817962050437927, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6842336654663086, 'reward_repeat_penalty_std': 0.09754367172718048, 'reward_repeat_floor_mean': 0.9721978902816772, 'reward_repeat_floor_std': 0.01882293075323105, 'reward_near_duplicate_penalty_mean': 0.9228942394256592, 'reward_near_duplicate_penalty_std': 0.11104470491409302, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.8350000381469727, 'reward_judge_quality_std': 0.27717968821525574, 'reward_total_composite_mean': 0.42997312545776367, 'reward_total_composite_std': 0.44086191058158875, 'epoch': 0.04}

 24%|██▍       | 808/3300 [3:00:25<7:49:54, 11.31s/it]INFO 04-12 18:58:51 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 809/3300 [3:00:33<7:07:07, 10.29s/it]2026-04-12 18:58:59,414 | INFO | train_grpo_train | metrics_logged mode=train step=809

                                                      
{'loss': 0.0949, 'grad_norm': 25.848854064941406, 'learning_rate': 7.551515151515152e-06, 'num_tokens': 1453771.0, 'completions/mean_length': 20.25, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.6410830616950989, 'rewards/meter/std': 0.4693940281867981, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7562500238418579, 'rewards/judge_quality/std': 0.31717222929000854, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.40862423181533813, 'rewards/total_composite/std': 0.3837718963623047, 'reward': 0.40862423181533813, 'reward_std': 0.3837718665599823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17167606949806213, 'sampling/sampling_logp_difference/max': 1.4852440357208252, 'sampling/importance_sampling_ratio/min': 0.22644707560539246, 'sampling/importance_sampling_ratio/mean': 1.0281985998153687, 'sampling/importance_sampling_ratio/max': 1.703879475593567, 'entropy': 1.3407299667596817, 'clip_ratio/low_mean': 0.08878105599433184, 'clip_ratio/low_min': 0.08878105599433184, 'clip_ratio/high_mean': 0.09265541657805443, 'clip_ratio/high_max': 0.09265541657805443, 'clip_ratio/region_mean': 0.18143647257238626, 'reward_total_mean': 0.40862423181533813, 'reward_meter_mean': 0.6410830616950989, 'reward_meter_std': 0.4693940281867981, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7562500238418579, 'reward_judge_quality_std': 0.31717222929000854, 'reward_total_composite_mean': 0.40862423181533813, 'reward_total_composite_std': 0.3837718963623047, 'epoch': 0.04}

 25%|██▍       | 809/3300 [3:00:33<7:07:07, 10.29s/it]INFO 04-12 18:58:59 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 810/3300 [3:00:41<6:46:31,  9.80s/it]2026-04-12 18:59:08,017 | INFO | train_grpo_train | metrics_logged mode=train step=810

                                                      
{'loss': -0.004, 'grad_norm': 19.20218276977539, 'learning_rate': 7.548484848484849e-06, 'num_tokens': 1455207.0, 'completions/mean_length': 22.5, 'completions/min_length': 16.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9356511831283569, 'rewards/meter/std': 0.06340444833040237, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.6993392705917358, 'rewards/repeat_penalty/std': 0.07444862276315689, 'rewards/repeat_floor/mean': 0.9758293628692627, 'rewards/repeat_floor/std': 0.014460408128798008, 'rewards/near_duplicate_penalty/mean': 0.9484779834747314, 'rewards/near_duplicate_penalty/std': 0.09841844439506531, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.3734077513217926, 'rewards/total_composite/std': 0.2502562701702118, 'reward': 0.3734077513217926, 'reward_std': 0.2502562701702118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15154768526554108, 'sampling/sampling_logp_difference/max': 1.2297334671020508, 'sampling/importance_sampling_ratio/min': 0.292370468378067, 'sampling/importance_sampling_ratio/mean': 1.00874924659729, 'sampling/importance_sampling_ratio/max': 1.7784370183944702, 'entropy': 0.9116416499018669, 'clip_ratio/low_mean': 0.11185723450034857, 'clip_ratio/low_min': 0.11185723450034857, 'clip_ratio/high_mean': 0.00657894741743803, 'clip_ratio/high_max': 0.00657894741743803, 'clip_ratio/region_mean': 0.1184361819177866, 'reward_total_mean': 0.3734077513217926, 'reward_meter_mean': 0.9356511831283569, 'reward_meter_std': 0.06340444833040237, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6993392705917358, 'reward_repeat_penalty_std': 0.07444862276315689, 'reward_repeat_floor_mean': 0.9758293628692627, 'reward_repeat_floor_std': 0.014460408128798008, 'reward_near_duplicate_penalty_mean': 0.9484779834747314, 'reward_near_duplicate_penalty_std': 0.09841844439506531, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.3734077513217926, 'reward_total_composite_std': 0.2502562701702118, 'epoch': 0.04}

 25%|██▍       | 810/3300 [3:00:41<6:46:31,  9.80s/it]INFO 04-12 18:59:08 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 811/3300 [3:00:51<6:39:44,  9.64s/it]2026-04-12 18:59:17,279 | INFO | train_grpo_train | metrics_logged mode=train step=811

                                                      
{'loss': 0.4295, 'grad_norm': 14.159217834472656, 'learning_rate': 7.545454545454546e-06, 'num_tokens': 1457361.0, 'completions/mean_length': 47.25, 'completions/min_length': 23.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9171833992004395, 'rewards/meter/std': 0.2167324274778366, 'rewards/count_adherence/mean': 0.2678571343421936, 'rewards/count_adherence/std': 0.23458294570446014, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7803571224212646, 'rewards/count_floor/std': 0.07037486881017685, 'rewards/lexical_plausibility/mean': 0.96875, 'rewards/lexical_plausibility/std': 0.0883883461356163, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.6697031259536743, 'rewards/repeat_penalty/std': 0.11168906092643738, 'rewards/repeat_floor/mean': 0.9688522219657898, 'rewards/repeat_floor/std': 0.02090632915496826, 'rewards/near_duplicate_penalty/mean': 0.9268824458122253, 'rewards/near_duplicate_penalty/std': 0.10454464703798294, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9007759690284729, 'rewards/distinct_2/std': 0.13997726142406464, 'rewards/total_composite/mean': 0.20380830764770508, 'rewards/total_composite/std': 0.09110172092914581, 'reward': 0.20380830764770508, 'reward_std': 0.09110172837972641, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17390233278274536, 'sampling/sampling_logp_difference/max': 1.7461357116699219, 'sampling/importance_sampling_ratio/min': 0.2534649968147278, 'sampling/importance_sampling_ratio/mean': 1.001904845237732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0507089421153069, 'clip_ratio/low_mean': 0.04173261485993862, 'clip_ratio/low_min': 0.04173261485993862, 'clip_ratio/high_mean': 0.0699902456253767, 'clip_ratio/high_max': 0.0699902456253767, 'clip_ratio/region_mean': 0.11172286048531532, 'reward_total_mean': 0.20380830764770508, 'reward_meter_mean': 0.9171833992004395, 'reward_meter_std': 0.2167324274778366, 'reward_count_adherence_mean': 0.2678571343421936, 'reward_count_adherence_std': 0.23458294570446014, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7803571224212646, 'reward_count_floor_std': 0.07037486881017685, 'reward_lexical_plausibility_mean': 0.96875, 'reward_lexical_plausibility_std': 0.0883883461356163, 'reward_repeat_penalty_mean': 0.6697031259536743, 'reward_repeat_penalty_std': 0.11168906092643738, 'reward_repeat_floor_mean': 0.9688522219657898, 'reward_repeat_floor_std': 0.02090632915496826, 'reward_near_duplicate_penalty_mean': 0.9268824458122253, 'reward_near_duplicate_penalty_std': 0.10454464703798294, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9007759690284729, 'reward_distinct_2_std': 0.13997726142406464, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.20380830764770508, 'reward_total_composite_std': 0.09110172092914581, 'epoch': 0.04}

 25%|██▍       | 811/3300 [3:00:51<6:39:44,  9.64s/it]INFO 04-12 18:59:17 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 812/3300 [3:00:59<6:27:28,  9.34s/it]2026-04-12 18:59:25,942 | INFO | train_grpo_train | metrics_logged mode=train step=812

                                                      
{'loss': 0.1732, 'grad_norm': 16.96632957458496, 'learning_rate': 7.542424242424244e-06, 'num_tokens': 1458933.0, 'completions/mean_length': 28.5, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8495146632194519, 'rewards/meter/std': 0.2902800738811493, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9752674102783203, 'rewards/lexical_plausibility/std': 0.06995441019535065, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.268537700176239, 'rewards/repeat_penalty/mean': 0.6847589612007141, 'rewards/repeat_penalty/std': 0.1333535611629486, 'rewards/repeat_floor/mean': 0.9746805429458618, 'rewards/repeat_floor/std': 0.02138371393084526, 'rewards/near_duplicate_penalty/mean': 0.9695277214050293, 'rewards/near_duplicate_penalty/std': 0.062313929200172424, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9152930974960327, 'rewards/distinct_2/std': 0.19306930899620056, 'rewards/total_composite/mean': 0.2691442668437958, 'rewards/total_composite/std': 0.23518584668636322, 'reward': 0.2691442668437958, 'reward_std': 0.23518583178520203, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1595873236656189, 'sampling/sampling_logp_difference/max': 1.6597552299499512, 'sampling/importance_sampling_ratio/min': 0.1901855319738388, 'sampling/importance_sampling_ratio/mean': 1.0176295042037964, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0433703511953354, 'clip_ratio/low_mean': 0.06702722702175379, 'clip_ratio/low_min': 0.06702722702175379, 'clip_ratio/high_mean': 0.05598343722522259, 'clip_ratio/high_max': 0.05598343722522259, 'clip_ratio/region_mean': 0.12301066424697638, 'reward_total_mean': 0.2691442668437958, 'reward_meter_mean': 0.8495146632194519, 'reward_meter_std': 0.2902800738811493, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9752674102783203, 'reward_lexical_plausibility_std': 0.06995441019535065, 'reward_repeat_penalty_mean': 0.6847589612007141, 'reward_repeat_penalty_std': 0.1333535611629486, 'reward_repeat_floor_mean': 0.9746805429458618, 'reward_repeat_floor_std': 0.02138371393084526, 'reward_near_duplicate_penalty_mean': 0.9695277214050293, 'reward_near_duplicate_penalty_std': 0.062313929200172424, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9152930974960327, 'reward_distinct_2_std': 0.19306930899620056, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.268537700176239, 'reward_total_composite_mean': 0.2691442668437958, 'reward_total_composite_std': 0.23518584668636322, 'epoch': 0.04}

 25%|██▍       | 812/3300 [3:00:59<6:27:28,  9.34s/it]INFO 04-12 18:59:26 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 813/3300 [3:01:13<7:21:07, 10.64s/it]2026-04-12 18:59:39,618 | INFO | train_grpo_train | metrics_logged mode=train step=813

                                                      
{'loss': -0.0351, 'grad_norm': 4.430045127868652, 'learning_rate': 7.53939393939394e-06, 'num_tokens': 1460651.0, 'completions/mean_length': 83.75, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.571430206298828, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8289089202880859, 'rewards/meter/std': 0.26676511764526367, 'rewards/count_adherence/mean': 0.28125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7843749523162842, 'rewards/count_floor/std': 0.026516521349549294, 'rewards/lexical_plausibility/mean': 0.9499424695968628, 'rewards/lexical_plausibility/std': 0.14158403873443604, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.7628153562545776, 'rewards/repeat_penalty/std': 0.10193809121847153, 'rewards/repeat_floor/mean': 0.9840642213821411, 'rewards/repeat_floor/std': 0.008201684802770615, 'rewards/near_duplicate_penalty/mean': 0.9901372194290161, 'rewards/near_duplicate_penalty/std': 0.025879662483930588, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.21908780932426453, 'rewards/total_composite/std': 0.13027232885360718, 'reward': 0.21908780932426453, 'reward_std': 0.13027232885360718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1934020221233368, 'sampling/sampling_logp_difference/max': 1.2168521881103516, 'sampling/importance_sampling_ratio/min': 0.2961609661579132, 'sampling/importance_sampling_ratio/mean': 1.024492859840393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3589426279067993, 'clip_ratio/low_mean': 0.03375000040978193, 'clip_ratio/low_min': 0.03375000040978193, 'clip_ratio/high_mean': 0.14882544800639153, 'clip_ratio/high_max': 0.14882544800639153, 'clip_ratio/region_mean': 0.18257544841617346, 'reward_total_mean': 0.21908780932426453, 'reward_meter_mean': 0.8289089202880859, 'reward_meter_std': 0.26676511764526367, 'reward_count_adherence_mean': 0.28125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7843749523162842, 'reward_count_floor_std': 0.026516521349549294, 'reward_lexical_plausibility_mean': 0.9499424695968628, 'reward_lexical_plausibility_std': 0.14158403873443604, 'reward_repeat_penalty_mean': 0.7628153562545776, 'reward_repeat_penalty_std': 0.10193809121847153, 'reward_repeat_floor_mean': 0.9840642213821411, 'reward_repeat_floor_std': 0.008201684802770615, 'reward_near_duplicate_penalty_mean': 0.9901372194290161, 'reward_near_duplicate_penalty_std': 0.025879662483930588, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.21908780932426453, 'reward_total_composite_std': 0.13027232885360718, 'epoch': 0.04}

 25%|██▍       | 813/3300 [3:01:13<7:21:07, 10.64s/it]INFO 04-12 18:59:39 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 814/3300 [3:01:27<8:03:20, 11.67s/it]2026-04-12 18:59:53,665 | INFO | train_grpo_train | metrics_logged mode=train step=814

                                                      
{'loss': -0.1966, 'grad_norm': 4.813582420349121, 'learning_rate': 7.536363636363637e-06, 'num_tokens': 1462491.0, 'completions/mean_length': 107.0, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.142860412597656, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.7007884383201599, 'rewards/meter/std': 0.27218306064605713, 'rewards/count_adherence/mean': 0.44999998807907104, 'rewards/count_adherence/std': 0.20701968669891357, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8350000381469727, 'rewards/count_floor/std': 0.06210590526461601, 'rewards/lexical_plausibility/mean': 0.9431955814361572, 'rewards/lexical_plausibility/std': 0.16066719591617584, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.8224967122077942, 'rewards/repeat_penalty/std': 0.15461789071559906, 'rewards/repeat_floor/mean': 0.9810358881950378, 'rewards/repeat_floor/std': 0.018225310370326042, 'rewards/near_duplicate_penalty/mean': 0.9384067058563232, 'rewards/near_duplicate_penalty/std': 0.0786098837852478, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9336100220680237, 'rewards/distinct_2/std': 0.08975629508495331, 'rewards/total_composite/mean': 0.16635443270206451, 'rewards/total_composite/std': 0.09484672546386719, 'reward': 0.16635443270206451, 'reward_std': 0.09484672546386719, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14266599714756012, 'sampling/sampling_logp_difference/max': 1.5299382209777832, 'sampling/importance_sampling_ratio/min': 0.21654905378818512, 'sampling/importance_sampling_ratio/mean': 1.0361651182174683, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9007492735981941, 'clip_ratio/low_mean': 0.019736841320991516, 'clip_ratio/low_min': 0.019736841320991516, 'clip_ratio/high_mean': 0.11878337897360325, 'clip_ratio/high_max': 0.11878337897360325, 'clip_ratio/region_mean': 0.13852022029459476, 'reward_total_mean': 0.16635443270206451, 'reward_meter_mean': 0.7007884383201599, 'reward_meter_std': 0.27218306064605713, 'reward_count_adherence_mean': 0.44999998807907104, 'reward_count_adherence_std': 0.20701968669891357, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8350000381469727, 'reward_count_floor_std': 0.06210590526461601, 'reward_lexical_plausibility_mean': 0.9431955814361572, 'reward_lexical_plausibility_std': 0.16066719591617584, 'reward_repeat_penalty_mean': 0.8224967122077942, 'reward_repeat_penalty_std': 0.15461789071559906, 'reward_repeat_floor_mean': 0.9810358881950378, 'reward_repeat_floor_std': 0.018225310370326042, 'reward_near_duplicate_penalty_mean': 0.9384067058563232, 'reward_near_duplicate_penalty_std': 0.0786098837852478, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9336100220680237, 'reward_distinct_2_std': 0.08975629508495331, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.16635443270206451, 'reward_total_composite_std': 0.09484672546386719, 'epoch': 0.04}

 25%|██▍       | 814/3300 [3:01:27<8:03:20, 11.67s/it]INFO 04-12 18:59:53 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 815/3300 [3:01:37<7:37:46, 11.05s/it]2026-04-12 19:00:03,294 | INFO | train_grpo_train | metrics_logged mode=train step=815

                                                      
{'loss': 0.0999, 'grad_norm': 12.627212524414062, 'learning_rate': 7.533333333333334e-06, 'num_tokens': 1464546.0, 'completions/mean_length': 64.875, 'completions/min_length': 48.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.5496289730072021, 'rewards/meter/std': 0.29180029034614563, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136146903038025, 'rewards/repeat_penalty/mean': 0.8907186388969421, 'rewards/repeat_penalty/std': 0.08900643140077591, 'rewards/repeat_floor/mean': 0.9878484010696411, 'rewards/repeat_floor/std': 0.009585205465555191, 'rewards/near_duplicate_penalty/mean': 0.9671115875244141, 'rewards/near_duplicate_penalty/std': 0.0205213725566864, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9197964668273926, 'rewards/distinct_2/std': 0.07553186267614365, 'rewards/total_composite/mean': 0.2369551658630371, 'rewards/total_composite/std': 0.2612629532814026, 'reward': 0.2369551658630371, 'reward_std': 0.2612629532814026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19140778481960297, 'sampling/sampling_logp_difference/max': 1.78289794921875, 'sampling/importance_sampling_ratio/min': 0.17033539712429047, 'sampling/importance_sampling_ratio/mean': 0.9856112003326416, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0083779841661453, 'clip_ratio/low_mean': 0.09884848818182945, 'clip_ratio/low_min': 0.09884848818182945, 'clip_ratio/high_mean': 0.06275434233248234, 'clip_ratio/high_max': 0.06275434233248234, 'clip_ratio/region_mean': 0.1616028305143118, 'reward_total_mean': 0.2369551658630371, 'reward_meter_mean': 0.5496289730072021, 'reward_meter_std': 0.29180029034614563, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8907186388969421, 'reward_repeat_penalty_std': 0.08900643140077591, 'reward_repeat_floor_mean': 0.9878484010696411, 'reward_repeat_floor_std': 0.009585205465555191, 'reward_near_duplicate_penalty_mean': 0.9671115875244141, 'reward_near_duplicate_penalty_std': 0.0205213725566864, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9197964668273926, 'reward_distinct_2_std': 0.07553186267614365, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136146903038025, 'reward_total_composite_mean': 0.2369551658630371, 'reward_total_composite_std': 0.2612629532814026, 'epoch': 0.04}

 25%|██▍       | 815/3300 [3:01:37<7:37:46, 11.05s/it]INFO 04-12 19:00:03 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 816/3300 [3:01:45<7:04:05, 10.24s/it]2026-04-12 19:00:11,647 | INFO | train_grpo_train | metrics_logged mode=train step=816

                                                      
{'loss': 0.0736, 'grad_norm': 22.224498748779297, 'learning_rate': 7.530303030303031e-06, 'num_tokens': 1465961.0, 'completions/mean_length': 19.875, 'completions/min_length': 18.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.6954392194747925, 'rewards/meter/std': 0.411810040473938, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.731249988079071, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.7489533424377441, 'rewards/repeat_penalty/std': 0.0029603736475110054, 'rewards/repeat_floor/mean': 0.9847906827926636, 'rewards/repeat_floor/std': 0.0005920789553783834, 'rewards/near_duplicate_penalty/mean': 0.9986044764518738, 'rewards/near_duplicate_penalty/std': 0.003947157878428698, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.546711802482605, 'rewards/total_composite/std': 0.38904139399528503, 'reward': 0.546711802482605, 'reward_std': 0.38904139399528503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.216852068901062, 'sampling/sampling_logp_difference/max': 1.6973917484283447, 'sampling/importance_sampling_ratio/min': 0.18316063284873962, 'sampling/importance_sampling_ratio/mean': 1.036706805229187, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4984114170074463, 'clip_ratio/low_mean': 0.08459596149623394, 'clip_ratio/low_min': 0.08459596149623394, 'clip_ratio/high_mean': 0.0734127014875412, 'clip_ratio/high_max': 0.0734127014875412, 'clip_ratio/region_mean': 0.15800866298377514, 'reward_total_mean': 0.546711802482605, 'reward_meter_mean': 0.6954392194747925, 'reward_meter_std': 0.411810040473938, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7489533424377441, 'reward_repeat_penalty_std': 0.0029603736475110054, 'reward_repeat_floor_mean': 0.9847906827926636, 'reward_repeat_floor_std': 0.0005920789553783834, 'reward_near_duplicate_penalty_mean': 0.9986044764518738, 'reward_near_duplicate_penalty_std': 0.003947157878428698, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.731249988079071, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.546711802482605, 'reward_total_composite_std': 0.38904139399528503, 'epoch': 0.04}

 25%|██▍       | 816/3300 [3:01:45<7:04:05, 10.24s/it]INFO 04-12 19:00:11 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 817/3300 [3:01:53<6:39:11,  9.65s/it]2026-04-12 19:00:19,896 | INFO | train_grpo_train | metrics_logged mode=train step=817

                                                      
{'loss': -0.0327, 'grad_norm': 20.363027572631836, 'learning_rate': 7.5272727272727274e-06, 'num_tokens': 1467463.0, 'completions/mean_length': 18.75, 'completions/min_length': 16.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9814982414245605, 'rewards/meter/std': 0.012693759053945541, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.684036135673523, 'rewards/repeat_penalty/std': 0.11495808511972427, 'rewards/repeat_floor/mean': 0.9692908525466919, 'rewards/repeat_floor/std': 0.01916888728737831, 'rewards/near_duplicate_penalty/mean': 0.899031400680542, 'rewards/near_duplicate_penalty/std': 0.10563550889492035, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9729020595550537, 'rewards/distinct_2/std': 0.05291995406150818, 'rewards/total_composite/mean': 0.2501251697540283, 'rewards/total_composite/std': 0.1132427453994751, 'reward': 0.2501251697540283, 'reward_std': 0.1132427453994751, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1789841502904892, 'sampling/sampling_logp_difference/max': 1.7048187255859375, 'sampling/importance_sampling_ratio/min': 0.1818053424358368, 'sampling/importance_sampling_ratio/mean': 1.0218631029129028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2871474474668503, 'clip_ratio/low_mean': 0.06093750009313226, 'clip_ratio/low_min': 0.06093750009313226, 'clip_ratio/high_mean': 0.11934603471308947, 'clip_ratio/high_max': 0.11934603471308947, 'clip_ratio/region_mean': 0.18028353480622172, 'reward_total_mean': 0.2501251697540283, 'reward_meter_mean': 0.9814982414245605, 'reward_meter_std': 0.012693759053945541, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.684036135673523, 'reward_repeat_penalty_std': 0.11495808511972427, 'reward_repeat_floor_mean': 0.9692908525466919, 'reward_repeat_floor_std': 0.01916888728737831, 'reward_near_duplicate_penalty_mean': 0.899031400680542, 'reward_near_duplicate_penalty_std': 0.10563550889492035, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9729020595550537, 'reward_distinct_2_std': 0.05291995406150818, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.2501251697540283, 'reward_total_composite_std': 0.1132427453994751, 'epoch': 0.04}

 25%|██▍       | 817/3300 [3:01:53<6:39:11,  9.65s/it]INFO 04-12 19:00:20 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 818/3300 [3:02:02<6:24:59,  9.31s/it]2026-04-12 19:00:28,412 | INFO | train_grpo_train | metrics_logged mode=train step=818

                                                      
{'loss': 0.0782, 'grad_norm': 18.549962997436523, 'learning_rate': 7.524242424242425e-06, 'num_tokens': 1468952.0, 'completions/mean_length': 32.125, 'completions/min_length': 29.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.3971840739250183, 'rewards/meter/std': 0.3852846622467041, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.9718365669250488, 'rewards/repeat_penalty/std': 0.06021438539028168, 'rewards/repeat_floor/mean': 0.996390700340271, 'rewards/repeat_floor/std': 0.007331717759370804, 'rewards/near_duplicate_penalty/mean': 0.983630359172821, 'rewards/near_duplicate_penalty/std': 0.028202861547470093, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.17902232706546783, 'rewards/total_composite/std': 0.17599532008171082, 'reward': 0.17902232706546783, 'reward_std': 0.17599530518054962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1546567678451538, 'sampling/sampling_logp_difference/max': 1.7123899459838867, 'sampling/importance_sampling_ratio/min': 0.18043406307697296, 'sampling/importance_sampling_ratio/mean': 1.0096814632415771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8197788000106812, 'clip_ratio/low_mean': 0.08430119324475527, 'clip_ratio/low_min': 0.08430119324475527, 'clip_ratio/high_mean': 0.054463294334709644, 'clip_ratio/high_max': 0.054463294334709644, 'clip_ratio/region_mean': 0.1387644875794649, 'reward_total_mean': 0.17902232706546783, 'reward_meter_mean': 0.3971840739250183, 'reward_meter_std': 0.3852846622467041, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9718365669250488, 'reward_repeat_penalty_std': 0.06021438539028168, 'reward_repeat_floor_mean': 0.996390700340271, 'reward_repeat_floor_std': 0.007331717759370804, 'reward_near_duplicate_penalty_mean': 0.983630359172821, 'reward_near_duplicate_penalty_std': 0.028202861547470093, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.17902232706546783, 'reward_total_composite_std': 0.17599532008171082, 'epoch': 0.04}

 25%|██▍       | 818/3300 [3:02:02<6:24:59,  9.31s/it]INFO 04-12 19:00:28 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 819/3300 [3:02:10<6:11:58,  9.00s/it]2026-04-12 19:00:36,681 | INFO | train_grpo_train | metrics_logged mode=train step=819

                                                      
{'loss': 0.0129, 'grad_norm': 17.536701202392578, 'learning_rate': 7.521212121212121e-06, 'num_tokens': 1470437.0, 'completions/mean_length': 22.625, 'completions/min_length': 17.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.8637288808822632, 'rewards/meter/std': 0.2945989668369293, 'rewards/count_adherence/mean': 0.3333333432674408, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.800000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.966911792755127, 'rewards/lexical_plausibility/std': 0.09358766674995422, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14961546659469604, 'rewards/total_composite/std': 0.10412398725748062, 'reward': 0.14961546659469604, 'reward_std': 0.10412398725748062, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20420442521572113, 'sampling/sampling_logp_difference/max': 2.08994197845459, 'sampling/importance_sampling_ratio/min': 0.12369430810213089, 'sampling/importance_sampling_ratio/mean': 1.0211299657821655, 'sampling/importance_sampling_ratio/max': 1.8783628940582275, 'entropy': 1.736636683344841, 'clip_ratio/low_mean': 0.1538101639598608, 'clip_ratio/low_min': 0.1538101639598608, 'clip_ratio/high_mean': 0.047173911705613136, 'clip_ratio/high_max': 0.047173911705613136, 'clip_ratio/region_mean': 0.20098407566547394, 'reward_total_mean': 0.14961546659469604, 'reward_meter_mean': 0.8637288808822632, 'reward_meter_std': 0.2945989668369293, 'reward_count_adherence_mean': 0.3333333432674408, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.800000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.966911792755127, 'reward_lexical_plausibility_std': 0.09358766674995422, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.14961546659469604, 'reward_total_composite_std': 0.10412398725748062, 'epoch': 0.04}

 25%|██▍       | 819/3300 [3:02:10<6:11:58,  9.00s/it]INFO 04-12 19:00:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 820/3300 [3:02:23<7:00:40, 10.18s/it]2026-04-12 19:00:49,618 | INFO | train_grpo_train | metrics_logged mode=train step=820

                                                      
{'loss': -0.0244, 'grad_norm': 4.032950401306152, 'learning_rate': 7.518181818181819e-06, 'num_tokens': 1471966.0, 'completions/mean_length': 90.125, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 29.85714340209961, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7555403709411621, 'rewards/meter/std': 0.4275420606136322, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9685947895050049, 'rewards/lexical_plausibility/std': 0.08882737159729004, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.377548485994339, 'rewards/repeat_penalty/mean': 0.7073977589607239, 'rewards/repeat_penalty/std': 0.10778601467609406, 'rewards/repeat_floor/mean': 0.974085807800293, 'rewards/repeat_floor/std': 0.01888308674097061, 'rewards/near_duplicate_penalty/mean': 0.9309778213500977, 'rewards/near_duplicate_penalty/std': 0.10524194687604904, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9729344844818115, 'rewards/distinct_2/std': 0.05020807683467865, 'rewards/total_composite/mean': 0.3957326114177704, 'rewards/total_composite/std': 0.3144020140171051, 'reward': 0.3957326114177704, 'reward_std': 0.3144020140171051, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1674242466688156, 'sampling/sampling_logp_difference/max': 1.392251968383789, 'sampling/importance_sampling_ratio/min': 0.2485150396823883, 'sampling/importance_sampling_ratio/mean': 1.0011268854141235, 'sampling/importance_sampling_ratio/max': 1.9608087539672852, 'entropy': 1.1507271006703377, 'clip_ratio/low_mean': 0.06778190191835165, 'clip_ratio/low_min': 0.06778190191835165, 'clip_ratio/high_mean': 0.05017857253551483, 'clip_ratio/high_max': 0.05017857253551483, 'clip_ratio/region_mean': 0.11796047445386648, 'reward_total_mean': 0.3957326114177704, 'reward_meter_mean': 0.7555403709411621, 'reward_meter_std': 0.4275420606136322, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9685947895050049, 'reward_lexical_plausibility_std': 0.08882737159729004, 'reward_repeat_penalty_mean': 0.7073977589607239, 'reward_repeat_penalty_std': 0.10778601467609406, 'reward_repeat_floor_mean': 0.974085807800293, 'reward_repeat_floor_std': 0.01888308674097061, 'reward_near_duplicate_penalty_mean': 0.9309778213500977, 'reward_near_duplicate_penalty_std': 0.10524194687604904, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9729344844818115, 'reward_distinct_2_std': 0.05020807683467865, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.377548485994339, 'reward_total_composite_mean': 0.3957326114177704, 'reward_total_composite_std': 0.3144020140171051, 'epoch': 0.04}

 25%|██▍       | 820/3300 [3:02:23<7:00:40, 10.18s/it]INFO 04-12 19:00:49 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 821/3300 [3:02:36<7:35:50, 11.03s/it]2026-04-12 19:01:02,644 | INFO | train_grpo_train | metrics_logged mode=train step=821

                                                      
{'loss': -0.017, 'grad_norm': 4.432583808898926, 'learning_rate': 7.515151515151516e-06, 'num_tokens': 1473251.0, 'completions/mean_length': 85.625, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.71428680419922, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7224617600440979, 'rewards/meter/std': 0.444905549287796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9313076734542847, 'rewards/lexical_plausibility/std': 0.1397034376859665, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.2728912830352783, 'rewards/repeat_penalty/mean': 0.7402052283287048, 'rewards/repeat_penalty/std': 0.027703819796442986, 'rewards/repeat_floor/mean': 0.9830410480499268, 'rewards/repeat_floor/std': 0.005540772341191769, 'rewards/near_duplicate_penalty/mean': 0.9869402647018433, 'rewards/near_duplicate_penalty/std': 0.03693840652704239, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.226249560713768, 'rewards/total_composite/std': 0.19705277681350708, 'reward': 0.226249560713768, 'reward_std': 0.19705277681350708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21654000878334045, 'sampling/sampling_logp_difference/max': 1.2811155319213867, 'sampling/importance_sampling_ratio/min': 0.27772730588912964, 'sampling/importance_sampling_ratio/mean': 1.099709153175354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6918454021215439, 'clip_ratio/low_mean': 0.054286833852529526, 'clip_ratio/low_min': 0.054286833852529526, 'clip_ratio/high_mean': 0.09520121151581407, 'clip_ratio/high_max': 0.09520121151581407, 'clip_ratio/region_mean': 0.1494880453683436, 'reward_total_mean': 0.226249560713768, 'reward_meter_mean': 0.7224617600440979, 'reward_meter_std': 0.444905549287796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9313076734542847, 'reward_lexical_plausibility_std': 0.1397034376859665, 'reward_repeat_penalty_mean': 0.7402052283287048, 'reward_repeat_penalty_std': 0.027703819796442986, 'reward_repeat_floor_mean': 0.9830410480499268, 'reward_repeat_floor_std': 0.005540772341191769, 'reward_near_duplicate_penalty_mean': 0.9869402647018433, 'reward_near_duplicate_penalty_std': 0.03693840652704239, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.2728912830352783, 'reward_total_composite_mean': 0.226249560713768, 'reward_total_composite_std': 0.19705277681350708, 'epoch': 0.04}

 25%|██▍       | 821/3300 [3:02:36<7:35:50, 11.03s/it]INFO 04-12 19:01:02 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 822/3300 [3:02:50<8:13:34, 11.95s/it]2026-04-12 19:01:16,739 | INFO | train_grpo_train | metrics_logged mode=train step=822

                                                      
{'loss': 0.0059, 'grad_norm': 4.892788410186768, 'learning_rate': 7.512121212121213e-06, 'num_tokens': 1474815.0, 'completions/mean_length': 92.5, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.57143020629883, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8654338121414185, 'rewards/meter/std': 0.28509634733200073, 'rewards/count_adherence/mean': 0.40625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8218749761581421, 'rewards/count_floor/std': 0.038816213607788086, 'rewards/lexical_plausibility/mean': 0.9496781229972839, 'rewards/lexical_plausibility/std': 0.1423318088054657, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3160809278488159, 'rewards/repeat_penalty/mean': 0.789899468421936, 'rewards/repeat_penalty/std': 0.2299429476261139, 'rewards/repeat_floor/mean': 0.976887047290802, 'rewards/repeat_floor/std': 0.029898732900619507, 'rewards/near_duplicate_penalty/mean': 0.9189907312393188, 'rewards/near_duplicate_penalty/std': 0.11814222484827042, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9615384340286255, 'rewards/distinct_2/std': 0.10878565162420273, 'rewards/total_composite/mean': 0.3553149104118347, 'rewards/total_composite/std': 0.23925107717514038, 'reward': 0.3553149104118347, 'reward_std': 0.23925107717514038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15001611411571503, 'sampling/sampling_logp_difference/max': 1.400146245956421, 'sampling/importance_sampling_ratio/min': 0.24656091630458832, 'sampling/importance_sampling_ratio/mean': 1.0442984104156494, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9598062038421631, 'clip_ratio/low_mean': 0.05434782616794109, 'clip_ratio/low_min': 0.05434782616794109, 'clip_ratio/high_mean': 0.10035935696214437, 'clip_ratio/high_max': 0.10035935696214437, 'clip_ratio/region_mean': 0.15470718313008547, 'reward_total_mean': 0.3553149104118347, 'reward_meter_mean': 0.8654338121414185, 'reward_meter_std': 0.28509634733200073, 'reward_count_adherence_mean': 0.40625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8218749761581421, 'reward_count_floor_std': 0.038816213607788086, 'reward_lexical_plausibility_mean': 0.9496781229972839, 'reward_lexical_plausibility_std': 0.1423318088054657, 'reward_repeat_penalty_mean': 0.789899468421936, 'reward_repeat_penalty_std': 0.2299429476261139, 'reward_repeat_floor_mean': 0.976887047290802, 'reward_repeat_floor_std': 0.029898732900619507, 'reward_near_duplicate_penalty_mean': 0.9189907312393188, 'reward_near_duplicate_penalty_std': 0.11814222484827042, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9615384340286255, 'reward_distinct_2_std': 0.10878565162420273, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3160809278488159, 'reward_total_composite_mean': 0.3553149104118347, 'reward_total_composite_std': 0.23925107717514038, 'epoch': 0.04}

 25%|██▍       | 822/3300 [3:02:50<8:13:34, 11.95s/it]INFO 04-12 19:01:16 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 823/3300 [3:03:04<8:37:59, 12.55s/it]2026-04-12 19:01:30,680 | INFO | train_grpo_train | metrics_logged mode=train step=823

                                                      
{'loss': -0.0354, 'grad_norm': 1.747131586074829, 'learning_rate': 7.509090909090909e-06, 'num_tokens': 1476076.0, 'completions/mean_length': 213.625, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 34.60000228881836, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5255786776542664, 'rewards/meter/std': 0.42836275696754456, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.8589390516281128, 'rewards/lexical_plausibility/std': 0.2013278305530548, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.3026755750179291, 'rewards/repeat_penalty/mean': 0.7820367813110352, 'rewards/repeat_penalty/std': 0.13554362952709198, 'rewards/repeat_floor/mean': 0.983207106590271, 'rewards/repeat_floor/std': 0.011739317327737808, 'rewards/near_duplicate_penalty/mean': 0.9688335657119751, 'rewards/near_duplicate_penalty/std': 0.03714458644390106, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9903568625450134, 'rewards/distinct_2/std': 0.02574104629456997, 'rewards/total_composite/mean': 0.21174120903015137, 'rewards/total_composite/std': 0.24407526850700378, 'reward': 0.21174120903015137, 'reward_std': 0.24407526850700378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17617204785346985, 'sampling/sampling_logp_difference/max': 1.3265776634216309, 'sampling/importance_sampling_ratio/min': 0.2653839588165283, 'sampling/importance_sampling_ratio/mean': 1.0312964916229248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9010612815618515, 'clip_ratio/low_mean': 0.03791291266679764, 'clip_ratio/low_min': 0.03791291266679764, 'clip_ratio/high_mean': 0.05412115249782801, 'clip_ratio/high_max': 0.05412115249782801, 'clip_ratio/region_mean': 0.09203406516462564, 'reward_total_mean': 0.21174120903015137, 'reward_meter_mean': 0.5255786776542664, 'reward_meter_std': 0.42836275696754456, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.8589390516281128, 'reward_lexical_plausibility_std': 0.2013278305530548, 'reward_repeat_penalty_mean': 0.7820367813110352, 'reward_repeat_penalty_std': 0.13554362952709198, 'reward_repeat_floor_mean': 0.983207106590271, 'reward_repeat_floor_std': 0.011739317327737808, 'reward_near_duplicate_penalty_mean': 0.9688335657119751, 'reward_near_duplicate_penalty_std': 0.03714458644390106, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9903568625450134, 'reward_distinct_2_std': 0.02574104629456997, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.3026755750179291, 'reward_total_composite_mean': 0.21174120903015137, 'reward_total_composite_std': 0.24407526850700378, 'epoch': 0.04}

 25%|██▍       | 823/3300 [3:03:04<8:37:59, 12.55s/it]INFO 04-12 19:01:30 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 824/3300 [3:03:18<9:00:43, 13.10s/it]2026-04-12 19:01:45,078 | INFO | train_grpo_train | metrics_logged mode=train step=824

                                                      
{'loss': -0.023, 'grad_norm': 2.2957868576049805, 'learning_rate': 7.5060606060606065e-06, 'num_tokens': 1477451.0, 'completions/mean_length': 207.875, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 25.399999618530273, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.435158371925354, 'rewards/meter/std': 0.4625847637653351, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.7725337743759155, 'rewards/lexical_plausibility/std': 0.2591196894645691, 'rewards/judge_quality/mean': 0.36125001311302185, 'rewards/judge_quality/std': 0.3782077431678772, 'rewards/repeat_penalty/mean': 0.8146295547485352, 'rewards/repeat_penalty/std': 0.1268990933895111, 'rewards/repeat_floor/mean': 0.9856278896331787, 'rewards/repeat_floor/std': 0.010767897590994835, 'rewards/near_duplicate_penalty/mean': 0.9666858911514282, 'rewards/near_duplicate_penalty/std': 0.05830582231283188, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1266794353723526, 'rewards/total_composite/std': 0.24408572912216187, 'reward': 0.1266794353723526, 'reward_std': 0.24408569931983948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.224935382604599, 'sampling/sampling_logp_difference/max': 1.225947380065918, 'sampling/importance_sampling_ratio/min': 0.2934795320034027, 'sampling/importance_sampling_ratio/mean': 1.0713961124420166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2053080052137375, 'clip_ratio/low_mean': 0.07659837417304516, 'clip_ratio/low_min': 0.07659837417304516, 'clip_ratio/high_mean': 0.036445013247430325, 'clip_ratio/high_max': 0.036445013247430325, 'clip_ratio/region_mean': 0.11304338742047548, 'reward_total_mean': 0.1266794353723526, 'reward_meter_mean': 0.435158371925354, 'reward_meter_std': 0.4625847637653351, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.7725337743759155, 'reward_lexical_plausibility_std': 0.2591196894645691, 'reward_repeat_penalty_mean': 0.8146295547485352, 'reward_repeat_penalty_std': 0.1268990933895111, 'reward_repeat_floor_mean': 0.9856278896331787, 'reward_repeat_floor_std': 0.010767897590994835, 'reward_near_duplicate_penalty_mean': 0.9666858911514282, 'reward_near_duplicate_penalty_std': 0.05830582231283188, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36125001311302185, 'reward_judge_quality_std': 0.3782077431678772, 'reward_total_composite_mean': 0.1266794353723526, 'reward_total_composite_std': 0.24408572912216187, 'epoch': 0.04}

 25%|██▍       | 824/3300 [3:03:18<9:00:43, 13.10s/it]INFO 04-12 19:01:45 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 825/3300 [3:03:32<9:04:41, 13.20s/it]2026-04-12 19:01:58,568 | INFO | train_grpo_train | metrics_logged mode=train step=825

                                                      
{'loss': -0.0418, 'grad_norm': 4.6135382652282715, 'learning_rate': 7.503030303030303e-06, 'num_tokens': 1478858.0, 'completions/mean_length': 83.875, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.71428680419922, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9282773733139038, 'rewards/meter/std': 0.1416764110326767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9474941492080688, 'rewards/lexical_plausibility/std': 0.1485089808702469, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.33161941170692444, 'rewards/repeat_penalty/mean': 0.7089676260948181, 'rewards/repeat_penalty/std': 0.08527421951293945, 'rewards/repeat_floor/mean': 0.977798342704773, 'rewards/repeat_floor/std': 0.014461321756243706, 'rewards/near_duplicate_penalty/mean': 0.9673733711242676, 'rewards/near_duplicate_penalty/std': 0.06057348847389221, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9743589758872986, 'rewards/distinct_2/std': 0.07252377271652222, 'rewards/total_composite/mean': 0.5850933790206909, 'rewards/total_composite/std': 0.3361518085002899, 'reward': 0.5850933790206909, 'reward_std': 0.33615177869796753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1841418296098709, 'sampling/sampling_logp_difference/max': 1.4703569412231445, 'sampling/importance_sampling_ratio/min': 0.22984343767166138, 'sampling/importance_sampling_ratio/mean': 1.0410012006759644, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.361395351588726, 'clip_ratio/low_mean': 0.06971343979239464, 'clip_ratio/low_min': 0.06971343979239464, 'clip_ratio/high_mean': 0.06624579289928079, 'clip_ratio/high_max': 0.06624579289928079, 'clip_ratio/region_mean': 0.13595923269167542, 'reward_total_mean': 0.5850933790206909, 'reward_meter_mean': 0.9282773733139038, 'reward_meter_std': 0.1416764110326767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9474941492080688, 'reward_lexical_plausibility_std': 0.1485089808702469, 'reward_repeat_penalty_mean': 0.7089676260948181, 'reward_repeat_penalty_std': 0.08527421951293945, 'reward_repeat_floor_mean': 0.977798342704773, 'reward_repeat_floor_std': 0.014461321756243706, 'reward_near_duplicate_penalty_mean': 0.9673733711242676, 'reward_near_duplicate_penalty_std': 0.06057348847389221, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9743589758872986, 'reward_distinct_2_std': 0.07252377271652222, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.33161941170692444, 'reward_total_composite_mean': 0.5850933790206909, 'reward_total_composite_std': 0.3361518085002899, 'epoch': 0.04}

 25%|██▌       | 825/3300 [3:03:32<9:04:41, 13.20s/it]INFO 04-12 19:01:58 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 826/3300 [3:03:46<9:14:15, 13.44s/it]2026-04-12 19:02:12,514 | INFO | train_grpo_train | metrics_logged mode=train step=826

                                                      
{'loss': -0.0405, 'grad_norm': 5.854129791259766, 'learning_rate': 7.500000000000001e-06, 'num_tokens': 1480228.0, 'completions/mean_length': 82.25, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.85714340209961, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7147685289382935, 'rewards/meter/std': 0.37668249011039734, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9501920938491821, 'rewards/lexical_plausibility/std': 0.11767436563968658, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.7453800439834595, 'rewards/repeat_penalty/std': 0.11828261613845825, 'rewards/repeat_floor/mean': 0.9798123836517334, 'rewards/repeat_floor/std': 0.014985439367592335, 'rewards/near_duplicate_penalty/mean': 0.9535229206085205, 'rewards/near_duplicate_penalty/std': 0.08709587901830673, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9989878535270691, 'rewards/distinct_2/std': 0.002862782683223486, 'rewards/total_composite/mean': 0.1379672884941101, 'rewards/total_composite/std': 0.12172599136829376, 'reward': 0.1379672884941101, 'reward_std': 0.12172599136829376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20663116872310638, 'sampling/sampling_logp_difference/max': 1.4144797325134277, 'sampling/importance_sampling_ratio/min': 0.24305203557014465, 'sampling/importance_sampling_ratio/mean': 1.019883632659912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.525448977947235, 'clip_ratio/low_mean': 0.12825261894613504, 'clip_ratio/low_min': 0.12825261894613504, 'clip_ratio/high_mean': 0.04304029420018196, 'clip_ratio/high_max': 0.04304029420018196, 'clip_ratio/region_mean': 0.171292913146317, 'reward_total_mean': 0.1379672884941101, 'reward_meter_mean': 0.7147685289382935, 'reward_meter_std': 0.37668249011039734, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9501920938491821, 'reward_lexical_plausibility_std': 0.11767436563968658, 'reward_repeat_penalty_mean': 0.7453800439834595, 'reward_repeat_penalty_std': 0.11828261613845825, 'reward_repeat_floor_mean': 0.9798123836517334, 'reward_repeat_floor_std': 0.014985439367592335, 'reward_near_duplicate_penalty_mean': 0.9535229206085205, 'reward_near_duplicate_penalty_std': 0.08709587901830673, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9989878535270691, 'reward_distinct_2_std': 0.002862782683223486, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.1379672884941101, 'reward_total_composite_std': 0.12172599136829376, 'epoch': 0.04}

 25%|██▌       | 826/3300 [3:03:46<9:14:15, 13.44s/it]INFO 04-12 19:02:12 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 827/3300 [3:03:59<9:11:54, 13.39s/it]2026-04-12 19:02:25,783 | INFO | train_grpo_train | metrics_logged mode=train step=827

                                                      
{'loss': -0.1259, 'grad_norm': 3.371136426925659, 'learning_rate': 7.496969696969698e-06, 'num_tokens': 1482059.0, 'completions/mean_length': 167.875, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.16666793823242, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.649419903755188, 'rewards/meter/std': 0.3531789481639862, 'rewards/count_adherence/mean': 0.4750000238418579, 'rewards/count_adherence/std': 0.2121320515871048, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8424999713897705, 'rewards/count_floor/std': 0.06363961845636368, 'rewards/lexical_plausibility/mean': 0.896161675453186, 'rewards/lexical_plausibility/std': 0.19566257297992706, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/repeat_penalty/mean': 0.8028115034103394, 'rewards/repeat_penalty/std': 0.18067368865013123, 'rewards/repeat_floor/mean': 0.9785290360450745, 'rewards/repeat_floor/std': 0.0209575854241848, 'rewards/near_duplicate_penalty/mean': 0.9310684204101562, 'rewards/near_duplicate_penalty/std': 0.07439356297254562, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9179863929748535, 'rewards/distinct_2/std': 0.13890817761421204, 'rewards/total_composite/mean': 0.15409380197525024, 'rewards/total_composite/std': 0.1144634485244751, 'reward': 0.15409380197525024, 'reward_std': 0.1144634485244751, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17851325869560242, 'sampling/sampling_logp_difference/max': 1.5433738231658936, 'sampling/importance_sampling_ratio/min': 0.25876614451408386, 'sampling/importance_sampling_ratio/mean': 1.0627481937408447, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2568029761314392, 'clip_ratio/low_mean': 0.0489986059255898, 'clip_ratio/low_min': 0.0489986059255898, 'clip_ratio/high_mean': 0.06642530579119921, 'clip_ratio/high_max': 0.06642530579119921, 'clip_ratio/region_mean': 0.11542391171678901, 'reward_total_mean': 0.15409380197525024, 'reward_meter_mean': 0.649419903755188, 'reward_meter_std': 0.3531789481639862, 'reward_count_adherence_mean': 0.4750000238418579, 'reward_count_adherence_std': 0.2121320515871048, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8424999713897705, 'reward_count_floor_std': 0.06363961845636368, 'reward_lexical_plausibility_mean': 0.896161675453186, 'reward_lexical_plausibility_std': 0.19566257297992706, 'reward_repeat_penalty_mean': 0.8028115034103394, 'reward_repeat_penalty_std': 0.18067368865013123, 'reward_repeat_floor_mean': 0.9785290360450745, 'reward_repeat_floor_std': 0.0209575854241848, 'reward_near_duplicate_penalty_mean': 0.9310684204101562, 'reward_near_duplicate_penalty_std': 0.07439356297254562, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9179863929748535, 'reward_distinct_2_std': 0.13890817761421204, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.15409380197525024, 'reward_total_composite_std': 0.1144634485244751, 'epoch': 0.04}

 25%|██▌       | 827/3300 [3:03:59<9:11:54, 13.39s/it]INFO 04-12 19:02:26 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 828/3300 [3:04:12<9:11:21, 13.38s/it]2026-04-12 19:02:39,148 | INFO | train_grpo_train | metrics_logged mode=train step=828

                                                      
{'loss': -0.0484, 'grad_norm': 4.912237167358398, 'learning_rate': 7.493939393939395e-06, 'num_tokens': 1483665.0, 'completions/mean_length': 100.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.3973170518875122, 'rewards/meter/std': 0.3878973424434662, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9215686917304993, 'rewards/lexical_plausibility/std': 0.15877632796764374, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.2721048891544342, 'rewards/repeat_penalty/mean': 0.9658288955688477, 'rewards/repeat_penalty/std': 0.03947308659553528, 'rewards/repeat_floor/mean': 0.9948743581771851, 'rewards/repeat_floor/std': 0.005920974537730217, 'rewards/near_duplicate_penalty/mean': 0.9658288955688477, 'rewards/near_duplicate_penalty/std': 0.03947308659553528, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15168119966983795, 'rewards/total_composite/std': 0.1791079193353653, 'reward': 0.15168119966983795, 'reward_std': 0.1791079193353653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16855451464653015, 'sampling/sampling_logp_difference/max': 1.5621957778930664, 'sampling/importance_sampling_ratio/min': 0.2096751630306244, 'sampling/importance_sampling_ratio/mean': 1.021122932434082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2351923137903214, 'clip_ratio/low_mean': 0.09063656628131866, 'clip_ratio/low_min': 0.09063656628131866, 'clip_ratio/high_mean': 0.04285037983208895, 'clip_ratio/high_max': 0.04285037983208895, 'clip_ratio/region_mean': 0.1334869461134076, 'reward_total_mean': 0.15168119966983795, 'reward_meter_mean': 0.3973170518875122, 'reward_meter_std': 0.3878973424434662, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9215686917304993, 'reward_lexical_plausibility_std': 0.15877632796764374, 'reward_repeat_penalty_mean': 0.9658288955688477, 'reward_repeat_penalty_std': 0.03947308659553528, 'reward_repeat_floor_mean': 0.9948743581771851, 'reward_repeat_floor_std': 0.005920974537730217, 'reward_near_duplicate_penalty_mean': 0.9658288955688477, 'reward_near_duplicate_penalty_std': 0.03947308659553528, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.2721048891544342, 'reward_total_composite_mean': 0.15168119966983795, 'reward_total_composite_std': 0.1791079193353653, 'epoch': 0.04}

 25%|██▌       | 828/3300 [3:04:13<9:11:21, 13.38s/it]INFO 04-12 19:02:39 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 829/3300 [3:04:26<9:10:43, 13.37s/it]2026-04-12 19:02:52,498 | INFO | train_grpo_train | metrics_logged mode=train step=829

                                                      
{'loss': -0.0172, 'grad_norm': 3.951854944229126, 'learning_rate': 7.490909090909092e-06, 'num_tokens': 1485180.0, 'completions/mean_length': 85.375, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.428571701049805, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8694266080856323, 'rewards/meter/std': 0.34493935108184814, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9419781565666199, 'rewards/lexical_plausibility/std': 0.16411055624485016, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.6539329290390015, 'rewards/repeat_penalty/std': 0.10353017598390579, 'rewards/repeat_floor/mean': 0.9666889905929565, 'rewards/repeat_floor/std': 0.01966666430234909, 'rewards/near_duplicate_penalty/mean': 0.8944945335388184, 'rewards/near_duplicate_penalty/std': 0.11550140380859375, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9723865985870361, 'rewards/distinct_2/std': 0.051303714513778687, 'rewards/total_composite/mean': 0.2763935625553131, 'rewards/total_composite/std': 0.23508165776729584, 'reward': 0.2763935625553131, 'reward_std': 0.23508165776729584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16341540217399597, 'sampling/sampling_logp_difference/max': 1.1937992572784424, 'sampling/importance_sampling_ratio/min': 0.3030676245689392, 'sampling/importance_sampling_ratio/mean': 1.0650129318237305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2643106579780579, 'clip_ratio/low_mean': 0.06530423276126385, 'clip_ratio/low_min': 0.06530423276126385, 'clip_ratio/high_mean': 0.08446558006107807, 'clip_ratio/high_max': 0.08446558006107807, 'clip_ratio/region_mean': 0.14976981282234192, 'reward_total_mean': 0.2763935625553131, 'reward_meter_mean': 0.8694266080856323, 'reward_meter_std': 0.34493935108184814, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9419781565666199, 'reward_lexical_plausibility_std': 0.16411055624485016, 'reward_repeat_penalty_mean': 0.6539329290390015, 'reward_repeat_penalty_std': 0.10353017598390579, 'reward_repeat_floor_mean': 0.9666889905929565, 'reward_repeat_floor_std': 0.01966666430234909, 'reward_near_duplicate_penalty_mean': 0.8944945335388184, 'reward_near_duplicate_penalty_std': 0.11550140380859375, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9723865985870361, 'reward_distinct_2_std': 0.051303714513778687, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.2763935625553131, 'reward_total_composite_std': 0.23508165776729584, 'epoch': 0.04}

 25%|██▌       | 829/3300 [3:04:26<9:10:43, 13.37s/it]INFO 04-12 19:02:52 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 830/3300 [3:04:33<7:59:55, 11.66s/it]2026-04-12 19:03:00,160 | INFO | train_grpo_train | metrics_logged mode=train step=830

                                                      
{'loss': -0.0495, 'grad_norm': 15.028091430664062, 'learning_rate': 7.487878787878788e-06, 'num_tokens': 1486573.0, 'completions/mean_length': 30.125, 'completions/min_length': 18.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7560732364654541, 'rewards/meter/std': 0.38195741176605225, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.32275325059890747, 'rewards/repeat_penalty/mean': 0.8598579168319702, 'rewards/repeat_penalty/std': 0.11858725547790527, 'rewards/repeat_floor/mean': 0.9902286529541016, 'rewards/repeat_floor/std': 0.006109093781560659, 'rewards/near_duplicate_penalty/mean': 0.9848579168319702, 'rewards/near_duplicate_penalty/std': 0.023067161440849304, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4428815245628357, 'rewards/total_composite/std': 0.3357965052127838, 'reward': 0.4428815245628357, 'reward_std': 0.3357965052127838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18034382164478302, 'sampling/sampling_logp_difference/max': 2.771963357925415, 'sampling/importance_sampling_ratio/min': 0.06253910064697266, 'sampling/importance_sampling_ratio/mean': 1.0282163619995117, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4569306150078773, 'clip_ratio/low_mean': 0.10260999295860529, 'clip_ratio/low_min': 0.10260999295860529, 'clip_ratio/high_mean': 0.07610336877405643, 'clip_ratio/high_max': 0.07610336877405643, 'clip_ratio/region_mean': 0.17871336173266172, 'reward_total_mean': 0.4428815245628357, 'reward_meter_mean': 0.7560732364654541, 'reward_meter_std': 0.38195741176605225, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8598579168319702, 'reward_repeat_penalty_std': 0.11858725547790527, 'reward_repeat_floor_mean': 0.9902286529541016, 'reward_repeat_floor_std': 0.006109093781560659, 'reward_near_duplicate_penalty_mean': 0.9848579168319702, 'reward_near_duplicate_penalty_std': 0.023067161440849304, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6937500238418579, 'reward_judge_quality_std': 0.32275325059890747, 'reward_total_composite_mean': 0.4428815245628357, 'reward_total_composite_std': 0.3357965052127838, 'epoch': 0.04}

 25%|██▌       | 830/3300 [3:04:34<7:59:55, 11.66s/it]INFO 04-12 19:03:00 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 831/3300 [3:04:46<8:13:00, 11.98s/it]2026-04-12 19:03:12,901 | INFO | train_grpo_train | metrics_logged mode=train step=831

                                                      
{'loss': -0.0022, 'grad_norm': 4.951153755187988, 'learning_rate': 7.484848484848486e-06, 'num_tokens': 1487930.0, 'completions/mean_length': 82.625, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.285715103149414, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.4101417660713196, 'rewards/meter/std': 0.4823291301727295, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9279738068580627, 'rewards/lexical_plausibility/std': 0.17980211973190308, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.3414648175239563, 'rewards/repeat_penalty/mean': 0.7275309562683105, 'rewards/repeat_penalty/std': 0.06355194747447968, 'rewards/repeat_floor/mean': 0.980506181716919, 'rewards/repeat_floor/std': 0.012710394337773323, 'rewards/near_duplicate_penalty/mean': 0.9700413346290588, 'rewards/near_duplicate_penalty/std': 0.0847359448671341, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2975137233734131, 'rewards/total_composite/std': 0.38895219564437866, 'reward': 0.2975137233734131, 'reward_std': 0.38895219564437866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1963077336549759, 'sampling/sampling_logp_difference/max': 1.481719970703125, 'sampling/importance_sampling_ratio/min': 0.2272465080022812, 'sampling/importance_sampling_ratio/mean': 1.046219825744629, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4422542601823807, 'clip_ratio/low_mean': 0.07865530252456665, 'clip_ratio/low_min': 0.07865530252456665, 'clip_ratio/high_mean': 0.0575396828353405, 'clip_ratio/high_max': 0.0575396828353405, 'clip_ratio/region_mean': 0.13619498535990715, 'reward_total_mean': 0.2975137233734131, 'reward_meter_mean': 0.4101417660713196, 'reward_meter_std': 0.4823291301727295, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9279738068580627, 'reward_lexical_plausibility_std': 0.17980211973190308, 'reward_repeat_penalty_mean': 0.7275309562683105, 'reward_repeat_penalty_std': 0.06355194747447968, 'reward_repeat_floor_mean': 0.980506181716919, 'reward_repeat_floor_std': 0.012710394337773323, 'reward_near_duplicate_penalty_mean': 0.9700413346290588, 'reward_near_duplicate_penalty_std': 0.0847359448671341, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.3414648175239563, 'reward_total_composite_mean': 0.2975137233734131, 'reward_total_composite_std': 0.38895219564437866, 'epoch': 0.04}

 25%|██▌       | 831/3300 [3:04:46<8:13:00, 11.98s/it]INFO 04-12 19:03:13 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 832/3300 [3:04:54<7:25:28, 10.83s/it]2026-04-12 19:03:21,040 | INFO | train_grpo_train | metrics_logged mode=train step=832

                                                      
{'loss': 0.1042, 'grad_norm': 21.915010452270508, 'learning_rate': 7.481818181818182e-06, 'num_tokens': 1489385.0, 'completions/mean_length': 28.875, 'completions/min_length': 15.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7020453810691833, 'rewards/meter/std': 0.4006826877593994, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.8433008790016174, 'rewards/repeat_penalty/std': 0.11093761026859283, 'rewards/repeat_floor/mean': 0.9854938387870789, 'rewards/repeat_floor/std': 0.009750976227223873, 'rewards/near_duplicate_penalty/mean': 0.9518063068389893, 'rewards/near_duplicate_penalty/std': 0.05563024803996086, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9816433191299438, 'rewards/distinct_2/std': 0.03404124453663826, 'rewards/total_composite/mean': 0.2792567312717438, 'rewards/total_composite/std': 0.16881141066551208, 'reward': 0.2792567312717438, 'reward_std': 0.16881141066551208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19659152626991272, 'sampling/sampling_logp_difference/max': 1.4823060035705566, 'sampling/importance_sampling_ratio/min': 0.22711338102817535, 'sampling/importance_sampling_ratio/mean': 1.0340520143508911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2338368371129036, 'clip_ratio/low_mean': 0.05188230890780687, 'clip_ratio/low_min': 0.05188230890780687, 'clip_ratio/high_mean': 0.10351192578673363, 'clip_ratio/high_max': 0.10351192578673363, 'clip_ratio/region_mean': 0.1553942346945405, 'reward_total_mean': 0.2792567312717438, 'reward_meter_mean': 0.7020453810691833, 'reward_meter_std': 0.4006826877593994, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8433008790016174, 'reward_repeat_penalty_std': 0.11093761026859283, 'reward_repeat_floor_mean': 0.9854938387870789, 'reward_repeat_floor_std': 0.009750976227223873, 'reward_near_duplicate_penalty_mean': 0.9518063068389893, 'reward_near_duplicate_penalty_std': 0.05563024803996086, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9816433191299438, 'reward_distinct_2_std': 0.03404124453663826, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2792567312717438, 'reward_total_composite_std': 0.16881141066551208, 'epoch': 0.04}

 25%|██▌       | 832/3300 [3:04:54<7:25:28, 10.83s/it]INFO 04-12 19:03:21 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 833/3300 [3:05:08<7:57:45, 11.62s/it]2026-04-12 19:03:34,505 | INFO | train_grpo_train | metrics_logged mode=train step=833

                                                      
{'loss': -0.0401, 'grad_norm': 1.4332927465438843, 'learning_rate': 7.47878787878788e-06, 'num_tokens': 1491117.0, 'completions/mean_length': 273.5, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.894349217414856, 'rewards/meter/std': 0.19056905806064606, 'rewards/count_adherence/mean': 0.45000001788139343, 'rewards/count_adherence/std': 0.23299294710159302, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8349999785423279, 'rewards/count_floor/std': 0.06989789009094238, 'rewards/lexical_plausibility/mean': 0.8141950368881226, 'rewards/lexical_plausibility/std': 0.20056737959384918, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.8730959892272949, 'rewards/repeat_penalty/std': 0.1404113918542862, 'rewards/repeat_floor/mean': 0.9892293810844421, 'rewards/repeat_floor/std': 0.01229516789317131, 'rewards/near_duplicate_penalty/mean': 0.9713583588600159, 'rewards/near_duplicate_penalty/std': 0.04979393258690834, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.990562379360199, 'rewards/distinct_2/std': 0.02513767220079899, 'rewards/total_composite/mean': 0.1059030294418335, 'rewards/total_composite/std': 0.08801984786987305, 'reward': 0.1059030294418335, 'reward_std': 0.08801984041929245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18387942016124725, 'sampling/sampling_logp_difference/max': 0.9939770698547363, 'sampling/importance_sampling_ratio/min': 0.4053056538105011, 'sampling/importance_sampling_ratio/mean': 1.087215542793274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9216035902500153, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06770982779562473, 'clip_ratio/high_max': 0.06770982779562473, 'clip_ratio/region_mean': 0.06770982779562473, 'reward_total_mean': 0.1059030294418335, 'reward_meter_mean': 0.894349217414856, 'reward_meter_std': 0.19056905806064606, 'reward_count_adherence_mean': 0.45000001788139343, 'reward_count_adherence_std': 0.23299294710159302, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8349999785423279, 'reward_count_floor_std': 0.06989789009094238, 'reward_lexical_plausibility_mean': 0.8141950368881226, 'reward_lexical_plausibility_std': 0.20056737959384918, 'reward_repeat_penalty_mean': 0.8730959892272949, 'reward_repeat_penalty_std': 0.1404113918542862, 'reward_repeat_floor_mean': 0.9892293810844421, 'reward_repeat_floor_std': 0.01229516789317131, 'reward_near_duplicate_penalty_mean': 0.9713583588600159, 'reward_near_duplicate_penalty_std': 0.04979393258690834, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.990562379360199, 'reward_distinct_2_std': 0.02513767220079899, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.1059030294418335, 'reward_total_composite_std': 0.08801984786987305, 'epoch': 0.04}

 25%|██▌       | 833/3300 [3:05:08<7:57:45, 11.62s/it]INFO 04-12 19:03:34 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 834/3300 [3:05:23<8:38:54, 12.63s/it]2026-04-12 19:03:49,471 | INFO | train_grpo_train | metrics_logged mode=train step=834

                                                      
{'loss': -0.092, 'grad_norm': 7.298940658569336, 'learning_rate': 7.4757575757575765e-06, 'num_tokens': 1493009.0, 'completions/mean_length': 109.5, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.000003814697266, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6640127897262573, 'rewards/meter/std': 0.3662632405757904, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9455872178077698, 'rewards/lexical_plausibility/std': 0.15390260517597198, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1922609806060791, 'rewards/repeat_penalty/mean': 0.8738936185836792, 'rewards/repeat_penalty/std': 0.07922649383544922, 'rewards/repeat_floor/mean': 0.9838800430297852, 'rewards/repeat_floor/std': 0.009779521264135838, 'rewards/near_duplicate_penalty/mean': 0.9276025295257568, 'rewards/near_duplicate_penalty/std': 0.04741376265883446, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9415521025657654, 'rewards/distinct_2/std': 0.06024651601910591, 'rewards/total_composite/mean': 0.1933053433895111, 'rewards/total_composite/std': 0.1433020532131195, 'reward': 0.1933053433895111, 'reward_std': 0.1433020681142807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.156411275267601, 'sampling/sampling_logp_difference/max': 1.525448203086853, 'sampling/importance_sampling_ratio/min': 0.21752354502677917, 'sampling/importance_sampling_ratio/mean': 1.0328177213668823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9312049373984337, 'clip_ratio/low_mean': 0.0673882495611906, 'clip_ratio/low_min': 0.0673882495611906, 'clip_ratio/high_mean': 0.05761074647307396, 'clip_ratio/high_max': 0.05761074647307396, 'clip_ratio/region_mean': 0.12499899603426456, 'reward_total_mean': 0.1933053433895111, 'reward_meter_mean': 0.6640127897262573, 'reward_meter_std': 0.3662632405757904, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9455872178077698, 'reward_lexical_plausibility_std': 0.15390260517597198, 'reward_repeat_penalty_mean': 0.8738936185836792, 'reward_repeat_penalty_std': 0.07922649383544922, 'reward_repeat_floor_mean': 0.9838800430297852, 'reward_repeat_floor_std': 0.009779521264135838, 'reward_near_duplicate_penalty_mean': 0.9276025295257568, 'reward_near_duplicate_penalty_std': 0.04741376265883446, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9415521025657654, 'reward_distinct_2_std': 0.06024651601910591, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1922609806060791, 'reward_total_composite_mean': 0.1933053433895111, 'reward_total_composite_std': 0.1433020532131195, 'epoch': 0.04}

 25%|██▌       | 834/3300 [3:05:23<8:38:54, 12.63s/it]INFO 04-12 19:03:49 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 835/3300 [3:05:37<8:56:13, 13.05s/it]2026-04-12 19:04:03,520 | INFO | train_grpo_train | metrics_logged mode=train step=835

                                                      
{'loss': -0.0257, 'grad_norm': 3.7628729343414307, 'learning_rate': 7.472727272727274e-06, 'num_tokens': 1494785.0, 'completions/mean_length': 92.0, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9261429905891418, 'rewards/meter/std': 0.1502811312675476, 'rewards/count_adherence/mean': 0.2750000059604645, 'rewards/count_adherence/std': 0.2121320366859436, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7825000286102295, 'rewards/count_floor/std': 0.06363961845636368, 'rewards/lexical_plausibility/mean': 0.9328755140304565, 'rewards/lexical_plausibility/std': 0.18985673785209656, 'rewards/judge_quality/mean': 0.4674999713897705, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.7128247618675232, 'rewards/repeat_penalty/std': 0.05981914699077606, 'rewards/repeat_floor/mean': 0.9769024848937988, 'rewards/repeat_floor/std': 0.011665260419249535, 'rewards/near_duplicate_penalty/mean': 0.9585079550743103, 'rewards/near_duplicate_penalty/std': 0.05751679837703705, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9583480954170227, 'rewards/distinct_2/std': 0.08286898583173752, 'rewards/total_composite/mean': 0.3452942371368408, 'rewards/total_composite/std': 0.2219136357307434, 'reward': 0.3452942371368408, 'reward_std': 0.22191362082958221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13392537832260132, 'sampling/sampling_logp_difference/max': 1.2734620571136475, 'sampling/importance_sampling_ratio/min': 0.2798610329627991, 'sampling/importance_sampling_ratio/mean': 1.0565990209579468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8899519741535187, 'clip_ratio/low_mean': 0.10612781345844269, 'clip_ratio/low_min': 0.10612781345844269, 'clip_ratio/high_mean': 0.030208333861082792, 'clip_ratio/high_max': 0.030208333861082792, 'clip_ratio/region_mean': 0.13633614731952548, 'reward_total_mean': 0.3452942371368408, 'reward_meter_mean': 0.9261429905891418, 'reward_meter_std': 0.1502811312675476, 'reward_count_adherence_mean': 0.2750000059604645, 'reward_count_adherence_std': 0.2121320366859436, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7825000286102295, 'reward_count_floor_std': 0.06363961845636368, 'reward_lexical_plausibility_mean': 0.9328755140304565, 'reward_lexical_plausibility_std': 0.18985673785209656, 'reward_repeat_penalty_mean': 0.7128247618675232, 'reward_repeat_penalty_std': 0.05981914699077606, 'reward_repeat_floor_mean': 0.9769024848937988, 'reward_repeat_floor_std': 0.011665260419249535, 'reward_near_duplicate_penalty_mean': 0.9585079550743103, 'reward_near_duplicate_penalty_std': 0.05751679837703705, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9583480954170227, 'reward_distinct_2_std': 0.08286898583173752, 'reward_judge_quality_mean': 0.4674999713897705, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.3452942371368408, 'reward_total_composite_std': 0.2219136357307434, 'epoch': 0.04}

 25%|██▌       | 835/3300 [3:05:37<8:56:13, 13.05s/it]INFO 04-12 19:04:03 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 836/3300 [3:05:45<7:57:08, 11.62s/it]2026-04-12 19:04:11,797 | INFO | train_grpo_train | metrics_logged mode=train step=836

                                                      
{'loss': 0.0434, 'grad_norm': 15.561553001403809, 'learning_rate': 7.46969696969697e-06, 'num_tokens': 1496297.0, 'completions/mean_length': 35.0, 'completions/min_length': 22.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.859635591506958, 'rewards/meter/std': 0.3367502987384796, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.27969372272491455, 'rewards/repeat_penalty/mean': 0.9460607767105103, 'rewards/repeat_penalty/std': 0.0891919881105423, 'rewards/repeat_floor/mean': 0.9952001571655273, 'rewards/repeat_floor/std': 0.006399350706487894, 'rewards/near_duplicate_penalty/mean': 0.9859956502914429, 'rewards/near_duplicate_penalty/std': 0.020329361781477928, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.42820051312446594, 'rewards/total_composite/std': 0.3306257426738739, 'reward': 0.42820051312446594, 'reward_std': 0.3306257426738739, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16682808101177216, 'sampling/sampling_logp_difference/max': 1.4321937561035156, 'sampling/importance_sampling_ratio/min': 0.238784521818161, 'sampling/importance_sampling_ratio/mean': 1.0416346788406372, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.441406436264515, 'clip_ratio/low_mean': 0.06633158819749951, 'clip_ratio/low_min': 0.06633158819749951, 'clip_ratio/high_mean': 0.07771610654890537, 'clip_ratio/high_max': 0.07771610654890537, 'clip_ratio/region_mean': 0.1440476947464049, 'reward_total_mean': 0.42820051312446594, 'reward_meter_mean': 0.859635591506958, 'reward_meter_std': 0.3367502987384796, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9460607767105103, 'reward_repeat_penalty_std': 0.0891919881105423, 'reward_repeat_floor_mean': 0.9952001571655273, 'reward_repeat_floor_std': 0.006399350706487894, 'reward_near_duplicate_penalty_mean': 0.9859956502914429, 'reward_near_duplicate_penalty_std': 0.020329361781477928, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.27969372272491455, 'reward_total_composite_mean': 0.42820051312446594, 'reward_total_composite_std': 0.3306257426738739, 'epoch': 0.04}

 25%|██▌       | 836/3300 [3:05:45<7:57:08, 11.62s/it]INFO 04-12 19:04:12 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 837/3300 [3:05:58<8:17:56, 12.13s/it]2026-04-12 19:04:25,124 | INFO | train_grpo_train | metrics_logged mode=train step=837

                                                      
{'loss': -0.0316, 'grad_norm': 5.0076398849487305, 'learning_rate': 7.4666666666666675e-06, 'num_tokens': 1497699.0, 'completions/mean_length': 86.25, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.428571701049805, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.863968014717102, 'rewards/meter/std': 0.3293401300907135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9573074579238892, 'rewards/lexical_plausibility/std': 0.1207527220249176, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.3882538676261902, 'rewards/repeat_penalty/mean': 0.7407181262969971, 'rewards/repeat_penalty/std': 0.023893872275948524, 'rewards/repeat_floor/mean': 0.9832043647766113, 'rewards/repeat_floor/std': 0.0047911605797708035, 'rewards/near_duplicate_penalty/mean': 0.9886363744735718, 'rewards/near_duplicate_penalty/std': 0.03214120864868164, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9989878535270691, 'rewards/distinct_2/std': 0.002862782683223486, 'rewards/total_composite/mean': 0.5511993169784546, 'rewards/total_composite/std': 0.38762637972831726, 'reward': 0.5511993169784546, 'reward_std': 0.38762637972831726, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19810850918293, 'sampling/sampling_logp_difference/max': 1.3831443786621094, 'sampling/importance_sampling_ratio/min': 0.25078874826431274, 'sampling/importance_sampling_ratio/mean': 1.0525456666946411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4116116017103195, 'clip_ratio/low_mean': 0.0786894503980875, 'clip_ratio/low_min': 0.0786894503980875, 'clip_ratio/high_mean': 0.08216579630970955, 'clip_ratio/high_max': 0.08216579630970955, 'clip_ratio/region_mean': 0.16085524670779705, 'reward_total_mean': 0.5511993169784546, 'reward_meter_mean': 0.863968014717102, 'reward_meter_std': 0.3293401300907135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9573074579238892, 'reward_lexical_plausibility_std': 0.1207527220249176, 'reward_repeat_penalty_mean': 0.7407181262969971, 'reward_repeat_penalty_std': 0.023893872275948524, 'reward_repeat_floor_mean': 0.9832043647766113, 'reward_repeat_floor_std': 0.0047911605797708035, 'reward_near_duplicate_penalty_mean': 0.9886363744735718, 'reward_near_duplicate_penalty_std': 0.03214120864868164, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9989878535270691, 'reward_distinct_2_std': 0.002862782683223486, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.3882538676261902, 'reward_total_composite_mean': 0.5511993169784546, 'reward_total_composite_std': 0.38762637972831726, 'epoch': 0.04}

 25%|██▌       | 837/3300 [3:05:58<8:17:56, 12.13s/it]INFO 04-12 19:04:25 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 838/3300 [3:06:12<8:33:53, 12.52s/it]2026-04-12 19:04:38,565 | INFO | train_grpo_train | metrics_logged mode=train step=838

                                                      
{'loss': -0.1328, 'grad_norm': 2.026712417602539, 'learning_rate': 7.463636363636364e-06, 'num_tokens': 1499613.0, 'completions/mean_length': 109.25, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.71428680419922, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.8534479141235352, 'rewards/meter/std': 0.346881240606308, 'rewards/count_adherence/mean': 0.2916666865348816, 'rewards/count_adherence/std': 0.2920915186405182, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.7875000238418579, 'rewards/count_floor/std': 0.08762745559215546, 'rewards/lexical_plausibility/mean': 0.9461221694946289, 'rewards/lexical_plausibility/std': 0.15238943696022034, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.6931389570236206, 'rewards/repeat_penalty/std': 0.22456727921962738, 'rewards/repeat_floor/mean': 0.9684528112411499, 'rewards/repeat_floor/std': 0.037759359925985336, 'rewards/near_duplicate_penalty/mean': 0.913510799407959, 'rewards/near_duplicate_penalty/std': 0.1730053871870041, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8977908492088318, 'rewards/distinct_2/std': 0.15520550310611725, 'rewards/total_composite/mean': 0.25702720880508423, 'rewards/total_composite/std': 0.10999058932065964, 'reward': 0.25702720880508423, 'reward_std': 0.10999057441949844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14010226726531982, 'sampling/sampling_logp_difference/max': 1.5988950729370117, 'sampling/importance_sampling_ratio/min': 0.20211972296237946, 'sampling/importance_sampling_ratio/mean': 1.0227127075195312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.005738541483879, 'clip_ratio/low_mean': 0.020208333153277636, 'clip_ratio/low_min': 0.020208333153277636, 'clip_ratio/high_mean': 0.06495158607140183, 'clip_ratio/high_max': 0.06495158607140183, 'clip_ratio/region_mean': 0.08515991922467947, 'reward_total_mean': 0.25702720880508423, 'reward_meter_mean': 0.8534479141235352, 'reward_meter_std': 0.346881240606308, 'reward_count_adherence_mean': 0.2916666865348816, 'reward_count_adherence_std': 0.2920915186405182, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.7875000238418579, 'reward_count_floor_std': 0.08762745559215546, 'reward_lexical_plausibility_mean': 0.9461221694946289, 'reward_lexical_plausibility_std': 0.15238943696022034, 'reward_repeat_penalty_mean': 0.6931389570236206, 'reward_repeat_penalty_std': 0.22456727921962738, 'reward_repeat_floor_mean': 0.9684528112411499, 'reward_repeat_floor_std': 0.037759359925985336, 'reward_near_duplicate_penalty_mean': 0.913510799407959, 'reward_near_duplicate_penalty_std': 0.1730053871870041, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8977908492088318, 'reward_distinct_2_std': 0.15520550310611725, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.25702720880508423, 'reward_total_composite_std': 0.10999058932065964, 'epoch': 0.04}

 25%|██▌       | 838/3300 [3:06:12<8:33:53, 12.52s/it]INFO 04-12 19:04:38 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 839/3300 [3:06:26<8:49:13, 12.90s/it]2026-04-12 19:04:52,345 | INFO | train_grpo_train | metrics_logged mode=train step=839

                                                      
{'loss': -0.0398, 'grad_norm': 2.6468305587768555, 'learning_rate': 7.460606060606061e-06, 'num_tokens': 1501046.0, 'completions/mean_length': 148.125, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 26.83333396911621, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.5430874824523926, 'rewards/meter/std': 0.42536190152168274, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.11338934302330017, 'rewards/lexical_plausibility/mean': 0.8956800103187561, 'rewards/lexical_plausibility/std': 0.19437187910079956, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.3543177545070648, 'rewards/repeat_penalty/mean': 0.9212819337844849, 'rewards/repeat_penalty/std': 0.10672668367624283, 'rewards/repeat_floor/mean': 0.9938173294067383, 'rewards/repeat_floor/std': 0.005868623033165932, 'rewards/near_duplicate_penalty/mean': 0.9837819337844849, 'rewards/near_duplicate_penalty/std': 0.017736751586198807, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1994948834180832, 'rewards/total_composite/std': 0.1772877722978592, 'reward': 0.1994948834180832, 'reward_std': 0.1772877722978592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1593082994222641, 'sampling/sampling_logp_difference/max': 1.3116185665130615, 'sampling/importance_sampling_ratio/min': 0.26938366889953613, 'sampling/importance_sampling_ratio/mean': 1.0183340311050415, 'sampling/importance_sampling_ratio/max': 1.8916113376617432, 'entropy': 0.7173252999782562, 'clip_ratio/low_mean': 0.031547619961202145, 'clip_ratio/low_min': 0.031547619961202145, 'clip_ratio/high_mean': 0.09393939655274153, 'clip_ratio/high_max': 0.09393939655274153, 'clip_ratio/region_mean': 0.12548701651394367, 'reward_total_mean': 0.1994948834180832, 'reward_meter_mean': 0.5430874824523926, 'reward_meter_std': 0.42536190152168274, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.11338934302330017, 'reward_lexical_plausibility_mean': 0.8956800103187561, 'reward_lexical_plausibility_std': 0.19437187910079956, 'reward_repeat_penalty_mean': 0.9212819337844849, 'reward_repeat_penalty_std': 0.10672668367624283, 'reward_repeat_floor_mean': 0.9938173294067383, 'reward_repeat_floor_std': 0.005868623033165932, 'reward_near_duplicate_penalty_mean': 0.9837819337844849, 'reward_near_duplicate_penalty_std': 0.017736751586198807, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.3543177545070648, 'reward_total_composite_mean': 0.1994948834180832, 'reward_total_composite_std': 0.1772877722978592, 'epoch': 0.04}

 25%|██▌       | 839/3300 [3:06:26<8:49:13, 12.90s/it]INFO 04-12 19:04:52 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 840/3300 [3:06:39<8:55:56, 13.07s/it]2026-04-12 19:05:05,811 | INFO | train_grpo_train | metrics_logged mode=train step=840

                                                      
{'loss': -0.0813, 'grad_norm': 2.4004197120666504, 'learning_rate': 7.4575757575757575e-06, 'num_tokens': 1502738.0, 'completions/mean_length': 153.5, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7789326906204224, 'rewards/meter/std': 0.39128315448760986, 'rewards/count_adherence/mean': 0.40625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8218749761581421, 'rewards/count_floor/std': 0.05580180510878563, 'rewards/lexical_plausibility/mean': 0.8987300395965576, 'rewards/lexical_plausibility/std': 0.18911586701869965, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.8428071737289429, 'rewards/repeat_penalty/std': 0.1011408269405365, 'rewards/repeat_floor/mean': 0.988522469997406, 'rewards/repeat_floor/std': 0.004091145470738411, 'rewards/near_duplicate_penalty/mean': 0.9826459884643555, 'rewards/near_duplicate_penalty/std': 0.019330358132719994, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9847284555435181, 'rewards/distinct_2/std': 0.029517238959670067, 'rewards/total_composite/mean': 0.20652464032173157, 'rewards/total_composite/std': 0.14455179870128632, 'reward': 0.20652464032173157, 'reward_std': 0.14455178380012512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16030967235565186, 'sampling/sampling_logp_difference/max': 1.276998519897461, 'sampling/importance_sampling_ratio/min': 0.2788730561733246, 'sampling/importance_sampling_ratio/mean': 0.9825403690338135, 'sampling/importance_sampling_ratio/max': 1.8128416538238525, 'entropy': 0.8350357115268707, 'clip_ratio/low_mean': 0.012195121496915817, 'clip_ratio/low_min': 0.012195121496915817, 'clip_ratio/high_mean': 0.11779505293816328, 'clip_ratio/high_max': 0.11779505293816328, 'clip_ratio/region_mean': 0.1299901744350791, 'reward_total_mean': 0.20652464032173157, 'reward_meter_mean': 0.7789326906204224, 'reward_meter_std': 0.39128315448760986, 'reward_count_adherence_mean': 0.40625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8218749761581421, 'reward_count_floor_std': 0.05580180510878563, 'reward_lexical_plausibility_mean': 0.8987300395965576, 'reward_lexical_plausibility_std': 0.18911586701869965, 'reward_repeat_penalty_mean': 0.8428071737289429, 'reward_repeat_penalty_std': 0.1011408269405365, 'reward_repeat_floor_mean': 0.988522469997406, 'reward_repeat_floor_std': 0.004091145470738411, 'reward_near_duplicate_penalty_mean': 0.9826459884643555, 'reward_near_duplicate_penalty_std': 0.019330358132719994, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9847284555435181, 'reward_distinct_2_std': 0.029517238959670067, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.20652464032173157, 'reward_total_composite_std': 0.14455179870128632, 'epoch': 0.04}

 25%|██▌       | 840/3300 [3:06:39<8:55:56, 13.07s/it]INFO 04-12 19:05:06 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 841/3300 [3:06:47<7:57:14, 11.64s/it]2026-04-12 19:05:14,127 | INFO | train_grpo_train | metrics_logged mode=train step=841

                                                      
{'loss': 0.0059, 'grad_norm': 16.900815963745117, 'learning_rate': 7.454545454545456e-06, 'num_tokens': 1504198.0, 'completions/mean_length': 24.5, 'completions/min_length': 21.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.8961358070373535, 'rewards/meter/std': 0.26502135396003723, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.6599999666213989, 'rewards/judge_quality/std': 0.30956190824508667, 'rewards/repeat_penalty/mean': 0.7792832255363464, 'rewards/repeat_penalty/std': 0.08935283869504929, 'rewards/repeat_floor/mean': 0.9864816665649414, 'rewards/repeat_floor/std': 0.005572130437940359, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5365504026412964, 'rewards/total_composite/std': 0.28162628412246704, 'reward': 0.5365504026412964, 'reward_std': 0.28162631392478943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13926811516284943, 'sampling/sampling_logp_difference/max': 1.2954695224761963, 'sampling/importance_sampling_ratio/min': 0.2737692892551422, 'sampling/importance_sampling_ratio/mean': 1.0077086687088013, 'sampling/importance_sampling_ratio/max': 1.8900736570358276, 'entropy': 1.2510548532009125, 'clip_ratio/low_mean': 0.04569361265748739, 'clip_ratio/low_min': 0.04569361265748739, 'clip_ratio/high_mean': 0.07435035659000278, 'clip_ratio/high_max': 0.07435035659000278, 'clip_ratio/region_mean': 0.12004396924749017, 'reward_total_mean': 0.5365504026412964, 'reward_meter_mean': 0.8961358070373535, 'reward_meter_std': 0.26502135396003723, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.7792832255363464, 'reward_repeat_penalty_std': 0.08935283869504929, 'reward_repeat_floor_mean': 0.9864816665649414, 'reward_repeat_floor_std': 0.005572130437940359, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6599999666213989, 'reward_judge_quality_std': 0.30956190824508667, 'reward_total_composite_mean': 0.5365504026412964, 'reward_total_composite_std': 0.28162628412246704, 'epoch': 0.04}

 25%|██▌       | 841/3300 [3:06:47<7:57:14, 11.64s/it]INFO 04-12 19:05:14 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 842/3300 [3:07:01<8:25:18, 12.33s/it]2026-04-12 19:05:28,070 | INFO | train_grpo_train | metrics_logged mode=train step=842

                                                      
{'loss': -0.1469, 'grad_norm': 3.475796937942505, 'learning_rate': 7.451515151515152e-06, 'num_tokens': 1506110.0, 'completions/mean_length': 104.0, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.71428680419922, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8337839841842651, 'rewards/meter/std': 0.31833702325820923, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.22160132229328156, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.06648041307926178, 'rewards/lexical_plausibility/mean': 0.9493201971054077, 'rewards/lexical_plausibility/std': 0.1433442085981369, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.18850918114185333, 'rewards/repeat_penalty/mean': 0.8963936567306519, 'rewards/repeat_penalty/std': 0.11243993043899536, 'rewards/repeat_floor/mean': 0.9904098510742188, 'rewards/repeat_floor/std': 0.008146839216351509, 'rewards/near_duplicate_penalty/mean': 0.9678531289100647, 'rewards/near_duplicate_penalty/std': 0.02587353065609932, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.29200267791748047, 'rewards/total_composite/std': 0.16657792031764984, 'reward': 0.29200267791748047, 'reward_std': 0.16657790541648865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14423060417175293, 'sampling/sampling_logp_difference/max': 1.0531961917877197, 'sampling/importance_sampling_ratio/min': 0.3488210439682007, 'sampling/importance_sampling_ratio/mean': 1.0366005897521973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1190107315778732, 'clip_ratio/low_mean': 0.0438988097012043, 'clip_ratio/low_min': 0.0438988097012043, 'clip_ratio/high_mean': 0.08323159534484148, 'clip_ratio/high_max': 0.08323159534484148, 'clip_ratio/region_mean': 0.12713040504604578, 'reward_total_mean': 0.29200267791748047, 'reward_meter_mean': 0.8337839841842651, 'reward_meter_std': 0.31833702325820923, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.22160132229328156, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.06648041307926178, 'reward_lexical_plausibility_mean': 0.9493201971054077, 'reward_lexical_plausibility_std': 0.1433442085981369, 'reward_repeat_penalty_mean': 0.8963936567306519, 'reward_repeat_penalty_std': 0.11243993043899536, 'reward_repeat_floor_mean': 0.9904098510742188, 'reward_repeat_floor_std': 0.008146839216351509, 'reward_near_duplicate_penalty_mean': 0.9678531289100647, 'reward_near_duplicate_penalty_std': 0.02587353065609932, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.18850918114185333, 'reward_total_composite_mean': 0.29200267791748047, 'reward_total_composite_std': 0.16657792031764984, 'epoch': 0.04}

 26%|██▌       | 842/3300 [3:07:01<8:25:18, 12.33s/it]INFO 04-12 19:05:28 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 843/3300 [3:07:15<8:46:05, 12.85s/it]2026-04-12 19:05:42,119 | INFO | train_grpo_train | metrics_logged mode=train step=843

                                                      
{'loss': -0.0795, 'grad_norm': 3.2899844646453857, 'learning_rate': 7.448484848484849e-06, 'num_tokens': 1507628.0, 'completions/mean_length': 87.75, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 27.142858505249023, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.8925532102584839, 'rewards/meter/std': 0.2252667397260666, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9658455848693848, 'rewards/lexical_plausibility/std': 0.09660319238901138, 'rewards/judge_quality/mean': 0.7524999976158142, 'rewards/judge_quality/std': 0.32805708050727844, 'rewards/repeat_penalty/mean': 0.7741244435310364, 'rewards/repeat_penalty/std': 0.09342192858457565, 'rewards/repeat_floor/mean': 0.9854499101638794, 'rewards/repeat_floor/std': 0.0071049463003873825, 'rewards/near_duplicate_penalty/mean': 0.9904992580413818, 'rewards/near_duplicate_penalty/std': 0.02687215805053711, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6136301755905151, 'rewards/total_composite/std': 0.28744563460350037, 'reward': 0.6136301755905151, 'reward_std': 0.28744563460350037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17705973982810974, 'sampling/sampling_logp_difference/max': 2.3522167205810547, 'sampling/importance_sampling_ratio/min': 0.09515798836946487, 'sampling/importance_sampling_ratio/mean': 1.0532846450805664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0713282823562622, 'clip_ratio/low_mean': 0.024038461968302727, 'clip_ratio/low_min': 0.024038461968302727, 'clip_ratio/high_mean': 0.14989417977631092, 'clip_ratio/high_max': 0.14989417977631092, 'clip_ratio/region_mean': 0.17393264174461365, 'reward_total_mean': 0.6136301755905151, 'reward_meter_mean': 0.8925532102584839, 'reward_meter_std': 0.2252667397260666, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9658455848693848, 'reward_lexical_plausibility_std': 0.09660319238901138, 'reward_repeat_penalty_mean': 0.7741244435310364, 'reward_repeat_penalty_std': 0.09342192858457565, 'reward_repeat_floor_mean': 0.9854499101638794, 'reward_repeat_floor_std': 0.0071049463003873825, 'reward_near_duplicate_penalty_mean': 0.9904992580413818, 'reward_near_duplicate_penalty_std': 0.02687215805053711, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7524999976158142, 'reward_judge_quality_std': 0.32805708050727844, 'reward_total_composite_mean': 0.6136301755905151, 'reward_total_composite_std': 0.28744563460350037, 'epoch': 0.04}

 26%|██▌       | 843/3300 [3:07:15<8:46:05, 12.85s/it]INFO 04-12 19:05:42 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 844/3300 [3:07:29<8:55:10, 13.07s/it]2026-04-12 19:05:55,718 | INFO | train_grpo_train | metrics_logged mode=train step=844

                                                      
{'loss': -0.0654, 'grad_norm': 3.3934338092803955, 'learning_rate': 7.445454545454546e-06, 'num_tokens': 1509111.0, 'completions/mean_length': 85.375, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.428571701049805, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.859717845916748, 'rewards/meter/std': 0.34734803438186646, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9550939798355103, 'rewards/lexical_plausibility/std': 0.1270134598016739, 'rewards/judge_quality/mean': 0.5637499690055847, 'rewards/judge_quality/std': 0.32275325059890747, 'rewards/repeat_penalty/mean': 0.6278575658798218, 'rewards/repeat_penalty/std': 0.2103126049041748, 'rewards/repeat_floor/mean': 0.9592983722686768, 'rewards/repeat_floor/std': 0.03865781053900719, 'rewards/near_duplicate_penalty/mean': 0.8725221157073975, 'rewards/near_duplicate_penalty/std': 0.14677858352661133, 'rewards/opening_diversity/mean': 0.734375, 'rewards/opening_diversity/std': 0.16952534019947052, 'rewards/distinct_2/mean': 0.93730628490448, 'rewards/distinct_2/std': 0.12876898050308228, 'rewards/total_composite/mean': 0.4694211483001709, 'rewards/total_composite/std': 0.29148513078689575, 'reward': 0.4694211483001709, 'reward_std': 0.29148513078689575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10539094358682632, 'sampling/sampling_logp_difference/max': 1.079148292541504, 'sampling/importance_sampling_ratio/min': 0.339884877204895, 'sampling/importance_sampling_ratio/mean': 1.0035425424575806, 'sampling/importance_sampling_ratio/max': 1.9466650485992432, 'entropy': 0.7305092141032219, 'clip_ratio/low_mean': 0.043347952887415886, 'clip_ratio/low_min': 0.043347952887415886, 'clip_ratio/high_mean': 0.04401119751855731, 'clip_ratio/high_max': 0.04401119751855731, 'clip_ratio/region_mean': 0.0873591504059732, 'reward_total_mean': 0.4694211483001709, 'reward_meter_mean': 0.859717845916748, 'reward_meter_std': 0.34734803438186646, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9550939798355103, 'reward_lexical_plausibility_std': 0.1270134598016739, 'reward_repeat_penalty_mean': 0.6278575658798218, 'reward_repeat_penalty_std': 0.2103126049041748, 'reward_repeat_floor_mean': 0.9592983722686768, 'reward_repeat_floor_std': 0.03865781053900719, 'reward_near_duplicate_penalty_mean': 0.8725221157073975, 'reward_near_duplicate_penalty_std': 0.14677858352661133, 'reward_opening_diversity_mean': 0.734375, 'reward_opening_diversity_std': 0.16952534019947052, 'reward_distinct_2_mean': 0.93730628490448, 'reward_distinct_2_std': 0.12876898050308228, 'reward_judge_quality_mean': 0.5637499690055847, 'reward_judge_quality_std': 0.32275325059890747, 'reward_total_composite_mean': 0.4694211483001709, 'reward_total_composite_std': 0.29148513078689575, 'epoch': 0.04}

 26%|██▌       | 844/3300 [3:07:29<8:55:10, 13.07s/it]INFO 04-12 19:05:55 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 845/3300 [3:07:43<9:00:50, 13.22s/it]2026-04-12 19:06:09,272 | INFO | train_grpo_train | metrics_logged mode=train step=845

                                                      
{'loss': -0.0596, 'grad_norm': 2.4339377880096436, 'learning_rate': 7.442424242424243e-06, 'num_tokens': 1510783.0, 'completions/mean_length': 167.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.2763864994049072, 'rewards/meter/std': 0.2573143243789673, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.882985532283783, 'rewards/lexical_plausibility/std': 0.16474157571792603, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9563765525817871, 'rewards/repeat_penalty/std': 0.06135902926325798, 'rewards/repeat_floor/mean': 0.9949256777763367, 'rewards/repeat_floor/std': 0.006504016928374767, 'rewards/near_duplicate_penalty/mean': 0.9824104905128479, 'rewards/near_duplicate_penalty/std': 0.014516990631818771, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9729344844818115, 'rewards/distinct_2/std': 0.05020807683467865, 'rewards/total_composite/mean': 0.05921619385480881, 'rewards/total_composite/std': 0.04154956713318825, 'reward': 0.05921619385480881, 'reward_std': 0.04154956713318825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13828082382678986, 'sampling/sampling_logp_difference/max': 1.2101831436157227, 'sampling/importance_sampling_ratio/min': 0.298142671585083, 'sampling/importance_sampling_ratio/mean': 1.0267951488494873, 'sampling/importance_sampling_ratio/max': 1.8265711069107056, 'entropy': 0.9513723403215408, 'clip_ratio/low_mean': 0.037676410749554634, 'clip_ratio/low_min': 0.037676410749554634, 'clip_ratio/high_mean': 0.0447845128364861, 'clip_ratio/high_max': 0.0447845128364861, 'clip_ratio/region_mean': 0.08246092358604074, 'reward_total_mean': 0.05921619385480881, 'reward_meter_mean': 0.2763864994049072, 'reward_meter_std': 0.2573143243789673, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.882985532283783, 'reward_lexical_plausibility_std': 0.16474157571792603, 'reward_repeat_penalty_mean': 0.9563765525817871, 'reward_repeat_penalty_std': 0.06135902926325798, 'reward_repeat_floor_mean': 0.9949256777763367, 'reward_repeat_floor_std': 0.006504016928374767, 'reward_near_duplicate_penalty_mean': 0.9824104905128479, 'reward_near_duplicate_penalty_std': 0.014516990631818771, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9729344844818115, 'reward_distinct_2_std': 0.05020807683467865, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.05921619385480881, 'reward_total_composite_std': 0.04154956713318825, 'epoch': 0.04}

 26%|██▌       | 845/3300 [3:07:43<9:00:50, 13.22s/it]INFO 04-12 19:06:09 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 846/3300 [3:07:51<7:57:06, 11.67s/it]2026-04-12 19:06:17,317 | INFO | train_grpo_train | metrics_logged mode=train step=846

                                                      
{'loss': 0.0301, 'grad_norm': 18.674097061157227, 'learning_rate': 7.439393939393939e-06, 'num_tokens': 1512310.0, 'completions/mean_length': 27.875, 'completions/min_length': 15.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7191743850708008, 'rewards/meter/std': 0.29383939504623413, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5762500166893005, 'rewards/judge_quality/std': 0.30123260617256165, 'rewards/repeat_penalty/mean': 0.7390232086181641, 'rewards/repeat_penalty/std': 0.16861271858215332, 'rewards/repeat_floor/mean': 0.9765577912330627, 'rewards/repeat_floor/std': 0.01894167996942997, 'rewards/near_duplicate_penalty/mean': 0.9490758180618286, 'rewards/near_duplicate_penalty/std': 0.06547746807336807, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9494047164916992, 'rewards/distinct_2/std': 0.07473616302013397, 'rewards/total_composite/mean': 0.3512076139450073, 'rewards/total_composite/std': 0.23876644670963287, 'reward': 0.3512076139450073, 'reward_std': 0.23876643180847168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12690965831279755, 'sampling/sampling_logp_difference/max': 2.050009250640869, 'sampling/importance_sampling_ratio/min': 0.12873370945453644, 'sampling/importance_sampling_ratio/mean': 0.9970605969429016, 'sampling/importance_sampling_ratio/max': 1.8812642097473145, 'entropy': 0.7203164547681808, 'clip_ratio/low_mean': 0.07454182393848896, 'clip_ratio/low_min': 0.07454182393848896, 'clip_ratio/high_mean': 0.06880877818912268, 'clip_ratio/high_max': 0.06880877818912268, 'clip_ratio/region_mean': 0.14335060212761164, 'reward_total_mean': 0.3512076139450073, 'reward_meter_mean': 0.7191743850708008, 'reward_meter_std': 0.29383939504623413, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7390232086181641, 'reward_repeat_penalty_std': 0.16861271858215332, 'reward_repeat_floor_mean': 0.9765577912330627, 'reward_repeat_floor_std': 0.01894167996942997, 'reward_near_duplicate_penalty_mean': 0.9490758180618286, 'reward_near_duplicate_penalty_std': 0.06547746807336807, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9494047164916992, 'reward_distinct_2_std': 0.07473616302013397, 'reward_judge_quality_mean': 0.5762500166893005, 'reward_judge_quality_std': 0.30123260617256165, 'reward_total_composite_mean': 0.3512076139450073, 'reward_total_composite_std': 0.23876644670963287, 'epoch': 0.04}

 26%|██▌       | 846/3300 [3:07:51<7:57:06, 11.67s/it]INFO 04-12 19:06:17 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 847/3300 [3:07:59<7:13:53, 10.61s/it]2026-04-12 19:06:25,473 | INFO | train_grpo_train | metrics_logged mode=train step=847

                                                      
{'loss': 0.0944, 'grad_norm': 18.91111183166504, 'learning_rate': 7.4363636363636375e-06, 'num_tokens': 1513697.0, 'completions/mean_length': 25.375, 'completions/min_length': 17.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8375080823898315, 'rewards/meter/std': 0.32897061109542847, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.6353042721748352, 'rewards/repeat_penalty/std': 0.1329401582479477, 'rewards/repeat_floor/mean': 0.9605028629302979, 'rewards/repeat_floor/std': 0.019854148849844933, 'rewards/near_duplicate_penalty/mean': 0.8545929193496704, 'rewards/near_duplicate_penalty/std': 0.11463391035795212, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9493215084075928, 'rewards/distinct_2/std': 0.07687065005302429, 'rewards/total_composite/mean': 0.25261372327804565, 'rewards/total_composite/std': 0.12922576069831848, 'reward': 0.25261372327804565, 'reward_std': 0.12922576069831848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1647508293390274, 'sampling/sampling_logp_difference/max': 1.8061351776123047, 'sampling/importance_sampling_ratio/min': 0.16428786516189575, 'sampling/importance_sampling_ratio/mean': 1.0099539756774902, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8246480226516724, 'clip_ratio/low_mean': 0.054459065198898315, 'clip_ratio/low_min': 0.054459065198898315, 'clip_ratio/high_mean': 0.11747966380789876, 'clip_ratio/high_max': 0.11747966380789876, 'clip_ratio/region_mean': 0.17193872900679708, 'reward_total_mean': 0.25261372327804565, 'reward_meter_mean': 0.8375080823898315, 'reward_meter_std': 0.32897061109542847, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6353042721748352, 'reward_repeat_penalty_std': 0.1329401582479477, 'reward_repeat_floor_mean': 0.9605028629302979, 'reward_repeat_floor_std': 0.019854148849844933, 'reward_near_duplicate_penalty_mean': 0.8545929193496704, 'reward_near_duplicate_penalty_std': 0.11463391035795212, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9493215084075928, 'reward_distinct_2_std': 0.07687065005302429, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.25261372327804565, 'reward_total_composite_std': 0.12922576069831848, 'epoch': 0.04}

 26%|██▌       | 847/3300 [3:07:59<7:13:53, 10.61s/it]INFO 04-12 19:06:25 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 848/3300 [3:08:08<6:59:58, 10.28s/it]2026-04-12 19:06:34,966 | INFO | train_grpo_train | metrics_logged mode=train step=848

                                                      
{'loss': 0.0029, 'grad_norm': 15.614718437194824, 'learning_rate': 7.433333333333334e-06, 'num_tokens': 1515224.0, 'completions/mean_length': 33.875, 'completions/min_length': 28.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5524039268493652, 'rewards/meter/std': 0.4421142339706421, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.9797122478485107, 'rewards/repeat_penalty/std': 0.030837321653962135, 'rewards/repeat_floor/mean': 0.9969568252563477, 'rewards/repeat_floor/std': 0.0046256063506007195, 'rewards/near_duplicate_penalty/mean': 0.9797122478485107, 'rewards/near_duplicate_penalty/std': 0.030837321653962135, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2150241732597351, 'rewards/total_composite/std': 0.18644297122955322, 'reward': 0.2150241732597351, 'reward_std': 0.18644297122955322, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1725483387708664, 'sampling/sampling_logp_difference/max': 2.9017302989959717, 'sampling/importance_sampling_ratio/min': 0.054928094148635864, 'sampling/importance_sampling_ratio/mean': 1.0425009727478027, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1987946033477783, 'clip_ratio/low_mean': 0.0832839971408248, 'clip_ratio/low_min': 0.0832839971408248, 'clip_ratio/high_mean': 0.07823642995208502, 'clip_ratio/high_max': 0.07823642995208502, 'clip_ratio/region_mean': 0.1615204270929098, 'reward_total_mean': 0.2150241732597351, 'reward_meter_mean': 0.5524039268493652, 'reward_meter_std': 0.4421142339706421, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9797122478485107, 'reward_repeat_penalty_std': 0.030837321653962135, 'reward_repeat_floor_mean': 0.9969568252563477, 'reward_repeat_floor_std': 0.0046256063506007195, 'reward_near_duplicate_penalty_mean': 0.9797122478485107, 'reward_near_duplicate_penalty_std': 0.030837321653962135, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.2150241732597351, 'reward_total_composite_std': 0.18644297122955322, 'epoch': 0.04}

 26%|██▌       | 848/3300 [3:08:08<6:59:58, 10.28s/it]INFO 04-12 19:06:35 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 849/3300 [3:08:22<7:39:26, 11.25s/it]2026-04-12 19:06:48,478 | INFO | train_grpo_train | metrics_logged mode=train step=849

                                                      
{'loss': -0.0386, 'grad_norm': 5.820818901062012, 'learning_rate': 7.430303030303031e-06, 'num_tokens': 1516771.0, 'completions/mean_length': 89.375, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 29.000001907348633, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.43082544207572937, 'rewards/meter/std': 0.41046732664108276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9545419812202454, 'rewards/lexical_plausibility/std': 0.1056101992726326, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.2585087716579437, 'rewards/repeat_penalty/mean': 0.944037675857544, 'rewards/repeat_penalty/std': 0.04582136869430542, 'rewards/repeat_floor/mean': 0.992429792881012, 'rewards/repeat_floor/std': 0.006118534132838249, 'rewards/near_duplicate_penalty/mean': 0.9577739834785461, 'rewards/near_duplicate_penalty/std': 0.04377095773816109, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.18160486221313477, 'rewards/total_composite/std': 0.17931190133094788, 'reward': 0.18160486221313477, 'reward_std': 0.17931188642978668, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1891181915998459, 'sampling/sampling_logp_difference/max': 2.019103527069092, 'sampling/importance_sampling_ratio/min': 0.1327744424343109, 'sampling/importance_sampling_ratio/mean': 0.9903258085250854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7362756803631783, 'clip_ratio/low_mean': 0.0722704716026783, 'clip_ratio/low_min': 0.0722704716026783, 'clip_ratio/high_mean': 0.12103262916207314, 'clip_ratio/high_max': 0.12103262916207314, 'clip_ratio/region_mean': 0.19330310076475143, 'reward_total_mean': 0.18160486221313477, 'reward_meter_mean': 0.43082544207572937, 'reward_meter_std': 0.41046732664108276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9545419812202454, 'reward_lexical_plausibility_std': 0.1056101992726326, 'reward_repeat_penalty_mean': 0.944037675857544, 'reward_repeat_penalty_std': 0.04582136869430542, 'reward_repeat_floor_mean': 0.992429792881012, 'reward_repeat_floor_std': 0.006118534132838249, 'reward_near_duplicate_penalty_mean': 0.9577739834785461, 'reward_near_duplicate_penalty_std': 0.04377095773816109, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.2585087716579437, 'reward_total_composite_mean': 0.18160486221313477, 'reward_total_composite_std': 0.17931190133094788, 'epoch': 0.04}

 26%|██▌       | 849/3300 [3:08:22<7:39:26, 11.25s/it]INFO 04-12 19:06:48 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 850/3300 [3:08:30<7:01:45, 10.33s/it]2026-04-12 19:06:56,664 | INFO | train_grpo_train | metrics_logged mode=train step=850

                                                      
{'loss': -0.0926, 'grad_norm': 29.220003128051758, 'learning_rate': 7.4272727272727275e-06, 'num_tokens': 1518245.0, 'completions/mean_length': 24.25, 'completions/min_length': 19.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9903374910354614, 'rewards/meter/std': 0.006952945608645678, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.7323012948036194, 'rewards/repeat_penalty/std': 0.11840531975030899, 'rewards/repeat_floor/mean': 0.9781587719917297, 'rewards/repeat_floor/std': 0.017474928870797157, 'rewards/near_duplicate_penalty/mean': 0.9538556337356567, 'rewards/near_duplicate_penalty/std': 0.08808069676160812, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9800602793693542, 'rewards/distinct_2/std': 0.04028116911649704, 'rewards/total_composite/mean': 0.32705873250961304, 'rewards/total_composite/std': 0.09604796022176743, 'reward': 0.32705873250961304, 'reward_std': 0.09604796022176743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15786370635032654, 'sampling/sampling_logp_difference/max': 2.3714892864227295, 'sampling/importance_sampling_ratio/min': 0.09334160387516022, 'sampling/importance_sampling_ratio/mean': 1.0329214334487915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0143160820007324, 'clip_ratio/low_mean': 0.046266235411167145, 'clip_ratio/low_min': 0.046266235411167145, 'clip_ratio/high_mean': 0.08138676919043064, 'clip_ratio/high_max': 0.08138676919043064, 'clip_ratio/region_mean': 0.12765300460159779, 'reward_total_mean': 0.32705873250961304, 'reward_meter_mean': 0.9903374910354614, 'reward_meter_std': 0.006952945608645678, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7323012948036194, 'reward_repeat_penalty_std': 0.11840531975030899, 'reward_repeat_floor_mean': 0.9781587719917297, 'reward_repeat_floor_std': 0.017474928870797157, 'reward_near_duplicate_penalty_mean': 0.9538556337356567, 'reward_near_duplicate_penalty_std': 0.08808069676160812, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9800602793693542, 'reward_distinct_2_std': 0.04028116911649704, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.32705873250961304, 'reward_total_composite_std': 0.09604796022176743, 'epoch': 0.04}

 26%|██▌       | 850/3300 [3:08:30<7:01:45, 10.33s/it]INFO 04-12 19:06:56 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:01,  5.55s/it]

 23%|██▎       | 3/13 [00:22<01:20,  8.08s/it]

 31%|███       | 4/13 [00:32<01:16,  8.54s/it]

 38%|███▊      | 5/13 [00:40<01:08,  8.52s/it]

 46%|████▌     | 6/13 [00:47<00:56,  8.10s/it]

 54%|█████▍    | 7/13 [01:00<00:57,  9.53s/it]

 62%|██████▏   | 8/13 [01:07<00:44,  8.91s/it]

 69%|██████▉   | 9/13 [01:19<00:39,  9.82s/it]

 77%|███████▋  | 10/13 [01:26<00:26,  8.74s/it]

 85%|████████▍ | 11/13 [01:33<00:16,  8.48s/it]

 92%|█████████▏| 12/13 [01:45<00:09,  9.32s/it]

100%|██████████| 13/13 [01:52<00:00,  8.79s/it]2026-04-12 19:08:59,066 | INFO | train_grpo_train | metrics_logged mode=eval step=850

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 122.3217, 'eval_samples_per_second': 0.85, 'eval_steps_per_second': 0.106, 'eval_num_tokens': 1518245.0, 'eval_completions/mean_length': 119.89423076923077, 'eval_completions/min_length': 24.153846153846153, 'eval_completions/max_length': 344.0769230769231, 'eval_completions/clipped_ratio': 0.04807692307692308, 'eval_completions/mean_terminated_length': 99.59478114201472, 'eval_completions/min_terminated_length': 24.153846153846153, 'eval_completions/max_terminated_length': 236.0, 'eval_rewards/meter/mean': 0.5940083219454839, 'eval_rewards/meter/std': 0.3709836029089414, 'eval_rewards/count_adherence/mean': 0.7718354509426997, 'eval_rewards/count_adherence/std': 0.21803476202946442, 'eval_rewards/arabic_clean/mean': 0.9230769230769231, 'eval_rewards/arabic_clean/std': 0.21757131356459397, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9956730741720933, 'eval_rewards/arabic_floor/std': 0.012238385012516608, 'eval_rewards/count_floor/mean': 0.9315506219863892, 'eval_rewards/count_floor/std': 0.0654104257432314, 'eval_rewards/lexical_plausibility/mean': 0.9737811409510099, 'eval_rewards/lexical_plausibility/std': 0.07229006003874999, 'eval_rewards/judge_quality/mean': 0.3463461444928096, 'eval_rewards/judge_quality/std': 0.1750354331273299, 'eval_rewards/repeat_penalty/mean': 0.6587543166600741, 'eval_rewards/repeat_penalty/std': 0.31394536048173904, 'eval_rewards/repeat_floor/mean': 0.9617704107211187, 'eval_rewards/repeat_floor/std': 0.03857633424922824, 'eval_rewards/near_duplicate_penalty/mean': 0.916577540911161, 'eval_rewards/near_duplicate_penalty/std': 0.0802127426633468, 'eval_rewards/opening_diversity/mean': 0.9571698904037476, 'eval_rewards/opening_diversity/std': 0.08214899517882329, 'eval_rewards/distinct_2/mean': 0.7428175623600299, 'eval_rewards/distinct_2/std': 0.3078934856905387, 'eval_rewards/total_composite/mean': 0.1832546075949302, 'eval_rewards/total_composite/std': 0.15364239307550284, 'eval_reward': 0.1832546075949302, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0761807497877341, 'eval_sampling/sampling_logp_difference/max': 1.1380724540123572, 'eval_sampling/importance_sampling_ratio/min': 0.32706615443413073, 'eval_sampling/importance_sampling_ratio/mean': 1.0208074037845318, 'eval_sampling/importance_sampling_ratio/max': 1.5621135968428392, 'eval_entropy': 0.9077165310199444, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1832546075949302, 'eval_reward_meter_mean': 0.5940083219454839, 'eval_reward_meter_std': 0.3709836029089414, 'eval_reward_count_adherence_mean': 0.7718354509426997, 'eval_reward_count_adherence_std': 0.21803476202946442, 'eval_reward_arabic_clean_mean': 0.9230769230769231, 'eval_reward_arabic_clean_std': 0.21757131356459397, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9956730741720933, 'eval_reward_arabic_floor_std': 0.012238385012516608, 'eval_reward_count_floor_mean': 0.9315506219863892, 'eval_reward_count_floor_std': 0.0654104257432314, 'eval_reward_lexical_plausibility_mean': 0.9737811409510099, 'eval_reward_lexical_plausibility_std': 0.07229006003874999, 'eval_reward_repeat_penalty_mean': 0.6587543166600741, 'eval_reward_repeat_penalty_std': 0.31394536048173904, 'eval_reward_repeat_floor_mean': 0.9617704107211187, 'eval_reward_repeat_floor_std': 0.03857633424922824, 'eval_reward_near_duplicate_penalty_mean': 0.916577540911161, 'eval_reward_near_duplicate_penalty_std': 0.0802127426633468, 'eval_reward_opening_diversity_mean': 0.9571698904037476, 'eval_reward_opening_diversity_std': 0.08214899517882329, 'eval_reward_distinct_2_mean': 0.7428175623600299, 'eval_reward_distinct_2_std': 0.3078934856905387, 'eval_reward_judge_quality_mean': 0.3463461444928096, 'eval_reward_judge_quality_std': 0.1750354331273299, 'eval_reward_total_composite_mean': 0.1832546075949302, 'eval_reward_total_composite_std': 0.15364239307550284, 'epoch': 0.04}

 26%|██▌       | 850/3300 [3:10:32<7:01:45, 10.33s/it]

100%|██████████| 13/13 [01:52<00:00,  8.79s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 26%|██▌       | 851/3300 [3:10:44<32:17:55, 47.48s/it]2026-04-12 19:09:10,830 | INFO | train_grpo_train | metrics_logged mode=train step=851

                                                       
{'loss': 0.0403, 'grad_norm': 16.429882049560547, 'learning_rate': 7.424242424242425e-06, 'num_tokens': 1519875.0, 'completions/mean_length': 42.75, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6129672527313232, 'rewards/meter/std': 0.3593083620071411, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9181331396102905, 'rewards/repeat_penalty/std': 0.10355643182992935, 'rewards/repeat_floor/mean': 0.9904378652572632, 'rewards/repeat_floor/std': 0.010766570456326008, 'rewards/near_duplicate_penalty/mean': 0.9653908014297485, 'rewards/near_duplicate_penalty/std': 0.023855308070778847, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9597697257995605, 'rewards/distinct_2/std': 0.06292783468961716, 'rewards/total_composite/mean': 0.2604284882545471, 'rewards/total_composite/std': 0.14449924230575562, 'reward': 0.2604284882545471, 'reward_std': 0.14449924230575562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15844325721263885, 'sampling/sampling_logp_difference/max': 1.6276932954788208, 'sampling/importance_sampling_ratio/min': 0.1963820457458496, 'sampling/importance_sampling_ratio/mean': 0.9931143522262573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9619476944208145, 'clip_ratio/low_mean': 0.05755591858178377, 'clip_ratio/low_min': 0.05755591858178377, 'clip_ratio/high_mean': 0.11028680298477411, 'clip_ratio/high_max': 0.11028680298477411, 'clip_ratio/region_mean': 0.16784272156655788, 'reward_total_mean': 0.2604284882545471, 'reward_meter_mean': 0.6129672527313232, 'reward_meter_std': 0.3593083620071411, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9181331396102905, 'reward_repeat_penalty_std': 0.10355643182992935, 'reward_repeat_floor_mean': 0.9904378652572632, 'reward_repeat_floor_std': 0.010766570456326008, 'reward_near_duplicate_penalty_mean': 0.9653908014297485, 'reward_near_duplicate_penalty_std': 0.023855308070778847, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9597697257995605, 'reward_distinct_2_std': 0.06292783468961716, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.2604284882545471, 'reward_total_composite_std': 0.14449924230575562, 'epoch': 0.04}

 26%|██▌       | 851/3300 [3:10:44<32:17:55, 47.48s/it]INFO 04-12 19:09:11 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 852/3300 [3:10:52<24:14:02, 35.64s/it]2026-04-12 19:09:18,839 | INFO | train_grpo_train | metrics_logged mode=train step=852

                                                       
{'loss': 0.0665, 'grad_norm': 18.8461971282959, 'learning_rate': 7.421212121212121e-06, 'num_tokens': 1521304.0, 'completions/mean_length': 28.625, 'completions/min_length': 24.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.48669493198394775, 'rewards/meter/std': 0.23172593116760254, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.9849555492401123, 'rewards/repeat_penalty/std': 0.015864934772253036, 'rewards/repeat_floor/mean': 0.9977433681488037, 'rewards/repeat_floor/std': 0.0023797410540282726, 'rewards/near_duplicate_penalty/mean': 0.9849555492401123, 'rewards/near_duplicate_penalty/std': 0.015864934772253036, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2910352945327759, 'rewards/total_composite/std': 0.1671823114156723, 'reward': 0.2910352945327759, 'reward_std': 0.1671823114156723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1417986899614334, 'sampling/sampling_logp_difference/max': 1.6381053924560547, 'sampling/importance_sampling_ratio/min': 0.19434790313243866, 'sampling/importance_sampling_ratio/mean': 1.0088468790054321, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5967523828148842, 'clip_ratio/low_mean': 0.08824086841195822, 'clip_ratio/low_min': 0.08824086841195822, 'clip_ratio/high_mean': 0.05666666757315397, 'clip_ratio/high_max': 0.05666666757315397, 'clip_ratio/region_mean': 0.1449075359851122, 'reward_total_mean': 0.2910352945327759, 'reward_meter_mean': 0.48669493198394775, 'reward_meter_std': 0.23172593116760254, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9849555492401123, 'reward_repeat_penalty_std': 0.015864934772253036, 'reward_repeat_floor_mean': 0.9977433681488037, 'reward_repeat_floor_std': 0.0023797410540282726, 'reward_near_duplicate_penalty_mean': 0.9849555492401123, 'reward_near_duplicate_penalty_std': 0.015864934772253036, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.2910352945327759, 'reward_total_composite_std': 0.1671823114156723, 'epoch': 0.04}

 26%|██▌       | 852/3300 [3:10:52<24:14:02, 35.64s/it]INFO 04-12 19:09:19 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 853/3300 [3:11:11<20:54:31, 30.76s/it]2026-04-12 19:09:38,216 | INFO | train_grpo_train | metrics_logged mode=train step=853

                                                       
{'loss': 0.1965, 'grad_norm': 17.48162841796875, 'learning_rate': 7.4181818181818185e-06, 'num_tokens': 1522861.0, 'completions/mean_length': 33.625, 'completions/min_length': 17.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7731637358665466, 'rewards/meter/std': 0.32478925585746765, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.881607711315155, 'rewards/repeat_penalty/std': 0.08464014530181885, 'rewards/repeat_floor/mean': 0.9868205785751343, 'rewards/repeat_floor/std': 0.008759504184126854, 'rewards/near_duplicate_penalty/mean': 0.9466404318809509, 'rewards/near_duplicate_penalty/std': 0.049804121255874634, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9633283615112305, 'rewards/distinct_2/std': 0.039248231798410416, 'rewards/total_composite/mean': 0.42103052139282227, 'rewards/total_composite/std': 0.21232201159000397, 'reward': 0.42103052139282227, 'reward_std': 0.21232201159000397, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1721251755952835, 'sampling/sampling_logp_difference/max': 1.676363468170166, 'sampling/importance_sampling_ratio/min': 0.18705295026302338, 'sampling/importance_sampling_ratio/mean': 1.0126187801361084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0633443146944046, 'clip_ratio/low_mean': 0.04480971675366163, 'clip_ratio/low_min': 0.04480971675366163, 'clip_ratio/high_mean': 0.09278519544750452, 'clip_ratio/high_max': 0.09278519544750452, 'clip_ratio/region_mean': 0.13759491220116615, 'reward_total_mean': 0.42103052139282227, 'reward_meter_mean': 0.7731637358665466, 'reward_meter_std': 0.32478925585746765, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.881607711315155, 'reward_repeat_penalty_std': 0.08464014530181885, 'reward_repeat_floor_mean': 0.9868205785751343, 'reward_repeat_floor_std': 0.008759504184126854, 'reward_near_duplicate_penalty_mean': 0.9466404318809509, 'reward_near_duplicate_penalty_std': 0.049804121255874634, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9633283615112305, 'reward_distinct_2_std': 0.039248231798410416, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.42103052139282227, 'reward_total_composite_std': 0.21232201159000397, 'epoch': 0.04}

 26%|██▌       | 853/3300 [3:11:12<20:54:31, 30.76s/it]INFO 04-12 19:09:38 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 854/3300 [3:11:26<17:30:21, 25.77s/it]2026-04-12 19:09:52,335 | INFO | train_grpo_train | metrics_logged mode=train step=854

                                                       
{'loss': -0.058, 'grad_norm': 4.439054489135742, 'learning_rate': 7.415151515151515e-06, 'num_tokens': 1524362.0, 'completions/mean_length': 97.625, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.42857360839844, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7083907723426819, 'rewards/meter/std': 0.41697582602500916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.953191876411438, 'rewards/lexical_plausibility/std': 0.13239336013793945, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9339249134063721, 'rewards/repeat_penalty/std': 0.0925673320889473, 'rewards/repeat_floor/mean': 0.9913167953491211, 'rewards/repeat_floor/std': 0.011540849693119526, 'rewards/near_duplicate_penalty/mean': 0.9588955640792847, 'rewards/near_duplicate_penalty/std': 0.04725971445441246, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9720279574394226, 'rewards/distinct_2/std': 0.05644134432077408, 'rewards/total_composite/mean': 0.28051459789276123, 'rewards/total_composite/std': 0.1796312928199768, 'reward': 0.28051459789276123, 'reward_std': 0.1796312779188156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14446556568145752, 'sampling/sampling_logp_difference/max': 1.145294189453125, 'sampling/importance_sampling_ratio/min': 0.3181303143501282, 'sampling/importance_sampling_ratio/mean': 1.0005576610565186, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9032696560025215, 'clip_ratio/low_mean': 0.02806441579014063, 'clip_ratio/low_min': 0.02806441579014063, 'clip_ratio/high_mean': 0.09225909784436226, 'clip_ratio/high_max': 0.09225909784436226, 'clip_ratio/region_mean': 0.12032351363450289, 'reward_total_mean': 0.28051459789276123, 'reward_meter_mean': 0.7083907723426819, 'reward_meter_std': 0.41697582602500916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.953191876411438, 'reward_lexical_plausibility_std': 0.13239336013793945, 'reward_repeat_penalty_mean': 0.9339249134063721, 'reward_repeat_penalty_std': 0.0925673320889473, 'reward_repeat_floor_mean': 0.9913167953491211, 'reward_repeat_floor_std': 0.011540849693119526, 'reward_near_duplicate_penalty_mean': 0.9588955640792847, 'reward_near_duplicate_penalty_std': 0.04725971445441246, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9720279574394226, 'reward_distinct_2_std': 0.05644134432077408, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.28051459789276123, 'reward_total_composite_std': 0.1796312928199768, 'epoch': 0.04}

 26%|██▌       | 854/3300 [3:11:26<17:30:21, 25.77s/it]INFO 04-12 19:09:52 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 855/3300 [3:11:39<15:00:51, 22.11s/it]2026-04-12 19:10:05,906 | INFO | train_grpo_train | metrics_logged mode=train step=855

                                                       
{'loss': -0.0112, 'grad_norm': 4.827568531036377, 'learning_rate': 7.412121212121213e-06, 'num_tokens': 1525901.0, 'completions/mean_length': 90.375, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 30.142858505249023, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6936649084091187, 'rewards/meter/std': 0.41528838872909546, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.9342831969261169, 'rewards/lexical_plausibility/std': 0.11025092750787735, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.25753986835479736, 'rewards/repeat_penalty/mean': 0.8837237358093262, 'rewards/repeat_penalty/std': 0.12036402523517609, 'rewards/repeat_floor/mean': 0.9920949935913086, 'rewards/repeat_floor/std': 0.007197240367531776, 'rewards/near_duplicate_penalty/mean': 0.9957888126373291, 'rewards/near_duplicate_penalty/std': 0.008139593526721, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9816849827766418, 'rewards/distinct_2/std': 0.05180269479751587, 'rewards/total_composite/mean': 0.2913930416107178, 'rewards/total_composite/std': 0.2611802816390991, 'reward': 0.2913930416107178, 'reward_std': 0.26118025183677673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16113024950027466, 'sampling/sampling_logp_difference/max': 1.1160130500793457, 'sampling/importance_sampling_ratio/min': 0.3275832533836365, 'sampling/importance_sampling_ratio/mean': 1.023898959159851, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.16659776866436, 'clip_ratio/low_mean': 0.04225318972021341, 'clip_ratio/low_min': 0.04225318972021341, 'clip_ratio/high_mean': 0.05241228174418211, 'clip_ratio/high_max': 0.05241228174418211, 'clip_ratio/region_mean': 0.09466547146439552, 'reward_total_mean': 0.2913930416107178, 'reward_meter_mean': 0.6936649084091187, 'reward_meter_std': 0.41528838872909546, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.9342831969261169, 'reward_lexical_plausibility_std': 0.11025092750787735, 'reward_repeat_penalty_mean': 0.8837237358093262, 'reward_repeat_penalty_std': 0.12036402523517609, 'reward_repeat_floor_mean': 0.9920949935913086, 'reward_repeat_floor_std': 0.007197240367531776, 'reward_near_duplicate_penalty_mean': 0.9957888126373291, 'reward_near_duplicate_penalty_std': 0.008139593526721, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9816849827766418, 'reward_distinct_2_std': 0.05180269479751587, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.25753986835479736, 'reward_total_composite_mean': 0.2913930416107178, 'reward_total_composite_std': 0.2611802816390991, 'epoch': 0.04}

 26%|██▌       | 855/3300 [3:11:39<15:00:51, 22.11s/it]INFO 04-12 19:10:06 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 856/3300 [3:11:48<12:18:42, 18.14s/it]2026-04-12 19:10:14,774 | INFO | train_grpo_train | metrics_logged mode=train step=856

                                                       
{'loss': -0.024, 'grad_norm': 18.634891510009766, 'learning_rate': 7.40909090909091e-06, 'num_tokens': 1527271.0, 'completions/mean_length': 22.25, 'completions/min_length': 20.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9949920773506165, 'rewards/meter/std': 0.001823427272029221, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9047181606292725, 'rewards/lexical_plausibility/std': 0.17645813524723053, 'rewards/judge_quality/mean': 0.4674999713897705, 'rewards/judge_quality/std': 0.3022179901599884, 'rewards/repeat_penalty/mean': 0.667992115020752, 'rewards/repeat_penalty/std': 0.1552628129720688, 'rewards/repeat_floor/mean': 0.9699516296386719, 'rewards/repeat_floor/std': 0.031173698604106903, 'rewards/near_duplicate_penalty/mean': 0.9386887550354004, 'rewards/near_duplicate_penalty/std': 0.14516063034534454, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9349817037582397, 'rewards/distinct_2/std': 0.10918577760457993, 'rewards/total_composite/mean': 0.4552363157272339, 'rewards/total_composite/std': 0.29889899492263794, 'reward': 0.4552363157272339, 'reward_std': 0.29889899492263794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19031181931495667, 'sampling/sampling_logp_difference/max': 1.3144540786743164, 'sampling/importance_sampling_ratio/min': 0.26862093806266785, 'sampling/importance_sampling_ratio/mean': 0.9980427622795105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0378254055976868, 'clip_ratio/low_mean': 0.07619047863408923, 'clip_ratio/low_min': 0.07619047863408923, 'clip_ratio/high_mean': 0.07451923191547394, 'clip_ratio/high_max': 0.07451923191547394, 'clip_ratio/region_mean': 0.15070971054956317, 'reward_total_mean': 0.4552363157272339, 'reward_meter_mean': 0.9949920773506165, 'reward_meter_std': 0.001823427272029221, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9047181606292725, 'reward_lexical_plausibility_std': 0.17645813524723053, 'reward_repeat_penalty_mean': 0.667992115020752, 'reward_repeat_penalty_std': 0.1552628129720688, 'reward_repeat_floor_mean': 0.9699516296386719, 'reward_repeat_floor_std': 0.031173698604106903, 'reward_near_duplicate_penalty_mean': 0.9386887550354004, 'reward_near_duplicate_penalty_std': 0.14516063034534454, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9349817037582397, 'reward_distinct_2_std': 0.10918577760457993, 'reward_judge_quality_mean': 0.4674999713897705, 'reward_judge_quality_std': 0.3022179901599884, 'reward_total_composite_mean': 0.4552363157272339, 'reward_total_composite_std': 0.29889899492263794, 'epoch': 0.05}

 26%|██▌       | 856/3300 [3:11:48<12:18:42, 18.14s/it]INFO 04-12 19:10:15 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 857/3300 [3:12:02<11:26:20, 16.86s/it]2026-04-12 19:10:28,644 | INFO | train_grpo_train | metrics_logged mode=train step=857

                                                       
{'loss': -0.064, 'grad_norm': 5.120342254638672, 'learning_rate': 7.406060606060607e-06, 'num_tokens': 1528686.0, 'completions/mean_length': 94.875, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.28571701049805, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.43202874064445496, 'rewards/meter/std': 0.2548648416996002, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9484158158302307, 'rewards/lexical_plausibility/std': 0.1459021121263504, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.18850918114185333, 'rewards/repeat_penalty/mean': 0.9160827398300171, 'rewards/repeat_penalty/std': 0.18514809012413025, 'rewards/repeat_floor/mean': 0.9891101121902466, 'rewards/repeat_floor/std': 0.02300209552049637, 'rewards/near_duplicate_penalty/mean': 0.9556350708007812, 'rewards/near_duplicate_penalty/std': 0.07480573654174805, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9737762212753296, 'rewards/distinct_2/std': 0.07417204976081848, 'rewards/total_composite/mean': 0.1700025349855423, 'rewards/total_composite/std': 0.12248846888542175, 'reward': 0.1700025349855423, 'reward_std': 0.12248846143484116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16652749478816986, 'sampling/sampling_logp_difference/max': 1.5494086742401123, 'sampling/importance_sampling_ratio/min': 0.2123735249042511, 'sampling/importance_sampling_ratio/mean': 1.01360023021698, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9056715667247772, 'clip_ratio/low_mean': 0.03661616239696741, 'clip_ratio/low_min': 0.03661616239696741, 'clip_ratio/high_mean': 0.0843220204114914, 'clip_ratio/high_max': 0.0843220204114914, 'clip_ratio/region_mean': 0.1209381828084588, 'reward_total_mean': 0.1700025349855423, 'reward_meter_mean': 0.43202874064445496, 'reward_meter_std': 0.2548648416996002, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9484158158302307, 'reward_lexical_plausibility_std': 0.1459021121263504, 'reward_repeat_penalty_mean': 0.9160827398300171, 'reward_repeat_penalty_std': 0.18514809012413025, 'reward_repeat_floor_mean': 0.9891101121902466, 'reward_repeat_floor_std': 0.02300209552049637, 'reward_near_duplicate_penalty_mean': 0.9556350708007812, 'reward_near_duplicate_penalty_std': 0.07480573654174805, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9737762212753296, 'reward_distinct_2_std': 0.07417204976081848, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.18850918114185333, 'reward_total_composite_mean': 0.1700025349855423, 'reward_total_composite_std': 0.12248846888542175, 'epoch': 0.05}

 26%|██▌       | 857/3300 [3:12:02<11:26:20, 16.86s/it]INFO 04-12 19:10:28 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 858/3300 [3:12:15<10:37:57, 15.67s/it]2026-04-12 19:10:41,553 | INFO | train_grpo_train | metrics_logged mode=train step=858

                                                       
{'loss': -0.0678, 'grad_norm': 4.318520545959473, 'learning_rate': 7.403030303030304e-06, 'num_tokens': 1530227.0, 'completions/mean_length': 91.625, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.571430206298828, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7368934154510498, 'rewards/meter/std': 0.41320157051086426, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9618405103683472, 'rewards/lexical_plausibility/std': 0.10793144255876541, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.27139851450920105, 'rewards/repeat_penalty/mean': 0.8467693328857422, 'rewards/repeat_penalty/std': 0.1273772269487381, 'rewards/repeat_floor/mean': 0.985733687877655, 'rewards/repeat_floor/std': 0.011612470261752605, 'rewards/near_duplicate_penalty/mean': 0.9524245262145996, 'rewards/near_duplicate_penalty/std': 0.05922013148665428, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.22769123315811157, 'rewards/total_composite/std': 0.1319492608308792, 'reward': 0.22769123315811157, 'reward_std': 0.13194924592971802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1802298128604889, 'sampling/sampling_logp_difference/max': 1.210554599761963, 'sampling/importance_sampling_ratio/min': 0.3580833077430725, 'sampling/importance_sampling_ratio/mean': 1.0397518873214722, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9694603458046913, 'clip_ratio/low_mean': 0.056907894089818, 'clip_ratio/low_min': 0.056907894089818, 'clip_ratio/high_mean': 0.11477100383490324, 'clip_ratio/high_max': 0.11477100383490324, 'clip_ratio/region_mean': 0.17167889792472124, 'reward_total_mean': 0.22769123315811157, 'reward_meter_mean': 0.7368934154510498, 'reward_meter_std': 0.41320157051086426, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9618405103683472, 'reward_lexical_plausibility_std': 0.10793144255876541, 'reward_repeat_penalty_mean': 0.8467693328857422, 'reward_repeat_penalty_std': 0.1273772269487381, 'reward_repeat_floor_mean': 0.985733687877655, 'reward_repeat_floor_std': 0.011612470261752605, 'reward_near_duplicate_penalty_mean': 0.9524245262145996, 'reward_near_duplicate_penalty_std': 0.05922013148665428, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.27139851450920105, 'reward_total_composite_mean': 0.22769123315811157, 'reward_total_composite_std': 0.1319492608308792, 'epoch': 0.05}

 26%|██▌       | 858/3300 [3:12:15<10:37:57, 15.67s/it]INFO 04-12 19:10:41 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 859/3300 [3:12:24<9:15:39, 13.66s/it] 2026-04-12 19:10:50,506 | INFO | train_grpo_train | metrics_logged mode=train step=859

                                                      
{'loss': 0.0709, 'grad_norm': 17.332473754882812, 'learning_rate': 7.4e-06, 'num_tokens': 1531665.0, 'completions/mean_length': 30.75, 'completions/min_length': 16.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7409610748291016, 'rewards/meter/std': 0.4215104281902313, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.837775468826294, 'rewards/repeat_penalty/std': 0.15606820583343506, 'rewards/repeat_floor/mean': 0.984778881072998, 'rewards/repeat_floor/std': 0.01476488821208477, 'rewards/near_duplicate_penalty/mean': 0.9613751769065857, 'rewards/near_duplicate_penalty/std': 0.041865818202495575, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9577505588531494, 'rewards/distinct_2/std': 0.061652857810258865, 'rewards/total_composite/mean': 0.34708037972450256, 'rewards/total_composite/std': 0.23397400975227356, 'reward': 0.34708037972450256, 'reward_std': 0.23397399485111237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17871926724910736, 'sampling/sampling_logp_difference/max': 2.5530452728271484, 'sampling/importance_sampling_ratio/min': 0.07784424722194672, 'sampling/importance_sampling_ratio/mean': 1.0164144039154053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9876235574483871, 'clip_ratio/low_mean': 0.043884568847715855, 'clip_ratio/low_min': 0.043884568847715855, 'clip_ratio/high_mean': 0.10760150942951441, 'clip_ratio/high_max': 0.10760150942951441, 'clip_ratio/region_mean': 0.15148607827723026, 'reward_total_mean': 0.34708037972450256, 'reward_meter_mean': 0.7409610748291016, 'reward_meter_std': 0.4215104281902313, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.837775468826294, 'reward_repeat_penalty_std': 0.15606820583343506, 'reward_repeat_floor_mean': 0.984778881072998, 'reward_repeat_floor_std': 0.01476488821208477, 'reward_near_duplicate_penalty_mean': 0.9613751769065857, 'reward_near_duplicate_penalty_std': 0.041865818202495575, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9577505588531494, 'reward_distinct_2_std': 0.061652857810258865, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.34708037972450256, 'reward_total_composite_std': 0.23397400975227356, 'epoch': 0.05}

 26%|██▌       | 859/3300 [3:12:24<9:15:39, 13.66s/it]INFO 04-12 19:10:50 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 860/3300 [3:12:33<8:18:32, 12.26s/it]2026-04-12 19:10:59,502 | INFO | train_grpo_train | metrics_logged mode=train step=860

                                                      
{'loss': 0.109, 'grad_norm': 24.974897384643555, 'learning_rate': 7.396969696969698e-06, 'num_tokens': 1533306.0, 'completions/mean_length': 33.125, 'completions/min_length': 24.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6976901292800903, 'rewards/meter/std': 0.2908549904823303, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9856323003768921, 'rewards/repeat_penalty/std': 0.022808540612459183, 'rewards/repeat_floor/mean': 0.9978448152542114, 'rewards/repeat_floor/std': 0.0034212765749543905, 'rewards/near_duplicate_penalty/mean': 0.9856323003768921, 'rewards/near_duplicate_penalty/std': 0.022808540612459183, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.25643959641456604, 'rewards/total_composite/std': 0.10720512270927429, 'reward': 0.25643959641456604, 'reward_std': 0.1072051152586937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2010492980480194, 'sampling/sampling_logp_difference/max': 2.270589828491211, 'sampling/importance_sampling_ratio/min': 0.10325125604867935, 'sampling/importance_sampling_ratio/mean': 1.0106351375579834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9329140707850456, 'clip_ratio/low_mean': 0.08431525342166424, 'clip_ratio/low_min': 0.08431525342166424, 'clip_ratio/high_mean': 0.12070085853338242, 'clip_ratio/high_max': 0.12070085853338242, 'clip_ratio/region_mean': 0.20501611195504665, 'reward_total_mean': 0.25643959641456604, 'reward_meter_mean': 0.6976901292800903, 'reward_meter_std': 0.2908549904823303, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9856323003768921, 'reward_repeat_penalty_std': 0.022808540612459183, 'reward_repeat_floor_mean': 0.9978448152542114, 'reward_repeat_floor_std': 0.0034212765749543905, 'reward_near_duplicate_penalty_mean': 0.9856323003768921, 'reward_near_duplicate_penalty_std': 0.022808540612459183, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.25643959641456604, 'reward_total_composite_std': 0.10720512270927429, 'epoch': 0.05}

 26%|██▌       | 860/3300 [3:12:33<8:18:32, 12.26s/it]INFO 04-12 19:10:59 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 861/3300 [3:12:42<7:44:41, 11.43s/it]2026-04-12 19:11:09,003 | INFO | train_grpo_train | metrics_logged mode=train step=861

                                                      
{'loss': 0.0785, 'grad_norm': 16.73429298400879, 'learning_rate': 7.393939393939395e-06, 'num_tokens': 1534955.0, 'completions/mean_length': 37.125, 'completions/min_length': 18.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8543132543563843, 'rewards/meter/std': 0.3133929371833801, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6212499737739563, 'rewards/judge_quality/std': 0.26626721024513245, 'rewards/repeat_penalty/mean': 0.9071429967880249, 'rewards/repeat_penalty/std': 0.09249752014875412, 'rewards/repeat_floor/mean': 0.9906789064407349, 'rewards/repeat_floor/std': 0.008230502717196941, 'rewards/near_duplicate_penalty/mean': 0.9711740016937256, 'rewards/near_duplicate_penalty/std': 0.032110974192619324, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9653087854385376, 'rewards/distinct_2/std': 0.049989111721515656, 'rewards/total_composite/mean': 0.5404928922653198, 'rewards/total_composite/std': 0.331277996301651, 'reward': 0.5404928922653198, 'reward_std': 0.3312779664993286, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13502705097198486, 'sampling/sampling_logp_difference/max': 1.8205838203430176, 'sampling/importance_sampling_ratio/min': 0.16193118691444397, 'sampling/importance_sampling_ratio/mean': 0.9921262264251709, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8673032820224762, 'clip_ratio/low_mean': 0.051648352295160294, 'clip_ratio/low_min': 0.051648352295160294, 'clip_ratio/high_mean': 0.04333420004695654, 'clip_ratio/high_max': 0.04333420004695654, 'clip_ratio/region_mean': 0.09498255234211683, 'reward_total_mean': 0.5404928922653198, 'reward_meter_mean': 0.8543132543563843, 'reward_meter_std': 0.3133929371833801, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9071429967880249, 'reward_repeat_penalty_std': 0.09249752014875412, 'reward_repeat_floor_mean': 0.9906789064407349, 'reward_repeat_floor_std': 0.008230502717196941, 'reward_near_duplicate_penalty_mean': 0.9711740016937256, 'reward_near_duplicate_penalty_std': 0.032110974192619324, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9653087854385376, 'reward_distinct_2_std': 0.049989111721515656, 'reward_judge_quality_mean': 0.6212499737739563, 'reward_judge_quality_std': 0.26626721024513245, 'reward_total_composite_mean': 0.5404928922653198, 'reward_total_composite_std': 0.331277996301651, 'epoch': 0.05}

 26%|██▌       | 861/3300 [3:12:42<7:44:41, 11.43s/it]INFO 04-12 19:11:09 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 862/3300 [3:12:56<8:08:13, 12.02s/it]2026-04-12 19:11:22,381 | INFO | train_grpo_train | metrics_logged mode=train step=862

                                                      
{'loss': -0.0413, 'grad_norm': 5.6746039390563965, 'learning_rate': 7.390909090909092e-06, 'num_tokens': 1536590.0, 'completions/mean_length': 107.375, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.57143020629883, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5889490842819214, 'rewards/meter/std': 0.40843716263771057, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.9559023976325989, 'rewards/lexical_plausibility/std': 0.12472687661647797, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.2773052752017975, 'rewards/repeat_penalty/mean': 0.9557989239692688, 'rewards/repeat_penalty/std': 0.04557191953063011, 'rewards/repeat_floor/mean': 0.9942553043365479, 'rewards/repeat_floor/std': 0.005345368292182684, 'rewards/near_duplicate_penalty/mean': 0.9718118906021118, 'rewards/near_duplicate_penalty/std': 0.0204855315387249, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9831501841545105, 'rewards/distinct_2/std': 0.031356681138277054, 'rewards/total_composite/mean': 0.20572629570960999, 'rewards/total_composite/std': 0.24233824014663696, 'reward': 0.20572629570960999, 'reward_std': 0.24233825504779816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1550081968307495, 'sampling/sampling_logp_difference/max': 1.634211778640747, 'sampling/importance_sampling_ratio/min': 0.19510610401630402, 'sampling/importance_sampling_ratio/mean': 1.0181303024291992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7492427676916122, 'clip_ratio/low_mean': 0.06888403557240963, 'clip_ratio/low_min': 0.06888403557240963, 'clip_ratio/high_mean': 0.05245192348957062, 'clip_ratio/high_max': 0.05245192348957062, 'clip_ratio/region_mean': 0.12133595906198025, 'reward_total_mean': 0.20572629570960999, 'reward_meter_mean': 0.5889490842819214, 'reward_meter_std': 0.40843716263771057, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.9559023976325989, 'reward_lexical_plausibility_std': 0.12472687661647797, 'reward_repeat_penalty_mean': 0.9557989239692688, 'reward_repeat_penalty_std': 0.04557191953063011, 'reward_repeat_floor_mean': 0.9942553043365479, 'reward_repeat_floor_std': 0.005345368292182684, 'reward_near_duplicate_penalty_mean': 0.9718118906021118, 'reward_near_duplicate_penalty_std': 0.0204855315387249, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9831501841545105, 'reward_distinct_2_std': 0.031356681138277054, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.2773052752017975, 'reward_total_composite_mean': 0.20572629570960999, 'reward_total_composite_std': 0.24233824014663696, 'epoch': 0.05}

 26%|██▌       | 862/3300 [3:12:56<8:08:13, 12.02s/it]INFO 04-12 19:11:22 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 863/3300 [3:13:04<7:27:54, 11.03s/it]2026-04-12 19:11:31,100 | INFO | train_grpo_train | metrics_logged mode=train step=863

                                                      
{'loss': 0.0265, 'grad_norm': 13.635979652404785, 'learning_rate': 7.3878787878787885e-06, 'num_tokens': 1538093.0, 'completions/mean_length': 36.875, 'completions/min_length': 19.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.755847692489624, 'rewards/meter/std': 0.25042101740837097, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.9187597036361694, 'rewards/repeat_penalty/std': 0.08145544677972794, 'rewards/repeat_floor/mean': 0.9906264543533325, 'rewards/repeat_floor/std': 0.007059456314891577, 'rewards/near_duplicate_penalty/mean': 0.9500097036361694, 'rewards/near_duplicate_penalty/std': 0.04892057552933693, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4313575029373169, 'rewards/total_composite/std': 0.21279199421405792, 'reward': 0.4313575029373169, 'reward_std': 0.21279196441173553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1790132075548172, 'sampling/sampling_logp_difference/max': 2.265141248703003, 'sampling/importance_sampling_ratio/min': 0.10381536930799484, 'sampling/importance_sampling_ratio/mean': 1.0040229558944702, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0147500038146973, 'clip_ratio/low_mean': 0.08506417460739613, 'clip_ratio/low_min': 0.08506417460739613, 'clip_ratio/high_mean': 0.07293621078133583, 'clip_ratio/high_max': 0.07293621078133583, 'clip_ratio/region_mean': 0.15800038538873196, 'reward_total_mean': 0.4313575029373169, 'reward_meter_mean': 0.755847692489624, 'reward_meter_std': 0.25042101740837097, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9187597036361694, 'reward_repeat_penalty_std': 0.08145544677972794, 'reward_repeat_floor_mean': 0.9906264543533325, 'reward_repeat_floor_std': 0.007059456314891577, 'reward_near_duplicate_penalty_mean': 0.9500097036361694, 'reward_near_duplicate_penalty_std': 0.04892057552933693, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.4313575029373169, 'reward_total_composite_std': 0.21279199421405792, 'epoch': 0.05}

 26%|██▌       | 863/3300 [3:13:04<7:27:54, 11.03s/it]INFO 04-12 19:11:31 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 864/3300 [3:13:12<6:47:50, 10.05s/it]2026-04-12 19:11:38,855 | INFO | train_grpo_train | metrics_logged mode=train step=864

                                                      
{'loss': 0.0288, 'grad_norm': 21.118820190429688, 'learning_rate': 7.384848484848486e-06, 'num_tokens': 1539482.0, 'completions/mean_length': 20.625, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7514198422431946, 'rewards/meter/std': 0.39766955375671387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6799999475479126, 'rewards/judge_quality/std': 0.2747986316680908, 'rewards/repeat_penalty/mean': 0.7259435653686523, 'rewards/repeat_penalty/std': 0.04471856728196144, 'rewards/repeat_floor/mean': 0.9801887273788452, 'rewards/repeat_floor/std': 0.008943723514676094, 'rewards/near_duplicate_penalty/mean': 0.9679248332977295, 'rewards/near_duplicate_penalty/std': 0.05962476134300232, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.48981940746307373, 'rewards/total_composite/std': 0.34963035583496094, 'reward': 0.48981940746307373, 'reward_std': 0.34963035583496094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13045021891593933, 'sampling/sampling_logp_difference/max': 1.0673611164093018, 'sampling/importance_sampling_ratio/min': 0.34391486644744873, 'sampling/importance_sampling_ratio/mean': 1.041421890258789, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8075439110398293, 'clip_ratio/low_mean': 0.06896146619692445, 'clip_ratio/low_min': 0.06896146619692445, 'clip_ratio/high_mean': 0.04229323333129287, 'clip_ratio/high_max': 0.04229323333129287, 'clip_ratio/region_mean': 0.11125469952821732, 'reward_total_mean': 0.48981940746307373, 'reward_meter_mean': 0.7514198422431946, 'reward_meter_std': 0.39766955375671387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7259435653686523, 'reward_repeat_penalty_std': 0.04471856728196144, 'reward_repeat_floor_mean': 0.9801887273788452, 'reward_repeat_floor_std': 0.008943723514676094, 'reward_near_duplicate_penalty_mean': 0.9679248332977295, 'reward_near_duplicate_penalty_std': 0.05962476134300232, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6799999475479126, 'reward_judge_quality_std': 0.2747986316680908, 'reward_total_composite_mean': 0.48981940746307373, 'reward_total_composite_std': 0.34963035583496094, 'epoch': 0.05}

 26%|██▌       | 864/3300 [3:13:12<6:47:50, 10.05s/it]INFO 04-12 19:11:39 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 865/3300 [3:13:26<7:33:47, 11.18s/it]2026-04-12 19:11:52,688 | INFO | train_grpo_train | metrics_logged mode=train step=865

                                                      
{'loss': -0.0887, 'grad_norm': 5.104220390319824, 'learning_rate': 7.381818181818182e-06, 'num_tokens': 1541910.0, 'completions/mean_length': 162.5, 'completions/min_length': 95.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 112.5714340209961, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.8013204336166382, 'rewards/meter/std': 0.2830294072628021, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9571428298950195, 'rewards/count_floor/std': 0.022908104583621025, 'rewards/lexical_plausibility/mean': 0.9757699966430664, 'rewards/lexical_plausibility/std': 0.0685328021645546, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.27742189168930054, 'rewards/repeat_penalty/std': 0.23724889755249023, 'rewards/repeat_floor/mean': 0.8951751589775085, 'rewards/repeat_floor/std': 0.07160430401563644, 'rewards/near_duplicate_penalty/mean': 0.7448532581329346, 'rewards/near_duplicate_penalty/std': 0.22116361558437347, 'rewards/opening_diversity/mean': 0.8401442170143127, 'rewards/opening_diversity/std': 0.2642263174057007, 'rewards/distinct_2/mean': 0.36709436774253845, 'rewards/distinct_2/std': 0.2868894338607788, 'rewards/total_composite/mean': 0.09057766199111938, 'rewards/total_composite/std': 0.03686390817165375, 'reward': 0.09057766199111938, 'reward_std': 0.03686390817165375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11240290105342865, 'sampling/sampling_logp_difference/max': 3.373406171798706, 'sampling/importance_sampling_ratio/min': 0.034272700548172, 'sampling/importance_sampling_ratio/mean': 1.00165593624115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5945057608187199, 'clip_ratio/low_mean': 0.0151869161054492, 'clip_ratio/low_min': 0.0151869161054492, 'clip_ratio/high_mean': 0.0843690219335258, 'clip_ratio/high_max': 0.0843690219335258, 'clip_ratio/region_mean': 0.099555938038975, 'reward_total_mean': 0.09057766199111938, 'reward_meter_mean': 0.8013204336166382, 'reward_meter_std': 0.2830294072628021, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9571428298950195, 'reward_count_floor_std': 0.022908104583621025, 'reward_lexical_plausibility_mean': 0.9757699966430664, 'reward_lexical_plausibility_std': 0.0685328021645546, 'reward_repeat_penalty_mean': 0.27742189168930054, 'reward_repeat_penalty_std': 0.23724889755249023, 'reward_repeat_floor_mean': 0.8951751589775085, 'reward_repeat_floor_std': 0.07160430401563644, 'reward_near_duplicate_penalty_mean': 0.7448532581329346, 'reward_near_duplicate_penalty_std': 0.22116361558437347, 'reward_opening_diversity_mean': 0.8401442170143127, 'reward_opening_diversity_std': 0.2642263174057007, 'reward_distinct_2_mean': 0.36709436774253845, 'reward_distinct_2_std': 0.2868894338607788, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.09057766199111938, 'reward_total_composite_std': 0.03686390817165375, 'epoch': 0.05}

 26%|██▌       | 865/3300 [3:13:26<7:33:47, 11.18s/it]INFO 04-12 19:11:52 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 866/3300 [3:13:34<6:55:47, 10.25s/it]2026-04-12 19:12:00,773 | INFO | train_grpo_train | metrics_logged mode=train step=866

                                                      
{'loss': 0.0512, 'grad_norm': 23.65655517578125, 'learning_rate': 7.378787878787879e-06, 'num_tokens': 1543481.0, 'completions/mean_length': 27.375, 'completions/min_length': 25.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5662153959274292, 'rewards/meter/std': 0.34931254386901855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.23384979367256165, 'rewards/repeat_penalty/mean': 0.9695297479629517, 'rewards/repeat_penalty/std': 0.062252357602119446, 'rewards/repeat_floor/mean': 0.9961713552474976, 'rewards/repeat_floor/std': 0.007383969612419605, 'rewards/near_duplicate_penalty/mean': 0.9833513498306274, 'rewards/near_duplicate_penalty/std': 0.02826262079179287, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.3762473464012146, 'rewards/total_composite/std': 0.2378154695034027, 'reward': 0.3762473464012146, 'reward_std': 0.2378154695034027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16126790642738342, 'sampling/sampling_logp_difference/max': 2.372018814086914, 'sampling/importance_sampling_ratio/min': 0.09329219162464142, 'sampling/importance_sampling_ratio/mean': 0.9962642192840576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.903374083340168, 'clip_ratio/low_mean': 0.07348901219666004, 'clip_ratio/low_min': 0.07348901219666004, 'clip_ratio/high_mean': 0.08042203634977341, 'clip_ratio/high_max': 0.08042203634977341, 'clip_ratio/region_mean': 0.15391104854643345, 'reward_total_mean': 0.3762473464012146, 'reward_meter_mean': 0.5662153959274292, 'reward_meter_std': 0.34931254386901855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9695297479629517, 'reward_repeat_penalty_std': 0.062252357602119446, 'reward_repeat_floor_mean': 0.9961713552474976, 'reward_repeat_floor_std': 0.007383969612419605, 'reward_near_duplicate_penalty_mean': 0.9833513498306274, 'reward_near_duplicate_penalty_std': 0.02826262079179287, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.23384979367256165, 'reward_total_composite_mean': 0.3762473464012146, 'reward_total_composite_std': 0.2378154695034027, 'epoch': 0.05}

 26%|██▌       | 866/3300 [3:13:34<6:55:47, 10.25s/it]INFO 04-12 19:12:01 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 867/3300 [3:13:42<6:21:42,  9.41s/it]2026-04-12 19:12:08,226 | INFO | train_grpo_train | metrics_logged mode=train step=867

                                                      
{'loss': 0.0654, 'grad_norm': 23.61690330505371, 'learning_rate': 7.375757575757576e-06, 'num_tokens': 1545039.0, 'completions/mean_length': 29.75, 'completions/min_length': 27.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.75, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.46958133578300476, 'rewards/meter/std': 0.3885764181613922, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9350000023841858, 'rewards/judge_quality/std': 0.0160356592386961, 'rewards/repeat_penalty/mean': 0.9995629191398621, 'rewards/repeat_penalty/std': 0.0012362091802060604, 'rewards/repeat_floor/mean': 0.9999344348907471, 'rewards/repeat_floor/std': 0.0001854250585893169, 'rewards/near_duplicate_penalty/mean': 0.9995629191398621, 'rewards/near_duplicate_penalty/std': 0.0012362091802060604, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4419003129005432, 'rewards/total_composite/std': 0.36857014894485474, 'reward': 0.4419003129005432, 'reward_std': 0.36857014894485474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11949307471513748, 'sampling/sampling_logp_difference/max': 2.093111038208008, 'sampling/importance_sampling_ratio/min': 0.12330294400453568, 'sampling/importance_sampling_ratio/mean': 0.9952176213264465, 'sampling/importance_sampling_ratio/max': 1.5812885761260986, 'entropy': 0.5339246280491352, 'clip_ratio/low_mean': 0.04908218001946807, 'clip_ratio/low_min': 0.04908218001946807, 'clip_ratio/high_mean': 0.09117965400218964, 'clip_ratio/high_max': 0.09117965400218964, 'clip_ratio/region_mean': 0.1402618340216577, 'reward_total_mean': 0.4419003129005432, 'reward_meter_mean': 0.46958133578300476, 'reward_meter_std': 0.3885764181613922, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9995629191398621, 'reward_repeat_penalty_std': 0.0012362091802060604, 'reward_repeat_floor_mean': 0.9999344348907471, 'reward_repeat_floor_std': 0.0001854250585893169, 'reward_near_duplicate_penalty_mean': 0.9995629191398621, 'reward_near_duplicate_penalty_std': 0.0012362091802060604, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9350000023841858, 'reward_judge_quality_std': 0.0160356592386961, 'reward_total_composite_mean': 0.4419003129005432, 'reward_total_composite_std': 0.36857014894485474, 'epoch': 0.05}

 26%|██▋       | 867/3300 [3:13:42<6:21:42,  9.41s/it]INFO 04-12 19:12:08 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 868/3300 [3:13:55<7:05:44, 10.50s/it]2026-04-12 19:12:21,275 | INFO | train_grpo_train | metrics_logged mode=train step=868

                                                      
{'loss': -0.0494, 'grad_norm': 5.872175693511963, 'learning_rate': 7.372727272727274e-06, 'num_tokens': 1546567.0, 'completions/mean_length': 98.0, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.85714340209961, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.2153160572052002, 'rewards/meter/std': 0.3381102681159973, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9605072140693665, 'rewards/lexical_plausibility/std': 0.11170248687267303, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9477797746658325, 'rewards/repeat_penalty/std': 0.04624565318226814, 'rewards/repeat_floor/mean': 0.9939553737640381, 'rewards/repeat_floor/std': 0.0046444605104625225, 'rewards/near_duplicate_penalty/mean': 0.97865891456604, 'rewards/near_duplicate_penalty/std': 0.01389516144990921, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9684065580368042, 'rewards/distinct_2/std': 0.044430434703826904, 'rewards/total_composite/mean': 0.08430136740207672, 'rewards/total_composite/std': 0.14996010065078735, 'reward': 0.08430136740207672, 'reward_std': 0.14996007084846497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.148210808634758, 'sampling/sampling_logp_difference/max': 1.4876470565795898, 'sampling/importance_sampling_ratio/min': 0.22590357065200806, 'sampling/importance_sampling_ratio/mean': 0.9940487146377563, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6033209711313248, 'clip_ratio/low_mean': 0.08633532654494047, 'clip_ratio/low_min': 0.08633532654494047, 'clip_ratio/high_mean': 0.04233511723577976, 'clip_ratio/high_max': 0.04233511723577976, 'clip_ratio/region_mean': 0.12867044378072023, 'reward_total_mean': 0.08430136740207672, 'reward_meter_mean': 0.2153160572052002, 'reward_meter_std': 0.3381102681159973, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9605072140693665, 'reward_lexical_plausibility_std': 0.11170248687267303, 'reward_repeat_penalty_mean': 0.9477797746658325, 'reward_repeat_penalty_std': 0.04624565318226814, 'reward_repeat_floor_mean': 0.9939553737640381, 'reward_repeat_floor_std': 0.0046444605104625225, 'reward_near_duplicate_penalty_mean': 0.97865891456604, 'reward_near_duplicate_penalty_std': 0.01389516144990921, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9684065580368042, 'reward_distinct_2_std': 0.044430434703826904, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.08430136740207672, 'reward_total_composite_std': 0.14996010065078735, 'epoch': 0.05}

 26%|██▋       | 868/3300 [3:13:55<7:05:44, 10.50s/it]INFO 04-12 19:12:21 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 869/3300 [3:14:08<7:44:39, 11.47s/it]2026-04-12 19:12:34,990 | INFO | train_grpo_train | metrics_logged mode=train step=869

                                                      
{'loss': -0.0467, 'grad_norm': 4.988476753234863, 'learning_rate': 7.36969696969697e-06, 'num_tokens': 1548184.0, 'completions/mean_length': 106.125, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.142860412597656, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6716678142547607, 'rewards/meter/std': 0.359855055809021, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9578310251235962, 'rewards/lexical_plausibility/std': 0.11927188187837601, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.9183489680290222, 'rewards/repeat_penalty/std': 0.08904451131820679, 'rewards/repeat_floor/mean': 0.9906047582626343, 'rewards/repeat_floor/std': 0.009944959543645382, 'rewards/near_duplicate_penalty/mean': 0.9705090522766113, 'rewards/near_duplicate_penalty/std': 0.02776431478559971, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9447595477104187, 'rewards/distinct_2/std': 0.06738081574440002, 'rewards/total_composite/mean': 0.17086133360862732, 'rewards/total_composite/std': 0.13022080063819885, 'reward': 0.17086133360862732, 'reward_std': 0.13022080063819885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18647263944149017, 'sampling/sampling_logp_difference/max': 1.8996200561523438, 'sampling/importance_sampling_ratio/min': 0.14962546527385712, 'sampling/importance_sampling_ratio/mean': 0.9904005527496338, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.973294161260128, 'clip_ratio/low_mean': 0.09361335076391697, 'clip_ratio/low_min': 0.09361335076391697, 'clip_ratio/high_mean': 0.03989813104271889, 'clip_ratio/high_max': 0.03989813104271889, 'clip_ratio/region_mean': 0.13351148180663586, 'reward_total_mean': 0.17086133360862732, 'reward_meter_mean': 0.6716678142547607, 'reward_meter_std': 0.359855055809021, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9578310251235962, 'reward_lexical_plausibility_std': 0.11927188187837601, 'reward_repeat_penalty_mean': 0.9183489680290222, 'reward_repeat_penalty_std': 0.08904451131820679, 'reward_repeat_floor_mean': 0.9906047582626343, 'reward_repeat_floor_std': 0.009944959543645382, 'reward_near_duplicate_penalty_mean': 0.9705090522766113, 'reward_near_duplicate_penalty_std': 0.02776431478559971, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9447595477104187, 'reward_distinct_2_std': 0.06738081574440002, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.17086133360862732, 'reward_total_composite_std': 0.13022080063819885, 'epoch': 0.05}

 26%|██▋       | 869/3300 [3:14:08<7:44:39, 11.47s/it]INFO 04-12 19:12:35 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 870/3300 [3:14:22<8:08:34, 12.06s/it]2026-04-12 19:12:48,445 | INFO | train_grpo_train | metrics_logged mode=train step=870

                                                      
{'loss': -0.043, 'grad_norm': 3.4905426502227783, 'learning_rate': 7.3666666666666676e-06, 'num_tokens': 1550053.0, 'completions/mean_length': 189.625, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 82.16667175292969, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.22927355766296387, 'rewards/meter/std': 0.17436650395393372, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9318444728851318, 'rewards/lexical_plausibility/std': 0.12274092435836792, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.6991857290267944, 'rewards/repeat_penalty/std': 0.3495412766933441, 'rewards/repeat_floor/mean': 0.9626133441925049, 'rewards/repeat_floor/std': 0.04534195736050606, 'rewards/near_duplicate_penalty/mean': 0.9125322103500366, 'rewards/near_duplicate_penalty/std': 0.0874033272266388, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.17598575353622437, 'rewards/distinct_2/mean': 0.7824562788009644, 'rewards/distinct_2/std': 0.2682415246963501, 'rewards/total_composite/mean': 0.04961996525526047, 'rewards/total_composite/std': 0.057361848652362823, 'reward': 0.04961996525526047, 'reward_std': 0.057361844927072525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20023545622825623, 'sampling/sampling_logp_difference/max': 1.6970524787902832, 'sampling/importance_sampling_ratio/min': 0.18322278559207916, 'sampling/importance_sampling_ratio/mean': 1.036204218864441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1511788666248322, 'clip_ratio/low_mean': 0.048418208956718445, 'clip_ratio/low_min': 0.048418208956718445, 'clip_ratio/high_mean': 0.08592762146145105, 'clip_ratio/high_max': 0.08592762146145105, 'clip_ratio/region_mean': 0.1343458304181695, 'reward_total_mean': 0.04961996525526047, 'reward_meter_mean': 0.22927355766296387, 'reward_meter_std': 0.17436650395393372, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9318444728851318, 'reward_lexical_plausibility_std': 0.12274092435836792, 'reward_repeat_penalty_mean': 0.6991857290267944, 'reward_repeat_penalty_std': 0.3495412766933441, 'reward_repeat_floor_mean': 0.9626133441925049, 'reward_repeat_floor_std': 0.04534195736050606, 'reward_near_duplicate_penalty_mean': 0.9125322103500366, 'reward_near_duplicate_penalty_std': 0.0874033272266388, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.17598575353622437, 'reward_distinct_2_mean': 0.7824562788009644, 'reward_distinct_2_std': 0.2682415246963501, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.04961996525526047, 'reward_total_composite_std': 0.057361848652362823, 'epoch': 0.05}

 26%|██▋       | 870/3300 [3:14:22<8:08:34, 12.06s/it]INFO 04-12 19:12:48 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 871/3300 [3:14:31<7:30:16, 11.12s/it]2026-04-12 19:12:57,371 | INFO | train_grpo_train | metrics_logged mode=train step=871

                                                      
{'loss': -0.0201, 'grad_norm': 16.500547409057617, 'learning_rate': 7.363636363636364e-06, 'num_tokens': 1551586.0, 'completions/mean_length': 37.625, 'completions/min_length': 18.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7750052809715271, 'rewards/meter/std': 0.3972945809364319, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.8672630190849304, 'rewards/repeat_penalty/std': 0.10418357700109482, 'rewards/repeat_floor/mean': 0.9845211505889893, 'rewards/repeat_floor/std': 0.01130314078181982, 'rewards/near_duplicate_penalty/mean': 0.931851863861084, 'rewards/near_duplicate_penalty/std': 0.05389080569148064, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9624260663986206, 'rewards/distinct_2/std': 0.07107843458652496, 'rewards/total_composite/mean': 0.3773953914642334, 'rewards/total_composite/std': 0.24919314682483673, 'reward': 0.3773953914642334, 'reward_std': 0.24919313192367554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16073720157146454, 'sampling/sampling_logp_difference/max': 2.2238855361938477, 'sampling/importance_sampling_ratio/min': 0.10818792134523392, 'sampling/importance_sampling_ratio/mean': 1.0057411193847656, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8519498854875565, 'clip_ratio/low_mean': 0.0675589069724083, 'clip_ratio/low_min': 0.0675589069724083, 'clip_ratio/high_mean': 0.04549406748265028, 'clip_ratio/high_max': 0.04549406748265028, 'clip_ratio/region_mean': 0.11305297445505857, 'reward_total_mean': 0.3773953914642334, 'reward_meter_mean': 0.7750052809715271, 'reward_meter_std': 0.3972945809364319, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8672630190849304, 'reward_repeat_penalty_std': 0.10418357700109482, 'reward_repeat_floor_mean': 0.9845211505889893, 'reward_repeat_floor_std': 0.01130314078181982, 'reward_near_duplicate_penalty_mean': 0.931851863861084, 'reward_near_duplicate_penalty_std': 0.05389080569148064, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9624260663986206, 'reward_distinct_2_std': 0.07107843458652496, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.3773953914642334, 'reward_total_composite_std': 0.24919314682483673, 'epoch': 0.05}

 26%|██▋       | 871/3300 [3:14:31<7:30:16, 11.12s/it]INFO 04-12 19:12:57 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 872/3300 [3:14:40<7:10:26, 10.64s/it]2026-04-12 19:13:06,877 | INFO | train_grpo_train | metrics_logged mode=train step=872

                                                      
{'loss': -0.0567, 'grad_norm': 19.357362747192383, 'learning_rate': 7.360606060606061e-06, 'num_tokens': 1553228.0, 'completions/mean_length': 32.25, 'completions/min_length': 16.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8034600615501404, 'rewards/meter/std': 0.30390915274620056, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.8482853174209595, 'rewards/repeat_penalty/std': 0.16881361603736877, 'rewards/repeat_floor/mean': 0.979885458946228, 'rewards/repeat_floor/std': 0.02209554985165596, 'rewards/near_duplicate_penalty/mean': 0.9054011702537537, 'rewards/near_duplicate_penalty/std': 0.09223775565624237, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9550033211708069, 'rewards/distinct_2/std': 0.05283082276582718, 'rewards/total_composite/mean': 0.4801190495491028, 'rewards/total_composite/std': 0.33508336544036865, 'reward': 0.4801190495491028, 'reward_std': 0.33508336544036865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1537906527519226, 'sampling/sampling_logp_difference/max': 2.0069661140441895, 'sampling/importance_sampling_ratio/min': 0.1343958079814911, 'sampling/importance_sampling_ratio/mean': 1.0011324882507324, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8896751999855042, 'clip_ratio/low_mean': 0.07757065468467772, 'clip_ratio/low_min': 0.07757065468467772, 'clip_ratio/high_mean': 0.05026374477893114, 'clip_ratio/high_max': 0.05026374477893114, 'clip_ratio/region_mean': 0.12783439946360886, 'reward_total_mean': 0.4801190495491028, 'reward_meter_mean': 0.8034600615501404, 'reward_meter_std': 0.30390915274620056, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8482853174209595, 'reward_repeat_penalty_std': 0.16881361603736877, 'reward_repeat_floor_mean': 0.979885458946228, 'reward_repeat_floor_std': 0.02209554985165596, 'reward_near_duplicate_penalty_mean': 0.9054011702537537, 'reward_near_duplicate_penalty_std': 0.09223775565624237, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9550033211708069, 'reward_distinct_2_std': 0.05283082276582718, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.4801190495491028, 'reward_total_composite_std': 0.33508336544036865, 'epoch': 0.05}

 26%|██▋       | 872/3300 [3:14:40<7:10:26, 10.64s/it]INFO 04-12 19:13:07 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 873/3300 [3:14:55<7:59:52, 11.86s/it]2026-04-12 19:13:21,606 | INFO | train_grpo_train | metrics_logged mode=train step=873

                                                      
{'loss': 0.0241, 'grad_norm': 7.038979530334473, 'learning_rate': 7.357575757575758e-06, 'num_tokens': 1554761.0, 'completions/mean_length': 85.625, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.71428680419922, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.749538779258728, 'rewards/meter/std': 0.4065682888031006, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 0.9514477252960205, 'rewards/lexical_plausibility/std': 0.1141868606209755, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.2840491831302643, 'rewards/repeat_penalty/mean': 0.7762438654899597, 'rewards/repeat_penalty/std': 0.1221478059887886, 'rewards/repeat_floor/mean': 0.9819340705871582, 'rewards/repeat_floor/std': 0.010640923865139484, 'rewards/near_duplicate_penalty/mean': 0.9545605182647705, 'rewards/near_duplicate_penalty/std': 0.04778008162975311, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3044961094856262, 'rewards/total_composite/std': 0.2873344123363495, 'reward': 0.3044961094856262, 'reward_std': 0.2873344123363495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17623257637023926, 'sampling/sampling_logp_difference/max': 1.8010578155517578, 'sampling/importance_sampling_ratio/min': 0.16512413322925568, 'sampling/importance_sampling_ratio/mean': 1.0417358875274658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2829299345612526, 'clip_ratio/low_mean': 0.028685898054391146, 'clip_ratio/low_min': 0.028685898054391146, 'clip_ratio/high_mean': 0.08574879262596369, 'clip_ratio/high_max': 0.08574879262596369, 'clip_ratio/region_mean': 0.11443469068035483, 'reward_total_mean': 0.3044961094856262, 'reward_meter_mean': 0.749538779258728, 'reward_meter_std': 0.4065682888031006, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 0.9514477252960205, 'reward_lexical_plausibility_std': 0.1141868606209755, 'reward_repeat_penalty_mean': 0.7762438654899597, 'reward_repeat_penalty_std': 0.1221478059887886, 'reward_repeat_floor_mean': 0.9819340705871582, 'reward_repeat_floor_std': 0.010640923865139484, 'reward_near_duplicate_penalty_mean': 0.9545605182647705, 'reward_near_duplicate_penalty_std': 0.04778008162975311, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.2840491831302643, 'reward_total_composite_mean': 0.3044961094856262, 'reward_total_composite_std': 0.2873344123363495, 'epoch': 0.05}

 26%|██▋       | 873/3300 [3:14:55<7:59:52, 11.86s/it]INFO 04-12 19:13:21 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 874/3300 [3:15:06<7:46:47, 11.54s/it]2026-04-12 19:13:32,402 | INFO | train_grpo_train | metrics_logged mode=train step=874

                                                      
{'loss': 0.0361, 'grad_norm': 6.368078231811523, 'learning_rate': 7.354545454545456e-06, 'num_tokens': 1558003.0, 'completions/mean_length': 166.25, 'completions/min_length': 135.0, 'completions/max_length': 205.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.25, 'completions/min_terminated_length': 135.0, 'completions/max_terminated_length': 205.0, 'rewards/meter/mean': 0.7665997743606567, 'rewards/meter/std': 0.19188617169857025, 'rewards/count_adherence/mean': 0.925000011920929, 'rewards/count_adherence/std': 0.10350984334945679, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9775000214576721, 'rewards/count_floor/std': 0.03105293959379196, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.4261830449104309, 'rewards/repeat_penalty/std': 0.2072986215353012, 'rewards/repeat_floor/mean': 0.9325563907623291, 'rewards/repeat_floor/std': 0.0313393771648407, 'rewards/near_duplicate_penalty/mean': 0.8621490597724915, 'rewards/near_duplicate_penalty/std': 0.0647788792848587, 'rewards/opening_diversity/mean': 0.9624999761581421, 'rewards/opening_diversity/std': 0.1060660257935524, 'rewards/distinct_2/mean': 0.505378246307373, 'rewards/distinct_2/std': 0.19852392375469208, 'rewards/total_composite/mean': 0.20223039388656616, 'rewards/total_composite/std': 0.10624415427446365, 'reward': 0.20223039388656616, 'reward_std': 0.10624415427446365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1410389542579651, 'sampling/sampling_logp_difference/max': 3.2784266471862793, 'sampling/importance_sampling_ratio/min': 0.037687502801418304, 'sampling/importance_sampling_ratio/mean': 1.0044747591018677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9199883863329887, 'clip_ratio/low_mean': 0.0591319790109992, 'clip_ratio/low_min': 0.0591319790109992, 'clip_ratio/high_mean': 0.0784632358700037, 'clip_ratio/high_max': 0.0784632358700037, 'clip_ratio/region_mean': 0.1375952148810029, 'reward_total_mean': 0.20223039388656616, 'reward_meter_mean': 0.7665997743606567, 'reward_meter_std': 0.19188617169857025, 'reward_count_adherence_mean': 0.925000011920929, 'reward_count_adherence_std': 0.10350984334945679, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9775000214576721, 'reward_count_floor_std': 0.03105293959379196, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4261830449104309, 'reward_repeat_penalty_std': 0.2072986215353012, 'reward_repeat_floor_mean': 0.9325563907623291, 'reward_repeat_floor_std': 0.0313393771648407, 'reward_near_duplicate_penalty_mean': 0.8621490597724915, 'reward_near_duplicate_penalty_std': 0.0647788792848587, 'reward_opening_diversity_mean': 0.9624999761581421, 'reward_opening_diversity_std': 0.1060660257935524, 'reward_distinct_2_mean': 0.505378246307373, 'reward_distinct_2_std': 0.19852392375469208, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.20223039388656616, 'reward_total_composite_std': 0.10624415427446365, 'epoch': 0.05}

 26%|██▋       | 874/3300 [3:15:06<7:46:47, 11.54s/it]INFO 04-12 19:13:32 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 875/3300 [3:15:15<7:17:58, 10.84s/it]2026-04-12 19:13:41,588 | INFO | train_grpo_train | metrics_logged mode=train step=875

                                                      
{'loss': -0.092, 'grad_norm': 11.503551483154297, 'learning_rate': 7.351515151515151e-06, 'num_tokens': 1560242.0, 'completions/mean_length': 71.875, 'completions/min_length': 22.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9875743389129639, 'rewards/meter/std': 0.018453437834978104, 'rewards/count_adherence/mean': 0.6499999761581421, 'rewards/count_adherence/std': 0.2777460515499115, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8949999809265137, 'rewards/count_floor/std': 0.08332381397485733, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.7421561479568481, 'rewards/repeat_penalty/std': 0.1999952644109726, 'rewards/repeat_floor/mean': 0.9704427719116211, 'rewards/repeat_floor/std': 0.028293592855334282, 'rewards/near_duplicate_penalty/mean': 0.9162959456443787, 'rewards/near_duplicate_penalty/std': 0.11142666637897491, 'rewards/opening_diversity/mean': 0.9296875, 'rewards/opening_diversity/std': 0.1129826009273529, 'rewards/distinct_2/mean': 0.8579676151275635, 'rewards/distinct_2/std': 0.12484433501958847, 'rewards/total_composite/mean': 0.3413340449333191, 'rewards/total_composite/std': 0.04577485844492912, 'reward': 0.3413340449333191, 'reward_std': 0.04577486589550972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14321988821029663, 'sampling/sampling_logp_difference/max': 1.9944705963134766, 'sampling/importance_sampling_ratio/min': 0.13608567416667938, 'sampling/importance_sampling_ratio/mean': 0.9968971014022827, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9085503965616226, 'clip_ratio/low_mean': 0.1005949666723609, 'clip_ratio/low_min': 0.1005949666723609, 'clip_ratio/high_mean': 0.055285826325416565, 'clip_ratio/high_max': 0.055285826325416565, 'clip_ratio/region_mean': 0.15588079299777746, 'reward_total_mean': 0.3413340449333191, 'reward_meter_mean': 0.9875743389129639, 'reward_meter_std': 0.018453437834978104, 'reward_count_adherence_mean': 0.6499999761581421, 'reward_count_adherence_std': 0.2777460515499115, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8949999809265137, 'reward_count_floor_std': 0.08332381397485733, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7421561479568481, 'reward_repeat_penalty_std': 0.1999952644109726, 'reward_repeat_floor_mean': 0.9704427719116211, 'reward_repeat_floor_std': 0.028293592855334282, 'reward_near_duplicate_penalty_mean': 0.9162959456443787, 'reward_near_duplicate_penalty_std': 0.11142666637897491, 'reward_opening_diversity_mean': 0.9296875, 'reward_opening_diversity_std': 0.1129826009273529, 'reward_distinct_2_mean': 0.8579676151275635, 'reward_distinct_2_std': 0.12484433501958847, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.3413340449333191, 'reward_total_composite_std': 0.04577485844492912, 'epoch': 0.05}

 27%|██▋       | 875/3300 [3:15:15<7:17:58, 10.84s/it]INFO 04-12 19:13:41 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 876/3300 [3:15:26<7:24:31, 11.00s/it]2026-04-12 19:13:52,981 | INFO | train_grpo_train | metrics_logged mode=train step=876

                                                      
{'loss': -0.095, 'grad_norm': 9.276765823364258, 'learning_rate': 7.348484848484849e-06, 'num_tokens': 1562697.0, 'completions/mean_length': 73.875, 'completions/min_length': 28.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.9826794862747192, 'rewards/meter/std': 0.02156369574368, 'rewards/count_adherence/mean': 0.675000011920929, 'rewards/count_adherence/std': 0.2121320515871048, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9024999737739563, 'rewards/count_floor/std': 0.06363961100578308, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.31512606143951416, 'rewards/repeat_penalty/std': 0.24519988894462585, 'rewards/repeat_floor/mean': 0.9055250883102417, 'rewards/repeat_floor/std': 0.05129530653357506, 'rewards/near_duplicate_penalty/mean': 0.7299025058746338, 'rewards/near_duplicate_penalty/std': 0.2182547152042389, 'rewards/opening_diversity/mean': 0.9084821343421936, 'rewards/opening_diversity/std': 0.1367267221212387, 'rewards/distinct_2/mean': 0.4614526629447937, 'rewards/distinct_2/std': 0.23033969104290009, 'rewards/total_composite/mean': 0.2057122439146042, 'rewards/total_composite/std': 0.09926880896091461, 'reward': 0.2057122439146042, 'reward_std': 0.09926880151033401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10832500457763672, 'sampling/sampling_logp_difference/max': 3.4130635261535645, 'sampling/importance_sampling_ratio/min': 0.35713186860084534, 'sampling/importance_sampling_ratio/mean': 1.035207748413086, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8915101364254951, 'clip_ratio/low_mean': 0.04373576492071152, 'clip_ratio/low_min': 0.04373576492071152, 'clip_ratio/high_mean': 0.04673337284475565, 'clip_ratio/high_max': 0.04673337284475565, 'clip_ratio/region_mean': 0.09046913776546717, 'reward_total_mean': 0.2057122439146042, 'reward_meter_mean': 0.9826794862747192, 'reward_meter_std': 0.02156369574368, 'reward_count_adherence_mean': 0.675000011920929, 'reward_count_adherence_std': 0.2121320515871048, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9024999737739563, 'reward_count_floor_std': 0.06363961100578308, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.31512606143951416, 'reward_repeat_penalty_std': 0.24519988894462585, 'reward_repeat_floor_mean': 0.9055250883102417, 'reward_repeat_floor_std': 0.05129530653357506, 'reward_near_duplicate_penalty_mean': 0.7299025058746338, 'reward_near_duplicate_penalty_std': 0.2182547152042389, 'reward_opening_diversity_mean': 0.9084821343421936, 'reward_opening_diversity_std': 0.1367267221212387, 'reward_distinct_2_mean': 0.4614526629447937, 'reward_distinct_2_std': 0.23033969104290009, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.2057122439146042, 'reward_total_composite_std': 0.09926880896091461, 'epoch': 0.05}

 27%|██▋       | 876/3300 [3:15:26<7:24:31, 11.00s/it]INFO 04-12 19:13:53 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 877/3300 [3:15:34<6:45:01, 10.03s/it]2026-04-12 19:14:00,746 | INFO | train_grpo_train | metrics_logged mode=train step=877

                                                      
{'loss': 0.031, 'grad_norm': 26.927339553833008, 'learning_rate': 7.345454545454546e-06, 'num_tokens': 1564002.0, 'completions/mean_length': 16.125, 'completions/min_length': 14.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.125, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.9322580099105835, 'rewards/meter/std': 0.10376407206058502, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7649999856948853, 'rewards/judge_quality/std': 0.29799333214759827, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6960526704788208, 'rewards/total_composite/std': 0.2776948809623718, 'reward': 0.6960526704788208, 'reward_std': 0.27769485116004944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.137740820646286, 'sampling/sampling_logp_difference/max': 1.3927980661392212, 'sampling/importance_sampling_ratio/min': 0.24837934970855713, 'sampling/importance_sampling_ratio/mean': 1.0237650871276855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9306323155760765, 'clip_ratio/low_mean': 0.0396446087397635, 'clip_ratio/low_min': 0.0396446087397635, 'clip_ratio/high_mean': 0.10573792364448309, 'clip_ratio/high_max': 0.10573792364448309, 'clip_ratio/region_mean': 0.1453825323842466, 'reward_total_mean': 0.6960526704788208, 'reward_meter_mean': 0.9322580099105835, 'reward_meter_std': 0.10376407206058502, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7649999856948853, 'reward_judge_quality_std': 0.29799333214759827, 'reward_total_composite_mean': 0.6960526704788208, 'reward_total_composite_std': 0.2776948809623718, 'epoch': 0.05}

 27%|██▋       | 877/3300 [3:15:34<6:45:01, 10.03s/it]INFO 04-12 19:14:00 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 878/3300 [3:15:44<6:46:56, 10.08s/it]2026-04-12 19:14:10,995 | INFO | train_grpo_train | metrics_logged mode=train step=878

                                                      
{'loss': -0.0036, 'grad_norm': 12.388447761535645, 'learning_rate': 7.342424242424243e-06, 'num_tokens': 1565817.0, 'completions/mean_length': 57.875, 'completions/min_length': 49.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.542432427406311, 'rewards/meter/std': 0.371094673871994, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4387500286102295, 'rewards/judge_quality/std': 0.3985128700733185, 'rewards/repeat_penalty/mean': 0.9952296018600464, 'rewards/repeat_penalty/std': 0.009135845117270947, 'rewards/repeat_floor/mean': 0.9992844462394714, 'rewards/repeat_floor/std': 0.001370378420688212, 'rewards/near_duplicate_penalty/mean': 0.9952296018600464, 'rewards/near_duplicate_penalty/std': 0.009135845117270947, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21066735684871674, 'rewards/total_composite/std': 0.27830126881599426, 'reward': 0.21066735684871674, 'reward_std': 0.27830126881599426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11187071353197098, 'sampling/sampling_logp_difference/max': 1.6696062088012695, 'sampling/importance_sampling_ratio/min': 0.18832120299339294, 'sampling/importance_sampling_ratio/mean': 1.0375959873199463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7789190337061882, 'clip_ratio/low_mean': 0.0972730815410614, 'clip_ratio/low_min': 0.0972730815410614, 'clip_ratio/high_mean': 0.008196720853447914, 'clip_ratio/high_max': 0.008196720853447914, 'clip_ratio/region_mean': 0.10546980239450932, 'reward_total_mean': 0.21066735684871674, 'reward_meter_mean': 0.542432427406311, 'reward_meter_std': 0.371094673871994, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9952296018600464, 'reward_repeat_penalty_std': 0.009135845117270947, 'reward_repeat_floor_mean': 0.9992844462394714, 'reward_repeat_floor_std': 0.001370378420688212, 'reward_near_duplicate_penalty_mean': 0.9952296018600464, 'reward_near_duplicate_penalty_std': 0.009135845117270947, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4387500286102295, 'reward_judge_quality_std': 0.3985128700733185, 'reward_total_composite_mean': 0.21066735684871674, 'reward_total_composite_std': 0.27830126881599426, 'epoch': 0.05}

 27%|██▋       | 878/3300 [3:15:44<6:46:56, 10.08s/it]INFO 04-12 19:14:11 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 879/3300 [3:15:52<6:21:56,  9.47s/it]2026-04-12 19:14:18,969 | INFO | train_grpo_train | metrics_logged mode=train step=879

                                                      
{'loss': 0.0157, 'grad_norm': 25.1612606048584, 'learning_rate': 7.3393939393939395e-06, 'num_tokens': 1567193.0, 'completions/mean_length': 24.0, 'completions/min_length': 21.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.607140064239502, 'rewards/meter/std': 0.3807227909564972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6599999666213989, 'rewards/judge_quality/std': 0.2805097699165344, 'rewards/repeat_penalty/mean': 0.9357510805130005, 'rewards/repeat_penalty/std': 0.08802963048219681, 'rewards/repeat_floor/mean': 0.9910063147544861, 'rewards/repeat_floor/std': 0.011473841033875942, 'rewards/near_duplicate_penalty/mean': 0.9505317211151123, 'rewards/near_duplicate_penalty/std': 0.04951261356472969, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.3787187933921814, 'rewards/total_composite/std': 0.2884226143360138, 'reward': 0.3787187933921814, 'reward_std': 0.2884226143360138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.190205916762352, 'sampling/sampling_logp_difference/max': 2.4015331268310547, 'sampling/importance_sampling_ratio/min': 0.09057898074388504, 'sampling/importance_sampling_ratio/mean': 1.0315871238708496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.704744964838028, 'clip_ratio/low_mean': 0.08622023742645979, 'clip_ratio/low_min': 0.08622023742645979, 'clip_ratio/high_mean': 0.08312920713797212, 'clip_ratio/high_max': 0.08312920713797212, 'clip_ratio/region_mean': 0.1693494445644319, 'reward_total_mean': 0.3787187933921814, 'reward_meter_mean': 0.607140064239502, 'reward_meter_std': 0.3807227909564972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9357510805130005, 'reward_repeat_penalty_std': 0.08802963048219681, 'reward_repeat_floor_mean': 0.9910063147544861, 'reward_repeat_floor_std': 0.011473841033875942, 'reward_near_duplicate_penalty_mean': 0.9505317211151123, 'reward_near_duplicate_penalty_std': 0.04951261356472969, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.6599999666213989, 'reward_judge_quality_std': 0.2805097699165344, 'reward_total_composite_mean': 0.3787187933921814, 'reward_total_composite_std': 0.2884226143360138, 'epoch': 0.05}

 27%|██▋       | 879/3300 [3:15:52<6:21:56,  9.47s/it]INFO 04-12 19:14:19 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 880/3300 [3:16:02<6:31:12,  9.70s/it]2026-04-12 19:14:29,262 | INFO | train_grpo_train | metrics_logged mode=train step=880

                                                      
{'loss': 0.0227, 'grad_norm': 5.945997714996338, 'learning_rate': 7.336363636363637e-06, 'num_tokens': 1569920.0, 'completions/mean_length': 132.875, 'completions/min_length': 103.0, 'completions/max_length': 165.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.875, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.5385571718215942, 'rewards/meter/std': 0.37893539667129517, 'rewards/count_adherence/mean': 0.9125000238418579, 'rewards/count_adherence/std': 0.06408699601888657, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9737499952316284, 'rewards/count_floor/std': 0.01922609470784664, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.08628641068935394, 'rewards/repeat_penalty/std': 0.09096816182136536, 'rewards/repeat_floor/mean': 0.8775551915168762, 'rewards/repeat_floor/std': 0.024921759963035583, 'rewards/near_duplicate_penalty/mean': 0.7272076606750488, 'rewards/near_duplicate_penalty/std': 0.06567627936601639, 'rewards/opening_diversity/mean': 0.9173793792724609, 'rewards/opening_diversity/std': 0.08697489649057388, 'rewards/distinct_2/mean': 0.14923608303070068, 'rewards/distinct_2/std': 0.146689772605896, 'rewards/total_composite/mean': 0.0990256667137146, 'rewards/total_composite/std': 0.08351548761129379, 'reward': 0.0990256667137146, 'reward_std': 0.08351548761129379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09959546476602554, 'sampling/sampling_logp_difference/max': 2.3838906288146973, 'sampling/importance_sampling_ratio/min': 0.0921911969780922, 'sampling/importance_sampling_ratio/mean': 1.007084608078003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5856452696025372, 'clip_ratio/low_mean': 0.034149602986872196, 'clip_ratio/low_min': 0.034149602986872196, 'clip_ratio/high_mean': 0.036335236858576536, 'clip_ratio/high_max': 0.036335236858576536, 'clip_ratio/region_mean': 0.07048483984544873, 'reward_total_mean': 0.0990256667137146, 'reward_meter_mean': 0.5385571718215942, 'reward_meter_std': 0.37893539667129517, 'reward_count_adherence_mean': 0.9125000238418579, 'reward_count_adherence_std': 0.06408699601888657, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9737499952316284, 'reward_count_floor_std': 0.01922609470784664, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.08628641068935394, 'reward_repeat_penalty_std': 0.09096816182136536, 'reward_repeat_floor_mean': 0.8775551915168762, 'reward_repeat_floor_std': 0.024921759963035583, 'reward_near_duplicate_penalty_mean': 0.7272076606750488, 'reward_near_duplicate_penalty_std': 0.06567627936601639, 'reward_opening_diversity_mean': 0.9173793792724609, 'reward_opening_diversity_std': 0.08697489649057388, 'reward_distinct_2_mean': 0.14923608303070068, 'reward_distinct_2_std': 0.146689772605896, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.0990256667137146, 'reward_total_composite_std': 0.08351548761129379, 'epoch': 0.05}

 27%|██▋       | 880/3300 [3:16:03<6:31:12,  9.70s/it]INFO 04-12 19:14:29 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 881/3300 [3:16:11<6:11:17,  9.21s/it]2026-04-12 19:14:37,299 | INFO | train_grpo_train | metrics_logged mode=train step=881

                                                      
{'loss': 0.0372, 'grad_norm': 21.14667510986328, 'learning_rate': 7.333333333333333e-06, 'num_tokens': 1571431.0, 'completions/mean_length': 18.875, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.7041109800338745, 'rewards/meter/std': 0.43137630820274353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.3123185336589813, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.446069598197937, 'rewards/total_composite/std': 0.3866676986217499, 'reward': 0.446069598197937, 'reward_std': 0.3866676986217499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15217608213424683, 'sampling/sampling_logp_difference/max': 1.1450872421264648, 'sampling/importance_sampling_ratio/min': 0.3181961476802826, 'sampling/importance_sampling_ratio/mean': 0.9864500761032104, 'sampling/importance_sampling_ratio/max': 1.7654838562011719, 'entropy': 1.2468656301498413, 'clip_ratio/low_mean': 0.0709261978045106, 'clip_ratio/low_min': 0.0709261978045106, 'clip_ratio/high_mean': 0.05462184967473149, 'clip_ratio/high_max': 0.05462184967473149, 'clip_ratio/region_mean': 0.1255480474792421, 'reward_total_mean': 0.446069598197937, 'reward_meter_mean': 0.7041109800338745, 'reward_meter_std': 0.43137630820274353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.3123185336589813, 'reward_total_composite_mean': 0.446069598197937, 'reward_total_composite_std': 0.3866676986217499, 'epoch': 0.05}

 27%|██▋       | 881/3300 [3:16:11<6:11:17,  9.21s/it]INFO 04-12 19:14:37 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 882/3300 [3:16:24<7:04:11, 10.53s/it]2026-04-12 19:14:50,881 | INFO | train_grpo_train | metrics_logged mode=train step=882

                                                      
{'loss': -0.0125, 'grad_norm': 1.9595142602920532, 'learning_rate': 7.330303030303031e-06, 'num_tokens': 1572818.0, 'completions/mean_length': 263.375, 'completions/min_length': 14.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 14.75, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.1281031370162964, 'rewards/meter/std': 0.3201996684074402, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.774090051651001, 'rewards/lexical_plausibility/std': 0.18709072470664978, 'rewards/judge_quality/mean': 0.38625001907348633, 'rewards/judge_quality/std': 0.45226216316223145, 'rewards/repeat_penalty/mean': 0.8080371618270874, 'rewards/repeat_penalty/std': 0.11912072449922562, 'rewards/repeat_floor/mean': 0.9878881573677063, 'rewards/repeat_floor/std': 0.007844987325370312, 'rewards/near_duplicate_penalty/mean': 0.9945987462997437, 'rewards/near_duplicate_penalty/std': 0.01527698989957571, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9994259476661682, 'rewards/distinct_2/std': 0.0016236650990322232, 'rewards/total_composite/mean': 0.02596518024802208, 'rewards/total_composite/std': 0.0494607649743557, 'reward': 0.02596518024802208, 'reward_std': 0.049460768699645996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16541264951229095, 'sampling/sampling_logp_difference/max': 0.8870255947113037, 'sampling/importance_sampling_ratio/min': 0.4118790328502655, 'sampling/importance_sampling_ratio/mean': 1.0033903121948242, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5232321918010712, 'clip_ratio/low_mean': 0.07142857275903225, 'clip_ratio/low_min': 0.07142857275903225, 'clip_ratio/high_mean': 0.023634454235434532, 'clip_ratio/high_max': 0.023634454235434532, 'clip_ratio/region_mean': 0.09506302699446678, 'reward_total_mean': 0.02596518024802208, 'reward_meter_mean': 0.1281031370162964, 'reward_meter_std': 0.3201996684074402, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.774090051651001, 'reward_lexical_plausibility_std': 0.18709072470664978, 'reward_repeat_penalty_mean': 0.8080371618270874, 'reward_repeat_penalty_std': 0.11912072449922562, 'reward_repeat_floor_mean': 0.9878881573677063, 'reward_repeat_floor_std': 0.007844987325370312, 'reward_near_duplicate_penalty_mean': 0.9945987462997437, 'reward_near_duplicate_penalty_std': 0.01527698989957571, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9994259476661682, 'reward_distinct_2_std': 0.0016236650990322232, 'reward_judge_quality_mean': 0.38625001907348633, 'reward_judge_quality_std': 0.45226216316223145, 'reward_total_composite_mean': 0.02596518024802208, 'reward_total_composite_std': 0.0494607649743557, 'epoch': 0.05}

 27%|██▋       | 882/3300 [3:16:24<7:04:11, 10.53s/it]INFO 04-12 19:14:51 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 883/3300 [3:16:38<7:40:57, 11.44s/it]2026-04-12 19:15:04,461 | INFO | train_grpo_train | metrics_logged mode=train step=883

                                                      
{'loss': -0.0228, 'grad_norm': 3.174278736114502, 'learning_rate': 7.3272727272727285e-06, 'num_tokens': 1574342.0, 'completions/mean_length': 157.5, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 39.333335876464844, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.11787182092666626, 'rewards/meter/std': 0.1051332876086235, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8745204210281372, 'rewards/lexical_plausibility/std': 0.20271295309066772, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3703642189502716, 'rewards/repeat_penalty/mean': 0.9847168922424316, 'rewards/repeat_penalty/std': 0.03703960031270981, 'rewards/repeat_floor/mean': 0.9978303909301758, 'rewards/repeat_floor/std': 0.0052101886831223965, 'rewards/near_duplicate_penalty/mean': 0.9871079921722412, 'rewards/near_duplicate_penalty/std': 0.030315378680825233, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9973800182342529, 'rewards/distinct_2/std': 0.007410469930619001, 'rewards/total_composite/mean': 0.07036422193050385, 'rewards/total_composite/std': 0.10603836923837662, 'reward': 0.07036422193050385, 'reward_std': 0.10603835433721542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18420745432376862, 'sampling/sampling_logp_difference/max': 2.062805652618408, 'sampling/importance_sampling_ratio/min': 0.12709687650203705, 'sampling/importance_sampling_ratio/mean': 1.0141772031784058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9522935003042221, 'clip_ratio/low_mean': 0.0976778669282794, 'clip_ratio/low_min': 0.0976778669282794, 'clip_ratio/high_mean': 0.05233253724873066, 'clip_ratio/high_max': 0.05233253724873066, 'clip_ratio/region_mean': 0.15001040417701006, 'reward_total_mean': 0.07036422193050385, 'reward_meter_mean': 0.11787182092666626, 'reward_meter_std': 0.1051332876086235, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8745204210281372, 'reward_lexical_plausibility_std': 0.20271295309066772, 'reward_repeat_penalty_mean': 0.9847168922424316, 'reward_repeat_penalty_std': 0.03703960031270981, 'reward_repeat_floor_mean': 0.9978303909301758, 'reward_repeat_floor_std': 0.0052101886831223965, 'reward_near_duplicate_penalty_mean': 0.9871079921722412, 'reward_near_duplicate_penalty_std': 0.030315378680825233, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9973800182342529, 'reward_distinct_2_std': 0.007410469930619001, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3703642189502716, 'reward_total_composite_mean': 0.07036422193050385, 'reward_total_composite_std': 0.10603836923837662, 'epoch': 0.05}

 27%|██▋       | 883/3300 [3:16:38<7:40:57, 11.44s/it]INFO 04-12 19:15:04 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 884/3300 [3:16:52<8:17:49, 12.36s/it]2026-04-12 19:15:19,007 | INFO | train_grpo_train | metrics_logged mode=train step=884

                                                      
{'loss': -0.1543, 'grad_norm': 3.139519214630127, 'learning_rate': 7.324242424242425e-06, 'num_tokens': 1576288.0, 'completions/mean_length': 189.25, 'completions/min_length': 76.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 81.66667175292969, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.8730491399765015, 'rewards/meter/std': 0.1392127275466919, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9090895652770996, 'rewards/lexical_plausibility/std': 0.1704242080450058, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.7702693939208984, 'rewards/repeat_penalty/std': 0.12146493047475815, 'rewards/repeat_floor/mean': 0.9743232727050781, 'rewards/repeat_floor/std': 0.012096121907234192, 'rewards/near_duplicate_penalty/mean': 0.9344332218170166, 'rewards/near_duplicate_penalty/std': 0.011053124442696571, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8291887044906616, 'rewards/distinct_2/std': 0.11819184571504593, 'rewards/total_composite/mean': 0.23385152220726013, 'rewards/total_composite/std': 0.12986473739147186, 'reward': 0.23385152220726013, 'reward_std': 0.12986473739147186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14850276708602905, 'sampling/sampling_logp_difference/max': 2.5442075729370117, 'sampling/importance_sampling_ratio/min': 0.07853526622056961, 'sampling/importance_sampling_ratio/mean': 1.0233300924301147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7174096181988716, 'clip_ratio/low_mean': 0.01151315774768591, 'clip_ratio/low_min': 0.01151315774768591, 'clip_ratio/high_mean': 0.07357755722478032, 'clip_ratio/high_max': 0.07357755722478032, 'clip_ratio/region_mean': 0.08509071497246623, 'reward_total_mean': 0.23385152220726013, 'reward_meter_mean': 0.8730491399765015, 'reward_meter_std': 0.1392127275466919, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9090895652770996, 'reward_lexical_plausibility_std': 0.1704242080450058, 'reward_repeat_penalty_mean': 0.7702693939208984, 'reward_repeat_penalty_std': 0.12146493047475815, 'reward_repeat_floor_mean': 0.9743232727050781, 'reward_repeat_floor_std': 0.012096121907234192, 'reward_near_duplicate_penalty_mean': 0.9344332218170166, 'reward_near_duplicate_penalty_std': 0.011053124442696571, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8291887044906616, 'reward_distinct_2_std': 0.11819184571504593, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.23385152220726013, 'reward_total_composite_std': 0.12986473739147186, 'epoch': 0.05}

 27%|██▋       | 884/3300 [3:16:52<8:17:49, 12.36s/it]INFO 04-12 19:15:19 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 885/3300 [3:17:07<8:46:03, 13.07s/it]2026-04-12 19:15:33,692 | INFO | train_grpo_train | metrics_logged mode=train step=885

                                                      
{'loss': -0.0652, 'grad_norm': 1.9510021209716797, 'learning_rate': 7.321212121212122e-06, 'num_tokens': 1577752.0, 'completions/mean_length': 218.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 41.60000228881836, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7423527240753174, 'rewards/meter/std': 0.24679777026176453, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 0.8737708926200867, 'rewards/lexical_plausibility/std': 0.16459837555885315, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.9060587882995605, 'rewards/repeat_penalty/std': 0.14735999703407288, 'rewards/repeat_floor/mean': 0.9902509450912476, 'rewards/repeat_floor/std': 0.011406295001506805, 'rewards/near_duplicate_penalty/mean': 0.957493782043457, 'rewards/near_duplicate_penalty/std': 0.031095990911126137, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9625208377838135, 'rewards/distinct_2/std': 0.08306626975536346, 'rewards/total_composite/mean': 0.16588640213012695, 'rewards/total_composite/std': 0.1439305692911148, 'reward': 0.16588640213012695, 'reward_std': 0.1439305692911148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1448756605386734, 'sampling/sampling_logp_difference/max': 1.2516169548034668, 'sampling/importance_sampling_ratio/min': 0.28604191541671753, 'sampling/importance_sampling_ratio/mean': 1.014423131942749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6253417953848839, 'clip_ratio/low_mean': 0.04021598957479, 'clip_ratio/low_min': 0.04021598957479, 'clip_ratio/high_mean': 0.061585801653563976, 'clip_ratio/high_max': 0.061585801653563976, 'clip_ratio/region_mean': 0.10180179122835398, 'reward_total_mean': 0.16588640213012695, 'reward_meter_mean': 0.7423527240753174, 'reward_meter_std': 0.24679777026176453, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 0.8737708926200867, 'reward_lexical_plausibility_std': 0.16459837555885315, 'reward_repeat_penalty_mean': 0.9060587882995605, 'reward_repeat_penalty_std': 0.14735999703407288, 'reward_repeat_floor_mean': 0.9902509450912476, 'reward_repeat_floor_std': 0.011406295001506805, 'reward_near_duplicate_penalty_mean': 0.957493782043457, 'reward_near_duplicate_penalty_std': 0.031095990911126137, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9625208377838135, 'reward_distinct_2_std': 0.08306626975536346, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.16588640213012695, 'reward_total_composite_std': 0.1439305692911148, 'epoch': 0.05}

 27%|██▋       | 885/3300 [3:17:07<8:46:03, 13.07s/it]INFO 04-12 19:15:34 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 886/3300 [3:17:22<9:05:34, 13.56s/it]2026-04-12 19:15:48,391 | INFO | train_grpo_train | metrics_logged mode=train step=886

                                                      
{'loss': -0.0421, 'grad_norm': 5.369659423828125, 'learning_rate': 7.3181818181818186e-06, 'num_tokens': 1579313.0, 'completions/mean_length': 96.125, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.3362264633178711, 'rewards/meter/std': 0.38697749376296997, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9562819600105286, 'rewards/lexical_plausibility/std': 0.11593372374773026, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9162084460258484, 'rewards/repeat_penalty/std': 0.08807915449142456, 'rewards/repeat_floor/mean': 0.9913227558135986, 'rewards/repeat_floor/std': 0.007231482770293951, 'rewards/near_duplicate_penalty/mean': 0.9675122499465942, 'rewards/near_duplicate_penalty/std': 0.029377853497862816, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9785662889480591, 'rewards/distinct_2/std': 0.03976902738213539, 'rewards/total_composite/mean': 0.1362835168838501, 'rewards/total_composite/std': 0.1556205302476883, 'reward': 0.1362835168838501, 'reward_std': 0.1556205302476883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13743160665035248, 'sampling/sampling_logp_difference/max': 1.3717479705810547, 'sampling/importance_sampling_ratio/min': 0.25366318225860596, 'sampling/importance_sampling_ratio/mean': 1.024179458618164, 'sampling/importance_sampling_ratio/max': 1.7470049858093262, 'entropy': 0.7761737704277039, 'clip_ratio/low_mean': 0.07000825181603432, 'clip_ratio/low_min': 0.07000825181603432, 'clip_ratio/high_mean': 0.05300951562821865, 'clip_ratio/high_max': 0.05300951562821865, 'clip_ratio/region_mean': 0.12301776744425297, 'reward_total_mean': 0.1362835168838501, 'reward_meter_mean': 0.3362264633178711, 'reward_meter_std': 0.38697749376296997, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9562819600105286, 'reward_lexical_plausibility_std': 0.11593372374773026, 'reward_repeat_penalty_mean': 0.9162084460258484, 'reward_repeat_penalty_std': 0.08807915449142456, 'reward_repeat_floor_mean': 0.9913227558135986, 'reward_repeat_floor_std': 0.007231482770293951, 'reward_near_duplicate_penalty_mean': 0.9675122499465942, 'reward_near_duplicate_penalty_std': 0.029377853497862816, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9785662889480591, 'reward_distinct_2_std': 0.03976902738213539, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.1362835168838501, 'reward_total_composite_std': 0.1556205302476883, 'epoch': 0.05}

 27%|██▋       | 886/3300 [3:17:22<9:05:34, 13.56s/it]INFO 04-12 19:15:48 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 887/3300 [3:17:30<8:07:14, 12.12s/it]2026-04-12 19:15:57,152 | INFO | train_grpo_train | metrics_logged mode=train step=887

                                                      
{'loss': 0.083, 'grad_norm': 10.571259498596191, 'learning_rate': 7.315151515151516e-06, 'num_tokens': 1581050.0, 'completions/mean_length': 60.125, 'completions/min_length': 50.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.33594268560409546, 'rewards/meter/std': 0.2705962359905243, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.3208470940589905, 'rewards/repeat_penalty/mean': 0.7862205505371094, 'rewards/repeat_penalty/std': 0.2621435225009918, 'rewards/repeat_floor/mean': 0.9746034741401672, 'rewards/repeat_floor/std': 0.030592938885092735, 'rewards/near_duplicate_penalty/mean': 0.9286086559295654, 'rewards/near_duplicate_penalty/std': 0.05637591704726219, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8368022441864014, 'rewards/distinct_2/std': 0.2516466975212097, 'rewards/total_composite/mean': 0.19157472252845764, 'rewards/total_composite/std': 0.22701068222522736, 'reward': 0.19157472252845764, 'reward_std': 0.22701066732406616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13616812229156494, 'sampling/sampling_logp_difference/max': 2.5768704414367676, 'sampling/importance_sampling_ratio/min': 0.07601151615381241, 'sampling/importance_sampling_ratio/mean': 1.007124900817871, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8027744293212891, 'clip_ratio/low_mean': 0.06500709755346179, 'clip_ratio/low_min': 0.06500709755346179, 'clip_ratio/high_mean': 0.04464669618755579, 'clip_ratio/high_max': 0.04464669618755579, 'clip_ratio/region_mean': 0.10965379374101758, 'reward_total_mean': 0.19157472252845764, 'reward_meter_mean': 0.33594268560409546, 'reward_meter_std': 0.2705962359905243, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7862205505371094, 'reward_repeat_penalty_std': 0.2621435225009918, 'reward_repeat_floor_mean': 0.9746034741401672, 'reward_repeat_floor_std': 0.030592938885092735, 'reward_near_duplicate_penalty_mean': 0.9286086559295654, 'reward_near_duplicate_penalty_std': 0.05637591704726219, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8368022441864014, 'reward_distinct_2_std': 0.2516466975212097, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.3208470940589905, 'reward_total_composite_mean': 0.19157472252845764, 'reward_total_composite_std': 0.22701068222522736, 'epoch': 0.05}

 27%|██▋       | 887/3300 [3:17:31<8:07:14, 12.12s/it]INFO 04-12 19:15:57 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 888/3300 [3:17:45<8:36:28, 12.85s/it]2026-04-12 19:16:11,695 | INFO | train_grpo_train | metrics_logged mode=train step=888

                                                      
{'loss': -0.0432, 'grad_norm': 1.5164231061935425, 'learning_rate': 7.312121212121212e-06, 'num_tokens': 1582644.0, 'completions/mean_length': 456.25, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6885676383972168, 'rewards/meter/std': 0.4023575782775879, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.29124119877815247, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.08737237006425858, 'rewards/lexical_plausibility/mean': 0.7418726682662964, 'rewards/lexical_plausibility/std': 0.11394708603620529, 'rewards/judge_quality/mean': 0.07500000298023224, 'rewards/judge_quality/std': 0.11338933557271957, 'rewards/repeat_penalty/mean': 0.8229809999465942, 'rewards/repeat_penalty/std': 0.14408570528030396, 'rewards/repeat_floor/mean': 0.9849258661270142, 'rewards/repeat_floor/std': 0.012617240659892559, 'rewards/near_duplicate_penalty/mean': 0.9663236737251282, 'rewards/near_duplicate_penalty/std': 0.03803425282239914, 'rewards/opening_diversity/mean': 0.8779761791229248, 'rewards/opening_diversity/std': 0.14010241627693176, 'rewards/distinct_2/mean': 0.9699857831001282, 'rewards/distinct_2/std': 0.05932611972093582, 'rewards/total_composite/mean': 0.0594598613679409, 'rewards/total_composite/std': 0.10353420674800873, 'reward': 0.0594598613679409, 'reward_std': 0.10353420674800873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16618087887763977, 'sampling/sampling_logp_difference/max': 1.071451187133789, 'sampling/importance_sampling_ratio/min': 0.3425111174583435, 'sampling/importance_sampling_ratio/mean': 1.049685001373291, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.14603425562381744, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.011363636702299118, 'clip_ratio/high_max': 0.011363636702299118, 'clip_ratio/region_mean': 0.011363636702299118, 'reward_total_mean': 0.0594598613679409, 'reward_meter_mean': 0.6885676383972168, 'reward_meter_std': 0.4023575782775879, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.29124119877815247, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.08737237006425858, 'reward_lexical_plausibility_mean': 0.7418726682662964, 'reward_lexical_plausibility_std': 0.11394708603620529, 'reward_repeat_penalty_mean': 0.8229809999465942, 'reward_repeat_penalty_std': 0.14408570528030396, 'reward_repeat_floor_mean': 0.9849258661270142, 'reward_repeat_floor_std': 0.012617240659892559, 'reward_near_duplicate_penalty_mean': 0.9663236737251282, 'reward_near_duplicate_penalty_std': 0.03803425282239914, 'reward_opening_diversity_mean': 0.8779761791229248, 'reward_opening_diversity_std': 0.14010241627693176, 'reward_distinct_2_mean': 0.9699857831001282, 'reward_distinct_2_std': 0.05932611972093582, 'reward_judge_quality_mean': 0.07500000298023224, 'reward_judge_quality_std': 0.11338933557271957, 'reward_total_composite_mean': 0.0594598613679409, 'reward_total_composite_std': 0.10353420674800873, 'epoch': 0.05}

 27%|██▋       | 888/3300 [3:17:45<8:36:28, 12.85s/it]INFO 04-12 19:16:11 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 889/3300 [3:17:59<8:46:17, 13.10s/it]2026-04-12 19:16:25,415 | INFO | train_grpo_train | metrics_logged mode=train step=889

                                                      
{'loss': 0.0156, 'grad_norm': 5.1369194984436035, 'learning_rate': 7.30909090909091e-06, 'num_tokens': 1585146.0, 'completions/mean_length': 161.75, 'completions/min_length': 97.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 111.71429443359375, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.520438551902771, 'rewards/meter/std': 0.44028639793395996, 'rewards/count_adherence/mean': 0.8035714626312256, 'rewards/count_adherence/std': 0.10628911107778549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9410713911056519, 'rewards/count_floor/std': 0.03188672661781311, 'rewards/lexical_plausibility/mean': 0.9573183059692383, 'rewards/lexical_plausibility/std': 0.12072212249040604, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.22867102921009064, 'rewards/repeat_penalty/std': 0.2610289752483368, 'rewards/repeat_floor/mean': 0.8779460191726685, 'rewards/repeat_floor/std': 0.06991404294967651, 'rewards/near_duplicate_penalty/mean': 0.6592497825622559, 'rewards/near_duplicate_penalty/std': 0.2152763456106186, 'rewards/opening_diversity/mean': 0.807508111000061, 'rewards/opening_diversity/std': 0.23668500781059265, 'rewards/distinct_2/mean': 0.34008944034576416, 'rewards/distinct_2/std': 0.3077734410762787, 'rewards/total_composite/mean': 0.09737002849578857, 'rewards/total_composite/std': 0.10020838677883148, 'reward': 0.09737002849578857, 'reward_std': 0.10020838677883148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11012234538793564, 'sampling/sampling_logp_difference/max': 2.440465211868286, 'sampling/importance_sampling_ratio/min': 0.08712030947208405, 'sampling/importance_sampling_ratio/mean': 0.99152010679245, 'sampling/importance_sampling_ratio/max': 1.9557863473892212, 'entropy': 0.6145061030983925, 'clip_ratio/low_mean': 0.026263192296028137, 'clip_ratio/low_min': 0.026263192296028137, 'clip_ratio/high_mean': 0.07299808692187071, 'clip_ratio/high_max': 0.07299808692187071, 'clip_ratio/region_mean': 0.09926127921789885, 'reward_total_mean': 0.09737002849578857, 'reward_meter_mean': 0.520438551902771, 'reward_meter_std': 0.44028639793395996, 'reward_count_adherence_mean': 0.8035714626312256, 'reward_count_adherence_std': 0.10628911107778549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9410713911056519, 'reward_count_floor_std': 0.03188672661781311, 'reward_lexical_plausibility_mean': 0.9573183059692383, 'reward_lexical_plausibility_std': 0.12072212249040604, 'reward_repeat_penalty_mean': 0.22867102921009064, 'reward_repeat_penalty_std': 0.2610289752483368, 'reward_repeat_floor_mean': 0.8779460191726685, 'reward_repeat_floor_std': 0.06991404294967651, 'reward_near_duplicate_penalty_mean': 0.6592497825622559, 'reward_near_duplicate_penalty_std': 0.2152763456106186, 'reward_opening_diversity_mean': 0.807508111000061, 'reward_opening_diversity_std': 0.23668500781059265, 'reward_distinct_2_mean': 0.34008944034576416, 'reward_distinct_2_std': 0.3077734410762787, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.09737002849578857, 'reward_total_composite_std': 0.10020838677883148, 'epoch': 0.05}

 27%|██▋       | 889/3300 [3:17:59<8:46:17, 13.10s/it]INFO 04-12 19:16:25 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 890/3300 [3:18:13<8:57:35, 13.38s/it]2026-04-12 19:16:39,428 | INFO | train_grpo_train | metrics_logged mode=train step=890

                                                      
{'loss': -0.1014, 'grad_norm': 1.881895661354065, 'learning_rate': 7.306060606060607e-06, 'num_tokens': 1586957.0, 'completions/mean_length': 297.375, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 82.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.6624069809913635, 'rewards/meter/std': 0.40717947483062744, 'rewards/count_adherence/mean': 0.7000000476837158, 'rewards/count_adherence/std': 0.3023715913295746, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9099999666213989, 'rewards/count_floor/std': 0.09071147441864014, 'rewards/lexical_plausibility/mean': 0.8344358801841736, 'rewards/lexical_plausibility/std': 0.18998073041439056, 'rewards/judge_quality/mean': 0.18125000596046448, 'rewards/judge_quality/std': 0.17512750625610352, 'rewards/repeat_penalty/mean': 0.7679789066314697, 'rewards/repeat_penalty/std': 0.284628689289093, 'rewards/repeat_floor/mean': 0.9727243781089783, 'rewards/repeat_floor/std': 0.03439345583319664, 'rewards/near_duplicate_penalty/mean': 0.9201740026473999, 'rewards/near_duplicate_penalty/std': 0.07724828273057938, 'rewards/opening_diversity/mean': 0.871874988079071, 'rewards/opening_diversity/std': 0.24474385380744934, 'rewards/distinct_2/mean': 0.9153976440429688, 'rewards/distinct_2/std': 0.1134895533323288, 'rewards/total_composite/mean': 0.1020907312631607, 'rewards/total_composite/std': 0.11728008091449738, 'reward': 0.1020907312631607, 'reward_std': 0.11728007346391678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14078360795974731, 'sampling/sampling_logp_difference/max': 1.666956901550293, 'sampling/importance_sampling_ratio/min': 0.18882077932357788, 'sampling/importance_sampling_ratio/mean': 1.0064233541488647, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5357254147529602, 'clip_ratio/low_mean': 0.009615384973585606, 'clip_ratio/low_min': 0.009615384973585606, 'clip_ratio/high_mean': 0.05034242570400238, 'clip_ratio/high_max': 0.05034242570400238, 'clip_ratio/region_mean': 0.059957810677587986, 'reward_total_mean': 0.1020907312631607, 'reward_meter_mean': 0.6624069809913635, 'reward_meter_std': 0.40717947483062744, 'reward_count_adherence_mean': 0.7000000476837158, 'reward_count_adherence_std': 0.3023715913295746, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9099999666213989, 'reward_count_floor_std': 0.09071147441864014, 'reward_lexical_plausibility_mean': 0.8344358801841736, 'reward_lexical_plausibility_std': 0.18998073041439056, 'reward_repeat_penalty_mean': 0.7679789066314697, 'reward_repeat_penalty_std': 0.284628689289093, 'reward_repeat_floor_mean': 0.9727243781089783, 'reward_repeat_floor_std': 0.03439345583319664, 'reward_near_duplicate_penalty_mean': 0.9201740026473999, 'reward_near_duplicate_penalty_std': 0.07724828273057938, 'reward_opening_diversity_mean': 0.871874988079071, 'reward_opening_diversity_std': 0.24474385380744934, 'reward_distinct_2_mean': 0.9153976440429688, 'reward_distinct_2_std': 0.1134895533323288, 'reward_judge_quality_mean': 0.18125000596046448, 'reward_judge_quality_std': 0.17512750625610352, 'reward_total_composite_mean': 0.1020907312631607, 'reward_total_composite_std': 0.11728008091449738, 'epoch': 0.05}

 27%|██▋       | 890/3300 [3:18:13<8:57:35, 13.38s/it]INFO 04-12 19:16:39 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 891/3300 [3:18:27<9:09:17, 13.68s/it]2026-04-12 19:16:53,804 | INFO | train_grpo_train | metrics_logged mode=train step=891

                                                      
{'loss': -0.1214, 'grad_norm': 4.463621616363525, 'learning_rate': 7.303030303030304e-06, 'num_tokens': 1588814.0, 'completions/mean_length': 125.125, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.85714721679688, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.8792005777359009, 'rewards/meter/std': 0.1741088628768921, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9465920925140381, 'rewards/lexical_plausibility/std': 0.13618627190589905, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.5569837093353271, 'rewards/repeat_penalty/std': 0.3384759724140167, 'rewards/repeat_floor/mean': 0.9390017986297607, 'rewards/repeat_floor/std': 0.05249625816941261, 'rewards/near_duplicate_penalty/mean': 0.8076207637786865, 'rewards/near_duplicate_penalty/std': 0.14858025312423706, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.31160587072372437, 'rewards/distinct_2/mean': 0.7887073755264282, 'rewards/distinct_2/std': 0.18778012692928314, 'rewards/total_composite/mean': 0.2400050163269043, 'rewards/total_composite/std': 0.11523940414190292, 'reward': 0.2400050163269043, 'reward_std': 0.11523939669132233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14740332961082458, 'sampling/sampling_logp_difference/max': 2.7489712238311768, 'sampling/importance_sampling_ratio/min': 0.0639936625957489, 'sampling/importance_sampling_ratio/mean': 1.0004315376281738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7213696390390396, 'clip_ratio/low_mean': 0.02547417301684618, 'clip_ratio/low_min': 0.02547417301684618, 'clip_ratio/high_mean': 0.085636455565691, 'clip_ratio/high_max': 0.085636455565691, 'clip_ratio/region_mean': 0.11111062858253717, 'reward_total_mean': 0.2400050163269043, 'reward_meter_mean': 0.8792005777359009, 'reward_meter_std': 0.1741088628768921, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9465920925140381, 'reward_lexical_plausibility_std': 0.13618627190589905, 'reward_repeat_penalty_mean': 0.5569837093353271, 'reward_repeat_penalty_std': 0.3384759724140167, 'reward_repeat_floor_mean': 0.9390017986297607, 'reward_repeat_floor_std': 0.05249625816941261, 'reward_near_duplicate_penalty_mean': 0.8076207637786865, 'reward_near_duplicate_penalty_std': 0.14858025312423706, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.31160587072372437, 'reward_distinct_2_mean': 0.7887073755264282, 'reward_distinct_2_std': 0.18778012692928314, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.2400050163269043, 'reward_total_composite_std': 0.11523940414190292, 'epoch': 0.05}

 27%|██▋       | 891/3300 [3:18:27<9:09:17, 13.68s/it]INFO 04-12 19:16:54 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 892/3300 [3:18:46<10:11:15, 15.23s/it]2026-04-12 19:17:12,713 | INFO | train_grpo_train | metrics_logged mode=train step=892

                                                       
{'loss': -0.0341, 'grad_norm': 2.5688624382019043, 'learning_rate': 7.3e-06, 'num_tokens': 1590283.0, 'completions/mean_length': 154.625, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5335460901260376, 'rewards/meter/std': 0.48206907510757446, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9092497229576111, 'rewards/lexical_plausibility/std': 0.168043315410614, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.3340979516506195, 'rewards/repeat_penalty/mean': 0.8412985801696777, 'rewards/repeat_penalty/std': 0.18187516927719116, 'rewards/repeat_floor/mean': 0.9852879047393799, 'rewards/repeat_floor/std': 0.019465075805783272, 'rewards/near_duplicate_penalty/mean': 0.97127366065979, 'rewards/near_duplicate_penalty/std': 0.0466519370675087, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9469095468521118, 'rewards/distinct_2/std': 0.09718659520149231, 'rewards/total_composite/mean': 0.2621089816093445, 'rewards/total_composite/std': 0.2744298279285431, 'reward': 0.2621089816093445, 'reward_std': 0.2744298279285431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16270951926708221, 'sampling/sampling_logp_difference/max': 1.796198844909668, 'sampling/importance_sampling_ratio/min': 0.1659284085035324, 'sampling/importance_sampling_ratio/mean': 1.0161446332931519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8777428865432739, 'clip_ratio/low_mean': 0.03945652116090059, 'clip_ratio/low_min': 0.03945652116090059, 'clip_ratio/high_mean': 0.08518800977617502, 'clip_ratio/high_max': 0.08518800977617502, 'clip_ratio/region_mean': 0.12464453093707561, 'reward_total_mean': 0.2621089816093445, 'reward_meter_mean': 0.5335460901260376, 'reward_meter_std': 0.48206907510757446, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9092497229576111, 'reward_lexical_plausibility_std': 0.168043315410614, 'reward_repeat_penalty_mean': 0.8412985801696777, 'reward_repeat_penalty_std': 0.18187516927719116, 'reward_repeat_floor_mean': 0.9852879047393799, 'reward_repeat_floor_std': 0.019465075805783272, 'reward_near_duplicate_penalty_mean': 0.97127366065979, 'reward_near_duplicate_penalty_std': 0.0466519370675087, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9469095468521118, 'reward_distinct_2_std': 0.09718659520149231, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.3340979516506195, 'reward_total_composite_mean': 0.2621089816093445, 'reward_total_composite_std': 0.2744298279285431, 'epoch': 0.05}

 27%|██▋       | 892/3300 [3:18:46<10:11:15, 15.23s/it]INFO 04-12 19:17:12 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 893/3300 [3:18:55<9:01:00, 13.49s/it] 2026-04-12 19:17:22,070 | INFO | train_grpo_train | metrics_logged mode=train step=893

                                                      
{'loss': -0.0423, 'grad_norm': 16.598234176635742, 'learning_rate': 7.296969696969698e-06, 'num_tokens': 1591919.0, 'completions/mean_length': 37.5, 'completions/min_length': 29.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8038854598999023, 'rewards/meter/std': 0.3313772976398468, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24706491827964783, 'rewards/repeat_penalty/mean': 0.6009575128555298, 'rewards/repeat_penalty/std': 0.39226582646369934, 'rewards/repeat_floor/mean': 0.9393091201782227, 'rewards/repeat_floor/std': 0.07334894686937332, 'rewards/near_duplicate_penalty/mean': 0.8244130611419678, 'rewards/near_duplicate_penalty/std': 0.22831372916698456, 'rewards/opening_diversity/mean': 0.768750011920929, 'rewards/opening_diversity/std': 0.26415297389030457, 'rewards/distinct_2/mean': 0.7724685668945312, 'rewards/distinct_2/std': 0.2670818865299225, 'rewards/total_composite/mean': 0.2674075961112976, 'rewards/total_composite/std': 0.15334029495716095, 'reward': 0.2674075961112976, 'reward_std': 0.15334029495716095, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19589951634407043, 'sampling/sampling_logp_difference/max': 2.0359065532684326, 'sampling/importance_sampling_ratio/min': 0.13056208193302155, 'sampling/importance_sampling_ratio/mean': 1.026831030845642, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1646473929286003, 'clip_ratio/low_mean': 0.06571322120726109, 'clip_ratio/low_min': 0.06571322120726109, 'clip_ratio/high_mean': 0.08612309023737907, 'clip_ratio/high_max': 0.08612309023737907, 'clip_ratio/region_mean': 0.15183631144464016, 'reward_total_mean': 0.2674075961112976, 'reward_meter_mean': 0.8038854598999023, 'reward_meter_std': 0.3313772976398468, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.6009575128555298, 'reward_repeat_penalty_std': 0.39226582646369934, 'reward_repeat_floor_mean': 0.9393091201782227, 'reward_repeat_floor_std': 0.07334894686937332, 'reward_near_duplicate_penalty_mean': 0.8244130611419678, 'reward_near_duplicate_penalty_std': 0.22831372916698456, 'reward_opening_diversity_mean': 0.768750011920929, 'reward_opening_diversity_std': 0.26415297389030457, 'reward_distinct_2_mean': 0.7724685668945312, 'reward_distinct_2_std': 0.2670818865299225, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24706491827964783, 'reward_total_composite_mean': 0.2674075961112976, 'reward_total_composite_std': 0.15334029495716095, 'epoch': 0.05}

 27%|██▋       | 893/3300 [3:18:55<9:01:00, 13.49s/it]INFO 04-12 19:17:22 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 894/3300 [3:19:05<8:16:54, 12.39s/it]2026-04-12 19:17:31,907 | INFO | train_grpo_train | metrics_logged mode=train step=894

                                                      
{'loss': 0.1559, 'grad_norm': 10.017705917358398, 'learning_rate': 7.293939393939394e-06, 'num_tokens': 1594311.0, 'completions/mean_length': 97.0, 'completions/min_length': 41.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.7279128432273865, 'rewards/meter/std': 0.3662586510181427, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.1767766773700714, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.918749988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.24379447102546692, 'rewards/repeat_penalty/std': 0.23227240145206451, 'rewards/repeat_floor/mean': 0.8775709867477417, 'rewards/repeat_floor/std': 0.07003941386938095, 'rewards/near_duplicate_penalty/mean': 0.6747478246688843, 'rewards/near_duplicate_penalty/std': 0.20920851826667786, 'rewards/opening_diversity/mean': 0.7244791984558105, 'rewards/opening_diversity/std': 0.3409140706062317, 'rewards/distinct_2/mean': 0.36544474959373474, 'rewards/distinct_2/std': 0.2325313836336136, 'rewards/total_composite/mean': 0.0848173201084137, 'rewards/total_composite/std': 0.039417754858732224, 'reward': 0.0848173201084137, 'reward_std': 0.039417751133441925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1081172302365303, 'sampling/sampling_logp_difference/max': 4.45536994934082, 'sampling/importance_sampling_ratio/min': 0.011616022326052189, 'sampling/importance_sampling_ratio/mean': 1.0248385667800903, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7483176216483116, 'clip_ratio/low_mean': 0.04287295136600733, 'clip_ratio/low_min': 0.04287295136600733, 'clip_ratio/high_mean': 0.07209587749093771, 'clip_ratio/high_max': 0.07209587749093771, 'clip_ratio/region_mean': 0.11496882885694504, 'reward_total_mean': 0.0848173201084137, 'reward_meter_mean': 0.7279128432273865, 'reward_meter_std': 0.3662586510181427, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.1767766773700714, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.918749988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.24379447102546692, 'reward_repeat_penalty_std': 0.23227240145206451, 'reward_repeat_floor_mean': 0.8775709867477417, 'reward_repeat_floor_std': 0.07003941386938095, 'reward_near_duplicate_penalty_mean': 0.6747478246688843, 'reward_near_duplicate_penalty_std': 0.20920851826667786, 'reward_opening_diversity_mean': 0.7244791984558105, 'reward_opening_diversity_std': 0.3409140706062317, 'reward_distinct_2_mean': 0.36544474959373474, 'reward_distinct_2_std': 0.2325313836336136, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.0848173201084137, 'reward_total_composite_std': 0.039417754858732224, 'epoch': 0.05}

 27%|██▋       | 894/3300 [3:19:05<8:16:54, 12.39s/it]INFO 04-12 19:17:32 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 895/3300 [3:19:18<8:23:25, 12.56s/it]2026-04-12 19:17:44,971 | INFO | train_grpo_train | metrics_logged mode=train step=895

                                                      
{'loss': -0.0245, 'grad_norm': 10.260034561157227, 'learning_rate': 7.290909090909092e-06, 'num_tokens': 1595761.0, 'completions/mean_length': 91.25, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.142858505249023, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.7115216255187988, 'rewards/meter/std': 0.3363083302974701, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9287829399108887, 'rewards/lexical_plausibility/std': 0.20143233239650726, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.7423636317253113, 'rewards/repeat_penalty/std': 0.28592798113822937, 'rewards/repeat_floor/mean': 0.9711732864379883, 'rewards/repeat_floor/std': 0.030263379216194153, 'rewards/near_duplicate_penalty/mean': 0.9118344783782959, 'rewards/near_duplicate_penalty/std': 0.07457102090120316, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8891459703445435, 'rewards/distinct_2/std': 0.1339576095342636, 'rewards/total_composite/mean': 0.21806921064853668, 'rewards/total_composite/std': 0.1697385460138321, 'reward': 0.21806921064853668, 'reward_std': 0.1697385460138321, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14132149517536163, 'sampling/sampling_logp_difference/max': 1.754323959350586, 'sampling/importance_sampling_ratio/min': 0.17302417755126953, 'sampling/importance_sampling_ratio/mean': 1.0358550548553467, 'sampling/importance_sampling_ratio/max': 1.7814950942993164, 'entropy': 0.8144881129264832, 'clip_ratio/low_mean': 0.05141128972172737, 'clip_ratio/low_min': 0.05141128972172737, 'clip_ratio/high_mean': 0.06525601074099541, 'clip_ratio/high_max': 0.06525601074099541, 'clip_ratio/region_mean': 0.11666730046272278, 'reward_total_mean': 0.21806921064853668, 'reward_meter_mean': 0.7115216255187988, 'reward_meter_std': 0.3363083302974701, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9287829399108887, 'reward_lexical_plausibility_std': 0.20143233239650726, 'reward_repeat_penalty_mean': 0.7423636317253113, 'reward_repeat_penalty_std': 0.28592798113822937, 'reward_repeat_floor_mean': 0.9711732864379883, 'reward_repeat_floor_std': 0.030263379216194153, 'reward_near_duplicate_penalty_mean': 0.9118344783782959, 'reward_near_duplicate_penalty_std': 0.07457102090120316, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8891459703445435, 'reward_distinct_2_std': 0.1339576095342636, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.21806921064853668, 'reward_total_composite_std': 0.1697385460138321, 'epoch': 0.05}

 27%|██▋       | 895/3300 [3:19:18<8:23:25, 12.56s/it]INFO 04-12 19:17:45 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 896/3300 [3:19:32<8:42:04, 13.03s/it]2026-04-12 19:17:58,993 | INFO | train_grpo_train | metrics_logged mode=train step=896

                                                      
{'loss': -0.0778, 'grad_norm': 1.8823866844177246, 'learning_rate': 7.287878787878789e-06, 'num_tokens': 1597392.0, 'completions/mean_length': 164.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 49.16666793823242, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9314152002334595, 'rewards/meter/std': 0.15830190479755402, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9106531739234924, 'rewards/lexical_plausibility/std': 0.16617372632026672, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.2314550131559372, 'rewards/repeat_penalty/mean': 0.7977739572525024, 'rewards/repeat_penalty/std': 0.25435206294059753, 'rewards/repeat_floor/mean': 0.9763270020484924, 'rewards/repeat_floor/std': 0.030593838542699814, 'rewards/near_duplicate_penalty/mean': 0.9393112659454346, 'rewards/near_duplicate_penalty/std': 0.0704958438873291, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.888114333152771, 'rewards/distinct_2/std': 0.15373705327510834, 'rewards/total_composite/mean': 0.3430384397506714, 'rewards/total_composite/std': 0.22983165085315704, 'reward': 0.3430384397506714, 'reward_std': 0.22983163595199585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13058602809906006, 'sampling/sampling_logp_difference/max': 1.4625039100646973, 'sampling/importance_sampling_ratio/min': 0.2316555231809616, 'sampling/importance_sampling_ratio/mean': 1.0431114435195923, 'sampling/importance_sampling_ratio/max': 1.8963088989257812, 'entropy': 0.8591323718428612, 'clip_ratio/low_mean': 0.018145160749554634, 'clip_ratio/low_min': 0.018145160749554634, 'clip_ratio/high_mean': 0.07727300748229027, 'clip_ratio/high_max': 0.07727300748229027, 'clip_ratio/region_mean': 0.0954181682318449, 'reward_total_mean': 0.3430384397506714, 'reward_meter_mean': 0.9314152002334595, 'reward_meter_std': 0.15830190479755402, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9106531739234924, 'reward_lexical_plausibility_std': 0.16617372632026672, 'reward_repeat_penalty_mean': 0.7977739572525024, 'reward_repeat_penalty_std': 0.25435206294059753, 'reward_repeat_floor_mean': 0.9763270020484924, 'reward_repeat_floor_std': 0.030593838542699814, 'reward_near_duplicate_penalty_mean': 0.9393112659454346, 'reward_near_duplicate_penalty_std': 0.0704958438873291, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.888114333152771, 'reward_distinct_2_std': 0.15373705327510834, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.2314550131559372, 'reward_total_composite_mean': 0.3430384397506714, 'reward_total_composite_std': 0.22983165085315704, 'epoch': 0.05}

 27%|██▋       | 896/3300 [3:19:32<8:42:04, 13.03s/it]INFO 04-12 19:17:59 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 897/3300 [3:19:41<7:49:08, 11.71s/it]2026-04-12 19:18:07,622 | INFO | train_grpo_train | metrics_logged mode=train step=897

                                                      
{'loss': 0.041, 'grad_norm': 16.027273178100586, 'learning_rate': 7.284848484848486e-06, 'num_tokens': 1598876.0, 'completions/mean_length': 40.5, 'completions/min_length': 31.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.9046456813812256, 'rewards/meter/std': 0.13706478476524353, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.7206360697746277, 'rewards/repeat_penalty/std': 0.28052711486816406, 'rewards/repeat_floor/mean': 0.9680224657058716, 'rewards/repeat_floor/std': 0.03394155576825142, 'rewards/near_duplicate_penalty/mean': 0.8939769268035889, 'rewards/near_duplicate_penalty/std': 0.10728864371776581, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8630656003952026, 'rewards/distinct_2/std': 0.15047280490398407, 'rewards/total_composite/mean': 0.25389695167541504, 'rewards/total_composite/std': 0.11473076045513153, 'reward': 0.25389695167541504, 'reward_std': 0.11473074555397034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15610337257385254, 'sampling/sampling_logp_difference/max': 1.4549541473388672, 'sampling/importance_sampling_ratio/min': 0.2334110587835312, 'sampling/importance_sampling_ratio/mean': 1.0133109092712402, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.24332857131958, 'clip_ratio/low_mean': 0.06171528622508049, 'clip_ratio/low_min': 0.06171528622508049, 'clip_ratio/high_mean': 0.09191965963691473, 'clip_ratio/high_max': 0.09191965963691473, 'clip_ratio/region_mean': 0.15363494586199522, 'reward_total_mean': 0.25389695167541504, 'reward_meter_mean': 0.9046456813812256, 'reward_meter_std': 0.13706478476524353, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7206360697746277, 'reward_repeat_penalty_std': 0.28052711486816406, 'reward_repeat_floor_mean': 0.9680224657058716, 'reward_repeat_floor_std': 0.03394155576825142, 'reward_near_duplicate_penalty_mean': 0.8939769268035889, 'reward_near_duplicate_penalty_std': 0.10728864371776581, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8630656003952026, 'reward_distinct_2_std': 0.15047280490398407, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.25389695167541504, 'reward_total_composite_std': 0.11473076045513153, 'epoch': 0.05}

 27%|██▋       | 897/3300 [3:19:41<7:49:08, 11.71s/it]INFO 04-12 19:18:07 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 898/3300 [3:19:49<7:10:16, 10.75s/it]2026-04-12 19:18:16,135 | INFO | train_grpo_train | metrics_logged mode=train step=898

                                                      
{'loss': -0.0027, 'grad_norm': 14.411706924438477, 'learning_rate': 7.281818181818182e-06, 'num_tokens': 1600376.0, 'completions/mean_length': 41.5, 'completions/min_length': 38.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7468670010566711, 'rewards/meter/std': 0.3747076392173767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.542788028717041, 'rewards/repeat_penalty/std': 0.2442883551120758, 'rewards/repeat_floor/mean': 0.9461007118225098, 'rewards/repeat_floor/std': 0.029961973428726196, 'rewards/near_duplicate_penalty/mean': 0.8603014945983887, 'rewards/near_duplicate_penalty/std': 0.0887211263179779, 'rewards/opening_diversity/mean': 0.765625, 'rewards/opening_diversity/std': 0.19408094882965088, 'rewards/distinct_2/mean': 0.7902001738548279, 'rewards/distinct_2/std': 0.10351910442113876, 'rewards/total_composite/mean': 0.39872249960899353, 'rewards/total_composite/std': 0.3332768380641937, 'reward': 0.39872249960899353, 'reward_std': 0.3332768380641937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11950510740280151, 'sampling/sampling_logp_difference/max': 1.0952110290527344, 'sampling/importance_sampling_ratio/min': 0.3344690203666687, 'sampling/importance_sampling_ratio/mean': 0.9956321716308594, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7607012093067169, 'clip_ratio/low_mean': 0.07318657310679555, 'clip_ratio/low_min': 0.07318657310679555, 'clip_ratio/high_mean': 0.050496033392846584, 'clip_ratio/high_max': 0.050496033392846584, 'clip_ratio/region_mean': 0.12368260649964213, 'reward_total_mean': 0.39872249960899353, 'reward_meter_mean': 0.7468670010566711, 'reward_meter_std': 0.3747076392173767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.542788028717041, 'reward_repeat_penalty_std': 0.2442883551120758, 'reward_repeat_floor_mean': 0.9461007118225098, 'reward_repeat_floor_std': 0.029961973428726196, 'reward_near_duplicate_penalty_mean': 0.8603014945983887, 'reward_near_duplicate_penalty_std': 0.0887211263179779, 'reward_opening_diversity_mean': 0.765625, 'reward_opening_diversity_std': 0.19408094882965088, 'reward_distinct_2_mean': 0.7902001738548279, 'reward_distinct_2_std': 0.10351910442113876, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.39872249960899353, 'reward_total_composite_std': 0.3332768380641937, 'epoch': 0.05}

 27%|██▋       | 898/3300 [3:19:49<7:10:16, 10.75s/it]INFO 04-12 19:18:16 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 899/3300 [3:20:03<7:45:40, 11.64s/it]2026-04-12 19:18:29,827 | INFO | train_grpo_train | metrics_logged mode=train step=899

                                                      
{'loss': -0.0262, 'grad_norm': 4.141161918640137, 'learning_rate': 7.2787878787878795e-06, 'num_tokens': 1601723.0, 'completions/mean_length': 81.375, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.85714340209961, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.6805827617645264, 'rewards/meter/std': 0.42947661876678467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9612054824829102, 'rewards/lexical_plausibility/std': 0.10972744971513748, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.3543202877044678, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4398835599422455, 'rewards/total_composite/std': 0.37205806374549866, 'reward': 0.4398835599422455, 'reward_std': 0.37205806374549866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19168579578399658, 'sampling/sampling_logp_difference/max': 1.7755107879638672, 'sampling/importance_sampling_ratio/min': 0.16939689218997955, 'sampling/importance_sampling_ratio/mean': 1.0172382593154907, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3149924129247665, 'clip_ratio/low_mean': 0.10765664838254452, 'clip_ratio/low_min': 0.10765664838254452, 'clip_ratio/high_mean': 0.03098370973020792, 'clip_ratio/high_max': 0.03098370973020792, 'clip_ratio/region_mean': 0.13864035811275244, 'reward_total_mean': 0.4398835599422455, 'reward_meter_mean': 0.6805827617645264, 'reward_meter_std': 0.42947661876678467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9612054824829102, 'reward_lexical_plausibility_std': 0.10972744971513748, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.3543202877044678, 'reward_total_composite_mean': 0.4398835599422455, 'reward_total_composite_std': 0.37205806374549866, 'epoch': 0.05}

 27%|██▋       | 899/3300 [3:20:03<7:45:40, 11.64s/it]INFO 04-12 19:18:30 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 900/3300 [3:20:12<7:16:00, 10.90s/it]2026-04-12 19:18:39,007 | INFO | train_grpo_train | metrics_logged mode=train step=900

                                                      
{'loss': 0.1995, 'grad_norm': 14.915257453918457, 'learning_rate': 7.275757575757576e-06, 'num_tokens': 1603219.0, 'completions/mean_length': 36.0, 'completions/min_length': 20.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.40367305278778076, 'rewards/meter/std': 0.37825754284858704, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.909568727016449, 'rewards/repeat_penalty/std': 0.1032397598028183, 'rewards/repeat_floor/mean': 0.9930744171142578, 'rewards/repeat_floor/std': 0.005789777263998985, 'rewards/near_duplicate_penalty/mean': 0.9893404245376587, 'rewards/near_duplicate_penalty/std': 0.009705417789518833, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9824812412261963, 'rewards/distinct_2/std': 0.03248300403356552, 'rewards/total_composite/mean': 0.1482512354850769, 'rewards/total_composite/std': 0.11045444011688232, 'reward': 0.1482512354850769, 'reward_std': 0.11045444011688232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16865186393260956, 'sampling/sampling_logp_difference/max': 1.6643738746643066, 'sampling/importance_sampling_ratio/min': 0.18930914998054504, 'sampling/importance_sampling_ratio/mean': 1.0059375762939453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1139141097664833, 'clip_ratio/low_mean': 0.07056170515716076, 'clip_ratio/low_min': 0.07056170515716076, 'clip_ratio/high_mean': 0.05067640822380781, 'clip_ratio/high_max': 0.05067640822380781, 'clip_ratio/region_mean': 0.12123811338096857, 'reward_total_mean': 0.1482512354850769, 'reward_meter_mean': 0.40367305278778076, 'reward_meter_std': 0.37825754284858704, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.909568727016449, 'reward_repeat_penalty_std': 0.1032397598028183, 'reward_repeat_floor_mean': 0.9930744171142578, 'reward_repeat_floor_std': 0.005789777263998985, 'reward_near_duplicate_penalty_mean': 0.9893404245376587, 'reward_near_duplicate_penalty_std': 0.009705417789518833, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9824812412261963, 'reward_distinct_2_std': 0.03248300403356552, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1482512354850769, 'reward_total_composite_std': 0.11045444011688232, 'epoch': 0.05}

 27%|██▋       | 900/3300 [3:20:12<7:16:00, 10.90s/it]INFO 04-12 19:18:39 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:03,  5.76s/it]

 23%|██▎       | 3/13 [00:22<01:20,  8.03s/it]

 31%|███       | 4/13 [00:36<01:31, 10.17s/it]

 38%|███▊      | 5/13 [00:44<01:15,  9.46s/it]

 46%|████▌     | 6/13 [00:58<01:16, 10.87s/it]

 54%|█████▍    | 7/13 [01:11<01:09, 11.64s/it]

 62%|██████▏   | 8/13 [01:22<00:57, 11.52s/it]

 69%|██████▉   | 9/13 [01:34<00:46, 11.67s/it]

 77%|███████▋  | 10/13 [01:50<00:38, 12.82s/it]

 85%|████████▍ | 11/13 [02:01<00:24, 12.37s/it]

 92%|█████████▏| 12/13 [02:13<00:12, 12.27s/it]

100%|██████████| 13/13 [02:25<00:00, 12.00s/it]2026-04-12 19:21:15,970 | INFO | train_grpo_train | metrics_logged mode=eval step=900

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 156.8828, 'eval_samples_per_second': 0.663, 'eval_steps_per_second': 0.083, 'eval_num_tokens': 1603219.0, 'eval_completions/mean_length': 163.98076923076923, 'eval_completions/min_length': 31.23076923076923, 'eval_completions/max_length': 471.3076923076923, 'eval_completions/clipped_ratio': 0.14423076923076922, 'eval_completions/mean_terminated_length': 104.68800706129808, 'eval_completions/min_terminated_length': 31.23076923076923, 'eval_completions/max_terminated_length': 216.84615384615384, 'eval_rewards/meter/mean': 0.42765475465701175, 'eval_rewards/meter/std': 0.37613093852996826, 'eval_rewards/count_adherence/mean': 0.8173322219115037, 'eval_rewards/count_adherence/std': 0.19084912767777076, 'eval_rewards/arabic_clean/mean': 0.8269230769230769, 'eval_rewards/arabic_clean/std': 0.35383587846389186, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9971153827813956, 'eval_rewards/arabic_floor/std': 0.008158923341677738, 'eval_rewards/count_floor/mean': 0.9451996729924128, 'eval_rewards/count_floor/std': 0.05725473781617788, 'eval_rewards/lexical_plausibility/mean': 0.9505092685039227, 'eval_rewards/lexical_plausibility/std': 0.10913157921570998, 'eval_rewards/judge_quality/mean': 0.2812500022924863, 'eval_rewards/judge_quality/std': 0.21458769618318632, 'eval_rewards/repeat_penalty/mean': 0.4704929475600903, 'eval_rewards/repeat_penalty/std': 0.35082630125375897, 'eval_rewards/repeat_floor/mean': 0.9265837761072012, 'eval_rewards/repeat_floor/std': 0.06258979840920521, 'eval_rewards/near_duplicate_penalty/mean': 0.8103658235990084, 'eval_rewards/near_duplicate_penalty/std': 0.18106824732743776, 'eval_rewards/opening_diversity/mean': 0.8588235561664288, 'eval_rewards/opening_diversity/std': 0.20346174732996866, 'eval_rewards/distinct_2/mean': 0.5944387568877294, 'eval_rewards/distinct_2/std': 0.3562886818097188, 'eval_rewards/total_composite/mean': 0.1092161713884427, 'eval_rewards/total_composite/std': 0.13914034028465933, 'eval_reward': 0.1092161713884427, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.059689754190353245, 'eval_sampling/sampling_logp_difference/max': 1.034017196068397, 'eval_sampling/importance_sampling_ratio/min': 0.3621820853306697, 'eval_sampling/importance_sampling_ratio/mean': 1.0169028410544763, 'eval_sampling/importance_sampling_ratio/max': 1.5208639731773963, 'eval_entropy': 0.6558135564510639, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1092161713884427, 'eval_reward_meter_mean': 0.42765475465701175, 'eval_reward_meter_std': 0.37613093852996826, 'eval_reward_count_adherence_mean': 0.8173322219115037, 'eval_reward_count_adherence_std': 0.19084912767777076, 'eval_reward_arabic_clean_mean': 0.8269230769230769, 'eval_reward_arabic_clean_std': 0.35383587846389186, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9971153827813956, 'eval_reward_arabic_floor_std': 0.008158923341677738, 'eval_reward_count_floor_mean': 0.9451996729924128, 'eval_reward_count_floor_std': 0.05725473781617788, 'eval_reward_lexical_plausibility_mean': 0.9505092685039227, 'eval_reward_lexical_plausibility_std': 0.10913157921570998, 'eval_reward_repeat_penalty_mean': 0.4704929475600903, 'eval_reward_repeat_penalty_std': 0.35082630125375897, 'eval_reward_repeat_floor_mean': 0.9265837761072012, 'eval_reward_repeat_floor_std': 0.06258979840920521, 'eval_reward_near_duplicate_penalty_mean': 0.8103658235990084, 'eval_reward_near_duplicate_penalty_std': 0.18106824732743776, 'eval_reward_opening_diversity_mean': 0.8588235561664288, 'eval_reward_opening_diversity_std': 0.20346174732996866, 'eval_reward_distinct_2_mean': 0.5944387568877294, 'eval_reward_distinct_2_std': 0.3562886818097188, 'eval_reward_judge_quality_mean': 0.2812500022924863, 'eval_reward_judge_quality_std': 0.21458769618318632, 'eval_reward_total_composite_mean': 0.1092161713884427, 'eval_reward_total_composite_std': 0.13914034028465933, 'epoch': 0.05}

 27%|██▋       | 900/3300 [3:22:49<7:16:00, 10.90s/it]

100%|██████████| 13/13 [02:25<00:00, 12.00s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 27%|██▋       | 901/3300 [3:23:02<38:55:43, 58.42s/it]2026-04-12 19:21:28,303 | INFO | train_grpo_train | metrics_logged mode=train step=901

                                                       
{'loss': 0.0128, 'grad_norm': 14.927328109741211, 'learning_rate': 7.272727272727273e-06, 'num_tokens': 1604996.0, 'completions/mean_length': 46.125, 'completions/min_length': 33.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6611801981925964, 'rewards/meter/std': 0.3047296702861786, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.2285943329334259, 'rewards/repeat_penalty/mean': 0.6565814018249512, 'rewards/repeat_penalty/std': 0.2637769281864166, 'rewards/repeat_floor/mean': 0.9578031301498413, 'rewards/repeat_floor/std': 0.034190621227025986, 'rewards/near_duplicate_penalty/mean': 0.883475661277771, 'rewards/near_duplicate_penalty/std': 0.09240245819091797, 'rewards/opening_diversity/mean': 0.8984375, 'rewards/opening_diversity/std': 0.12472067028284073, 'rewards/distinct_2/mean': 0.7930607795715332, 'rewards/distinct_2/std': 0.17247456312179565, 'rewards/total_composite/mean': 0.24497410655021667, 'rewards/total_composite/std': 0.22584471106529236, 'reward': 0.24497410655021667, 'reward_std': 0.22584472596645355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13213300704956055, 'sampling/sampling_logp_difference/max': 2.4206185340881348, 'sampling/importance_sampling_ratio/min': 0.08886662870645523, 'sampling/importance_sampling_ratio/mean': 1.0153093338012695, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9030940532684326, 'clip_ratio/low_mean': 0.04073049547150731, 'clip_ratio/low_min': 0.04073049547150731, 'clip_ratio/high_mean': 0.06071013957262039, 'clip_ratio/high_max': 0.06071013957262039, 'clip_ratio/region_mean': 0.1014406350441277, 'reward_total_mean': 0.24497410655021667, 'reward_meter_mean': 0.6611801981925964, 'reward_meter_std': 0.3047296702861786, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6565814018249512, 'reward_repeat_penalty_std': 0.2637769281864166, 'reward_repeat_floor_mean': 0.9578031301498413, 'reward_repeat_floor_std': 0.034190621227025986, 'reward_near_duplicate_penalty_mean': 0.883475661277771, 'reward_near_duplicate_penalty_std': 0.09240245819091797, 'reward_opening_diversity_mean': 0.8984375, 'reward_opening_diversity_std': 0.12472067028284073, 'reward_distinct_2_mean': 0.7930607795715332, 'reward_distinct_2_std': 0.17247456312179565, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.2285943329334259, 'reward_total_composite_mean': 0.24497410655021667, 'reward_total_composite_std': 0.22584471106529236, 'epoch': 0.05}

 27%|██▋       | 901/3300 [3:23:02<38:55:43, 58.42s/it]INFO 04-12 19:21:28 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 902/3300 [3:23:14<29:42:57, 44.61s/it]2026-04-12 19:21:40,700 | INFO | train_grpo_train | metrics_logged mode=train step=902

                                                       
{'loss': -0.0042, 'grad_norm': 10.568100929260254, 'learning_rate': 7.26969696969697e-06, 'num_tokens': 1606848.0, 'completions/mean_length': 62.5, 'completions/min_length': 45.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.8505071997642517, 'rewards/meter/std': 0.28576141595840454, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.6060097813606262, 'rewards/repeat_penalty/std': 0.2631150186061859, 'rewards/repeat_floor/mean': 0.9498291015625, 'rewards/repeat_floor/std': 0.043091047555208206, 'rewards/near_duplicate_penalty/mean': 0.850010871887207, 'rewards/near_duplicate_penalty/std': 0.13662436604499817, 'rewards/opening_diversity/mean': 0.9196428656578064, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.7460992336273193, 'rewards/distinct_2/std': 0.20074522495269775, 'rewards/total_composite/mean': 0.25891658663749695, 'rewards/total_composite/std': 0.11678348481655121, 'reward': 0.25891658663749695, 'reward_std': 0.1167834922671318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.134932741522789, 'sampling/sampling_logp_difference/max': 1.8648953437805176, 'sampling/importance_sampling_ratio/min': 0.15491242706775665, 'sampling/importance_sampling_ratio/mean': 1.0250189304351807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9565348699688911, 'clip_ratio/low_mean': 0.03921568766236305, 'clip_ratio/low_min': 0.03921568766236305, 'clip_ratio/high_mean': 0.09171149600297213, 'clip_ratio/high_max': 0.09171149600297213, 'clip_ratio/region_mean': 0.13092718366533518, 'reward_total_mean': 0.25891658663749695, 'reward_meter_mean': 0.8505071997642517, 'reward_meter_std': 0.28576141595840454, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6060097813606262, 'reward_repeat_penalty_std': 0.2631150186061859, 'reward_repeat_floor_mean': 0.9498291015625, 'reward_repeat_floor_std': 0.043091047555208206, 'reward_near_duplicate_penalty_mean': 0.850010871887207, 'reward_near_duplicate_penalty_std': 0.13662436604499817, 'reward_opening_diversity_mean': 0.9196428656578064, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.7460992336273193, 'reward_distinct_2_std': 0.20074522495269775, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.25891658663749695, 'reward_total_composite_std': 0.11678348481655121, 'epoch': 0.05}

 27%|██▋       | 902/3300 [3:23:14<29:42:57, 44.61s/it]INFO 04-12 19:21:40 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 903/3300 [3:23:31<24:16:29, 36.46s/it]2026-04-12 19:21:58,146 | INFO | train_grpo_train | metrics_logged mode=train step=903

                                                       
{'loss': -0.0929, 'grad_norm': 2.662440776824951, 'learning_rate': 7.266666666666668e-06, 'num_tokens': 1608512.0, 'completions/mean_length': 230.0, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 60.79999923706055, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.7251635789871216, 'rewards/meter/std': 0.3493713140487671, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.1602174937725067, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.934374988079071, 'rewards/count_floor/std': 0.04806524142622948, 'rewards/lexical_plausibility/mean': 0.8431609869003296, 'rewards/lexical_plausibility/std': 0.22001656889915466, 'rewards/judge_quality/mean': 0.16250000894069672, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.884110152721405, 'rewards/repeat_penalty/std': 0.0890854001045227, 'rewards/repeat_floor/mean': 0.9862861633300781, 'rewards/repeat_floor/std': 0.01029918435961008, 'rewards/near_duplicate_penalty/mean': 0.9471815824508667, 'rewards/near_duplicate_penalty/std': 0.037550702691078186, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9439880847930908, 'rewards/distinct_2/std': 0.06266674399375916, 'rewards/total_composite/mean': 0.10940331220626831, 'rewards/total_composite/std': 0.09299742430448532, 'reward': 0.10940331220626831, 'reward_std': 0.09299741685390472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19197878241539001, 'sampling/sampling_logp_difference/max': 1.816019058227539, 'sampling/importance_sampling_ratio/min': 0.1626720428466797, 'sampling/importance_sampling_ratio/mean': 1.0411927700042725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.026407465338707, 'clip_ratio/low_mean': 0.02358490601181984, 'clip_ratio/low_min': 0.02358490601181984, 'clip_ratio/high_mean': 0.09216949716210365, 'clip_ratio/high_max': 0.09216949716210365, 'clip_ratio/region_mean': 0.11575440317392349, 'reward_total_mean': 0.10940331220626831, 'reward_meter_mean': 0.7251635789871216, 'reward_meter_std': 0.3493713140487671, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.1602174937725067, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.934374988079071, 'reward_count_floor_std': 0.04806524142622948, 'reward_lexical_plausibility_mean': 0.8431609869003296, 'reward_lexical_plausibility_std': 0.22001656889915466, 'reward_repeat_penalty_mean': 0.884110152721405, 'reward_repeat_penalty_std': 0.0890854001045227, 'reward_repeat_floor_mean': 0.9862861633300781, 'reward_repeat_floor_std': 0.01029918435961008, 'reward_near_duplicate_penalty_mean': 0.9471815824508667, 'reward_near_duplicate_penalty_std': 0.037550702691078186, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9439880847930908, 'reward_distinct_2_std': 0.06266674399375916, 'reward_judge_quality_mean': 0.16250000894069672, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.10940331220626831, 'reward_total_composite_std': 0.09299742430448532, 'epoch': 0.05}

 27%|██▋       | 903/3300 [3:23:32<24:16:29, 36.46s/it]INFO 04-12 19:21:58 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 904/3300 [3:23:45<19:45:07, 29.68s/it]2026-04-12 19:22:11,989 | INFO | train_grpo_train | metrics_logged mode=train step=904

                                                       
{'loss': -0.0469, 'grad_norm': 5.912710189819336, 'learning_rate': 7.263636363636364e-06, 'num_tokens': 1609908.0, 'completions/mean_length': 93.5, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.71428680419922, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5049070119857788, 'rewards/meter/std': 0.4599471390247345, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.957805871963501, 'rewards/lexical_plausibility/std': 0.11934302002191544, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.30514341592788696, 'rewards/repeat_penalty/mean': 0.9901989698410034, 'rewards/repeat_penalty/std': 0.015121985226869583, 'rewards/repeat_floor/mean': 0.998529851436615, 'rewards/repeat_floor/std': 0.0022683022543787956, 'rewards/near_duplicate_penalty/mean': 0.9901989698410034, 'rewards/near_duplicate_penalty/std': 0.015121985226869583, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.253612756729126, 'rewards/total_composite/std': 0.25815820693969727, 'reward': 0.253612756729126, 'reward_std': 0.25815820693969727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14818434417247772, 'sampling/sampling_logp_difference/max': 1.5815479755401611, 'sampling/importance_sampling_ratio/min': 0.21555957198143005, 'sampling/importance_sampling_ratio/mean': 1.011784315109253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7499367445707321, 'clip_ratio/low_mean': 0.016145833767950535, 'clip_ratio/low_min': 0.016145833767950535, 'clip_ratio/high_mean': 0.05803346075117588, 'clip_ratio/high_max': 0.05803346075117588, 'clip_ratio/region_mean': 0.07417929451912642, 'reward_total_mean': 0.253612756729126, 'reward_meter_mean': 0.5049070119857788, 'reward_meter_std': 0.4599471390247345, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.957805871963501, 'reward_lexical_plausibility_std': 0.11934302002191544, 'reward_repeat_penalty_mean': 0.9901989698410034, 'reward_repeat_penalty_std': 0.015121985226869583, 'reward_repeat_floor_mean': 0.998529851436615, 'reward_repeat_floor_std': 0.0022683022543787956, 'reward_near_duplicate_penalty_mean': 0.9901989698410034, 'reward_near_duplicate_penalty_std': 0.015121985226869583, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.30514341592788696, 'reward_total_composite_mean': 0.253612756729126, 'reward_total_composite_std': 0.25815820693969727, 'epoch': 0.05}

 27%|██▋       | 904/3300 [3:23:45<19:45:07, 29.68s/it]INFO 04-12 19:22:12 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 905/3300 [3:23:54<15:29:32, 23.29s/it]2026-04-12 19:22:20,373 | INFO | train_grpo_train | metrics_logged mode=train step=905

                                                       
{'loss': 0.0399, 'grad_norm': 13.982442855834961, 'learning_rate': 7.260606060606061e-06, 'num_tokens': 1611771.0, 'completions/mean_length': 40.875, 'completions/min_length': 30.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.1939105987548828, 'rewards/meter/std': 0.31769174337387085, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.05175493285059929, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.5864450335502625, 'rewards/repeat_penalty/std': 0.19441701471805573, 'rewards/repeat_floor/mean': 0.9565856456756592, 'rewards/repeat_floor/std': 0.02301929146051407, 'rewards/near_duplicate_penalty/mean': 0.8873894810676575, 'rewards/near_duplicate_penalty/std': 0.06870521605014801, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8094695210456848, 'rewards/distinct_2/std': 0.1260429471731186, 'rewards/total_composite/mean': 0.06114654242992401, 'rewards/total_composite/std': 0.09451024234294891, 'reward': 0.06114654242992401, 'reward_std': 0.09451024234294891, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1358642429113388, 'sampling/sampling_logp_difference/max': 1.6681842803955078, 'sampling/importance_sampling_ratio/min': 0.1885891705751419, 'sampling/importance_sampling_ratio/mean': 1.0365381240844727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7639429569244385, 'clip_ratio/low_mean': 0.07807393930852413, 'clip_ratio/low_min': 0.07807393930852413, 'clip_ratio/high_mean': 0.030741626396775246, 'clip_ratio/high_max': 0.030741626396775246, 'clip_ratio/region_mean': 0.10881556570529938, 'reward_total_mean': 0.06114654242992401, 'reward_meter_mean': 0.1939105987548828, 'reward_meter_std': 0.31769174337387085, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.05175493285059929, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.5864450335502625, 'reward_repeat_penalty_std': 0.19441701471805573, 'reward_repeat_floor_mean': 0.9565856456756592, 'reward_repeat_floor_std': 0.02301929146051407, 'reward_near_duplicate_penalty_mean': 0.8873894810676575, 'reward_near_duplicate_penalty_std': 0.06870521605014801, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8094695210456848, 'reward_distinct_2_std': 0.1260429471731186, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.06114654242992401, 'reward_total_composite_std': 0.09451024234294891, 'epoch': 0.05}

 27%|██▋       | 905/3300 [3:23:54<15:29:32, 23.29s/it]INFO 04-12 19:22:20 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 906/3300 [3:24:08<13:44:45, 20.67s/it]2026-04-12 19:22:34,929 | INFO | train_grpo_train | metrics_logged mode=train step=906

                                                       
{'loss': -0.1174, 'grad_norm': 3.9887447357177734, 'learning_rate': 7.257575757575758e-06, 'num_tokens': 1613798.0, 'completions/mean_length': 139.375, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.14286041259766, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.6583786606788635, 'rewards/meter/std': 0.35448578000068665, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.9676902294158936, 'rewards/lexical_plausibility/std': 0.09138572961091995, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.7100282907485962, 'rewards/repeat_penalty/std': 0.27022063732147217, 'rewards/repeat_floor/mean': 0.9633342027664185, 'rewards/repeat_floor/std': 0.042829982936382294, 'rewards/near_duplicate_penalty/mean': 0.9010794162750244, 'rewards/near_duplicate_penalty/std': 0.13472360372543335, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.7783033847808838, 'rewards/distinct_2/std': 0.20225897431373596, 'rewards/total_composite/mean': 0.1825007200241089, 'rewards/total_composite/std': 0.11758004128932953, 'reward': 0.1825007200241089, 'reward_std': 0.11758004128932953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13918501138687134, 'sampling/sampling_logp_difference/max': 2.9872148036956787, 'sampling/importance_sampling_ratio/min': 0.05042769014835358, 'sampling/importance_sampling_ratio/mean': 1.006597638130188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8042197227478027, 'clip_ratio/low_mean': 0.07502768700942397, 'clip_ratio/low_min': 0.07502768700942397, 'clip_ratio/high_mean': 0.05760060250759125, 'clip_ratio/high_max': 0.05760060250759125, 'clip_ratio/region_mean': 0.13262828951701522, 'reward_total_mean': 0.1825007200241089, 'reward_meter_mean': 0.6583786606788635, 'reward_meter_std': 0.35448578000068665, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.9676902294158936, 'reward_lexical_plausibility_std': 0.09138572961091995, 'reward_repeat_penalty_mean': 0.7100282907485962, 'reward_repeat_penalty_std': 0.27022063732147217, 'reward_repeat_floor_mean': 0.9633342027664185, 'reward_repeat_floor_std': 0.042829982936382294, 'reward_near_duplicate_penalty_mean': 0.9010794162750244, 'reward_near_duplicate_penalty_std': 0.13472360372543335, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.7783033847808838, 'reward_distinct_2_std': 0.20225897431373596, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.1825007200241089, 'reward_total_composite_std': 0.11758004128932953, 'epoch': 0.05}

 27%|██▋       | 906/3300 [3:24:08<13:44:45, 20.67s/it]INFO 04-12 19:22:35 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 907/3300 [3:24:17<11:18:46, 17.02s/it]2026-04-12 19:22:43,426 | INFO | train_grpo_train | metrics_logged mode=train step=907

                                                       
{'loss': 0.2505, 'grad_norm': 19.770160675048828, 'learning_rate': 7.254545454545455e-06, 'num_tokens': 1615236.0, 'completions/mean_length': 25.75, 'completions/min_length': 16.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.06526508927345276, 'rewards/meter/std': 0.04139859601855278, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.15526476502418518, 'rewards/lexical_plausibility/mean': 0.9713541269302368, 'rewards/lexical_plausibility/std': 0.05990393087267876, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.6929816007614136, 'rewards/repeat_penalty/std': 0.21932563185691833, 'rewards/repeat_floor/mean': 0.9714167714118958, 'rewards/repeat_floor/std': 0.025788577273488045, 'rewards/near_duplicate_penalty/mean': 0.9219865798950195, 'rewards/near_duplicate_penalty/std': 0.10779546201229095, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9374309778213501, 'rewards/distinct_2/std': 0.14701072871685028, 'rewards/total_composite/mean': 0.017753588035702705, 'rewards/total_composite/std': 0.016435202211141586, 'reward': 0.017753588035702705, 'reward_std': 0.016435202211141586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16366185247898102, 'sampling/sampling_logp_difference/max': 1.3756822347640991, 'sampling/importance_sampling_ratio/min': 0.2526671588420868, 'sampling/importance_sampling_ratio/mean': 1.012420654296875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.24896989017725, 'clip_ratio/low_mean': 0.0977272754535079, 'clip_ratio/low_min': 0.0977272754535079, 'clip_ratio/high_mean': 0.03125, 'clip_ratio/high_max': 0.03125, 'clip_ratio/region_mean': 0.1289772754535079, 'reward_total_mean': 0.017753588035702705, 'reward_meter_mean': 0.06526508927345276, 'reward_meter_std': 0.04139859601855278, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.15526476502418518, 'reward_lexical_plausibility_mean': 0.9713541269302368, 'reward_lexical_plausibility_std': 0.05990393087267876, 'reward_repeat_penalty_mean': 0.6929816007614136, 'reward_repeat_penalty_std': 0.21932563185691833, 'reward_repeat_floor_mean': 0.9714167714118958, 'reward_repeat_floor_std': 0.025788577273488045, 'reward_near_duplicate_penalty_mean': 0.9219865798950195, 'reward_near_duplicate_penalty_std': 0.10779546201229095, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9374309778213501, 'reward_distinct_2_std': 0.14701072871685028, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.017753588035702705, 'reward_total_composite_std': 0.016435202211141586, 'epoch': 0.05}

 27%|██▋       | 907/3300 [3:24:17<11:18:46, 17.02s/it]INFO 04-12 19:22:43 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 908/3300 [3:24:31<10:40:01, 16.05s/it]2026-04-12 19:22:57,231 | INFO | train_grpo_train | metrics_logged mode=train step=908

                                                       
{'loss': -0.0337, 'grad_norm': 15.10283374786377, 'learning_rate': 7.251515151515151e-06, 'num_tokens': 1616807.0, 'completions/mean_length': 98.375, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.28571701049805, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6321123838424683, 'rewards/meter/std': 0.2811214029788971, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9574733972549438, 'rewards/lexical_plausibility/std': 0.1202833354473114, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690458357334137, 'rewards/repeat_penalty/mean': 0.8837242722511292, 'rewards/repeat_penalty/std': 0.24979446828365326, 'rewards/repeat_floor/mean': 0.9870565533638, 'rewards/repeat_floor/std': 0.024856988340616226, 'rewards/near_duplicate_penalty/mean': 0.9571062326431274, 'rewards/near_duplicate_penalty/std': 0.04875440523028374, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.05725561082363129, 'rewards/total_composite/mean': 0.1716843545436859, 'rewards/total_composite/std': 0.11949339509010315, 'reward': 0.1716843545436859, 'reward_std': 0.11949339509010315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18133169412612915, 'sampling/sampling_logp_difference/max': 2.038975715637207, 'sampling/importance_sampling_ratio/min': 0.13016197085380554, 'sampling/importance_sampling_ratio/mean': 1.0067718029022217, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9734255075454712, 'clip_ratio/low_mean': 0.06737266108393669, 'clip_ratio/low_min': 0.06737266108393669, 'clip_ratio/high_mean': 0.10661562532186508, 'clip_ratio/high_max': 0.10661562532186508, 'clip_ratio/region_mean': 0.17398828640580177, 'reward_total_mean': 0.1716843545436859, 'reward_meter_mean': 0.6321123838424683, 'reward_meter_std': 0.2811214029788971, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9574733972549438, 'reward_lexical_plausibility_std': 0.1202833354473114, 'reward_repeat_penalty_mean': 0.8837242722511292, 'reward_repeat_penalty_std': 0.24979446828365326, 'reward_repeat_floor_mean': 0.9870565533638, 'reward_repeat_floor_std': 0.024856988340616226, 'reward_near_duplicate_penalty_mean': 0.9571062326431274, 'reward_near_duplicate_penalty_std': 0.04875440523028374, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.05725561082363129, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690458357334137, 'reward_total_composite_mean': 0.1716843545436859, 'reward_total_composite_std': 0.11949339509010315, 'epoch': 0.05}

 28%|██▊       | 908/3300 [3:24:31<10:40:01, 16.05s/it]INFO 04-12 19:22:57 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 909/3300 [3:24:46<10:27:59, 15.76s/it]2026-04-12 19:23:12,302 | INFO | train_grpo_train | metrics_logged mode=train step=909

                                                       
{'loss': -0.0551, 'grad_norm': 3.51112699508667, 'learning_rate': 7.2484848484848495e-06, 'num_tokens': 1618883.0, 'completions/mean_length': 244.5, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 84.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.5140498876571655, 'rewards/meter/std': 0.36049869656562805, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.18516401946544647, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.940000057220459, 'rewards/count_floor/std': 0.05554921180009842, 'rewards/lexical_plausibility/mean': 0.8891571760177612, 'rewards/lexical_plausibility/std': 0.16540585458278656, 'rewards/judge_quality/mean': 0.13124999403953552, 'rewards/judge_quality/std': 0.1066954955458641, 'rewards/repeat_penalty/mean': 0.6409416198730469, 'rewards/repeat_penalty/std': 0.39120203256607056, 'rewards/repeat_floor/mean': 0.9426136016845703, 'rewards/repeat_floor/std': 0.07456691563129425, 'rewards/near_duplicate_penalty/mean': 0.8410364389419556, 'rewards/near_duplicate_penalty/std': 0.1904403567314148, 'rewards/opening_diversity/mean': 0.8742897510528564, 'rewards/opening_diversity/std': 0.22970367968082428, 'rewards/distinct_2/mean': 0.7111193537712097, 'rewards/distinct_2/std': 0.36678844690322876, 'rewards/total_composite/mean': 0.06978833675384521, 'rewards/total_composite/std': 0.1033511683344841, 'reward': 0.06978833675384521, 'reward_std': 0.1033511683344841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14205285906791687, 'sampling/sampling_logp_difference/max': 2.3772497177124023, 'sampling/importance_sampling_ratio/min': 0.09280547499656677, 'sampling/importance_sampling_ratio/mean': 1.0200109481811523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.558080542832613, 'clip_ratio/low_mean': 0.036580849438905716, 'clip_ratio/low_min': 0.036580849438905716, 'clip_ratio/high_mean': 0.024484342895448208, 'clip_ratio/high_max': 0.024484342895448208, 'clip_ratio/region_mean': 0.061065192334353924, 'reward_total_mean': 0.06978833675384521, 'reward_meter_mean': 0.5140498876571655, 'reward_meter_std': 0.36049869656562805, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.18516401946544647, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.940000057220459, 'reward_count_floor_std': 0.05554921180009842, 'reward_lexical_plausibility_mean': 0.8891571760177612, 'reward_lexical_plausibility_std': 0.16540585458278656, 'reward_repeat_penalty_mean': 0.6409416198730469, 'reward_repeat_penalty_std': 0.39120203256607056, 'reward_repeat_floor_mean': 0.9426136016845703, 'reward_repeat_floor_std': 0.07456691563129425, 'reward_near_duplicate_penalty_mean': 0.8410364389419556, 'reward_near_duplicate_penalty_std': 0.1904403567314148, 'reward_opening_diversity_mean': 0.8742897510528564, 'reward_opening_diversity_std': 0.22970367968082428, 'reward_distinct_2_mean': 0.7111193537712097, 'reward_distinct_2_std': 0.36678844690322876, 'reward_judge_quality_mean': 0.13124999403953552, 'reward_judge_quality_std': 0.1066954955458641, 'reward_total_composite_mean': 0.06978833675384521, 'reward_total_composite_std': 0.1033511683344841, 'epoch': 0.05}

 28%|██▊       | 909/3300 [3:24:46<10:27:59, 15.76s/it]INFO 04-12 19:23:12 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 910/3300 [3:25:02<10:36:17, 15.97s/it]2026-04-12 19:23:28,778 | INFO | train_grpo_train | metrics_logged mode=train step=910

                                                       
{'loss': -0.0488, 'grad_norm': 3.6321957111358643, 'learning_rate': 7.245454545454546e-06, 'num_tokens': 1620482.0, 'completions/mean_length': 159.875, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5518388152122498, 'rewards/meter/std': 0.431120365858078, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9289302229881287, 'rewards/lexical_plausibility/std': 0.1316879838705063, 'rewards/judge_quality/mean': 0.48875001072883606, 'rewards/judge_quality/std': 0.38926443457603455, 'rewards/repeat_penalty/mean': 0.9867567420005798, 'rewards/repeat_penalty/std': 0.03295497968792915, 'rewards/repeat_floor/mean': 0.9985635280609131, 'rewards/repeat_floor/std': 0.0034054333809763193, 'rewards/near_duplicate_penalty/mean': 0.9961925745010376, 'rewards/near_duplicate_penalty/std': 0.007286599837243557, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.24825844168663025, 'rewards/total_composite/std': 0.31368914246559143, 'reward': 0.24825844168663025, 'reward_std': 0.31368911266326904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13701516389846802, 'sampling/sampling_logp_difference/max': 1.120093822479248, 'sampling/importance_sampling_ratio/min': 0.37801021337509155, 'sampling/importance_sampling_ratio/mean': 1.0397045612335205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6110581383109093, 'clip_ratio/low_mean': 0.04854899179190397, 'clip_ratio/low_min': 0.04854899179190397, 'clip_ratio/high_mean': 0.04152823984622955, 'clip_ratio/high_max': 0.04152823984622955, 'clip_ratio/region_mean': 0.09007723163813353, 'reward_total_mean': 0.24825844168663025, 'reward_meter_mean': 0.5518388152122498, 'reward_meter_std': 0.431120365858078, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9289302229881287, 'reward_lexical_plausibility_std': 0.1316879838705063, 'reward_repeat_penalty_mean': 0.9867567420005798, 'reward_repeat_penalty_std': 0.03295497968792915, 'reward_repeat_floor_mean': 0.9985635280609131, 'reward_repeat_floor_std': 0.0034054333809763193, 'reward_near_duplicate_penalty_mean': 0.9961925745010376, 'reward_near_duplicate_penalty_std': 0.007286599837243557, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.48875001072883606, 'reward_judge_quality_std': 0.38926443457603455, 'reward_total_composite_mean': 0.24825844168663025, 'reward_total_composite_std': 0.31368914246559143, 'epoch': 0.05}

 28%|██▊       | 910/3300 [3:25:02<10:36:17, 15.97s/it]INFO 04-12 19:23:29 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 911/3300 [3:25:12<9:27:22, 14.25s/it] 2026-04-12 19:23:39,061 | INFO | train_grpo_train | metrics_logged mode=train step=911

                                                      
{'loss': 0.1152, 'grad_norm': 17.10910415649414, 'learning_rate': 7.242424242424243e-06, 'num_tokens': 1621955.0, 'completions/mean_length': 33.125, 'completions/min_length': 16.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8175361752510071, 'rewards/meter/std': 0.2913716435432434, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9928355813026428, 'rewards/lexical_plausibility/std': 0.020263994112610817, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.9153796434402466, 'rewards/repeat_penalty/std': 0.11607661843299866, 'rewards/repeat_floor/mean': 0.9938516616821289, 'rewards/repeat_floor/std': 0.007939220406115055, 'rewards/near_duplicate_penalty/mean': 0.9940446019172668, 'rewards/near_duplicate_penalty/std': 0.011787589639425278, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.40861278772354126, 'rewards/total_composite/std': 0.24489811062812805, 'reward': 0.40861278772354126, 'reward_std': 0.24489811062812805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19116827845573425, 'sampling/sampling_logp_difference/max': 1.6085829734802246, 'sampling/importance_sampling_ratio/min': 0.2001710683107376, 'sampling/importance_sampling_ratio/mean': 1.0210844278335571, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2049486115574837, 'clip_ratio/low_mean': 0.08096088282763958, 'clip_ratio/low_min': 0.08096088282763958, 'clip_ratio/high_mean': 0.14603892713785172, 'clip_ratio/high_max': 0.14603892713785172, 'clip_ratio/region_mean': 0.2269998099654913, 'reward_total_mean': 0.40861278772354126, 'reward_meter_mean': 0.8175361752510071, 'reward_meter_std': 0.2913716435432434, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9928355813026428, 'reward_lexical_plausibility_std': 0.020263994112610817, 'reward_repeat_penalty_mean': 0.9153796434402466, 'reward_repeat_penalty_std': 0.11607661843299866, 'reward_repeat_floor_mean': 0.9938516616821289, 'reward_repeat_floor_std': 0.007939220406115055, 'reward_near_duplicate_penalty_mean': 0.9940446019172668, 'reward_near_duplicate_penalty_std': 0.011787589639425278, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.40861278772354126, 'reward_total_composite_std': 0.24489811062812805, 'epoch': 0.05}

 28%|██▊       | 911/3300 [3:25:12<9:27:22, 14.25s/it]INFO 04-12 19:23:39 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 912/3300 [3:25:22<8:32:07, 12.87s/it]2026-04-12 19:23:48,649 | INFO | train_grpo_train | metrics_logged mode=train step=912

                                                      
{'loss': -0.0437, 'grad_norm': 15.735276222229004, 'learning_rate': 7.2393939393939404e-06, 'num_tokens': 1623304.0, 'completions/mean_length': 28.625, 'completions/min_length': 24.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.5129578113555908, 'rewards/meter/std': 0.37983524799346924, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.9517156481742859, 'rewards/repeat_penalty/std': 0.13215535879135132, 'rewards/repeat_floor/mean': 0.9952853918075562, 'rewards/repeat_floor/std': 0.012677213177084923, 'rewards/near_duplicate_penalty/mean': 0.987479567527771, 'rewards/near_duplicate_penalty/std': 0.031165635213255882, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.22096094489097595, 'rewards/total_composite/std': 0.19184307754039764, 'reward': 0.22096094489097595, 'reward_std': 0.19184306263923645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16369909048080444, 'sampling/sampling_logp_difference/max': 1.1884865760803223, 'sampling/importance_sampling_ratio/min': 0.30468204617500305, 'sampling/importance_sampling_ratio/mean': 1.0196163654327393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9141949564218521, 'clip_ratio/low_mean': 0.07289101183414459, 'clip_ratio/low_min': 0.07289101183414459, 'clip_ratio/high_mean': 0.0879049114882946, 'clip_ratio/high_max': 0.0879049114882946, 'clip_ratio/region_mean': 0.1607959233224392, 'reward_total_mean': 0.22096094489097595, 'reward_meter_mean': 0.5129578113555908, 'reward_meter_std': 0.37983524799346924, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9517156481742859, 'reward_repeat_penalty_std': 0.13215535879135132, 'reward_repeat_floor_mean': 0.9952853918075562, 'reward_repeat_floor_std': 0.012677213177084923, 'reward_near_duplicate_penalty_mean': 0.987479567527771, 'reward_near_duplicate_penalty_std': 0.031165635213255882, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.22096094489097595, 'reward_total_composite_std': 0.19184307754039764, 'epoch': 0.05}

 28%|██▊       | 912/3300 [3:25:22<8:32:07, 12.87s/it]INFO 04-12 19:23:48 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 913/3300 [3:25:36<8:42:09, 13.13s/it]2026-04-12 19:24:02,412 | INFO | train_grpo_train | metrics_logged mode=train step=913

                                                      
{'loss': -0.0474, 'grad_norm': 5.727684020996094, 'learning_rate': 7.236363636363637e-06, 'num_tokens': 1625039.0, 'completions/mean_length': 106.875, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.41863495111465454, 'rewards/meter/std': 0.4023495316505432, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9539152383804321, 'rewards/lexical_plausibility/std': 0.13034749031066895, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.6094663143157959, 'rewards/repeat_penalty/std': 0.22540996968746185, 'rewards/repeat_floor/mean': 0.9550163149833679, 'rewards/repeat_floor/std': 0.027856511995196342, 'rewards/near_duplicate_penalty/mean': 0.8774460554122925, 'rewards/near_duplicate_penalty/std': 0.08431324362754822, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.7669378519058228, 'rewards/distinct_2/std': 0.18185408413410187, 'rewards/total_composite/mean': 0.055691204965114594, 'rewards/total_composite/std': 0.05289734899997711, 'reward': 0.055691204965114594, 'reward_std': 0.05289734899997711, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1409750133752823, 'sampling/sampling_logp_difference/max': 1.8814221620559692, 'sampling/importance_sampling_ratio/min': 0.15237325429916382, 'sampling/importance_sampling_ratio/mean': 1.0115469694137573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7962162271142006, 'clip_ratio/low_mean': 0.044459800235927105, 'clip_ratio/low_min': 0.044459800235927105, 'clip_ratio/high_mean': 0.046875000931322575, 'clip_ratio/high_max': 0.046875000931322575, 'clip_ratio/region_mean': 0.09133480116724968, 'reward_total_mean': 0.055691204965114594, 'reward_meter_mean': 0.41863495111465454, 'reward_meter_std': 0.4023495316505432, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9539152383804321, 'reward_lexical_plausibility_std': 0.13034749031066895, 'reward_repeat_penalty_mean': 0.6094663143157959, 'reward_repeat_penalty_std': 0.22540996968746185, 'reward_repeat_floor_mean': 0.9550163149833679, 'reward_repeat_floor_std': 0.027856511995196342, 'reward_near_duplicate_penalty_mean': 0.8774460554122925, 'reward_near_duplicate_penalty_std': 0.08431324362754822, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.7669378519058228, 'reward_distinct_2_std': 0.18185408413410187, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.055691204965114594, 'reward_total_composite_std': 0.05289734899997711, 'epoch': 0.05}

 28%|██▊       | 913/3300 [3:25:36<8:42:09, 13.13s/it]INFO 04-12 19:24:02 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 914/3300 [3:25:48<8:30:18, 12.83s/it]2026-04-12 19:24:14,638 | INFO | train_grpo_train | metrics_logged mode=train step=914

                                                      
{'loss': -0.0209, 'grad_norm': 8.038836479187012, 'learning_rate': 7.233333333333334e-06, 'num_tokens': 1627553.0, 'completions/mean_length': 119.25, 'completions/min_length': 23.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.8334822654724121, 'rewards/meter/std': 0.23967531323432922, 'rewards/count_adherence/mean': 0.671875, 'rewards/count_adherence/std': 0.22097088396549225, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.901562511920929, 'rewards/count_floor/std': 0.06629125773906708, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.32043540477752686, 'rewards/repeat_penalty/std': 0.19843009114265442, 'rewards/repeat_floor/mean': 0.9110767841339111, 'rewards/repeat_floor/std': 0.044244445860385895, 'rewards/near_duplicate_penalty/mean': 0.7726829051971436, 'rewards/near_duplicate_penalty/std': 0.12587784230709076, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.18221724033355713, 'rewards/distinct_2/mean': 0.4637429416179657, 'rewards/distinct_2/std': 0.2513948678970337, 'rewards/total_composite/mean': 0.1342175006866455, 'rewards/total_composite/std': 0.07074952870607376, 'reward': 0.1342175006866455, 'reward_std': 0.07074952125549316, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13308842480182648, 'sampling/sampling_logp_difference/max': 1.8935027122497559, 'sampling/importance_sampling_ratio/min': 0.15054357051849365, 'sampling/importance_sampling_ratio/mean': 1.0097712278366089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6534483358263969, 'clip_ratio/low_mean': 0.0673923552967608, 'clip_ratio/low_min': 0.0673923552967608, 'clip_ratio/high_mean': 0.031618923880159855, 'clip_ratio/high_max': 0.031618923880159855, 'clip_ratio/region_mean': 0.09901127917692065, 'reward_total_mean': 0.1342175006866455, 'reward_meter_mean': 0.8334822654724121, 'reward_meter_std': 0.23967531323432922, 'reward_count_adherence_mean': 0.671875, 'reward_count_adherence_std': 0.22097088396549225, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.901562511920929, 'reward_count_floor_std': 0.06629125773906708, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.32043540477752686, 'reward_repeat_penalty_std': 0.19843009114265442, 'reward_repeat_floor_mean': 0.9110767841339111, 'reward_repeat_floor_std': 0.044244445860385895, 'reward_near_duplicate_penalty_mean': 0.7726829051971436, 'reward_near_duplicate_penalty_std': 0.12587784230709076, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.18221724033355713, 'reward_distinct_2_mean': 0.4637429416179657, 'reward_distinct_2_std': 0.2513948678970337, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1342175006866455, 'reward_total_composite_std': 0.07074952870607376, 'epoch': 0.05}

 28%|██▊       | 914/3300 [3:25:48<8:30:18, 12.83s/it]INFO 04-12 19:24:14 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 915/3300 [3:25:56<7:40:05, 11.57s/it]2026-04-12 19:24:23,181 | INFO | train_grpo_train | metrics_logged mode=train step=915

                                                      
{'loss': 0.048, 'grad_norm': 10.597201347351074, 'learning_rate': 7.2303030303030305e-06, 'num_tokens': 1629415.0, 'completions/mean_length': 55.75, 'completions/min_length': 46.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.11782373487949371, 'rewards/meter/std': 0.15586894750595093, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.3094436228275299, 'rewards/repeat_penalty/mean': 0.8395186066627502, 'rewards/repeat_penalty/std': 0.2252216339111328, 'rewards/repeat_floor/mean': 0.9809199571609497, 'rewards/repeat_floor/std': 0.026184694841504097, 'rewards/near_duplicate_penalty/mean': 0.9340975284576416, 'rewards/near_duplicate_penalty/std': 0.07946949452161789, 'rewards/opening_diversity/mean': 0.9765625, 'rewards/opening_diversity/std': 0.032346826046705246, 'rewards/distinct_2/mean': 0.9134615659713745, 'rewards/distinct_2/std': 0.14002475142478943, 'rewards/total_composite/mean': 0.08347926288843155, 'rewards/total_composite/std': 0.13637875020503998, 'reward': 0.08347926288843155, 'reward_std': 0.13637875020503998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10143669694662094, 'sampling/sampling_logp_difference/max': 1.5763344764709473, 'sampling/importance_sampling_ratio/min': 0.20673148334026337, 'sampling/importance_sampling_ratio/mean': 1.0131306648254395, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.576223261654377, 'clip_ratio/low_mean': 0.05960719892755151, 'clip_ratio/low_min': 0.05960719892755151, 'clip_ratio/high_mean': 0.019539337139576674, 'clip_ratio/high_max': 0.019539337139576674, 'clip_ratio/region_mean': 0.07914653606712818, 'reward_total_mean': 0.08347926288843155, 'reward_meter_mean': 0.11782373487949371, 'reward_meter_std': 0.15586894750595093, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8395186066627502, 'reward_repeat_penalty_std': 0.2252216339111328, 'reward_repeat_floor_mean': 0.9809199571609497, 'reward_repeat_floor_std': 0.026184694841504097, 'reward_near_duplicate_penalty_mean': 0.9340975284576416, 'reward_near_duplicate_penalty_std': 0.07946949452161789, 'reward_opening_diversity_mean': 0.9765625, 'reward_opening_diversity_std': 0.032346826046705246, 'reward_distinct_2_mean': 0.9134615659713745, 'reward_distinct_2_std': 0.14002475142478943, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.3094436228275299, 'reward_total_composite_mean': 0.08347926288843155, 'reward_total_composite_std': 0.13637875020503998, 'epoch': 0.05}

 28%|██▊       | 915/3300 [3:25:57<7:40:05, 11.57s/it]INFO 04-12 19:24:23 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 916/3300 [3:26:11<8:11:05, 12.36s/it]2026-04-12 19:24:37,356 | INFO | train_grpo_train | metrics_logged mode=train step=916

                                                      
{'loss': -0.0626, 'grad_norm': 7.1197357177734375, 'learning_rate': 7.227272727272729e-06, 'num_tokens': 1631803.0, 'completions/mean_length': 183.5, 'completions/min_length': 114.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 136.57144165039062, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.6531333923339844, 'rewards/meter/std': 0.33623582124710083, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.23754698038101196, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.07126409560441971, 'rewards/lexical_plausibility/mean': 0.9599512815475464, 'rewards/lexical_plausibility/std': 0.07999832928180695, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.32311004400253296, 'rewards/repeat_penalty/std': 0.35683465003967285, 'rewards/repeat_floor/mean': 0.9192783236503601, 'rewards/repeat_floor/std': 0.05014077201485634, 'rewards/near_duplicate_penalty/mean': 0.8367998600006104, 'rewards/near_duplicate_penalty/std': 0.09752444922924042, 'rewards/opening_diversity/mean': 0.9750000238418579, 'rewards/opening_diversity/std': 0.04629101976752281, 'rewards/distinct_2/mean': 0.3917592465877533, 'rewards/distinct_2/std': 0.38457441329956055, 'rewards/total_composite/mean': 0.09356064349412918, 'rewards/total_composite/std': 0.09753771871328354, 'reward': 0.09356064349412918, 'reward_std': 0.09753771871328354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12213119864463806, 'sampling/sampling_logp_difference/max': 2.6143012046813965, 'sampling/importance_sampling_ratio/min': 0.073218934237957, 'sampling/importance_sampling_ratio/mean': 1.0072952508926392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6170133873820305, 'clip_ratio/low_mean': 0.03932244423776865, 'clip_ratio/low_min': 0.03932244423776865, 'clip_ratio/high_mean': 0.061362046748399734, 'clip_ratio/high_max': 0.061362046748399734, 'clip_ratio/region_mean': 0.10068449098616838, 'reward_total_mean': 0.09356064349412918, 'reward_meter_mean': 0.6531333923339844, 'reward_meter_std': 0.33623582124710083, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.23754698038101196, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.07126409560441971, 'reward_lexical_plausibility_mean': 0.9599512815475464, 'reward_lexical_plausibility_std': 0.07999832928180695, 'reward_repeat_penalty_mean': 0.32311004400253296, 'reward_repeat_penalty_std': 0.35683465003967285, 'reward_repeat_floor_mean': 0.9192783236503601, 'reward_repeat_floor_std': 0.05014077201485634, 'reward_near_duplicate_penalty_mean': 0.8367998600006104, 'reward_near_duplicate_penalty_std': 0.09752444922924042, 'reward_opening_diversity_mean': 0.9750000238418579, 'reward_opening_diversity_std': 0.04629101976752281, 'reward_distinct_2_mean': 0.3917592465877533, 'reward_distinct_2_std': 0.38457441329956055, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.09356064349412918, 'reward_total_composite_std': 0.09753771871328354, 'epoch': 0.05}

 28%|██▊       | 916/3300 [3:26:11<8:11:05, 12.36s/it]INFO 04-12 19:24:37 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 917/3300 [3:26:28<9:04:42, 13.71s/it]2026-04-12 19:24:54,235 | INFO | train_grpo_train | metrics_logged mode=train step=917

                                                      
{'loss': 0.0158, 'grad_norm': 4.974209308624268, 'learning_rate': 7.224242424242425e-06, 'num_tokens': 1633506.0, 'completions/mean_length': 102.875, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.42857360839844, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.3541560173034668, 'rewards/meter/std': 0.4033718407154083, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9388551712036133, 'rewards/lexical_plausibility/std': 0.17294375598430634, 'rewards/judge_quality/mean': 0.27125000953674316, 'rewards/judge_quality/std': 0.28667742013931274, 'rewards/repeat_penalty/mean': 0.6628930568695068, 'rewards/repeat_penalty/std': 0.20800966024398804, 'rewards/repeat_floor/mean': 0.9601958394050598, 'rewards/repeat_floor/std': 0.024883992969989777, 'rewards/near_duplicate_penalty/mean': 0.890661895275116, 'rewards/near_duplicate_penalty/std': 0.06234469637274742, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8024621605873108, 'rewards/distinct_2/std': 0.13615424931049347, 'rewards/total_composite/mean': 0.15337839722633362, 'rewards/total_composite/std': 0.27499744296073914, 'reward': 0.15337839722633362, 'reward_std': 0.27499744296073914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18709489703178406, 'sampling/sampling_logp_difference/max': 1.9674787521362305, 'sampling/importance_sampling_ratio/min': 0.13980890810489655, 'sampling/importance_sampling_ratio/mean': 1.0322145223617554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9502277672290802, 'clip_ratio/low_mean': 0.08572155516594648, 'clip_ratio/low_min': 0.08572155516594648, 'clip_ratio/high_mean': 0.05032051540911198, 'clip_ratio/high_max': 0.05032051540911198, 'clip_ratio/region_mean': 0.13604207057505846, 'reward_total_mean': 0.15337839722633362, 'reward_meter_mean': 0.3541560173034668, 'reward_meter_std': 0.4033718407154083, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9388551712036133, 'reward_lexical_plausibility_std': 0.17294375598430634, 'reward_repeat_penalty_mean': 0.6628930568695068, 'reward_repeat_penalty_std': 0.20800966024398804, 'reward_repeat_floor_mean': 0.9601958394050598, 'reward_repeat_floor_std': 0.024883992969989777, 'reward_near_duplicate_penalty_mean': 0.890661895275116, 'reward_near_duplicate_penalty_std': 0.06234469637274742, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8024621605873108, 'reward_distinct_2_std': 0.13615424931049347, 'reward_judge_quality_mean': 0.27125000953674316, 'reward_judge_quality_std': 0.28667742013931274, 'reward_total_composite_mean': 0.15337839722633362, 'reward_total_composite_std': 0.27499744296073914, 'epoch': 0.05}

 28%|██▊       | 917/3300 [3:26:28<9:04:42, 13.71s/it]INFO 04-12 19:24:54 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 918/3300 [3:26:42<9:16:25, 14.02s/it]2026-04-12 19:25:08,954 | INFO | train_grpo_train | metrics_logged mode=train step=918

                                                      
{'loss': -0.0404, 'grad_norm': 3.647557258605957, 'learning_rate': 7.221212121212122e-06, 'num_tokens': 1634856.0, 'completions/mean_length': 83.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.571430206298828, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.5454115867614746, 'rewards/meter/std': 0.4758901000022888, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9623986482620239, 'rewards/lexical_plausibility/std': 0.08399605005979538, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690458357334137, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16828002035617828, 'rewards/total_composite/std': 0.17661982774734497, 'reward': 0.16828002035617828, 'reward_std': 0.17661982774734497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14703625440597534, 'sampling/sampling_logp_difference/max': 1.6368846893310547, 'sampling/importance_sampling_ratio/min': 0.19458529353141785, 'sampling/importance_sampling_ratio/mean': 1.0103979110717773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.860602855682373, 'clip_ratio/low_mean': 0.06764069385826588, 'clip_ratio/low_min': 0.06764069385826588, 'clip_ratio/high_mean': 0.06905594561249018, 'clip_ratio/high_max': 0.06905594561249018, 'clip_ratio/region_mean': 0.13669663947075605, 'reward_total_mean': 0.16828002035617828, 'reward_meter_mean': 0.5454115867614746, 'reward_meter_std': 0.4758901000022888, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9623986482620239, 'reward_lexical_plausibility_std': 0.08399605005979538, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690458357334137, 'reward_total_composite_mean': 0.16828002035617828, 'reward_total_composite_std': 0.17661982774734497, 'epoch': 0.05}

 28%|██▊       | 918/3300 [3:26:42<9:16:25, 14.02s/it]INFO 04-12 19:25:09 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 919/3300 [3:26:53<8:38:31, 13.07s/it]2026-04-12 19:25:19,803 | INFO | train_grpo_train | metrics_logged mode=train step=919

                                                      
{'loss': -0.0298, 'grad_norm': 11.700063705444336, 'learning_rate': 7.218181818181819e-06, 'num_tokens': 1636297.0, 'completions/mean_length': 35.125, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.24146994948387146, 'rewards/meter/std': 0.2094639390707016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.7204203605651855, 'rewards/repeat_penalty/std': 0.23886586725711823, 'rewards/repeat_floor/mean': 0.9617202281951904, 'rewards/repeat_floor/std': 0.03448401391506195, 'rewards/near_duplicate_penalty/mean': 0.8409971594810486, 'rewards/near_duplicate_penalty/std': 0.12362983077764511, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8813408017158508, 'rewards/distinct_2/std': 0.12492332607507706, 'rewards/total_composite/mean': 0.1139337420463562, 'rewards/total_composite/std': 0.13872671127319336, 'reward': 0.1139337420463562, 'reward_std': 0.13872671127319336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11376723647117615, 'sampling/sampling_logp_difference/max': 3.1247811317443848, 'sampling/importance_sampling_ratio/min': 0.04394655302166939, 'sampling/importance_sampling_ratio/mean': 1.0136206150054932, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7129223123192787, 'clip_ratio/low_mean': 0.04757504374720156, 'clip_ratio/low_min': 0.04757504374720156, 'clip_ratio/high_mean': 0.04945064848288894, 'clip_ratio/high_max': 0.04945064848288894, 'clip_ratio/region_mean': 0.0970256922300905, 'reward_total_mean': 0.1139337420463562, 'reward_meter_mean': 0.24146994948387146, 'reward_meter_std': 0.2094639390707016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7204203605651855, 'reward_repeat_penalty_std': 0.23886586725711823, 'reward_repeat_floor_mean': 0.9617202281951904, 'reward_repeat_floor_std': 0.03448401391506195, 'reward_near_duplicate_penalty_mean': 0.8409971594810486, 'reward_near_duplicate_penalty_std': 0.12362983077764511, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8813408017158508, 'reward_distinct_2_std': 0.12492332607507706, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.1139337420463562, 'reward_total_composite_std': 0.13872671127319336, 'epoch': 0.05}

 28%|██▊       | 919/3300 [3:26:53<8:38:31, 13.07s/it]INFO 04-12 19:25:20 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 920/3300 [3:27:01<7:42:04, 11.65s/it]2026-04-12 19:25:28,138 | INFO | train_grpo_train | metrics_logged mode=train step=920

                                                      
{'loss': 0.2338, 'grad_norm': 20.58114242553711, 'learning_rate': 7.215151515151516e-06, 'num_tokens': 1637623.0, 'completions/mean_length': 23.75, 'completions/min_length': 17.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6520994901657104, 'rewards/meter/std': 0.3819913864135742, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2558424770832062, 'rewards/repeat_penalty/mean': 0.7563884258270264, 'rewards/repeat_penalty/std': 0.06048721447587013, 'rewards/repeat_floor/mean': 0.9820820093154907, 'rewards/repeat_floor/std': 0.0050341845490038395, 'rewards/near_duplicate_penalty/mean': 0.982519805431366, 'rewards/near_duplicate_penalty/std': 0.032421257346868515, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9550449848175049, 'rewards/distinct_2/std': 0.06571410596370697, 'rewards/total_composite/mean': 0.3310341238975525, 'rewards/total_composite/std': 0.21665610373020172, 'reward': 0.3310341238975525, 'reward_std': 0.21665608882904053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16083046793937683, 'sampling/sampling_logp_difference/max': 1.895230770111084, 'sampling/importance_sampling_ratio/min': 0.30498144030570984, 'sampling/importance_sampling_ratio/mean': 1.015960931777954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9854451939463615, 'clip_ratio/low_mean': 0.0378510388545692, 'clip_ratio/low_min': 0.0378510388545692, 'clip_ratio/high_mean': 0.1030723275616765, 'clip_ratio/high_max': 0.1030723275616765, 'clip_ratio/region_mean': 0.1409233664162457, 'reward_total_mean': 0.3310341238975525, 'reward_meter_mean': 0.6520994901657104, 'reward_meter_std': 0.3819913864135742, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7563884258270264, 'reward_repeat_penalty_std': 0.06048721447587013, 'reward_repeat_floor_mean': 0.9820820093154907, 'reward_repeat_floor_std': 0.0050341845490038395, 'reward_near_duplicate_penalty_mean': 0.982519805431366, 'reward_near_duplicate_penalty_std': 0.032421257346868515, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9550449848175049, 'reward_distinct_2_std': 0.06571410596370697, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2558424770832062, 'reward_total_composite_mean': 0.3310341238975525, 'reward_total_composite_std': 0.21665610373020172, 'epoch': 0.05}

 28%|██▊       | 920/3300 [3:27:01<7:42:04, 11.65s/it]INFO 04-12 19:25:28 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 921/3300 [3:27:11<7:20:56, 11.12s/it]2026-04-12 19:25:38,032 | INFO | train_grpo_train | metrics_logged mode=train step=921

                                                      
{'loss': 0.0547, 'grad_norm': 10.257909774780273, 'learning_rate': 7.212121212121212e-06, 'num_tokens': 1639603.0, 'completions/mean_length': 70.5, 'completions/min_length': 66.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.24497263133525848, 'rewards/meter/std': 0.2819669544696808, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.3018253445625305, 'rewards/repeat_penalty/std': 0.25132858753204346, 'rewards/repeat_floor/mean': 0.9001807570457458, 'rewards/repeat_floor/std': 0.059470199048519135, 'rewards/near_duplicate_penalty/mean': 0.7772849202156067, 'rewards/near_duplicate_penalty/std': 0.16345053911209106, 'rewards/opening_diversity/mean': 0.628125011920929, 'rewards/opening_diversity/std': 0.2580619752407074, 'rewards/distinct_2/mean': 0.5100057125091553, 'rewards/distinct_2/std': 0.27947521209716797, 'rewards/total_composite/mean': 0.052619487047195435, 'rewards/total_composite/std': 0.07346183061599731, 'reward': 0.052619487047195435, 'reward_std': 0.07346183061599731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13194158673286438, 'sampling/sampling_logp_difference/max': 2.478618621826172, 'sampling/importance_sampling_ratio/min': 0.08385898172855377, 'sampling/importance_sampling_ratio/mean': 1.0079710483551025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7592105269432068, 'clip_ratio/low_mean': 0.08315952587872744, 'clip_ratio/low_min': 0.08315952587872744, 'clip_ratio/high_mean': 0.02538364566862583, 'clip_ratio/high_max': 0.02538364566862583, 'clip_ratio/region_mean': 0.10854317154735327, 'reward_total_mean': 0.052619487047195435, 'reward_meter_mean': 0.24497263133525848, 'reward_meter_std': 0.2819669544696808, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3018253445625305, 'reward_repeat_penalty_std': 0.25132858753204346, 'reward_repeat_floor_mean': 0.9001807570457458, 'reward_repeat_floor_std': 0.059470199048519135, 'reward_near_duplicate_penalty_mean': 0.7772849202156067, 'reward_near_duplicate_penalty_std': 0.16345053911209106, 'reward_opening_diversity_mean': 0.628125011920929, 'reward_opening_diversity_std': 0.2580619752407074, 'reward_distinct_2_mean': 0.5100057125091553, 'reward_distinct_2_std': 0.27947521209716797, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.052619487047195435, 'reward_total_composite_std': 0.07346183061599731, 'epoch': 0.05}

 28%|██▊       | 921/3300 [3:27:11<7:20:56, 11.12s/it]INFO 04-12 19:25:38 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 922/3300 [3:27:20<6:47:36, 10.28s/it]2026-04-12 19:25:46,362 | INFO | train_grpo_train | metrics_logged mode=train step=922

                                                      
{'loss': -0.0017, 'grad_norm': 15.81523609161377, 'learning_rate': 7.2090909090909104e-06, 'num_tokens': 1641206.0, 'completions/mean_length': 33.375, 'completions/min_length': 30.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.3166966736316681, 'rewards/meter/std': 0.38948431611061096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5087499618530273, 'rewards/judge_quality/std': 0.2811424434185028, 'rewards/repeat_penalty/mean': 0.8077449798583984, 'rewards/repeat_penalty/std': 0.16329263150691986, 'rewards/repeat_floor/mean': 0.9776030778884888, 'rewards/repeat_floor/std': 0.015670295804739, 'rewards/near_duplicate_penalty/mean': 0.9165524840354919, 'rewards/near_duplicate_penalty/std': 0.04324489086866379, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9318910241127014, 'rewards/distinct_2/std': 0.0807490199804306, 'rewards/total_composite/mean': 0.12235023081302643, 'rewards/total_composite/std': 0.13654646277427673, 'reward': 0.12235023081302643, 'reward_std': 0.13654646277427673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12699806690216064, 'sampling/sampling_logp_difference/max': 1.7624311447143555, 'sampling/importance_sampling_ratio/min': 0.17162710428237915, 'sampling/importance_sampling_ratio/mean': 1.005696415901184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7293168902397156, 'clip_ratio/low_mean': 0.0682261842302978, 'clip_ratio/low_min': 0.0682261842302978, 'clip_ratio/high_mean': 0.06333085335791111, 'clip_ratio/high_max': 0.06333085335791111, 'clip_ratio/region_mean': 0.1315570375882089, 'reward_total_mean': 0.12235023081302643, 'reward_meter_mean': 0.3166966736316681, 'reward_meter_std': 0.38948431611061096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8077449798583984, 'reward_repeat_penalty_std': 0.16329263150691986, 'reward_repeat_floor_mean': 0.9776030778884888, 'reward_repeat_floor_std': 0.015670295804739, 'reward_near_duplicate_penalty_mean': 0.9165524840354919, 'reward_near_duplicate_penalty_std': 0.04324489086866379, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9318910241127014, 'reward_distinct_2_std': 0.0807490199804306, 'reward_judge_quality_mean': 0.5087499618530273, 'reward_judge_quality_std': 0.2811424434185028, 'reward_total_composite_mean': 0.12235023081302643, 'reward_total_composite_std': 0.13654646277427673, 'epoch': 0.05}

 28%|██▊       | 922/3300 [3:27:20<6:47:36, 10.28s/it]INFO 04-12 19:25:46 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 923/3300 [3:27:29<6:33:07,  9.92s/it]2026-04-12 19:25:55,463 | INFO | train_grpo_train | metrics_logged mode=train step=923

                                                      
{'loss': 0.023, 'grad_norm': 18.954689025878906, 'learning_rate': 7.206060606060606e-06, 'num_tokens': 1642783.0, 'completions/mean_length': 27.125, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.125, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5961880683898926, 'rewards/meter/std': 0.4088229835033417, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.7672998905181885, 'rewards/repeat_penalty/std': 0.14997050166130066, 'rewards/repeat_floor/mean': 0.9812946319580078, 'rewards/repeat_floor/std': 0.019519859924912453, 'rewards/near_duplicate_penalty/mean': 0.9676051139831543, 'rewards/near_duplicate_penalty/std': 0.08506648987531662, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9711538553237915, 'rewards/distinct_2/std': 0.05723259598016739, 'rewards/total_composite/mean': 0.22934095561504364, 'rewards/total_composite/std': 0.2348758727312088, 'reward': 0.22934095561504364, 'reward_std': 0.2348758727312088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14124515652656555, 'sampling/sampling_logp_difference/max': 1.8496522903442383, 'sampling/importance_sampling_ratio/min': 0.15729184448719025, 'sampling/importance_sampling_ratio/mean': 1.015867829322815, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8634953424334526, 'clip_ratio/low_mean': 0.05253623332828283, 'clip_ratio/low_min': 0.05253623332828283, 'clip_ratio/high_mean': 0.053854302037507296, 'clip_ratio/high_max': 0.053854302037507296, 'clip_ratio/region_mean': 0.10639053536579013, 'reward_total_mean': 0.22934095561504364, 'reward_meter_mean': 0.5961880683898926, 'reward_meter_std': 0.4088229835033417, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7672998905181885, 'reward_repeat_penalty_std': 0.14997050166130066, 'reward_repeat_floor_mean': 0.9812946319580078, 'reward_repeat_floor_std': 0.019519859924912453, 'reward_near_duplicate_penalty_mean': 0.9676051139831543, 'reward_near_duplicate_penalty_std': 0.08506648987531662, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9711538553237915, 'reward_distinct_2_std': 0.05723259598016739, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.22934095561504364, 'reward_total_composite_std': 0.2348758727312088, 'epoch': 0.05}

 28%|██▊       | 923/3300 [3:27:29<6:33:07,  9.92s/it]INFO 04-12 19:25:55 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 924/3300 [3:27:43<7:26:28, 11.27s/it]2026-04-12 19:26:09,889 | INFO | train_grpo_train | metrics_logged mode=train step=924

                                                      
{'loss': -0.0842, 'grad_norm': 5.369450092315674, 'learning_rate': 7.203030303030304e-06, 'num_tokens': 1644342.0, 'completions/mean_length': 91.875, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.85714340209961, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7182058095932007, 'rewards/meter/std': 0.3672333359718323, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9414045810699463, 'rewards/lexical_plausibility/std': 0.16573286056518555, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.8725728988647461, 'rewards/repeat_penalty/std': 0.13889358937740326, 'rewards/repeat_floor/mean': 0.9896508455276489, 'rewards/repeat_floor/std': 0.011627145111560822, 'rewards/near_duplicate_penalty/mean': 0.9781209230422974, 'rewards/near_duplicate_penalty/std': 0.03966702148318291, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.1930149793624878, 'rewards/total_composite/std': 0.14683732390403748, 'reward': 0.1930149793624878, 'reward_std': 0.14683730900287628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17910854518413544, 'sampling/sampling_logp_difference/max': 1.6146667003631592, 'sampling/importance_sampling_ratio/min': 0.19895696640014648, 'sampling/importance_sampling_ratio/mean': 1.0262075662612915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1495482996106148, 'clip_ratio/low_mean': 0.06895678862929344, 'clip_ratio/low_min': 0.06895678862929344, 'clip_ratio/high_mean': 0.06418128684163094, 'clip_ratio/high_max': 0.06418128684163094, 'clip_ratio/region_mean': 0.13313807547092438, 'reward_total_mean': 0.1930149793624878, 'reward_meter_mean': 0.7182058095932007, 'reward_meter_std': 0.3672333359718323, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9414045810699463, 'reward_lexical_plausibility_std': 0.16573286056518555, 'reward_repeat_penalty_mean': 0.8725728988647461, 'reward_repeat_penalty_std': 0.13889358937740326, 'reward_repeat_floor_mean': 0.9896508455276489, 'reward_repeat_floor_std': 0.011627145111560822, 'reward_near_duplicate_penalty_mean': 0.9781209230422974, 'reward_near_duplicate_penalty_std': 0.03966702148318291, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.1930149793624878, 'reward_total_composite_std': 0.14683732390403748, 'epoch': 0.05}

 28%|██▊       | 924/3300 [3:27:43<7:26:28, 11.27s/it]INFO 04-12 19:26:10 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 925/3300 [3:27:57<7:54:42, 11.99s/it]2026-04-12 19:26:23,541 | INFO | train_grpo_train | metrics_logged mode=train step=925

                                                      
{'loss': -0.0084, 'grad_norm': 4.7378830909729, 'learning_rate': 7.2000000000000005e-06, 'num_tokens': 1645972.0, 'completions/mean_length': 101.75, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.142860412597656, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.15823863446712494, 'rewards/meter/std': 0.17890150845050812, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9619423747062683, 'rewards/lexical_plausibility/std': 0.10764322429895401, 'rewards/judge_quality/mean': 0.5162500143051147, 'rewards/judge_quality/std': 0.2672044336795807, 'rewards/repeat_penalty/mean': 0.7766191959381104, 'rewards/repeat_penalty/std': 0.2135908305644989, 'rewards/repeat_floor/mean': 0.9733813405036926, 'rewards/repeat_floor/std': 0.023051032796502113, 'rewards/near_duplicate_penalty/mean': 0.903794527053833, 'rewards/near_duplicate_penalty/std': 0.0657821074128151, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8645795583724976, 'rewards/distinct_2/std': 0.16796272993087769, 'rewards/total_composite/mean': 0.06841225922107697, 'rewards/total_composite/std': 0.08258438110351562, 'reward': 0.06841225922107697, 'reward_std': 0.08258438110351562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12552478909492493, 'sampling/sampling_logp_difference/max': 1.2614984512329102, 'sampling/importance_sampling_ratio/min': 0.2832292914390564, 'sampling/importance_sampling_ratio/mean': 1.014461636543274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.833333745598793, 'clip_ratio/low_mean': 0.07326135924085975, 'clip_ratio/low_min': 0.07326135924085975, 'clip_ratio/high_mean': 0.03728070296347141, 'clip_ratio/high_max': 0.03728070296347141, 'clip_ratio/region_mean': 0.11054206220433116, 'reward_total_mean': 0.06841225922107697, 'reward_meter_mean': 0.15823863446712494, 'reward_meter_std': 0.17890150845050812, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9619423747062683, 'reward_lexical_plausibility_std': 0.10764322429895401, 'reward_repeat_penalty_mean': 0.7766191959381104, 'reward_repeat_penalty_std': 0.2135908305644989, 'reward_repeat_floor_mean': 0.9733813405036926, 'reward_repeat_floor_std': 0.023051032796502113, 'reward_near_duplicate_penalty_mean': 0.903794527053833, 'reward_near_duplicate_penalty_std': 0.0657821074128151, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8645795583724976, 'reward_distinct_2_std': 0.16796272993087769, 'reward_judge_quality_mean': 0.5162500143051147, 'reward_judge_quality_std': 0.2672044336795807, 'reward_total_composite_mean': 0.06841225922107697, 'reward_total_composite_std': 0.08258438110351562, 'epoch': 0.05}

 28%|██▊       | 925/3300 [3:27:57<7:54:42, 11.99s/it]INFO 04-12 19:26:23 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 926/3300 [3:28:11<8:21:39, 12.68s/it]2026-04-12 19:26:37,819 | INFO | train_grpo_train | metrics_logged mode=train step=926

                                                      
{'loss': -0.0729, 'grad_norm': 2.65804123878479, 'learning_rate': 7.196969696969698e-06, 'num_tokens': 1647528.0, 'completions/mean_length': 225.5, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 53.60000228881836, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.3696368932723999, 'rewards/meter/std': 0.3880865275859833, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.07559289783239365, 'rewards/lexical_plausibility/mean': 0.8377869129180908, 'rewards/lexical_plausibility/std': 0.22775839269161224, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.936210036277771, 'rewards/repeat_penalty/std': 0.09846557676792145, 'rewards/repeat_floor/mean': 0.9940980076789856, 'rewards/repeat_floor/std': 0.008051811717450619, 'rewards/near_duplicate_penalty/mean': 0.9833345413208008, 'rewards/near_duplicate_penalty/std': 0.02137635461986065, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.983031690120697, 'rewards/distinct_2/std': 0.04799367114901543, 'rewards/total_composite/mean': 0.13293004035949707, 'rewards/total_composite/std': 0.16459985077381134, 'reward': 0.13293004035949707, 'reward_std': 0.16459985077381134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19071801006793976, 'sampling/sampling_logp_difference/max': 2.3512825965881348, 'sampling/importance_sampling_ratio/min': 0.09524691849946976, 'sampling/importance_sampling_ratio/mean': 1.0463134050369263, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8124346882104874, 'clip_ratio/low_mean': 0.022188347298651934, 'clip_ratio/low_min': 0.022188347298651934, 'clip_ratio/high_mean': 0.03940536640584469, 'clip_ratio/high_max': 0.03940536640584469, 'clip_ratio/region_mean': 0.06159371370449662, 'reward_total_mean': 0.13293004035949707, 'reward_meter_mean': 0.3696368932723999, 'reward_meter_std': 0.3880865275859833, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.07559289783239365, 'reward_lexical_plausibility_mean': 0.8377869129180908, 'reward_lexical_plausibility_std': 0.22775839269161224, 'reward_repeat_penalty_mean': 0.936210036277771, 'reward_repeat_penalty_std': 0.09846557676792145, 'reward_repeat_floor_mean': 0.9940980076789856, 'reward_repeat_floor_std': 0.008051811717450619, 'reward_near_duplicate_penalty_mean': 0.9833345413208008, 'reward_near_duplicate_penalty_std': 0.02137635461986065, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.983031690120697, 'reward_distinct_2_std': 0.04799367114901543, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.13293004035949707, 'reward_total_composite_std': 0.16459985077381134, 'epoch': 0.05}

 28%|██▊       | 926/3300 [3:28:11<8:21:39, 12.68s/it]INFO 04-12 19:26:38 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 927/3300 [3:28:24<8:28:13, 12.85s/it]2026-04-12 19:26:51,071 | INFO | train_grpo_train | metrics_logged mode=train step=927

                                                      
{'loss': -0.0277, 'grad_norm': 4.465095520019531, 'learning_rate': 7.193939393939394e-06, 'num_tokens': 1648892.0, 'completions/mean_length': 81.5, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9119614958763123, 'rewards/meter/std': 0.21772506833076477, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9544011354446411, 'rewards/lexical_plausibility/std': 0.1289730668067932, 'rewards/judge_quality/mean': 0.24625001847743988, 'rewards/judge_quality/std': 0.2774340510368347, 'rewards/repeat_penalty/mean': 0.7469173073768616, 'rewards/repeat_penalty/std': 0.00871915090829134, 'rewards/repeat_floor/mean': 0.9843834638595581, 'rewards/repeat_floor/std': 0.0017438469221815467, 'rewards/near_duplicate_penalty/mean': 0.9958897829055786, 'rewards/near_duplicate_penalty/std': 0.011625555343925953, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20090323686599731, 'rewards/total_composite/std': 0.23400413990020752, 'reward': 0.20090323686599731, 'reward_std': 0.23400413990020752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14231418073177338, 'sampling/sampling_logp_difference/max': 1.2070648670196533, 'sampling/importance_sampling_ratio/min': 0.29907381534576416, 'sampling/importance_sampling_ratio/mean': 1.0327887535095215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1380844712257385, 'clip_ratio/low_mean': 0.06239557405933738, 'clip_ratio/low_min': 0.06239557405933738, 'clip_ratio/high_mean': 0.03060640674084425, 'clip_ratio/high_max': 0.03060640674084425, 'clip_ratio/region_mean': 0.09300198080018163, 'reward_total_mean': 0.20090323686599731, 'reward_meter_mean': 0.9119614958763123, 'reward_meter_std': 0.21772506833076477, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9544011354446411, 'reward_lexical_plausibility_std': 0.1289730668067932, 'reward_repeat_penalty_mean': 0.7469173073768616, 'reward_repeat_penalty_std': 0.00871915090829134, 'reward_repeat_floor_mean': 0.9843834638595581, 'reward_repeat_floor_std': 0.0017438469221815467, 'reward_near_duplicate_penalty_mean': 0.9958897829055786, 'reward_near_duplicate_penalty_std': 0.011625555343925953, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.24625001847743988, 'reward_judge_quality_std': 0.2774340510368347, 'reward_total_composite_mean': 0.20090323686599731, 'reward_total_composite_std': 0.23400413990020752, 'epoch': 0.05}

 28%|██▊       | 927/3300 [3:28:24<8:28:13, 12.85s/it]INFO 04-12 19:26:51 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 928/3300 [3:28:33<7:36:33, 11.55s/it]2026-04-12 19:26:59,595 | INFO | train_grpo_train | metrics_logged mode=train step=928

                                                      
{'loss': 0.1078, 'grad_norm': 15.49630069732666, 'learning_rate': 7.1909090909090914e-06, 'num_tokens': 1650338.0, 'completions/mean_length': 38.75, 'completions/min_length': 32.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.42397141456604004, 'rewards/meter/std': 0.4718492925167084, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.35602855682373047, 'rewards/repeat_penalty/std': 0.3165453374385834, 'rewards/repeat_floor/mean': 0.9027478694915771, 'rewards/repeat_floor/std': 0.07226752489805222, 'rewards/near_duplicate_penalty/mean': 0.7224303483963013, 'rewards/near_duplicate_penalty/std': 0.27404770255088806, 'rewards/opening_diversity/mean': 0.640625, 'rewards/opening_diversity/std': 0.23563650250434875, 'rewards/distinct_2/mean': 0.6216205954551697, 'rewards/distinct_2/std': 0.2848329246044159, 'rewards/total_composite/mean': 0.09111328423023224, 'rewards/total_composite/std': 0.11265230178833008, 'reward': 0.09111328423023224, 'reward_std': 0.11265230178833008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11080250889062881, 'sampling/sampling_logp_difference/max': 1.2789342403411865, 'sampling/importance_sampling_ratio/min': 0.27833378314971924, 'sampling/importance_sampling_ratio/mean': 1.0121363401412964, 'sampling/importance_sampling_ratio/max': 1.9798465967178345, 'entropy': 0.751460611820221, 'clip_ratio/low_mean': 0.042764355428516865, 'clip_ratio/low_min': 0.042764355428516865, 'clip_ratio/high_mean': 0.04218073608353734, 'clip_ratio/high_max': 0.04218073608353734, 'clip_ratio/region_mean': 0.0849450915120542, 'reward_total_mean': 0.09111328423023224, 'reward_meter_mean': 0.42397141456604004, 'reward_meter_std': 0.4718492925167084, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.35602855682373047, 'reward_repeat_penalty_std': 0.3165453374385834, 'reward_repeat_floor_mean': 0.9027478694915771, 'reward_repeat_floor_std': 0.07226752489805222, 'reward_near_duplicate_penalty_mean': 0.7224303483963013, 'reward_near_duplicate_penalty_std': 0.27404770255088806, 'reward_opening_diversity_mean': 0.640625, 'reward_opening_diversity_std': 0.23563650250434875, 'reward_distinct_2_mean': 0.6216205954551697, 'reward_distinct_2_std': 0.2848329246044159, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.09111328423023224, 'reward_total_composite_std': 0.11265230178833008, 'epoch': 0.05}

 28%|██▊       | 928/3300 [3:28:33<7:36:33, 11.55s/it]INFO 04-12 19:26:59 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 929/3300 [3:28:41<7:01:23, 10.66s/it]2026-04-12 19:27:08,179 | INFO | train_grpo_train | metrics_logged mode=train step=929

                                                      
{'loss': 0.2541, 'grad_norm': 11.848160743713379, 'learning_rate': 7.187878787878788e-06, 'num_tokens': 1651993.0, 'completions/mean_length': 41.875, 'completions/min_length': 20.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6769824028015137, 'rewards/meter/std': 0.3021802306175232, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.29546841979026794, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.08864052593708038, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.855697751045227, 'rewards/repeat_penalty/std': 0.1298636645078659, 'rewards/repeat_floor/mean': 0.9862238168716431, 'rewards/repeat_floor/std': 0.014152740128338337, 'rewards/near_duplicate_penalty/mean': 0.96401047706604, 'rewards/near_duplicate_penalty/std': 0.05547286197543144, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9485810995101929, 'rewards/distinct_2/std': 0.08081252127885818, 'rewards/total_composite/mean': 0.2899732291698456, 'rewards/total_composite/std': 0.1964443475008011, 'reward': 0.2899732291698456, 'reward_std': 0.1964443176984787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11191190034151077, 'sampling/sampling_logp_difference/max': 2.764425039291382, 'sampling/importance_sampling_ratio/min': 0.06301231682300568, 'sampling/importance_sampling_ratio/mean': 1.0081783533096313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6256960667669773, 'clip_ratio/low_mean': 0.048545082565397024, 'clip_ratio/low_min': 0.048545082565397024, 'clip_ratio/high_mean': 0.05867794528603554, 'clip_ratio/high_max': 0.05867794528603554, 'clip_ratio/region_mean': 0.10722302785143256, 'reward_total_mean': 0.2899732291698456, 'reward_meter_mean': 0.6769824028015137, 'reward_meter_std': 0.3021802306175232, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.29546841979026794, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.08864052593708038, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.855697751045227, 'reward_repeat_penalty_std': 0.1298636645078659, 'reward_repeat_floor_mean': 0.9862238168716431, 'reward_repeat_floor_std': 0.014152740128338337, 'reward_near_duplicate_penalty_mean': 0.96401047706604, 'reward_near_duplicate_penalty_std': 0.05547286197543144, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9485810995101929, 'reward_distinct_2_std': 0.08081252127885818, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.2899732291698456, 'reward_total_composite_std': 0.1964443475008011, 'epoch': 0.05}

 28%|██▊       | 929/3300 [3:28:42<7:01:23, 10.66s/it]INFO 04-12 19:27:08 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 930/3300 [3:28:56<7:43:58, 11.75s/it]2026-04-12 19:27:22,464 | INFO | train_grpo_train | metrics_logged mode=train step=930

                                                      
{'loss': -0.055, 'grad_norm': 3.5176053047180176, 'learning_rate': 7.184848484848486e-06, 'num_tokens': 1653591.0, 'completions/mean_length': 167.75, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.3827265501022339, 'rewards/meter/std': 0.3244231641292572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9070299863815308, 'rewards/lexical_plausibility/std': 0.17688804864883423, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.8969683647155762, 'rewards/repeat_penalty/std': 0.09577226638793945, 'rewards/repeat_floor/mean': 0.9901719093322754, 'rewards/repeat_floor/std': 0.0072926925495266914, 'rewards/near_duplicate_penalty/mean': 0.9601371884346008, 'rewards/near_duplicate_penalty/std': 0.026828937232494354, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9572371244430542, 'rewards/distinct_2/std': 0.053154509514570236, 'rewards/total_composite/mean': 0.0810088962316513, 'rewards/total_composite/std': 0.08484596014022827, 'reward': 0.0810088962316513, 'reward_std': 0.08484595268964767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1922956258058548, 'sampling/sampling_logp_difference/max': 1.2328667640686035, 'sampling/importance_sampling_ratio/min': 0.29145583510398865, 'sampling/importance_sampling_ratio/mean': 1.0240886211395264, 'sampling/importance_sampling_ratio/max': 1.9131968021392822, 'entropy': 1.2628989219665527, 'clip_ratio/low_mean': 0.06400027498602867, 'clip_ratio/low_min': 0.06400027498602867, 'clip_ratio/high_mean': 0.08033769205212593, 'clip_ratio/high_max': 0.08033769205212593, 'clip_ratio/region_mean': 0.1443379670381546, 'reward_total_mean': 0.0810088962316513, 'reward_meter_mean': 0.3827265501022339, 'reward_meter_std': 0.3244231641292572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9070299863815308, 'reward_lexical_plausibility_std': 0.17688804864883423, 'reward_repeat_penalty_mean': 0.8969683647155762, 'reward_repeat_penalty_std': 0.09577226638793945, 'reward_repeat_floor_mean': 0.9901719093322754, 'reward_repeat_floor_std': 0.0072926925495266914, 'reward_near_duplicate_penalty_mean': 0.9601371884346008, 'reward_near_duplicate_penalty_std': 0.026828937232494354, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9572371244430542, 'reward_distinct_2_std': 0.053154509514570236, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.0810088962316513, 'reward_total_composite_std': 0.08484596014022827, 'epoch': 0.05}

 28%|██▊       | 930/3300 [3:28:56<7:43:58, 11.75s/it]INFO 04-12 19:27:22 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 931/3300 [3:29:05<7:12:44, 10.96s/it]2026-04-12 19:27:31,580 | INFO | train_grpo_train | metrics_logged mode=train step=931

                                                      
{'loss': 0.0754, 'grad_norm': 14.462099075317383, 'learning_rate': 7.181818181818182e-06, 'num_tokens': 1655152.0, 'completions/mean_length': 40.125, 'completions/min_length': 30.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.5972735285758972, 'rewards/meter/std': 0.4042917788028717, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.8369426131248474, 'rewards/repeat_penalty/std': 0.2847599387168884, 'rewards/repeat_floor/mean': 0.9788985848426819, 'rewards/repeat_floor/std': 0.038716524839401245, 'rewards/near_duplicate_penalty/mean': 0.9422450065612793, 'rewards/near_duplicate_penalty/std': 0.09756415337324142, 'rewards/opening_diversity/mean': 0.921875, 'rewards/opening_diversity/std': 0.22097088396549225, 'rewards/distinct_2/mean': 0.9138815402984619, 'rewards/distinct_2/std': 0.13255546987056732, 'rewards/total_composite/mean': 0.2229437232017517, 'rewards/total_composite/std': 0.12855421006679535, 'reward': 0.2229437232017517, 'reward_std': 0.12855419516563416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11774379760026932, 'sampling/sampling_logp_difference/max': 1.618084192276001, 'sampling/importance_sampling_ratio/min': 0.19827820360660553, 'sampling/importance_sampling_ratio/mean': 1.0261934995651245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8934935554862022, 'clip_ratio/low_mean': 0.04492550063878298, 'clip_ratio/low_min': 0.04492550063878298, 'clip_ratio/high_mean': 0.10349396243691444, 'clip_ratio/high_max': 0.10349396243691444, 'clip_ratio/region_mean': 0.14841946307569742, 'reward_total_mean': 0.2229437232017517, 'reward_meter_mean': 0.5972735285758972, 'reward_meter_std': 0.4042917788028717, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8369426131248474, 'reward_repeat_penalty_std': 0.2847599387168884, 'reward_repeat_floor_mean': 0.9788985848426819, 'reward_repeat_floor_std': 0.038716524839401245, 'reward_near_duplicate_penalty_mean': 0.9422450065612793, 'reward_near_duplicate_penalty_std': 0.09756415337324142, 'reward_opening_diversity_mean': 0.921875, 'reward_opening_diversity_std': 0.22097088396549225, 'reward_distinct_2_mean': 0.9138815402984619, 'reward_distinct_2_std': 0.13255546987056732, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.2229437232017517, 'reward_total_composite_std': 0.12855421006679535, 'epoch': 0.05}

 28%|██▊       | 931/3300 [3:29:05<7:12:44, 10.96s/it]INFO 04-12 19:27:31 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 932/3300 [3:29:20<8:06:08, 12.32s/it]2026-04-12 19:27:47,064 | INFO | train_grpo_train | metrics_logged mode=train step=932

                                                      
{'loss': -0.0486, 'grad_norm': 1.9425721168518066, 'learning_rate': 7.17878787878788e-06, 'num_tokens': 1656700.0, 'completions/mean_length': 145.5, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 23.33333396911621, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9154078364372253, 'rewards/meter/std': 0.12859348952770233, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9230225086212158, 'rewards/lexical_plausibility/std': 0.14301204681396484, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.8125, 'rewards/repeat_penalty/std': 0.1157275140285492, 'rewards/repeat_floor/mean': 0.9887499809265137, 'rewards/repeat_floor/std': 0.0069436440244317055, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.23230081796646118, 'rewards/total_composite/std': 0.1508869230747223, 'reward': 0.23230081796646118, 'reward_std': 0.1508869230747223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15069572627544403, 'sampling/sampling_logp_difference/max': 0.9685931205749512, 'sampling/importance_sampling_ratio/min': 0.37961673736572266, 'sampling/importance_sampling_ratio/mean': 1.0240278244018555, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.934409961104393, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.10361413098871708, 'clip_ratio/high_max': 0.10361413098871708, 'clip_ratio/region_mean': 0.11923913098871708, 'reward_total_mean': 0.23230081796646118, 'reward_meter_mean': 0.9154078364372253, 'reward_meter_std': 0.12859348952770233, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9230225086212158, 'reward_lexical_plausibility_std': 0.14301204681396484, 'reward_repeat_penalty_mean': 0.8125, 'reward_repeat_penalty_std': 0.1157275140285492, 'reward_repeat_floor_mean': 0.9887499809265137, 'reward_repeat_floor_std': 0.0069436440244317055, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.23230081796646118, 'reward_total_composite_std': 0.1508869230747223, 'epoch': 0.05}

 28%|██▊       | 932/3300 [3:29:20<8:06:08, 12.32s/it]INFO 04-12 19:27:47 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 933/3300 [3:29:34<8:25:58, 12.83s/it]2026-04-12 19:28:01,080 | INFO | train_grpo_train | metrics_logged mode=train step=933

                                                      
{'loss': -0.0018, 'grad_norm': 4.698115825653076, 'learning_rate': 7.175757575757576e-06, 'num_tokens': 1658291.0, 'completions/mean_length': 104.875, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 46.71428680419922, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.22150354087352753, 'rewards/meter/std': 0.2616729438304901, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9202638864517212, 'rewards/lexical_plausibility/std': 0.15705610811710358, 'rewards/judge_quality/mean': 0.6412500143051147, 'rewards/judge_quality/std': 0.40417245030403137, 'rewards/repeat_penalty/mean': 0.9221000671386719, 'rewards/repeat_penalty/std': 0.07147147506475449, 'rewards/repeat_floor/mean': 0.991619348526001, 'rewards/repeat_floor/std': 0.007712433114647865, 'rewards/near_duplicate_penalty/mean': 0.9801720380783081, 'rewards/near_duplicate_penalty/std': 0.019732806831598282, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9399279356002808, 'rewards/distinct_2/std': 0.05639173462986946, 'rewards/total_composite/mean': 0.16178405284881592, 'rewards/total_composite/std': 0.2549450397491455, 'reward': 0.16178405284881592, 'reward_std': 0.2549450397491455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15619610249996185, 'sampling/sampling_logp_difference/max': 3.191709518432617, 'sampling/importance_sampling_ratio/min': 0.12825708091259003, 'sampling/importance_sampling_ratio/mean': 1.0248557329177856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9056680053472519, 'clip_ratio/low_mean': 0.08119824156165123, 'clip_ratio/low_min': 0.08119824156165123, 'clip_ratio/high_mean': 0.04029933363199234, 'clip_ratio/high_max': 0.04029933363199234, 'clip_ratio/region_mean': 0.12149757519364357, 'reward_total_mean': 0.16178405284881592, 'reward_meter_mean': 0.22150354087352753, 'reward_meter_std': 0.2616729438304901, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9202638864517212, 'reward_lexical_plausibility_std': 0.15705610811710358, 'reward_repeat_penalty_mean': 0.9221000671386719, 'reward_repeat_penalty_std': 0.07147147506475449, 'reward_repeat_floor_mean': 0.991619348526001, 'reward_repeat_floor_std': 0.007712433114647865, 'reward_near_duplicate_penalty_mean': 0.9801720380783081, 'reward_near_duplicate_penalty_std': 0.019732806831598282, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9399279356002808, 'reward_distinct_2_std': 0.05639173462986946, 'reward_judge_quality_mean': 0.6412500143051147, 'reward_judge_quality_std': 0.40417245030403137, 'reward_total_composite_mean': 0.16178405284881592, 'reward_total_composite_std': 0.2549450397491455, 'epoch': 0.05}

 28%|██▊       | 933/3300 [3:29:34<8:25:58, 12.83s/it]INFO 04-12 19:28:01 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 934/3300 [3:29:43<7:34:53, 11.54s/it]2026-04-12 19:28:09,657 | INFO | train_grpo_train | metrics_logged mode=train step=934

                                                      
{'loss': 0.0094, 'grad_norm': 22.757076263427734, 'learning_rate': 7.172727272727273e-06, 'num_tokens': 1659744.0, 'completions/mean_length': 18.625, 'completions/min_length': 16.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7352614402770996, 'rewards/meter/std': 0.29680031538009644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8362500667572021, 'rewards/judge_quality/std': 0.23688077926635742, 'rewards/repeat_penalty/mean': 0.6889147162437439, 'rewards/repeat_penalty/std': 0.10295876860618591, 'rewards/repeat_floor/mean': 0.9727829694747925, 'rewards/repeat_floor/std': 0.020591754466295242, 'rewards/near_duplicate_penalty/mean': 0.9185529947280884, 'rewards/near_duplicate_penalty/std': 0.13727836310863495, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5794228911399841, 'rewards/total_composite/std': 0.287330687046051, 'reward': 0.5794228911399841, 'reward_std': 0.28733065724372864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13337747752666473, 'sampling/sampling_logp_difference/max': 1.0140495300292969, 'sampling/importance_sampling_ratio/min': 0.36274704337120056, 'sampling/importance_sampling_ratio/mean': 1.0323998928070068, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0619853287935257, 'clip_ratio/low_mean': 0.04961622832342982, 'clip_ratio/low_min': 0.04961622832342982, 'clip_ratio/high_mean': 0.08143520634621382, 'clip_ratio/high_max': 0.08143520634621382, 'clip_ratio/region_mean': 0.13105143466964364, 'reward_total_mean': 0.5794228911399841, 'reward_meter_mean': 0.7352614402770996, 'reward_meter_std': 0.29680031538009644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6889147162437439, 'reward_repeat_penalty_std': 0.10295876860618591, 'reward_repeat_floor_mean': 0.9727829694747925, 'reward_repeat_floor_std': 0.020591754466295242, 'reward_near_duplicate_penalty_mean': 0.9185529947280884, 'reward_near_duplicate_penalty_std': 0.13727836310863495, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8362500667572021, 'reward_judge_quality_std': 0.23688077926635742, 'reward_total_composite_mean': 0.5794228911399841, 'reward_total_composite_std': 0.287330687046051, 'epoch': 0.05}

 28%|██▊       | 934/3300 [3:29:43<7:34:53, 11.54s/it]INFO 04-12 19:28:10 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 935/3300 [3:29:54<7:29:01, 11.39s/it]2026-04-12 19:28:20,680 | INFO | train_grpo_train | metrics_logged mode=train step=935

                                                      
{'loss': 0.0096, 'grad_norm': 22.159223556518555, 'learning_rate': 7.16969696969697e-06, 'num_tokens': 1661219.0, 'completions/mean_length': 28.375, 'completions/min_length': 25.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6256688833236694, 'rewards/meter/std': 0.3753491938114166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9661613702774048, 'rewards/repeat_penalty/std': 0.0639868751168251, 'rewards/repeat_floor/mean': 0.995513379573822, 'rewards/repeat_floor/std': 0.008002697490155697, 'rewards/near_duplicate_penalty/mean': 0.9777815341949463, 'rewards/near_duplicate_penalty/std': 0.03342784568667412, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.4539850354194641, 'rewards/total_composite/std': 0.3794321119785309, 'reward': 0.4539850354194641, 'reward_std': 0.3794320821762085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.126794695854187, 'sampling/sampling_logp_difference/max': 1.9264278411865234, 'sampling/importance_sampling_ratio/min': 0.14566762745380402, 'sampling/importance_sampling_ratio/mean': 1.0171736478805542, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6241744309663773, 'clip_ratio/low_mean': 0.0650077536702156, 'clip_ratio/low_min': 0.0650077536702156, 'clip_ratio/high_mean': 0.04758361866697669, 'clip_ratio/high_max': 0.04758361866697669, 'clip_ratio/region_mean': 0.1125913723371923, 'reward_total_mean': 0.4539850354194641, 'reward_meter_mean': 0.6256688833236694, 'reward_meter_std': 0.3753491938114166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9661613702774048, 'reward_repeat_penalty_std': 0.0639868751168251, 'reward_repeat_floor_mean': 0.995513379573822, 'reward_repeat_floor_std': 0.008002697490155697, 'reward_near_duplicate_penalty_mean': 0.9777815341949463, 'reward_near_duplicate_penalty_std': 0.03342784568667412, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.4539850354194641, 'reward_total_composite_std': 0.3794321119785309, 'epoch': 0.05}

 28%|██▊       | 935/3300 [3:29:54<7:29:01, 11.39s/it]INFO 04-12 19:28:20 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 936/3300 [3:30:07<7:54:00, 12.03s/it]2026-04-12 19:28:34,183 | INFO | train_grpo_train | metrics_logged mode=train step=936

                                                      
{'loss': -0.0706, 'grad_norm': 2.846550941467285, 'learning_rate': 7.166666666666667e-06, 'num_tokens': 1662629.0, 'completions/mean_length': 158.25, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 40.333335876464844, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6912367343902588, 'rewards/meter/std': 0.3497598171234131, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8528701663017273, 'rewards/lexical_plausibility/std': 0.2725129723548889, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9377641677856445, 'rewards/repeat_penalty/std': 0.08006878197193146, 'rewards/repeat_floor/mean': 0.9920282363891602, 'rewards/repeat_floor/std': 0.009480367414653301, 'rewards/near_duplicate_penalty/mean': 0.9638878107070923, 'rewards/near_duplicate_penalty/std': 0.03365929797291756, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.971611738204956, 'rewards/distinct_2/std': 0.056806087493896484, 'rewards/total_composite/mean': 0.1744149923324585, 'rewards/total_composite/std': 0.1357843279838562, 'reward': 0.1744149923324585, 'reward_std': 0.1357843279838562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1601579189300537, 'sampling/sampling_logp_difference/max': 1.1780023574829102, 'sampling/importance_sampling_ratio/min': 0.30789318680763245, 'sampling/importance_sampling_ratio/mean': 1.0185701847076416, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0627938359975815, 'clip_ratio/low_mean': 0.029645269736647606, 'clip_ratio/low_min': 0.029645269736647606, 'clip_ratio/high_mean': 0.07594627793878317, 'clip_ratio/high_max': 0.07594627793878317, 'clip_ratio/region_mean': 0.10559154767543077, 'reward_total_mean': 0.1744149923324585, 'reward_meter_mean': 0.6912367343902588, 'reward_meter_std': 0.3497598171234131, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8528701663017273, 'reward_lexical_plausibility_std': 0.2725129723548889, 'reward_repeat_penalty_mean': 0.9377641677856445, 'reward_repeat_penalty_std': 0.08006878197193146, 'reward_repeat_floor_mean': 0.9920282363891602, 'reward_repeat_floor_std': 0.009480367414653301, 'reward_near_duplicate_penalty_mean': 0.9638878107070923, 'reward_near_duplicate_penalty_std': 0.03365929797291756, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.971611738204956, 'reward_distinct_2_std': 0.056806087493896484, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.1744149923324585, 'reward_total_composite_std': 0.1357843279838562, 'epoch': 0.05}

 28%|██▊       | 936/3300 [3:30:08<7:54:00, 12.03s/it]INFO 04-12 19:28:34 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 937/3300 [3:30:21<8:13:37, 12.53s/it]2026-04-12 19:28:47,887 | INFO | train_grpo_train | metrics_logged mode=train step=937

                                                      
{'loss': -0.0842, 'grad_norm': 4.684035778045654, 'learning_rate': 7.163636363636363e-06, 'num_tokens': 1664126.0, 'completions/mean_length': 97.125, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.85714340209961, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7876128554344177, 'rewards/meter/std': 0.26310455799102783, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.957014799118042, 'rewards/lexical_plausibility/std': 0.12158055603504181, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.8698775768280029, 'rewards/repeat_penalty/std': 0.17616939544677734, 'rewards/repeat_floor/mean': 0.9855564832687378, 'rewards/repeat_floor/std': 0.01942961849272251, 'rewards/near_duplicate_penalty/mean': 0.9460172057151794, 'rewards/near_duplicate_penalty/std': 0.06474689394235611, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9711538553237915, 'rewards/distinct_2/std': 0.0815892368555069, 'rewards/total_composite/mean': 0.23081491887569427, 'rewards/total_composite/std': 0.13306322693824768, 'reward': 0.23081491887569427, 'reward_std': 0.13306322693824768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1624113917350769, 'sampling/sampling_logp_difference/max': 1.279693365097046, 'sampling/importance_sampling_ratio/min': 0.27812254428863525, 'sampling/importance_sampling_ratio/mean': 1.0237486362457275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0300191193819046, 'clip_ratio/low_mean': 0.04493092838674784, 'clip_ratio/low_min': 0.04493092838674784, 'clip_ratio/high_mean': 0.08559286501258612, 'clip_ratio/high_max': 0.08559286501258612, 'clip_ratio/region_mean': 0.13052379339933395, 'reward_total_mean': 0.23081491887569427, 'reward_meter_mean': 0.7876128554344177, 'reward_meter_std': 0.26310455799102783, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.957014799118042, 'reward_lexical_plausibility_std': 0.12158055603504181, 'reward_repeat_penalty_mean': 0.8698775768280029, 'reward_repeat_penalty_std': 0.17616939544677734, 'reward_repeat_floor_mean': 0.9855564832687378, 'reward_repeat_floor_std': 0.01942961849272251, 'reward_near_duplicate_penalty_mean': 0.9460172057151794, 'reward_near_duplicate_penalty_std': 0.06474689394235611, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9711538553237915, 'reward_distinct_2_std': 0.0815892368555069, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.23081491887569427, 'reward_total_composite_std': 0.13306322693824768, 'epoch': 0.05}

 28%|██▊       | 937/3300 [3:30:21<8:13:37, 12.53s/it]INFO 04-12 19:28:48 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 938/3300 [3:30:35<8:28:00, 12.90s/it]2026-04-12 19:29:01,656 | INFO | train_grpo_train | metrics_logged mode=train step=938

                                                      
{'loss': -0.0571, 'grad_norm': 2.6473395824432373, 'learning_rate': 7.1606060606060615e-06, 'num_tokens': 1665748.0, 'completions/mean_length': 154.75, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.66666793823242, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.3612874150276184, 'rewards/meter/std': 0.39760667085647583, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.907730221748352, 'rewards/lexical_plausibility/std': 0.1780407875776291, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.335644394159317, 'rewards/repeat_penalty/mean': 0.8570314645767212, 'rewards/repeat_penalty/std': 0.13955850899219513, 'rewards/repeat_floor/mean': 0.9853149652481079, 'rewards/repeat_floor/std': 0.013978006318211555, 'rewards/near_duplicate_penalty/mean': 0.950524091720581, 'rewards/near_duplicate_penalty/std': 0.06395603716373444, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9609594345092773, 'rewards/distinct_2/std': 0.0558197945356369, 'rewards/total_composite/mean': 0.13955768942832947, 'rewards/total_composite/std': 0.13091062009334564, 'reward': 0.13955768942832947, 'reward_std': 0.13091060519218445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1684911847114563, 'sampling/sampling_logp_difference/max': 2.065189838409424, 'sampling/importance_sampling_ratio/min': 0.1267942190170288, 'sampling/importance_sampling_ratio/mean': 1.0257047414779663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6408987268805504, 'clip_ratio/low_mean': 0.0464743603952229, 'clip_ratio/low_min': 0.0464743603952229, 'clip_ratio/high_mean': 0.05412051361054182, 'clip_ratio/high_max': 0.05412051361054182, 'clip_ratio/region_mean': 0.10059487400576472, 'reward_total_mean': 0.13955768942832947, 'reward_meter_mean': 0.3612874150276184, 'reward_meter_std': 0.39760667085647583, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.907730221748352, 'reward_lexical_plausibility_std': 0.1780407875776291, 'reward_repeat_penalty_mean': 0.8570314645767212, 'reward_repeat_penalty_std': 0.13955850899219513, 'reward_repeat_floor_mean': 0.9853149652481079, 'reward_repeat_floor_std': 0.013978006318211555, 'reward_near_duplicate_penalty_mean': 0.950524091720581, 'reward_near_duplicate_penalty_std': 0.06395603716373444, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9609594345092773, 'reward_distinct_2_std': 0.0558197945356369, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.335644394159317, 'reward_total_composite_mean': 0.13955768942832947, 'reward_total_composite_std': 0.13091062009334564, 'epoch': 0.05}

 28%|██▊       | 938/3300 [3:30:35<8:28:00, 12.90s/it]INFO 04-12 19:29:01 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 939/3300 [3:30:48<8:34:05, 13.06s/it]2026-04-12 19:29:15,099 | INFO | train_grpo_train | metrics_logged mode=train step=939

                                                      
{'loss': -0.0472, 'grad_norm': 4.579878330230713, 'learning_rate': 7.157575757575758e-06, 'num_tokens': 1667290.0, 'completions/mean_length': 95.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.28571701049805, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5754709243774414, 'rewards/meter/std': 0.35846441984176636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9475219249725342, 'rewards/lexical_plausibility/std': 0.14843030273914337, 'rewards/judge_quality/mean': 0.6262500286102295, 'rewards/judge_quality/std': 0.3452095687389374, 'rewards/repeat_penalty/mean': 0.6471267342567444, 'rewards/repeat_penalty/std': 0.17849737405776978, 'rewards/repeat_floor/mean': 0.963615894317627, 'rewards/repeat_floor/std': 0.021028872579336166, 'rewards/near_duplicate_penalty/mean': 0.9201711416244507, 'rewards/near_duplicate_penalty/std': 0.04978715255856514, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8537800908088684, 'rewards/distinct_2/std': 0.11982870101928711, 'rewards/total_composite/mean': 0.38300296664237976, 'rewards/total_composite/std': 0.3078805208206177, 'reward': 0.38300296664237976, 'reward_std': 0.3078805208206177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1256883144378662, 'sampling/sampling_logp_difference/max': 0.9241501092910767, 'sampling/importance_sampling_ratio/min': 0.39686858654022217, 'sampling/importance_sampling_ratio/mean': 1.0176739692687988, 'sampling/importance_sampling_ratio/max': 1.9200845956802368, 'entropy': 0.6841282397508621, 'clip_ratio/low_mean': 0.02996472967788577, 'clip_ratio/low_min': 0.02996472967788577, 'clip_ratio/high_mean': 0.06550312554463744, 'clip_ratio/high_max': 0.06550312554463744, 'clip_ratio/region_mean': 0.09546785522252321, 'reward_total_mean': 0.38300296664237976, 'reward_meter_mean': 0.5754709243774414, 'reward_meter_std': 0.35846441984176636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9475219249725342, 'reward_lexical_plausibility_std': 0.14843030273914337, 'reward_repeat_penalty_mean': 0.6471267342567444, 'reward_repeat_penalty_std': 0.17849737405776978, 'reward_repeat_floor_mean': 0.963615894317627, 'reward_repeat_floor_std': 0.021028872579336166, 'reward_near_duplicate_penalty_mean': 0.9201711416244507, 'reward_near_duplicate_penalty_std': 0.04978715255856514, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8537800908088684, 'reward_distinct_2_std': 0.11982870101928711, 'reward_judge_quality_mean': 0.6262500286102295, 'reward_judge_quality_std': 0.3452095687389374, 'reward_total_composite_mean': 0.38300296664237976, 'reward_total_composite_std': 0.3078805208206177, 'epoch': 0.05}

 28%|██▊       | 939/3300 [3:30:48<8:34:05, 13.06s/it]INFO 04-12 19:29:15 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 940/3300 [3:31:02<8:44:46, 13.34s/it]2026-04-12 19:29:29,084 | INFO | train_grpo_train | metrics_logged mode=train step=940

                                                      
{'loss': -0.0455, 'grad_norm': 2.4794535636901855, 'learning_rate': 7.154545454545455e-06, 'num_tokens': 1668995.0, 'completions/mean_length': 225.125, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8009392023086548, 'rewards/meter/std': 0.34026679396629333, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.30860671401023865, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.09258200228214264, 'rewards/lexical_plausibility/mean': 0.853161096572876, 'rewards/lexical_plausibility/std': 0.20306220650672913, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.7446501851081848, 'rewards/repeat_penalty/std': 0.29569900035858154, 'rewards/repeat_floor/mean': 0.970337986946106, 'rewards/repeat_floor/std': 0.04130903631448746, 'rewards/near_duplicate_penalty/mean': 0.9298611879348755, 'rewards/near_duplicate_penalty/std': 0.09926348179578781, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.155838742852211, 'rewards/distinct_2/mean': 0.8706530928611755, 'rewards/distinct_2/std': 0.21183568239212036, 'rewards/total_composite/mean': 0.13488975167274475, 'rewards/total_composite/std': 0.12976008653640747, 'reward': 0.13488975167274475, 'reward_std': 0.12976008653640747, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1553594022989273, 'sampling/sampling_logp_difference/max': 2.8208189010620117, 'sampling/importance_sampling_ratio/min': 0.059557151049375534, 'sampling/importance_sampling_ratio/mean': 1.025700330734253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7399625405669212, 'clip_ratio/low_mean': 0.045493897050619125, 'clip_ratio/low_min': 0.045493897050619125, 'clip_ratio/high_mean': 0.02914167894050479, 'clip_ratio/high_max': 0.02914167894050479, 'clip_ratio/region_mean': 0.07463557599112391, 'reward_total_mean': 0.13488975167274475, 'reward_meter_mean': 0.8009392023086548, 'reward_meter_std': 0.34026679396629333, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.30860671401023865, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.09258200228214264, 'reward_lexical_plausibility_mean': 0.853161096572876, 'reward_lexical_plausibility_std': 0.20306220650672913, 'reward_repeat_penalty_mean': 0.7446501851081848, 'reward_repeat_penalty_std': 0.29569900035858154, 'reward_repeat_floor_mean': 0.970337986946106, 'reward_repeat_floor_std': 0.04130903631448746, 'reward_near_duplicate_penalty_mean': 0.9298611879348755, 'reward_near_duplicate_penalty_std': 0.09926348179578781, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.155838742852211, 'reward_distinct_2_mean': 0.8706530928611755, 'reward_distinct_2_std': 0.21183568239212036, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.13488975167274475, 'reward_total_composite_std': 0.12976008653640747, 'epoch': 0.05}

 28%|██▊       | 940/3300 [3:31:02<8:44:46, 13.34s/it]INFO 04-12 19:29:29 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 941/3300 [3:31:18<9:11:36, 14.03s/it]2026-04-12 19:29:44,724 | INFO | train_grpo_train | metrics_logged mode=train step=941

                                                      
{'loss': -0.0797, 'grad_norm': 1.528396725654602, 'learning_rate': 7.151515151515152e-06, 'num_tokens': 1670416.0, 'completions/mean_length': 280.625, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6363770961761475, 'rewards/meter/std': 0.35401707887649536, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.05175493285059929, 'rewards/lexical_plausibility/mean': 0.780329704284668, 'rewards/lexical_plausibility/std': 0.23699282109737396, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.21380899846553802, 'rewards/repeat_penalty/mean': 0.901418924331665, 'rewards/repeat_penalty/std': 0.12053182721138, 'rewards/repeat_floor/mean': 0.988824188709259, 'rewards/repeat_floor/std': 0.013290772214531898, 'rewards/near_duplicate_penalty/mean': 0.965886652469635, 'rewards/near_duplicate_penalty/std': 0.03909238800406456, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.941012978553772, 'rewards/distinct_2/std': 0.07635874301195145, 'rewards/total_composite/mean': 0.195618137717247, 'rewards/total_composite/std': 0.19238482415676117, 'reward': 0.195618137717247, 'reward_std': 0.19238482415676117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13922935724258423, 'sampling/sampling_logp_difference/max': 1.459977149963379, 'sampling/importance_sampling_ratio/min': 0.23224158585071564, 'sampling/importance_sampling_ratio/mean': 1.022443413734436, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5502031892538071, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08355611423030496, 'clip_ratio/high_max': 0.08355611423030496, 'clip_ratio/region_mean': 0.08355611423030496, 'reward_total_mean': 0.195618137717247, 'reward_meter_mean': 0.6363770961761475, 'reward_meter_std': 0.35401707887649536, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.05175493285059929, 'reward_lexical_plausibility_mean': 0.780329704284668, 'reward_lexical_plausibility_std': 0.23699282109737396, 'reward_repeat_penalty_mean': 0.901418924331665, 'reward_repeat_penalty_std': 0.12053182721138, 'reward_repeat_floor_mean': 0.988824188709259, 'reward_repeat_floor_std': 0.013290772214531898, 'reward_near_duplicate_penalty_mean': 0.965886652469635, 'reward_near_duplicate_penalty_std': 0.03909238800406456, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.941012978553772, 'reward_distinct_2_std': 0.07635874301195145, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.21380899846553802, 'reward_total_composite_mean': 0.195618137717247, 'reward_total_composite_std': 0.19238482415676117, 'epoch': 0.05}

 29%|██▊       | 941/3300 [3:31:18<9:11:36, 14.03s/it]INFO 04-12 19:29:44 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 942/3300 [3:31:33<9:21:41, 14.29s/it]2026-04-12 19:29:59,681 | INFO | train_grpo_train | metrics_logged mode=train step=942

                                                      
{'loss': -0.06, 'grad_norm': 4.63368558883667, 'learning_rate': 7.148484848484849e-06, 'num_tokens': 1672114.0, 'completions/mean_length': 173.25, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 60.333335876464844, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5495532751083374, 'rewards/meter/std': 0.3860201835632324, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.8999386429786682, 'rewards/lexical_plausibility/std': 0.18630100786685944, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8315969109535217, 'rewards/repeat_penalty/std': 0.1564721018075943, 'rewards/repeat_floor/mean': 0.9792554974555969, 'rewards/repeat_floor/std': 0.019118284806609154, 'rewards/near_duplicate_penalty/mean': 0.926944375038147, 'rewards/near_duplicate_penalty/std': 0.05962273105978966, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8912646770477295, 'rewards/distinct_2/std': 0.12045878916978836, 'rewards/total_composite/mean': 0.10068076848983765, 'rewards/total_composite/std': 0.09677880257368088, 'reward': 0.10068076848983765, 'reward_std': 0.09677879512310028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16810256242752075, 'sampling/sampling_logp_difference/max': 2.708258628845215, 'sampling/importance_sampling_ratio/min': 0.06665277481079102, 'sampling/importance_sampling_ratio/mean': 1.0213956832885742, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7796542420983315, 'clip_ratio/low_mean': 0.06385566294193268, 'clip_ratio/low_min': 0.06385566294193268, 'clip_ratio/high_mean': 0.037737186066806316, 'clip_ratio/high_max': 0.037737186066806316, 'clip_ratio/region_mean': 0.101592849008739, 'reward_total_mean': 0.10068076848983765, 'reward_meter_mean': 0.5495532751083374, 'reward_meter_std': 0.3860201835632324, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.8999386429786682, 'reward_lexical_plausibility_std': 0.18630100786685944, 'reward_repeat_penalty_mean': 0.8315969109535217, 'reward_repeat_penalty_std': 0.1564721018075943, 'reward_repeat_floor_mean': 0.9792554974555969, 'reward_repeat_floor_std': 0.019118284806609154, 'reward_near_duplicate_penalty_mean': 0.926944375038147, 'reward_near_duplicate_penalty_std': 0.05962273105978966, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8912646770477295, 'reward_distinct_2_std': 0.12045878916978836, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.10068076848983765, 'reward_total_composite_std': 0.09677880257368088, 'epoch': 0.05}

 29%|██▊       | 942/3300 [3:31:33<9:21:41, 14.29s/it]INFO 04-12 19:30:00 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 943/3300 [3:31:48<9:29:13, 14.49s/it]2026-04-12 19:30:14,588 | INFO | train_grpo_train | metrics_logged mode=train step=943

                                                      
{'loss': -0.0734, 'grad_norm': 1.6541037559509277, 'learning_rate': 7.145454545454547e-06, 'num_tokens': 1673680.0, 'completions/mean_length': 219.75, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 44.400001525878906, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5324200391769409, 'rewards/meter/std': 0.34049493074417114, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8362206816673279, 'rewards/lexical_plausibility/std': 0.23431777954101562, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.17512750625610352, 'rewards/repeat_penalty/mean': 0.8054771423339844, 'rewards/repeat_penalty/std': 0.16741535067558289, 'rewards/repeat_floor/mean': 0.9801266193389893, 'rewards/repeat_floor/std': 0.019427379593253136, 'rewards/near_duplicate_penalty/mean': 0.9438799619674683, 'rewards/near_duplicate_penalty/std': 0.0614447183907032, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9352185726165771, 'rewards/distinct_2/std': 0.09087558090686798, 'rewards/total_composite/mean': 0.14134326577186584, 'rewards/total_composite/std': 0.13784575462341309, 'reward': 0.14134326577186584, 'reward_std': 0.13784575462341309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16657239198684692, 'sampling/sampling_logp_difference/max': 1.5688648223876953, 'sampling/importance_sampling_ratio/min': 0.2082814872264862, 'sampling/importance_sampling_ratio/mean': 1.0320324897766113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8716687113046646, 'clip_ratio/low_mean': 0.02445652149617672, 'clip_ratio/low_min': 0.02445652149617672, 'clip_ratio/high_mean': 0.07158573903143406, 'clip_ratio/high_max': 0.07158573903143406, 'clip_ratio/region_mean': 0.09604226052761078, 'reward_total_mean': 0.14134326577186584, 'reward_meter_mean': 0.5324200391769409, 'reward_meter_std': 0.34049493074417114, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8362206816673279, 'reward_lexical_plausibility_std': 0.23431777954101562, 'reward_repeat_penalty_mean': 0.8054771423339844, 'reward_repeat_penalty_std': 0.16741535067558289, 'reward_repeat_floor_mean': 0.9801266193389893, 'reward_repeat_floor_std': 0.019427379593253136, 'reward_near_duplicate_penalty_mean': 0.9438799619674683, 'reward_near_duplicate_penalty_std': 0.0614447183907032, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9352185726165771, 'reward_distinct_2_std': 0.09087558090686798, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.17512750625610352, 'reward_total_composite_mean': 0.14134326577186584, 'reward_total_composite_std': 0.13784575462341309, 'epoch': 0.05}

 29%|██▊       | 943/3300 [3:31:48<9:29:13, 14.49s/it]INFO 04-12 19:30:15 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 944/3300 [3:31:57<8:28:07, 12.94s/it]2026-04-12 19:30:23,951 | INFO | train_grpo_train | metrics_logged mode=train step=944

                                                      
{'loss': 0.0123, 'grad_norm': 15.0864896774292, 'learning_rate': 7.142424242424243e-06, 'num_tokens': 1675394.0, 'completions/mean_length': 26.25, 'completions/min_length': 24.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.47471821308135986, 'rewards/meter/std': 0.3976638913154602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.2988907992839813, 'rewards/repeat_penalty/mean': 0.745326578617096, 'rewards/repeat_penalty/std': 0.1761200875043869, 'rewards/repeat_floor/mean': 0.9686992168426514, 'rewards/repeat_floor/std': 0.022518165409564972, 'rewards/near_duplicate_penalty/mean': 0.8890085220336914, 'rewards/near_duplicate_penalty/std': 0.06082123890519142, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8788655400276184, 'rewards/distinct_2/std': 0.09237746149301529, 'rewards/total_composite/mean': 0.24461592733860016, 'rewards/total_composite/std': 0.2185964286327362, 'reward': 0.24461592733860016, 'reward_std': 0.218596413731575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1147134006023407, 'sampling/sampling_logp_difference/max': 1.083981990814209, 'sampling/importance_sampling_ratio/min': 0.3382459282875061, 'sampling/importance_sampling_ratio/mean': 1.0304076671600342, 'sampling/importance_sampling_ratio/max': 1.9772380590438843, 'entropy': 0.6096719056367874, 'clip_ratio/low_mean': 0.04892857093364, 'clip_ratio/low_min': 0.04892857093364, 'clip_ratio/high_mean': 0.05568356392905116, 'clip_ratio/high_max': 0.05568356392905116, 'clip_ratio/region_mean': 0.10461213486269116, 'reward_total_mean': 0.24461592733860016, 'reward_meter_mean': 0.47471821308135986, 'reward_meter_std': 0.3976638913154602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.745326578617096, 'reward_repeat_penalty_std': 0.1761200875043869, 'reward_repeat_floor_mean': 0.9686992168426514, 'reward_repeat_floor_std': 0.022518165409564972, 'reward_near_duplicate_penalty_mean': 0.8890085220336914, 'reward_near_duplicate_penalty_std': 0.06082123890519142, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8788655400276184, 'reward_distinct_2_std': 0.09237746149301529, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.2988907992839813, 'reward_total_composite_mean': 0.24461592733860016, 'reward_total_composite_std': 0.2185964286327362, 'epoch': 0.05}

 29%|██▊       | 944/3300 [3:31:57<8:28:07, 12.94s/it]INFO 04-12 19:30:24 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 945/3300 [3:32:07<7:50:34, 11.99s/it]2026-04-12 19:30:33,669 | INFO | train_grpo_train | metrics_logged mode=train step=945

                                                      
{'loss': 0.0621, 'grad_norm': 12.124125480651855, 'learning_rate': 7.1393939393939405e-06, 'num_tokens': 1677534.0, 'completions/mean_length': 86.5, 'completions/min_length': 80.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.5, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.7432481646537781, 'rewards/meter/std': 0.28944098949432373, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.74357008934021, 'rewards/repeat_penalty/std': 0.17146047949790955, 'rewards/repeat_floor/mean': 0.9709051251411438, 'rewards/repeat_floor/std': 0.01959552988409996, 'rewards/near_duplicate_penalty/mean': 0.9234217405319214, 'rewards/near_duplicate_penalty/std': 0.04708775505423546, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.804354190826416, 'rewards/distinct_2/std': 0.14099173247814178, 'rewards/total_composite/mean': 0.29450181126594543, 'rewards/total_composite/std': 0.1559145450592041, 'reward': 0.29450181126594543, 'reward_std': 0.1559145450592041, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1454910784959793, 'sampling/sampling_logp_difference/max': 2.71665358543396, 'sampling/importance_sampling_ratio/min': 0.06609556823968887, 'sampling/importance_sampling_ratio/mean': 1.0035830736160278, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6880360022187233, 'clip_ratio/low_mean': 0.03264713007956743, 'clip_ratio/low_min': 0.03264713007956743, 'clip_ratio/high_mean': 0.08935249224305153, 'clip_ratio/high_max': 0.08935249224305153, 'clip_ratio/region_mean': 0.12199962232261896, 'reward_total_mean': 0.29450181126594543, 'reward_meter_mean': 0.7432481646537781, 'reward_meter_std': 0.28944098949432373, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.74357008934021, 'reward_repeat_penalty_std': 0.17146047949790955, 'reward_repeat_floor_mean': 0.9709051251411438, 'reward_repeat_floor_std': 0.01959552988409996, 'reward_near_duplicate_penalty_mean': 0.9234217405319214, 'reward_near_duplicate_penalty_std': 0.04708775505423546, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.804354190826416, 'reward_distinct_2_std': 0.14099173247814178, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.29450181126594543, 'reward_total_composite_std': 0.1559145450592041, 'epoch': 0.05}

 29%|██▊       | 945/3300 [3:32:07<7:50:34, 11.99s/it]INFO 04-12 19:30:33 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 946/3300 [3:32:16<7:13:17, 11.04s/it]2026-04-12 19:30:42,508 | INFO | train_grpo_train | metrics_logged mode=train step=946

                                                      
{'loss': -0.0179, 'grad_norm': 10.980562210083008, 'learning_rate': 7.136363636363637e-06, 'num_tokens': 1679246.0, 'completions/mean_length': 39.0, 'completions/min_length': 23.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.776240885257721, 'rewards/meter/std': 0.33104467391967773, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3366537094116211, 'rewards/repeat_penalty/mean': 0.7612777948379517, 'rewards/repeat_penalty/std': 0.19857537746429443, 'rewards/repeat_floor/mean': 0.9745221138000488, 'rewards/repeat_floor/std': 0.022732319310307503, 'rewards/near_duplicate_penalty/mean': 0.941365659236908, 'rewards/near_duplicate_penalty/std': 0.05276405066251755, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8563030958175659, 'rewards/distinct_2/std': 0.16082361340522766, 'rewards/total_composite/mean': 0.2662997841835022, 'rewards/total_composite/std': 0.25945690274238586, 'reward': 0.2662997841835022, 'reward_std': 0.25945690274238586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14707358181476593, 'sampling/sampling_logp_difference/max': 2.3706791400909424, 'sampling/importance_sampling_ratio/min': 0.09341725707054138, 'sampling/importance_sampling_ratio/mean': 1.0250524282455444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9912973940372467, 'clip_ratio/low_mean': 0.11727049294859171, 'clip_ratio/low_min': 0.11727049294859171, 'clip_ratio/high_mean': 0.0397529061883688, 'clip_ratio/high_max': 0.0397529061883688, 'clip_ratio/region_mean': 0.1570233991369605, 'reward_total_mean': 0.2662997841835022, 'reward_meter_mean': 0.776240885257721, 'reward_meter_std': 0.33104467391967773, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7612777948379517, 'reward_repeat_penalty_std': 0.19857537746429443, 'reward_repeat_floor_mean': 0.9745221138000488, 'reward_repeat_floor_std': 0.022732319310307503, 'reward_near_duplicate_penalty_mean': 0.941365659236908, 'reward_near_duplicate_penalty_std': 0.05276405066251755, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8563030958175659, 'reward_distinct_2_std': 0.16082361340522766, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3366537094116211, 'reward_total_composite_mean': 0.2662997841835022, 'reward_total_composite_std': 0.25945690274238586, 'epoch': 0.05}

 29%|██▊       | 946/3300 [3:32:16<7:13:17, 11.04s/it]INFO 04-12 19:30:42 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 947/3300 [3:32:25<6:55:56, 10.61s/it]2026-04-12 19:30:52,142 | INFO | train_grpo_train | metrics_logged mode=train step=947

                                                      
{'loss': 0.0416, 'grad_norm': 11.462865829467773, 'learning_rate': 7.133333333333334e-06, 'num_tokens': 1680896.0, 'completions/mean_length': 40.25, 'completions/min_length': 36.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.524586021900177, 'rewards/meter/std': 0.4112315773963928, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.8523638248443604, 'rewards/repeat_penalty/std': 0.1372341513633728, 'rewards/repeat_floor/mean': 0.9850330352783203, 'rewards/repeat_floor/std': 0.012364797294139862, 'rewards/near_duplicate_penalty/mean': 0.956728458404541, 'rewards/near_duplicate_penalty/std': 0.03769984096288681, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9474858045578003, 'rewards/distinct_2/std': 0.0465642586350441, 'rewards/total_composite/mean': 0.18610750138759613, 'rewards/total_composite/std': 0.18198786675930023, 'reward': 0.18610750138759613, 'reward_std': 0.18198786675930023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18754582107067108, 'sampling/sampling_logp_difference/max': 1.3779113292694092, 'sampling/importance_sampling_ratio/min': 0.25210458040237427, 'sampling/importance_sampling_ratio/mean': 1.0283889770507812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3841508403420448, 'clip_ratio/low_mean': 0.12969746626913548, 'clip_ratio/low_min': 0.12969746626913548, 'clip_ratio/high_mean': 0.07959495019167662, 'clip_ratio/high_max': 0.07959495019167662, 'clip_ratio/region_mean': 0.2092924164608121, 'reward_total_mean': 0.18610750138759613, 'reward_meter_mean': 0.524586021900177, 'reward_meter_std': 0.4112315773963928, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8523638248443604, 'reward_repeat_penalty_std': 0.1372341513633728, 'reward_repeat_floor_mean': 0.9850330352783203, 'reward_repeat_floor_std': 0.012364797294139862, 'reward_near_duplicate_penalty_mean': 0.956728458404541, 'reward_near_duplicate_penalty_std': 0.03769984096288681, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9474858045578003, 'reward_distinct_2_std': 0.0465642586350441, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.18610750138759613, 'reward_total_composite_std': 0.18198786675930023, 'epoch': 0.05}

 29%|██▊       | 947/3300 [3:32:25<6:55:56, 10.61s/it]INFO 04-12 19:30:52 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 948/3300 [3:32:36<7:00:18, 10.72s/it]2026-04-12 19:31:03,138 | INFO | train_grpo_train | metrics_logged mode=train step=948

                                                      
{'loss': 0.0791, 'grad_norm': 10.601950645446777, 'learning_rate': 7.130303030303031e-06, 'num_tokens': 1683210.0, 'completions/mean_length': 103.25, 'completions/min_length': 90.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.25, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.730411171913147, 'rewards/meter/std': 0.15826690196990967, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.020044606178998947, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.5941283702850342, 'rewards/repeat_penalty/std': 0.14179567992687225, 'rewards/repeat_floor/mean': 0.9539738297462463, 'rewards/repeat_floor/std': 0.01815166510641575, 'rewards/near_duplicate_penalty/mean': 0.8893653154373169, 'rewards/near_duplicate_penalty/std': 0.05428140237927437, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.6781972646713257, 'rewards/distinct_2/std': 0.1103530302643776, 'rewards/total_composite/mean': 0.22807839512825012, 'rewards/total_composite/std': 0.11001968383789062, 'reward': 0.22807839512825012, 'reward_std': 0.11001968383789062, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14287881553173065, 'sampling/sampling_logp_difference/max': 2.320979118347168, 'sampling/importance_sampling_ratio/min': 0.09817741066217422, 'sampling/importance_sampling_ratio/mean': 1.0042535066604614, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.743267148733139, 'clip_ratio/low_mean': 0.06109267380088568, 'clip_ratio/low_min': 0.06109267380088568, 'clip_ratio/high_mean': 0.06792892795056105, 'clip_ratio/high_max': 0.06792892795056105, 'clip_ratio/region_mean': 0.12902160175144672, 'reward_total_mean': 0.22807839512825012, 'reward_meter_mean': 0.730411171913147, 'reward_meter_std': 0.15826690196990967, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.020044606178998947, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5941283702850342, 'reward_repeat_penalty_std': 0.14179567992687225, 'reward_repeat_floor_mean': 0.9539738297462463, 'reward_repeat_floor_std': 0.01815166510641575, 'reward_near_duplicate_penalty_mean': 0.8893653154373169, 'reward_near_duplicate_penalty_std': 0.05428140237927437, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.6781972646713257, 'reward_distinct_2_std': 0.1103530302643776, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.22807839512825012, 'reward_total_composite_std': 0.11001968383789062, 'epoch': 0.05}

 29%|██▊       | 948/3300 [3:32:36<7:00:18, 10.72s/it]INFO 04-12 19:31:03 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 949/3300 [3:32:50<7:39:17, 11.72s/it]2026-04-12 19:31:17,138 | INFO | train_grpo_train | metrics_logged mode=train step=949

                                                      
{'loss': -0.0433, 'grad_norm': 2.3729870319366455, 'learning_rate': 7.127272727272728e-06, 'num_tokens': 1684658.0, 'completions/mean_length': 214.0, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 35.20000076293945, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.4223981499671936, 'rewards/meter/std': 0.3667108416557312, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8473978042602539, 'rewards/lexical_plausibility/std': 0.21464623510837555, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.21876277029514313, 'rewards/repeat_penalty/mean': 0.7196487784385681, 'rewards/repeat_penalty/std': 0.26018139719963074, 'rewards/repeat_floor/mean': 0.9653791189193726, 'rewards/repeat_floor/std': 0.035732757300138474, 'rewards/near_duplicate_penalty/mean': 0.8930052518844604, 'rewards/near_duplicate_penalty/std': 0.10914258658885956, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8561482429504395, 'rewards/distinct_2/std': 0.16621184349060059, 'rewards/total_composite/mean': 0.1590559184551239, 'rewards/total_composite/std': 0.21661260724067688, 'reward': 0.1590559184551239, 'reward_std': 0.21661259233951569, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11724242568016052, 'sampling/sampling_logp_difference/max': 1.46563720703125, 'sampling/importance_sampling_ratio/min': 0.23093080520629883, 'sampling/importance_sampling_ratio/mean': 1.0067551136016846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46586231887340546, 'clip_ratio/low_mean': 0.02888655522838235, 'clip_ratio/low_min': 0.02888655522838235, 'clip_ratio/high_mean': 0.027413126546889544, 'clip_ratio/high_max': 0.027413126546889544, 'clip_ratio/region_mean': 0.05629968177527189, 'reward_total_mean': 0.1590559184551239, 'reward_meter_mean': 0.4223981499671936, 'reward_meter_std': 0.3667108416557312, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8473978042602539, 'reward_lexical_plausibility_std': 0.21464623510837555, 'reward_repeat_penalty_mean': 0.7196487784385681, 'reward_repeat_penalty_std': 0.26018139719963074, 'reward_repeat_floor_mean': 0.9653791189193726, 'reward_repeat_floor_std': 0.035732757300138474, 'reward_near_duplicate_penalty_mean': 0.8930052518844604, 'reward_near_duplicate_penalty_std': 0.10914258658885956, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8561482429504395, 'reward_distinct_2_std': 0.16621184349060059, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.21876277029514313, 'reward_total_composite_mean': 0.1590559184551239, 'reward_total_composite_std': 0.21661260724067688, 'epoch': 0.05}

 29%|██▉       | 949/3300 [3:32:50<7:39:17, 11.72s/it]INFO 04-12 19:31:17 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 950/3300 [3:32:59<7:01:27, 10.76s/it]2026-04-12 19:31:25,658 | INFO | train_grpo_train | metrics_logged mode=train step=950

                                                      
{'loss': 0.0298, 'grad_norm': 14.033048629760742, 'learning_rate': 7.124242424242424e-06, 'num_tokens': 1686225.0, 'completions/mean_length': 39.875, 'completions/min_length': 36.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.5826082825660706, 'rewards/meter/std': 0.3956049382686615, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.755547285079956, 'rewards/repeat_penalty/std': 0.20465554296970367, 'rewards/repeat_floor/mean': 0.9698851108551025, 'rewards/repeat_floor/std': 0.027316758409142494, 'rewards/near_duplicate_penalty/mean': 0.8884701132774353, 'rewards/near_duplicate_penalty/std': 0.1049058735370636, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8929404020309448, 'rewards/distinct_2/std': 0.1175127923488617, 'rewards/total_composite/mean': 0.34095045924186707, 'rewards/total_composite/std': 0.30790162086486816, 'reward': 0.34095045924186707, 'reward_std': 0.30790165066719055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12967446446418762, 'sampling/sampling_logp_difference/max': 1.8172540664672852, 'sampling/importance_sampling_ratio/min': 0.16247127950191498, 'sampling/importance_sampling_ratio/mean': 1.0189299583435059, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8970945179462433, 'clip_ratio/low_mean': 0.072966150008142, 'clip_ratio/low_min': 0.072966150008142, 'clip_ratio/high_mean': 0.05367823038250208, 'clip_ratio/high_max': 0.05367823038250208, 'clip_ratio/region_mean': 0.12664438039064407, 'reward_total_mean': 0.34095045924186707, 'reward_meter_mean': 0.5826082825660706, 'reward_meter_std': 0.3956049382686615, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.755547285079956, 'reward_repeat_penalty_std': 0.20465554296970367, 'reward_repeat_floor_mean': 0.9698851108551025, 'reward_repeat_floor_std': 0.027316758409142494, 'reward_near_duplicate_penalty_mean': 0.8884701132774353, 'reward_near_duplicate_penalty_std': 0.1049058735370636, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8929404020309448, 'reward_distinct_2_std': 0.1175127923488617, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.34095045924186707, 'reward_total_composite_std': 0.30790162086486816, 'epoch': 0.05}

 29%|██▉       | 950/3300 [3:32:59<7:01:27, 10.76s/it]INFO 04-12 19:31:25 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:03,  5.79s/it]

 23%|██▎       | 3/13 [00:23<01:25,  8.54s/it]

 31%|███       | 4/13 [00:35<01:26,  9.64s/it]

 38%|███▊      | 5/13 [00:44<01:15,  9.45s/it]

 46%|████▌     | 6/13 [00:55<01:09,  9.99s/it]

 54%|█████▍    | 7/13 [01:07<01:03, 10.50s/it]

 62%|██████▏   | 8/13 [01:19<00:55, 11.13s/it]

 69%|██████▉   | 9/13 [01:31<00:45, 11.33s/it]

 77%|███████▋  | 10/13 [01:43<00:34, 11.47s/it]

 85%|████████▍ | 11/13 [01:51<00:21, 10.58s/it]

 92%|█████████▏| 12/13 [02:03<00:11, 11.03s/it]

100%|██████████| 13/13 [02:11<00:00, 10.14s/it]2026-04-12 19:33:46,037 | INFO | train_grpo_train | metrics_logged mode=eval step=950

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 140.2932, 'eval_samples_per_second': 0.741, 'eval_steps_per_second': 0.093, 'eval_num_tokens': 1686225.0, 'eval_completions/mean_length': 142.03846153846155, 'eval_completions/min_length': 28.307692307692307, 'eval_completions/max_length': 409.84615384615387, 'eval_completions/clipped_ratio': 0.09615384615384616, 'eval_completions/mean_terminated_length': 102.19551555926984, 'eval_completions/min_terminated_length': 28.307692307692307, 'eval_completions/max_terminated_length': 221.69230769230768, 'eval_rewards/meter/mean': 0.5910043991529025, 'eval_rewards/meter/std': 0.38212204667238087, 'eval_rewards/count_adherence/mean': 0.7742666693834158, 'eval_rewards/count_adherence/std': 0.22555504796596673, 'eval_rewards/arabic_clean/mean': 0.8846153846153846, 'eval_rewards/arabic_clean/std': 0.2657948434352875, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9322800040245056, 'eval_rewards/count_floor/std': 0.0676665109797166, 'eval_rewards/lexical_plausibility/mean': 0.9623410380803622, 'eval_rewards/lexical_plausibility/std': 0.08675260125444485, 'eval_rewards/judge_quality/mean': 0.27961538388178897, 'eval_rewards/judge_quality/std': 0.16308964559665093, 'eval_rewards/repeat_penalty/mean': 0.5226171566889837, 'eval_rewards/repeat_penalty/std': 0.320600233398951, 'eval_rewards/repeat_floor/mean': 0.9376098742851844, 'eval_rewards/repeat_floor/std': 0.05426772254017683, 'eval_rewards/near_duplicate_penalty/mean': 0.8426430317071768, 'eval_rewards/near_duplicate_penalty/std': 0.15295882179186895, 'eval_rewards/opening_diversity/mean': 0.9035141651447003, 'eval_rewards/opening_diversity/std': 0.1447479403935946, 'eval_rewards/distinct_2/mean': 0.6333618897658128, 'eval_rewards/distinct_2/std': 0.340181895173513, 'eval_rewards/total_composite/mean': 0.14772453502966806, 'eval_rewards/total_composite/std': 0.1381625120456402, 'eval_reward': 0.14772453502966806, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06588262348220898, 'eval_sampling/sampling_logp_difference/max': 1.0992378821739783, 'eval_sampling/importance_sampling_ratio/min': 0.340822676053414, 'eval_sampling/importance_sampling_ratio/mean': 1.015147511775677, 'eval_sampling/importance_sampling_ratio/max': 1.544888395529527, 'eval_entropy': 0.6938141309298002, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.14772453502966806, 'eval_reward_meter_mean': 0.5910043991529025, 'eval_reward_meter_std': 0.38212204667238087, 'eval_reward_count_adherence_mean': 0.7742666693834158, 'eval_reward_count_adherence_std': 0.22555504796596673, 'eval_reward_arabic_clean_mean': 0.8846153846153846, 'eval_reward_arabic_clean_std': 0.2657948434352875, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9322800040245056, 'eval_reward_count_floor_std': 0.0676665109797166, 'eval_reward_lexical_plausibility_mean': 0.9623410380803622, 'eval_reward_lexical_plausibility_std': 0.08675260125444485, 'eval_reward_repeat_penalty_mean': 0.5226171566889837, 'eval_reward_repeat_penalty_std': 0.320600233398951, 'eval_reward_repeat_floor_mean': 0.9376098742851844, 'eval_reward_repeat_floor_std': 0.05426772254017683, 'eval_reward_near_duplicate_penalty_mean': 0.8426430317071768, 'eval_reward_near_duplicate_penalty_std': 0.15295882179186895, 'eval_reward_opening_diversity_mean': 0.9035141651447003, 'eval_reward_opening_diversity_std': 0.1447479403935946, 'eval_reward_distinct_2_mean': 0.6333618897658128, 'eval_reward_distinct_2_std': 0.340181895173513, 'eval_reward_judge_quality_mean': 0.27961538388178897, 'eval_reward_judge_quality_std': 0.16308964559665093, 'eval_reward_total_composite_mean': 0.14772453502966806, 'eval_reward_total_composite_std': 0.1381625120456402, 'epoch': 0.05}

 29%|██▉       | 950/3300 [3:35:19<7:01:27, 10.76s/it]

100%|██████████| 13/13 [02:12<00:00, 10.14s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 29%|██▉       | 951/3300 [3:35:34<35:21:09, 54.18s/it]2026-04-12 19:34:01,153 | INFO | train_grpo_train | metrics_logged mode=train step=951

                                                       
{'loss': 0.0189, 'grad_norm': 10.40640640258789, 'learning_rate': 7.121212121212122e-06, 'num_tokens': 1688968.0, 'completions/mean_length': 127.875, 'completions/min_length': 110.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.875, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.6116044521331787, 'rewards/meter/std': 0.21812674403190613, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.020044606178998947, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.6276565194129944, 'rewards/repeat_penalty/std': 0.1394803524017334, 'rewards/repeat_floor/mean': 0.9618325233459473, 'rewards/repeat_floor/std': 0.012549640610814095, 'rewards/near_duplicate_penalty/mean': 0.9100544452667236, 'rewards/near_duplicate_penalty/std': 0.03262824937701225, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7258261442184448, 'rewards/distinct_2/std': 0.11192677170038223, 'rewards/total_composite/mean': 0.19275856018066406, 'rewards/total_composite/std': 0.08759541064500809, 'reward': 0.19275856018066406, 'reward_std': 0.08759541064500809, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16588589549064636, 'sampling/sampling_logp_difference/max': 3.5576331615448, 'sampling/importance_sampling_ratio/min': 0.028506215661764145, 'sampling/importance_sampling_ratio/mean': 1.0032833814620972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7086888179183006, 'clip_ratio/low_mean': 0.06634479109197855, 'clip_ratio/low_min': 0.06634479109197855, 'clip_ratio/high_mean': 0.06301605701446533, 'clip_ratio/high_max': 0.06301605701446533, 'clip_ratio/region_mean': 0.12936084810644388, 'reward_total_mean': 0.19275856018066406, 'reward_meter_mean': 0.6116044521331787, 'reward_meter_std': 0.21812674403190613, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.020044606178998947, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6276565194129944, 'reward_repeat_penalty_std': 0.1394803524017334, 'reward_repeat_floor_mean': 0.9618325233459473, 'reward_repeat_floor_std': 0.012549640610814095, 'reward_near_duplicate_penalty_mean': 0.9100544452667236, 'reward_near_duplicate_penalty_std': 0.03262824937701225, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7258261442184448, 'reward_distinct_2_std': 0.11192677170038223, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.19275856018066406, 'reward_total_composite_std': 0.08759541064500809, 'epoch': 0.05}

 29%|██▉       | 951/3300 [3:35:35<35:21:09, 54.18s/it]INFO 04-12 19:34:01 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 952/3300 [3:35:45<26:42:27, 40.95s/it]2026-04-12 19:34:11,231 | INFO | train_grpo_train | metrics_logged mode=train step=952

                                                       
{'loss': 0.083, 'grad_norm': 11.282236099243164, 'learning_rate': 7.118181818181819e-06, 'num_tokens': 1691552.0, 'completions/mean_length': 98.0, 'completions/min_length': 21.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.8767145872116089, 'rewards/meter/std': 0.20181696116924286, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.3234066069126129, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.09702197462320328, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.5589762926101685, 'rewards/repeat_penalty/std': 0.21816739439964294, 'rewards/repeat_floor/mean': 0.9461507797241211, 'rewards/repeat_floor/std': 0.033819179981946945, 'rewards/near_duplicate_penalty/mean': 0.8495292067527771, 'rewards/near_duplicate_penalty/std': 0.10360148549079895, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.6941260099411011, 'rewards/distinct_2/std': 0.21561336517333984, 'rewards/total_composite/mean': 0.22477945685386658, 'rewards/total_composite/std': 0.07185804843902588, 'reward': 0.22477945685386658, 'reward_std': 0.07185804843902588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1414160281419754, 'sampling/sampling_logp_difference/max': 1.7998636960983276, 'sampling/importance_sampling_ratio/min': 0.16532142460346222, 'sampling/importance_sampling_ratio/mean': 1.0004907846450806, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8585029914975166, 'clip_ratio/low_mean': 0.06142281927168369, 'clip_ratio/low_min': 0.06142281927168369, 'clip_ratio/high_mean': 0.07327659148722887, 'clip_ratio/high_max': 0.07327659148722887, 'clip_ratio/region_mean': 0.13469941075891256, 'reward_total_mean': 0.22477945685386658, 'reward_meter_mean': 0.8767145872116089, 'reward_meter_std': 0.20181696116924286, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.3234066069126129, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.09702197462320328, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5589762926101685, 'reward_repeat_penalty_std': 0.21816739439964294, 'reward_repeat_floor_mean': 0.9461507797241211, 'reward_repeat_floor_std': 0.033819179981946945, 'reward_near_duplicate_penalty_mean': 0.8495292067527771, 'reward_near_duplicate_penalty_std': 0.10360148549079895, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.6941260099411011, 'reward_distinct_2_std': 0.21561336517333984, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.22477945685386658, 'reward_total_composite_std': 0.07185804843902588, 'epoch': 0.05}

 29%|██▉       | 952/3300 [3:35:45<26:42:27, 40.95s/it]INFO 04-12 19:34:11 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 953/3300 [3:35:52<20:14:00, 31.04s/it]2026-04-12 19:34:19,134 | INFO | train_grpo_train | metrics_logged mode=train step=953

                                                       
{'loss': 0.0997, 'grad_norm': 18.836973190307617, 'learning_rate': 7.115151515151516e-06, 'num_tokens': 1692917.0, 'completions/mean_length': 20.625, 'completions/min_length': 20.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.5763106346130371, 'rewards/meter/std': 0.42897507548332214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3586059808731079, 'rewards/repeat_penalty/mean': 0.6704980134963989, 'rewards/repeat_penalty/std': 0.14287002384662628, 'rewards/repeat_floor/mean': 0.9698734283447266, 'rewards/repeat_floor/std': 0.027681956067681313, 'rewards/near_duplicate_penalty/mean': 0.9247974157333374, 'rewards/near_duplicate_penalty/std': 0.14086396992206573, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9572649598121643, 'rewards/distinct_2/std': 0.07912991940975189, 'rewards/total_composite/mean': 0.35372188687324524, 'rewards/total_composite/std': 0.3583594858646393, 'reward': 0.35372188687324524, 'reward_std': 0.3583594560623169, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16290335357189178, 'sampling/sampling_logp_difference/max': 3.2261862754821777, 'sampling/importance_sampling_ratio/min': 0.03970864787697792, 'sampling/importance_sampling_ratio/mean': 0.9987561702728271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8055507615208626, 'clip_ratio/low_mean': 0.03497670870274305, 'clip_ratio/low_min': 0.03497670870274305, 'clip_ratio/high_mean': 0.07500000204890966, 'clip_ratio/high_max': 0.07500000204890966, 'clip_ratio/region_mean': 0.10997671075165272, 'reward_total_mean': 0.35372188687324524, 'reward_meter_mean': 0.5763106346130371, 'reward_meter_std': 0.42897507548332214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6704980134963989, 'reward_repeat_penalty_std': 0.14287002384662628, 'reward_repeat_floor_mean': 0.9698734283447266, 'reward_repeat_floor_std': 0.027681956067681313, 'reward_near_duplicate_penalty_mean': 0.9247974157333374, 'reward_near_duplicate_penalty_std': 0.14086396992206573, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9572649598121643, 'reward_distinct_2_std': 0.07912991940975189, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3586059808731079, 'reward_total_composite_mean': 0.35372188687324524, 'reward_total_composite_std': 0.3583594858646393, 'epoch': 0.05}

 29%|██▉       | 953/3300 [3:35:52<20:14:00, 31.04s/it]INFO 04-12 19:34:19 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 954/3300 [3:36:01<15:56:03, 24.45s/it]2026-04-12 19:34:28,218 | INFO | train_grpo_train | metrics_logged mode=train step=954

                                                       
{'loss': 0.0262, 'grad_norm': 14.238551139831543, 'learning_rate': 7.1121212121212125e-06, 'num_tokens': 1694827.0, 'completions/mean_length': 63.75, 'completions/min_length': 39.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.44905251264572144, 'rewards/meter/std': 0.30701690912246704, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.6292611956596375, 'rewards/repeat_penalty/std': 0.2200477570295334, 'rewards/repeat_floor/mean': 0.9567912817001343, 'rewards/repeat_floor/std': 0.027628019452095032, 'rewards/near_duplicate_penalty/mean': 0.8850417137145996, 'rewards/near_duplicate_penalty/std': 0.09129218757152557, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.7531673908233643, 'rewards/distinct_2/std': 0.1234266459941864, 'rewards/total_composite/mean': 0.140194833278656, 'rewards/total_composite/std': 0.09795386344194412, 'reward': 0.140194833278656, 'reward_std': 0.09795386344194412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13733336329460144, 'sampling/sampling_logp_difference/max': 1.8037546873092651, 'sampling/importance_sampling_ratio/min': 0.1646794080734253, 'sampling/importance_sampling_ratio/mean': 1.0126349925994873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8165108859539032, 'clip_ratio/low_mean': 0.047445680014789104, 'clip_ratio/low_min': 0.047445680014789104, 'clip_ratio/high_mean': 0.06441685557365417, 'clip_ratio/high_max': 0.06441685557365417, 'clip_ratio/region_mean': 0.11186253558844328, 'reward_total_mean': 0.140194833278656, 'reward_meter_mean': 0.44905251264572144, 'reward_meter_std': 0.30701690912246704, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.6292611956596375, 'reward_repeat_penalty_std': 0.2200477570295334, 'reward_repeat_floor_mean': 0.9567912817001343, 'reward_repeat_floor_std': 0.027628019452095032, 'reward_near_duplicate_penalty_mean': 0.8850417137145996, 'reward_near_duplicate_penalty_std': 0.09129218757152557, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.7531673908233643, 'reward_distinct_2_std': 0.1234266459941864, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.140194833278656, 'reward_total_composite_std': 0.09795386344194412, 'epoch': 0.05}

 29%|██▉       | 954/3300 [3:36:02<15:56:03, 24.45s/it]INFO 04-12 19:34:28 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 955/3300 [3:36:13<13:19:23, 20.45s/it]2026-04-12 19:34:39,345 | INFO | train_grpo_train | metrics_logged mode=train step=955

                                                       
{'loss': 0.1342, 'grad_norm': 9.032503128051758, 'learning_rate': 7.10909090909091e-06, 'num_tokens': 1697331.0, 'completions/mean_length': 112.0, 'completions/min_length': 19.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.8889793753623962, 'rewards/meter/std': 0.27543914318084717, 'rewards/count_adherence/mean': 0.6964285373687744, 'rewards/count_adherence/std': 0.23458294570446014, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9089285135269165, 'rewards/count_floor/std': 0.07037486881017685, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.4074024558067322, 'rewards/repeat_penalty/std': 0.2968716323375702, 'rewards/repeat_floor/mean': 0.9225919246673584, 'rewards/repeat_floor/std': 0.05010543391108513, 'rewards/near_duplicate_penalty/mean': 0.8128306269645691, 'rewards/near_duplicate_penalty/std': 0.1494600623846054, 'rewards/opening_diversity/mean': 0.8491476774215698, 'rewards/opening_diversity/std': 0.1471789926290512, 'rewards/distinct_2/mean': 0.5524275302886963, 'rewards/distinct_2/std': 0.2812095880508423, 'rewards/total_composite/mean': 0.2147814929485321, 'rewards/total_composite/std': 0.09185987710952759, 'reward': 0.2147814929485321, 'reward_std': 0.09185986965894699, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12279205024242401, 'sampling/sampling_logp_difference/max': 3.0420215129852295, 'sampling/importance_sampling_ratio/min': 0.04773828759789467, 'sampling/importance_sampling_ratio/mean': 1.0091530084609985, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7275544628500938, 'clip_ratio/low_mean': 0.04513888945803046, 'clip_ratio/low_min': 0.04513888945803046, 'clip_ratio/high_mean': 0.0845271460711956, 'clip_ratio/high_max': 0.0845271460711956, 'clip_ratio/region_mean': 0.12966603552922606, 'reward_total_mean': 0.2147814929485321, 'reward_meter_mean': 0.8889793753623962, 'reward_meter_std': 0.27543914318084717, 'reward_count_adherence_mean': 0.6964285373687744, 'reward_count_adherence_std': 0.23458294570446014, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9089285135269165, 'reward_count_floor_std': 0.07037486881017685, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.4074024558067322, 'reward_repeat_penalty_std': 0.2968716323375702, 'reward_repeat_floor_mean': 0.9225919246673584, 'reward_repeat_floor_std': 0.05010543391108513, 'reward_near_duplicate_penalty_mean': 0.8128306269645691, 'reward_near_duplicate_penalty_std': 0.1494600623846054, 'reward_opening_diversity_mean': 0.8491476774215698, 'reward_opening_diversity_std': 0.1471789926290512, 'reward_distinct_2_mean': 0.5524275302886963, 'reward_distinct_2_std': 0.2812095880508423, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.2147814929485321, 'reward_total_composite_std': 0.09185987710952759, 'epoch': 0.05}

 29%|██▉       | 955/3300 [3:36:13<13:19:23, 20.45s/it]INFO 04-12 19:34:39 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 956/3300 [3:36:27<12:08:32, 18.65s/it]2026-04-12 19:34:53,785 | INFO | train_grpo_train | metrics_logged mode=train step=956

                                                       
{'loss': -0.192, 'grad_norm': 5.031750202178955, 'learning_rate': 7.106060606060606e-06, 'num_tokens': 1699358.0, 'completions/mean_length': 101.375, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.71428680419922, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9130992293357849, 'rewards/meter/std': 0.14417938888072968, 'rewards/count_adherence/mean': 0.3392857313156128, 'rewards/count_adherence/std': 0.3234066069126129, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8017857074737549, 'rewards/count_floor/std': 0.09702196717262268, 'rewards/lexical_plausibility/mean': 0.9637529850006104, 'rewards/lexical_plausibility/std': 0.10252200067043304, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.7165063619613647, 'rewards/repeat_penalty/std': 0.07188376039266586, 'rewards/repeat_floor/mean': 0.9738814234733582, 'rewards/repeat_floor/std': 0.011144919320940971, 'rewards/near_duplicate_penalty/mean': 0.9363481998443604, 'rewards/near_duplicate_penalty/std': 0.06483633071184158, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9200469255447388, 'rewards/distinct_2/std': 0.10844539850950241, 'rewards/total_composite/mean': 0.1309385895729065, 'rewards/total_composite/std': 0.07075480371713638, 'reward': 0.1309385895729065, 'reward_std': 0.07075481116771698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14706915616989136, 'sampling/sampling_logp_difference/max': 1.3654050827026367, 'sampling/importance_sampling_ratio/min': 0.25527724623680115, 'sampling/importance_sampling_ratio/mean': 1.0193216800689697, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9359106048941612, 'clip_ratio/low_mean': 0.11711270874366164, 'clip_ratio/low_min': 0.11711270874366164, 'clip_ratio/high_mean': 0.02961753774434328, 'clip_ratio/high_max': 0.02961753774434328, 'clip_ratio/region_mean': 0.14673024648800492, 'reward_total_mean': 0.1309385895729065, 'reward_meter_mean': 0.9130992293357849, 'reward_meter_std': 0.14417938888072968, 'reward_count_adherence_mean': 0.3392857313156128, 'reward_count_adherence_std': 0.3234066069126129, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8017857074737549, 'reward_count_floor_std': 0.09702196717262268, 'reward_lexical_plausibility_mean': 0.9637529850006104, 'reward_lexical_plausibility_std': 0.10252200067043304, 'reward_repeat_penalty_mean': 0.7165063619613647, 'reward_repeat_penalty_std': 0.07188376039266586, 'reward_repeat_floor_mean': 0.9738814234733582, 'reward_repeat_floor_std': 0.011144919320940971, 'reward_near_duplicate_penalty_mean': 0.9363481998443604, 'reward_near_duplicate_penalty_std': 0.06483633071184158, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9200469255447388, 'reward_distinct_2_std': 0.10844539850950241, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.1309385895729065, 'reward_total_composite_std': 0.07075480371713638, 'epoch': 0.05}

 29%|██▉       | 956/3300 [3:36:27<12:08:32, 18.65s/it]INFO 04-12 19:34:54 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 957/3300 [3:36:37<10:21:15, 15.91s/it]2026-04-12 19:35:03,302 | INFO | train_grpo_train | metrics_logged mode=train step=957

                                                       
{'loss': 0.1632, 'grad_norm': 11.119133949279785, 'learning_rate': 7.103030303030304e-06, 'num_tokens': 1701265.0, 'completions/mean_length': 54.375, 'completions/min_length': 34.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.3305176794528961, 'rewards/meter/std': 0.3714480698108673, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9886363744735718, 'rewards/lexical_plausibility/std': 0.03214120864868164, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.13887302577495575, 'rewards/repeat_penalty/mean': 0.8939952850341797, 'rewards/repeat_penalty/std': 0.1726941168308258, 'rewards/repeat_floor/mean': 0.9871750473976135, 'rewards/repeat_floor/std': 0.02102658897638321, 'rewards/near_duplicate_penalty/mean': 0.959881067276001, 'rewards/near_duplicate_penalty/std': 0.05803714320063591, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9243653416633606, 'rewards/distinct_2/std': 0.1401679813861847, 'rewards/total_composite/mean': 0.1598445475101471, 'rewards/total_composite/std': 0.22067265212535858, 'reward': 0.1598445475101471, 'reward_std': 0.2206726223230362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1448253095149994, 'sampling/sampling_logp_difference/max': 1.810577392578125, 'sampling/importance_sampling_ratio/min': 0.1635596752166748, 'sampling/importance_sampling_ratio/mean': 1.0112669467926025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7519679144024849, 'clip_ratio/low_mean': 0.06484001688659191, 'clip_ratio/low_min': 0.06484001688659191, 'clip_ratio/high_mean': 0.05810453463345766, 'clip_ratio/high_max': 0.05810453463345766, 'clip_ratio/region_mean': 0.12294455152004957, 'reward_total_mean': 0.1598445475101471, 'reward_meter_mean': 0.3305176794528961, 'reward_meter_std': 0.3714480698108673, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9886363744735718, 'reward_lexical_plausibility_std': 0.03214120864868164, 'reward_repeat_penalty_mean': 0.8939952850341797, 'reward_repeat_penalty_std': 0.1726941168308258, 'reward_repeat_floor_mean': 0.9871750473976135, 'reward_repeat_floor_std': 0.02102658897638321, 'reward_near_duplicate_penalty_mean': 0.959881067276001, 'reward_near_duplicate_penalty_std': 0.05803714320063591, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9243653416633606, 'reward_distinct_2_std': 0.1401679813861847, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.13887302577495575, 'reward_total_composite_mean': 0.1598445475101471, 'reward_total_composite_std': 0.22067265212535858, 'epoch': 0.05}

 29%|██▉       | 957/3300 [3:36:37<10:21:15, 15.91s/it]INFO 04-12 19:35:03 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 958/3300 [3:36:51<10:00:34, 15.39s/it]2026-04-12 19:35:17,467 | INFO | train_grpo_train | metrics_logged mode=train step=958

                                                       
{'loss': 0.1987, 'grad_norm': 3.029975175857544, 'learning_rate': 7.100000000000001e-06, 'num_tokens': 1703211.0, 'completions/mean_length': 143.25, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 90.5714340209961, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 406.0, 'rewards/meter/mean': 0.7041539549827576, 'rewards/meter/std': 0.31757858395576477, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9418036937713623, 'rewards/lexical_plausibility/std': 0.16460399329662323, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.2773052752017975, 'rewards/repeat_penalty/mean': 0.656030535697937, 'rewards/repeat_penalty/std': 0.3646494150161743, 'rewards/repeat_floor/mean': 0.9413734674453735, 'rewards/repeat_floor/std': 0.0894436240196228, 'rewards/near_duplicate_penalty/mean': 0.8343397378921509, 'rewards/near_duplicate_penalty/std': 0.2797500789165497, 'rewards/opening_diversity/mean': 0.7979166507720947, 'rewards/opening_diversity/std': 0.29566141963005066, 'rewards/distinct_2/mean': 0.7594171762466431, 'rewards/distinct_2/std': 0.33755066990852356, 'rewards/total_composite/mean': 0.25579366087913513, 'rewards/total_composite/std': 0.27933651208877563, 'reward': 0.25579366087913513, 'reward_std': 0.27933651208877563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.056202176958322525, 'sampling/sampling_logp_difference/max': 1.220952033996582, 'sampling/importance_sampling_ratio/min': 0.2949492335319519, 'sampling/importance_sampling_ratio/mean': 1.0123591423034668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6898953504860401, 'clip_ratio/low_mean': 0.04925617319531739, 'clip_ratio/low_min': 0.04925617319531739, 'clip_ratio/high_mean': 0.04855974344536662, 'clip_ratio/high_max': 0.04855974344536662, 'clip_ratio/region_mean': 0.09781591664068401, 'reward_total_mean': 0.25579366087913513, 'reward_meter_mean': 0.7041539549827576, 'reward_meter_std': 0.31757858395576477, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9418036937713623, 'reward_lexical_plausibility_std': 0.16460399329662323, 'reward_repeat_penalty_mean': 0.656030535697937, 'reward_repeat_penalty_std': 0.3646494150161743, 'reward_repeat_floor_mean': 0.9413734674453735, 'reward_repeat_floor_std': 0.0894436240196228, 'reward_near_duplicate_penalty_mean': 0.8343397378921509, 'reward_near_duplicate_penalty_std': 0.2797500789165497, 'reward_opening_diversity_mean': 0.7979166507720947, 'reward_opening_diversity_std': 0.29566141963005066, 'reward_distinct_2_mean': 0.7594171762466431, 'reward_distinct_2_std': 0.33755066990852356, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.2773052752017975, 'reward_total_composite_mean': 0.25579366087913513, 'reward_total_composite_std': 0.27933651208877563, 'epoch': 0.05}

 29%|██▉       | 958/3300 [3:36:51<10:00:34, 15.39s/it]INFO 04-12 19:35:17 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 959/3300 [3:37:04<9:32:16, 14.67s/it] 2026-04-12 19:35:30,453 | INFO | train_grpo_train | metrics_logged mode=train step=959

                                                      
{'loss': -0.031, 'grad_norm': 3.7746758460998535, 'learning_rate': 7.096969696969698e-06, 'num_tokens': 1704660.0, 'completions/mean_length': 91.125, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.000001907348633, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8713250756263733, 'rewards/meter/std': 0.3392218351364136, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9499614238739014, 'rewards/lexical_plausibility/std': 0.14153039455413818, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.3301055133342743, 'rewards/repeat_penalty/mean': 0.7269412279129028, 'rewards/repeat_penalty/std': 0.14046072959899902, 'rewards/repeat_floor/mean': 0.9756722450256348, 'rewards/repeat_floor/std': 0.020906880497932434, 'rewards/near_duplicate_penalty/mean': 0.941906750202179, 'rewards/near_duplicate_penalty/std': 0.10029266774654388, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9515135288238525, 'rewards/distinct_2/std': 0.0954609140753746, 'rewards/total_composite/mean': 0.371695339679718, 'rewards/total_composite/std': 0.27089422941207886, 'reward': 0.371695339679718, 'reward_std': 0.27089422941207886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12153531610965729, 'sampling/sampling_logp_difference/max': 1.0987062454223633, 'sampling/importance_sampling_ratio/min': 0.4633790850639343, 'sampling/importance_sampling_ratio/mean': 1.0648107528686523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9690554961562157, 'clip_ratio/low_mean': 0.06534963380545378, 'clip_ratio/low_min': 0.06534963380545378, 'clip_ratio/high_mean': 0.04171892395243049, 'clip_ratio/high_max': 0.04171892395243049, 'clip_ratio/region_mean': 0.10706855775788426, 'reward_total_mean': 0.371695339679718, 'reward_meter_mean': 0.8713250756263733, 'reward_meter_std': 0.3392218351364136, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9499614238739014, 'reward_lexical_plausibility_std': 0.14153039455413818, 'reward_repeat_penalty_mean': 0.7269412279129028, 'reward_repeat_penalty_std': 0.14046072959899902, 'reward_repeat_floor_mean': 0.9756722450256348, 'reward_repeat_floor_std': 0.020906880497932434, 'reward_near_duplicate_penalty_mean': 0.941906750202179, 'reward_near_duplicate_penalty_std': 0.10029266774654388, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9515135288238525, 'reward_distinct_2_std': 0.0954609140753746, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.3301055133342743, 'reward_total_composite_mean': 0.371695339679718, 'reward_total_composite_std': 0.27089422941207886, 'epoch': 0.05}

 29%|██▉       | 959/3300 [3:37:04<9:32:16, 14.67s/it]INFO 04-12 19:35:30 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 960/3300 [3:37:17<9:20:52, 14.38s/it]2026-04-12 19:35:44,172 | INFO | train_grpo_train | metrics_logged mode=train step=960

                                                      
{'loss': -0.0403, 'grad_norm': 2.2011475563049316, 'learning_rate': 7.093939393939394e-06, 'num_tokens': 1706178.0, 'completions/mean_length': 215.75, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.4090723395347595, 'rewards/meter/std': 0.3334019184112549, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8459243774414062, 'rewards/lexical_plausibility/std': 0.2214832901954651, 'rewards/judge_quality/mean': 0.2900000214576721, 'rewards/judge_quality/std': 0.304396390914917, 'rewards/repeat_penalty/mean': 0.9365726709365845, 'rewards/repeat_penalty/std': 0.08800051361322403, 'rewards/repeat_floor/mean': 0.99422287940979, 'rewards/repeat_floor/std': 0.006382799707353115, 'rewards/near_duplicate_penalty/mean': 0.9842880368232727, 'rewards/near_duplicate_penalty/std': 0.01846707984805107, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9828296899795532, 'rewards/distinct_2/std': 0.03188734129071236, 'rewards/total_composite/mean': 0.12301649153232574, 'rewards/total_composite/std': 0.1511751115322113, 'reward': 0.12301649153232574, 'reward_std': 0.1511751264333725, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1577311009168625, 'sampling/sampling_logp_difference/max': 1.4751026630401611, 'sampling/importance_sampling_ratio/min': 0.2548242211341858, 'sampling/importance_sampling_ratio/mean': 1.0386005640029907, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7927848324179649, 'clip_ratio/low_mean': 0.01666666753590107, 'clip_ratio/low_min': 0.01666666753590107, 'clip_ratio/high_mean': 0.06475761369802058, 'clip_ratio/high_max': 0.06475761369802058, 'clip_ratio/region_mean': 0.08142428123392165, 'reward_total_mean': 0.12301649153232574, 'reward_meter_mean': 0.4090723395347595, 'reward_meter_std': 0.3334019184112549, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8459243774414062, 'reward_lexical_plausibility_std': 0.2214832901954651, 'reward_repeat_penalty_mean': 0.9365726709365845, 'reward_repeat_penalty_std': 0.08800051361322403, 'reward_repeat_floor_mean': 0.99422287940979, 'reward_repeat_floor_std': 0.006382799707353115, 'reward_near_duplicate_penalty_mean': 0.9842880368232727, 'reward_near_duplicate_penalty_std': 0.01846707984805107, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9828296899795532, 'reward_distinct_2_std': 0.03188734129071236, 'reward_judge_quality_mean': 0.2900000214576721, 'reward_judge_quality_std': 0.304396390914917, 'reward_total_composite_mean': 0.12301649153232574, 'reward_total_composite_std': 0.1511751115322113, 'epoch': 0.05}

 29%|██▉       | 960/3300 [3:37:18<9:20:52, 14.38s/it]INFO 04-12 19:35:44 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 961/3300 [3:37:27<8:18:35, 12.79s/it]2026-04-12 19:35:53,248 | INFO | train_grpo_train | metrics_logged mode=train step=961

                                                      
{'loss': 0.0526, 'grad_norm': 18.03604507446289, 'learning_rate': 7.0909090909090916e-06, 'num_tokens': 1707653.0, 'completions/mean_length': 27.375, 'completions/min_length': 18.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9818100929260254, 'rewards/meter/std': 0.011301834136247635, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6675000190734863, 'rewards/judge_quality/std': 0.26162952184677124, 'rewards/repeat_penalty/mean': 0.7778453826904297, 'rewards/repeat_penalty/std': 0.09889805316925049, 'rewards/repeat_floor/mean': 0.9818924069404602, 'rewards/repeat_floor/std': 0.010265972465276718, 'rewards/near_duplicate_penalty/mean': 0.9638105630874634, 'rewards/near_duplicate_penalty/std': 0.052360448986291885, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9632866978645325, 'rewards/distinct_2/std': 0.06808248907327652, 'rewards/total_composite/mean': 0.5793253779411316, 'rewards/total_composite/std': 0.22016698122024536, 'reward': 0.5793253779411316, 'reward_std': 0.22016699612140656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14183680713176727, 'sampling/sampling_logp_difference/max': 6.309340476989746, 'sampling/importance_sampling_ratio/min': 0.0018192327115684748, 'sampling/importance_sampling_ratio/mean': 1.0047463178634644, 'sampling/importance_sampling_ratio/max': 1.5898181200027466, 'entropy': 0.7553734555840492, 'clip_ratio/low_mean': 0.06564555549994111, 'clip_ratio/low_min': 0.06564555549994111, 'clip_ratio/high_mean': 0.06134962756186724, 'clip_ratio/high_max': 0.06134962756186724, 'clip_ratio/region_mean': 0.12699518306180835, 'reward_total_mean': 0.5793253779411316, 'reward_meter_mean': 0.9818100929260254, 'reward_meter_std': 0.011301834136247635, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7778453826904297, 'reward_repeat_penalty_std': 0.09889805316925049, 'reward_repeat_floor_mean': 0.9818924069404602, 'reward_repeat_floor_std': 0.010265972465276718, 'reward_near_duplicate_penalty_mean': 0.9638105630874634, 'reward_near_duplicate_penalty_std': 0.052360448986291885, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9632866978645325, 'reward_distinct_2_std': 0.06808248907327652, 'reward_judge_quality_mean': 0.6675000190734863, 'reward_judge_quality_std': 0.26162952184677124, 'reward_total_composite_mean': 0.5793253779411316, 'reward_total_composite_std': 0.22016698122024536, 'epoch': 0.05}

 29%|██▉       | 961/3300 [3:37:27<8:18:35, 12.79s/it]INFO 04-12 19:35:53 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 962/3300 [3:37:35<7:30:39, 11.57s/it]2026-04-12 19:36:01,954 | INFO | train_grpo_train | metrics_logged mode=train step=962

                                                      
{'loss': 0.0372, 'grad_norm': 20.44489097595215, 'learning_rate': 7.087878787878788e-06, 'num_tokens': 1709195.0, 'completions/mean_length': 20.75, 'completions/min_length': 18.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8680230975151062, 'rewards/meter/std': 0.33155015110969543, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.7146035432815552, 'rewards/repeat_penalty/std': 0.05397561192512512, 'rewards/repeat_floor/mean': 0.9801807403564453, 'rewards/repeat_floor/std': 0.006781466770917177, 'rewards/near_duplicate_penalty/mean': 0.9909484386444092, 'rewards/near_duplicate_penalty/std': 0.01611136458814144, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9615384340286255, 'rewards/distinct_2/std': 0.07121692597866058, 'rewards/total_composite/mean': 0.4299643039703369, 'rewards/total_composite/std': 0.2295384556055069, 'reward': 0.4299643039703369, 'reward_std': 0.2295384407043457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11307244747877121, 'sampling/sampling_logp_difference/max': 1.1255760192871094, 'sampling/importance_sampling_ratio/min': 0.3244655132293701, 'sampling/importance_sampling_ratio/mean': 1.0188649892807007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6549806296825409, 'clip_ratio/low_mean': 0.03147546947002411, 'clip_ratio/low_min': 0.03147546947002411, 'clip_ratio/high_mean': 0.047945951111614704, 'clip_ratio/high_max': 0.047945951111614704, 'clip_ratio/region_mean': 0.07942142058163881, 'reward_total_mean': 0.4299643039703369, 'reward_meter_mean': 0.8680230975151062, 'reward_meter_std': 0.33155015110969543, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7146035432815552, 'reward_repeat_penalty_std': 0.05397561192512512, 'reward_repeat_floor_mean': 0.9801807403564453, 'reward_repeat_floor_std': 0.006781466770917177, 'reward_near_duplicate_penalty_mean': 0.9909484386444092, 'reward_near_duplicate_penalty_std': 0.01611136458814144, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9615384340286255, 'reward_distinct_2_std': 0.07121692597866058, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.4299643039703369, 'reward_total_composite_std': 0.2295384556055069, 'epoch': 0.05}

 29%|██▉       | 962/3300 [3:37:35<7:30:39, 11.57s/it]INFO 04-12 19:36:02 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 963/3300 [3:37:47<7:30:29, 11.57s/it]2026-04-12 19:36:13,520 | INFO | train_grpo_train | metrics_logged mode=train step=963

                                                      
{'loss': 0.0286, 'grad_norm': 13.568646430969238, 'learning_rate': 7.084848484848485e-06, 'num_tokens': 1710796.0, 'completions/mean_length': 46.125, 'completions/min_length': 38.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.39017507433891296, 'rewards/meter/std': 0.36206796765327454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9397321343421936, 'rewards/lexical_plausibility/std': 0.13258253037929535, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9037748575210571, 'rewards/repeat_penalty/std': 0.06213401257991791, 'rewards/repeat_floor/mean': 0.9888076782226562, 'rewards/repeat_floor/std': 0.007460832595825195, 'rewards/near_duplicate_penalty/mean': 0.9621427059173584, 'rewards/near_duplicate_penalty/std': 0.032039932906627655, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9387362599372864, 'rewards/distinct_2/std': 0.04270491376519203, 'rewards/total_composite/mean': 0.16202372312545776, 'rewards/total_composite/std': 0.14978615939617157, 'reward': 0.16202372312545776, 'reward_std': 0.14978615939617157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1728232353925705, 'sampling/sampling_logp_difference/max': 3.8115999698638916, 'sampling/importance_sampling_ratio/min': 0.02211277186870575, 'sampling/importance_sampling_ratio/mean': 1.0311328172683716, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7943053096532822, 'clip_ratio/low_mean': 0.09817915316671133, 'clip_ratio/low_min': 0.09817915316671133, 'clip_ratio/high_mean': 0.06024638283997774, 'clip_ratio/high_max': 0.06024638283997774, 'clip_ratio/region_mean': 0.15842553600668907, 'reward_total_mean': 0.16202372312545776, 'reward_meter_mean': 0.39017507433891296, 'reward_meter_std': 0.36206796765327454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9397321343421936, 'reward_lexical_plausibility_std': 0.13258253037929535, 'reward_repeat_penalty_mean': 0.9037748575210571, 'reward_repeat_penalty_std': 0.06213401257991791, 'reward_repeat_floor_mean': 0.9888076782226562, 'reward_repeat_floor_std': 0.007460832595825195, 'reward_near_duplicate_penalty_mean': 0.9621427059173584, 'reward_near_duplicate_penalty_std': 0.032039932906627655, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9387362599372864, 'reward_distinct_2_std': 0.04270491376519203, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.16202372312545776, 'reward_total_composite_std': 0.14978615939617157, 'epoch': 0.05}

 29%|██▉       | 963/3300 [3:37:47<7:30:29, 11.57s/it]INFO 04-12 19:36:13 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 964/3300 [3:37:55<6:51:41, 10.57s/it]2026-04-12 19:36:21,782 | INFO | train_grpo_train | metrics_logged mode=train step=964

                                                      
{'loss': -0.036, 'grad_norm': 22.8244686126709, 'learning_rate': 7.081818181818182e-06, 'num_tokens': 1712131.0, 'completions/mean_length': 19.875, 'completions/min_length': 18.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.5185908675193787, 'rewards/meter/std': 0.3965068757534027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.3634728491306305, 'rewards/repeat_penalty/mean': 0.7266660928726196, 'rewards/repeat_penalty/std': 0.03757713735103607, 'rewards/repeat_floor/mean': 0.9803332090377808, 'rewards/repeat_floor/std': 0.007515429984778166, 'rewards/near_duplicate_penalty/mean': 0.9688881635665894, 'rewards/near_duplicate_penalty/std': 0.05010281875729561, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17059214413166046, 'rewards/total_composite/std': 0.22592337429523468, 'reward': 0.17059214413166046, 'reward_std': 0.2259233593940735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10648006945848465, 'sampling/sampling_logp_difference/max': 1.299217700958252, 'sampling/importance_sampling_ratio/min': 0.2727450728416443, 'sampling/importance_sampling_ratio/mean': 1.0147902965545654, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5916283205151558, 'clip_ratio/low_mean': 0.05069444561377168, 'clip_ratio/low_min': 0.05069444561377168, 'clip_ratio/high_mean': 0.04408212564885616, 'clip_ratio/high_max': 0.04408212564885616, 'clip_ratio/region_mean': 0.09477657126262784, 'reward_total_mean': 0.17059214413166046, 'reward_meter_mean': 0.5185908675193787, 'reward_meter_std': 0.3965068757534027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7266660928726196, 'reward_repeat_penalty_std': 0.03757713735103607, 'reward_repeat_floor_mean': 0.9803332090377808, 'reward_repeat_floor_std': 0.007515429984778166, 'reward_near_duplicate_penalty_mean': 0.9688881635665894, 'reward_near_duplicate_penalty_std': 0.05010281875729561, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.3634728491306305, 'reward_total_composite_mean': 0.17059214413166046, 'reward_total_composite_std': 0.22592337429523468, 'epoch': 0.05}

 29%|██▉       | 964/3300 [3:37:55<6:51:41, 10.57s/it]INFO 04-12 19:36:22 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 965/3300 [3:38:09<7:36:06, 11.72s/it]2026-04-12 19:36:36,176 | INFO | train_grpo_train | metrics_logged mode=train step=965

                                                      
{'loss': -0.0327, 'grad_norm': 5.6483283042907715, 'learning_rate': 7.07878787878788e-06, 'num_tokens': 1713802.0, 'completions/mean_length': 97.875, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.71428680419922, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.4031611680984497, 'rewards/meter/std': 0.3963141143321991, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9448670148849487, 'rewards/lexical_plausibility/std': 0.15593969821929932, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.27105283737182617, 'rewards/repeat_penalty/mean': 0.8642938137054443, 'rewards/repeat_penalty/std': 0.24190878868103027, 'rewards/repeat_floor/mean': 0.983004093170166, 'rewards/repeat_floor/std': 0.03545467182993889, 'rewards/near_duplicate_penalty/mean': 0.9417939186096191, 'rewards/near_duplicate_penalty/std': 0.13483747839927673, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9498327970504761, 'rewards/distinct_2/std': 0.11722736805677414, 'rewards/total_composite/mean': 0.12933379411697388, 'rewards/total_composite/std': 0.14334839582443237, 'reward': 0.12933379411697388, 'reward_std': 0.14334839582443237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19400382041931152, 'sampling/sampling_logp_difference/max': 1.928518295288086, 'sampling/importance_sampling_ratio/min': 0.14536342024803162, 'sampling/importance_sampling_ratio/mean': 1.01447594165802, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3471220135688782, 'clip_ratio/low_mean': 0.09430984780192375, 'clip_ratio/low_min': 0.09430984780192375, 'clip_ratio/high_mean': 0.04414682649075985, 'clip_ratio/high_max': 0.04414682649075985, 'clip_ratio/region_mean': 0.1384566742926836, 'reward_total_mean': 0.12933379411697388, 'reward_meter_mean': 0.4031611680984497, 'reward_meter_std': 0.3963141143321991, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9448670148849487, 'reward_lexical_plausibility_std': 0.15593969821929932, 'reward_repeat_penalty_mean': 0.8642938137054443, 'reward_repeat_penalty_std': 0.24190878868103027, 'reward_repeat_floor_mean': 0.983004093170166, 'reward_repeat_floor_std': 0.03545467182993889, 'reward_near_duplicate_penalty_mean': 0.9417939186096191, 'reward_near_duplicate_penalty_std': 0.13483747839927673, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9498327970504761, 'reward_distinct_2_std': 0.11722736805677414, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.27105283737182617, 'reward_total_composite_mean': 0.12933379411697388, 'reward_total_composite_std': 0.14334839582443237, 'epoch': 0.05}

 29%|██▉       | 965/3300 [3:38:10<7:36:06, 11.72s/it]INFO 04-12 19:36:36 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 966/3300 [3:38:24<8:05:36, 12.48s/it]2026-04-12 19:36:50,444 | INFO | train_grpo_train | metrics_logged mode=train step=966

                                                      
{'loss': -0.131, 'grad_norm': 5.304625988006592, 'learning_rate': 7.075757575757576e-06, 'num_tokens': 1715623.0, 'completions/mean_length': 125.625, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.42857360839844, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.717265248298645, 'rewards/meter/std': 0.2712932229042053, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9700000286102295, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.9640973210334778, 'rewards/lexical_plausibility/std': 0.10154811292886734, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.8364830017089844, 'rewards/repeat_penalty/std': 0.12496007978916168, 'rewards/repeat_floor/mean': 0.984601616859436, 'rewards/repeat_floor/std': 0.007827362976968288, 'rewards/near_duplicate_penalty/mean': 0.9498469829559326, 'rewards/near_duplicate_penalty/std': 0.03668202832341194, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9177031517028809, 'rewards/distinct_2/std': 0.05820242688059807, 'rewards/total_composite/mean': 0.22700145840644836, 'rewards/total_composite/std': 0.11999580264091492, 'reward': 0.22700145840644836, 'reward_std': 0.11999580264091492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17169775068759918, 'sampling/sampling_logp_difference/max': 1.4786462783813477, 'sampling/importance_sampling_ratio/min': 0.22794605791568756, 'sampling/importance_sampling_ratio/mean': 0.9984003305435181, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7991107404232025, 'clip_ratio/low_mean': 0.03892543911933899, 'clip_ratio/low_min': 0.03892543911933899, 'clip_ratio/high_mean': 0.1201695678755641, 'clip_ratio/high_max': 0.1201695678755641, 'clip_ratio/region_mean': 0.1590950069949031, 'reward_total_mean': 0.22700145840644836, 'reward_meter_mean': 0.717265248298645, 'reward_meter_std': 0.2712932229042053, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9700000286102295, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.9640973210334778, 'reward_lexical_plausibility_std': 0.10154811292886734, 'reward_repeat_penalty_mean': 0.8364830017089844, 'reward_repeat_penalty_std': 0.12496007978916168, 'reward_repeat_floor_mean': 0.984601616859436, 'reward_repeat_floor_std': 0.007827362976968288, 'reward_near_duplicate_penalty_mean': 0.9498469829559326, 'reward_near_duplicate_penalty_std': 0.03668202832341194, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9177031517028809, 'reward_distinct_2_std': 0.05820242688059807, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.22700145840644836, 'reward_total_composite_std': 0.11999580264091492, 'epoch': 0.05}

 29%|██▉       | 966/3300 [3:38:24<8:05:36, 12.48s/it]INFO 04-12 19:36:50 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 967/3300 [3:38:38<8:30:03, 13.12s/it]2026-04-12 19:37:05,040 | INFO | train_grpo_train | metrics_logged mode=train step=967

                                                      
{'loss': -0.2086, 'grad_norm': 5.202090263366699, 'learning_rate': 7.072727272727273e-06, 'num_tokens': 1717764.0, 'completions/mean_length': 107.625, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.85714340209961, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.9686695337295532, 'rewards/meter/std': 0.04694174602627754, 'rewards/count_adherence/mean': 0.3125, 'rewards/count_adherence/std': 0.2834733724594116, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.793749988079071, 'rewards/count_floor/std': 0.08504199981689453, 'rewards/lexical_plausibility/mean': 0.9523479342460632, 'rewards/lexical_plausibility/std': 0.11168882250785828, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.6907267570495605, 'rewards/repeat_penalty/std': 0.11627142876386642, 'rewards/repeat_floor/mean': 0.9688509702682495, 'rewards/repeat_floor/std': 0.014714229851961136, 'rewards/near_duplicate_penalty/mean': 0.9194468259811401, 'rewards/near_duplicate_penalty/std': 0.058872126042842865, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8923208713531494, 'rewards/distinct_2/std': 0.08279687166213989, 'rewards/total_composite/mean': 0.20435890555381775, 'rewards/total_composite/std': 0.10451614111661911, 'reward': 0.20435890555381775, 'reward_std': 0.10451613366603851, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15677496790885925, 'sampling/sampling_logp_difference/max': 1.6174089908599854, 'sampling/importance_sampling_ratio/min': 0.19841212034225464, 'sampling/importance_sampling_ratio/mean': 1.0298054218292236, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2037087976932526, 'clip_ratio/low_mean': 0.0291422288864851, 'clip_ratio/low_min': 0.0291422288864851, 'clip_ratio/high_mean': 0.07766798511147499, 'clip_ratio/high_max': 0.07766798511147499, 'clip_ratio/region_mean': 0.10681021399796009, 'reward_total_mean': 0.20435890555381775, 'reward_meter_mean': 0.9686695337295532, 'reward_meter_std': 0.04694174602627754, 'reward_count_adherence_mean': 0.3125, 'reward_count_adherence_std': 0.2834733724594116, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.793749988079071, 'reward_count_floor_std': 0.08504199981689453, 'reward_lexical_plausibility_mean': 0.9523479342460632, 'reward_lexical_plausibility_std': 0.11168882250785828, 'reward_repeat_penalty_mean': 0.6907267570495605, 'reward_repeat_penalty_std': 0.11627142876386642, 'reward_repeat_floor_mean': 0.9688509702682495, 'reward_repeat_floor_std': 0.014714229851961136, 'reward_near_duplicate_penalty_mean': 0.9194468259811401, 'reward_near_duplicate_penalty_std': 0.058872126042842865, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8923208713531494, 'reward_distinct_2_std': 0.08279687166213989, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.20435890555381775, 'reward_total_composite_std': 0.10451614111661911, 'epoch': 0.05}

 29%|██▉       | 967/3300 [3:38:38<8:30:03, 13.12s/it]INFO 04-12 19:37:05 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 968/3300 [3:38:52<8:36:38, 13.29s/it]2026-04-12 19:37:18,740 | INFO | train_grpo_train | metrics_logged mode=train step=968

                                                      
{'loss': -0.0438, 'grad_norm': 5.238527774810791, 'learning_rate': 7.06969696969697e-06, 'num_tokens': 1719289.0, 'completions/mean_length': 92.625, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.71428680419922, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7131102085113525, 'rewards/meter/std': 0.3417653441429138, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9248441457748413, 'rewards/lexical_plausibility/std': 0.12555167078971863, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.3160809278488159, 'rewards/repeat_penalty/mean': 0.8501054048538208, 'rewards/repeat_penalty/std': 0.1546521633863449, 'rewards/repeat_floor/mean': 0.9813551306724548, 'rewards/repeat_floor/std': 0.018153000622987747, 'rewards/near_duplicate_penalty/mean': 0.9203752279281616, 'rewards/near_duplicate_penalty/std': 0.07198642194271088, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9463757276535034, 'rewards/distinct_2/std': 0.07744858413934708, 'rewards/total_composite/mean': 0.31176048517227173, 'rewards/total_composite/std': 0.220302551984787, 'reward': 0.31176048517227173, 'reward_std': 0.220302551984787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19009913504123688, 'sampling/sampling_logp_difference/max': 2.8461923599243164, 'sampling/importance_sampling_ratio/min': 0.05806499347090721, 'sampling/importance_sampling_ratio/mean': 0.9877751469612122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6971128806471825, 'clip_ratio/low_mean': 0.018067227210849524, 'clip_ratio/low_min': 0.018067227210849524, 'clip_ratio/high_mean': 0.12980854138731956, 'clip_ratio/high_max': 0.12980854138731956, 'clip_ratio/region_mean': 0.1478757685981691, 'reward_total_mean': 0.31176048517227173, 'reward_meter_mean': 0.7131102085113525, 'reward_meter_std': 0.3417653441429138, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9248441457748413, 'reward_lexical_plausibility_std': 0.12555167078971863, 'reward_repeat_penalty_mean': 0.8501054048538208, 'reward_repeat_penalty_std': 0.1546521633863449, 'reward_repeat_floor_mean': 0.9813551306724548, 'reward_repeat_floor_std': 0.018153000622987747, 'reward_near_duplicate_penalty_mean': 0.9203752279281616, 'reward_near_duplicate_penalty_std': 0.07198642194271088, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9463757276535034, 'reward_distinct_2_std': 0.07744858413934708, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.3160809278488159, 'reward_total_composite_mean': 0.31176048517227173, 'reward_total_composite_std': 0.220302551984787, 'epoch': 0.05}

 29%|██▉       | 968/3300 [3:38:52<8:36:38, 13.29s/it]INFO 04-12 19:37:18 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 969/3300 [3:39:01<7:43:39, 11.93s/it]2026-04-12 19:37:27,504 | INFO | train_grpo_train | metrics_logged mode=train step=969

                                                      
{'loss': -0.0781, 'grad_norm': 18.737300872802734, 'learning_rate': 7.066666666666667e-06, 'num_tokens': 1720934.0, 'completions/mean_length': 27.625, 'completions/min_length': 20.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.954509973526001, 'rewards/meter/std': 0.09417146444320679, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.26250001788139343, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.7761447429656982, 'rewards/repeat_penalty/std': 0.1170436441898346, 'rewards/repeat_floor/mean': 0.9829691052436829, 'rewards/repeat_floor/std': 0.01006452925503254, 'rewards/near_duplicate_penalty/mean': 0.9710761904716492, 'rewards/near_duplicate_penalty/std': 0.04073138162493706, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.21585124731063843, 'rewards/total_composite/std': 0.13087515532970428, 'reward': 0.21585124731063843, 'reward_std': 0.13087515532970428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16987186670303345, 'sampling/sampling_logp_difference/max': 1.384159803390503, 'sampling/importance_sampling_ratio/min': 0.25053420662879944, 'sampling/importance_sampling_ratio/mean': 1.0210856199264526, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.413540206849575, 'clip_ratio/low_mean': 0.09910714440047741, 'clip_ratio/low_min': 0.09910714440047741, 'clip_ratio/high_mean': 0.06668862700462341, 'clip_ratio/high_max': 0.06668862700462341, 'clip_ratio/region_mean': 0.16579577140510082, 'reward_total_mean': 0.21585124731063843, 'reward_meter_mean': 0.954509973526001, 'reward_meter_std': 0.09417146444320679, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7761447429656982, 'reward_repeat_penalty_std': 0.1170436441898346, 'reward_repeat_floor_mean': 0.9829691052436829, 'reward_repeat_floor_std': 0.01006452925503254, 'reward_near_duplicate_penalty_mean': 0.9710761904716492, 'reward_near_duplicate_penalty_std': 0.04073138162493706, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.26250001788139343, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.21585124731063843, 'reward_total_composite_std': 0.13087515532970428, 'epoch': 0.05}

 29%|██▉       | 969/3300 [3:39:01<7:43:39, 11.93s/it]INFO 04-12 19:37:27 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 970/3300 [3:39:14<7:58:22, 12.32s/it]2026-04-12 19:37:40,719 | INFO | train_grpo_train | metrics_logged mode=train step=970

                                                      
{'loss': -0.037, 'grad_norm': 3.005152940750122, 'learning_rate': 7.063636363636365e-06, 'num_tokens': 1722556.0, 'completions/mean_length': 157.75, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 39.66666793823242, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.423380970954895, 'rewards/meter/std': 0.3594983220100403, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8948670625686646, 'rewards/lexical_plausibility/std': 0.17127196490764618, 'rewards/judge_quality/mean': 0.2587500214576721, 'rewards/judge_quality/std': 0.2947850525379181, 'rewards/repeat_penalty/mean': 0.5997943878173828, 'rewards/repeat_penalty/std': 0.22570596635341644, 'rewards/repeat_floor/mean': 0.9517270922660828, 'rewards/repeat_floor/std': 0.037120044231414795, 'rewards/near_duplicate_penalty/mean': 0.8516302704811096, 'rewards/near_duplicate_penalty/std': 0.1349460482597351, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8359174728393555, 'rewards/distinct_2/std': 0.17480257153511047, 'rewards/total_composite/mean': 0.17451606690883636, 'rewards/total_composite/std': 0.29133304953575134, 'reward': 0.17451606690883636, 'reward_std': 0.29133301973342896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15425877273082733, 'sampling/sampling_logp_difference/max': 1.5117311477661133, 'sampling/importance_sampling_ratio/min': 0.22052788734436035, 'sampling/importance_sampling_ratio/mean': 1.0376791954040527, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9024818316102028, 'clip_ratio/low_mean': 0.0920450184494257, 'clip_ratio/low_min': 0.0920450184494257, 'clip_ratio/high_mean': 0.018597560934722424, 'clip_ratio/high_max': 0.018597560934722424, 'clip_ratio/region_mean': 0.11064257938414812, 'reward_total_mean': 0.17451606690883636, 'reward_meter_mean': 0.423380970954895, 'reward_meter_std': 0.3594983220100403, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8948670625686646, 'reward_lexical_plausibility_std': 0.17127196490764618, 'reward_repeat_penalty_mean': 0.5997943878173828, 'reward_repeat_penalty_std': 0.22570596635341644, 'reward_repeat_floor_mean': 0.9517270922660828, 'reward_repeat_floor_std': 0.037120044231414795, 'reward_near_duplicate_penalty_mean': 0.8516302704811096, 'reward_near_duplicate_penalty_std': 0.1349460482597351, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8359174728393555, 'reward_distinct_2_std': 0.17480257153511047, 'reward_judge_quality_mean': 0.2587500214576721, 'reward_judge_quality_std': 0.2947850525379181, 'reward_total_composite_mean': 0.17451606690883636, 'reward_total_composite_std': 0.29133304953575134, 'epoch': 0.05}

 29%|██▉       | 970/3300 [3:39:14<7:58:22, 12.32s/it]INFO 04-12 19:37:40 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 971/3300 [3:39:24<7:29:21, 11.58s/it]2026-04-12 19:37:50,565 | INFO | train_grpo_train | metrics_logged mode=train step=971

                                                      
{'loss': -0.0028, 'grad_norm': 14.2072172164917, 'learning_rate': 7.060606060606061e-06, 'num_tokens': 1724298.0, 'completions/mean_length': 51.75, 'completions/min_length': 38.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7715051174163818, 'rewards/meter/std': 0.30648893117904663, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8485367298126221, 'rewards/repeat_penalty/std': 0.10371339321136475, 'rewards/repeat_floor/mean': 0.9831637144088745, 'rewards/repeat_floor/std': 0.009831927716732025, 'rewards/near_duplicate_penalty/mean': 0.9448339343070984, 'rewards/near_duplicate_penalty/std': 0.0315127931535244, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9257071614265442, 'rewards/distinct_2/std': 0.042042117565870285, 'rewards/total_composite/mean': 0.2650087773799896, 'rewards/total_composite/std': 0.12917882204055786, 'reward': 0.2650087773799896, 'reward_std': 0.12917882204055786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16324123740196228, 'sampling/sampling_logp_difference/max': 1.9861998558044434, 'sampling/importance_sampling_ratio/min': 0.13721586763858795, 'sampling/importance_sampling_ratio/mean': 1.0014806985855103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9240273982286453, 'clip_ratio/low_mean': 0.02583712339401245, 'clip_ratio/low_min': 0.02583712339401245, 'clip_ratio/high_mean': 0.11590507440268993, 'clip_ratio/high_max': 0.11590507440268993, 'clip_ratio/region_mean': 0.14174219779670238, 'reward_total_mean': 0.2650087773799896, 'reward_meter_mean': 0.7715051174163818, 'reward_meter_std': 0.30648893117904663, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8485367298126221, 'reward_repeat_penalty_std': 0.10371339321136475, 'reward_repeat_floor_mean': 0.9831637144088745, 'reward_repeat_floor_std': 0.009831927716732025, 'reward_near_duplicate_penalty_mean': 0.9448339343070984, 'reward_near_duplicate_penalty_std': 0.0315127931535244, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9257071614265442, 'reward_distinct_2_std': 0.042042117565870285, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.2650087773799896, 'reward_total_composite_std': 0.12917882204055786, 'epoch': 0.05}

 29%|██▉       | 971/3300 [3:39:24<7:29:21, 11.58s/it]INFO 04-12 19:37:50 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 972/3300 [3:39:38<7:59:14, 12.35s/it]2026-04-12 19:38:04,725 | INFO | train_grpo_train | metrics_logged mode=train step=972

                                                      
{'loss': -0.0731, 'grad_norm': 4.164340972900391, 'learning_rate': 7.057575757575759e-06, 'num_tokens': 1725689.0, 'completions/mean_length': 93.875, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.142860412597656, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6997895240783691, 'rewards/meter/std': 0.433590292930603, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9476128816604614, 'rewards/lexical_plausibility/std': 0.1481732279062271, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.8653319478034973, 'rewards/repeat_penalty/std': 0.16824647784233093, 'rewards/repeat_floor/mean': 0.9855325222015381, 'rewards/repeat_floor/std': 0.018374187871813774, 'rewards/near_duplicate_penalty/mean': 0.9500287175178528, 'rewards/near_duplicate_penalty/std': 0.06329852342605591, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9642024636268616, 'rewards/distinct_2/std': 0.06630805134773254, 'rewards/total_composite/mean': 0.2770465314388275, 'rewards/total_composite/std': 0.18259696662425995, 'reward': 0.2770465314388275, 'reward_std': 0.18259696662425995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16110078990459442, 'sampling/sampling_logp_difference/max': 3.9171507358551025, 'sampling/importance_sampling_ratio/min': 0.0198977068066597, 'sampling/importance_sampling_ratio/mean': 1.0118138790130615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8980119079351425, 'clip_ratio/low_mean': 0.033928572200238705, 'clip_ratio/low_min': 0.033928572200238705, 'clip_ratio/high_mean': 0.11240511201322079, 'clip_ratio/high_max': 0.11240511201322079, 'clip_ratio/region_mean': 0.1463336842134595, 'reward_total_mean': 0.2770465314388275, 'reward_meter_mean': 0.6997895240783691, 'reward_meter_std': 0.433590292930603, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9476128816604614, 'reward_lexical_plausibility_std': 0.1481732279062271, 'reward_repeat_penalty_mean': 0.8653319478034973, 'reward_repeat_penalty_std': 0.16824647784233093, 'reward_repeat_floor_mean': 0.9855325222015381, 'reward_repeat_floor_std': 0.018374187871813774, 'reward_near_duplicate_penalty_mean': 0.9500287175178528, 'reward_near_duplicate_penalty_std': 0.06329852342605591, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9642024636268616, 'reward_distinct_2_std': 0.06630805134773254, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2770465314388275, 'reward_total_composite_std': 0.18259696662425995, 'epoch': 0.05}

 29%|██▉       | 972/3300 [3:39:38<7:59:14, 12.35s/it]INFO 04-12 19:38:04 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 973/3300 [3:39:52<8:14:53, 12.76s/it]2026-04-12 19:38:18,441 | INFO | train_grpo_train | metrics_logged mode=train step=973

                                                      
{'loss': -0.0693, 'grad_norm': 3.999750852584839, 'learning_rate': 7.054545454545455e-06, 'num_tokens': 1727155.0, 'completions/mean_length': 96.25, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.85714340209961, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8346974849700928, 'rewards/meter/std': 0.27647799253463745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9555260539054871, 'rewards/lexical_plausibility/std': 0.12579134106636047, 'rewards/judge_quality/mean': 0.5300000309944153, 'rewards/judge_quality/std': 0.2929163873195648, 'rewards/repeat_penalty/mean': 0.96783846616745, 'rewards/repeat_penalty/std': 0.061910782009363174, 'rewards/repeat_floor/mean': 0.9961574077606201, 'rewards/repeat_floor/std': 0.006572483107447624, 'rewards/near_duplicate_penalty/mean': 0.985371470451355, 'rewards/near_duplicate_penalty/std': 0.015771590173244476, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9816849827766418, 'rewards/distinct_2/std': 0.05180269479751587, 'rewards/total_composite/mean': 0.4928828477859497, 'rewards/total_composite/std': 0.3025922477245331, 'reward': 0.4928828477859497, 'reward_std': 0.3025922477245331, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14396162331104279, 'sampling/sampling_logp_difference/max': 1.389665126800537, 'sampling/importance_sampling_ratio/min': 0.24915872514247894, 'sampling/importance_sampling_ratio/mean': 1.0058937072753906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7554473727941513, 'clip_ratio/low_mean': 0.07631270308047533, 'clip_ratio/low_min': 0.07631270308047533, 'clip_ratio/high_mean': 0.05304666329175234, 'clip_ratio/high_max': 0.05304666329175234, 'clip_ratio/region_mean': 0.12935936637222767, 'reward_total_mean': 0.4928828477859497, 'reward_meter_mean': 0.8346974849700928, 'reward_meter_std': 0.27647799253463745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9555260539054871, 'reward_lexical_plausibility_std': 0.12579134106636047, 'reward_repeat_penalty_mean': 0.96783846616745, 'reward_repeat_penalty_std': 0.061910782009363174, 'reward_repeat_floor_mean': 0.9961574077606201, 'reward_repeat_floor_std': 0.006572483107447624, 'reward_near_duplicate_penalty_mean': 0.985371470451355, 'reward_near_duplicate_penalty_std': 0.015771590173244476, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9816849827766418, 'reward_distinct_2_std': 0.05180269479751587, 'reward_judge_quality_mean': 0.5300000309944153, 'reward_judge_quality_std': 0.2929163873195648, 'reward_total_composite_mean': 0.4928828477859497, 'reward_total_composite_std': 0.3025922477245331, 'epoch': 0.05}

 29%|██▉       | 973/3300 [3:39:52<8:14:53, 12.76s/it]INFO 04-12 19:38:18 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 974/3300 [3:40:02<7:44:37, 11.99s/it]2026-04-12 19:38:28,621 | INFO | train_grpo_train | metrics_logged mode=train step=974

                                                      
{'loss': -0.0492, 'grad_norm': 8.404296875, 'learning_rate': 7.0515151515151525e-06, 'num_tokens': 1729572.0, 'completions/mean_length': 114.125, 'completions/min_length': 29.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.9692351222038269, 'rewards/meter/std': 0.04848598688840866, 'rewards/count_adherence/mean': 0.6785714626312256, 'rewards/count_adherence/std': 0.22587698698043823, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9035714268684387, 'rewards/count_floor/std': 0.0677630826830864, 'rewards/lexical_plausibility/mean': 0.9994343519210815, 'rewards/lexical_plausibility/std': 0.0015997890150174499, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.32866936922073364, 'rewards/repeat_penalty/std': 0.28666263818740845, 'rewards/repeat_floor/mean': 0.9035038352012634, 'rewards/repeat_floor/std': 0.06129074469208717, 'rewards/near_duplicate_penalty/mean': 0.7400773167610168, 'rewards/near_duplicate_penalty/std': 0.1981327086687088, 'rewards/opening_diversity/mean': 0.8542613983154297, 'rewards/opening_diversity/std': 0.2397436797618866, 'rewards/distinct_2/mean': 0.4581839442253113, 'rewards/distinct_2/std': 0.3071747422218323, 'rewards/total_composite/mean': 0.14026211202144623, 'rewards/total_composite/std': 0.06768891960382462, 'reward': 0.14026211202144623, 'reward_std': 0.06768891960382462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11227668821811676, 'sampling/sampling_logp_difference/max': 1.5684280395507812, 'sampling/importance_sampling_ratio/min': 0.20837247371673584, 'sampling/importance_sampling_ratio/mean': 0.9983904361724854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8551279902458191, 'clip_ratio/low_mean': 0.08559842547401786, 'clip_ratio/low_min': 0.08559842547401786, 'clip_ratio/high_mean': 0.02178030274808407, 'clip_ratio/high_max': 0.02178030274808407, 'clip_ratio/region_mean': 0.10737872822210193, 'reward_total_mean': 0.14026211202144623, 'reward_meter_mean': 0.9692351222038269, 'reward_meter_std': 0.04848598688840866, 'reward_count_adherence_mean': 0.6785714626312256, 'reward_count_adherence_std': 0.22587698698043823, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9035714268684387, 'reward_count_floor_std': 0.0677630826830864, 'reward_lexical_plausibility_mean': 0.9994343519210815, 'reward_lexical_plausibility_std': 0.0015997890150174499, 'reward_repeat_penalty_mean': 0.32866936922073364, 'reward_repeat_penalty_std': 0.28666263818740845, 'reward_repeat_floor_mean': 0.9035038352012634, 'reward_repeat_floor_std': 0.06129074469208717, 'reward_near_duplicate_penalty_mean': 0.7400773167610168, 'reward_near_duplicate_penalty_std': 0.1981327086687088, 'reward_opening_diversity_mean': 0.8542613983154297, 'reward_opening_diversity_std': 0.2397436797618866, 'reward_distinct_2_mean': 0.4581839442253113, 'reward_distinct_2_std': 0.3071747422218323, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.14026211202144623, 'reward_total_composite_std': 0.06768891960382462, 'epoch': 0.05}

 30%|██▉       | 974/3300 [3:40:02<7:44:37, 11.99s/it]INFO 04-12 19:38:28 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 975/3300 [3:40:17<8:21:42, 12.95s/it]2026-04-12 19:38:43,809 | INFO | train_grpo_train | metrics_logged mode=train step=975

                                                      
{'loss': -0.0623, 'grad_norm': 6.282349586486816, 'learning_rate': 7.048484848484849e-06, 'num_tokens': 1731370.0, 'completions/mean_length': 93.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8668069243431091, 'rewards/meter/std': 0.22566258907318115, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9519826769828796, 'rewards/lexical_plausibility/std': 0.11270204931497574, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.7283657193183899, 'rewards/repeat_penalty/std': 0.15732789039611816, 'rewards/repeat_floor/mean': 0.9734642505645752, 'rewards/repeat_floor/std': 0.020226595923304558, 'rewards/near_duplicate_penalty/mean': 0.9339142441749573, 'rewards/near_duplicate_penalty/std': 0.06464814394712448, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9194676876068115, 'rewards/distinct_2/std': 0.11235395073890686, 'rewards/total_composite/mean': 0.238030344247818, 'rewards/total_composite/std': 0.12912023067474365, 'reward': 0.238030344247818, 'reward_std': 0.12912021577358246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16155895590782166, 'sampling/sampling_logp_difference/max': 1.6713132858276367, 'sampling/importance_sampling_ratio/min': 0.18800000846385956, 'sampling/importance_sampling_ratio/mean': 1.0223023891448975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1112970113754272, 'clip_ratio/low_mean': 0.042171718552708626, 'clip_ratio/low_min': 0.042171718552708626, 'clip_ratio/high_mean': 0.10932516120374203, 'clip_ratio/high_max': 0.10932516120374203, 'clip_ratio/region_mean': 0.15149687975645065, 'reward_total_mean': 0.238030344247818, 'reward_meter_mean': 0.8668069243431091, 'reward_meter_std': 0.22566258907318115, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9519826769828796, 'reward_lexical_plausibility_std': 0.11270204931497574, 'reward_repeat_penalty_mean': 0.7283657193183899, 'reward_repeat_penalty_std': 0.15732789039611816, 'reward_repeat_floor_mean': 0.9734642505645752, 'reward_repeat_floor_std': 0.020226595923304558, 'reward_near_duplicate_penalty_mean': 0.9339142441749573, 'reward_near_duplicate_penalty_std': 0.06464814394712448, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9194676876068115, 'reward_distinct_2_std': 0.11235395073890686, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.238030344247818, 'reward_total_composite_std': 0.12912023067474365, 'epoch': 0.05}

 30%|██▉       | 975/3300 [3:40:17<8:21:42, 12.95s/it]INFO 04-12 19:38:44 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 976/3300 [3:40:31<8:35:05, 13.30s/it]2026-04-12 19:38:57,927 | INFO | train_grpo_train | metrics_logged mode=train step=976

                                                      
{'loss': -0.1167, 'grad_norm': 4.999024868011475, 'learning_rate': 7.045454545454546e-06, 'num_tokens': 1733106.0, 'completions/mean_length': 107.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.142860412597656, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7845979928970337, 'rewards/meter/std': 0.3847057521343231, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 0.9493359327316284, 'rewards/lexical_plausibility/std': 0.14329960942268372, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.08345229923725128, 'rewards/repeat_penalty/mean': 0.5883820056915283, 'rewards/repeat_penalty/std': 0.2849542796611786, 'rewards/repeat_floor/mean': 0.9443279504776001, 'rewards/repeat_floor/std': 0.0443461611866951, 'rewards/near_duplicate_penalty/mean': 0.8288168907165527, 'rewards/near_duplicate_penalty/std': 0.13788726925849915, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.7292263507843018, 'rewards/distinct_2/std': 0.18950118124485016, 'rewards/total_composite/mean': 0.13117390871047974, 'rewards/total_composite/std': 0.09778672456741333, 'reward': 0.13117390871047974, 'reward_std': 0.09778672456741333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13256482779979706, 'sampling/sampling_logp_difference/max': 1.0620343685150146, 'sampling/importance_sampling_ratio/min': 0.34575170278549194, 'sampling/importance_sampling_ratio/mean': 1.019872784614563, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9142376482486725, 'clip_ratio/low_mean': 0.053039186634123325, 'clip_ratio/low_min': 0.053039186634123325, 'clip_ratio/high_mean': 0.07695492263883352, 'clip_ratio/high_max': 0.07695492263883352, 'clip_ratio/region_mean': 0.12999410927295685, 'reward_total_mean': 0.13117390871047974, 'reward_meter_mean': 0.7845979928970337, 'reward_meter_std': 0.3847057521343231, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 0.9493359327316284, 'reward_lexical_plausibility_std': 0.14329960942268372, 'reward_repeat_penalty_mean': 0.5883820056915283, 'reward_repeat_penalty_std': 0.2849542796611786, 'reward_repeat_floor_mean': 0.9443279504776001, 'reward_repeat_floor_std': 0.0443461611866951, 'reward_near_duplicate_penalty_mean': 0.8288168907165527, 'reward_near_duplicate_penalty_std': 0.13788726925849915, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.7292263507843018, 'reward_distinct_2_std': 0.18950118124485016, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.08345229923725128, 'reward_total_composite_mean': 0.13117390871047974, 'reward_total_composite_std': 0.09778672456741333, 'epoch': 0.05}

 30%|██▉       | 976/3300 [3:40:31<8:35:05, 13.30s/it]INFO 04-12 19:38:58 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 977/3300 [3:40:41<7:53:36, 12.23s/it]2026-04-12 19:39:07,675 | INFO | train_grpo_train | metrics_logged mode=train step=977

                                                      
{'loss': -0.0165, 'grad_norm': 9.522601127624512, 'learning_rate': 7.0424242424242426e-06, 'num_tokens': 1735170.0, 'completions/mean_length': 79.0, 'completions/min_length': 66.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.7234905958175659, 'rewards/meter/std': 0.22925277054309845, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9625000357627869, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.7888998985290527, 'rewards/repeat_penalty/std': 0.16465048491954803, 'rewards/repeat_floor/mean': 0.9782580733299255, 'rewards/repeat_floor/std': 0.017079001292586327, 'rewards/near_duplicate_penalty/mean': 0.9523767232894897, 'rewards/near_duplicate_penalty/std': 0.03251688554883003, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.8461284637451172, 'rewards/distinct_2/std': 0.1207500472664833, 'rewards/total_composite/mean': 0.16878071427345276, 'rewards/total_composite/std': 0.0734722912311554, 'reward': 0.16878071427345276, 'reward_std': 0.0734722912311554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15607404708862305, 'sampling/sampling_logp_difference/max': 1.5497758388519287, 'sampling/importance_sampling_ratio/min': 0.2122955471277237, 'sampling/importance_sampling_ratio/mean': 1.0087289810180664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0343920290470123, 'clip_ratio/low_mean': 0.04488448053598404, 'clip_ratio/low_min': 0.04488448053598404, 'clip_ratio/high_mean': 0.08986167516559362, 'clip_ratio/high_max': 0.08986167516559362, 'clip_ratio/region_mean': 0.13474615570157766, 'reward_total_mean': 0.16878071427345276, 'reward_meter_mean': 0.7234905958175659, 'reward_meter_std': 0.22925277054309845, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9625000357627869, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7888998985290527, 'reward_repeat_penalty_std': 0.16465048491954803, 'reward_repeat_floor_mean': 0.9782580733299255, 'reward_repeat_floor_std': 0.017079001292586327, 'reward_near_duplicate_penalty_mean': 0.9523767232894897, 'reward_near_duplicate_penalty_std': 0.03251688554883003, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.8461284637451172, 'reward_distinct_2_std': 0.1207500472664833, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.16878071427345276, 'reward_total_composite_std': 0.0734722912311554, 'epoch': 0.05}

 30%|██▉       | 977/3300 [3:40:41<7:53:36, 12.23s/it]INFO 04-12 19:39:07 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 978/3300 [3:40:55<8:20:01, 12.92s/it]2026-04-12 19:39:22,249 | INFO | train_grpo_train | metrics_logged mode=train step=978

                                                      
{'loss': -0.0334, 'grad_norm': 2.9417386054992676, 'learning_rate': 7.039393939393941e-06, 'num_tokens': 1737045.0, 'completions/mean_length': 174.375, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 61.833335876464844, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.6534391641616821, 'rewards/meter/std': 0.4237983524799347, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.053032997995615005, 'rewards/lexical_plausibility/mean': 0.8672837615013123, 'rewards/lexical_plausibility/std': 0.22072960436344147, 'rewards/judge_quality/mean': 0.23125000298023224, 'rewards/judge_quality/std': 0.2103356271982193, 'rewards/repeat_penalty/mean': 0.7841565012931824, 'rewards/repeat_penalty/std': 0.13806259632110596, 'rewards/repeat_floor/mean': 0.9757506847381592, 'rewards/repeat_floor/std': 0.01494610495865345, 'rewards/near_duplicate_penalty/mean': 0.9242197275161743, 'rewards/near_duplicate_penalty/std': 0.0468585379421711, 'rewards/opening_diversity/mean': 0.9508928656578064, 'rewards/opening_diversity/std': 0.09533105045557022, 'rewards/distinct_2/mean': 0.88960200548172, 'rewards/distinct_2/std': 0.08879511058330536, 'rewards/total_composite/mean': 0.16443876922130585, 'rewards/total_composite/std': 0.18450936675071716, 'reward': 0.16443876922130585, 'reward_std': 0.18450935184955597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.162736177444458, 'sampling/sampling_logp_difference/max': 1.5083184242248535, 'sampling/importance_sampling_ratio/min': 0.2212817668914795, 'sampling/importance_sampling_ratio/mean': 1.0156605243682861, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.789588488638401, 'clip_ratio/low_mean': 0.06831929553300142, 'clip_ratio/low_min': 0.06831929553300142, 'clip_ratio/high_mean': 0.04146765172481537, 'clip_ratio/high_max': 0.04146765172481537, 'clip_ratio/region_mean': 0.10978694725781679, 'reward_total_mean': 0.16443876922130585, 'reward_meter_mean': 0.6534391641616821, 'reward_meter_std': 0.4237983524799347, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.053032997995615005, 'reward_lexical_plausibility_mean': 0.8672837615013123, 'reward_lexical_plausibility_std': 0.22072960436344147, 'reward_repeat_penalty_mean': 0.7841565012931824, 'reward_repeat_penalty_std': 0.13806259632110596, 'reward_repeat_floor_mean': 0.9757506847381592, 'reward_repeat_floor_std': 0.01494610495865345, 'reward_near_duplicate_penalty_mean': 0.9242197275161743, 'reward_near_duplicate_penalty_std': 0.0468585379421711, 'reward_opening_diversity_mean': 0.9508928656578064, 'reward_opening_diversity_std': 0.09533105045557022, 'reward_distinct_2_mean': 0.88960200548172, 'reward_distinct_2_std': 0.08879511058330536, 'reward_judge_quality_mean': 0.23125000298023224, 'reward_judge_quality_std': 0.2103356271982193, 'reward_total_composite_mean': 0.16443876922130585, 'reward_total_composite_std': 0.18450936675071716, 'epoch': 0.05}

 30%|██▉       | 978/3300 [3:40:56<8:20:01, 12.92s/it]INFO 04-12 19:39:22 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 979/3300 [3:41:10<8:38:23, 13.40s/it]2026-04-12 19:39:36,776 | INFO | train_grpo_train | metrics_logged mode=train step=979

                                                      
{'loss': -0.0126, 'grad_norm': 4.711750030517578, 'learning_rate': 7.036363636363637e-06, 'num_tokens': 1738775.0, 'completions/mean_length': 113.25, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.28571701049805, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9221183657646179, 'rewards/meter/std': 0.10711797326803207, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9554562568664551, 'rewards/lexical_plausibility/std': 0.12598878145217896, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.2728912830352783, 'rewards/repeat_penalty/mean': 0.7102515697479248, 'rewards/repeat_penalty/std': 0.24917364120483398, 'rewards/repeat_floor/mean': 0.966413140296936, 'rewards/repeat_floor/std': 0.025598620995879173, 'rewards/near_duplicate_penalty/mean': 0.8830002546310425, 'rewards/near_duplicate_penalty/std': 0.06831125169992447, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8218122720718384, 'rewards/distinct_2/std': 0.20058032870292664, 'rewards/total_composite/mean': 0.2271711528301239, 'rewards/total_composite/std': 0.1514083296060562, 'reward': 0.2271711528301239, 'reward_std': 0.15140831470489502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15027040243148804, 'sampling/sampling_logp_difference/max': 2.3592848777770996, 'sampling/importance_sampling_ratio/min': 0.09448777139186859, 'sampling/importance_sampling_ratio/mean': 1.0147895812988281, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7766230329871178, 'clip_ratio/low_mean': 0.04790704511106014, 'clip_ratio/low_min': 0.04790704511106014, 'clip_ratio/high_mean': 0.08284418471157551, 'clip_ratio/high_max': 0.08284418471157551, 'clip_ratio/region_mean': 0.13075122982263565, 'reward_total_mean': 0.2271711528301239, 'reward_meter_mean': 0.9221183657646179, 'reward_meter_std': 0.10711797326803207, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9554562568664551, 'reward_lexical_plausibility_std': 0.12598878145217896, 'reward_repeat_penalty_mean': 0.7102515697479248, 'reward_repeat_penalty_std': 0.24917364120483398, 'reward_repeat_floor_mean': 0.966413140296936, 'reward_repeat_floor_std': 0.025598620995879173, 'reward_near_duplicate_penalty_mean': 0.8830002546310425, 'reward_near_duplicate_penalty_std': 0.06831125169992447, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8218122720718384, 'reward_distinct_2_std': 0.20058032870292664, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.2728912830352783, 'reward_total_composite_mean': 0.2271711528301239, 'reward_total_composite_std': 0.1514083296060562, 'epoch': 0.05}

 30%|██▉       | 979/3300 [3:41:10<8:38:23, 13.40s/it]INFO 04-12 19:39:37 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 980/3300 [3:41:25<9:00:19, 13.97s/it]2026-04-12 19:39:52,089 | INFO | train_grpo_train | metrics_logged mode=train step=980

                                                      
{'loss': -0.0726, 'grad_norm': 7.281307220458984, 'learning_rate': 7.033333333333334e-06, 'num_tokens': 1740339.0, 'completions/mean_length': 99.5, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.57143020629883, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5013697743415833, 'rewards/meter/std': 0.42095452547073364, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9430537819862366, 'rewards/lexical_plausibility/std': 0.16106823086738586, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3296508193016052, 'rewards/repeat_penalty/mean': 0.9501906633377075, 'rewards/repeat_penalty/std': 0.08315180242061615, 'rewards/repeat_floor/mean': 0.9953410625457764, 'rewards/repeat_floor/std': 0.005080376751720905, 'rewards/near_duplicate_penalty/mean': 0.9814406633377075, 'rewards/near_duplicate_penalty/std': 0.02067345753312111, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.30800536274909973, 'rewards/total_composite/std': 0.29107582569122314, 'reward': 0.30800536274909973, 'reward_std': 0.29107579588890076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.142264723777771, 'sampling/sampling_logp_difference/max': 1.2008533477783203, 'sampling/importance_sampling_ratio/min': 0.300937294960022, 'sampling/importance_sampling_ratio/mean': 1.0181927680969238, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8764906153082848, 'clip_ratio/low_mean': 0.050627183401957154, 'clip_ratio/low_min': 0.050627183401957154, 'clip_ratio/high_mean': 0.06795885041356087, 'clip_ratio/high_max': 0.06795885041356087, 'clip_ratio/region_mean': 0.11858603381551802, 'reward_total_mean': 0.30800536274909973, 'reward_meter_mean': 0.5013697743415833, 'reward_meter_std': 0.42095452547073364, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9430537819862366, 'reward_lexical_plausibility_std': 0.16106823086738586, 'reward_repeat_penalty_mean': 0.9501906633377075, 'reward_repeat_penalty_std': 0.08315180242061615, 'reward_repeat_floor_mean': 0.9953410625457764, 'reward_repeat_floor_std': 0.005080376751720905, 'reward_near_duplicate_penalty_mean': 0.9814406633377075, 'reward_near_duplicate_penalty_std': 0.02067345753312111, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3296508193016052, 'reward_total_composite_mean': 0.30800536274909973, 'reward_total_composite_std': 0.29107582569122314, 'epoch': 0.05}

 30%|██▉       | 980/3300 [3:41:25<9:00:19, 13.97s/it]INFO 04-12 19:39:52 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 981/3300 [3:41:41<9:23:29, 14.58s/it]2026-04-12 19:40:08,024 | INFO | train_grpo_train | metrics_logged mode=train step=981

                                                      
{'loss': -0.1136, 'grad_norm': 5.602263927459717, 'learning_rate': 7.030303030303031e-06, 'num_tokens': 1742416.0, 'completions/mean_length': 143.625, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 91.00000762939453, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.6096323132514954, 'rewards/meter/std': 0.21670779585838318, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9562499523162842, 'rewards/count_floor/std': 0.01767767407000065, 'rewards/lexical_plausibility/mean': 0.9634765386581421, 'rewards/lexical_plausibility/std': 0.10330399125814438, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.523923397064209, 'rewards/repeat_penalty/std': 0.3207071125507355, 'rewards/repeat_floor/mean': 0.9451455473899841, 'rewards/repeat_floor/std': 0.04579712823033333, 'rewards/near_duplicate_penalty/mean': 0.8665324449539185, 'rewards/near_duplicate_penalty/std': 0.12881141901016235, 'rewards/opening_diversity/mean': 0.8772727251052856, 'rewards/opening_diversity/std': 0.18227215111255646, 'rewards/distinct_2/mean': 0.6947696805000305, 'rewards/distinct_2/std': 0.24594655632972717, 'rewards/total_composite/mean': 0.15561579167842865, 'rewards/total_composite/std': 0.061649389564991, 'reward': 0.15561579167842865, 'reward_std': 0.061649389564991, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16433143615722656, 'sampling/sampling_logp_difference/max': 1.791287899017334, 'sampling/importance_sampling_ratio/min': 0.16674529016017914, 'sampling/importance_sampling_ratio/mean': 1.0020898580551147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8947829157114029, 'clip_ratio/low_mean': 0.04905950138345361, 'clip_ratio/low_min': 0.04905950138345361, 'clip_ratio/high_mean': 0.09206949733197689, 'clip_ratio/high_max': 0.09206949733197689, 'clip_ratio/region_mean': 0.1411289987154305, 'reward_total_mean': 0.15561579167842865, 'reward_meter_mean': 0.6096323132514954, 'reward_meter_std': 0.21670779585838318, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9562499523162842, 'reward_count_floor_std': 0.01767767407000065, 'reward_lexical_plausibility_mean': 0.9634765386581421, 'reward_lexical_plausibility_std': 0.10330399125814438, 'reward_repeat_penalty_mean': 0.523923397064209, 'reward_repeat_penalty_std': 0.3207071125507355, 'reward_repeat_floor_mean': 0.9451455473899841, 'reward_repeat_floor_std': 0.04579712823033333, 'reward_near_duplicate_penalty_mean': 0.8665324449539185, 'reward_near_duplicate_penalty_std': 0.12881141901016235, 'reward_opening_diversity_mean': 0.8772727251052856, 'reward_opening_diversity_std': 0.18227215111255646, 'reward_distinct_2_mean': 0.6947696805000305, 'reward_distinct_2_std': 0.24594655632972717, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.15561579167842865, 'reward_total_composite_std': 0.061649389564991, 'epoch': 0.05}

 30%|██▉       | 981/3300 [3:41:41<9:23:29, 14.58s/it]INFO 04-12 19:40:08 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 982/3300 [3:41:55<9:12:05, 14.29s/it]2026-04-12 19:40:21,640 | INFO | train_grpo_train | metrics_logged mode=train step=982

                                                      
{'loss': -0.0648, 'grad_norm': 2.4111146926879883, 'learning_rate': 7.027272727272728e-06, 'num_tokens': 1744056.0, 'completions/mean_length': 221.0, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 46.400001525878906, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5399047136306763, 'rewards/meter/std': 0.3687581419944763, 'rewards/count_adherence/mean': 0.53125, 'rewards/count_adherence/std': 0.2086307406425476, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.859375, 'rewards/count_floor/std': 0.06258923560380936, 'rewards/lexical_plausibility/mean': 0.8683696985244751, 'rewards/lexical_plausibility/std': 0.18313650786876678, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.91753089427948, 'rewards/repeat_penalty/std': 0.08553789556026459, 'rewards/repeat_floor/mean': 0.9917012453079224, 'rewards/repeat_floor/std': 0.006855242885649204, 'rewards/near_duplicate_penalty/mean': 0.9715892672538757, 'rewards/near_duplicate_penalty/std': 0.029122455045580864, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9759762287139893, 'rewards/distinct_2/std': 0.03664376586675644, 'rewards/total_composite/mean': 0.10129118710756302, 'rewards/total_composite/std': 0.09476771205663681, 'reward': 0.10129118710756302, 'reward_std': 0.09476771205663681, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19497662782669067, 'sampling/sampling_logp_difference/max': 1.427119255065918, 'sampling/importance_sampling_ratio/min': 0.23999930918216705, 'sampling/importance_sampling_ratio/mean': 0.9906262159347534, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0351803004741669, 'clip_ratio/low_mean': 0.013461538590490818, 'clip_ratio/low_min': 0.013461538590490818, 'clip_ratio/high_mean': 0.13716032542288303, 'clip_ratio/high_max': 0.13716032542288303, 'clip_ratio/region_mean': 0.15062186401337385, 'reward_total_mean': 0.10129118710756302, 'reward_meter_mean': 0.5399047136306763, 'reward_meter_std': 0.3687581419944763, 'reward_count_adherence_mean': 0.53125, 'reward_count_adherence_std': 0.2086307406425476, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.859375, 'reward_count_floor_std': 0.06258923560380936, 'reward_lexical_plausibility_mean': 0.8683696985244751, 'reward_lexical_plausibility_std': 0.18313650786876678, 'reward_repeat_penalty_mean': 0.91753089427948, 'reward_repeat_penalty_std': 0.08553789556026459, 'reward_repeat_floor_mean': 0.9917012453079224, 'reward_repeat_floor_std': 0.006855242885649204, 'reward_near_duplicate_penalty_mean': 0.9715892672538757, 'reward_near_duplicate_penalty_std': 0.029122455045580864, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9759762287139893, 'reward_distinct_2_std': 0.03664376586675644, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.10129118710756302, 'reward_total_composite_std': 0.09476771205663681, 'epoch': 0.05}

 30%|██▉       | 982/3300 [3:41:55<9:12:05, 14.29s/it]INFO 04-12 19:40:21 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 983/3300 [3:42:10<9:20:57, 14.53s/it]2026-04-12 19:40:36,717 | INFO | train_grpo_train | metrics_logged mode=train step=983

                                                      
{'loss': -0.1172, 'grad_norm': 3.152083158493042, 'learning_rate': 7.024242424242424e-06, 'num_tokens': 1745985.0, 'completions/mean_length': 257.125, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 104.20000457763672, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.4858367443084717, 'rewards/meter/std': 0.3778248429298401, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.19416078925132751, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.058248236775398254, 'rewards/lexical_plausibility/mean': 0.8242735862731934, 'rewards/lexical_plausibility/std': 0.2443189173936844, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.709581732749939, 'rewards/repeat_penalty/std': 0.33633026480674744, 'rewards/repeat_floor/mean': 0.9660537838935852, 'rewards/repeat_floor/std': 0.04854338616132736, 'rewards/near_duplicate_penalty/mean': 0.9026554226875305, 'rewards/near_duplicate_penalty/std': 0.10118906944990158, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.04419417306780815, 'rewards/distinct_2/mean': 0.7954776287078857, 'rewards/distinct_2/std': 0.3431466221809387, 'rewards/total_composite/mean': 0.15342482924461365, 'rewards/total_composite/std': 0.15523755550384521, 'reward': 0.15342482924461365, 'reward_std': 0.15523755550384521, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1514987349510193, 'sampling/sampling_logp_difference/max': 1.7979998588562012, 'sampling/importance_sampling_ratio/min': 0.1758684664964676, 'sampling/importance_sampling_ratio/mean': 1.0104864835739136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5023180171847343, 'clip_ratio/low_mean': 0.00826446246355772, 'clip_ratio/low_min': 0.00826446246355772, 'clip_ratio/high_mean': 0.0801688376814127, 'clip_ratio/high_max': 0.0801688376814127, 'clip_ratio/region_mean': 0.08843330014497042, 'reward_total_mean': 0.15342482924461365, 'reward_meter_mean': 0.4858367443084717, 'reward_meter_std': 0.3778248429298401, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.19416078925132751, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.058248236775398254, 'reward_lexical_plausibility_mean': 0.8242735862731934, 'reward_lexical_plausibility_std': 0.2443189173936844, 'reward_repeat_penalty_mean': 0.709581732749939, 'reward_repeat_penalty_std': 0.33633026480674744, 'reward_repeat_floor_mean': 0.9660537838935852, 'reward_repeat_floor_std': 0.04854338616132736, 'reward_near_duplicate_penalty_mean': 0.9026554226875305, 'reward_near_duplicate_penalty_std': 0.10118906944990158, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.04419417306780815, 'reward_distinct_2_mean': 0.7954776287078857, 'reward_distinct_2_std': 0.3431466221809387, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.15342482924461365, 'reward_total_composite_std': 0.15523755550384521, 'epoch': 0.05}

 30%|██▉       | 983/3300 [3:42:10<9:20:57, 14.53s/it]INFO 04-12 19:40:36 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 984/3300 [3:42:19<8:21:51, 13.00s/it]2026-04-12 19:40:46,162 | INFO | train_grpo_train | metrics_logged mode=train step=984

                                                      
{'loss': -0.1284, 'grad_norm': 12.46694564819336, 'learning_rate': 7.021212121212122e-06, 'num_tokens': 1747856.0, 'completions/mean_length': 57.875, 'completions/min_length': 40.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9671827554702759, 'rewards/meter/std': 0.03809695690870285, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.7879979014396667, 'rewards/repeat_penalty/std': 0.16411684453487396, 'rewards/repeat_floor/mean': 0.9735134243965149, 'rewards/repeat_floor/std': 0.02506200224161148, 'rewards/near_duplicate_penalty/mean': 0.9086606502532959, 'rewards/near_duplicate_penalty/std': 0.11212155967950821, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8579369783401489, 'rewards/distinct_2/std': 0.1026187464594841, 'rewards/total_composite/mean': 0.3434147238731384, 'rewards/total_composite/std': 0.19259168207645416, 'reward': 0.3434147238731384, 'reward_std': 0.19259168207645416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16055570542812347, 'sampling/sampling_logp_difference/max': 2.009434700012207, 'sampling/importance_sampling_ratio/min': 0.1340644508600235, 'sampling/importance_sampling_ratio/mean': 1.0233654975891113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9670621529221535, 'clip_ratio/low_mean': 0.09576516877859831, 'clip_ratio/low_min': 0.09576516877859831, 'clip_ratio/high_mean': 0.06828634068369865, 'clip_ratio/high_max': 0.06828634068369865, 'clip_ratio/region_mean': 0.16405150946229696, 'reward_total_mean': 0.3434147238731384, 'reward_meter_mean': 0.9671827554702759, 'reward_meter_std': 0.03809695690870285, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7879979014396667, 'reward_repeat_penalty_std': 0.16411684453487396, 'reward_repeat_floor_mean': 0.9735134243965149, 'reward_repeat_floor_std': 0.02506200224161148, 'reward_near_duplicate_penalty_mean': 0.9086606502532959, 'reward_near_duplicate_penalty_std': 0.11212155967950821, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8579369783401489, 'reward_distinct_2_std': 0.1026187464594841, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.3434147238731384, 'reward_total_composite_std': 0.19259168207645416, 'epoch': 0.05}

 30%|██▉       | 984/3300 [3:42:20<8:21:51, 13.00s/it]INFO 04-12 19:40:46 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 985/3300 [3:42:34<8:35:29, 13.36s/it]2026-04-12 19:41:00,360 | INFO | train_grpo_train | metrics_logged mode=train step=985

                                                      
{'loss': -0.0363, 'grad_norm': 5.490208625793457, 'learning_rate': 7.018181818181818e-06, 'num_tokens': 1749594.0, 'completions/mean_length': 117.25, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.857147216796875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8051995038986206, 'rewards/meter/std': 0.27309978008270264, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.05580177903175354, 'rewards/lexical_plausibility/mean': 0.972457766532898, 'rewards/lexical_plausibility/std': 0.07790117710828781, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.2653266191482544, 'rewards/repeat_penalty/mean': 0.9663611650466919, 'rewards/repeat_penalty/std': 0.02206224761903286, 'rewards/repeat_floor/mean': 0.9958232641220093, 'rewards/repeat_floor/std': 0.0021291167940944433, 'rewards/near_duplicate_penalty/mean': 0.9812142848968506, 'rewards/near_duplicate_penalty/std': 0.008108322508633137, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9849016666412354, 'rewards/distinct_2/std': 0.022879112511873245, 'rewards/total_composite/mean': 0.28808528184890747, 'rewards/total_composite/std': 0.2510199248790741, 'reward': 0.28808528184890747, 'reward_std': 0.2510199248790741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1580878496170044, 'sampling/sampling_logp_difference/max': 1.9391846656799316, 'sampling/importance_sampling_ratio/min': 0.14382116496562958, 'sampling/importance_sampling_ratio/mean': 1.0177786350250244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6177263595163822, 'clip_ratio/low_mean': 0.045895135030150414, 'clip_ratio/low_min': 0.045895135030150414, 'clip_ratio/high_mean': 0.08070050645619631, 'clip_ratio/high_max': 0.08070050645619631, 'clip_ratio/region_mean': 0.12659564148634672, 'reward_total_mean': 0.28808528184890747, 'reward_meter_mean': 0.8051995038986206, 'reward_meter_std': 0.27309978008270264, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.05580177903175354, 'reward_lexical_plausibility_mean': 0.972457766532898, 'reward_lexical_plausibility_std': 0.07790117710828781, 'reward_repeat_penalty_mean': 0.9663611650466919, 'reward_repeat_penalty_std': 0.02206224761903286, 'reward_repeat_floor_mean': 0.9958232641220093, 'reward_repeat_floor_std': 0.0021291167940944433, 'reward_near_duplicate_penalty_mean': 0.9812142848968506, 'reward_near_duplicate_penalty_std': 0.008108322508633137, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9849016666412354, 'reward_distinct_2_std': 0.022879112511873245, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.2653266191482544, 'reward_total_composite_mean': 0.28808528184890747, 'reward_total_composite_std': 0.2510199248790741, 'epoch': 0.05}

 30%|██▉       | 985/3300 [3:42:34<8:35:29, 13.36s/it]INFO 04-12 19:41:00 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 986/3300 [3:42:48<8:48:48, 13.71s/it]2026-04-12 19:41:14,891 | INFO | train_grpo_train | metrics_logged mode=train step=986

                                                      
{'loss': -0.1813, 'grad_norm': 3.5879359245300293, 'learning_rate': 7.015151515151516e-06, 'num_tokens': 1752183.0, 'completions/mean_length': 184.625, 'completions/min_length': 116.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 137.85714721679688, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.7311665415763855, 'rewards/meter/std': 0.33589377999305725, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.017359109595417976, 'rewards/lexical_plausibility/mean': 0.946574866771698, 'rewards/lexical_plausibility/std': 0.15110905468463898, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.3428924083709717, 'rewards/repeat_penalty/std': 0.27344465255737305, 'rewards/repeat_floor/mean': 0.9182778000831604, 'rewards/repeat_floor/std': 0.04620883986353874, 'rewards/near_duplicate_penalty/mean': 0.8172633051872253, 'rewards/near_duplicate_penalty/std': 0.10879214107990265, 'rewards/opening_diversity/mean': 0.9143544435501099, 'rewards/opening_diversity/std': 0.12935413420200348, 'rewards/distinct_2/mean': 0.4536336660385132, 'rewards/distinct_2/std': 0.2980527877807617, 'rewards/total_composite/mean': 0.09469848871231079, 'rewards/total_composite/std': 0.044970057904720306, 'reward': 0.09469848871231079, 'reward_std': 0.044970057904720306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13201110064983368, 'sampling/sampling_logp_difference/max': 2.124917507171631, 'sampling/importance_sampling_ratio/min': 0.11944282054901123, 'sampling/importance_sampling_ratio/mean': 1.0128896236419678, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7170674204826355, 'clip_ratio/low_mean': 0.024541988968849182, 'clip_ratio/low_min': 0.024541988968849182, 'clip_ratio/high_mean': 0.07692741509526968, 'clip_ratio/high_max': 0.07692741509526968, 'clip_ratio/region_mean': 0.10146940406411886, 'reward_total_mean': 0.09469848871231079, 'reward_meter_mean': 0.7311665415763855, 'reward_meter_std': 0.33589377999305725, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.017359109595417976, 'reward_lexical_plausibility_mean': 0.946574866771698, 'reward_lexical_plausibility_std': 0.15110905468463898, 'reward_repeat_penalty_mean': 0.3428924083709717, 'reward_repeat_penalty_std': 0.27344465255737305, 'reward_repeat_floor_mean': 0.9182778000831604, 'reward_repeat_floor_std': 0.04620883986353874, 'reward_near_duplicate_penalty_mean': 0.8172633051872253, 'reward_near_duplicate_penalty_std': 0.10879214107990265, 'reward_opening_diversity_mean': 0.9143544435501099, 'reward_opening_diversity_std': 0.12935413420200348, 'reward_distinct_2_mean': 0.4536336660385132, 'reward_distinct_2_std': 0.2980527877807617, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.09469848871231079, 'reward_total_composite_std': 0.044970057904720306, 'epoch': 0.05}

 30%|██▉       | 986/3300 [3:42:48<8:48:48, 13.71s/it]INFO 04-12 19:41:15 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 987/3300 [3:42:56<7:45:23, 12.07s/it]2026-04-12 19:41:23,135 | INFO | train_grpo_train | metrics_logged mode=train step=987

                                                      
{'loss': 0.1011, 'grad_norm': 14.259634971618652, 'learning_rate': 7.0121212121212126e-06, 'num_tokens': 1753798.0, 'completions/mean_length': 43.875, 'completions/min_length': 38.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8553590178489685, 'rewards/meter/std': 0.21360167860984802, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4287499785423279, 'rewards/judge_quality/std': 0.20746342837810516, 'rewards/repeat_penalty/mean': 0.8162665367126465, 'rewards/repeat_penalty/std': 0.14816901087760925, 'rewards/repeat_floor/mean': 0.9770795106887817, 'rewards/repeat_floor/std': 0.017852913588285446, 'rewards/near_duplicate_penalty/mean': 0.9080779552459717, 'rewards/near_duplicate_penalty/std': 0.06150315701961517, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9193644523620605, 'rewards/distinct_2/std': 0.05791359022259712, 'rewards/total_composite/mean': 0.3761236071586609, 'rewards/total_composite/std': 0.2318309098482132, 'reward': 0.3761236071586609, 'reward_std': 0.2318309098482132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16257882118225098, 'sampling/sampling_logp_difference/max': 1.484790325164795, 'sampling/importance_sampling_ratio/min': 0.22654984891414642, 'sampling/importance_sampling_ratio/mean': 1.0188466310501099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0776085332036018, 'clip_ratio/low_mean': 0.0735879885032773, 'clip_ratio/low_min': 0.0735879885032773, 'clip_ratio/high_mean': 0.07214896846562624, 'clip_ratio/high_max': 0.07214896846562624, 'clip_ratio/region_mean': 0.14573695696890354, 'reward_total_mean': 0.3761236071586609, 'reward_meter_mean': 0.8553590178489685, 'reward_meter_std': 0.21360167860984802, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8162665367126465, 'reward_repeat_penalty_std': 0.14816901087760925, 'reward_repeat_floor_mean': 0.9770795106887817, 'reward_repeat_floor_std': 0.017852913588285446, 'reward_near_duplicate_penalty_mean': 0.9080779552459717, 'reward_near_duplicate_penalty_std': 0.06150315701961517, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9193644523620605, 'reward_distinct_2_std': 0.05791359022259712, 'reward_judge_quality_mean': 0.4287499785423279, 'reward_judge_quality_std': 0.20746342837810516, 'reward_total_composite_mean': 0.3761236071586609, 'reward_total_composite_std': 0.2318309098482132, 'epoch': 0.05}

 30%|██▉       | 987/3300 [3:42:56<7:45:23, 12.07s/it]INFO 04-12 19:41:23 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 988/3300 [3:43:11<8:10:29, 12.73s/it]2026-04-12 19:41:37,397 | INFO | train_grpo_train | metrics_logged mode=train step=988

                                                      
{'loss': -0.0333, 'grad_norm': 5.760228157043457, 'learning_rate': 7.00909090909091e-06, 'num_tokens': 1755222.0, 'completions/mean_length': 99.0, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.7354631423950195, 'rewards/meter/std': 0.4045969843864441, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9584840536117554, 'rewards/lexical_plausibility/std': 0.1174248680472374, 'rewards/judge_quality/mean': 0.38374999165534973, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.8071895241737366, 'rewards/repeat_penalty/std': 0.1083652675151825, 'rewards/repeat_floor/mean': 0.9800229072570801, 'rewards/repeat_floor/std': 0.014075199142098427, 'rewards/near_duplicate_penalty/mean': 0.9385877251625061, 'rewards/near_duplicate_penalty/std': 0.0692981407046318, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9220526218414307, 'rewards/distinct_2/std': 0.07810666412115097, 'rewards/total_composite/mean': 0.3047555088996887, 'rewards/total_composite/std': 0.26785439252853394, 'reward': 0.3047555088996887, 'reward_std': 0.26785439252853394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1661999672651291, 'sampling/sampling_logp_difference/max': 1.9121756553649902, 'sampling/importance_sampling_ratio/min': 0.14775855839252472, 'sampling/importance_sampling_ratio/mean': 0.9966478943824768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7666777074337006, 'clip_ratio/low_mean': 0.06960220634937286, 'clip_ratio/low_min': 0.06960220634937286, 'clip_ratio/high_mean': 0.09643602184951305, 'clip_ratio/high_max': 0.09643602184951305, 'clip_ratio/region_mean': 0.16603822819888592, 'reward_total_mean': 0.3047555088996887, 'reward_meter_mean': 0.7354631423950195, 'reward_meter_std': 0.4045969843864441, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9584840536117554, 'reward_lexical_plausibility_std': 0.1174248680472374, 'reward_repeat_penalty_mean': 0.8071895241737366, 'reward_repeat_penalty_std': 0.1083652675151825, 'reward_repeat_floor_mean': 0.9800229072570801, 'reward_repeat_floor_std': 0.014075199142098427, 'reward_near_duplicate_penalty_mean': 0.9385877251625061, 'reward_near_duplicate_penalty_std': 0.0692981407046318, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9220526218414307, 'reward_distinct_2_std': 0.07810666412115097, 'reward_judge_quality_mean': 0.38374999165534973, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.3047555088996887, 'reward_total_composite_std': 0.26785439252853394, 'epoch': 0.05}

 30%|██▉       | 988/3300 [3:43:11<8:10:29, 12.73s/it]INFO 04-12 19:41:37 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 989/3300 [3:43:19<7:24:20, 11.54s/it]2026-04-12 19:41:46,154 | INFO | train_grpo_train | metrics_logged mode=train step=989

                                                      
{'loss': 0.0084, 'grad_norm': 14.503719329833984, 'learning_rate': 7.006060606060606e-06, 'num_tokens': 1757201.0, 'completions/mean_length': 59.375, 'completions/min_length': 52.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.753119945526123, 'rewards/meter/std': 0.35611605644226074, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.7723912000656128, 'rewards/repeat_penalty/std': 0.23510518670082092, 'rewards/repeat_floor/mean': 0.9704902172088623, 'rewards/repeat_floor/std': 0.03147847205400467, 'rewards/near_duplicate_penalty/mean': 0.8949650526046753, 'rewards/near_duplicate_penalty/std': 0.09759186208248138, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.847171425819397, 'rewards/distinct_2/std': 0.20438216626644135, 'rewards/total_composite/mean': 0.2648583650588989, 'rewards/total_composite/std': 0.17425023019313812, 'reward': 0.2648583650588989, 'reward_std': 0.17425024509429932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14865384995937347, 'sampling/sampling_logp_difference/max': 2.655883312225342, 'sampling/importance_sampling_ratio/min': 0.07023677229881287, 'sampling/importance_sampling_ratio/mean': 0.9875383973121643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8590160459280014, 'clip_ratio/low_mean': 0.04135649371892214, 'clip_ratio/low_min': 0.04135649371892214, 'clip_ratio/high_mean': 0.10494442097842693, 'clip_ratio/high_max': 0.10494442097842693, 'clip_ratio/region_mean': 0.14630091469734907, 'reward_total_mean': 0.2648583650588989, 'reward_meter_mean': 0.753119945526123, 'reward_meter_std': 0.35611605644226074, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.7723912000656128, 'reward_repeat_penalty_std': 0.23510518670082092, 'reward_repeat_floor_mean': 0.9704902172088623, 'reward_repeat_floor_std': 0.03147847205400467, 'reward_near_duplicate_penalty_mean': 0.8949650526046753, 'reward_near_duplicate_penalty_std': 0.09759186208248138, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.847171425819397, 'reward_distinct_2_std': 0.20438216626644135, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.2648583650588989, 'reward_total_composite_std': 0.17425023019313812, 'epoch': 0.05}

 30%|██▉       | 989/3300 [3:43:20<7:24:20, 11.54s/it]INFO 04-12 19:41:46 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 990/3300 [3:43:29<7:00:44, 10.93s/it]2026-04-12 19:41:55,664 | INFO | train_grpo_train | metrics_logged mode=train step=990

                                                      
{'loss': 0.0212, 'grad_norm': 10.293787956237793, 'learning_rate': 7.0030303030303035e-06, 'num_tokens': 1759252.0, 'completions/mean_length': 76.375, 'completions/min_length': 52.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.3753584921360016, 'rewards/meter/std': 0.3555986285209656, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.6902063488960266, 'rewards/repeat_penalty/std': 0.09471764415502548, 'rewards/repeat_floor/mean': 0.9640001058578491, 'rewards/repeat_floor/std': 0.011091821826994419, 'rewards/near_duplicate_penalty/mean': 0.9010778665542603, 'rewards/near_duplicate_penalty/std': 0.03414170444011688, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7648714780807495, 'rewards/distinct_2/std': 0.09027978032827377, 'rewards/total_composite/mean': 0.1398789882659912, 'rewards/total_composite/std': 0.14445915818214417, 'reward': 0.1398789882659912, 'reward_std': 0.14445915818214417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14950519800186157, 'sampling/sampling_logp_difference/max': 4.399341583251953, 'sampling/importance_sampling_ratio/min': 0.012285426259040833, 'sampling/importance_sampling_ratio/mean': 1.0108695030212402, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.802264578640461, 'clip_ratio/low_mean': 0.08543698769062757, 'clip_ratio/low_min': 0.08543698769062757, 'clip_ratio/high_mean': 0.05243348143994808, 'clip_ratio/high_max': 0.05243348143994808, 'clip_ratio/region_mean': 0.13787046913057566, 'reward_total_mean': 0.1398789882659912, 'reward_meter_mean': 0.3753584921360016, 'reward_meter_std': 0.3555986285209656, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6902063488960266, 'reward_repeat_penalty_std': 0.09471764415502548, 'reward_repeat_floor_mean': 0.9640001058578491, 'reward_repeat_floor_std': 0.011091821826994419, 'reward_near_duplicate_penalty_mean': 0.9010778665542603, 'reward_near_duplicate_penalty_std': 0.03414170444011688, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7648714780807495, 'reward_distinct_2_std': 0.09027978032827377, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.1398789882659912, 'reward_total_composite_std': 0.14445915818214417, 'epoch': 0.05}

 30%|███       | 990/3300 [3:43:29<7:00:44, 10.93s/it]INFO 04-12 19:41:55 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 991/3300 [3:43:43<7:40:23, 11.96s/it]2026-04-12 19:42:10,041 | INFO | train_grpo_train | metrics_logged mode=train step=991

                                                      
{'loss': -0.0592, 'grad_norm': 6.651369094848633, 'learning_rate': 7e-06, 'num_tokens': 1761192.0, 'completions/mean_length': 117.5, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.142860412597656, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.7795392870903015, 'rewards/meter/std': 0.21611040830612183, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.07763238251209259, 'rewards/lexical_plausibility/mean': 0.9535982608795166, 'rewards/lexical_plausibility/std': 0.11648311465978622, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.6481847167015076, 'rewards/repeat_penalty/std': 0.31310275197029114, 'rewards/repeat_floor/mean': 0.9519377946853638, 'rewards/repeat_floor/std': 0.05035944655537605, 'rewards/near_duplicate_penalty/mean': 0.8536094427108765, 'rewards/near_duplicate_penalty/std': 0.13116014003753662, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.18600596487522125, 'rewards/distinct_2/mean': 0.7724599838256836, 'rewards/distinct_2/std': 0.25449228286743164, 'rewards/total_composite/mean': 0.18278560042381287, 'rewards/total_composite/std': 0.1323985755443573, 'reward': 0.18278560042381287, 'reward_std': 0.1323985755443573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16739000380039215, 'sampling/sampling_logp_difference/max': 1.6728935241699219, 'sampling/importance_sampling_ratio/min': 0.18770314753055573, 'sampling/importance_sampling_ratio/mean': 0.9968866109848022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.926862820982933, 'clip_ratio/low_mean': 0.033057416789233685, 'clip_ratio/low_min': 0.033057416789233685, 'clip_ratio/high_mean': 0.07861966080963612, 'clip_ratio/high_max': 0.07861966080963612, 'clip_ratio/region_mean': 0.1116770775988698, 'reward_total_mean': 0.18278560042381287, 'reward_meter_mean': 0.7795392870903015, 'reward_meter_std': 0.21611040830612183, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.07763238251209259, 'reward_lexical_plausibility_mean': 0.9535982608795166, 'reward_lexical_plausibility_std': 0.11648311465978622, 'reward_repeat_penalty_mean': 0.6481847167015076, 'reward_repeat_penalty_std': 0.31310275197029114, 'reward_repeat_floor_mean': 0.9519377946853638, 'reward_repeat_floor_std': 0.05035944655537605, 'reward_near_duplicate_penalty_mean': 0.8536094427108765, 'reward_near_duplicate_penalty_std': 0.13116014003753662, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.18600596487522125, 'reward_distinct_2_mean': 0.7724599838256836, 'reward_distinct_2_std': 0.25449228286743164, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.18278560042381287, 'reward_total_composite_std': 0.1323985755443573, 'epoch': 0.05}

 30%|███       | 991/3300 [3:43:43<7:40:23, 11.96s/it]INFO 04-12 19:42:10 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 992/3300 [3:43:51<6:48:54, 10.63s/it]2026-04-12 19:42:17,559 | INFO | train_grpo_train | metrics_logged mode=train step=992

                                                      
{'loss': 0.3272, 'grad_norm': 16.29084014892578, 'learning_rate': 6.996969696969698e-06, 'num_tokens': 1762630.0, 'completions/mean_length': 26.75, 'completions/min_length': 15.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9486788511276245, 'rewards/meter/std': 0.07560297101736069, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.15526476502418518, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5712500214576721, 'rewards/judge_quality/std': 0.4023657739162445, 'rewards/repeat_penalty/mean': 0.6705222129821777, 'rewards/repeat_penalty/std': 0.19232770800590515, 'rewards/repeat_floor/mean': 0.9719693064689636, 'rewards/repeat_floor/std': 0.02538437396287918, 'rewards/near_duplicate_penalty/mean': 0.9534604549407959, 'rewards/near_duplicate_penalty/std': 0.076200470328331, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9119470119476318, 'rewards/distinct_2/std': 0.17741413414478302, 'rewards/total_composite/mean': 0.5146055817604065, 'rewards/total_composite/std': 0.4119769334793091, 'reward': 0.5146055817604065, 'reward_std': 0.4119769036769867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12871712446212769, 'sampling/sampling_logp_difference/max': 1.6771574020385742, 'sampling/importance_sampling_ratio/min': 0.18690451979637146, 'sampling/importance_sampling_ratio/mean': 1.0465173721313477, 'sampling/importance_sampling_ratio/max': 1.9479018449783325, 'entropy': 0.7988493666052818, 'clip_ratio/low_mean': 0.08833384374156594, 'clip_ratio/low_min': 0.08833384374156594, 'clip_ratio/high_mean': 0.047437612898647785, 'clip_ratio/high_max': 0.047437612898647785, 'clip_ratio/region_mean': 0.13577145664021373, 'reward_total_mean': 0.5146055817604065, 'reward_meter_mean': 0.9486788511276245, 'reward_meter_std': 0.07560297101736069, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.15526476502418518, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6705222129821777, 'reward_repeat_penalty_std': 0.19232770800590515, 'reward_repeat_floor_mean': 0.9719693064689636, 'reward_repeat_floor_std': 0.02538437396287918, 'reward_near_duplicate_penalty_mean': 0.9534604549407959, 'reward_near_duplicate_penalty_std': 0.076200470328331, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9119470119476318, 'reward_distinct_2_std': 0.17741413414478302, 'reward_judge_quality_mean': 0.5712500214576721, 'reward_judge_quality_std': 0.4023657739162445, 'reward_total_composite_mean': 0.5146055817604065, 'reward_total_composite_std': 0.4119769334793091, 'epoch': 0.05}

 30%|███       | 992/3300 [3:43:51<6:48:54, 10.63s/it]INFO 04-12 19:42:17 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 993/3300 [3:43:59<6:19:05,  9.86s/it]2026-04-12 19:42:25,620 | INFO | train_grpo_train | metrics_logged mode=train step=993

                                                      
{'loss': 0.0287, 'grad_norm': 15.378118515014648, 'learning_rate': 6.993939393939394e-06, 'num_tokens': 1764121.0, 'completions/mean_length': 40.375, 'completions/min_length': 36.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.47582417726516724, 'rewards/meter/std': 0.46113091707229614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9162414073944092, 'rewards/repeat_penalty/std': 0.08564253151416779, 'rewards/repeat_floor/mean': 0.9898476600646973, 'rewards/repeat_floor/std': 0.011220183223485947, 'rewards/near_duplicate_penalty/mean': 0.9616491794586182, 'rewards/near_duplicate_penalty/std': 0.051408372819423676, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9511142373085022, 'rewards/distinct_2/std': 0.045760173350572586, 'rewards/total_composite/mean': 0.370554655790329, 'rewards/total_composite/std': 0.3914250433444977, 'reward': 0.370554655790329, 'reward_std': 0.39142507314682007, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19196295738220215, 'sampling/sampling_logp_difference/max': 1.846895694732666, 'sampling/importance_sampling_ratio/min': 0.15772603452205658, 'sampling/importance_sampling_ratio/mean': 0.9863186478614807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9971630051732063, 'clip_ratio/low_mean': 0.08436782564967871, 'clip_ratio/low_min': 0.08436782564967871, 'clip_ratio/high_mean': 0.08469890104606748, 'clip_ratio/high_max': 0.08469890104606748, 'clip_ratio/region_mean': 0.16906672669574618, 'reward_total_mean': 0.370554655790329, 'reward_meter_mean': 0.47582417726516724, 'reward_meter_std': 0.46113091707229614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9162414073944092, 'reward_repeat_penalty_std': 0.08564253151416779, 'reward_repeat_floor_mean': 0.9898476600646973, 'reward_repeat_floor_std': 0.011220183223485947, 'reward_near_duplicate_penalty_mean': 0.9616491794586182, 'reward_near_duplicate_penalty_std': 0.051408372819423676, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9511142373085022, 'reward_distinct_2_std': 0.045760173350572586, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.370554655790329, 'reward_total_composite_std': 0.3914250433444977, 'epoch': 0.05}

 30%|███       | 993/3300 [3:43:59<6:19:05,  9.86s/it]INFO 04-12 19:42:25 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 994/3300 [3:44:13<7:04:22, 11.04s/it]2026-04-12 19:42:39,423 | INFO | train_grpo_train | metrics_logged mode=train step=994

                                                      
{'loss': -0.1588, 'grad_norm': 3.641510009765625, 'learning_rate': 6.990909090909092e-06, 'num_tokens': 1766208.0, 'completions/mean_length': 215.875, 'completions/min_length': 103.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 117.16667175292969, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.7931424379348755, 'rewards/meter/std': 0.23350565135478973, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.10681165754795074, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9562499523162842, 'rewards/count_floor/std': 0.03204350546002388, 'rewards/lexical_plausibility/mean': 0.9017550945281982, 'rewards/lexical_plausibility/std': 0.18226182460784912, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.11952286213636398, 'rewards/repeat_penalty/mean': 0.7025879621505737, 'rewards/repeat_penalty/std': 0.12204108387231827, 'rewards/repeat_floor/mean': 0.9696890115737915, 'rewards/repeat_floor/std': 0.01690717414021492, 'rewards/near_duplicate_penalty/mean': 0.919126033782959, 'rewards/near_duplicate_penalty/std': 0.04020199924707413, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7980012893676758, 'rewards/distinct_2/std': 0.13514964282512665, 'rewards/total_composite/mean': 0.19511479139328003, 'rewards/total_composite/std': 0.1260329633951187, 'reward': 0.19511479139328003, 'reward_std': 0.12603294849395752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14866895973682404, 'sampling/sampling_logp_difference/max': 2.5168070793151855, 'sampling/importance_sampling_ratio/min': 0.08071692287921906, 'sampling/importance_sampling_ratio/mean': 1.002547025680542, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6413288563489914, 'clip_ratio/low_mean': 0.03732842393219471, 'clip_ratio/low_min': 0.03732842393219471, 'clip_ratio/high_mean': 0.07037896942347288, 'clip_ratio/high_max': 0.07037896942347288, 'clip_ratio/region_mean': 0.10770739335566759, 'reward_total_mean': 0.19511479139328003, 'reward_meter_mean': 0.7931424379348755, 'reward_meter_std': 0.23350565135478973, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.10681165754795074, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9562499523162842, 'reward_count_floor_std': 0.03204350546002388, 'reward_lexical_plausibility_mean': 0.9017550945281982, 'reward_lexical_plausibility_std': 0.18226182460784912, 'reward_repeat_penalty_mean': 0.7025879621505737, 'reward_repeat_penalty_std': 0.12204108387231827, 'reward_repeat_floor_mean': 0.9696890115737915, 'reward_repeat_floor_std': 0.01690717414021492, 'reward_near_duplicate_penalty_mean': 0.919126033782959, 'reward_near_duplicate_penalty_std': 0.04020199924707413, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7980012893676758, 'reward_distinct_2_std': 0.13514964282512665, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.11952286213636398, 'reward_total_composite_mean': 0.19511479139328003, 'reward_total_composite_std': 0.1260329633951187, 'epoch': 0.05}

 30%|███       | 994/3300 [3:44:13<7:04:22, 11.04s/it]INFO 04-12 19:42:39 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 995/3300 [3:44:27<7:46:03, 12.13s/it]2026-04-12 19:42:54,100 | INFO | train_grpo_train | metrics_logged mode=train step=995

                                                      
{'loss': -0.0701, 'grad_norm': 6.266446113586426, 'learning_rate': 6.987878787878788e-06, 'num_tokens': 1768113.0, 'completions/mean_length': 129.125, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.42857360839844, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.534123420715332, 'rewards/meter/std': 0.4553905427455902, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.06943651288747787, 'rewards/lexical_plausibility/mean': 0.9521560668945312, 'rewards/lexical_plausibility/std': 0.1353231817483902, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.18516400456428528, 'rewards/repeat_penalty/mean': 0.7730696797370911, 'rewards/repeat_penalty/std': 0.16656331717967987, 'rewards/repeat_floor/mean': 0.9777371883392334, 'rewards/repeat_floor/std': 0.015646226704120636, 'rewards/near_duplicate_penalty/mean': 0.9473599195480347, 'rewards/near_duplicate_penalty/std': 0.04339979588985443, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.8664110898971558, 'rewards/distinct_2/std': 0.1253872960805893, 'rewards/total_composite/mean': 0.21720297634601593, 'rewards/total_composite/std': 0.20966914296150208, 'reward': 0.21720297634601593, 'reward_std': 0.20966912806034088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1831740140914917, 'sampling/sampling_logp_difference/max': 2.044797420501709, 'sampling/importance_sampling_ratio/min': 0.16297850012779236, 'sampling/importance_sampling_ratio/mean': 0.9990149736404419, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.831375740468502, 'clip_ratio/low_mean': 0.04941620957106352, 'clip_ratio/low_min': 0.04941620957106352, 'clip_ratio/high_mean': 0.08083320036530495, 'clip_ratio/high_max': 0.08083320036530495, 'clip_ratio/region_mean': 0.13024940993636847, 'reward_total_mean': 0.21720297634601593, 'reward_meter_mean': 0.534123420715332, 'reward_meter_std': 0.4553905427455902, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.06943651288747787, 'reward_lexical_plausibility_mean': 0.9521560668945312, 'reward_lexical_plausibility_std': 0.1353231817483902, 'reward_repeat_penalty_mean': 0.7730696797370911, 'reward_repeat_penalty_std': 0.16656331717967987, 'reward_repeat_floor_mean': 0.9777371883392334, 'reward_repeat_floor_std': 0.015646226704120636, 'reward_near_duplicate_penalty_mean': 0.9473599195480347, 'reward_near_duplicate_penalty_std': 0.04339979588985443, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.8664110898971558, 'reward_distinct_2_std': 0.1253872960805893, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.18516400456428528, 'reward_total_composite_mean': 0.21720297634601593, 'reward_total_composite_std': 0.20966914296150208, 'epoch': 0.05}

 30%|███       | 995/3300 [3:44:27<7:46:03, 12.13s/it]INFO 04-12 19:42:54 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 996/3300 [3:44:42<8:14:34, 12.88s/it]2026-04-12 19:43:08,725 | INFO | train_grpo_train | metrics_logged mode=train step=996

                                                      
{'loss': -0.1379, 'grad_norm': 4.341999053955078, 'learning_rate': 6.984848484848485e-06, 'num_tokens': 1770580.0, 'completions/mean_length': 185.375, 'completions/min_length': 119.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 138.71429443359375, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.8178333044052124, 'rewards/meter/std': 0.33430203795433044, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.2645200192928314, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9357142448425293, 'rewards/count_floor/std': 0.07935599982738495, 'rewards/lexical_plausibility/mean': 0.9604979753494263, 'rewards/lexical_plausibility/std': 0.11172851920127869, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.5520014762878418, 'rewards/repeat_penalty/std': 0.21984051167964935, 'rewards/repeat_floor/mean': 0.9548934102058411, 'rewards/repeat_floor/std': 0.023419057950377464, 'rewards/near_duplicate_penalty/mean': 0.9242173433303833, 'rewards/near_duplicate_penalty/std': 0.03988896310329437, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.6459535360336304, 'rewards/distinct_2/std': 0.23264984786510468, 'rewards/total_composite/mean': 0.25405704975128174, 'rewards/total_composite/std': 0.13310401141643524, 'reward': 0.25405704975128174, 'reward_std': 0.13310401141643524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16121190786361694, 'sampling/sampling_logp_difference/max': 2.8203842639923096, 'sampling/importance_sampling_ratio/min': 0.059583041816949844, 'sampling/importance_sampling_ratio/mean': 1.0124974250793457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8249397203326225, 'clip_ratio/low_mean': 0.028817596845328808, 'clip_ratio/low_min': 0.028817596845328808, 'clip_ratio/high_mean': 0.08488169685006142, 'clip_ratio/high_max': 0.08488169685006142, 'clip_ratio/region_mean': 0.11369929369539022, 'reward_total_mean': 0.25405704975128174, 'reward_meter_mean': 0.8178333044052124, 'reward_meter_std': 0.33430203795433044, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.2645200192928314, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9357142448425293, 'reward_count_floor_std': 0.07935599982738495, 'reward_lexical_plausibility_mean': 0.9604979753494263, 'reward_lexical_plausibility_std': 0.11172851920127869, 'reward_repeat_penalty_mean': 0.5520014762878418, 'reward_repeat_penalty_std': 0.21984051167964935, 'reward_repeat_floor_mean': 0.9548934102058411, 'reward_repeat_floor_std': 0.023419057950377464, 'reward_near_duplicate_penalty_mean': 0.9242173433303833, 'reward_near_duplicate_penalty_std': 0.03988896310329437, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.6459535360336304, 'reward_distinct_2_std': 0.23264984786510468, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.25405704975128174, 'reward_total_composite_std': 0.13310401141643524, 'epoch': 0.05}

 30%|███       | 996/3300 [3:44:42<8:14:34, 12.88s/it]INFO 04-12 19:43:08 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 997/3300 [3:44:52<7:42:49, 12.06s/it]2026-04-12 19:43:18,864 | INFO | train_grpo_train | metrics_logged mode=train step=997

                                                      
{'loss': 0.0076, 'grad_norm': 23.889097213745117, 'learning_rate': 6.981818181818183e-06, 'num_tokens': 1772183.0, 'completions/mean_length': 40.375, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9363147020339966, 'rewards/meter/std': 0.07850109040737152, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.7469356060028076, 'rewards/repeat_penalty/std': 0.22237913310527802, 'rewards/repeat_floor/mean': 0.9729578495025635, 'rewards/repeat_floor/std': 0.021939516067504883, 'rewards/near_duplicate_penalty/mean': 0.9223943948745728, 'rewards/near_duplicate_penalty/std': 0.059935662895441055, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8705416321754456, 'rewards/distinct_2/std': 0.10770028084516525, 'rewards/total_composite/mean': 0.30584216117858887, 'rewards/total_composite/std': 0.07497670501470566, 'reward': 0.30584216117858887, 'reward_std': 0.07497670501470566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1816522479057312, 'sampling/sampling_logp_difference/max': 1.8009929656982422, 'sampling/importance_sampling_ratio/min': 0.16513483226299286, 'sampling/importance_sampling_ratio/mean': 1.0063318014144897, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1741241961717606, 'clip_ratio/low_mean': 0.04714912362396717, 'clip_ratio/low_min': 0.04714912362396717, 'clip_ratio/high_mean': 0.14426992181688547, 'clip_ratio/high_max': 0.14426992181688547, 'clip_ratio/region_mean': 0.19141904544085264, 'reward_total_mean': 0.30584216117858887, 'reward_meter_mean': 0.9363147020339966, 'reward_meter_std': 0.07850109040737152, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7469356060028076, 'reward_repeat_penalty_std': 0.22237913310527802, 'reward_repeat_floor_mean': 0.9729578495025635, 'reward_repeat_floor_std': 0.021939516067504883, 'reward_near_duplicate_penalty_mean': 0.9223943948745728, 'reward_near_duplicate_penalty_std': 0.059935662895441055, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8705416321754456, 'reward_distinct_2_std': 0.10770028084516525, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.30584216117858887, 'reward_total_composite_std': 0.07497670501470566, 'epoch': 0.05}

 30%|███       | 997/3300 [3:44:52<7:42:49, 12.06s/it]INFO 04-12 19:43:19 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 998/3300 [3:45:06<8:00:31, 12.52s/it]2026-04-12 19:43:32,477 | INFO | train_grpo_train | metrics_logged mode=train step=998

                                                      
{'loss': -0.0395, 'grad_norm': 5.007837772369385, 'learning_rate': 6.978787878787879e-06, 'num_tokens': 1773550.0, 'completions/mean_length': 84.875, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.85714340209961, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6104247570037842, 'rewards/meter/std': 0.40084177255630493, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9202219843864441, 'rewards/lexical_plausibility/std': 0.20158244669437408, 'rewards/judge_quality/mean': 0.643750011920929, 'rewards/judge_quality/std': 0.3899061381816864, 'rewards/repeat_penalty/mean': 0.6620035171508789, 'rewards/repeat_penalty/std': 0.267974317073822, 'rewards/repeat_floor/mean': 0.9725660085678101, 'rewards/repeat_floor/std': 0.031774453818798065, 'rewards/near_duplicate_penalty/mean': 0.9630855321884155, 'rewards/near_duplicate_penalty/std': 0.0787380039691925, 'rewards/opening_diversity/mean': 0.6875, 'rewards/opening_diversity/std': 0.29124119877815247, 'rewards/distinct_2/mean': 0.9650349617004395, 'rewards/distinct_2/std': 0.06474266946315765, 'rewards/total_composite/mean': 0.3806096911430359, 'rewards/total_composite/std': 0.328918993473053, 'reward': 0.3806096911430359, 'reward_std': 0.328918993473053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18157890439033508, 'sampling/sampling_logp_difference/max': 1.5199534893035889, 'sampling/importance_sampling_ratio/min': 0.21872207522392273, 'sampling/importance_sampling_ratio/mean': 1.0177454948425293, 'sampling/importance_sampling_ratio/max': 1.9309744834899902, 'entropy': 1.1548809856176376, 'clip_ratio/low_mean': 0.05364583386108279, 'clip_ratio/low_min': 0.05364583386108279, 'clip_ratio/high_mean': 0.08059417875483632, 'clip_ratio/high_max': 0.08059417875483632, 'clip_ratio/region_mean': 0.1342400126159191, 'reward_total_mean': 0.3806096911430359, 'reward_meter_mean': 0.6104247570037842, 'reward_meter_std': 0.40084177255630493, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9202219843864441, 'reward_lexical_plausibility_std': 0.20158244669437408, 'reward_repeat_penalty_mean': 0.6620035171508789, 'reward_repeat_penalty_std': 0.267974317073822, 'reward_repeat_floor_mean': 0.9725660085678101, 'reward_repeat_floor_std': 0.031774453818798065, 'reward_near_duplicate_penalty_mean': 0.9630855321884155, 'reward_near_duplicate_penalty_std': 0.0787380039691925, 'reward_opening_diversity_mean': 0.6875, 'reward_opening_diversity_std': 0.29124119877815247, 'reward_distinct_2_mean': 0.9650349617004395, 'reward_distinct_2_std': 0.06474266946315765, 'reward_judge_quality_mean': 0.643750011920929, 'reward_judge_quality_std': 0.3899061381816864, 'reward_total_composite_mean': 0.3806096911430359, 'reward_total_composite_std': 0.328918993473053, 'epoch': 0.05}

 30%|███       | 998/3300 [3:45:06<8:00:31, 12.52s/it]INFO 04-12 19:43:32 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 999/3300 [3:45:14<7:14:59, 11.34s/it]2026-04-12 19:43:41,059 | INFO | train_grpo_train | metrics_logged mode=train step=999

                                                      
{'loss': 0.0568, 'grad_norm': 27.54399299621582, 'learning_rate': 6.975757575757577e-06, 'num_tokens': 1774933.0, 'completions/mean_length': 14.875, 'completions/min_length': 12.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 14.875, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.7081332206726074, 'rewards/meter/std': 0.3522411286830902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9898897409439087, 'rewards/lexical_plausibility/std': 0.028596237301826477, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.2653266191482544, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19429558515548706, 'rewards/total_composite/std': 0.13851676881313324, 'reward': 0.19429558515548706, 'reward_std': 0.13851676881313324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1672811508178711, 'sampling/sampling_logp_difference/max': 1.6584110260009766, 'sampling/importance_sampling_ratio/min': 0.1904413402080536, 'sampling/importance_sampling_ratio/mean': 1.0127928256988525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7742867171764374, 'clip_ratio/low_mean': 0.06736111221835017, 'clip_ratio/low_min': 0.06736111221835017, 'clip_ratio/high_mean': 0.0651785722002387, 'clip_ratio/high_max': 0.0651785722002387, 'clip_ratio/region_mean': 0.13253968441858888, 'reward_total_mean': 0.19429558515548706, 'reward_meter_mean': 0.7081332206726074, 'reward_meter_std': 0.3522411286830902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9898897409439087, 'reward_lexical_plausibility_std': 0.028596237301826477, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.2653266191482544, 'reward_total_composite_mean': 0.19429558515548706, 'reward_total_composite_std': 0.13851676881313324, 'epoch': 0.05}

 30%|███       | 999/3300 [3:45:14<7:14:59, 11.34s/it]INFO 04-12 19:43:41 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1000/3300 [3:45:24<6:50:50, 10.72s/it]2026-04-12 19:43:50,322 | INFO | train_grpo_train | metrics_logged mode=train step=1000

                                                       
{'loss': 0.1831, 'grad_norm': 17.784151077270508, 'learning_rate': 6.9727272727272735e-06, 'num_tokens': 1776651.0, 'completions/mean_length': 50.75, 'completions/min_length': 39.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.847233235836029, 'rewards/meter/std': 0.2537863850593567, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.96875, 'rewards/lexical_plausibility/std': 0.0883883461356163, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.8717567324638367, 'rewards/repeat_penalty/std': 0.22724825143814087, 'rewards/repeat_floor/mean': 0.9831814169883728, 'rewards/repeat_floor/std': 0.03041817992925644, 'rewards/near_duplicate_penalty/mean': 0.9444844722747803, 'rewards/near_duplicate_penalty/std': 0.0802675411105156, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9264860153198242, 'rewards/distinct_2/std': 0.1488673985004425, 'rewards/total_composite/mean': 0.3012877404689789, 'rewards/total_composite/std': 0.10948627442121506, 'reward': 0.3012877404689789, 'reward_std': 0.10948627442121506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18483132123947144, 'sampling/sampling_logp_difference/max': 2.2926578521728516, 'sampling/importance_sampling_ratio/min': 0.1009976714849472, 'sampling/importance_sampling_ratio/mean': 1.024142861366272, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1258426532149315, 'clip_ratio/low_mean': 0.05139520764350891, 'clip_ratio/low_min': 0.05139520764350891, 'clip_ratio/high_mean': 0.10453180596232414, 'clip_ratio/high_max': 0.10453180596232414, 'clip_ratio/region_mean': 0.15592701360583305, 'reward_total_mean': 0.3012877404689789, 'reward_meter_mean': 0.847233235836029, 'reward_meter_std': 0.2537863850593567, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.96875, 'reward_lexical_plausibility_std': 0.0883883461356163, 'reward_repeat_penalty_mean': 0.8717567324638367, 'reward_repeat_penalty_std': 0.22724825143814087, 'reward_repeat_floor_mean': 0.9831814169883728, 'reward_repeat_floor_std': 0.03041817992925644, 'reward_near_duplicate_penalty_mean': 0.9444844722747803, 'reward_near_duplicate_penalty_std': 0.0802675411105156, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9264860153198242, 'reward_distinct_2_std': 0.1488673985004425, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.3012877404689789, 'reward_total_composite_std': 0.10948627442121506, 'epoch': 0.05}

 30%|███       | 1000/3300 [3:45:24<6:50:50, 10.72s/it]INFO 04-12 19:43:50 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:08<00:49,  4.49s/it]

 23%|██▎       | 3/13 [00:15<00:54,  5.42s/it]

 31%|███       | 4/13 [00:27<01:09,  7.67s/it]

 38%|███▊      | 5/13 [00:35<01:02,  7.78s/it]

 46%|████▌     | 6/13 [00:42<00:53,  7.59s/it]

 54%|█████▍    | 7/13 [00:49<00:45,  7.56s/it]

 62%|██████▏   | 8/13 [01:01<00:44,  8.80s/it]

 69%|██████▉   | 9/13 [01:11<00:36,  9.25s/it]

 77%|███████▋  | 10/13 [01:18<00:25,  8.61s/it]

 85%|████████▍ | 11/13 [01:26<00:16,  8.36s/it]

 92%|█████████▏| 12/13 [01:35<00:08,  8.45s/it]

100%|██████████| 13/13 [01:47<00:00,  9.61s/it]2026-04-12 19:45:47,624 | INFO | train_grpo_train | metrics_logged mode=eval step=1000

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 117.2097, 'eval_samples_per_second': 0.887, 'eval_steps_per_second': 0.111, 'eval_num_tokens': 1776651.0, 'eval_completions/mean_length': 133.19230769230768, 'eval_completions/min_length': 38.38461538461539, 'eval_completions/max_length': 305.7692307692308, 'eval_completions/clipped_ratio': 0.038461538461538464, 'eval_completions/mean_terminated_length': 118.62362670898438, 'eval_completions/min_terminated_length': 38.38461538461539, 'eval_completions/max_terminated_length': 253.53846153846155, 'eval_rewards/meter/mean': 0.5928477942943573, 'eval_rewards/meter/std': 0.3524536627989549, 'eval_rewards/count_adherence/mean': 0.8475735462628878, 'eval_rewards/count_adherence/std': 0.1999759771502935, 'eval_rewards/arabic_clean/mean': 0.9038461538461539, 'eval_rewards/arabic_clean/std': 0.19682669181090134, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9542720638788663, 'eval_rewards/count_floor/std': 0.05999279351761708, 'eval_rewards/lexical_plausibility/mean': 0.9799962318860568, 'eval_rewards/lexical_plausibility/std': 0.042006535025743336, 'eval_rewards/judge_quality/mean': 0.3322115380030412, 'eval_rewards/judge_quality/std': 0.17376625767120948, 'eval_rewards/repeat_penalty/mean': 0.6011631328326005, 'eval_rewards/repeat_penalty/std': 0.3092478123994974, 'eval_rewards/repeat_floor/mean': 0.950995922088623, 'eval_rewards/repeat_floor/std': 0.04671543435408519, 'eval_rewards/near_duplicate_penalty/mean': 0.8854492398408743, 'eval_rewards/near_duplicate_penalty/std': 0.11667734155288109, 'eval_rewards/opening_diversity/mean': 0.9396305221777695, 'eval_rewards/opening_diversity/std': 0.12207835955688587, 'eval_rewards/distinct_2/mean': 0.686674544444451, 'eval_rewards/distinct_2/std': 0.29635819162313753, 'eval_rewards/total_composite/mean': 0.18042223499371454, 'eval_rewards/total_composite/std': 0.14621163216921, 'eval_reward': 0.18042223499371454, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06332223375256245, 'eval_sampling/sampling_logp_difference/max': 1.0790344018202562, 'eval_sampling/importance_sampling_ratio/min': 0.34716252753367793, 'eval_sampling/importance_sampling_ratio/mean': 1.0149937226222112, 'eval_sampling/importance_sampling_ratio/max': 1.5712984158442571, 'eval_entropy': 0.6994791076733515, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.18042223499371454, 'eval_reward_meter_mean': 0.5928477942943573, 'eval_reward_meter_std': 0.3524536627989549, 'eval_reward_count_adherence_mean': 0.8475735462628878, 'eval_reward_count_adherence_std': 0.1999759771502935, 'eval_reward_arabic_clean_mean': 0.9038461538461539, 'eval_reward_arabic_clean_std': 0.19682669181090134, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9542720638788663, 'eval_reward_count_floor_std': 0.05999279351761708, 'eval_reward_lexical_plausibility_mean': 0.9799962318860568, 'eval_reward_lexical_plausibility_std': 0.042006535025743336, 'eval_reward_repeat_penalty_mean': 0.6011631328326005, 'eval_reward_repeat_penalty_std': 0.3092478123994974, 'eval_reward_repeat_floor_mean': 0.950995922088623, 'eval_reward_repeat_floor_std': 0.04671543435408519, 'eval_reward_near_duplicate_penalty_mean': 0.8854492398408743, 'eval_reward_near_duplicate_penalty_std': 0.11667734155288109, 'eval_reward_opening_diversity_mean': 0.9396305221777695, 'eval_reward_opening_diversity_std': 0.12207835955688587, 'eval_reward_distinct_2_mean': 0.686674544444451, 'eval_reward_distinct_2_std': 0.29635819162313753, 'eval_reward_judge_quality_mean': 0.3322115380030412, 'eval_reward_judge_quality_std': 0.17376625767120948, 'eval_reward_total_composite_mean': 0.18042223499371454, 'eval_reward_total_composite_std': 0.14621163216921, 'epoch': 0.05}

 30%|███       | 1000/3300 [3:47:21<6:50:50, 10.72s/it]

100%|██████████| 13/13 [01:47<00:00,  9.61s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 30%|███       | 1001/3300 [3:47:32<29:28:17, 46.15s/it]2026-04-12 19:45:59,153 | INFO | train_grpo_train | metrics_logged mode=train step=1001

                                                        
{'loss': 0.0049, 'grad_norm': 22.124954223632812, 'learning_rate': 6.969696969696971e-06, 'num_tokens': 1778329.0, 'completions/mean_length': 39.75, 'completions/min_length': 32.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6398776769638062, 'rewards/meter/std': 0.3831755816936493, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.953125, 'rewards/lexical_plausibility/std': 0.0867956355214119, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.9656601548194885, 'rewards/repeat_penalty/std': 0.05841750651597977, 'rewards/repeat_floor/mean': 0.9959844350814819, 'rewards/repeat_floor/std': 0.006667524576187134, 'rewards/near_duplicate_penalty/mean': 0.9860895872116089, 'rewards/near_duplicate_penalty/std': 0.0209298525005579, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9785662889480591, 'rewards/distinct_2/std': 0.03976903110742569, 'rewards/total_composite/mean': 0.2616726756095886, 'rewards/total_composite/std': 0.15978965163230896, 'reward': 0.2616726756095886, 'reward_std': 0.15978963673114777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18585973978042603, 'sampling/sampling_logp_difference/max': 1.7011492252349854, 'sampling/importance_sampling_ratio/min': 0.18247368931770325, 'sampling/importance_sampling_ratio/mean': 1.0141067504882812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9986731186509132, 'clip_ratio/low_mean': 0.06751769222319126, 'clip_ratio/low_min': 0.06751769222319126, 'clip_ratio/high_mean': 0.0593555336818099, 'clip_ratio/high_max': 0.0593555336818099, 'clip_ratio/region_mean': 0.12687322590500116, 'reward_total_mean': 0.2616726756095886, 'reward_meter_mean': 0.6398776769638062, 'reward_meter_std': 0.3831755816936493, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.953125, 'reward_lexical_plausibility_std': 0.0867956355214119, 'reward_repeat_penalty_mean': 0.9656601548194885, 'reward_repeat_penalty_std': 0.05841750651597977, 'reward_repeat_floor_mean': 0.9959844350814819, 'reward_repeat_floor_std': 0.006667524576187134, 'reward_near_duplicate_penalty_mean': 0.9860895872116089, 'reward_near_duplicate_penalty_std': 0.0209298525005579, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9785662889480591, 'reward_distinct_2_std': 0.03976903110742569, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.2616726756095886, 'reward_total_composite_std': 0.15978965163230896, 'epoch': 0.05}

 30%|███       | 1001/3300 [3:47:33<29:28:17, 46.15s/it]INFO 04-12 19:45:59 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1002/3300 [3:47:42<22:25:16, 35.12s/it]2026-04-12 19:46:08,553 | INFO | train_grpo_train | metrics_logged mode=train step=1002

                                                        
{'loss': -0.0115, 'grad_norm': 33.8516731262207, 'learning_rate': 6.966666666666667e-06, 'num_tokens': 1780059.0, 'completions/mean_length': 42.25, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7623470425605774, 'rewards/meter/std': 0.33664560317993164, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.6936088800430298, 'rewards/repeat_penalty/std': 0.19767102599143982, 'rewards/repeat_floor/mean': 0.9604244828224182, 'rewards/repeat_floor/std': 0.026541326195001602, 'rewards/near_duplicate_penalty/mean': 0.8602346181869507, 'rewards/near_duplicate_penalty/std': 0.09306368231773376, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7932140231132507, 'rewards/distinct_2/std': 0.1414918303489685, 'rewards/total_composite/mean': 0.22167715430259705, 'rewards/total_composite/std': 0.13051337003707886, 'reward': 0.22167715430259705, 'reward_std': 0.13051335513591766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15382272005081177, 'sampling/sampling_logp_difference/max': 1.674424171447754, 'sampling/importance_sampling_ratio/min': 0.18741606175899506, 'sampling/importance_sampling_ratio/mean': 0.9914189577102661, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7480546422302723, 'clip_ratio/low_mean': 0.04925073031336069, 'clip_ratio/low_min': 0.04925073031336069, 'clip_ratio/high_mean': 0.07198562938719988, 'clip_ratio/high_max': 0.07198562938719988, 'clip_ratio/region_mean': 0.12123635970056057, 'reward_total_mean': 0.22167715430259705, 'reward_meter_mean': 0.7623470425605774, 'reward_meter_std': 0.33664560317993164, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6936088800430298, 'reward_repeat_penalty_std': 0.19767102599143982, 'reward_repeat_floor_mean': 0.9604244828224182, 'reward_repeat_floor_std': 0.026541326195001602, 'reward_near_duplicate_penalty_mean': 0.8602346181869507, 'reward_near_duplicate_penalty_std': 0.09306368231773376, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7932140231132507, 'reward_distinct_2_std': 0.1414918303489685, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.22167715430259705, 'reward_total_composite_std': 0.13051337003707886, 'epoch': 0.05}

 30%|███       | 1002/3300 [3:47:42<22:25:16, 35.12s/it]INFO 04-12 19:46:08 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1003/3300 [3:47:55<18:12:48, 28.55s/it]2026-04-12 19:46:21,741 | INFO | train_grpo_train | metrics_logged mode=train step=1003

                                                        
{'loss': -0.0466, 'grad_norm': 5.782053470611572, 'learning_rate': 6.963636363636364e-06, 'num_tokens': 1781500.0, 'completions/mean_length': 101.125, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.42857360839844, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6858435869216919, 'rewards/meter/std': 0.2864830791950226, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9452910423278809, 'rewards/lexical_plausibility/std': 0.15474040806293488, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.17204131186008453, 'rewards/repeat_penalty/mean': 0.8735029697418213, 'rewards/repeat_penalty/std': 0.15507210791110992, 'rewards/repeat_floor/mean': 0.9849951267242432, 'rewards/repeat_floor/std': 0.01891426369547844, 'rewards/near_duplicate_penalty/mean': 0.9539567232131958, 'rewards/near_duplicate_penalty/std': 0.06166788935661316, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9100174903869629, 'rewards/distinct_2/std': 0.11439591646194458, 'rewards/total_composite/mean': 0.2047213762998581, 'rewards/total_composite/std': 0.1663588434457779, 'reward': 0.2047213762998581, 'reward_std': 0.1663588583469391, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.179252028465271, 'sampling/sampling_logp_difference/max': 1.431613564491272, 'sampling/importance_sampling_ratio/min': 0.2389231026172638, 'sampling/importance_sampling_ratio/mean': 0.9973267316818237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9459731131792068, 'clip_ratio/low_mean': 0.06421158276498318, 'clip_ratio/low_min': 0.06421158276498318, 'clip_ratio/high_mean': 0.06737321056425571, 'clip_ratio/high_max': 0.06737321056425571, 'clip_ratio/region_mean': 0.1315847933292389, 'reward_total_mean': 0.2047213762998581, 'reward_meter_mean': 0.6858435869216919, 'reward_meter_std': 0.2864830791950226, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9452910423278809, 'reward_lexical_plausibility_std': 0.15474040806293488, 'reward_repeat_penalty_mean': 0.8735029697418213, 'reward_repeat_penalty_std': 0.15507210791110992, 'reward_repeat_floor_mean': 0.9849951267242432, 'reward_repeat_floor_std': 0.01891426369547844, 'reward_near_duplicate_penalty_mean': 0.9539567232131958, 'reward_near_duplicate_penalty_std': 0.06166788935661316, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9100174903869629, 'reward_distinct_2_std': 0.11439591646194458, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.17204131186008453, 'reward_total_composite_mean': 0.2047213762998581, 'reward_total_composite_std': 0.1663588434457779, 'epoch': 0.05}

 30%|███       | 1003/3300 [3:47:55<18:12:48, 28.55s/it]INFO 04-12 19:46:21 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1004/3300 [3:48:06<14:49:46, 23.25s/it]2026-04-12 19:46:32,855 | INFO | train_grpo_train | metrics_logged mode=train step=1004

                                                        
{'loss': 0.0387, 'grad_norm': 8.3599214553833, 'learning_rate': 6.960606060606061e-06, 'num_tokens': 1784176.0, 'completions/mean_length': 128.5, 'completions/min_length': 117.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.5, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9748039245605469, 'rewards/meter/std': 0.04602670669555664, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.019838986918330193, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.3358052968978882, 'rewards/repeat_penalty/std': 0.2542286515235901, 'rewards/repeat_floor/mean': 0.9120162725448608, 'rewards/repeat_floor/std': 0.043963972479104996, 'rewards/near_duplicate_penalty/mean': 0.7990399599075317, 'rewards/near_duplicate_penalty/std': 0.11572938412427902, 'rewards/opening_diversity/mean': 0.9090908765792847, 'rewards/opening_diversity/std': 0.15932264924049377, 'rewards/distinct_2/mean': 0.4179425835609436, 'rewards/distinct_2/std': 0.22397363185882568, 'rewards/total_composite/mean': 0.21779005229473114, 'rewards/total_composite/std': 0.11563143879175186, 'reward': 0.21779005229473114, 'reward_std': 0.11563143879175186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14209409058094025, 'sampling/sampling_logp_difference/max': 2.6210408210754395, 'sampling/importance_sampling_ratio/min': 0.07272712141275406, 'sampling/importance_sampling_ratio/mean': 1.011225700378418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6245669387280941, 'clip_ratio/low_mean': 0.05218059569597244, 'clip_ratio/low_min': 0.05218059569597244, 'clip_ratio/high_mean': 0.061950214207172394, 'clip_ratio/high_max': 0.061950214207172394, 'clip_ratio/region_mean': 0.11413080990314484, 'reward_total_mean': 0.21779005229473114, 'reward_meter_mean': 0.9748039245605469, 'reward_meter_std': 0.04602670669555664, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.019838986918330193, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.3358052968978882, 'reward_repeat_penalty_std': 0.2542286515235901, 'reward_repeat_floor_mean': 0.9120162725448608, 'reward_repeat_floor_std': 0.043963972479104996, 'reward_near_duplicate_penalty_mean': 0.7990399599075317, 'reward_near_duplicate_penalty_std': 0.11572938412427902, 'reward_opening_diversity_mean': 0.9090908765792847, 'reward_opening_diversity_std': 0.15932264924049377, 'reward_distinct_2_mean': 0.4179425835609436, 'reward_distinct_2_std': 0.22397363185882568, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.21779005229473114, 'reward_total_composite_std': 0.11563143879175186, 'epoch': 0.05}

 30%|███       | 1004/3300 [3:48:06<14:49:46, 23.25s/it]INFO 04-12 19:46:33 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1005/3300 [3:48:15<12:09:04, 19.06s/it]2026-04-12 19:46:41,925 | INFO | train_grpo_train | metrics_logged mode=train step=1005

                                                        
{'loss': 0.0222, 'grad_norm': 16.885114669799805, 'learning_rate': 6.957575757575759e-06, 'num_tokens': 1785922.0, 'completions/mean_length': 37.25, 'completions/min_length': 33.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7541271448135376, 'rewards/meter/std': 0.3757603168487549, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.8378614783287048, 'rewards/repeat_penalty/std': 0.18811632692813873, 'rewards/repeat_floor/mean': 0.9812521934509277, 'rewards/repeat_floor/std': 0.01981162466108799, 'rewards/near_duplicate_penalty/mean': 0.9331832528114319, 'rewards/near_duplicate_penalty/std': 0.05320000275969505, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.944719672203064, 'rewards/distinct_2/std': 0.07666565477848053, 'rewards/total_composite/mean': 0.37645769119262695, 'rewards/total_composite/std': 0.25472790002822876, 'reward': 0.37645769119262695, 'reward_std': 0.25472790002822876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1486559808254242, 'sampling/sampling_logp_difference/max': 1.6582918167114258, 'sampling/importance_sampling_ratio/min': 0.190464049577713, 'sampling/importance_sampling_ratio/mean': 1.0120564699172974, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8786394447088242, 'clip_ratio/low_mean': 0.04948117770254612, 'clip_ratio/low_min': 0.04948117770254612, 'clip_ratio/high_mean': 0.06206894665956497, 'clip_ratio/high_max': 0.06206894665956497, 'clip_ratio/region_mean': 0.11155012436211109, 'reward_total_mean': 0.37645769119262695, 'reward_meter_mean': 0.7541271448135376, 'reward_meter_std': 0.3757603168487549, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8378614783287048, 'reward_repeat_penalty_std': 0.18811632692813873, 'reward_repeat_floor_mean': 0.9812521934509277, 'reward_repeat_floor_std': 0.01981162466108799, 'reward_near_duplicate_penalty_mean': 0.9331832528114319, 'reward_near_duplicate_penalty_std': 0.05320000275969505, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.944719672203064, 'reward_distinct_2_std': 0.07666565477848053, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.37645769119262695, 'reward_total_composite_std': 0.25472790002822876, 'epoch': 0.05}

 30%|███       | 1005/3300 [3:48:15<12:09:04, 19.06s/it]INFO 04-12 19:46:42 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1006/3300 [3:48:26<10:31:27, 16.52s/it]2026-04-12 19:46:52,502 | INFO | train_grpo_train | metrics_logged mode=train step=1006

                                                        
{'loss': 0.1351, 'grad_norm': 11.750435829162598, 'learning_rate': 6.954545454545455e-06, 'num_tokens': 1788394.0, 'completions/mean_length': 96.0, 'completions/min_length': 86.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.8809903860092163, 'rewards/meter/std': 0.20758835971355438, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9550000429153442, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.6056420207023621, 'rewards/repeat_penalty/std': 0.11947911977767944, 'rewards/repeat_floor/mean': 0.953300416469574, 'rewards/repeat_floor/std': 0.01842612773180008, 'rewards/near_duplicate_penalty/mean': 0.8755171298980713, 'rewards/near_duplicate_penalty/std': 0.06730354577302933, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.038382481783628464, 'rewards/distinct_2/mean': 0.717753529548645, 'rewards/distinct_2/std': 0.09489758312702179, 'rewards/total_composite/mean': 0.26976504921913147, 'rewards/total_composite/std': 0.0885532796382904, 'reward': 0.26976504921913147, 'reward_std': 0.0885532796382904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1504390984773636, 'sampling/sampling_logp_difference/max': 2.123899221420288, 'sampling/importance_sampling_ratio/min': 0.11956451088190079, 'sampling/importance_sampling_ratio/mean': 1.0095388889312744, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8073622062802315, 'clip_ratio/low_mean': 0.03410334698855877, 'clip_ratio/low_min': 0.03410334698855877, 'clip_ratio/high_mean': 0.09946299437433481, 'clip_ratio/high_max': 0.09946299437433481, 'clip_ratio/region_mean': 0.13356634136289358, 'reward_total_mean': 0.26976504921913147, 'reward_meter_mean': 0.8809903860092163, 'reward_meter_std': 0.20758835971355438, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9550000429153442, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.6056420207023621, 'reward_repeat_penalty_std': 0.11947911977767944, 'reward_repeat_floor_mean': 0.953300416469574, 'reward_repeat_floor_std': 0.01842612773180008, 'reward_near_duplicate_penalty_mean': 0.8755171298980713, 'reward_near_duplicate_penalty_std': 0.06730354577302933, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.038382481783628464, 'reward_distinct_2_mean': 0.717753529548645, 'reward_distinct_2_std': 0.09489758312702179, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.26976504921913147, 'reward_total_composite_std': 0.0885532796382904, 'epoch': 0.05}

 30%|███       | 1006/3300 [3:48:26<10:31:27, 16.52s/it]INFO 04-12 19:46:52 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1007/3300 [3:48:39<9:51:58, 15.49s/it] 2026-04-12 19:47:05,598 | INFO | train_grpo_train | metrics_logged mode=train step=1007

                                                       
{'loss': -0.0408, 'grad_norm': 2.1118760108947754, 'learning_rate': 6.951515151515153e-06, 'num_tokens': 1789791.0, 'completions/mean_length': 143.625, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 20.83333396911621, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9310407042503357, 'rewards/meter/std': 0.14807407557964325, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8842537999153137, 'rewards/lexical_plausibility/std': 0.19642283022403717, 'rewards/judge_quality/mean': 0.41750001907348633, 'rewards/judge_quality/std': 0.3491520583629608, 'rewards/repeat_penalty/mean': 0.7398061752319336, 'rewards/repeat_penalty/std': 0.028832554817199707, 'rewards/repeat_floor/mean': 0.9829612374305725, 'rewards/repeat_floor/std': 0.005766510963439941, 'rewards/near_duplicate_penalty/mean': 0.9864082336425781, 'rewards/near_duplicate_penalty/std': 0.0384434312582016, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4026138186454773, 'rewards/total_composite/std': 0.34245648980140686, 'reward': 0.4026138186454773, 'reward_std': 0.34245648980140686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14732153713703156, 'sampling/sampling_logp_difference/max': 1.240102767944336, 'sampling/importance_sampling_ratio/min': 0.28935450315475464, 'sampling/importance_sampling_ratio/mean': 1.0136672258377075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7386007532477379, 'clip_ratio/low_mean': 0.032894738018512726, 'clip_ratio/low_min': 0.032894738018512726, 'clip_ratio/high_mean': 0.07768433075398207, 'clip_ratio/high_max': 0.07768433075398207, 'clip_ratio/region_mean': 0.11057906877249479, 'reward_total_mean': 0.4026138186454773, 'reward_meter_mean': 0.9310407042503357, 'reward_meter_std': 0.14807407557964325, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8842537999153137, 'reward_lexical_plausibility_std': 0.19642283022403717, 'reward_repeat_penalty_mean': 0.7398061752319336, 'reward_repeat_penalty_std': 0.028832554817199707, 'reward_repeat_floor_mean': 0.9829612374305725, 'reward_repeat_floor_std': 0.005766510963439941, 'reward_near_duplicate_penalty_mean': 0.9864082336425781, 'reward_near_duplicate_penalty_std': 0.0384434312582016, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41750001907348633, 'reward_judge_quality_std': 0.3491520583629608, 'reward_total_composite_mean': 0.4026138186454773, 'reward_total_composite_std': 0.34245648980140686, 'epoch': 0.05}

 31%|███       | 1007/3300 [3:48:39<9:51:58, 15.49s/it]INFO 04-12 19:47:05 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1008/3300 [3:48:47<8:31:44, 13.40s/it]2026-04-12 19:47:14,108 | INFO | train_grpo_train | metrics_logged mode=train step=1008

                                                       
{'loss': 0.0978, 'grad_norm': 18.99734878540039, 'learning_rate': 6.948484848484849e-06, 'num_tokens': 1791109.0, 'completions/mean_length': 41.75, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7719502449035645, 'rewards/meter/std': 0.3665204644203186, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.22177450358867645, 'rewards/repeat_penalty/mean': 0.7883635759353638, 'rewards/repeat_penalty/std': 0.2875305116176605, 'rewards/repeat_floor/mean': 0.9761488437652588, 'rewards/repeat_floor/std': 0.03173362836241722, 'rewards/near_duplicate_penalty/mean': 0.9299284219741821, 'rewards/near_duplicate_penalty/std': 0.09230577945709229, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8934398889541626, 'rewards/distinct_2/std': 0.13211345672607422, 'rewards/total_composite/mean': 0.26631075143814087, 'rewards/total_composite/std': 0.1468190997838974, 'reward': 0.26631075143814087, 'reward_std': 0.1468190997838974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16248413920402527, 'sampling/sampling_logp_difference/max': 2.112684488296509, 'sampling/importance_sampling_ratio/min': 0.12091293931007385, 'sampling/importance_sampling_ratio/mean': 1.0183058977127075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0836613178253174, 'clip_ratio/low_mean': 0.036015139892697334, 'clip_ratio/low_min': 0.036015139892697334, 'clip_ratio/high_mean': 0.097270124591887, 'clip_ratio/high_max': 0.097270124591887, 'clip_ratio/region_mean': 0.13328526448458433, 'reward_total_mean': 0.26631075143814087, 'reward_meter_mean': 0.7719502449035645, 'reward_meter_std': 0.3665204644203186, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7883635759353638, 'reward_repeat_penalty_std': 0.2875305116176605, 'reward_repeat_floor_mean': 0.9761488437652588, 'reward_repeat_floor_std': 0.03173362836241722, 'reward_near_duplicate_penalty_mean': 0.9299284219741821, 'reward_near_duplicate_penalty_std': 0.09230577945709229, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8934398889541626, 'reward_distinct_2_std': 0.13211345672607422, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.26631075143814087, 'reward_total_composite_std': 0.1468190997838974, 'epoch': 0.05}

 31%|███       | 1008/3300 [3:48:47<8:31:44, 13.40s/it]INFO 04-12 19:47:14 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1009/3300 [3:48:55<7:25:02, 11.66s/it]2026-04-12 19:47:21,700 | INFO | train_grpo_train | metrics_logged mode=train step=1009

                                                       
{'loss': 0.0024, 'grad_norm': 19.639877319335938, 'learning_rate': 6.945454545454546e-06, 'num_tokens': 1792713.0, 'completions/mean_length': 38.5, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.20156092941761017, 'rewards/meter/std': 0.3161162734031677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8487499952316284, 'rewards/judge_quality/std': 0.20152544975280762, 'rewards/repeat_penalty/mean': 0.8401227593421936, 'rewards/repeat_penalty/std': 0.24954256415367126, 'rewards/repeat_floor/mean': 0.9801956415176392, 'rewards/repeat_floor/std': 0.03257455304265022, 'rewards/near_duplicate_penalty/mean': 0.9383005499839783, 'rewards/near_duplicate_penalty/std': 0.11043066531419754, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9244505167007446, 'rewards/distinct_2/std': 0.10807745158672333, 'rewards/total_composite/mean': 0.1712694615125656, 'rewards/total_composite/std': 0.29461145401000977, 'reward': 0.1712694615125656, 'reward_std': 0.29461145401000977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15079528093338013, 'sampling/sampling_logp_difference/max': 2.5437378883361816, 'sampling/importance_sampling_ratio/min': 0.07857215404510498, 'sampling/importance_sampling_ratio/mean': 1.0141090154647827, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7228083424270153, 'clip_ratio/low_mean': 0.11751789692789316, 'clip_ratio/low_min': 0.11751789692789316, 'clip_ratio/high_mean': 0.011627906933426857, 'clip_ratio/high_max': 0.011627906933426857, 'clip_ratio/region_mean': 0.12914580386132002, 'reward_total_mean': 0.1712694615125656, 'reward_meter_mean': 0.20156092941761017, 'reward_meter_std': 0.3161162734031677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8401227593421936, 'reward_repeat_penalty_std': 0.24954256415367126, 'reward_repeat_floor_mean': 0.9801956415176392, 'reward_repeat_floor_std': 0.03257455304265022, 'reward_near_duplicate_penalty_mean': 0.9383005499839783, 'reward_near_duplicate_penalty_std': 0.11043066531419754, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9244505167007446, 'reward_distinct_2_std': 0.10807745158672333, 'reward_judge_quality_mean': 0.8487499952316284, 'reward_judge_quality_std': 0.20152544975280762, 'reward_total_composite_mean': 0.1712694615125656, 'reward_total_composite_std': 0.29461145401000977, 'epoch': 0.05}

 31%|███       | 1009/3300 [3:48:55<7:25:02, 11.66s/it]INFO 04-12 19:47:21 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1010/3300 [3:49:08<7:43:30, 12.14s/it]2026-04-12 19:47:34,988 | INFO | train_grpo_train | metrics_logged mode=train step=1010

                                                       
{'loss': -0.0514, 'grad_norm': 6.53665018081665, 'learning_rate': 6.942424242424243e-06, 'num_tokens': 1794308.0, 'completions/mean_length': 100.375, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.57143020629883, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6300738453865051, 'rewards/meter/std': 0.4180264174938202, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9348697662353516, 'rewards/lexical_plausibility/std': 0.18421603739261627, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7733302116394043, 'rewards/repeat_penalty/std': 0.17437773942947388, 'rewards/repeat_floor/mean': 0.9716709852218628, 'rewards/repeat_floor/std': 0.021430013701319695, 'rewards/near_duplicate_penalty/mean': 0.8944940567016602, 'rewards/near_duplicate_penalty/std': 0.07561302930116653, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8819093704223633, 'rewards/distinct_2/std': 0.09303763508796692, 'rewards/total_composite/mean': 0.25638124346733093, 'rewards/total_composite/std': 0.17983435094356537, 'reward': 0.25638124346733093, 'reward_std': 0.17983435094356537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1843303143978119, 'sampling/sampling_logp_difference/max': 2.005678653717041, 'sampling/importance_sampling_ratio/min': 0.27497678995132446, 'sampling/importance_sampling_ratio/mean': 1.0134769678115845, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9222210869193077, 'clip_ratio/low_mean': 0.04940476268529892, 'clip_ratio/low_min': 0.04940476268529892, 'clip_ratio/high_mean': 0.08858220186084509, 'clip_ratio/high_max': 0.08858220186084509, 'clip_ratio/region_mean': 0.137986964546144, 'reward_total_mean': 0.25638124346733093, 'reward_meter_mean': 0.6300738453865051, 'reward_meter_std': 0.4180264174938202, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9348697662353516, 'reward_lexical_plausibility_std': 0.18421603739261627, 'reward_repeat_penalty_mean': 0.7733302116394043, 'reward_repeat_penalty_std': 0.17437773942947388, 'reward_repeat_floor_mean': 0.9716709852218628, 'reward_repeat_floor_std': 0.021430013701319695, 'reward_near_duplicate_penalty_mean': 0.8944940567016602, 'reward_near_duplicate_penalty_std': 0.07561302930116653, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8819093704223633, 'reward_distinct_2_std': 0.09303763508796692, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.25638124346733093, 'reward_total_composite_std': 0.17983435094356537, 'epoch': 0.05}

 31%|███       | 1010/3300 [3:49:08<7:43:30, 12.14s/it]INFO 04-12 19:47:35 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1011/3300 [3:49:23<8:09:00, 12.82s/it]2026-04-12 19:47:49,377 | INFO | train_grpo_train | metrics_logged mode=train step=1011

                                                       
{'loss': -0.0497, 'grad_norm': 2.913093328475952, 'learning_rate': 6.93939393939394e-06, 'num_tokens': 1795827.0, 'completions/mean_length': 158.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 41.16666793823242, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8286142349243164, 'rewards/meter/std': 0.26238933205604553, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9084140062332153, 'rewards/lexical_plausibility/std': 0.184202179312706, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.8880747556686401, 'rewards/repeat_penalty/std': 0.12806706130504608, 'rewards/repeat_floor/mean': 0.9884530901908875, 'rewards/repeat_floor/std': 0.011946422047913074, 'rewards/near_duplicate_penalty/mean': 0.9700514674186707, 'rewards/near_duplicate_penalty/std': 0.026091916486620903, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9424481391906738, 'rewards/distinct_2/std': 0.07517612725496292, 'rewards/total_composite/mean': 0.19156673550605774, 'rewards/total_composite/std': 0.1724543571472168, 'reward': 0.19156673550605774, 'reward_std': 0.1724543422460556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16360272467136383, 'sampling/sampling_logp_difference/max': 1.8231163024902344, 'sampling/importance_sampling_ratio/min': 0.16152162849903107, 'sampling/importance_sampling_ratio/mean': 0.9966424703598022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7443043738603592, 'clip_ratio/low_mean': 0.04181357566267252, 'clip_ratio/low_min': 0.04181357566267252, 'clip_ratio/high_mean': 0.0680640246719122, 'clip_ratio/high_max': 0.0680640246719122, 'clip_ratio/region_mean': 0.10987760033458471, 'reward_total_mean': 0.19156673550605774, 'reward_meter_mean': 0.8286142349243164, 'reward_meter_std': 0.26238933205604553, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9084140062332153, 'reward_lexical_plausibility_std': 0.184202179312706, 'reward_repeat_penalty_mean': 0.8880747556686401, 'reward_repeat_penalty_std': 0.12806706130504608, 'reward_repeat_floor_mean': 0.9884530901908875, 'reward_repeat_floor_std': 0.011946422047913074, 'reward_near_duplicate_penalty_mean': 0.9700514674186707, 'reward_near_duplicate_penalty_std': 0.026091916486620903, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9424481391906738, 'reward_distinct_2_std': 0.07517612725496292, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.19156673550605774, 'reward_total_composite_std': 0.1724543571472168, 'epoch': 0.05}

 31%|███       | 1011/3300 [3:49:23<8:09:00, 12.82s/it]INFO 04-12 19:47:49 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1012/3300 [3:49:32<7:29:05, 11.78s/it]2026-04-12 19:47:58,724 | INFO | train_grpo_train | metrics_logged mode=train step=1012

                                                       
{'loss': -0.0235, 'grad_norm': 16.611148834228516, 'learning_rate': 6.936363636363636e-06, 'num_tokens': 1797920.0, 'completions/mean_length': 79.625, 'completions/min_length': 56.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.625, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.5717673301696777, 'rewards/meter/std': 0.20936554670333862, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625821024179459, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.025877466425299644, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.6044003367424011, 'rewards/repeat_penalty/std': 0.30608245730400085, 'rewards/repeat_floor/mean': 0.9524610638618469, 'rewards/repeat_floor/std': 0.04029426351189613, 'rewards/near_duplicate_penalty/mean': 0.8641104102134705, 'rewards/near_duplicate_penalty/std': 0.1204640343785286, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.7066054940223694, 'rewards/distinct_2/std': 0.2334507554769516, 'rewards/total_composite/mean': 0.17899343371391296, 'rewards/total_composite/std': 0.09015974402427673, 'reward': 0.17899343371391296, 'reward_std': 0.09015974402427673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17611022293567657, 'sampling/sampling_logp_difference/max': 3.5170092582702637, 'sampling/importance_sampling_ratio/min': 0.029688091948628426, 'sampling/importance_sampling_ratio/mean': 0.9967480301856995, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6358252018690109, 'clip_ratio/low_mean': 0.07266506366431713, 'clip_ratio/low_min': 0.07266506366431713, 'clip_ratio/high_mean': 0.06133164744824171, 'clip_ratio/high_max': 0.06133164744824171, 'clip_ratio/region_mean': 0.13399671111255884, 'reward_total_mean': 0.17899343371391296, 'reward_meter_mean': 0.5717673301696777, 'reward_meter_std': 0.20936554670333862, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625821024179459, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.025877466425299644, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6044003367424011, 'reward_repeat_penalty_std': 0.30608245730400085, 'reward_repeat_floor_mean': 0.9524610638618469, 'reward_repeat_floor_std': 0.04029426351189613, 'reward_near_duplicate_penalty_mean': 0.8641104102134705, 'reward_near_duplicate_penalty_std': 0.1204640343785286, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.7066054940223694, 'reward_distinct_2_std': 0.2334507554769516, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.17899343371391296, 'reward_total_composite_std': 0.09015974402427673, 'epoch': 0.05}

 31%|███       | 1012/3300 [3:49:32<7:29:05, 11.78s/it]INFO 04-12 19:47:58 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1013/3300 [3:49:46<7:49:56, 12.33s/it]2026-04-12 19:48:12,346 | INFO | train_grpo_train | metrics_logged mode=train step=1013

                                                       
{'loss': -0.0772, 'grad_norm': 4.246950626373291, 'learning_rate': 6.9333333333333344e-06, 'num_tokens': 1800689.0, 'completions/mean_length': 194.125, 'completions/min_length': 140.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 148.71429443359375, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.8107649683952332, 'rewards/meter/std': 0.2995389997959137, 'rewards/count_adherence/mean': 0.6944444179534912, 'rewards/count_adherence/std': 0.19472770392894745, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9083333015441895, 'rewards/count_floor/std': 0.05841831862926483, 'rewards/lexical_plausibility/mean': 0.9449319839477539, 'rewards/lexical_plausibility/std': 0.15575581789016724, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.37832921743392944, 'rewards/repeat_penalty/std': 0.2577608823776245, 'rewards/repeat_floor/mean': 0.9157594442367554, 'rewards/repeat_floor/std': 0.055635854601860046, 'rewards/near_duplicate_penalty/mean': 0.8125444650650024, 'rewards/near_duplicate_penalty/std': 0.1331138014793396, 'rewards/opening_diversity/mean': 0.8142857551574707, 'rewards/opening_diversity/std': 0.2313605397939682, 'rewards/distinct_2/mean': 0.5002288818359375, 'rewards/distinct_2/std': 0.27545952796936035, 'rewards/total_composite/mean': 0.13160385191440582, 'rewards/total_composite/std': 0.1019962877035141, 'reward': 0.13160385191440582, 'reward_std': 0.1019962951540947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11599625647068024, 'sampling/sampling_logp_difference/max': 2.0500707626342773, 'sampling/importance_sampling_ratio/min': 0.1287257969379425, 'sampling/importance_sampling_ratio/mean': 1.0036892890930176, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6240360960364342, 'clip_ratio/low_mean': 0.0621154080145061, 'clip_ratio/low_min': 0.0621154080145061, 'clip_ratio/high_mean': 0.03458994813263416, 'clip_ratio/high_max': 0.03458994813263416, 'clip_ratio/region_mean': 0.09670535614714026, 'reward_total_mean': 0.13160385191440582, 'reward_meter_mean': 0.8107649683952332, 'reward_meter_std': 0.2995389997959137, 'reward_count_adherence_mean': 0.6944444179534912, 'reward_count_adherence_std': 0.19472770392894745, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9083333015441895, 'reward_count_floor_std': 0.05841831862926483, 'reward_lexical_plausibility_mean': 0.9449319839477539, 'reward_lexical_plausibility_std': 0.15575581789016724, 'reward_repeat_penalty_mean': 0.37832921743392944, 'reward_repeat_penalty_std': 0.2577608823776245, 'reward_repeat_floor_mean': 0.9157594442367554, 'reward_repeat_floor_std': 0.055635854601860046, 'reward_near_duplicate_penalty_mean': 0.8125444650650024, 'reward_near_duplicate_penalty_std': 0.1331138014793396, 'reward_opening_diversity_mean': 0.8142857551574707, 'reward_opening_diversity_std': 0.2313605397939682, 'reward_distinct_2_mean': 0.5002288818359375, 'reward_distinct_2_std': 0.27545952796936035, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.13160385191440582, 'reward_total_composite_std': 0.1019962877035141, 'epoch': 0.05}

 31%|███       | 1013/3300 [3:49:46<7:49:56, 12.33s/it]INFO 04-12 19:48:12 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1014/3300 [3:49:55<7:10:33, 11.30s/it]2026-04-12 19:48:21,242 | INFO | train_grpo_train | metrics_logged mode=train step=1014

                                                       
{'loss': -0.0059, 'grad_norm': 16.719514846801758, 'learning_rate': 6.930303030303031e-06, 'num_tokens': 1802249.0, 'completions/mean_length': 39.0, 'completions/min_length': 34.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6012445092201233, 'rewards/meter/std': 0.3797958791255951, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9591650366783142, 'rewards/repeat_penalty/std': 0.05876043438911438, 'rewards/repeat_floor/mean': 0.994678258895874, 'rewards/repeat_floor/std': 0.007383603136986494, 'rewards/near_duplicate_penalty/mean': 0.9743462800979614, 'rewards/near_duplicate_penalty/std': 0.03212657570838928, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9836259484291077, 'rewards/distinct_2/std': 0.03032796084880829, 'rewards/total_composite/mean': 0.2808791697025299, 'rewards/total_composite/std': 0.21103060245513916, 'reward': 0.2808791697025299, 'reward_std': 0.21103060245513916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14438484609127045, 'sampling/sampling_logp_difference/max': 1.0845024585723877, 'sampling/importance_sampling_ratio/min': 0.33806994557380676, 'sampling/importance_sampling_ratio/mean': 1.0362433195114136, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8803220018744469, 'clip_ratio/low_mean': 0.07116326969116926, 'clip_ratio/low_min': 0.07116326969116926, 'clip_ratio/high_mean': 0.06175364926457405, 'clip_ratio/high_max': 0.06175364926457405, 'clip_ratio/region_mean': 0.1329169189557433, 'reward_total_mean': 0.2808791697025299, 'reward_meter_mean': 0.6012445092201233, 'reward_meter_std': 0.3797958791255951, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9591650366783142, 'reward_repeat_penalty_std': 0.05876043438911438, 'reward_repeat_floor_mean': 0.994678258895874, 'reward_repeat_floor_std': 0.007383603136986494, 'reward_near_duplicate_penalty_mean': 0.9743462800979614, 'reward_near_duplicate_penalty_std': 0.03212657570838928, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9836259484291077, 'reward_distinct_2_std': 0.03032796084880829, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.2808791697025299, 'reward_total_composite_std': 0.21103060245513916, 'epoch': 0.05}

 31%|███       | 1014/3300 [3:49:55<7:10:33, 11.30s/it]INFO 04-12 19:48:21 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1015/3300 [3:50:04<6:44:43, 10.63s/it]2026-04-12 19:48:30,300 | INFO | train_grpo_train | metrics_logged mode=train step=1015

                                                       
{'loss': 0.0455, 'grad_norm': 15.736807823181152, 'learning_rate': 6.927272727272728e-06, 'num_tokens': 1804083.0, 'completions/mean_length': 62.25, 'completions/min_length': 50.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.25, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.5187026262283325, 'rewards/meter/std': 0.3501666188240051, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9775000214576721, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9289546608924866, 'rewards/repeat_penalty/std': 0.05026537925004959, 'rewards/repeat_floor/mean': 0.9915232062339783, 'rewards/repeat_floor/std': 0.005576802417635918, 'rewards/near_duplicate_penalty/mean': 0.9679281711578369, 'rewards/near_duplicate_penalty/std': 0.01552621554583311, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9592667818069458, 'rewards/distinct_2/std': 0.037985555827617645, 'rewards/total_composite/mean': 0.20378226041793823, 'rewards/total_composite/std': 0.13046510517597198, 'reward': 0.20378226041793823, 'reward_std': 0.13046510517597198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.191785991191864, 'sampling/sampling_logp_difference/max': 3.3755979537963867, 'sampling/importance_sampling_ratio/min': 0.03419766202569008, 'sampling/importance_sampling_ratio/mean': 0.9981364011764526, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8586732968688011, 'clip_ratio/low_mean': 0.075193808414042, 'clip_ratio/low_min': 0.075193808414042, 'clip_ratio/high_mean': 0.09183205105364323, 'clip_ratio/high_max': 0.09183205105364323, 'clip_ratio/region_mean': 0.16702585946768522, 'reward_total_mean': 0.20378226041793823, 'reward_meter_mean': 0.5187026262283325, 'reward_meter_std': 0.3501666188240051, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9775000214576721, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9289546608924866, 'reward_repeat_penalty_std': 0.05026537925004959, 'reward_repeat_floor_mean': 0.9915232062339783, 'reward_repeat_floor_std': 0.005576802417635918, 'reward_near_duplicate_penalty_mean': 0.9679281711578369, 'reward_near_duplicate_penalty_std': 0.01552621554583311, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9592667818069458, 'reward_distinct_2_std': 0.037985555827617645, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.20378226041793823, 'reward_total_composite_std': 0.13046510517597198, 'epoch': 0.05}

 31%|███       | 1015/3300 [3:50:04<6:44:43, 10.63s/it]INFO 04-12 19:48:30 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1016/3300 [3:50:18<7:28:01, 11.77s/it]2026-04-12 19:48:44,737 | INFO | train_grpo_train | metrics_logged mode=train step=1016

                                                       
{'loss': -0.0999, 'grad_norm': 2.734677314758301, 'learning_rate': 6.9242424242424245e-06, 'num_tokens': 1805953.0, 'completions/mean_length': 185.75, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 77.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.8243870139122009, 'rewards/meter/std': 0.3384326696395874, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9718749523162842, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 0.9324014186859131, 'rewards/lexical_plausibility/std': 0.1251707375049591, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.7513312101364136, 'rewards/repeat_penalty/std': 0.24689440429210663, 'rewards/repeat_floor/mean': 0.9705663919448853, 'rewards/repeat_floor/std': 0.02873048186302185, 'rewards/near_duplicate_penalty/mean': 0.9130241274833679, 'rewards/near_duplicate_penalty/std': 0.07745034992694855, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8179193735122681, 'rewards/distinct_2/std': 0.19593806564807892, 'rewards/total_composite/mean': 0.16487714648246765, 'rewards/total_composite/std': 0.1369059830904007, 'reward': 0.16487714648246765, 'reward_std': 0.1369059830904007, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15814648568630219, 'sampling/sampling_logp_difference/max': 1.503152847290039, 'sampling/importance_sampling_ratio/min': 0.22242778539657593, 'sampling/importance_sampling_ratio/mean': 1.0122147798538208, 'sampling/importance_sampling_ratio/max': 1.9908485412597656, 'entropy': 0.7580432891845703, 'clip_ratio/low_mean': 0.06812815926969051, 'clip_ratio/low_min': 0.06812815926969051, 'clip_ratio/high_mean': 0.04769197665154934, 'clip_ratio/high_max': 0.04769197665154934, 'clip_ratio/region_mean': 0.11582013592123985, 'reward_total_mean': 0.16487714648246765, 'reward_meter_mean': 0.8243870139122009, 'reward_meter_std': 0.3384326696395874, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9718749523162842, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 0.9324014186859131, 'reward_lexical_plausibility_std': 0.1251707375049591, 'reward_repeat_penalty_mean': 0.7513312101364136, 'reward_repeat_penalty_std': 0.24689440429210663, 'reward_repeat_floor_mean': 0.9705663919448853, 'reward_repeat_floor_std': 0.02873048186302185, 'reward_near_duplicate_penalty_mean': 0.9130241274833679, 'reward_near_duplicate_penalty_std': 0.07745034992694855, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8179193735122681, 'reward_distinct_2_std': 0.19593806564807892, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.16487714648246765, 'reward_total_composite_std': 0.1369059830904007, 'epoch': 0.05}

 31%|███       | 1016/3300 [3:50:18<7:28:01, 11.77s/it]INFO 04-12 19:48:44 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1017/3300 [3:50:27<6:58:11, 10.99s/it]2026-04-12 19:48:53,906 | INFO | train_grpo_train | metrics_logged mode=train step=1017

                                                       
{'loss': 0.0294, 'grad_norm': 15.249451637268066, 'learning_rate': 6.921212121212122e-06, 'num_tokens': 1807476.0, 'completions/mean_length': 42.375, 'completions/min_length': 36.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9520642757415771, 'rewards/meter/std': 0.039989177137613297, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.7672468423843384, 'rewards/repeat_penalty/std': 0.21213914453983307, 'rewards/repeat_floor/mean': 0.9698014259338379, 'rewards/repeat_floor/std': 0.026916010305285454, 'rewards/near_duplicate_penalty/mean': 0.8764516115188599, 'rewards/near_duplicate_penalty/std': 0.0908878818154335, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.912040114402771, 'rewards/distinct_2/std': 0.08792216330766678, 'rewards/total_composite/mean': 0.4185837209224701, 'rewards/total_composite/std': 0.19330984354019165, 'reward': 0.4185837209224701, 'reward_std': 0.19330984354019165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14242349565029144, 'sampling/sampling_logp_difference/max': 2.2091095447540283, 'sampling/importance_sampling_ratio/min': 0.1097983792424202, 'sampling/importance_sampling_ratio/mean': 1.016740083694458, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9091897383332253, 'clip_ratio/low_mean': 0.08053475618362427, 'clip_ratio/low_min': 0.08053475618362427, 'clip_ratio/high_mean': 0.02852564211934805, 'clip_ratio/high_max': 0.02852564211934805, 'clip_ratio/region_mean': 0.10906039830297232, 'reward_total_mean': 0.4185837209224701, 'reward_meter_mean': 0.9520642757415771, 'reward_meter_std': 0.039989177137613297, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7672468423843384, 'reward_repeat_penalty_std': 0.21213914453983307, 'reward_repeat_floor_mean': 0.9698014259338379, 'reward_repeat_floor_std': 0.026916010305285454, 'reward_near_duplicate_penalty_mean': 0.8764516115188599, 'reward_near_duplicate_penalty_std': 0.0908878818154335, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.912040114402771, 'reward_distinct_2_std': 0.08792216330766678, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.4185837209224701, 'reward_total_composite_std': 0.19330984354019165, 'epoch': 0.05}

 31%|███       | 1017/3300 [3:50:27<6:58:11, 10.99s/it]INFO 04-12 19:48:54 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1018/3300 [3:50:36<6:33:34, 10.35s/it]2026-04-12 19:49:02,756 | INFO | train_grpo_train | metrics_logged mode=train step=1018

                                                       
{'loss': 0.0507, 'grad_norm': 19.598302841186523, 'learning_rate': 6.918181818181818e-06, 'num_tokens': 1808956.0, 'completions/mean_length': 35.0, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5354318618774414, 'rewards/meter/std': 0.43367528915405273, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.9152252078056335, 'rewards/repeat_penalty/std': 0.16887253522872925, 'rewards/repeat_floor/mean': 0.9895817637443542, 'rewards/repeat_floor/std': 0.019032226875424385, 'rewards/near_duplicate_penalty/mean': 0.9587793350219727, 'rewards/near_duplicate_penalty/std': 0.05521485581994057, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9737762212753296, 'rewards/distinct_2/std': 0.07417204976081848, 'rewards/total_composite/mean': 0.24740704894065857, 'rewards/total_composite/std': 0.2626734972000122, 'reward': 0.24740704894065857, 'reward_std': 0.2626734972000122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1655251681804657, 'sampling/sampling_logp_difference/max': 1.6508779525756836, 'sampling/importance_sampling_ratio/min': 0.19188137352466583, 'sampling/importance_sampling_ratio/mean': 1.0032281875610352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0728318616747856, 'clip_ratio/low_mean': 0.07185245538130403, 'clip_ratio/low_min': 0.07185245538130403, 'clip_ratio/high_mean': 0.09028219431638718, 'clip_ratio/high_max': 0.09028219431638718, 'clip_ratio/region_mean': 0.1621346496976912, 'reward_total_mean': 0.24740704894065857, 'reward_meter_mean': 0.5354318618774414, 'reward_meter_std': 0.43367528915405273, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9152252078056335, 'reward_repeat_penalty_std': 0.16887253522872925, 'reward_repeat_floor_mean': 0.9895817637443542, 'reward_repeat_floor_std': 0.019032226875424385, 'reward_near_duplicate_penalty_mean': 0.9587793350219727, 'reward_near_duplicate_penalty_std': 0.05521485581994057, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9737762212753296, 'reward_distinct_2_std': 0.07417204976081848, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.24740704894065857, 'reward_total_composite_std': 0.2626734972000122, 'epoch': 0.05}

 31%|███       | 1018/3300 [3:50:36<6:33:34, 10.35s/it]INFO 04-12 19:49:03 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1019/3300 [3:50:44<6:09:11,  9.71s/it]2026-04-12 19:49:10,979 | INFO | train_grpo_train | metrics_logged mode=train step=1019

                                                       
{'loss': 0.0669, 'grad_norm': 23.339536666870117, 'learning_rate': 6.915151515151515e-06, 'num_tokens': 1810416.0, 'completions/mean_length': 18.5, 'completions/min_length': 17.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.9146131873130798, 'rewards/meter/std': 0.16223149001598358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2790483236312866, 'rewards/total_composite/std': 0.1188952848315239, 'reward': 0.2790483236312866, 'reward_std': 0.11889529228210449, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1691134124994278, 'sampling/sampling_logp_difference/max': 1.1980762481689453, 'sampling/importance_sampling_ratio/min': 0.30177417397499084, 'sampling/importance_sampling_ratio/mean': 1.0461504459381104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.437881775200367, 'clip_ratio/low_mean': 0.09756929334253073, 'clip_ratio/low_min': 0.09756929334253073, 'clip_ratio/high_mean': 0.04771671863272786, 'clip_ratio/high_max': 0.04771671863272786, 'clip_ratio/region_mean': 0.1452860119752586, 'reward_total_mean': 0.2790483236312866, 'reward_meter_mean': 0.9146131873130798, 'reward_meter_std': 0.16223149001598358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.2790483236312866, 'reward_total_composite_std': 0.1188952848315239, 'epoch': 0.05}

 31%|███       | 1019/3300 [3:50:44<6:09:11,  9.71s/it]INFO 04-12 19:49:11 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1020/3300 [3:50:58<6:49:50, 10.79s/it]2026-04-12 19:49:24,271 | INFO | train_grpo_train | metrics_logged mode=train step=1020

                                                       
{'loss': -0.0313, 'grad_norm': 4.8817291259765625, 'learning_rate': 6.912121212121212e-06, 'num_tokens': 1811742.0, 'completions/mean_length': 83.75, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.571430206298828, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7196645736694336, 'rewards/meter/std': 0.3615145683288574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9158893823623657, 'rewards/lexical_plausibility/std': 0.18216611444950104, 'rewards/judge_quality/mean': 0.28125, 'rewards/judge_quality/std': 0.1869635283946991, 'rewards/repeat_penalty/mean': 0.729889988899231, 'rewards/repeat_penalty/std': 0.056879594922065735, 'rewards/repeat_floor/mean': 0.9823212623596191, 'rewards/repeat_floor/std': 0.007576760370284319, 'rewards/near_duplicate_penalty/mean': 0.996564507484436, 'rewards/near_duplicate_penalty/std': 0.009717104025185108, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9759615659713745, 'rewards/distinct_2/std': 0.06799104809761047, 'rewards/total_composite/mean': 0.2241857647895813, 'rewards/total_composite/std': 0.17084917426109314, 'reward': 0.2241857647895813, 'reward_std': 0.17084917426109314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16490037739276886, 'sampling/sampling_logp_difference/max': 1.0908076763153076, 'sampling/importance_sampling_ratio/min': 0.3359450399875641, 'sampling/importance_sampling_ratio/mean': 1.0472825765609741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.959427073597908, 'clip_ratio/low_mean': 0.0788014093413949, 'clip_ratio/low_min': 0.0788014093413949, 'clip_ratio/high_mean': 0.06151185929775238, 'clip_ratio/high_max': 0.06151185929775238, 'clip_ratio/region_mean': 0.14031326863914728, 'reward_total_mean': 0.2241857647895813, 'reward_meter_mean': 0.7196645736694336, 'reward_meter_std': 0.3615145683288574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9158893823623657, 'reward_lexical_plausibility_std': 0.18216611444950104, 'reward_repeat_penalty_mean': 0.729889988899231, 'reward_repeat_penalty_std': 0.056879594922065735, 'reward_repeat_floor_mean': 0.9823212623596191, 'reward_repeat_floor_std': 0.007576760370284319, 'reward_near_duplicate_penalty_mean': 0.996564507484436, 'reward_near_duplicate_penalty_std': 0.009717104025185108, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9759615659713745, 'reward_distinct_2_std': 0.06799104809761047, 'reward_judge_quality_mean': 0.28125, 'reward_judge_quality_std': 0.1869635283946991, 'reward_total_composite_mean': 0.2241857647895813, 'reward_total_composite_std': 0.17084917426109314, 'epoch': 0.05}

 31%|███       | 1020/3300 [3:50:58<6:49:50, 10.79s/it]INFO 04-12 19:49:24 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1021/3300 [3:51:07<6:34:12, 10.38s/it]2026-04-12 19:49:33,705 | INFO | train_grpo_train | metrics_logged mode=train step=1021

                                                       
{'loss': 0.0312, 'grad_norm': 11.923295974731445, 'learning_rate': 6.90909090909091e-06, 'num_tokens': 1813520.0, 'completions/mean_length': 59.25, 'completions/min_length': 55.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5017757415771484, 'rewards/meter/std': 0.3954038619995117, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390502154827118, 'rewards/repeat_penalty/mean': 0.7688629627227783, 'rewards/repeat_penalty/std': 0.2558096945285797, 'rewards/repeat_floor/mean': 0.9724924564361572, 'rewards/repeat_floor/std': 0.03248051553964615, 'rewards/near_duplicate_penalty/mean': 0.9232298731803894, 'rewards/near_duplicate_penalty/std': 0.09298214316368103, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.822311282157898, 'rewards/distinct_2/std': 0.2100122720003128, 'rewards/total_composite/mean': 0.17784088850021362, 'rewards/total_composite/std': 0.12724155187606812, 'reward': 0.17784088850021362, 'reward_std': 0.12724153697490692, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1490003764629364, 'sampling/sampling_logp_difference/max': 3.624004364013672, 'sampling/importance_sampling_ratio/min': 0.026675643399357796, 'sampling/importance_sampling_ratio/mean': 1.003395676612854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7765489220619202, 'clip_ratio/low_mean': 0.05926316091790795, 'clip_ratio/low_min': 0.05926316091790795, 'clip_ratio/high_mean': 0.0503940312191844, 'clip_ratio/high_max': 0.0503940312191844, 'clip_ratio/region_mean': 0.10965719213709235, 'reward_total_mean': 0.17784088850021362, 'reward_meter_mean': 0.5017757415771484, 'reward_meter_std': 0.3954038619995117, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7688629627227783, 'reward_repeat_penalty_std': 0.2558096945285797, 'reward_repeat_floor_mean': 0.9724924564361572, 'reward_repeat_floor_std': 0.03248051553964615, 'reward_near_duplicate_penalty_mean': 0.9232298731803894, 'reward_near_duplicate_penalty_std': 0.09298214316368103, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.822311282157898, 'reward_distinct_2_std': 0.2100122720003128, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390502154827118, 'reward_total_composite_mean': 0.17784088850021362, 'reward_total_composite_std': 0.12724155187606812, 'epoch': 0.05}

 31%|███       | 1021/3300 [3:51:07<6:34:12, 10.38s/it]INFO 04-12 19:49:33 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1022/3300 [3:51:22<7:21:20, 11.62s/it]2026-04-12 19:49:48,235 | INFO | train_grpo_train | metrics_logged mode=train step=1022

                                                       
{'loss': -0.0738, 'grad_norm': 5.380892276763916, 'learning_rate': 6.906060606060606e-06, 'num_tokens': 1815317.0, 'completions/mean_length': 123.625, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 68.14286041259766, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6609570980072021, 'rewards/meter/std': 0.3926407992839813, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9517898559570312, 'rewards/lexical_plausibility/std': 0.13635894656181335, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.8896200656890869, 'rewards/repeat_penalty/std': 0.1802321821451187, 'rewards/repeat_floor/mean': 0.9887716174125671, 'rewards/repeat_floor/std': 0.016658827662467957, 'rewards/near_duplicate_penalty/mean': 0.9653176665306091, 'rewards/near_duplicate_penalty/std': 0.04678317531943321, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9330440759658813, 'rewards/distinct_2/std': 0.11006199568510056, 'rewards/total_composite/mean': 0.16526050865650177, 'rewards/total_composite/std': 0.12858688831329346, 'reward': 0.16526050865650177, 'reward_std': 0.12858688831329346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13404597342014313, 'sampling/sampling_logp_difference/max': 2.2708511352539062, 'sampling/importance_sampling_ratio/min': 0.10322428494691849, 'sampling/importance_sampling_ratio/mean': 1.0057851076126099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.734650507569313, 'clip_ratio/low_mean': 0.049931887770071626, 'clip_ratio/low_min': 0.049931887770071626, 'clip_ratio/high_mean': 0.05768774729222059, 'clip_ratio/high_max': 0.05768774729222059, 'clip_ratio/region_mean': 0.10761963506229222, 'reward_total_mean': 0.16526050865650177, 'reward_meter_mean': 0.6609570980072021, 'reward_meter_std': 0.3926407992839813, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9517898559570312, 'reward_lexical_plausibility_std': 0.13635894656181335, 'reward_repeat_penalty_mean': 0.8896200656890869, 'reward_repeat_penalty_std': 0.1802321821451187, 'reward_repeat_floor_mean': 0.9887716174125671, 'reward_repeat_floor_std': 0.016658827662467957, 'reward_near_duplicate_penalty_mean': 0.9653176665306091, 'reward_near_duplicate_penalty_std': 0.04678317531943321, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9330440759658813, 'reward_distinct_2_std': 0.11006199568510056, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.16526050865650177, 'reward_total_composite_std': 0.12858688831329346, 'epoch': 0.05}

 31%|███       | 1022/3300 [3:51:22<7:21:20, 11.62s/it]INFO 04-12 19:49:48 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1023/3300 [3:51:31<6:57:52, 11.01s/it]2026-04-12 19:49:57,819 | INFO | train_grpo_train | metrics_logged mode=train step=1023

                                                       
{'loss': -0.0068, 'grad_norm': 10.58643913269043, 'learning_rate': 6.903030303030304e-06, 'num_tokens': 1817268.0, 'completions/mean_length': 66.875, 'completions/min_length': 61.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9753686785697937, 'rewards/meter/std': 0.034470826387405396, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4662499725818634, 'rewards/judge_quality/std': 0.17435084283351898, 'rewards/repeat_penalty/mean': 0.8345916271209717, 'rewards/repeat_penalty/std': 0.09254156798124313, 'rewards/repeat_floor/mean': 0.9815016984939575, 'rewards/repeat_floor/std': 0.01192476786673069, 'rewards/near_duplicate_penalty/mean': 0.9519960880279541, 'rewards/near_duplicate_penalty/std': 0.04704742506146431, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8744693398475647, 'rewards/distinct_2/std': 0.058364640921354294, 'rewards/total_composite/mean': 0.4374304413795471, 'rewards/total_composite/std': 0.1469321846961975, 'reward': 0.4374304413795471, 'reward_std': 0.1469321846961975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14517521858215332, 'sampling/sampling_logp_difference/max': 2.105573892593384, 'sampling/importance_sampling_ratio/min': 0.12177576869726181, 'sampling/importance_sampling_ratio/mean': 1.0097715854644775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9674729481339455, 'clip_ratio/low_mean': 0.08652904396876693, 'clip_ratio/low_min': 0.08652904396876693, 'clip_ratio/high_mean': 0.034713342785835266, 'clip_ratio/high_max': 0.034713342785835266, 'clip_ratio/region_mean': 0.1212423867546022, 'reward_total_mean': 0.4374304413795471, 'reward_meter_mean': 0.9753686785697937, 'reward_meter_std': 0.034470826387405396, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8345916271209717, 'reward_repeat_penalty_std': 0.09254156798124313, 'reward_repeat_floor_mean': 0.9815016984939575, 'reward_repeat_floor_std': 0.01192476786673069, 'reward_near_duplicate_penalty_mean': 0.9519960880279541, 'reward_near_duplicate_penalty_std': 0.04704742506146431, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8744693398475647, 'reward_distinct_2_std': 0.058364640921354294, 'reward_judge_quality_mean': 0.4662499725818634, 'reward_judge_quality_std': 0.17435084283351898, 'reward_total_composite_mean': 0.4374304413795471, 'reward_total_composite_std': 0.1469321846961975, 'epoch': 0.05}

 31%|███       | 1023/3300 [3:51:31<6:57:52, 11.01s/it]INFO 04-12 19:49:58 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1024/3300 [3:51:45<7:27:20, 11.79s/it]2026-04-12 19:50:11,441 | INFO | train_grpo_train | metrics_logged mode=train step=1024

                                                       
{'loss': -0.0344, 'grad_norm': 5.592095851898193, 'learning_rate': 6.9e-06, 'num_tokens': 1819001.0, 'completions/mean_length': 108.625, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.000003814697266, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.5329739451408386, 'rewards/meter/std': 0.40733638405799866, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9345566630363464, 'rewards/lexical_plausibility/std': 0.18510174751281738, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.36371105909347534, 'rewards/repeat_penalty/mean': 0.8286395072937012, 'rewards/repeat_penalty/std': 0.17090271413326263, 'rewards/repeat_floor/mean': 0.9779955148696899, 'rewards/repeat_floor/std': 0.023956574499607086, 'rewards/near_duplicate_penalty/mean': 0.9162154197692871, 'rewards/near_duplicate_penalty/std': 0.10002227127552032, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8951464891433716, 'rewards/distinct_2/std': 0.10349778831005096, 'rewards/total_composite/mean': 0.27428188920021057, 'rewards/total_composite/std': 0.2637614607810974, 'reward': 0.27428188920021057, 'reward_std': 0.263761430978775, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14042514562606812, 'sampling/sampling_logp_difference/max': 3.151454448699951, 'sampling/importance_sampling_ratio/min': 0.042789846658706665, 'sampling/importance_sampling_ratio/mean': 1.0247440338134766, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6956662759184837, 'clip_ratio/low_mean': 0.03759612329304218, 'clip_ratio/low_min': 0.03759612329304218, 'clip_ratio/high_mean': 0.050842379219830036, 'clip_ratio/high_max': 0.050842379219830036, 'clip_ratio/region_mean': 0.08843850251287222, 'reward_total_mean': 0.27428188920021057, 'reward_meter_mean': 0.5329739451408386, 'reward_meter_std': 0.40733638405799866, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9345566630363464, 'reward_lexical_plausibility_std': 0.18510174751281738, 'reward_repeat_penalty_mean': 0.8286395072937012, 'reward_repeat_penalty_std': 0.17090271413326263, 'reward_repeat_floor_mean': 0.9779955148696899, 'reward_repeat_floor_std': 0.023956574499607086, 'reward_near_duplicate_penalty_mean': 0.9162154197692871, 'reward_near_duplicate_penalty_std': 0.10002227127552032, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8951464891433716, 'reward_distinct_2_std': 0.10349778831005096, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.36371105909347534, 'reward_total_composite_mean': 0.27428188920021057, 'reward_total_composite_std': 0.2637614607810974, 'epoch': 0.05}

 31%|███       | 1024/3300 [3:51:45<7:27:20, 11.79s/it]INFO 04-12 19:50:11 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1025/3300 [3:51:54<6:55:14, 10.95s/it]2026-04-12 19:50:20,424 | INFO | train_grpo_train | metrics_logged mode=train step=1025

                                                       
{'loss': -0.001, 'grad_norm': 13.66706371307373, 'learning_rate': 6.896969696969697e-06, 'num_tokens': 1820532.0, 'completions/mean_length': 38.375, 'completions/min_length': 34.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7257226705551147, 'rewards/meter/std': 0.40899670124053955, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6212500333786011, 'rewards/judge_quality/std': 0.2704724967479706, 'rewards/repeat_penalty/mean': 0.9570635557174683, 'rewards/repeat_penalty/std': 0.07852336764335632, 'rewards/repeat_floor/mean': 0.9944359660148621, 'rewards/repeat_floor/std': 0.009332636371254921, 'rewards/near_duplicate_penalty/mean': 0.9744448661804199, 'rewards/near_duplicate_penalty/std': 0.03071560338139534, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.37927505373954773, 'rewards/total_composite/std': 0.28317341208457947, 'reward': 0.37927505373954773, 'reward_std': 0.28317341208457947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15160652995109558, 'sampling/sampling_logp_difference/max': 1.9479103088378906, 'sampling/importance_sampling_ratio/min': 0.14257168769836426, 'sampling/importance_sampling_ratio/mean': 0.9949976801872253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8863668218255043, 'clip_ratio/low_mean': 0.08853199845179915, 'clip_ratio/low_min': 0.08853199845179915, 'clip_ratio/high_mean': 0.025894062593579292, 'clip_ratio/high_max': 0.025894062593579292, 'clip_ratio/region_mean': 0.11442606104537845, 'reward_total_mean': 0.37927505373954773, 'reward_meter_mean': 0.7257226705551147, 'reward_meter_std': 0.40899670124053955, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9570635557174683, 'reward_repeat_penalty_std': 0.07852336764335632, 'reward_repeat_floor_mean': 0.9944359660148621, 'reward_repeat_floor_std': 0.009332636371254921, 'reward_near_duplicate_penalty_mean': 0.9744448661804199, 'reward_near_duplicate_penalty_std': 0.03071560338139534, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.6212500333786011, 'reward_judge_quality_std': 0.2704724967479706, 'reward_total_composite_mean': 0.37927505373954773, 'reward_total_composite_std': 0.28317341208457947, 'epoch': 0.05}

 31%|███       | 1025/3300 [3:51:54<6:55:14, 10.95s/it]INFO 04-12 19:50:20 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1026/3300 [3:52:04<6:49:17, 10.80s/it]2026-04-12 19:50:30,877 | INFO | train_grpo_train | metrics_logged mode=train step=1026

                                                       
{'loss': 0.1032, 'grad_norm': 12.394335746765137, 'learning_rate': 6.893939393939395e-06, 'num_tokens': 1822868.0, 'completions/mean_length': 103.0, 'completions/min_length': 85.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.0, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.749626874923706, 'rewards/meter/std': 0.22611446678638458, 'rewards/count_adherence/mean': 0.9642857313156128, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9892857074737549, 'rewards/count_floor/std': 0.01983901485800743, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.7000752091407776, 'rewards/repeat_penalty/std': 0.2564781904220581, 'rewards/repeat_floor/mean': 0.9643749594688416, 'rewards/repeat_floor/std': 0.04147970303893089, 'rewards/near_duplicate_penalty/mean': 0.911603569984436, 'rewards/near_duplicate_penalty/std': 0.10075400769710541, 'rewards/opening_diversity/mean': 0.9553571343421936, 'rewards/opening_diversity/std': 0.12626907229423523, 'rewards/distinct_2/mean': 0.7812555432319641, 'rewards/distinct_2/std': 0.2117156982421875, 'rewards/total_composite/mean': 0.27177268266677856, 'rewards/total_composite/std': 0.09701024740934372, 'reward': 0.27177268266677856, 'reward_std': 0.09701025485992432, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16282899677753448, 'sampling/sampling_logp_difference/max': 1.6345796585083008, 'sampling/importance_sampling_ratio/min': 0.19503432512283325, 'sampling/importance_sampling_ratio/mean': 1.0140913724899292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9033370688557625, 'clip_ratio/low_mean': 0.058566014282405376, 'clip_ratio/low_min': 0.058566014282405376, 'clip_ratio/high_mean': 0.0894228583201766, 'clip_ratio/high_max': 0.0894228583201766, 'clip_ratio/region_mean': 0.14798887260258198, 'reward_total_mean': 0.27177268266677856, 'reward_meter_mean': 0.749626874923706, 'reward_meter_std': 0.22611446678638458, 'reward_count_adherence_mean': 0.9642857313156128, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9892857074737549, 'reward_count_floor_std': 0.01983901485800743, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7000752091407776, 'reward_repeat_penalty_std': 0.2564781904220581, 'reward_repeat_floor_mean': 0.9643749594688416, 'reward_repeat_floor_std': 0.04147970303893089, 'reward_near_duplicate_penalty_mean': 0.911603569984436, 'reward_near_duplicate_penalty_std': 0.10075400769710541, 'reward_opening_diversity_mean': 0.9553571343421936, 'reward_opening_diversity_std': 0.12626907229423523, 'reward_distinct_2_mean': 0.7812555432319641, 'reward_distinct_2_std': 0.2117156982421875, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.27177268266677856, 'reward_total_composite_std': 0.09701024740934372, 'epoch': 0.05}

 31%|███       | 1026/3300 [3:52:04<6:49:17, 10.80s/it]INFO 04-12 19:50:31 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1027/3300 [3:52:18<7:26:24, 11.78s/it]2026-04-12 19:50:44,943 | INFO | train_grpo_train | metrics_logged mode=train step=1027

                                                       
{'loss': -0.0351, 'grad_norm': 6.426838397979736, 'learning_rate': 6.890909090909092e-06, 'num_tokens': 1826362.0, 'completions/mean_length': 292.75, 'completions/min_length': 242.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 261.4285888671875, 'completions/min_terminated_length': 242.0, 'completions/max_terminated_length': 279.0, 'rewards/meter/mean': 0.7850620746612549, 'rewards/meter/std': 0.3157118260860443, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.08868516236543655, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9625000357627869, 'rewards/count_floor/std': 0.026605553925037384, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.1679842621088028, 'rewards/repeat_penalty/std': 0.1664513796567917, 'rewards/repeat_floor/mean': 0.9083834886550903, 'rewards/repeat_floor/std': 0.025251084938645363, 'rewards/near_duplicate_penalty/mean': 0.8698483109474182, 'rewards/near_duplicate_penalty/std': 0.06974322348833084, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.2041664719581604, 'rewards/distinct_2/std': 0.19479382038116455, 'rewards/total_composite/mean': 0.19514016807079315, 'rewards/total_composite/std': 0.10494624823331833, 'reward': 0.19514016807079315, 'reward_std': 0.10494624078273773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10818116366863251, 'sampling/sampling_logp_difference/max': 2.5145912170410156, 'sampling/importance_sampling_ratio/min': 0.08089597523212433, 'sampling/importance_sampling_ratio/mean': 1.0067905187606812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4590645357966423, 'clip_ratio/low_mean': 0.027754032518714666, 'clip_ratio/low_min': 0.027754032518714666, 'clip_ratio/high_mean': 0.05993103887885809, 'clip_ratio/high_max': 0.05993103887885809, 'clip_ratio/region_mean': 0.08768507139757276, 'reward_total_mean': 0.19514016807079315, 'reward_meter_mean': 0.7850620746612549, 'reward_meter_std': 0.3157118260860443, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.08868516236543655, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9625000357627869, 'reward_count_floor_std': 0.026605553925037384, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.1679842621088028, 'reward_repeat_penalty_std': 0.1664513796567917, 'reward_repeat_floor_mean': 0.9083834886550903, 'reward_repeat_floor_std': 0.025251084938645363, 'reward_near_duplicate_penalty_mean': 0.8698483109474182, 'reward_near_duplicate_penalty_std': 0.06974322348833084, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.2041664719581604, 'reward_distinct_2_std': 0.19479382038116455, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.19514016807079315, 'reward_total_composite_std': 0.10494624823331833, 'epoch': 0.05}

 31%|███       | 1027/3300 [3:52:18<7:26:24, 11.78s/it]INFO 04-12 19:50:45 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1028/3300 [3:52:32<7:44:46, 12.27s/it]2026-04-12 19:50:58,361 | INFO | train_grpo_train | metrics_logged mode=train step=1028

                                                       
{'loss': -0.0606, 'grad_norm': 4.061544895172119, 'learning_rate': 6.887878787878789e-06, 'num_tokens': 1827992.0, 'completions/mean_length': 96.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.42857360839844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7716999053955078, 'rewards/meter/std': 0.3507132828235626, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.932048499584198, 'rewards/lexical_plausibility/std': 0.192195862531662, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.7073366641998291, 'rewards/repeat_penalty/std': 0.15373623371124268, 'rewards/repeat_floor/mean': 0.9699632525444031, 'rewards/repeat_floor/std': 0.016544846817851067, 'rewards/near_duplicate_penalty/mean': 0.91960608959198, 'rewards/near_duplicate_penalty/std': 0.04806467518210411, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9044150114059448, 'rewards/distinct_2/std': 0.05708000436425209, 'rewards/total_composite/mean': 0.35969823598861694, 'rewards/total_composite/std': 0.25277888774871826, 'reward': 0.35969823598861694, 'reward_std': 0.25277888774871826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14263054728507996, 'sampling/sampling_logp_difference/max': 1.3940820693969727, 'sampling/importance_sampling_ratio/min': 0.24806064367294312, 'sampling/importance_sampling_ratio/mean': 1.0255366563796997, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8742780163884163, 'clip_ratio/low_mean': 0.06313793826848269, 'clip_ratio/low_min': 0.06313793826848269, 'clip_ratio/high_mean': 0.0343750009778887, 'clip_ratio/high_max': 0.0343750009778887, 'clip_ratio/region_mean': 0.09751293924637139, 'reward_total_mean': 0.35969823598861694, 'reward_meter_mean': 0.7716999053955078, 'reward_meter_std': 0.3507132828235626, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.932048499584198, 'reward_lexical_plausibility_std': 0.192195862531662, 'reward_repeat_penalty_mean': 0.7073366641998291, 'reward_repeat_penalty_std': 0.15373623371124268, 'reward_repeat_floor_mean': 0.9699632525444031, 'reward_repeat_floor_std': 0.016544846817851067, 'reward_near_duplicate_penalty_mean': 0.91960608959198, 'reward_near_duplicate_penalty_std': 0.04806467518210411, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9044150114059448, 'reward_distinct_2_std': 0.05708000436425209, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.35969823598861694, 'reward_total_composite_std': 0.25277888774871826, 'epoch': 0.05}

 31%|███       | 1028/3300 [3:52:32<7:44:46, 12.27s/it]INFO 04-12 19:50:58 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1029/3300 [3:52:45<7:56:25, 12.59s/it]2026-04-12 19:51:11,677 | INFO | train_grpo_train | metrics_logged mode=train step=1029

                                                       
{'loss': -0.0624, 'grad_norm': 3.5032265186309814, 'learning_rate': 6.8848484848484854e-06, 'num_tokens': 1829799.0, 'completions/mean_length': 180.875, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.668749213218689, 'rewards/meter/std': 0.4241233170032501, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.33905068039894104, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.934374988079071, 'rewards/count_floor/std': 0.10171520709991455, 'rewards/lexical_plausibility/mean': 0.9060420393943787, 'rewards/lexical_plausibility/std': 0.17521625757217407, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.8074693083763123, 'rewards/repeat_penalty/std': 0.18075814843177795, 'rewards/repeat_floor/mean': 0.978519082069397, 'rewards/repeat_floor/std': 0.019033988937735558, 'rewards/near_duplicate_penalty/mean': 0.9393673539161682, 'rewards/near_duplicate_penalty/std': 0.04634714126586914, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8675855398178101, 'rewards/distinct_2/std': 0.13593615591526031, 'rewards/total_composite/mean': 0.14603254199028015, 'rewards/total_composite/std': 0.1473514884710312, 'reward': 0.14603254199028015, 'reward_std': 0.1473514884710312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17409799993038177, 'sampling/sampling_logp_difference/max': 5.867944717407227, 'sampling/importance_sampling_ratio/min': 0.002828681143000722, 'sampling/importance_sampling_ratio/mean': 1.0262959003448486, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7266783341765404, 'clip_ratio/low_mean': 0.08197050169110298, 'clip_ratio/low_min': 0.08197050169110298, 'clip_ratio/high_mean': 0.04159544222056866, 'clip_ratio/high_max': 0.04159544222056866, 'clip_ratio/region_mean': 0.12356594391167164, 'reward_total_mean': 0.14603254199028015, 'reward_meter_mean': 0.668749213218689, 'reward_meter_std': 0.4241233170032501, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.33905068039894104, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.934374988079071, 'reward_count_floor_std': 0.10171520709991455, 'reward_lexical_plausibility_mean': 0.9060420393943787, 'reward_lexical_plausibility_std': 0.17521625757217407, 'reward_repeat_penalty_mean': 0.8074693083763123, 'reward_repeat_penalty_std': 0.18075814843177795, 'reward_repeat_floor_mean': 0.978519082069397, 'reward_repeat_floor_std': 0.019033988937735558, 'reward_near_duplicate_penalty_mean': 0.9393673539161682, 'reward_near_duplicate_penalty_std': 0.04634714126586914, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8675855398178101, 'reward_distinct_2_std': 0.13593615591526031, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.14603254199028015, 'reward_total_composite_std': 0.1473514884710312, 'epoch': 0.05}

 31%|███       | 1029/3300 [3:52:45<7:56:25, 12.59s/it]INFO 04-12 19:51:11 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1030/3300 [3:52:55<7:29:37, 11.88s/it]2026-04-12 19:51:21,923 | INFO | train_grpo_train | metrics_logged mode=train step=1030

                                                       
{'loss': 0.0701, 'grad_norm': 10.174373626708984, 'learning_rate': 6.881818181818183e-06, 'num_tokens': 1832335.0, 'completions/mean_length': 129.0, 'completions/min_length': 114.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.0, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.7598661184310913, 'rewards/meter/std': 0.10195587575435638, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.01178510207682848, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.8587624430656433, 'rewards/repeat_penalty/std': 0.1601112335920334, 'rewards/repeat_floor/mean': 0.9859421253204346, 'rewards/repeat_floor/std': 0.012671094387769699, 'rewards/near_duplicate_penalty/mean': 0.9625072479248047, 'rewards/near_duplicate_penalty/std': 0.020657068118453026, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9062890410423279, 'rewards/distinct_2/std': 0.11188486963510513, 'rewards/total_composite/mean': 0.361072301864624, 'rewards/total_composite/std': 0.19525328278541565, 'reward': 0.361072301864624, 'reward_std': 0.19525328278541565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16824322938919067, 'sampling/sampling_logp_difference/max': 3.170562744140625, 'sampling/importance_sampling_ratio/min': 0.041979968547821045, 'sampling/importance_sampling_ratio/mean': 0.9918917417526245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6613135971128941, 'clip_ratio/low_mean': 0.11224114056676626, 'clip_ratio/low_min': 0.11224114056676626, 'clip_ratio/high_mean': 0.02254098281264305, 'clip_ratio/high_max': 0.02254098281264305, 'clip_ratio/region_mean': 0.1347821233794093, 'reward_total_mean': 0.361072301864624, 'reward_meter_mean': 0.7598661184310913, 'reward_meter_std': 0.10195587575435638, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.01178510207682848, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8587624430656433, 'reward_repeat_penalty_std': 0.1601112335920334, 'reward_repeat_floor_mean': 0.9859421253204346, 'reward_repeat_floor_std': 0.012671094387769699, 'reward_near_duplicate_penalty_mean': 0.9625072479248047, 'reward_near_duplicate_penalty_std': 0.020657068118453026, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9062890410423279, 'reward_distinct_2_std': 0.11188486963510513, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.361072301864624, 'reward_total_composite_std': 0.19525328278541565, 'epoch': 0.05}

 31%|███       | 1030/3300 [3:52:55<7:29:37, 11.88s/it]INFO 04-12 19:51:22 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1031/3300 [3:53:04<6:52:54, 10.92s/it]2026-04-12 19:51:30,593 | INFO | train_grpo_train | metrics_logged mode=train step=1031

                                                       
{'loss': 0.0247, 'grad_norm': 12.100837707519531, 'learning_rate': 6.878787878787879e-06, 'num_tokens': 1834082.0, 'completions/mean_length': 68.375, 'completions/min_length': 18.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.99265056848526, 'rewards/meter/std': 0.003089885925874114, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.31160587072372437, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156250357627869, 'rewards/count_floor/std': 0.09348177164793015, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.724898099899292, 'rewards/repeat_penalty/std': 0.1089765653014183, 'rewards/repeat_floor/mean': 0.9708527326583862, 'rewards/repeat_floor/std': 0.016079066321253777, 'rewards/near_duplicate_penalty/mean': 0.9235653877258301, 'rewards/near_duplicate_penalty/std': 0.0582248792052269, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.845199704170227, 'rewards/distinct_2/std': 0.1347823441028595, 'rewards/total_composite/mean': 0.28273075819015503, 'rewards/total_composite/std': 0.13824431598186493, 'reward': 0.28273075819015503, 'reward_std': 0.13824431598186493, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15212556719779968, 'sampling/sampling_logp_difference/max': 1.730299472808838, 'sampling/importance_sampling_ratio/min': 0.1772313266992569, 'sampling/importance_sampling_ratio/mean': 1.0234389305114746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9372230246663094, 'clip_ratio/low_mean': 0.056656996719539165, 'clip_ratio/low_min': 0.056656996719539165, 'clip_ratio/high_mean': 0.07306154072284698, 'clip_ratio/high_max': 0.07306154072284698, 'clip_ratio/region_mean': 0.12971853744238615, 'reward_total_mean': 0.28273075819015503, 'reward_meter_mean': 0.99265056848526, 'reward_meter_std': 0.003089885925874114, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.31160587072372437, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156250357627869, 'reward_count_floor_std': 0.09348177164793015, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.724898099899292, 'reward_repeat_penalty_std': 0.1089765653014183, 'reward_repeat_floor_mean': 0.9708527326583862, 'reward_repeat_floor_std': 0.016079066321253777, 'reward_near_duplicate_penalty_mean': 0.9235653877258301, 'reward_near_duplicate_penalty_std': 0.0582248792052269, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.845199704170227, 'reward_distinct_2_std': 0.1347823441028595, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.28273075819015503, 'reward_total_composite_std': 0.13824431598186493, 'epoch': 0.05}

 31%|███       | 1031/3300 [3:53:04<6:52:54, 10.92s/it]INFO 04-12 19:51:30 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1032/3300 [3:53:17<7:22:45, 11.71s/it]2026-04-12 19:51:44,155 | INFO | train_grpo_train | metrics_logged mode=train step=1032

                                                       
{'loss': -0.1468, 'grad_norm': 3.566657543182373, 'learning_rate': 6.875757575757576e-06, 'num_tokens': 1836175.0, 'completions/mean_length': 128.625, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.85714721679688, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9461867213249207, 'rewards/meter/std': 0.06024564430117607, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 0.9610633850097656, 'rewards/lexical_plausibility/std': 0.11012942343950272, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.7090461850166321, 'rewards/repeat_penalty/std': 0.15860313177108765, 'rewards/repeat_floor/mean': 0.9681207537651062, 'rewards/repeat_floor/std': 0.01909528486430645, 'rewards/near_duplicate_penalty/mean': 0.918491542339325, 'rewards/near_duplicate_penalty/std': 0.04221451282501221, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.7920502424240112, 'rewards/distinct_2/std': 0.13314341008663177, 'rewards/total_composite/mean': 0.31375664472579956, 'rewards/total_composite/std': 0.12196913361549377, 'reward': 0.31375664472579956, 'reward_std': 0.12196913361549377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16305559873580933, 'sampling/sampling_logp_difference/max': 1.5273551940917969, 'sampling/importance_sampling_ratio/min': 0.21710912883281708, 'sampling/importance_sampling_ratio/mean': 1.0084128379821777, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0001589730381966, 'clip_ratio/low_mean': 0.013513513840734959, 'clip_ratio/low_min': 0.013513513840734959, 'clip_ratio/high_mean': 0.10479060281068087, 'clip_ratio/high_max': 0.10479060281068087, 'clip_ratio/region_mean': 0.11830411665141582, 'reward_total_mean': 0.31375664472579956, 'reward_meter_mean': 0.9461867213249207, 'reward_meter_std': 0.06024564430117607, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 0.9610633850097656, 'reward_lexical_plausibility_std': 0.11012942343950272, 'reward_repeat_penalty_mean': 0.7090461850166321, 'reward_repeat_penalty_std': 0.15860313177108765, 'reward_repeat_floor_mean': 0.9681207537651062, 'reward_repeat_floor_std': 0.01909528486430645, 'reward_near_duplicate_penalty_mean': 0.918491542339325, 'reward_near_duplicate_penalty_std': 0.04221451282501221, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.7920502424240112, 'reward_distinct_2_std': 0.13314341008663177, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.31375664472579956, 'reward_total_composite_std': 0.12196913361549377, 'epoch': 0.05}

 31%|███▏      | 1032/3300 [3:53:18<7:22:45, 11.71s/it]INFO 04-12 19:51:44 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1033/3300 [3:53:32<7:49:26, 12.42s/it]2026-04-12 19:51:58,238 | INFO | train_grpo_train | metrics_logged mode=train step=1033

                                                       
{'loss': -0.0589, 'grad_norm': 1.3373076915740967, 'learning_rate': 6.872727272727273e-06, 'num_tokens': 1837713.0, 'completions/mean_length': 338.25, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 48.66666793823242, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5539199113845825, 'rewards/meter/std': 0.3940742611885071, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.33034375309944153, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.09910311549901962, 'rewards/lexical_plausibility/mean': 0.7599345445632935, 'rewards/lexical_plausibility/std': 0.2054721713066101, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.8813475370407104, 'rewards/repeat_penalty/std': 0.11809955537319183, 'rewards/repeat_floor/mean': 0.9911878108978271, 'rewards/repeat_floor/std': 0.007555747404694557, 'rewards/near_duplicate_penalty/mean': 0.9853224158287048, 'rewards/near_duplicate_penalty/std': 0.024444254115223885, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9890491962432861, 'rewards/distinct_2/std': 0.029218412935733795, 'rewards/total_composite/mean': 0.12289460003376007, 'rewards/total_composite/std': 0.1493535190820694, 'reward': 0.12289460003376007, 'reward_std': 0.1493535190820694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1713695079088211, 'sampling/sampling_logp_difference/max': 1.3529753684997559, 'sampling/importance_sampling_ratio/min': 0.2584700584411621, 'sampling/importance_sampling_ratio/mean': 1.0064082145690918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44355134665966034, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07733955420553684, 'clip_ratio/high_max': 0.07733955420553684, 'clip_ratio/region_mean': 0.07733955420553684, 'reward_total_mean': 0.12289460003376007, 'reward_meter_mean': 0.5539199113845825, 'reward_meter_std': 0.3940742611885071, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.33034375309944153, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.09910311549901962, 'reward_lexical_plausibility_mean': 0.7599345445632935, 'reward_lexical_plausibility_std': 0.2054721713066101, 'reward_repeat_penalty_mean': 0.8813475370407104, 'reward_repeat_penalty_std': 0.11809955537319183, 'reward_repeat_floor_mean': 0.9911878108978271, 'reward_repeat_floor_std': 0.007555747404694557, 'reward_near_duplicate_penalty_mean': 0.9853224158287048, 'reward_near_duplicate_penalty_std': 0.024444254115223885, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9890491962432861, 'reward_distinct_2_std': 0.029218412935733795, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.12289460003376007, 'reward_total_composite_std': 0.1493535190820694, 'epoch': 0.05}

 31%|███▏      | 1033/3300 [3:53:32<7:49:26, 12.42s/it]INFO 04-12 19:51:58 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1034/3300 [3:53:45<8:03:53, 12.81s/it]2026-04-12 19:52:11,955 | INFO | train_grpo_train | metrics_logged mode=train step=1034

                                                       
{'loss': -0.1268, 'grad_norm': 3.4945671558380127, 'learning_rate': 6.869696969696971e-06, 'num_tokens': 1840223.0, 'completions/mean_length': 231.75, 'completions/min_length': 117.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 138.33334350585938, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.8629274368286133, 'rewards/meter/std': 0.17817242443561554, 'rewards/count_adherence/mean': 0.8214285373687744, 'rewards/count_adherence/std': 0.06613000482320786, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9464285373687744, 'rewards/count_floor/std': 0.019838986918330193, 'rewards/lexical_plausibility/mean': 0.9097108244895935, 'rewards/lexical_plausibility/std': 0.1690024882555008, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.5891968607902527, 'rewards/repeat_penalty/std': 0.21916449069976807, 'rewards/repeat_floor/mean': 0.951640248298645, 'rewards/repeat_floor/std': 0.02437850832939148, 'rewards/near_duplicate_penalty/mean': 0.8651718497276306, 'rewards/near_duplicate_penalty/std': 0.04946175217628479, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.6873830556869507, 'rewards/distinct_2/std': 0.21976767480373383, 'rewards/total_composite/mean': 0.1882256269454956, 'rewards/total_composite/std': 0.1159299686551094, 'reward': 0.1882256269454956, 'reward_std': 0.1159299686551094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12832891941070557, 'sampling/sampling_logp_difference/max': 4.318770408630371, 'sampling/importance_sampling_ratio/min': 0.013316246680915356, 'sampling/importance_sampling_ratio/mean': 1.0024572610855103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6344615891575813, 'clip_ratio/low_mean': 0.028120783157646656, 'clip_ratio/low_min': 0.028120783157646656, 'clip_ratio/high_mean': 0.07120212726294994, 'clip_ratio/high_max': 0.07120212726294994, 'clip_ratio/region_mean': 0.0993229104205966, 'reward_total_mean': 0.1882256269454956, 'reward_meter_mean': 0.8629274368286133, 'reward_meter_std': 0.17817242443561554, 'reward_count_adherence_mean': 0.8214285373687744, 'reward_count_adherence_std': 0.06613000482320786, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9464285373687744, 'reward_count_floor_std': 0.019838986918330193, 'reward_lexical_plausibility_mean': 0.9097108244895935, 'reward_lexical_plausibility_std': 0.1690024882555008, 'reward_repeat_penalty_mean': 0.5891968607902527, 'reward_repeat_penalty_std': 0.21916449069976807, 'reward_repeat_floor_mean': 0.951640248298645, 'reward_repeat_floor_std': 0.02437850832939148, 'reward_near_duplicate_penalty_mean': 0.8651718497276306, 'reward_near_duplicate_penalty_std': 0.04946175217628479, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.6873830556869507, 'reward_distinct_2_std': 0.21976767480373383, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.1882256269454956, 'reward_total_composite_std': 0.1159299686551094, 'epoch': 0.05}

 31%|███▏      | 1034/3300 [3:53:45<8:03:53, 12.81s/it]INFO 04-12 19:52:12 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1035/3300 [3:54:00<8:22:34, 13.31s/it]2026-04-12 19:52:26,437 | INFO | train_grpo_train | metrics_logged mode=train step=1035

                                                       
{'loss': -0.106, 'grad_norm': 4.426401138305664, 'learning_rate': 6.866666666666667e-06, 'num_tokens': 1842176.0, 'completions/mean_length': 134.125, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.14286041259766, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.6368252038955688, 'rewards/meter/std': 0.3716123402118683, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9403793215751648, 'rewards/lexical_plausibility/std': 0.16863277554512024, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9268494844436646, 'rewards/repeat_penalty/std': 0.04847259446978569, 'rewards/repeat_floor/mean': 0.9913119673728943, 'rewards/repeat_floor/std': 0.005451911594718695, 'rewards/near_duplicate_penalty/mean': 0.9675188660621643, 'rewards/near_duplicate_penalty/std': 0.018193546682596207, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9576012492179871, 'rewards/distinct_2/std': 0.037529367953538895, 'rewards/total_composite/mean': 0.23162199556827545, 'rewards/total_composite/std': 0.14037111401557922, 'reward': 0.23162199556827545, 'reward_std': 0.14037111401557922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1483355313539505, 'sampling/sampling_logp_difference/max': 1.8519296646118164, 'sampling/importance_sampling_ratio/min': 0.19051046669483185, 'sampling/importance_sampling_ratio/mean': 1.0096585750579834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8855811953544617, 'clip_ratio/low_mean': 0.04130448121577501, 'clip_ratio/low_min': 0.04130448121577501, 'clip_ratio/high_mean': 0.073572619818151, 'clip_ratio/high_max': 0.073572619818151, 'clip_ratio/region_mean': 0.11487710103392601, 'reward_total_mean': 0.23162199556827545, 'reward_meter_mean': 0.6368252038955688, 'reward_meter_std': 0.3716123402118683, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9403793215751648, 'reward_lexical_plausibility_std': 0.16863277554512024, 'reward_repeat_penalty_mean': 0.9268494844436646, 'reward_repeat_penalty_std': 0.04847259446978569, 'reward_repeat_floor_mean': 0.9913119673728943, 'reward_repeat_floor_std': 0.005451911594718695, 'reward_near_duplicate_penalty_mean': 0.9675188660621643, 'reward_near_duplicate_penalty_std': 0.018193546682596207, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9576012492179871, 'reward_distinct_2_std': 0.037529367953538895, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.23162199556827545, 'reward_total_composite_std': 0.14037111401557922, 'epoch': 0.05}

 31%|███▏      | 1035/3300 [3:54:00<8:22:34, 13.31s/it]INFO 04-12 19:52:26 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1036/3300 [3:54:14<8:31:15, 13.55s/it]2026-04-12 19:52:40,540 | INFO | train_grpo_train | metrics_logged mode=train step=1036

                                                       
{'loss': -0.0467, 'grad_norm': 5.379473686218262, 'learning_rate': 6.8636363636363645e-06, 'num_tokens': 1843858.0, 'completions/mean_length': 110.25, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.85714340209961, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.559983491897583, 'rewards/meter/std': 0.35088860988616943, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9512271881103516, 'rewards/lexical_plausibility/std': 0.13795040547847748, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.858528733253479, 'rewards/repeat_penalty/std': 0.19305704534053802, 'rewards/repeat_floor/mean': 0.983191728591919, 'rewards/repeat_floor/std': 0.022784627974033356, 'rewards/near_duplicate_penalty/mean': 0.9497088193893433, 'rewards/near_duplicate_penalty/std': 0.05611324682831764, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9178939461708069, 'rewards/distinct_2/std': 0.11625760793685913, 'rewards/total_composite/mean': 0.12528550624847412, 'rewards/total_composite/std': 0.1045026183128357, 'reward': 0.12528550624847412, 'reward_std': 0.1045026183128357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16183240711688995, 'sampling/sampling_logp_difference/max': 1.2830981016159058, 'sampling/importance_sampling_ratio/min': 0.27717724442481995, 'sampling/importance_sampling_ratio/mean': 1.0346561670303345, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3770179897546768, 'clip_ratio/low_mean': 0.04216533713042736, 'clip_ratio/low_min': 0.04216533713042736, 'clip_ratio/high_mean': 0.11152931675314903, 'clip_ratio/high_max': 0.11152931675314903, 'clip_ratio/region_mean': 0.1536946538835764, 'reward_total_mean': 0.12528550624847412, 'reward_meter_mean': 0.559983491897583, 'reward_meter_std': 0.35088860988616943, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9512271881103516, 'reward_lexical_plausibility_std': 0.13795040547847748, 'reward_repeat_penalty_mean': 0.858528733253479, 'reward_repeat_penalty_std': 0.19305704534053802, 'reward_repeat_floor_mean': 0.983191728591919, 'reward_repeat_floor_std': 0.022784627974033356, 'reward_near_duplicate_penalty_mean': 0.9497088193893433, 'reward_near_duplicate_penalty_std': 0.05611324682831764, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9178939461708069, 'reward_distinct_2_std': 0.11625760793685913, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.12528550624847412, 'reward_total_composite_std': 0.1045026183128357, 'epoch': 0.05}

 31%|███▏      | 1036/3300 [3:54:14<8:31:15, 13.55s/it]INFO 04-12 19:52:40 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1037/3300 [3:54:28<8:34:07, 13.63s/it]2026-04-12 19:52:54,361 | INFO | train_grpo_train | metrics_logged mode=train step=1037

                                                       
{'loss': -0.0856, 'grad_norm': 2.698458433151245, 'learning_rate': 6.860606060606061e-06, 'num_tokens': 1845916.0, 'completions/mean_length': 322.25, 'completions/min_length': 119.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 132.5, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.3715631365776062, 'rewards/meter/std': 0.35139113664627075, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.2781743109226227, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9124999642372131, 'rewards/count_floor/std': 0.08345229178667068, 'rewards/lexical_plausibility/mean': 0.7954964637756348, 'rewards/lexical_plausibility/std': 0.22257614135742188, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.8200597763061523, 'rewards/repeat_penalty/std': 0.19439402222633362, 'rewards/repeat_floor/mean': 0.9812372326850891, 'rewards/repeat_floor/std': 0.01872258074581623, 'rewards/near_duplicate_penalty/mean': 0.955824077129364, 'rewards/near_duplicate_penalty/std': 0.03457092121243477, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8651514053344727, 'rewards/distinct_2/std': 0.16505487263202667, 'rewards/total_composite/mean': 0.03896675631403923, 'rewards/total_composite/std': 0.025124112144112587, 'reward': 0.03896675631403923, 'reward_std': 0.025124112144112587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1449243426322937, 'sampling/sampling_logp_difference/max': 2.939000129699707, 'sampling/importance_sampling_ratio/min': 0.05291861295700073, 'sampling/importance_sampling_ratio/mean': 0.9852445721626282, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4010564833879471, 'clip_ratio/low_mean': 0.031799906864762306, 'clip_ratio/low_min': 0.031799906864762306, 'clip_ratio/high_mean': 0.044992998242378235, 'clip_ratio/high_max': 0.044992998242378235, 'clip_ratio/region_mean': 0.07679290510714054, 'reward_total_mean': 0.03896675631403923, 'reward_meter_mean': 0.3715631365776062, 'reward_meter_std': 0.35139113664627075, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.2781743109226227, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9124999642372131, 'reward_count_floor_std': 0.08345229178667068, 'reward_lexical_plausibility_mean': 0.7954964637756348, 'reward_lexical_plausibility_std': 0.22257614135742188, 'reward_repeat_penalty_mean': 0.8200597763061523, 'reward_repeat_penalty_std': 0.19439402222633362, 'reward_repeat_floor_mean': 0.9812372326850891, 'reward_repeat_floor_std': 0.01872258074581623, 'reward_near_duplicate_penalty_mean': 0.955824077129364, 'reward_near_duplicate_penalty_std': 0.03457092121243477, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8651514053344727, 'reward_distinct_2_std': 0.16505487263202667, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.03896675631403923, 'reward_total_composite_std': 0.025124112144112587, 'epoch': 0.05}

 31%|███▏      | 1037/3300 [3:54:28<8:34:07, 13.63s/it]INFO 04-12 19:52:54 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1038/3300 [3:54:37<7:48:30, 12.43s/it]2026-04-12 19:53:03,978 | INFO | train_grpo_train | metrics_logged mode=train step=1038

                                                       
{'loss': 0.1119, 'grad_norm': 11.717079162597656, 'learning_rate': 6.857575757575758e-06, 'num_tokens': 1848327.0, 'completions/mean_length': 101.375, 'completions/min_length': 87.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.375, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.9379615783691406, 'rewards/meter/std': 0.14111998677253723, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9700000286102295, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.6559946537017822, 'rewards/repeat_penalty/std': 0.2319997400045395, 'rewards/repeat_floor/mean': 0.9619393348693848, 'rewards/repeat_floor/std': 0.024726862087845802, 'rewards/near_duplicate_penalty/mean': 0.9073032140731812, 'rewards/near_duplicate_penalty/std': 0.046702027320861816, 'rewards/opening_diversity/mean': 0.9119791388511658, 'rewards/opening_diversity/std': 0.15455128252506256, 'rewards/distinct_2/mean': 0.7902786731719971, 'rewards/distinct_2/std': 0.1557294875383377, 'rewards/total_composite/mean': 0.2675013840198517, 'rewards/total_composite/std': 0.09637827426195145, 'reward': 0.2675013840198517, 'reward_std': 0.09637827426195145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13963530957698822, 'sampling/sampling_logp_difference/max': 1.5723917484283447, 'sampling/importance_sampling_ratio/min': 0.20754818618297577, 'sampling/importance_sampling_ratio/mean': 1.0107676982879639, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8562929853796959, 'clip_ratio/low_mean': 0.037640057504177094, 'clip_ratio/low_min': 0.037640057504177094, 'clip_ratio/high_mean': 0.09844038728624582, 'clip_ratio/high_max': 0.09844038728624582, 'clip_ratio/region_mean': 0.13608044479042292, 'reward_total_mean': 0.2675013840198517, 'reward_meter_mean': 0.9379615783691406, 'reward_meter_std': 0.14111998677253723, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9700000286102295, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6559946537017822, 'reward_repeat_penalty_std': 0.2319997400045395, 'reward_repeat_floor_mean': 0.9619393348693848, 'reward_repeat_floor_std': 0.024726862087845802, 'reward_near_duplicate_penalty_mean': 0.9073032140731812, 'reward_near_duplicate_penalty_std': 0.046702027320861816, 'reward_opening_diversity_mean': 0.9119791388511658, 'reward_opening_diversity_std': 0.15455128252506256, 'reward_distinct_2_mean': 0.7902786731719971, 'reward_distinct_2_std': 0.1557294875383377, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.2675013840198517, 'reward_total_composite_std': 0.09637827426195145, 'epoch': 0.05}

 31%|███▏      | 1038/3300 [3:54:37<7:48:30, 12.43s/it]INFO 04-12 19:53:04 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1039/3300 [3:54:47<7:21:38, 11.72s/it]2026-04-12 19:53:14,052 | INFO | train_grpo_train | metrics_logged mode=train step=1039

                                                       
{'loss': 0.0451, 'grad_norm': 7.77894926071167, 'learning_rate': 6.854545454545455e-06, 'num_tokens': 1850965.0, 'completions/mean_length': 132.75, 'completions/min_length': 123.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.75, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.7572954893112183, 'rewards/meter/std': 0.2965412735939026, 'rewards/count_adherence/mean': 0.890625, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9671874642372131, 'rewards/count_floor/std': 0.013258260674774647, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.37820857763290405, 'rewards/repeat_penalty/std': 0.2359757274389267, 'rewards/repeat_floor/mean': 0.9255591034889221, 'rewards/repeat_floor/std': 0.03798024356365204, 'rewards/near_duplicate_penalty/mean': 0.8441143035888672, 'rewards/near_duplicate_penalty/std': 0.08639401942491531, 'rewards/opening_diversity/mean': 0.9079241156578064, 'rewards/opening_diversity/std': 0.13796918094158173, 'rewards/distinct_2/mean': 0.4940720498561859, 'rewards/distinct_2/std': 0.2183462381362915, 'rewards/total_composite/mean': 0.1361137479543686, 'rewards/total_composite/std': 0.08358364552259445, 'reward': 0.1361137479543686, 'reward_std': 0.08358364552259445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1351195126771927, 'sampling/sampling_logp_difference/max': 2.775458335876465, 'sampling/importance_sampling_ratio/min': 0.062320906668901443, 'sampling/importance_sampling_ratio/mean': 1.0014041662216187, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6944403424859047, 'clip_ratio/low_mean': 0.05950515950098634, 'clip_ratio/low_min': 0.05950515950098634, 'clip_ratio/high_mean': 0.05474177189171314, 'clip_ratio/high_max': 0.05474177189171314, 'clip_ratio/region_mean': 0.11424693139269948, 'reward_total_mean': 0.1361137479543686, 'reward_meter_mean': 0.7572954893112183, 'reward_meter_std': 0.2965412735939026, 'reward_count_adherence_mean': 0.890625, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9671874642372131, 'reward_count_floor_std': 0.013258260674774647, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.37820857763290405, 'reward_repeat_penalty_std': 0.2359757274389267, 'reward_repeat_floor_mean': 0.9255591034889221, 'reward_repeat_floor_std': 0.03798024356365204, 'reward_near_duplicate_penalty_mean': 0.8441143035888672, 'reward_near_duplicate_penalty_std': 0.08639401942491531, 'reward_opening_diversity_mean': 0.9079241156578064, 'reward_opening_diversity_std': 0.13796918094158173, 'reward_distinct_2_mean': 0.4940720498561859, 'reward_distinct_2_std': 0.2183462381362915, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.1361137479543686, 'reward_total_composite_std': 0.08358364552259445, 'epoch': 0.05}

 31%|███▏      | 1039/3300 [3:54:47<7:21:38, 11.72s/it]INFO 04-12 19:53:14 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1040/3300 [3:54:57<7:02:53, 11.23s/it]2026-04-12 19:53:24,130 | INFO | train_grpo_train | metrics_logged mode=train step=1040

                                                       
{'loss': 0.0123, 'grad_norm': 9.252738952636719, 'learning_rate': 6.851515151515153e-06, 'num_tokens': 1853566.0, 'completions/mean_length': 129.125, 'completions/min_length': 107.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.125, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.9835991263389587, 'rewards/meter/std': 0.02488532103598118, 'rewards/count_adherence/mean': 0.8035714030265808, 'rewards/count_adherence/std': 0.07393559068441391, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9410713911056519, 'rewards/count_floor/std': 0.02218066342175007, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.4110645055770874, 'rewards/repeat_penalty/std': 0.22772547602653503, 'rewards/repeat_floor/mean': 0.9282637238502502, 'rewards/repeat_floor/std': 0.03510439395904541, 'rewards/near_duplicate_penalty/mean': 0.8382090330123901, 'rewards/near_duplicate_penalty/std': 0.08379925787448883, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1767766922712326, 'rewards/distinct_2/mean': 0.514248788356781, 'rewards/distinct_2/std': 0.1942049264907837, 'rewards/total_composite/mean': 0.18888162076473236, 'rewards/total_composite/std': 0.08576741814613342, 'reward': 0.18888162076473236, 'reward_std': 0.08576741814613342, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13730427622795105, 'sampling/sampling_logp_difference/max': 2.3340024948120117, 'sampling/importance_sampling_ratio/min': 0.09690709412097931, 'sampling/importance_sampling_ratio/mean': 1.0222182273864746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7903077602386475, 'clip_ratio/low_mean': 0.06582424882799387, 'clip_ratio/low_min': 0.06582424882799387, 'clip_ratio/high_mean': 0.041744185611605644, 'clip_ratio/high_max': 0.041744185611605644, 'clip_ratio/region_mean': 0.10756843443959951, 'reward_total_mean': 0.18888162076473236, 'reward_meter_mean': 0.9835991263389587, 'reward_meter_std': 0.02488532103598118, 'reward_count_adherence_mean': 0.8035714030265808, 'reward_count_adherence_std': 0.07393559068441391, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9410713911056519, 'reward_count_floor_std': 0.02218066342175007, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4110645055770874, 'reward_repeat_penalty_std': 0.22772547602653503, 'reward_repeat_floor_mean': 0.9282637238502502, 'reward_repeat_floor_std': 0.03510439395904541, 'reward_near_duplicate_penalty_mean': 0.8382090330123901, 'reward_near_duplicate_penalty_std': 0.08379925787448883, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1767766922712326, 'reward_distinct_2_mean': 0.514248788356781, 'reward_distinct_2_std': 0.1942049264907837, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.18888162076473236, 'reward_total_composite_std': 0.08576741814613342, 'epoch': 0.05}

 32%|███▏      | 1040/3300 [3:54:57<7:02:53, 11.23s/it]INFO 04-12 19:53:24 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1041/3300 [3:55:11<7:30:35, 11.97s/it]2026-04-12 19:53:37,826 | INFO | train_grpo_train | metrics_logged mode=train step=1041

                                                       
{'loss': -0.0459, 'grad_norm': 2.820596694946289, 'learning_rate': 6.848484848484849e-06, 'num_tokens': 1855036.0, 'completions/mean_length': 156.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 38.333335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.39307206869125366, 'rewards/meter/std': 0.37239328026771545, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8818939924240112, 'rewards/lexical_plausibility/std': 0.22319026291370392, 'rewards/judge_quality/mean': 0.35874998569488525, 'rewards/judge_quality/std': 0.2864281237125397, 'rewards/repeat_penalty/mean': 0.8335074782371521, 'rewards/repeat_penalty/std': 0.14929601550102234, 'rewards/repeat_floor/mean': 0.980891764163971, 'rewards/repeat_floor/std': 0.017063746228814125, 'rewards/near_duplicate_penalty/mean': 0.9217875003814697, 'rewards/near_duplicate_penalty/std': 0.079309843480587, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9597069025039673, 'rewards/distinct_2/std': 0.0338943749666214, 'rewards/total_composite/mean': 0.10745798796415329, 'rewards/total_composite/std': 0.12441365420818329, 'reward': 0.10745798796415329, 'reward_std': 0.12441365420818329, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15830090641975403, 'sampling/sampling_logp_difference/max': 1.7270796298980713, 'sampling/importance_sampling_ratio/min': 0.1778029054403305, 'sampling/importance_sampling_ratio/mean': 1.024382472038269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9981816112995148, 'clip_ratio/low_mean': 0.05997073370963335, 'clip_ratio/low_min': 0.05997073370963335, 'clip_ratio/high_mean': 0.07817834243178368, 'clip_ratio/high_max': 0.07817834243178368, 'clip_ratio/region_mean': 0.13814907614141703, 'reward_total_mean': 0.10745798796415329, 'reward_meter_mean': 0.39307206869125366, 'reward_meter_std': 0.37239328026771545, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8818939924240112, 'reward_lexical_plausibility_std': 0.22319026291370392, 'reward_repeat_penalty_mean': 0.8335074782371521, 'reward_repeat_penalty_std': 0.14929601550102234, 'reward_repeat_floor_mean': 0.980891764163971, 'reward_repeat_floor_std': 0.017063746228814125, 'reward_near_duplicate_penalty_mean': 0.9217875003814697, 'reward_near_duplicate_penalty_std': 0.079309843480587, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9597069025039673, 'reward_distinct_2_std': 0.0338943749666214, 'reward_judge_quality_mean': 0.35874998569488525, 'reward_judge_quality_std': 0.2864281237125397, 'reward_total_composite_mean': 0.10745798796415329, 'reward_total_composite_std': 0.12441365420818329, 'epoch': 0.05}

 32%|███▏      | 1041/3300 [3:55:11<7:30:35, 11.97s/it]INFO 04-12 19:53:38 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1042/3300 [3:55:27<8:17:39, 13.22s/it]2026-04-12 19:53:53,978 | INFO | train_grpo_train | metrics_logged mode=train step=1042

                                                       
{'loss': -0.0754, 'grad_norm': 4.727536201477051, 'learning_rate': 6.845454545454546e-06, 'num_tokens': 1856688.0, 'completions/mean_length': 103.5, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.142860412597656, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.5524774789810181, 'rewards/meter/std': 0.31214046478271484, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9506340026855469, 'rewards/lexical_plausibility/std': 0.1396281123161316, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.9423285126686096, 'rewards/repeat_penalty/std': 0.09359228610992432, 'rewards/repeat_floor/mean': 0.9948512315750122, 'rewards/repeat_floor/std': 0.007179405074566603, 'rewards/near_duplicate_penalty/mean': 0.9861325621604919, 'rewards/near_duplicate_penalty/std': 0.020935945212841034, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9867373704910278, 'rewards/distinct_2/std': 0.03751229867339134, 'rewards/total_composite/mean': 0.1782945692539215, 'rewards/total_composite/std': 0.14102791249752045, 'reward': 0.1782945692539215, 'reward_std': 0.14102791249752045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16874884068965912, 'sampling/sampling_logp_difference/max': 2.3880558013916016, 'sampling/importance_sampling_ratio/min': 0.0918080061674118, 'sampling/importance_sampling_ratio/mean': 1.0198297500610352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1007827296853065, 'clip_ratio/low_mean': 0.0689714988693595, 'clip_ratio/low_min': 0.0689714988693595, 'clip_ratio/high_mean': 0.05714285932481289, 'clip_ratio/high_max': 0.05714285932481289, 'clip_ratio/region_mean': 0.12611435819417238, 'reward_total_mean': 0.1782945692539215, 'reward_meter_mean': 0.5524774789810181, 'reward_meter_std': 0.31214046478271484, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9506340026855469, 'reward_lexical_plausibility_std': 0.1396281123161316, 'reward_repeat_penalty_mean': 0.9423285126686096, 'reward_repeat_penalty_std': 0.09359228610992432, 'reward_repeat_floor_mean': 0.9948512315750122, 'reward_repeat_floor_std': 0.007179405074566603, 'reward_near_duplicate_penalty_mean': 0.9861325621604919, 'reward_near_duplicate_penalty_std': 0.020935945212841034, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9867373704910278, 'reward_distinct_2_std': 0.03751229867339134, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.1782945692539215, 'reward_total_composite_std': 0.14102791249752045, 'epoch': 0.05}

 32%|███▏      | 1042/3300 [3:55:27<8:17:39, 13.22s/it]INFO 04-12 19:53:54 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1043/3300 [3:55:41<8:26:46, 13.47s/it]2026-04-12 19:54:08,035 | INFO | train_grpo_train | metrics_logged mode=train step=1043

                                                       
{'loss': -0.1677, 'grad_norm': 4.966696262359619, 'learning_rate': 6.842424242424243e-06, 'num_tokens': 1859480.0, 'completions/mean_length': 194.0, 'completions/min_length': 132.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 148.57144165039062, 'completions/min_terminated_length': 132.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.7360803484916687, 'rewards/meter/std': 0.38025182485580444, 'rewards/count_adherence/mean': 0.671875, 'rewards/count_adherence/std': 0.23085150122642517, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9015624523162842, 'rewards/count_floor/std': 0.06925544887781143, 'rewards/lexical_plausibility/mean': 0.9519443511962891, 'rewards/lexical_plausibility/std': 0.1359219253063202, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.5964800119400024, 'rewards/repeat_penalty/std': 0.24976219236850739, 'rewards/repeat_floor/mean': 0.9522773027420044, 'rewards/repeat_floor/std': 0.039006780833005905, 'rewards/near_duplicate_penalty/mean': 0.8929007053375244, 'rewards/near_duplicate_penalty/std': 0.08897947520017624, 'rewards/opening_diversity/mean': 0.827556848526001, 'rewards/opening_diversity/std': 0.16875314712524414, 'rewards/distinct_2/mean': 0.7632088661193848, 'rewards/distinct_2/std': 0.20450684428215027, 'rewards/total_composite/mean': 0.2121185064315796, 'rewards/total_composite/std': 0.12680518627166748, 'reward': 0.2121185064315796, 'reward_std': 0.12680518627166748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1323598176240921, 'sampling/sampling_logp_difference/max': 1.90132737159729, 'sampling/importance_sampling_ratio/min': 0.1493702232837677, 'sampling/importance_sampling_ratio/mean': 1.0072625875473022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7059270441532135, 'clip_ratio/low_mean': 0.028657877817749977, 'clip_ratio/low_min': 0.028657877817749977, 'clip_ratio/high_mean': 0.07214623875916004, 'clip_ratio/high_max': 0.07214623875916004, 'clip_ratio/region_mean': 0.10080411657691002, 'reward_total_mean': 0.2121185064315796, 'reward_meter_mean': 0.7360803484916687, 'reward_meter_std': 0.38025182485580444, 'reward_count_adherence_mean': 0.671875, 'reward_count_adherence_std': 0.23085150122642517, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9015624523162842, 'reward_count_floor_std': 0.06925544887781143, 'reward_lexical_plausibility_mean': 0.9519443511962891, 'reward_lexical_plausibility_std': 0.1359219253063202, 'reward_repeat_penalty_mean': 0.5964800119400024, 'reward_repeat_penalty_std': 0.24976219236850739, 'reward_repeat_floor_mean': 0.9522773027420044, 'reward_repeat_floor_std': 0.039006780833005905, 'reward_near_duplicate_penalty_mean': 0.8929007053375244, 'reward_near_duplicate_penalty_std': 0.08897947520017624, 'reward_opening_diversity_mean': 0.827556848526001, 'reward_opening_diversity_std': 0.16875314712524414, 'reward_distinct_2_mean': 0.7632088661193848, 'reward_distinct_2_std': 0.20450684428215027, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.2121185064315796, 'reward_total_composite_std': 0.12680518627166748, 'epoch': 0.05}

 32%|███▏      | 1043/3300 [3:55:41<8:26:46, 13.47s/it]INFO 04-12 19:54:08 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1044/3300 [3:55:58<9:01:54, 14.41s/it]2026-04-12 19:54:24,642 | INFO | train_grpo_train | metrics_logged mode=train step=1044

                                                       
{'loss': -0.012, 'grad_norm': 6.336915016174316, 'learning_rate': 6.83939393939394e-06, 'num_tokens': 1860958.0, 'completions/mean_length': 98.75, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.4518135190010071, 'rewards/meter/std': 0.39509329199790955, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9591144323348999, 'rewards/lexical_plausibility/std': 0.1156417652964592, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.23627693951129913, 'rewards/repeat_penalty/mean': 0.8956582546234131, 'rewards/repeat_penalty/std': 0.1424568146467209, 'rewards/repeat_floor/mean': 0.9878213405609131, 'rewards/repeat_floor/std': 0.015821415930986404, 'rewards/near_duplicate_penalty/mean': 0.9531832933425903, 'rewards/near_duplicate_penalty/std': 0.05720429867506027, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9427089691162109, 'rewards/distinct_2/std': 0.08827294409275055, 'rewards/total_composite/mean': 0.22185459733009338, 'rewards/total_composite/std': 0.28892993927001953, 'reward': 0.22185459733009338, 'reward_std': 0.2889299690723419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.198969766497612, 'sampling/sampling_logp_difference/max': 1.3652875423431396, 'sampling/importance_sampling_ratio/min': 0.25530725717544556, 'sampling/importance_sampling_ratio/mean': 1.0301458835601807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9624626711010933, 'clip_ratio/low_mean': 0.09740735590457916, 'clip_ratio/low_min': 0.09740735590457916, 'clip_ratio/high_mean': 0.06637935806065798, 'clip_ratio/high_max': 0.06637935806065798, 'clip_ratio/region_mean': 0.16378671396523714, 'reward_total_mean': 0.22185459733009338, 'reward_meter_mean': 0.4518135190010071, 'reward_meter_std': 0.39509329199790955, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9591144323348999, 'reward_lexical_plausibility_std': 0.1156417652964592, 'reward_repeat_penalty_mean': 0.8956582546234131, 'reward_repeat_penalty_std': 0.1424568146467209, 'reward_repeat_floor_mean': 0.9878213405609131, 'reward_repeat_floor_std': 0.015821415930986404, 'reward_near_duplicate_penalty_mean': 0.9531832933425903, 'reward_near_duplicate_penalty_std': 0.05720429867506027, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9427089691162109, 'reward_distinct_2_std': 0.08827294409275055, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.23627693951129913, 'reward_total_composite_mean': 0.22185459733009338, 'reward_total_composite_std': 0.28892993927001953, 'epoch': 0.05}

 32%|███▏      | 1044/3300 [3:55:58<9:01:54, 14.41s/it]INFO 04-12 19:54:24 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1045/3300 [3:56:12<8:57:11, 14.29s/it]2026-04-12 19:54:38,652 | INFO | train_grpo_train | metrics_logged mode=train step=1045

                                                       
{'loss': -0.0847, 'grad_norm': 4.317941188812256, 'learning_rate': 6.8363636363636364e-06, 'num_tokens': 1863046.0, 'completions/mean_length': 131.0, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 76.5714340209961, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.8516666889190674, 'rewards/meter/std': 0.30668655037879944, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.18516401946544647, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.940000057220459, 'rewards/count_floor/std': 0.05554921180009842, 'rewards/lexical_plausibility/mean': 0.9649927616119385, 'rewards/lexical_plausibility/std': 0.09901534020900726, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.7306001782417297, 'rewards/repeat_penalty/std': 0.09317930042743683, 'rewards/repeat_floor/mean': 0.970472514629364, 'rewards/repeat_floor/std': 0.009617557749152184, 'rewards/near_duplicate_penalty/mean': 0.9298416376113892, 'rewards/near_duplicate_penalty/std': 0.03050163947045803, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.7992637753486633, 'rewards/distinct_2/std': 0.10670207440853119, 'rewards/total_composite/mean': 0.17786066234111786, 'rewards/total_composite/std': 0.10766448080539703, 'reward': 0.17786066234111786, 'reward_std': 0.10766447335481644, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13103488087654114, 'sampling/sampling_logp_difference/max': 1.6882412433624268, 'sampling/importance_sampling_ratio/min': 0.18484432995319366, 'sampling/importance_sampling_ratio/mean': 1.0278335809707642, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7445037737488747, 'clip_ratio/low_mean': 0.06442466378211975, 'clip_ratio/low_min': 0.06442466378211975, 'clip_ratio/high_mean': 0.05242701154202223, 'clip_ratio/high_max': 0.05242701154202223, 'clip_ratio/region_mean': 0.11685167532414198, 'reward_total_mean': 0.17786066234111786, 'reward_meter_mean': 0.8516666889190674, 'reward_meter_std': 0.30668655037879944, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.18516401946544647, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.940000057220459, 'reward_count_floor_std': 0.05554921180009842, 'reward_lexical_plausibility_mean': 0.9649927616119385, 'reward_lexical_plausibility_std': 0.09901534020900726, 'reward_repeat_penalty_mean': 0.7306001782417297, 'reward_repeat_penalty_std': 0.09317930042743683, 'reward_repeat_floor_mean': 0.970472514629364, 'reward_repeat_floor_std': 0.009617557749152184, 'reward_near_duplicate_penalty_mean': 0.9298416376113892, 'reward_near_duplicate_penalty_std': 0.03050163947045803, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.7992637753486633, 'reward_distinct_2_std': 0.10670207440853119, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.17786066234111786, 'reward_total_composite_std': 0.10766448080539703, 'epoch': 0.05}

 32%|███▏      | 1045/3300 [3:56:12<8:57:11, 14.29s/it]INFO 04-12 19:54:38 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1046/3300 [3:56:27<9:01:41, 14.42s/it]2026-04-12 19:54:53,393 | INFO | train_grpo_train | metrics_logged mode=train step=1046

                                                       
{'loss': -0.0602, 'grad_norm': 5.8024163246154785, 'learning_rate': 6.833333333333334e-06, 'num_tokens': 1864710.0, 'completions/mean_length': 103.0, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.57143020629883, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6534960269927979, 'rewards/meter/std': 0.43770304322242737, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9470566511154175, 'rewards/lexical_plausibility/std': 0.1497463583946228, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.15454426407814026, 'rewards/repeat_penalty/mean': 0.9568419456481934, 'rewards/repeat_penalty/std': 0.06302206218242645, 'rewards/repeat_floor/mean': 0.9943863153457642, 'rewards/repeat_floor/std': 0.008068595081567764, 'rewards/near_duplicate_penalty/mean': 0.9730864763259888, 'rewards/near_duplicate_penalty/std': 0.037587083876132965, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9824813008308411, 'rewards/distinct_2/std': 0.03248300403356552, 'rewards/total_composite/mean': 0.242730051279068, 'rewards/total_composite/std': 0.18923543393611908, 'reward': 0.242730051279068, 'reward_std': 0.18923543393611908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18914373219013214, 'sampling/sampling_logp_difference/max': 1.8349785804748535, 'sampling/importance_sampling_ratio/min': 0.15961691737174988, 'sampling/importance_sampling_ratio/mean': 1.0350492000579834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2791118025779724, 'clip_ratio/low_mean': 0.04326923191547394, 'clip_ratio/low_min': 0.04326923191547394, 'clip_ratio/high_mean': 0.11115088500082493, 'clip_ratio/high_max': 0.11115088500082493, 'clip_ratio/region_mean': 0.15442011691629887, 'reward_total_mean': 0.242730051279068, 'reward_meter_mean': 0.6534960269927979, 'reward_meter_std': 0.43770304322242737, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9470566511154175, 'reward_lexical_plausibility_std': 0.1497463583946228, 'reward_repeat_penalty_mean': 0.9568419456481934, 'reward_repeat_penalty_std': 0.06302206218242645, 'reward_repeat_floor_mean': 0.9943863153457642, 'reward_repeat_floor_std': 0.008068595081567764, 'reward_near_duplicate_penalty_mean': 0.9730864763259888, 'reward_near_duplicate_penalty_std': 0.037587083876132965, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9824813008308411, 'reward_distinct_2_std': 0.03248300403356552, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.15454426407814026, 'reward_total_composite_mean': 0.242730051279068, 'reward_total_composite_std': 0.18923543393611908, 'epoch': 0.06}

 32%|███▏      | 1046/3300 [3:56:27<9:01:41, 14.42s/it]INFO 04-12 19:54:53 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1047/3300 [3:56:40<8:51:30, 14.15s/it]2026-04-12 19:55:07,031 | INFO | train_grpo_train | metrics_logged mode=train step=1047

                                                       
{'loss': -0.0319, 'grad_norm': 6.724843978881836, 'learning_rate': 6.83030303030303e-06, 'num_tokens': 1866284.0, 'completions/mean_length': 107.75, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.000003814697266, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6977684497833252, 'rewards/meter/std': 0.3932817280292511, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9323267936706543, 'rewards/lexical_plausibility/std': 0.13708946108818054, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.3127185106277466, 'rewards/repeat_penalty/mean': 0.8352065086364746, 'rewards/repeat_penalty/std': 0.20242764055728912, 'rewards/repeat_floor/mean': 0.9793391227722168, 'rewards/repeat_floor/std': 0.025525081902742386, 'rewards/near_duplicate_penalty/mean': 0.9230776429176331, 'rewards/near_duplicate_penalty/std': 0.0763208419084549, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.9278052449226379, 'rewards/distinct_2/std': 0.1118457093834877, 'rewards/total_composite/mean': 0.3243905305862427, 'rewards/total_composite/std': 0.36138296127319336, 'reward': 0.3243905305862427, 'reward_std': 0.36138296127319336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13323017954826355, 'sampling/sampling_logp_difference/max': 1.5657258033752441, 'sampling/importance_sampling_ratio/min': 0.20893631875514984, 'sampling/importance_sampling_ratio/mean': 1.0500448942184448, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7880965247750282, 'clip_ratio/low_mean': 0.06800894532352686, 'clip_ratio/low_min': 0.06800894532352686, 'clip_ratio/high_mean': 0.038596960715949535, 'clip_ratio/high_max': 0.038596960715949535, 'clip_ratio/region_mean': 0.1066059060394764, 'reward_total_mean': 0.3243905305862427, 'reward_meter_mean': 0.6977684497833252, 'reward_meter_std': 0.3932817280292511, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9323267936706543, 'reward_lexical_plausibility_std': 0.13708946108818054, 'reward_repeat_penalty_mean': 0.8352065086364746, 'reward_repeat_penalty_std': 0.20242764055728912, 'reward_repeat_floor_mean': 0.9793391227722168, 'reward_repeat_floor_std': 0.025525081902742386, 'reward_near_duplicate_penalty_mean': 0.9230776429176331, 'reward_near_duplicate_penalty_std': 0.0763208419084549, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.9278052449226379, 'reward_distinct_2_std': 0.1118457093834877, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.3127185106277466, 'reward_total_composite_mean': 0.3243905305862427, 'reward_total_composite_std': 0.36138296127319336, 'epoch': 0.06}

 32%|███▏      | 1047/3300 [3:56:40<8:51:30, 14.15s/it]INFO 04-12 19:55:07 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1048/3300 [3:56:56<9:13:02, 14.73s/it]2026-04-12 19:55:23,009 | INFO | train_grpo_train | metrics_logged mode=train step=1048

                                                       
{'loss': -0.0619, 'grad_norm': 2.6656150817871094, 'learning_rate': 6.827272727272728e-06, 'num_tokens': 1867766.0, 'completions/mean_length': 164.25, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 48.333335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.945351243019104, 'rewards/meter/std': 0.07080448418855667, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.11338934302330017, 'rewards/lexical_plausibility/mean': 0.8734819889068604, 'rewards/lexical_plausibility/std': 0.23427492380142212, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.3345252573490143, 'rewards/repeat_penalty/mean': 0.9593273997306824, 'rewards/repeat_penalty/std': 0.02835819311439991, 'rewards/repeat_floor/mean': 0.9940921664237976, 'rewards/repeat_floor/std': 0.00415258202701807, 'rewards/near_duplicate_penalty/mean': 0.9627646207809448, 'rewards/near_duplicate_penalty/std': 0.0270233191549778, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9964163303375244, 'rewards/distinct_2/std': 0.007133082952350378, 'rewards/total_composite/mean': 0.4152056872844696, 'rewards/total_composite/std': 0.3385762870311737, 'reward': 0.4152056872844696, 'reward_std': 0.3385762870311737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14341652393341064, 'sampling/sampling_logp_difference/max': 1.234467625617981, 'sampling/importance_sampling_ratio/min': 0.2909896671772003, 'sampling/importance_sampling_ratio/mean': 1.0214452743530273, 'sampling/importance_sampling_ratio/max': 1.8521702289581299, 'entropy': 0.7694663479924202, 'clip_ratio/low_mean': 0.06064156163483858, 'clip_ratio/low_min': 0.06064156163483858, 'clip_ratio/high_mean': 0.06093749962747097, 'clip_ratio/high_max': 0.06093749962747097, 'clip_ratio/region_mean': 0.12157906126230955, 'reward_total_mean': 0.4152056872844696, 'reward_meter_mean': 0.945351243019104, 'reward_meter_std': 0.07080448418855667, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.11338934302330017, 'reward_lexical_plausibility_mean': 0.8734819889068604, 'reward_lexical_plausibility_std': 0.23427492380142212, 'reward_repeat_penalty_mean': 0.9593273997306824, 'reward_repeat_penalty_std': 0.02835819311439991, 'reward_repeat_floor_mean': 0.9940921664237976, 'reward_repeat_floor_std': 0.00415258202701807, 'reward_near_duplicate_penalty_mean': 0.9627646207809448, 'reward_near_duplicate_penalty_std': 0.0270233191549778, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9964163303375244, 'reward_distinct_2_std': 0.007133082952350378, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.3345252573490143, 'reward_total_composite_mean': 0.4152056872844696, 'reward_total_composite_std': 0.3385762870311737, 'epoch': 0.06}

 32%|███▏      | 1048/3300 [3:56:56<9:13:02, 14.73s/it]INFO 04-12 19:55:23 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1049/3300 [3:57:11<9:14:53, 14.79s/it]2026-04-12 19:55:37,950 | INFO | train_grpo_train | metrics_logged mode=train step=1049

                                                       
{'loss': -0.0538, 'grad_norm': 4.804747581481934, 'learning_rate': 6.824242424242425e-06, 'num_tokens': 1869306.0, 'completions/mean_length': 85.5, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.571430206298828, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.7500928640365601, 'rewards/meter/std': 0.3757437765598297, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7450137138366699, 'rewards/lexical_plausibility/std': 0.2284269630908966, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.7430853247642517, 'rewards/repeat_penalty/std': 0.019557634368538857, 'rewards/repeat_floor/mean': 0.9836170673370361, 'rewards/repeat_floor/std': 0.0039115226827561855, 'rewards/near_duplicate_penalty/mean': 0.9907804727554321, 'rewards/near_duplicate_penalty/std': 0.026076845824718475, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19522154331207275, 'rewards/total_composite/std': 0.1586945652961731, 'reward': 0.19522154331207275, 'reward_std': 0.1586945503950119, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17666389048099518, 'sampling/sampling_logp_difference/max': 1.0350725650787354, 'sampling/importance_sampling_ratio/min': 0.3552005887031555, 'sampling/importance_sampling_ratio/mean': 1.041739821434021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2145957052707672, 'clip_ratio/low_mean': 0.11156832380220294, 'clip_ratio/low_min': 0.11156832380220294, 'clip_ratio/high_mean': 0.06845238246023655, 'clip_ratio/high_max': 0.06845238246023655, 'clip_ratio/region_mean': 0.1800207062624395, 'reward_total_mean': 0.19522154331207275, 'reward_meter_mean': 0.7500928640365601, 'reward_meter_std': 0.3757437765598297, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.7450137138366699, 'reward_lexical_plausibility_std': 0.2284269630908966, 'reward_repeat_penalty_mean': 0.7430853247642517, 'reward_repeat_penalty_std': 0.019557634368538857, 'reward_repeat_floor_mean': 0.9836170673370361, 'reward_repeat_floor_std': 0.0039115226827561855, 'reward_near_duplicate_penalty_mean': 0.9907804727554321, 'reward_near_duplicate_penalty_std': 0.026076845824718475, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.19522154331207275, 'reward_total_composite_std': 0.1586945652961731, 'epoch': 0.06}

 32%|███▏      | 1049/3300 [3:57:11<9:14:53, 14.79s/it]INFO 04-12 19:55:38 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1050/3300 [3:57:22<8:29:54, 13.60s/it]2026-04-12 19:55:48,788 | INFO | train_grpo_train | metrics_logged mode=train step=1050

                                                       
{'loss': -0.011, 'grad_norm': 9.942171096801758, 'learning_rate': 6.821212121212122e-06, 'num_tokens': 1871497.0, 'completions/mean_length': 94.875, 'completions/min_length': 82.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.875, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9846819639205933, 'rewards/meter/std': 0.005867636762559414, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.8275918960571289, 'rewards/repeat_penalty/std': 0.0879611074924469, 'rewards/repeat_floor/mean': 0.9796619415283203, 'rewards/repeat_floor/std': 0.010421667248010635, 'rewards/near_duplicate_penalty/mean': 0.933120846748352, 'rewards/near_duplicate_penalty/std': 0.034280579537153244, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8854864835739136, 'rewards/distinct_2/std': 0.07136402279138565, 'rewards/total_composite/mean': 0.35399487614631653, 'rewards/total_composite/std': 0.06711232662200928, 'reward': 0.35399487614631653, 'reward_std': 0.06711231917142868, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1373150497674942, 'sampling/sampling_logp_difference/max': 1.6900954246520996, 'sampling/importance_sampling_ratio/min': 0.18450191617012024, 'sampling/importance_sampling_ratio/mean': 1.02568519115448, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.009215459227562, 'clip_ratio/low_mean': 0.05434692371636629, 'clip_ratio/low_min': 0.05434692371636629, 'clip_ratio/high_mean': 0.06637095753103495, 'clip_ratio/high_max': 0.06637095753103495, 'clip_ratio/region_mean': 0.12071788124740124, 'reward_total_mean': 0.35399487614631653, 'reward_meter_mean': 0.9846819639205933, 'reward_meter_std': 0.005867636762559414, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8275918960571289, 'reward_repeat_penalty_std': 0.0879611074924469, 'reward_repeat_floor_mean': 0.9796619415283203, 'reward_repeat_floor_std': 0.010421667248010635, 'reward_near_duplicate_penalty_mean': 0.933120846748352, 'reward_near_duplicate_penalty_std': 0.034280579537153244, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8854864835739136, 'reward_distinct_2_std': 0.07136402279138565, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.35399487614631653, 'reward_total_composite_std': 0.06711232662200928, 'epoch': 0.06}

 32%|███▏      | 1050/3300 [3:57:22<8:29:54, 13.60s/it]INFO 04-12 19:55:49 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:04,  5.90s/it]

 23%|██▎       | 3/13 [00:26<01:34,  9.47s/it]

 31%|███       | 4/13 [00:39<01:38, 10.91s/it]

 38%|███▊      | 5/13 [00:48<01:21, 10.20s/it]

 46%|████▌     | 6/13 [01:03<01:22, 11.74s/it]

 54%|█████▍    | 7/13 [01:15<01:10, 11.78s/it]

 62%|██████▏   | 8/13 [01:27<00:59, 11.97s/it]

 69%|██████▉   | 9/13 [01:40<00:48, 12.17s/it]

 77%|███████▋  | 10/13 [01:53<00:37, 12.65s/it]

 85%|████████▍ | 11/13 [02:07<00:26, 13.04s/it]

 92%|█████████▏| 12/13 [02:22<00:13, 13.48s/it]

100%|██████████| 13/13 [02:34<00:00, 13.09s/it]2026-04-12 19:58:34,754 | INFO | train_grpo_train | metrics_logged mode=eval step=1050

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 165.8218, 'eval_samples_per_second': 0.627, 'eval_steps_per_second': 0.078, 'eval_num_tokens': 1871497.0, 'eval_completions/mean_length': 181.07692307692307, 'eval_completions/min_length': 48.69230769230769, 'eval_completions/max_length': 463.7692307692308, 'eval_completions/clipped_ratio': 0.1346153846153846, 'eval_completions/mean_terminated_length': 130.24368462195764, 'eval_completions/min_terminated_length': 48.69230769230769, 'eval_completions/max_terminated_length': 260.0, 'eval_rewards/meter/mean': 0.6325419086676377, 'eval_rewards/meter/std': 0.34741294040129733, 'eval_rewards/count_adherence/mean': 0.8383747935295105, 'eval_rewards/count_adherence/std': 0.1801246553659439, 'eval_rewards/arabic_clean/mean': 0.8173076923076923, 'eval_rewards/arabic_clean/std': 0.3538298515173105, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9515124376003559, 'eval_rewards/count_floor/std': 0.05403739500504274, 'eval_rewards/lexical_plausibility/mean': 0.9488471654745249, 'eval_rewards/lexical_plausibility/std': 0.1151565955235408, 'eval_rewards/judge_quality/mean': 0.29365384349456203, 'eval_rewards/judge_quality/std': 0.15624268696858332, 'eval_rewards/repeat_penalty/mean': 0.6301313638687134, 'eval_rewards/repeat_penalty/std': 0.30668880618535554, 'eval_rewards/repeat_floor/mean': 0.9571049717756418, 'eval_rewards/repeat_floor/std': 0.03760866577235552, 'eval_rewards/near_duplicate_penalty/mean': 0.9016651786290683, 'eval_rewards/near_duplicate_penalty/std': 0.08141242311551021, 'eval_rewards/opening_diversity/mean': 0.9554844727882972, 'eval_rewards/opening_diversity/std': 0.09711383541042988, 'eval_rewards/distinct_2/mean': 0.7169570968701289, 'eval_rewards/distinct_2/std': 0.277169399536573, 'eval_rewards/total_composite/mean': 0.1787545084953308, 'eval_rewards/total_composite/std': 0.14113237192997566, 'eval_reward': 0.1787545084953308, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06447494402527809, 'eval_sampling/sampling_logp_difference/max': 1.1991608509650598, 'eval_sampling/importance_sampling_ratio/min': 0.31045468953939587, 'eval_sampling/importance_sampling_ratio/mean': 1.0170835990172167, 'eval_sampling/importance_sampling_ratio/max': 1.5534480259968684, 'eval_entropy': 0.7405934104552636, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1787545084953308, 'eval_reward_meter_mean': 0.6325419086676377, 'eval_reward_meter_std': 0.34741294040129733, 'eval_reward_count_adherence_mean': 0.8383747935295105, 'eval_reward_count_adherence_std': 0.1801246553659439, 'eval_reward_arabic_clean_mean': 0.8173076923076923, 'eval_reward_arabic_clean_std': 0.3538298515173105, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9515124376003559, 'eval_reward_count_floor_std': 0.05403739500504274, 'eval_reward_lexical_plausibility_mean': 0.9488471654745249, 'eval_reward_lexical_plausibility_std': 0.1151565955235408, 'eval_reward_repeat_penalty_mean': 0.6301313638687134, 'eval_reward_repeat_penalty_std': 0.30668880618535554, 'eval_reward_repeat_floor_mean': 0.9571049717756418, 'eval_reward_repeat_floor_std': 0.03760866577235552, 'eval_reward_near_duplicate_penalty_mean': 0.9016651786290683, 'eval_reward_near_duplicate_penalty_std': 0.08141242311551021, 'eval_reward_opening_diversity_mean': 0.9554844727882972, 'eval_reward_opening_diversity_std': 0.09711383541042988, 'eval_reward_distinct_2_mean': 0.7169570968701289, 'eval_reward_distinct_2_std': 0.277169399536573, 'eval_reward_judge_quality_mean': 0.29365384349456203, 'eval_reward_judge_quality_std': 0.15624268696858332, 'eval_reward_total_composite_mean': 0.1787545084953308, 'eval_reward_total_composite_std': 0.14113237192997566, 'epoch': 0.06}

 32%|███▏      | 1050/3300 [4:00:08<8:29:54, 13.60s/it]

100%|██████████| 13/13 [02:34<00:00, 13.09s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 32%|███▏      | 1051/3300 [4:00:30<41:12:24, 65.96s/it]2026-04-12 19:58:56,920 | INFO | train_grpo_train | metrics_logged mode=train step=1051

                                                        
{'loss': 0.0689, 'grad_norm': 6.433602809906006, 'learning_rate': 6.818181818181818e-06, 'num_tokens': 1874586.0, 'completions/mean_length': 178.125, 'completions/min_length': 152.0, 'completions/max_length': 208.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.125, 'completions/min_terminated_length': 152.0, 'completions/max_terminated_length': 208.0, 'rewards/meter/mean': 0.892245888710022, 'rewards/meter/std': 0.11291775852441788, 'rewards/count_adherence/mean': 0.8181818127632141, 'rewards/count_adherence/std': 0.06872081756591797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9454545974731445, 'rewards/count_floor/std': 0.02061624825000763, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.010859579779207706, 'rewards/repeat_penalty/std': 0.02485370635986328, 'rewards/repeat_floor/mean': 0.7860385775566101, 'rewards/repeat_floor/std': 0.03931530937552452, 'rewards/near_duplicate_penalty/mean': 0.44362872838974, 'rewards/near_duplicate_penalty/std': 0.18571671843528748, 'rewards/opening_diversity/mean': 0.25591516494750977, 'rewards/opening_diversity/std': 0.11222255229949951, 'rewards/distinct_2/mean': 0.04599259793758392, 'rewards/distinct_2/std': 0.0893903449177742, 'rewards/total_composite/mean': 0.09890294820070267, 'rewards/total_composite/std': 0.008755313232541084, 'reward': 0.09890294820070267, 'reward_std': 0.008755314163863659, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.058146748691797256, 'sampling/sampling_logp_difference/max': 2.436023712158203, 'sampling/importance_sampling_ratio/min': 0.08750811964273453, 'sampling/importance_sampling_ratio/mean': 1.0135841369628906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3538610301911831, 'clip_ratio/low_mean': 0.03504520142450929, 'clip_ratio/low_min': 0.03504520142450929, 'clip_ratio/high_mean': 0.02121712057851255, 'clip_ratio/high_max': 0.02121712057851255, 'clip_ratio/region_mean': 0.056262322003021836, 'reward_total_mean': 0.09890294820070267, 'reward_meter_mean': 0.892245888710022, 'reward_meter_std': 0.11291775852441788, 'reward_count_adherence_mean': 0.8181818127632141, 'reward_count_adherence_std': 0.06872081756591797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9454545974731445, 'reward_count_floor_std': 0.02061624825000763, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.010859579779207706, 'reward_repeat_penalty_std': 0.02485370635986328, 'reward_repeat_floor_mean': 0.7860385775566101, 'reward_repeat_floor_std': 0.03931530937552452, 'reward_near_duplicate_penalty_mean': 0.44362872838974, 'reward_near_duplicate_penalty_std': 0.18571671843528748, 'reward_opening_diversity_mean': 0.25591516494750977, 'reward_opening_diversity_std': 0.11222255229949951, 'reward_distinct_2_mean': 0.04599259793758392, 'reward_distinct_2_std': 0.0893903449177742, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.09890294820070267, 'reward_total_composite_std': 0.008755313232541084, 'epoch': 0.06}

 32%|███▏      | 1051/3300 [4:00:30<41:12:24, 65.96s/it]INFO 04-12 19:58:57 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1052/3300 [4:00:49<32:20:01, 51.78s/it]2026-04-12 19:59:15,688 | INFO | train_grpo_train | metrics_logged mode=train step=1052

                                                        
{'loss': -0.0504, 'grad_norm': 5.15717887878418, 'learning_rate': 6.8151515151515155e-06, 'num_tokens': 1876170.0, 'completions/mean_length': 108.0, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.28571701049805, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9441964626312256, 'rewards/meter/std': 0.13102741539478302, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.942996621131897, 'rewards/lexical_plausibility/std': 0.16123001277446747, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.9034607410430908, 'rewards/repeat_penalty/std': 0.07770588248968124, 'rewards/repeat_floor/mean': 0.989519476890564, 'rewards/repeat_floor/std': 0.006607821211218834, 'rewards/near_duplicate_penalty/mean': 0.956121027469635, 'rewards/near_duplicate_penalty/std': 0.04314931482076645, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9775147438049316, 'rewards/distinct_2/std': 0.031041136011481285, 'rewards/total_composite/mean': 0.2695521116256714, 'rewards/total_composite/std': 0.15337784588336945, 'reward': 0.2695521116256714, 'reward_std': 0.15337783098220825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1628119796514511, 'sampling/sampling_logp_difference/max': 1.3122692108154297, 'sampling/importance_sampling_ratio/min': 0.26920846104621887, 'sampling/importance_sampling_ratio/mean': 1.034109115600586, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1611990332603455, 'clip_ratio/low_mean': 0.03667763154953718, 'clip_ratio/low_min': 0.03667763154953718, 'clip_ratio/high_mean': 0.09007105976343155, 'clip_ratio/high_max': 0.09007105976343155, 'clip_ratio/region_mean': 0.12674869131296873, 'reward_total_mean': 0.2695521116256714, 'reward_meter_mean': 0.9441964626312256, 'reward_meter_std': 0.13102741539478302, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.942996621131897, 'reward_lexical_plausibility_std': 0.16123001277446747, 'reward_repeat_penalty_mean': 0.9034607410430908, 'reward_repeat_penalty_std': 0.07770588248968124, 'reward_repeat_floor_mean': 0.989519476890564, 'reward_repeat_floor_std': 0.006607821211218834, 'reward_near_duplicate_penalty_mean': 0.956121027469635, 'reward_near_duplicate_penalty_std': 0.04314931482076645, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9775147438049316, 'reward_distinct_2_std': 0.031041136011481285, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.2695521116256714, 'reward_total_composite_std': 0.15337784588336945, 'epoch': 0.06}

 32%|███▏      | 1052/3300 [4:00:49<32:20:01, 51.78s/it]INFO 04-12 19:59:15 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1053/3300 [4:01:07<26:03:47, 41.76s/it]2026-04-12 19:59:34,032 | INFO | train_grpo_train | metrics_logged mode=train step=1053

                                                        
{'loss': -0.0273, 'grad_norm': 5.930370807647705, 'learning_rate': 6.812121212121212e-06, 'num_tokens': 1877840.0, 'completions/mean_length': 111.75, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.57143020629883, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.7181860208511353, 'rewards/meter/std': 0.3997362554073334, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9589493870735168, 'rewards/lexical_plausibility/std': 0.11610868573188782, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8715294599533081, 'rewards/repeat_penalty/std': 0.17832595109939575, 'rewards/repeat_floor/mean': 0.986426591873169, 'rewards/repeat_floor/std': 0.019042083993554115, 'rewards/near_duplicate_penalty/mean': 0.9549828767776489, 'rewards/near_duplicate_penalty/std': 0.05734385922551155, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9450459480285645, 'rewards/distinct_2/std': 0.07699161767959595, 'rewards/total_composite/mean': 0.2721409201622009, 'rewards/total_composite/std': 0.17866960167884827, 'reward': 0.2721409201622009, 'reward_std': 0.17866961658000946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15803660452365875, 'sampling/sampling_logp_difference/max': 1.1634140014648438, 'sampling/importance_sampling_ratio/min': 0.31848540902137756, 'sampling/importance_sampling_ratio/mean': 1.0133339166641235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9356729984283447, 'clip_ratio/low_mean': 0.03724747523665428, 'clip_ratio/low_min': 0.03724747523665428, 'clip_ratio/high_mean': 0.10755943227559328, 'clip_ratio/high_max': 0.10755943227559328, 'clip_ratio/region_mean': 0.14480690751224756, 'reward_total_mean': 0.2721409201622009, 'reward_meter_mean': 0.7181860208511353, 'reward_meter_std': 0.3997362554073334, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9589493870735168, 'reward_lexical_plausibility_std': 0.11610868573188782, 'reward_repeat_penalty_mean': 0.8715294599533081, 'reward_repeat_penalty_std': 0.17832595109939575, 'reward_repeat_floor_mean': 0.986426591873169, 'reward_repeat_floor_std': 0.019042083993554115, 'reward_near_duplicate_penalty_mean': 0.9549828767776489, 'reward_near_duplicate_penalty_std': 0.05734385922551155, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9450459480285645, 'reward_distinct_2_std': 0.07699161767959595, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.2721409201622009, 'reward_total_composite_std': 0.17866960167884827, 'epoch': 0.06}

 32%|███▏      | 1053/3300 [4:01:07<26:03:47, 41.76s/it]INFO 04-12 19:59:34 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1054/3300 [4:01:18<20:17:24, 32.52s/it]2026-04-12 19:59:44,906 | INFO | train_grpo_train | metrics_logged mode=train step=1054

                                                        
{'loss': 0.3981, 'grad_norm': 18.729524612426758, 'learning_rate': 6.80909090909091e-06, 'num_tokens': 1879511.0, 'completions/mean_length': 57.875, 'completions/min_length': 36.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.740469753742218, 'rewards/meter/std': 0.3820609748363495, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.11338934302330017, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.82399582862854, 'rewards/repeat_penalty/std': 0.27185317873954773, 'rewards/repeat_floor/mean': 0.9810456037521362, 'rewards/repeat_floor/std': 0.02700800821185112, 'rewards/near_duplicate_penalty/mean': 0.9464839696884155, 'rewards/near_duplicate_penalty/std': 0.048685867339372635, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8785883188247681, 'rewards/distinct_2/std': 0.22482581436634064, 'rewards/total_composite/mean': 0.27263057231903076, 'rewards/total_composite/std': 0.1562337726354599, 'reward': 0.27263057231903076, 'reward_std': 0.1562337726354599, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1589154154062271, 'sampling/sampling_logp_difference/max': 3.067474842071533, 'sampling/importance_sampling_ratio/min': 0.04653852432966232, 'sampling/importance_sampling_ratio/mean': 1.0194118022918701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9854426980018616, 'clip_ratio/low_mean': 0.03309714142233133, 'clip_ratio/low_min': 0.03309714142233133, 'clip_ratio/high_mean': 0.10806720703840256, 'clip_ratio/high_max': 0.10806720703840256, 'clip_ratio/region_mean': 0.1411643484607339, 'reward_total_mean': 0.27263057231903076, 'reward_meter_mean': 0.740469753742218, 'reward_meter_std': 0.3820609748363495, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.11338934302330017, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.82399582862854, 'reward_repeat_penalty_std': 0.27185317873954773, 'reward_repeat_floor_mean': 0.9810456037521362, 'reward_repeat_floor_std': 0.02700800821185112, 'reward_near_duplicate_penalty_mean': 0.9464839696884155, 'reward_near_duplicate_penalty_std': 0.048685867339372635, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8785883188247681, 'reward_distinct_2_std': 0.22482581436634064, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.27263057231903076, 'reward_total_composite_std': 0.1562337726354599, 'epoch': 0.06}

 32%|███▏      | 1054/3300 [4:01:18<20:17:24, 32.52s/it]INFO 04-12 19:59:45 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1055/3300 [4:01:36<17:29:41, 28.05s/it]2026-04-12 20:00:02,548 | INFO | train_grpo_train | metrics_logged mode=train step=1055

                                                        
{'loss': -0.1124, 'grad_norm': 1.7240773439407349, 'learning_rate': 6.806060606060607e-06, 'num_tokens': 1881232.0, 'completions/mean_length': 231.125, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 62.60000228881836, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7945102453231812, 'rewards/meter/std': 0.3211391866207123, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8764052391052246, 'rewards/lexical_plausibility/std': 0.1765848696231842, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.6604894995689392, 'rewards/repeat_penalty/std': 0.29806774854660034, 'rewards/repeat_floor/mean': 0.9599384069442749, 'rewards/repeat_floor/std': 0.03850284218788147, 'rewards/near_duplicate_penalty/mean': 0.8947899341583252, 'rewards/near_duplicate_penalty/std': 0.10041578114032745, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.7718883156776428, 'rewards/distinct_2/std': 0.20208999514579773, 'rewards/total_composite/mean': 0.23197248578071594, 'rewards/total_composite/std': 0.15391208231449127, 'reward': 0.23197248578071594, 'reward_std': 0.15391208231449127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14595472812652588, 'sampling/sampling_logp_difference/max': 3.4403138160705566, 'sampling/importance_sampling_ratio/min': 0.03205462545156479, 'sampling/importance_sampling_ratio/mean': 1.025005578994751, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6018383800983429, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0645558824762702, 'clip_ratio/high_max': 0.0645558824762702, 'clip_ratio/region_mean': 0.0645558824762702, 'reward_total_mean': 0.23197248578071594, 'reward_meter_mean': 0.7945102453231812, 'reward_meter_std': 0.3211391866207123, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8764052391052246, 'reward_lexical_plausibility_std': 0.1765848696231842, 'reward_repeat_penalty_mean': 0.6604894995689392, 'reward_repeat_penalty_std': 0.29806774854660034, 'reward_repeat_floor_mean': 0.9599384069442749, 'reward_repeat_floor_std': 0.03850284218788147, 'reward_near_duplicate_penalty_mean': 0.8947899341583252, 'reward_near_duplicate_penalty_std': 0.10041578114032745, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.7718883156776428, 'reward_distinct_2_std': 0.20208999514579773, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.23197248578071594, 'reward_total_composite_std': 0.15391208231449127, 'epoch': 0.06}

 32%|███▏      | 1055/3300 [4:01:36<17:29:41, 28.05s/it]INFO 04-12 20:00:02 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1056/3300 [4:01:46<14:12:55, 22.81s/it]2026-04-12 20:00:13,191 | INFO | train_grpo_train | metrics_logged mode=train step=1056

                                                        
{'loss': 0.1138, 'grad_norm': 13.494691848754883, 'learning_rate': 6.803030303030304e-06, 'num_tokens': 1882856.0, 'completions/mean_length': 44.0, 'completions/min_length': 36.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.6584372520446777, 'rewards/meter/std': 0.32397934794425964, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.97265625, 'rewards/lexical_plausibility/std': 0.07733980566263199, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.8241800665855408, 'rewards/repeat_penalty/std': 0.14683881402015686, 'rewards/repeat_floor/mean': 0.9783002734184265, 'rewards/repeat_floor/std': 0.018380578607320786, 'rewards/near_duplicate_penalty/mean': 0.908063530921936, 'rewards/near_duplicate_penalty/std': 0.09522882848978043, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9329525232315063, 'rewards/distinct_2/std': 0.05054157227277756, 'rewards/total_composite/mean': 0.3619273900985718, 'rewards/total_composite/std': 0.3119892477989197, 'reward': 0.3619273900985718, 'reward_std': 0.3119892477989197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1304035186767578, 'sampling/sampling_logp_difference/max': 1.7612848281860352, 'sampling/importance_sampling_ratio/min': 0.17182396352291107, 'sampling/importance_sampling_ratio/mean': 1.0198287963867188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.059627190232277, 'clip_ratio/low_mean': 0.11065081227570772, 'clip_ratio/low_min': 0.11065081227570772, 'clip_ratio/high_mean': 0.03814102616161108, 'clip_ratio/high_max': 0.03814102616161108, 'clip_ratio/region_mean': 0.1487918384373188, 'reward_total_mean': 0.3619273900985718, 'reward_meter_mean': 0.6584372520446777, 'reward_meter_std': 0.32397934794425964, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.97265625, 'reward_lexical_plausibility_std': 0.07733980566263199, 'reward_repeat_penalty_mean': 0.8241800665855408, 'reward_repeat_penalty_std': 0.14683881402015686, 'reward_repeat_floor_mean': 0.9783002734184265, 'reward_repeat_floor_std': 0.018380578607320786, 'reward_near_duplicate_penalty_mean': 0.908063530921936, 'reward_near_duplicate_penalty_std': 0.09522882848978043, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9329525232315063, 'reward_distinct_2_std': 0.05054157227277756, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.2522329092025757, 'reward_total_composite_mean': 0.3619273900985718, 'reward_total_composite_std': 0.3119892477989197, 'epoch': 0.06}

 32%|███▏      | 1056/3300 [4:01:47<14:12:55, 22.81s/it]INFO 04-12 20:00:13 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1057/3300 [4:02:04<13:16:37, 21.31s/it]2026-04-12 20:00:30,909 | INFO | train_grpo_train | metrics_logged mode=train step=1057

                                                        
{'loss': -0.0935, 'grad_norm': 2.5543081760406494, 'learning_rate': 6.800000000000001e-06, 'num_tokens': 1884421.0, 'completions/mean_length': 163.625, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6037691831588745, 'rewards/meter/std': 0.46449241042137146, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.8839333057403564, 'rewards/lexical_plausibility/std': 0.21582959592342377, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9102270603179932, 'rewards/repeat_penalty/std': 0.061639171093702316, 'rewards/repeat_floor/mean': 0.9885294437408447, 'rewards/repeat_floor/std': 0.008984849788248539, 'rewards/near_duplicate_penalty/mean': 0.9452494978904724, 'rewards/near_duplicate_penalty/std': 0.0642799362540245, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9637999534606934, 'rewards/distinct_2/std': 0.039412181824445724, 'rewards/total_composite/mean': 0.22545307874679565, 'rewards/total_composite/std': 0.174667090177536, 'reward': 0.22545307874679565, 'reward_std': 0.174667090177536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11902369558811188, 'sampling/sampling_logp_difference/max': 1.1344772577285767, 'sampling/importance_sampling_ratio/min': 0.3215901851654053, 'sampling/importance_sampling_ratio/mean': 1.0294219255447388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5526062697172165, 'clip_ratio/low_mean': 0.011627906933426857, 'clip_ratio/low_min': 0.011627906933426857, 'clip_ratio/high_mean': 0.08634344255551696, 'clip_ratio/high_max': 0.08634344255551696, 'clip_ratio/region_mean': 0.09797134948894382, 'reward_total_mean': 0.22545307874679565, 'reward_meter_mean': 0.6037691831588745, 'reward_meter_std': 0.46449241042137146, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.8839333057403564, 'reward_lexical_plausibility_std': 0.21582959592342377, 'reward_repeat_penalty_mean': 0.9102270603179932, 'reward_repeat_penalty_std': 0.061639171093702316, 'reward_repeat_floor_mean': 0.9885294437408447, 'reward_repeat_floor_std': 0.008984849788248539, 'reward_near_duplicate_penalty_mean': 0.9452494978904724, 'reward_near_duplicate_penalty_std': 0.0642799362540245, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9637999534606934, 'reward_distinct_2_std': 0.039412181824445724, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.22545307874679565, 'reward_total_composite_std': 0.174667090177536, 'epoch': 0.06}

 32%|███▏      | 1057/3300 [4:02:04<13:16:37, 21.31s/it]INFO 04-12 20:00:31 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1058/3300 [4:02:18<11:55:48, 19.16s/it]2026-04-12 20:00:45,043 | INFO | train_grpo_train | metrics_logged mode=train step=1058

                                                        
{'loss': -0.0657, 'grad_norm': 2.4039409160614014, 'learning_rate': 6.796969696969697e-06, 'num_tokens': 1886124.0, 'completions/mean_length': 162.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.629806399345398, 'rewards/meter/std': 0.41197389364242554, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.8900993466377258, 'rewards/lexical_plausibility/std': 0.20356714725494385, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.8640317916870117, 'rewards/repeat_penalty/std': 0.12923002243041992, 'rewards/repeat_floor/mean': 0.9841066002845764, 'rewards/repeat_floor/std': 0.014904729090631008, 'rewards/near_duplicate_penalty/mean': 0.9508353471755981, 'rewards/near_duplicate_penalty/std': 0.04037024453282356, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9053477048873901, 'rewards/distinct_2/std': 0.10357029736042023, 'rewards/total_composite/mean': 0.21735641360282898, 'rewards/total_composite/std': 0.17386916279792786, 'reward': 0.21735641360282898, 'reward_std': 0.17386916279792786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1386731117963791, 'sampling/sampling_logp_difference/max': 1.817184567451477, 'sampling/importance_sampling_ratio/min': 0.16248255968093872, 'sampling/importance_sampling_ratio/mean': 1.005149006843567, 'sampling/importance_sampling_ratio/max': 1.8159948587417603, 'entropy': 0.7190160900354385, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.08336841035634279, 'clip_ratio/high_max': 0.08336841035634279, 'clip_ratio/region_mean': 0.10122555401176214, 'reward_total_mean': 0.21735641360282898, 'reward_meter_mean': 0.629806399345398, 'reward_meter_std': 0.41197389364242554, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.8900993466377258, 'reward_lexical_plausibility_std': 0.20356714725494385, 'reward_repeat_penalty_mean': 0.8640317916870117, 'reward_repeat_penalty_std': 0.12923002243041992, 'reward_repeat_floor_mean': 0.9841066002845764, 'reward_repeat_floor_std': 0.014904729090631008, 'reward_near_duplicate_penalty_mean': 0.9508353471755981, 'reward_near_duplicate_penalty_std': 0.04037024453282356, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9053477048873901, 'reward_distinct_2_std': 0.10357029736042023, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.21735641360282898, 'reward_total_composite_std': 0.17386916279792786, 'epoch': 0.06}

 32%|███▏      | 1058/3300 [4:02:18<11:55:48, 19.16s/it]INFO 04-12 20:00:45 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1059/3300 [4:02:32<10:57:54, 17.61s/it]2026-04-12 20:00:59,060 | INFO | train_grpo_train | metrics_logged mode=train step=1059

                                                        
{'loss': -0.06, 'grad_norm': 1.3979827165603638, 'learning_rate': 6.793939393939395e-06, 'num_tokens': 1887704.0, 'completions/mean_length': 400.5, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.4304245710372925, 'rewards/meter/std': 0.42752981185913086, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.35634833574295044, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.10690450668334961, 'rewards/lexical_plausibility/mean': 0.7138697504997253, 'rewards/lexical_plausibility/std': 0.18141908943653107, 'rewards/judge_quality/mean': 0.08124999701976776, 'rewards/judge_quality/std': 0.05938675254583359, 'rewards/repeat_penalty/mean': 0.8546221256256104, 'rewards/repeat_penalty/std': 0.18983952701091766, 'rewards/repeat_floor/mean': 0.9849105477333069, 'rewards/repeat_floor/std': 0.01979179121553898, 'rewards/near_duplicate_penalty/mean': 0.9394791126251221, 'rewards/near_duplicate_penalty/std': 0.06317517161369324, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.9451121091842651, 'rewards/distinct_2/std': 0.09245479851961136, 'rewards/total_composite/mean': 0.04833013191819191, 'rewards/total_composite/std': 0.05893223360180855, 'reward': 0.04833013191819191, 'reward_std': 0.05893222987651825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1608784794807434, 'sampling/sampling_logp_difference/max': 1.230952262878418, 'sampling/importance_sampling_ratio/min': 0.29201436042785645, 'sampling/importance_sampling_ratio/mean': 1.0251495838165283, 'sampling/importance_sampling_ratio/max': 1.996578335762024, 'entropy': 0.2717258781194687, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03030998818576336, 'clip_ratio/high_max': 0.03030998818576336, 'clip_ratio/region_mean': 0.03030998818576336, 'reward_total_mean': 0.04833013191819191, 'reward_meter_mean': 0.4304245710372925, 'reward_meter_std': 0.42752981185913086, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.35634833574295044, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.10690450668334961, 'reward_lexical_plausibility_mean': 0.7138697504997253, 'reward_lexical_plausibility_std': 0.18141908943653107, 'reward_repeat_penalty_mean': 0.8546221256256104, 'reward_repeat_penalty_std': 0.18983952701091766, 'reward_repeat_floor_mean': 0.9849105477333069, 'reward_repeat_floor_std': 0.01979179121553898, 'reward_near_duplicate_penalty_mean': 0.9394791126251221, 'reward_near_duplicate_penalty_std': 0.06317517161369324, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.9451121091842651, 'reward_distinct_2_std': 0.09245479851961136, 'reward_judge_quality_mean': 0.08124999701976776, 'reward_judge_quality_std': 0.05938675254583359, 'reward_total_composite_mean': 0.04833013191819191, 'reward_total_composite_std': 0.05893223360180855, 'epoch': 0.06}

 32%|███▏      | 1059/3300 [4:02:32<10:57:54, 17.61s/it]INFO 04-12 20:00:59 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1060/3300 [4:02:51<11:07:35, 17.88s/it]2026-04-12 20:01:17,628 | INFO | train_grpo_train | metrics_logged mode=train step=1060

                                                        
{'loss': -0.0819, 'grad_norm': 1.4049897193908691, 'learning_rate': 6.790909090909091e-06, 'num_tokens': 1889125.0, 'completions/mean_length': 344.625, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 65.66667175292969, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8862758874893188, 'rewards/meter/std': 0.12985679507255554, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.1060660183429718, 'rewards/lexical_plausibility/mean': 0.7624388337135315, 'rewards/lexical_plausibility/std': 0.20639127492904663, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9160146713256836, 'rewards/repeat_penalty/std': 0.08384377509355545, 'rewards/repeat_floor/mean': 0.9922943711280823, 'rewards/repeat_floor/std': 0.005956006236374378, 'rewards/near_duplicate_penalty/mean': 0.9839053750038147, 'rewards/near_duplicate_penalty/std': 0.013218950480222702, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9620392322540283, 'rewards/distinct_2/std': 0.038459762930870056, 'rewards/total_composite/mean': 0.15728075802326202, 'rewards/total_composite/std': 0.14909547567367554, 'reward': 0.15728075802326202, 'reward_std': 0.14909547567367554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14358174800872803, 'sampling/sampling_logp_difference/max': 1.3824973106384277, 'sampling/importance_sampling_ratio/min': 0.2509510815143585, 'sampling/importance_sampling_ratio/mean': 1.0164670944213867, 'sampling/importance_sampling_ratio/max': 1.7339977025985718, 'entropy': 0.5054924339056015, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0542998481541872, 'clip_ratio/high_max': 0.0542998481541872, 'clip_ratio/region_mean': 0.0542998481541872, 'reward_total_mean': 0.15728075802326202, 'reward_meter_mean': 0.8862758874893188, 'reward_meter_std': 0.12985679507255554, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.1060660183429718, 'reward_lexical_plausibility_mean': 0.7624388337135315, 'reward_lexical_plausibility_std': 0.20639127492904663, 'reward_repeat_penalty_mean': 0.9160146713256836, 'reward_repeat_penalty_std': 0.08384377509355545, 'reward_repeat_floor_mean': 0.9922943711280823, 'reward_repeat_floor_std': 0.005956006236374378, 'reward_near_duplicate_penalty_mean': 0.9839053750038147, 'reward_near_duplicate_penalty_std': 0.013218950480222702, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9620392322540283, 'reward_distinct_2_std': 0.038459762930870056, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.15728075802326202, 'reward_total_composite_std': 0.14909547567367554, 'epoch': 0.06}

 32%|███▏      | 1060/3300 [4:02:51<11:07:35, 17.88s/it]INFO 04-12 20:01:17 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1061/3300 [4:03:06<10:33:00, 16.96s/it]2026-04-12 20:01:32,425 | INFO | train_grpo_train | metrics_logged mode=train step=1061

                                                        
{'loss': -0.1515, 'grad_norm': 4.107607364654541, 'learning_rate': 6.787878787878789e-06, 'num_tokens': 1891867.0, 'completions/mean_length': 199.75, 'completions/min_length': 136.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 155.1428680419922, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.9535471200942993, 'rewards/meter/std': 0.06693251430988312, 'rewards/count_adherence/mean': 0.7321428656578064, 'rewards/count_adherence/std': 0.141575887799263, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9196428060531616, 'rewards/count_floor/std': 0.042472753673791885, 'rewards/lexical_plausibility/mean': 0.9342121481895447, 'rewards/lexical_plausibility/std': 0.12252917885780334, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.12799972295761108, 'rewards/repeat_penalty/mean': 0.5980527400970459, 'rewards/repeat_penalty/std': 0.1840987354516983, 'rewards/repeat_floor/mean': 0.9544660449028015, 'rewards/repeat_floor/std': 0.024737274274230003, 'rewards/near_duplicate_penalty/mean': 0.9016295671463013, 'rewards/near_duplicate_penalty/std': 0.04615097492933273, 'rewards/opening_diversity/mean': 0.8622158765792847, 'rewards/opening_diversity/std': 0.15586212277412415, 'rewards/distinct_2/mean': 0.7498734593391418, 'rewards/distinct_2/std': 0.12536492943763733, 'rewards/total_composite/mean': 0.18040475249290466, 'rewards/total_composite/std': 0.10451984405517578, 'reward': 0.18040475249290466, 'reward_std': 0.10451985150575638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14744122326374054, 'sampling/sampling_logp_difference/max': 2.5274710655212402, 'sampling/importance_sampling_ratio/min': 0.07986072450876236, 'sampling/importance_sampling_ratio/mean': 1.0065003633499146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2474434822797775, 'clip_ratio/low_mean': 0.03400100348517299, 'clip_ratio/low_min': 0.03400100348517299, 'clip_ratio/high_mean': 0.06318632420152426, 'clip_ratio/high_max': 0.06318632420152426, 'clip_ratio/region_mean': 0.09718732768669724, 'reward_total_mean': 0.18040475249290466, 'reward_meter_mean': 0.9535471200942993, 'reward_meter_std': 0.06693251430988312, 'reward_count_adherence_mean': 0.7321428656578064, 'reward_count_adherence_std': 0.141575887799263, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9196428060531616, 'reward_count_floor_std': 0.042472753673791885, 'reward_lexical_plausibility_mean': 0.9342121481895447, 'reward_lexical_plausibility_std': 0.12252917885780334, 'reward_repeat_penalty_mean': 0.5980527400970459, 'reward_repeat_penalty_std': 0.1840987354516983, 'reward_repeat_floor_mean': 0.9544660449028015, 'reward_repeat_floor_std': 0.024737274274230003, 'reward_near_duplicate_penalty_mean': 0.9016295671463013, 'reward_near_duplicate_penalty_std': 0.04615097492933273, 'reward_opening_diversity_mean': 0.8622158765792847, 'reward_opening_diversity_std': 0.15586212277412415, 'reward_distinct_2_mean': 0.7498734593391418, 'reward_distinct_2_std': 0.12536492943763733, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.12799972295761108, 'reward_total_composite_mean': 0.18040475249290466, 'reward_total_composite_std': 0.10451984405517578, 'epoch': 0.06}

 32%|███▏      | 1061/3300 [4:03:06<10:33:00, 16.96s/it]INFO 04-12 20:01:32 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1062/3300 [4:03:14<9:00:43, 14.50s/it] 2026-04-12 20:01:41,245 | INFO | train_grpo_train | metrics_logged mode=train step=1062

                                                       
{'loss': 0.2148, 'grad_norm': 14.848766326904297, 'learning_rate': 6.7848484848484855e-06, 'num_tokens': 1893426.0, 'completions/mean_length': 44.875, 'completions/min_length': 37.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8891546726226807, 'rewards/meter/std': 0.2533186078071594, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7374999523162842, 'rewards/judge_quality/std': 0.29489707946777344, 'rewards/repeat_penalty/mean': 0.9520331621170044, 'rewards/repeat_penalty/std': 0.05258069187402725, 'rewards/repeat_floor/mean': 0.9941830635070801, 'rewards/repeat_floor/std': 0.006022192072123289, 'rewards/near_duplicate_penalty/mean': 0.9767163991928101, 'rewards/near_duplicate_penalty/std': 0.019991640001535416, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9741734266281128, 'rewards/distinct_2/std': 0.03608804941177368, 'rewards/total_composite/mean': 0.6819944977760315, 'rewards/total_composite/std': 0.356192022562027, 'reward': 0.6819944977760315, 'reward_std': 0.356192022562027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10999785363674164, 'sampling/sampling_logp_difference/max': 1.8745062351226807, 'sampling/importance_sampling_ratio/min': 0.2727104127407074, 'sampling/importance_sampling_ratio/mean': 1.0146538019180298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5823167040944099, 'clip_ratio/low_mean': 0.04365589935332537, 'clip_ratio/low_min': 0.04365589935332537, 'clip_ratio/high_mean': 0.027027026982977986, 'clip_ratio/high_max': 0.027027026982977986, 'clip_ratio/region_mean': 0.07068292633630335, 'reward_total_mean': 0.6819944977760315, 'reward_meter_mean': 0.8891546726226807, 'reward_meter_std': 0.2533186078071594, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9520331621170044, 'reward_repeat_penalty_std': 0.05258069187402725, 'reward_repeat_floor_mean': 0.9941830635070801, 'reward_repeat_floor_std': 0.006022192072123289, 'reward_near_duplicate_penalty_mean': 0.9767163991928101, 'reward_near_duplicate_penalty_std': 0.019991640001535416, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9741734266281128, 'reward_distinct_2_std': 0.03608804941177368, 'reward_judge_quality_mean': 0.7374999523162842, 'reward_judge_quality_std': 0.29489707946777344, 'reward_total_composite_mean': 0.6819944977760315, 'reward_total_composite_std': 0.356192022562027, 'epoch': 0.06}

 32%|███▏      | 1062/3300 [4:03:15<9:00:43, 14.50s/it]INFO 04-12 20:01:41 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1063/3300 [4:03:30<9:10:41, 14.77s/it]2026-04-12 20:01:56,590 | INFO | train_grpo_train | metrics_logged mode=train step=1063

                                                       
{'loss': -0.0645, 'grad_norm': 1.2782859802246094, 'learning_rate': 6.781818181818183e-06, 'num_tokens': 1895374.0, 'completions/mean_length': 468.5, 'completions/min_length': 164.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 164.0, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.7416023015975952, 'rewards/meter/std': 0.23459242284297943, 'rewards/count_adherence/mean': 0.6111111640930176, 'rewards/count_adherence/std': 0.2300218641757965, 'rewards/arabic_clean/mean': 0.125, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8833333253860474, 'rewards/count_floor/std': 0.06900654733181, 'rewards/lexical_plausibility/mean': 0.7000858187675476, 'rewards/lexical_plausibility/std': 0.16103827953338623, 'rewards/judge_quality/mean': 0.10000000149011612, 'rewards/judge_quality/std': 0.053452253341674805, 'rewards/repeat_penalty/mean': 0.8089038133621216, 'rewards/repeat_penalty/std': 0.2780887484550476, 'rewards/repeat_floor/mean': 0.9789993166923523, 'rewards/repeat_floor/std': 0.029382452368736267, 'rewards/near_duplicate_penalty/mean': 0.9428079128265381, 'rewards/near_duplicate_penalty/std': 0.05258138105273247, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8828124403953552, 'rewards/distinct_2/std': 0.26019859313964844, 'rewards/total_composite/mean': 0.06213457137346268, 'rewards/total_composite/std': 0.04056248441338539, 'reward': 0.06213457137346268, 'reward_std': 0.04056248441338539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09518899023532867, 'sampling/sampling_logp_difference/max': 2.308614492416382, 'sampling/importance_sampling_ratio/min': 0.09939887374639511, 'sampling/importance_sampling_ratio/mean': 1.024430513381958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.0841052234172821, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.009146341122686863, 'clip_ratio/high_max': 0.009146341122686863, 'clip_ratio/region_mean': 0.009146341122686863, 'reward_total_mean': 0.06213457137346268, 'reward_meter_mean': 0.7416023015975952, 'reward_meter_std': 0.23459242284297943, 'reward_count_adherence_mean': 0.6111111640930176, 'reward_count_adherence_std': 0.2300218641757965, 'reward_arabic_clean_mean': 0.125, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8833333253860474, 'reward_count_floor_std': 0.06900654733181, 'reward_lexical_plausibility_mean': 0.7000858187675476, 'reward_lexical_plausibility_std': 0.16103827953338623, 'reward_repeat_penalty_mean': 0.8089038133621216, 'reward_repeat_penalty_std': 0.2780887484550476, 'reward_repeat_floor_mean': 0.9789993166923523, 'reward_repeat_floor_std': 0.029382452368736267, 'reward_near_duplicate_penalty_mean': 0.9428079128265381, 'reward_near_duplicate_penalty_std': 0.05258138105273247, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8828124403953552, 'reward_distinct_2_std': 0.26019859313964844, 'reward_judge_quality_mean': 0.10000000149011612, 'reward_judge_quality_std': 0.053452253341674805, 'reward_total_composite_mean': 0.06213457137346268, 'reward_total_composite_std': 0.04056248441338539, 'epoch': 0.06}

 32%|███▏      | 1063/3300 [4:03:30<9:10:41, 14.77s/it]INFO 04-12 20:01:56 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1064/3300 [4:03:46<9:27:10, 15.22s/it]2026-04-12 20:02:12,821 | INFO | train_grpo_train | metrics_logged mode=train step=1064

                                                       
{'loss': -0.0474, 'grad_norm': 1.296054720878601, 'learning_rate': 6.778787878787879e-06, 'num_tokens': 1896773.0, 'completions/mean_length': 395.875, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.3521232604980469, 'rewards/meter/std': 0.38961490988731384, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.7096376419067383, 'rewards/lexical_plausibility/std': 0.18482549488544464, 'rewards/judge_quality/mean': 0.11249999701976776, 'rewards/judge_quality/std': 0.1482035368680954, 'rewards/repeat_penalty/mean': 0.9406920075416565, 'rewards/repeat_penalty/std': 0.08965983241796494, 'rewards/repeat_floor/mean': 0.9944422245025635, 'rewards/repeat_floor/std': 0.006898265331983566, 'rewards/near_duplicate_penalty/mean': 0.9816901683807373, 'rewards/near_duplicate_penalty/std': 0.026689769700169563, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9895961284637451, 'rewards/distinct_2/std': 0.024498410522937775, 'rewards/total_composite/mean': 0.08751553297042847, 'rewards/total_composite/std': 0.15225645899772644, 'reward': 0.08751553297042847, 'reward_std': 0.15225645899772644, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17415796220302582, 'sampling/sampling_logp_difference/max': 0.9947662353515625, 'sampling/importance_sampling_ratio/min': 0.3698098957538605, 'sampling/importance_sampling_ratio/mean': 1.0123945474624634, 'sampling/importance_sampling_ratio/max': 1.951388955116272, 'entropy': 0.2597910240292549, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.036846186965703964, 'clip_ratio/high_max': 0.036846186965703964, 'clip_ratio/region_mean': 0.036846186965703964, 'reward_total_mean': 0.08751553297042847, 'reward_meter_mean': 0.3521232604980469, 'reward_meter_std': 0.38961490988731384, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.7096376419067383, 'reward_lexical_plausibility_std': 0.18482549488544464, 'reward_repeat_penalty_mean': 0.9406920075416565, 'reward_repeat_penalty_std': 0.08965983241796494, 'reward_repeat_floor_mean': 0.9944422245025635, 'reward_repeat_floor_std': 0.006898265331983566, 'reward_near_duplicate_penalty_mean': 0.9816901683807373, 'reward_near_duplicate_penalty_std': 0.026689769700169563, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9895961284637451, 'reward_distinct_2_std': 0.024498410522937775, 'reward_judge_quality_mean': 0.11249999701976776, 'reward_judge_quality_std': 0.1482035368680954, 'reward_total_composite_mean': 0.08751553297042847, 'reward_total_composite_std': 0.15225645899772644, 'epoch': 0.06}

 32%|███▏      | 1064/3300 [4:03:46<9:27:10, 15.22s/it]INFO 04-12 20:02:13 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1065/3300 [4:04:01<9:22:31, 15.10s/it]2026-04-12 20:02:27,635 | INFO | train_grpo_train | metrics_logged mode=train step=1065

                                                       
{'loss': -0.1071, 'grad_norm': 2.836012840270996, 'learning_rate': 6.7757575757575765e-06, 'num_tokens': 1898597.0, 'completions/mean_length': 242.0, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.7483861446380615, 'rewards/meter/std': 0.3190765976905823, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 0.873509407043457, 'rewards/lexical_plausibility/std': 0.17771925032138824, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.4131460189819336, 'rewards/repeat_penalty/std': 0.2663205862045288, 'rewards/repeat_floor/mean': 0.9226658940315247, 'rewards/repeat_floor/std': 0.04311244934797287, 'rewards/near_duplicate_penalty/mean': 0.7836589813232422, 'rewards/near_duplicate_penalty/std': 0.09089522063732147, 'rewards/opening_diversity/mean': 0.7680398225784302, 'rewards/opening_diversity/std': 0.18792153894901276, 'rewards/distinct_2/mean': 0.6559408903121948, 'rewards/distinct_2/std': 0.23752181231975555, 'rewards/total_composite/mean': 0.1310032308101654, 'rewards/total_composite/std': 0.1257311999797821, 'reward': 0.1310032308101654, 'reward_std': 0.1257311999797821, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1252867728471756, 'sampling/sampling_logp_difference/max': 1.4811582565307617, 'sampling/importance_sampling_ratio/min': 0.22737418115139008, 'sampling/importance_sampling_ratio/mean': 1.019413948059082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5311301052570343, 'clip_ratio/low_mean': 0.028292838484048843, 'clip_ratio/low_min': 0.028292838484048843, 'clip_ratio/high_mean': 0.044053866527974606, 'clip_ratio/high_max': 0.044053866527974606, 'clip_ratio/region_mean': 0.07234670501202345, 'reward_total_mean': 0.1310032308101654, 'reward_meter_mean': 0.7483861446380615, 'reward_meter_std': 0.3190765976905823, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 0.873509407043457, 'reward_lexical_plausibility_std': 0.17771925032138824, 'reward_repeat_penalty_mean': 0.4131460189819336, 'reward_repeat_penalty_std': 0.2663205862045288, 'reward_repeat_floor_mean': 0.9226658940315247, 'reward_repeat_floor_std': 0.04311244934797287, 'reward_near_duplicate_penalty_mean': 0.7836589813232422, 'reward_near_duplicate_penalty_std': 0.09089522063732147, 'reward_opening_diversity_mean': 0.7680398225784302, 'reward_opening_diversity_std': 0.18792153894901276, 'reward_distinct_2_mean': 0.6559408903121948, 'reward_distinct_2_std': 0.23752181231975555, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.1310032308101654, 'reward_total_composite_std': 0.1257311999797821, 'epoch': 0.06}

 32%|███▏      | 1065/3300 [4:04:01<9:22:31, 15.10s/it]INFO 04-12 20:02:27 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1066/3300 [4:04:16<9:21:43, 15.09s/it]2026-04-12 20:02:42,697 | INFO | train_grpo_train | metrics_logged mode=train step=1066

                                                       
{'loss': -0.0411, 'grad_norm': 2.7471554279327393, 'learning_rate': 6.772727272727273e-06, 'num_tokens': 1899906.0, 'completions/mean_length': 397.625, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.3778165578842163, 'rewards/meter/std': 0.32139384746551514, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.7011655569076538, 'rewards/lexical_plausibility/std': 0.19055898487567902, 'rewards/judge_quality/mean': 0.10625000298023224, 'rewards/judge_quality/std': 0.14500616490840912, 'rewards/repeat_penalty/mean': 0.9135193824768066, 'rewards/repeat_penalty/std': 0.1693478226661682, 'rewards/repeat_floor/mean': 0.9898003339767456, 'rewards/repeat_floor/std': 0.01800340972840786, 'rewards/near_duplicate_penalty/mean': 0.9542979598045349, 'rewards/near_duplicate_penalty/std': 0.062135085463523865, 'rewards/opening_diversity/mean': 0.9553571343421936, 'rewards/opening_diversity/std': 0.12626907229423523, 'rewards/distinct_2/mean': 0.9926029443740845, 'rewards/distinct_2/std': 0.014667045325040817, 'rewards/total_composite/mean': 0.034173764288425446, 'rewards/total_composite/std': 0.058432213962078094, 'reward': 0.034173764288425446, 'reward_std': 0.058432210236787796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17999720573425293, 'sampling/sampling_logp_difference/max': 0.9929285049438477, 'sampling/importance_sampling_ratio/min': 0.37049010396003723, 'sampling/importance_sampling_ratio/mean': 1.0718178749084473, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4738563895225525, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.0234375, 'clip_ratio/high_max': 0.0234375, 'clip_ratio/region_mean': 0.037326388992369175, 'reward_total_mean': 0.034173764288425446, 'reward_meter_mean': 0.3778165578842163, 'reward_meter_std': 0.32139384746551514, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.7011655569076538, 'reward_lexical_plausibility_std': 0.19055898487567902, 'reward_repeat_penalty_mean': 0.9135193824768066, 'reward_repeat_penalty_std': 0.1693478226661682, 'reward_repeat_floor_mean': 0.9898003339767456, 'reward_repeat_floor_std': 0.01800340972840786, 'reward_near_duplicate_penalty_mean': 0.9542979598045349, 'reward_near_duplicate_penalty_std': 0.062135085463523865, 'reward_opening_diversity_mean': 0.9553571343421936, 'reward_opening_diversity_std': 0.12626907229423523, 'reward_distinct_2_mean': 0.9926029443740845, 'reward_distinct_2_std': 0.014667045325040817, 'reward_judge_quality_mean': 0.10625000298023224, 'reward_judge_quality_std': 0.14500616490840912, 'reward_total_composite_mean': 0.034173764288425446, 'reward_total_composite_std': 0.058432213962078094, 'epoch': 0.06}

 32%|███▏      | 1066/3300 [4:04:16<9:21:43, 15.09s/it]INFO 04-12 20:02:42 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1067/3300 [4:04:32<9:30:13, 15.32s/it]2026-04-12 20:02:58,595 | INFO | train_grpo_train | metrics_logged mode=train step=1067

                                                       
{'loss': -0.0823, 'grad_norm': 1.6724656820297241, 'learning_rate': 6.76969696969697e-06, 'num_tokens': 1901370.0, 'completions/mean_length': 222.0, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6790182590484619, 'rewards/meter/std': 0.32086798548698425, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.1060660183429718, 'rewards/lexical_plausibility/mean': 0.8339661359786987, 'rewards/lexical_plausibility/std': 0.23012177646160126, 'rewards/judge_quality/mean': 0.2562500238418579, 'rewards/judge_quality/std': 0.1898072361946106, 'rewards/repeat_penalty/mean': 0.9172540307044983, 'rewards/repeat_penalty/std': 0.06828992068767548, 'rewards/repeat_floor/mean': 0.9890499114990234, 'rewards/repeat_floor/std': 0.008878531865775585, 'rewards/near_duplicate_penalty/mean': 0.9455007314682007, 'rewards/near_duplicate_penalty/std': 0.0421881340444088, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9691647291183472, 'rewards/distinct_2/std': 0.03645267337560654, 'rewards/total_composite/mean': 0.15745678544044495, 'rewards/total_composite/std': 0.14194640517234802, 'reward': 0.15745678544044495, 'reward_std': 0.14194639027118683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14348484575748444, 'sampling/sampling_logp_difference/max': 1.2481822967529297, 'sampling/importance_sampling_ratio/min': 0.2870260775089264, 'sampling/importance_sampling_ratio/mean': 1.014376163482666, 'sampling/importance_sampling_ratio/max': 1.798269271850586, 'entropy': 0.680285319685936, 'clip_ratio/low_mean': 0.02812499925494194, 'clip_ratio/low_min': 0.02812499925494194, 'clip_ratio/high_mean': 0.10010830871760845, 'clip_ratio/high_max': 0.10010830871760845, 'clip_ratio/region_mean': 0.1282333079725504, 'reward_total_mean': 0.15745678544044495, 'reward_meter_mean': 0.6790182590484619, 'reward_meter_std': 0.32086798548698425, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.1060660183429718, 'reward_lexical_plausibility_mean': 0.8339661359786987, 'reward_lexical_plausibility_std': 0.23012177646160126, 'reward_repeat_penalty_mean': 0.9172540307044983, 'reward_repeat_penalty_std': 0.06828992068767548, 'reward_repeat_floor_mean': 0.9890499114990234, 'reward_repeat_floor_std': 0.008878531865775585, 'reward_near_duplicate_penalty_mean': 0.9455007314682007, 'reward_near_duplicate_penalty_std': 0.0421881340444088, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9691647291183472, 'reward_distinct_2_std': 0.03645267337560654, 'reward_judge_quality_mean': 0.2562500238418579, 'reward_judge_quality_std': 0.1898072361946106, 'reward_total_composite_mean': 0.15745678544044495, 'reward_total_composite_std': 0.14194640517234802, 'epoch': 0.06}

 32%|███▏      | 1067/3300 [4:04:32<9:30:13, 15.32s/it]INFO 04-12 20:02:58 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1068/3300 [4:04:41<8:21:19, 13.48s/it]2026-04-12 20:03:07,738 | INFO | train_grpo_train | metrics_logged mode=train step=1068

                                                       
{'loss': -0.0231, 'grad_norm': 13.393712043762207, 'learning_rate': 6.7666666666666665e-06, 'num_tokens': 1902919.0, 'completions/mean_length': 20.625, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9761309623718262, 'rewards/meter/std': 0.036653872579336166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.35419678688049316, 'rewards/repeat_penalty/mean': 0.7455357313156128, 'rewards/repeat_penalty/std': 0.012626901268959045, 'rewards/repeat_floor/mean': 0.9841071367263794, 'rewards/repeat_floor/std': 0.002525375923141837, 'rewards/near_duplicate_penalty/mean': 0.9940476417541504, 'rewards/near_duplicate_penalty/std': 0.016835875809192657, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5899828672409058, 'rewards/total_composite/std': 0.3566151559352875, 'reward': 0.5899828672409058, 'reward_std': 0.3566151559352875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1127404049038887, 'sampling/sampling_logp_difference/max': 0.7840399742126465, 'sampling/importance_sampling_ratio/min': 0.4565578103065491, 'sampling/importance_sampling_ratio/mean': 1.0180014371871948, 'sampling/importance_sampling_ratio/max': 1.9889965057373047, 'entropy': 0.6647360324859619, 'clip_ratio/low_mean': 0.0624323608353734, 'clip_ratio/low_min': 0.0624323608353734, 'clip_ratio/high_mean': 0.052759740967303514, 'clip_ratio/high_max': 0.052759740967303514, 'clip_ratio/region_mean': 0.11519210180267692, 'reward_total_mean': 0.5899828672409058, 'reward_meter_mean': 0.9761309623718262, 'reward_meter_std': 0.036653872579336166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7455357313156128, 'reward_repeat_penalty_std': 0.012626901268959045, 'reward_repeat_floor_mean': 0.9841071367263794, 'reward_repeat_floor_std': 0.002525375923141837, 'reward_near_duplicate_penalty_mean': 0.9940476417541504, 'reward_near_duplicate_penalty_std': 0.016835875809192657, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.35419678688049316, 'reward_total_composite_mean': 0.5899828672409058, 'reward_total_composite_std': 0.3566151559352875, 'epoch': 0.06}

 32%|███▏      | 1068/3300 [4:04:41<8:21:19, 13.48s/it]INFO 04-12 20:03:07 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1069/3300 [4:04:57<8:51:10, 14.29s/it]2026-04-12 20:03:23,918 | INFO | train_grpo_train | metrics_logged mode=train step=1069

                                                       
{'loss': -0.123, 'grad_norm': 1.8636703491210938, 'learning_rate': 6.763636363636365e-06, 'num_tokens': 1904484.0, 'completions/mean_length': 100.625, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.85714340209961, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8739906549453735, 'rewards/meter/std': 0.30079954862594604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9546717405319214, 'rewards/lexical_plausibility/std': 0.12820763885974884, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.710982084274292, 'rewards/repeat_penalty/std': 0.18087314069271088, 'rewards/repeat_floor/mean': 0.9623284339904785, 'rewards/repeat_floor/std': 0.022545693442225456, 'rewards/near_duplicate_penalty/mean': 0.8514623641967773, 'rewards/near_duplicate_penalty/std': 0.06669747829437256, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8498234748840332, 'rewards/distinct_2/std': 0.12484649568796158, 'rewards/total_composite/mean': 0.31140628457069397, 'rewards/total_composite/std': 0.12951219081878662, 'reward': 0.31140628457069397, 'reward_std': 0.12951219081878662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10668300092220306, 'sampling/sampling_logp_difference/max': 1.5623292922973633, 'sampling/importance_sampling_ratio/min': 0.20964716374874115, 'sampling/importance_sampling_ratio/mean': 1.0178141593933105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48977963626384735, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10859350766986609, 'clip_ratio/high_max': 0.10859350766986609, 'clip_ratio/region_mean': 0.10859350766986609, 'reward_total_mean': 0.31140628457069397, 'reward_meter_mean': 0.8739906549453735, 'reward_meter_std': 0.30079954862594604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9546717405319214, 'reward_lexical_plausibility_std': 0.12820763885974884, 'reward_repeat_penalty_mean': 0.710982084274292, 'reward_repeat_penalty_std': 0.18087314069271088, 'reward_repeat_floor_mean': 0.9623284339904785, 'reward_repeat_floor_std': 0.022545693442225456, 'reward_near_duplicate_penalty_mean': 0.8514623641967773, 'reward_near_duplicate_penalty_std': 0.06669747829437256, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8498234748840332, 'reward_distinct_2_std': 0.12484649568796158, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.31140628457069397, 'reward_total_composite_std': 0.12951219081878662, 'epoch': 0.06}

 32%|███▏      | 1069/3300 [4:04:57<8:51:10, 14.29s/it]INFO 04-12 20:03:24 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1070/3300 [4:05:16<9:38:24, 15.56s/it]2026-04-12 20:03:42,573 | INFO | train_grpo_train | metrics_logged mode=train step=1070

                                                       
{'loss': -0.0979, 'grad_norm': 4.400808334350586, 'learning_rate': 6.760606060606061e-06, 'num_tokens': 1907799.0, 'completions/mean_length': 263.375, 'completions/min_length': 210.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 227.85714721679688, 'completions/min_terminated_length': 210.0, 'completions/max_terminated_length': 251.0, 'rewards/meter/mean': 0.6059356927871704, 'rewards/meter/std': 0.23981502652168274, 'rewards/count_adherence/mean': 0.8229166269302368, 'rewards/count_adherence/std': 0.13684004545211792, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9468749761581421, 'rewards/count_floor/std': 0.04105200991034508, 'rewards/lexical_plausibility/mean': 0.961497962474823, 'rewards/lexical_plausibility/std': 0.10890025645494461, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.4514191150665283, 'rewards/repeat_penalty/std': 0.23794931173324585, 'rewards/repeat_floor/mean': 0.9416136741638184, 'rewards/repeat_floor/std': 0.02634624019265175, 'rewards/near_duplicate_penalty/mean': 0.9047547578811646, 'rewards/near_duplicate_penalty/std': 0.03628513216972351, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.5100048184394836, 'rewards/distinct_2/std': 0.2374163269996643, 'rewards/total_composite/mean': 0.10428134351968765, 'rewards/total_composite/std': 0.07745161652565002, 'reward': 0.10428134351968765, 'reward_std': 0.07745160907506943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12649211287498474, 'sampling/sampling_logp_difference/max': 1.844409465789795, 'sampling/importance_sampling_ratio/min': 0.15811868011951447, 'sampling/importance_sampling_ratio/mean': 0.9979678988456726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6061956956982613, 'clip_ratio/low_mean': 0.05695985537022352, 'clip_ratio/low_min': 0.05695985537022352, 'clip_ratio/high_mean': 0.05597740225493908, 'clip_ratio/high_max': 0.05597740225493908, 'clip_ratio/region_mean': 0.1129372576251626, 'reward_total_mean': 0.10428134351968765, 'reward_meter_mean': 0.6059356927871704, 'reward_meter_std': 0.23981502652168274, 'reward_count_adherence_mean': 0.8229166269302368, 'reward_count_adherence_std': 0.13684004545211792, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9468749761581421, 'reward_count_floor_std': 0.04105200991034508, 'reward_lexical_plausibility_mean': 0.961497962474823, 'reward_lexical_plausibility_std': 0.10890025645494461, 'reward_repeat_penalty_mean': 0.4514191150665283, 'reward_repeat_penalty_std': 0.23794931173324585, 'reward_repeat_floor_mean': 0.9416136741638184, 'reward_repeat_floor_std': 0.02634624019265175, 'reward_near_duplicate_penalty_mean': 0.9047547578811646, 'reward_near_duplicate_penalty_std': 0.03628513216972351, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.5100048184394836, 'reward_distinct_2_std': 0.2374163269996643, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.10428134351968765, 'reward_total_composite_std': 0.07745161652565002, 'epoch': 0.06}

 32%|███▏      | 1070/3300 [4:05:16<9:38:24, 15.56s/it]INFO 04-12 20:03:42 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1071/3300 [4:05:26<8:36:31, 13.90s/it]2026-04-12 20:03:52,482 | INFO | train_grpo_train | metrics_logged mode=train step=1071

                                                       
{'loss': 0.0531, 'grad_norm': 11.77622127532959, 'learning_rate': 6.757575757575758e-06, 'num_tokens': 1910020.0, 'completions/mean_length': 99.625, 'completions/min_length': 88.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.625, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.567720890045166, 'rewards/meter/std': 0.3172169625759125, 'rewards/count_adherence/mean': 0.8571428656578064, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9571428298950195, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.16690458357334137, 'rewards/repeat_penalty/mean': 0.9097242951393127, 'rewards/repeat_penalty/std': 0.08820004761219025, 'rewards/repeat_floor/mean': 0.9922217726707458, 'rewards/repeat_floor/std': 0.006347257178276777, 'rewards/near_duplicate_penalty/mean': 0.9796038269996643, 'rewards/near_duplicate_penalty/std': 0.009924345649778843, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9475690126419067, 'rewards/distinct_2/std': 0.05622003972530365, 'rewards/total_composite/mean': 0.2502760589122772, 'rewards/total_composite/std': 0.2120453417301178, 'reward': 0.2502760589122772, 'reward_std': 0.2120453417301178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1844819337129593, 'sampling/sampling_logp_difference/max': 3.5286781787872314, 'sampling/importance_sampling_ratio/min': 0.029343679547309875, 'sampling/importance_sampling_ratio/mean': 0.9935512542724609, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7695295810699463, 'clip_ratio/low_mean': 0.09332773555070162, 'clip_ratio/low_min': 0.09332773555070162, 'clip_ratio/high_mean': 0.06373801082372665, 'clip_ratio/high_max': 0.06373801082372665, 'clip_ratio/region_mean': 0.15706574637442827, 'reward_total_mean': 0.2502760589122772, 'reward_meter_mean': 0.567720890045166, 'reward_meter_std': 0.3172169625759125, 'reward_count_adherence_mean': 0.8571428656578064, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9571428298950195, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9097242951393127, 'reward_repeat_penalty_std': 0.08820004761219025, 'reward_repeat_floor_mean': 0.9922217726707458, 'reward_repeat_floor_std': 0.006347257178276777, 'reward_near_duplicate_penalty_mean': 0.9796038269996643, 'reward_near_duplicate_penalty_std': 0.009924345649778843, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9475690126419067, 'reward_distinct_2_std': 0.05622003972530365, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.16690458357334137, 'reward_total_composite_mean': 0.2502760589122772, 'reward_total_composite_std': 0.2120453417301178, 'epoch': 0.06}

 32%|███▏      | 1071/3300 [4:05:26<8:36:31, 13.90s/it]INFO 04-12 20:03:52 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1072/3300 [4:05:35<7:42:56, 12.47s/it]2026-04-12 20:04:01,636 | INFO | train_grpo_train | metrics_logged mode=train step=1072

                                                       
{'loss': 0.0181, 'grad_norm': 20.931276321411133, 'learning_rate': 6.754545454545455e-06, 'num_tokens': 1911435.0, 'completions/mean_length': 29.875, 'completions/min_length': 25.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.5035991072654724, 'rewards/meter/std': 0.4299202561378479, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/repeat_penalty/mean': 0.9292556047439575, 'rewards/repeat_penalty/std': 0.16393889486789703, 'rewards/repeat_floor/mean': 0.9918794631958008, 'rewards/repeat_floor/std': 0.01759110577404499, 'rewards/near_duplicate_penalty/mean': 0.9737478494644165, 'rewards/near_duplicate_penalty/std': 0.04055837541818619, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9743589758872986, 'rewards/distinct_2/std': 0.07252377271652222, 'rewards/total_composite/mean': 0.34068629145622253, 'rewards/total_composite/std': 0.3119940161705017, 'reward': 0.34068629145622253, 'reward_std': 0.3119940161705017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15305404365062714, 'sampling/sampling_logp_difference/max': 1.2168183326721191, 'sampling/importance_sampling_ratio/min': 0.2961709797382355, 'sampling/importance_sampling_ratio/mean': 1.0122547149658203, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.810894064605236, 'clip_ratio/low_mean': 0.05046419193968177, 'clip_ratio/low_min': 0.05046419193968177, 'clip_ratio/high_mean': 0.06753788003697991, 'clip_ratio/high_max': 0.06753788003697991, 'clip_ratio/region_mean': 0.11800207197666168, 'reward_total_mean': 0.34068629145622253, 'reward_meter_mean': 0.5035991072654724, 'reward_meter_std': 0.4299202561378479, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9292556047439575, 'reward_repeat_penalty_std': 0.16393889486789703, 'reward_repeat_floor_mean': 0.9918794631958008, 'reward_repeat_floor_std': 0.01759110577404499, 'reward_near_duplicate_penalty_mean': 0.9737478494644165, 'reward_near_duplicate_penalty_std': 0.04055837541818619, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9743589758872986, 'reward_distinct_2_std': 0.07252377271652222, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.34068629145622253, 'reward_total_composite_std': 0.3119940161705017, 'epoch': 0.06}

 32%|███▏      | 1072/3300 [4:05:35<7:42:56, 12.47s/it]INFO 04-12 20:04:01 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1073/3300 [4:05:51<8:29:03, 13.72s/it]2026-04-12 20:04:18,268 | INFO | train_grpo_train | metrics_logged mode=train step=1073

                                                       
{'loss': -0.0983, 'grad_norm': 4.400396823883057, 'learning_rate': 6.751515151515152e-06, 'num_tokens': 1914004.0, 'completions/mean_length': 182.125, 'completions/min_length': 115.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 135.0, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9867106676101685, 'rewards/meter/std': 0.013244129717350006, 'rewards/count_adherence/mean': 0.6428571343421936, 'rewards/count_adherence/std': 0.2159797102212906, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8928571343421936, 'rewards/count_floor/std': 0.0647939071059227, 'rewards/lexical_plausibility/mean': 0.947879433631897, 'rewards/lexical_plausibility/std': 0.14741921424865723, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.6394490599632263, 'rewards/repeat_penalty/std': 0.16927896440029144, 'rewards/repeat_floor/mean': 0.9610540270805359, 'rewards/repeat_floor/std': 0.017209012061357498, 'rewards/near_duplicate_penalty/mean': 0.9004266262054443, 'rewards/near_duplicate_penalty/std': 0.036961011588573456, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.7415555715560913, 'rewards/distinct_2/std': 0.13002930581569672, 'rewards/total_composite/mean': 0.18242649734020233, 'rewards/total_composite/std': 0.10402916371822357, 'reward': 0.18242649734020233, 'reward_std': 0.10402915626764297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13504239916801453, 'sampling/sampling_logp_difference/max': 2.2505435943603516, 'sampling/importance_sampling_ratio/min': 0.10534194856882095, 'sampling/importance_sampling_ratio/mean': 1.0122077465057373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7037039622664452, 'clip_ratio/low_mean': 0.07036673557013273, 'clip_ratio/low_min': 0.07036673557013273, 'clip_ratio/high_mean': 0.051998681388795376, 'clip_ratio/high_max': 0.051998681388795376, 'clip_ratio/region_mean': 0.12236541695892811, 'reward_total_mean': 0.18242649734020233, 'reward_meter_mean': 0.9867106676101685, 'reward_meter_std': 0.013244129717350006, 'reward_count_adherence_mean': 0.6428571343421936, 'reward_count_adherence_std': 0.2159797102212906, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8928571343421936, 'reward_count_floor_std': 0.0647939071059227, 'reward_lexical_plausibility_mean': 0.947879433631897, 'reward_lexical_plausibility_std': 0.14741921424865723, 'reward_repeat_penalty_mean': 0.6394490599632263, 'reward_repeat_penalty_std': 0.16927896440029144, 'reward_repeat_floor_mean': 0.9610540270805359, 'reward_repeat_floor_std': 0.017209012061357498, 'reward_near_duplicate_penalty_mean': 0.9004266262054443, 'reward_near_duplicate_penalty_std': 0.036961011588573456, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.7415555715560913, 'reward_distinct_2_std': 0.13002930581569672, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.18242649734020233, 'reward_total_composite_std': 0.10402916371822357, 'epoch': 0.06}

 33%|███▎      | 1073/3300 [4:05:52<8:29:03, 13.72s/it]INFO 04-12 20:04:18 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1074/3300 [4:05:59<7:21:10, 11.89s/it]2026-04-12 20:04:25,891 | INFO | train_grpo_train | metrics_logged mode=train step=1074

                                                       
{'loss': 0.0822, 'grad_norm': 19.359661102294922, 'learning_rate': 6.748484848484848e-06, 'num_tokens': 1915340.0, 'completions/mean_length': 18.0, 'completions/min_length': 16.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7420926094055176, 'rewards/meter/std': 0.4073685109615326, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6887499690055847, 'rewards/judge_quality/std': 0.34156936407089233, 'rewards/repeat_penalty/mean': 0.5175803899765015, 'rewards/repeat_penalty/std': 0.18043386936187744, 'rewards/repeat_floor/mean': 0.9380412697792053, 'rewards/repeat_floor/std': 0.03679239749908447, 'rewards/near_duplicate_penalty/mean': 0.7305313348770142, 'rewards/near_duplicate_penalty/std': 0.19719846546649933, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9273504018783569, 'rewards/distinct_2/std': 0.10447946935892105, 'rewards/total_composite/mean': 0.4954415261745453, 'rewards/total_composite/std': 0.3863968551158905, 'reward': 0.4954415261745453, 'reward_std': 0.3863968253135681, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.119955874979496, 'sampling/sampling_logp_difference/max': 1.0646991729736328, 'sampling/importance_sampling_ratio/min': 0.34483158588409424, 'sampling/importance_sampling_ratio/mean': 0.9877858757972717, 'sampling/importance_sampling_ratio/max': 1.8000118732452393, 'entropy': 0.8861508518457413, 'clip_ratio/low_mean': 0.012531328480690718, 'clip_ratio/low_min': 0.012531328480690718, 'clip_ratio/high_mean': 0.0705785620957613, 'clip_ratio/high_max': 0.0705785620957613, 'clip_ratio/region_mean': 0.08310989057645202, 'reward_total_mean': 0.4954415261745453, 'reward_meter_mean': 0.7420926094055176, 'reward_meter_std': 0.4073685109615326, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5175803899765015, 'reward_repeat_penalty_std': 0.18043386936187744, 'reward_repeat_floor_mean': 0.9380412697792053, 'reward_repeat_floor_std': 0.03679239749908447, 'reward_near_duplicate_penalty_mean': 0.7305313348770142, 'reward_near_duplicate_penalty_std': 0.19719846546649933, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9273504018783569, 'reward_distinct_2_std': 0.10447946935892105, 'reward_judge_quality_mean': 0.6887499690055847, 'reward_judge_quality_std': 0.34156936407089233, 'reward_total_composite_mean': 0.4954415261745453, 'reward_total_composite_std': 0.3863968551158905, 'epoch': 0.06}

 33%|███▎      | 1074/3300 [4:05:59<7:21:10, 11.89s/it]INFO 04-12 20:04:26 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1075/3300 [4:06:12<7:33:56, 12.24s/it]2026-04-12 20:04:38,919 | INFO | train_grpo_train | metrics_logged mode=train step=1075

                                                       
{'loss': -0.0389, 'grad_norm': 4.987910747528076, 'learning_rate': 6.7454545454545465e-06, 'num_tokens': 1916754.0, 'completions/mean_length': 87.75, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 27.142858505249023, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6227566003799438, 'rewards/meter/std': 0.4933727979660034, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.933384120464325, 'rewards/lexical_plausibility/std': 0.1884182095527649, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.3829747438430786, 'rewards/repeat_penalty/mean': 0.742737889289856, 'rewards/repeat_penalty/std': 0.12986721098423004, 'rewards/repeat_floor/mean': 0.9807339906692505, 'rewards/repeat_floor/std': 0.014583591371774673, 'rewards/near_duplicate_penalty/mean': 0.9782905578613281, 'rewards/near_duplicate_penalty/std': 0.037107914686203, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9679487347602844, 'rewards/distinct_2/std': 0.09065470844507217, 'rewards/total_composite/mean': 0.5589503049850464, 'rewards/total_composite/std': 0.45456141233444214, 'reward': 0.5589503049850464, 'reward_std': 0.45456141233444214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1434827744960785, 'sampling/sampling_logp_difference/max': 1.3663749694824219, 'sampling/importance_sampling_ratio/min': 0.2550297677516937, 'sampling/importance_sampling_ratio/mean': 1.0104016065597534, 'sampling/importance_sampling_ratio/max': 1.837349534034729, 'entropy': 0.8188060000538826, 'clip_ratio/low_mean': 0.0528273805975914, 'clip_ratio/low_min': 0.0528273805975914, 'clip_ratio/high_mean': 0.1226099543273449, 'clip_ratio/high_max': 0.1226099543273449, 'clip_ratio/region_mean': 0.1754373349249363, 'reward_total_mean': 0.5589503049850464, 'reward_meter_mean': 0.6227566003799438, 'reward_meter_std': 0.4933727979660034, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.933384120464325, 'reward_lexical_plausibility_std': 0.1884182095527649, 'reward_repeat_penalty_mean': 0.742737889289856, 'reward_repeat_penalty_std': 0.12986721098423004, 'reward_repeat_floor_mean': 0.9807339906692505, 'reward_repeat_floor_std': 0.014583591371774673, 'reward_near_duplicate_penalty_mean': 0.9782905578613281, 'reward_near_duplicate_penalty_std': 0.037107914686203, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9679487347602844, 'reward_distinct_2_std': 0.09065470844507217, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.3829747438430786, 'reward_total_composite_mean': 0.5589503049850464, 'reward_total_composite_std': 0.45456141233444214, 'epoch': 0.06}

 33%|███▎      | 1075/3300 [4:06:12<7:33:56, 12.24s/it]INFO 04-12 20:04:39 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1076/3300 [4:06:26<7:48:42, 12.64s/it]2026-04-12 20:04:52,503 | INFO | train_grpo_train | metrics_logged mode=train step=1076

                                                       
{'loss': -0.1246, 'grad_norm': 4.763963222503662, 'learning_rate': 6.742424242424243e-06, 'num_tokens': 1918812.0, 'completions/mean_length': 135.25, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.42857360839844, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.970775842666626, 'rewards/meter/std': 0.04196593537926674, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9694584608078003, 'rewards/lexical_plausibility/std': 0.08638457953929901, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.8369671106338501, 'rewards/repeat_penalty/std': 0.12449848651885986, 'rewards/repeat_floor/mean': 0.9803304076194763, 'rewards/repeat_floor/std': 0.01572480984032154, 'rewards/near_duplicate_penalty/mean': 0.9328269958496094, 'rewards/near_duplicate_penalty/std': 0.05469420179724693, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8934038877487183, 'rewards/distinct_2/std': 0.09271004050970078, 'rewards/total_composite/mean': 0.29090413451194763, 'rewards/total_composite/std': 0.09215669333934784, 'reward': 0.29090413451194763, 'reward_std': 0.09215668588876724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14479467272758484, 'sampling/sampling_logp_difference/max': 1.4905118942260742, 'sampling/importance_sampling_ratio/min': 0.22525732219219208, 'sampling/importance_sampling_ratio/mean': 1.0108000040054321, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8082137405872345, 'clip_ratio/low_mean': 0.020220588892698288, 'clip_ratio/low_min': 0.020220588892698288, 'clip_ratio/high_mean': 0.12136476766318083, 'clip_ratio/high_max': 0.12136476766318083, 'clip_ratio/region_mean': 0.14158535655587912, 'reward_total_mean': 0.29090413451194763, 'reward_meter_mean': 0.970775842666626, 'reward_meter_std': 0.04196593537926674, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9694584608078003, 'reward_lexical_plausibility_std': 0.08638457953929901, 'reward_repeat_penalty_mean': 0.8369671106338501, 'reward_repeat_penalty_std': 0.12449848651885986, 'reward_repeat_floor_mean': 0.9803304076194763, 'reward_repeat_floor_std': 0.01572480984032154, 'reward_near_duplicate_penalty_mean': 0.9328269958496094, 'reward_near_duplicate_penalty_std': 0.05469420179724693, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8934038877487183, 'reward_distinct_2_std': 0.09271004050970078, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.29090413451194763, 'reward_total_composite_std': 0.09215669333934784, 'epoch': 0.06}

 33%|███▎      | 1076/3300 [4:06:26<7:48:42, 12.64s/it]INFO 04-12 20:04:52 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1077/3300 [4:06:36<7:26:06, 12.04s/it]2026-04-12 20:05:03,150 | INFO | train_grpo_train | metrics_logged mode=train step=1077

                                                       
{'loss': 0.0823, 'grad_norm': 7.245699405670166, 'learning_rate': 6.73939393939394e-06, 'num_tokens': 1921592.0, 'completions/mean_length': 146.5, 'completions/min_length': 119.0, 'completions/max_length': 181.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 146.5, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.6044440269470215, 'rewards/meter/std': 0.23835039138793945, 'rewards/count_adherence/mean': 0.7777777910232544, 'rewards/count_adherence/std': 0.059391383081674576, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9333332777023315, 'rewards/count_floor/std': 0.017817415297031403, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.3985518515110016, 'rewards/repeat_penalty/std': 0.234026700258255, 'rewards/repeat_floor/mean': 0.9324367046356201, 'rewards/repeat_floor/std': 0.029193731024861336, 'rewards/near_duplicate_penalty/mean': 0.8773298263549805, 'rewards/near_duplicate_penalty/std': 0.057365674525499344, 'rewards/opening_diversity/mean': 0.87109375, 'rewards/opening_diversity/std': 0.2299424558877945, 'rewards/distinct_2/mean': 0.5396846532821655, 'rewards/distinct_2/std': 0.19586032629013062, 'rewards/total_composite/mean': 0.17250597476959229, 'rewards/total_composite/std': 0.08049008250236511, 'reward': 0.17250597476959229, 'reward_std': 0.08049008250236511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11890258640050888, 'sampling/sampling_logp_difference/max': 3.552889823913574, 'sampling/importance_sampling_ratio/min': 0.028641749173402786, 'sampling/importance_sampling_ratio/mean': 1.021666407585144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6128877066075802, 'clip_ratio/low_mean': 0.06310850847512484, 'clip_ratio/low_min': 0.06310850847512484, 'clip_ratio/high_mean': 0.05600618664175272, 'clip_ratio/high_max': 0.05600618664175272, 'clip_ratio/region_mean': 0.11911469511687756, 'reward_total_mean': 0.17250597476959229, 'reward_meter_mean': 0.6044440269470215, 'reward_meter_std': 0.23835039138793945, 'reward_count_adherence_mean': 0.7777777910232544, 'reward_count_adherence_std': 0.059391383081674576, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9333332777023315, 'reward_count_floor_std': 0.017817415297031403, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.3985518515110016, 'reward_repeat_penalty_std': 0.234026700258255, 'reward_repeat_floor_mean': 0.9324367046356201, 'reward_repeat_floor_std': 0.029193731024861336, 'reward_near_duplicate_penalty_mean': 0.8773298263549805, 'reward_near_duplicate_penalty_std': 0.057365674525499344, 'reward_opening_diversity_mean': 0.87109375, 'reward_opening_diversity_std': 0.2299424558877945, 'reward_distinct_2_mean': 0.5396846532821655, 'reward_distinct_2_std': 0.19586032629013062, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.17250597476959229, 'reward_total_composite_std': 0.08049008250236511, 'epoch': 0.06}

 33%|███▎      | 1077/3300 [4:06:37<7:26:06, 12.04s/it]INFO 04-12 20:05:03 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1078/3300 [4:06:50<7:45:16, 12.56s/it]2026-04-12 20:05:17,025 | INFO | train_grpo_train | metrics_logged mode=train step=1078

                                                       
{'loss': -0.0222, 'grad_norm': 3.444074869155884, 'learning_rate': 6.7363636363636365e-06, 'num_tokens': 1923161.0, 'completions/mean_length': 173.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 60.16666793823242, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.5488107800483704, 'rewards/meter/std': 0.3649800419807434, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.9025658965110779, 'rewards/lexical_plausibility/std': 0.18058884143829346, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.3561275005340576, 'rewards/repeat_penalty/mean': 0.8407871723175049, 'rewards/repeat_penalty/std': 0.27340683341026306, 'rewards/repeat_floor/mean': 0.9819720387458801, 'rewards/repeat_floor/std': 0.0298537015914917, 'rewards/near_duplicate_penalty/mean': 0.9548909664154053, 'rewards/near_duplicate_penalty/std': 0.05168209224939346, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.874870777130127, 'rewards/distinct_2/std': 0.25533223152160645, 'rewards/total_composite/mean': 0.2264440655708313, 'rewards/total_composite/std': 0.29111823439598083, 'reward': 0.2264440655708313, 'reward_std': 0.29111823439598083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15008863806724548, 'sampling/sampling_logp_difference/max': 1.7263131141662598, 'sampling/importance_sampling_ratio/min': 0.1779392510652542, 'sampling/importance_sampling_ratio/mean': 1.0275452136993408, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7353262901306152, 'clip_ratio/low_mean': 0.07197060901671648, 'clip_ratio/low_min': 0.07197060901671648, 'clip_ratio/high_mean': 0.029941861517727375, 'clip_ratio/high_max': 0.029941861517727375, 'clip_ratio/region_mean': 0.10191247053444386, 'reward_total_mean': 0.2264440655708313, 'reward_meter_mean': 0.5488107800483704, 'reward_meter_std': 0.3649800419807434, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.9025658965110779, 'reward_lexical_plausibility_std': 0.18058884143829346, 'reward_repeat_penalty_mean': 0.8407871723175049, 'reward_repeat_penalty_std': 0.27340683341026306, 'reward_repeat_floor_mean': 0.9819720387458801, 'reward_repeat_floor_std': 0.0298537015914917, 'reward_near_duplicate_penalty_mean': 0.9548909664154053, 'reward_near_duplicate_penalty_std': 0.05168209224939346, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.874870777130127, 'reward_distinct_2_std': 0.25533223152160645, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.3561275005340576, 'reward_total_composite_mean': 0.2264440655708313, 'reward_total_composite_std': 0.29111823439598083, 'epoch': 0.06}

 33%|███▎      | 1078/3300 [4:06:50<7:45:16, 12.56s/it]INFO 04-12 20:05:17 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1079/3300 [4:07:06<8:17:20, 13.44s/it]2026-04-12 20:05:32,434 | INFO | train_grpo_train | metrics_logged mode=train step=1079

                                                       
{'loss': -0.1521, 'grad_norm': 2.4961585998535156, 'learning_rate': 6.733333333333334e-06, 'num_tokens': 1925263.0, 'completions/mean_length': 259.75, 'completions/min_length': 104.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 108.4000015258789, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.9006693363189697, 'rewards/meter/std': 0.09917160868644714, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.08017838001251221, 'rewards/lexical_plausibility/mean': 0.8560393452644348, 'rewards/lexical_plausibility/std': 0.2027360051870346, 'rewards/judge_quality/mean': 0.3212500214576721, 'rewards/judge_quality/std': 0.2773052752017975, 'rewards/repeat_penalty/mean': 0.9579098224639893, 'rewards/repeat_penalty/std': 0.0416024886071682, 'rewards/repeat_floor/mean': 0.9952350854873657, 'rewards/repeat_floor/std': 0.003988961223512888, 'rewards/near_duplicate_penalty/mean': 0.9844063520431519, 'rewards/near_duplicate_penalty/std': 0.008151083253324032, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.973046600818634, 'rewards/distinct_2/std': 0.04050234705209732, 'rewards/total_composite/mean': 0.27432894706726074, 'rewards/total_composite/std': 0.20733150839805603, 'reward': 0.27432894706726074, 'reward_std': 0.20733147859573364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12474246323108673, 'sampling/sampling_logp_difference/max': 1.5720982551574707, 'sampling/importance_sampling_ratio/min': 0.20760910212993622, 'sampling/importance_sampling_ratio/mean': 1.0134941339492798, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44906847178936005, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08068029396235943, 'clip_ratio/high_max': 0.08068029396235943, 'clip_ratio/region_mean': 0.08068029396235943, 'reward_total_mean': 0.27432894706726074, 'reward_meter_mean': 0.9006693363189697, 'reward_meter_std': 0.09917160868644714, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.08017838001251221, 'reward_lexical_plausibility_mean': 0.8560393452644348, 'reward_lexical_plausibility_std': 0.2027360051870346, 'reward_repeat_penalty_mean': 0.9579098224639893, 'reward_repeat_penalty_std': 0.0416024886071682, 'reward_repeat_floor_mean': 0.9952350854873657, 'reward_repeat_floor_std': 0.003988961223512888, 'reward_near_duplicate_penalty_mean': 0.9844063520431519, 'reward_near_duplicate_penalty_std': 0.008151083253324032, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.973046600818634, 'reward_distinct_2_std': 0.04050234705209732, 'reward_judge_quality_mean': 0.3212500214576721, 'reward_judge_quality_std': 0.2773052752017975, 'reward_total_composite_mean': 0.27432894706726074, 'reward_total_composite_std': 0.20733150839805603, 'epoch': 0.06}

 33%|███▎      | 1079/3300 [4:07:06<8:17:20, 13.44s/it]INFO 04-12 20:05:32 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1080/3300 [4:07:14<7:15:29, 11.77s/it]2026-04-12 20:05:40,268 | INFO | train_grpo_train | metrics_logged mode=train step=1080

                                                       
{'loss': 0.0497, 'grad_norm': 16.204898834228516, 'learning_rate': 6.73030303030303e-06, 'num_tokens': 1926756.0, 'completions/mean_length': 27.625, 'completions/min_length': 20.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9221764206886292, 'rewards/meter/std': 0.19007368385791779, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8350000381469727, 'rewards/judge_quality/std': 0.27717968821525574, 'rewards/repeat_penalty/mean': 0.6980905532836914, 'rewards/repeat_penalty/std': 0.10256372392177582, 'rewards/repeat_floor/mean': 0.9735473394393921, 'rewards/repeat_floor/std': 0.019675888121128082, 'rewards/near_duplicate_penalty/mean': 0.9372479319572449, 'rewards/near_duplicate_penalty/std': 0.10603788495063782, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.956501841545105, 'rewards/distinct_2/std': 0.08417937159538269, 'rewards/total_composite/mean': 0.7104453444480896, 'rewards/total_composite/std': 0.2962467670440674, 'reward': 0.7104453444480896, 'reward_std': 0.2962467670440674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10239305347204208, 'sampling/sampling_logp_difference/max': 1.2080011367797852, 'sampling/importance_sampling_ratio/min': 0.2987939119338989, 'sampling/importance_sampling_ratio/mean': 1.0033000707626343, 'sampling/importance_sampling_ratio/max': 1.7388825416564941, 'entropy': 0.8475519418716431, 'clip_ratio/low_mean': 0.046653298661112785, 'clip_ratio/low_min': 0.046653298661112785, 'clip_ratio/high_mean': 0.08468259172514081, 'clip_ratio/high_max': 0.08468259172514081, 'clip_ratio/region_mean': 0.1313358903862536, 'reward_total_mean': 0.7104453444480896, 'reward_meter_mean': 0.9221764206886292, 'reward_meter_std': 0.19007368385791779, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6980905532836914, 'reward_repeat_penalty_std': 0.10256372392177582, 'reward_repeat_floor_mean': 0.9735473394393921, 'reward_repeat_floor_std': 0.019675888121128082, 'reward_near_duplicate_penalty_mean': 0.9372479319572449, 'reward_near_duplicate_penalty_std': 0.10603788495063782, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.956501841545105, 'reward_distinct_2_std': 0.08417937159538269, 'reward_judge_quality_mean': 0.8350000381469727, 'reward_judge_quality_std': 0.27717968821525574, 'reward_total_composite_mean': 0.7104453444480896, 'reward_total_composite_std': 0.2962467670440674, 'epoch': 0.06}

 33%|███▎      | 1080/3300 [4:07:14<7:15:29, 11.77s/it]INFO 04-12 20:05:40 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1081/3300 [4:07:28<7:44:16, 12.55s/it]2026-04-12 20:05:54,697 | INFO | train_grpo_train | metrics_logged mode=train step=1081

                                                       
{'loss': -0.0051, 'grad_norm': 5.6562700271606445, 'learning_rate': 6.7272727272727275e-06, 'num_tokens': 1928367.0, 'completions/mean_length': 100.375, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.57143020629883, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.41312259435653687, 'rewards/meter/std': 0.42788901925086975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9441838264465332, 'rewards/lexical_plausibility/std': 0.15787187218666077, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101401686668396, 'rewards/repeat_penalty/mean': 0.5048427581787109, 'rewards/repeat_penalty/std': 0.20335707068443298, 'rewards/repeat_floor/mean': 0.942783772945404, 'rewards/repeat_floor/std': 0.02365000732243061, 'rewards/near_duplicate_penalty/mean': 0.8348733186721802, 'rewards/near_duplicate_penalty/std': 0.06673448532819748, 'rewards/opening_diversity/mean': 0.737500011920929, 'rewards/opening_diversity/std': 0.24712635576725006, 'rewards/distinct_2/mean': 0.8144752979278564, 'rewards/distinct_2/std': 0.09979349374771118, 'rewards/total_composite/mean': 0.16769617795944214, 'rewards/total_composite/std': 0.25463947653770447, 'reward': 0.16769617795944214, 'reward_std': 0.25463947653770447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15469835698604584, 'sampling/sampling_logp_difference/max': 1.55474853515625, 'sampling/importance_sampling_ratio/min': 0.21124249696731567, 'sampling/importance_sampling_ratio/mean': 1.0261539220809937, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6518596932291985, 'clip_ratio/low_mean': 0.08085842244327068, 'clip_ratio/low_min': 0.08085842244327068, 'clip_ratio/high_mean': 0.033882785588502884, 'clip_ratio/high_max': 0.033882785588502884, 'clip_ratio/region_mean': 0.11474120803177357, 'reward_total_mean': 0.16769617795944214, 'reward_meter_mean': 0.41312259435653687, 'reward_meter_std': 0.42788901925086975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9441838264465332, 'reward_lexical_plausibility_std': 0.15787187218666077, 'reward_repeat_penalty_mean': 0.5048427581787109, 'reward_repeat_penalty_std': 0.20335707068443298, 'reward_repeat_floor_mean': 0.942783772945404, 'reward_repeat_floor_std': 0.02365000732243061, 'reward_near_duplicate_penalty_mean': 0.8348733186721802, 'reward_near_duplicate_penalty_std': 0.06673448532819748, 'reward_opening_diversity_mean': 0.737500011920929, 'reward_opening_diversity_std': 0.24712635576725006, 'reward_distinct_2_mean': 0.8144752979278564, 'reward_distinct_2_std': 0.09979349374771118, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101401686668396, 'reward_total_composite_mean': 0.16769617795944214, 'reward_total_composite_std': 0.25463947653770447, 'epoch': 0.06}

 33%|███▎      | 1081/3300 [4:07:28<7:44:16, 12.55s/it]INFO 04-12 20:05:55 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1082/3300 [4:07:40<7:39:22, 12.43s/it]2026-04-12 20:06:06,776 | INFO | train_grpo_train | metrics_logged mode=train step=1082

                                                       
{'loss': 0.232, 'grad_norm': 16.52114486694336, 'learning_rate': 6.724242424242424e-06, 'num_tokens': 1929907.0, 'completions/mean_length': 29.5, 'completions/min_length': 17.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6338993310928345, 'rewards/meter/std': 0.26156964898109436, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.16035674512386322, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.8179764747619629, 'rewards/repeat_penalty/std': 0.09199567139148712, 'rewards/repeat_floor/mean': 0.9851780533790588, 'rewards/repeat_floor/std': 0.006651438307017088, 'rewards/near_duplicate_penalty/mean': 0.966921329498291, 'rewards/near_duplicate_penalty/std': 0.04215977340936661, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9737762212753296, 'rewards/distinct_2/std': 0.05202964320778847, 'rewards/total_composite/mean': 0.17168530821800232, 'rewards/total_composite/std': 0.11954525113105774, 'reward': 0.17168530821800232, 'reward_std': 0.11954524368047714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17270267009735107, 'sampling/sampling_logp_difference/max': 1.880910873413086, 'sampling/importance_sampling_ratio/min': 0.15245117247104645, 'sampling/importance_sampling_ratio/mean': 0.9874579310417175, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8343949541449547, 'clip_ratio/low_mean': 0.09078109497204423, 'clip_ratio/low_min': 0.09078109497204423, 'clip_ratio/high_mean': 0.06691176537424326, 'clip_ratio/high_max': 0.06691176537424326, 'clip_ratio/region_mean': 0.1576928603462875, 'reward_total_mean': 0.17168530821800232, 'reward_meter_mean': 0.6338993310928345, 'reward_meter_std': 0.26156964898109436, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.16035674512386322, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8179764747619629, 'reward_repeat_penalty_std': 0.09199567139148712, 'reward_repeat_floor_mean': 0.9851780533790588, 'reward_repeat_floor_std': 0.006651438307017088, 'reward_near_duplicate_penalty_mean': 0.966921329498291, 'reward_near_duplicate_penalty_std': 0.04215977340936661, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9737762212753296, 'reward_distinct_2_std': 0.05202964320778847, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.17168530821800232, 'reward_total_composite_std': 0.11954525113105774, 'epoch': 0.06}

 33%|███▎      | 1082/3300 [4:07:40<7:39:22, 12.43s/it]INFO 04-12 20:06:07 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1083/3300 [4:07:54<7:52:25, 12.79s/it]2026-04-12 20:06:20,402 | INFO | train_grpo_train | metrics_logged mode=train step=1083

                                                       
{'loss': -0.0892, 'grad_norm': 3.384092092514038, 'learning_rate': 6.721212121212122e-06, 'num_tokens': 1931867.0, 'completions/mean_length': 193.0, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 86.66667175292969, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.6980332136154175, 'rewards/meter/std': 0.3264663815498352, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.07954952120780945, 'rewards/lexical_plausibility/mean': 0.9041027426719666, 'rewards/lexical_plausibility/std': 0.1783706098794937, 'rewards/judge_quality/mean': 0.14375001192092896, 'rewards/judge_quality/std': 0.10155048221349716, 'rewards/repeat_penalty/mean': 0.6313503980636597, 'rewards/repeat_penalty/std': 0.2825292646884918, 'rewards/repeat_floor/mean': 0.9609284400939941, 'rewards/repeat_floor/std': 0.0327136367559433, 'rewards/near_duplicate_penalty/mean': 0.9083307385444641, 'rewards/near_duplicate_penalty/std': 0.07224540412425995, 'rewards/opening_diversity/mean': 0.8671875, 'rewards/opening_diversity/std': 0.16173411905765533, 'rewards/distinct_2/mean': 0.8071950078010559, 'rewards/distinct_2/std': 0.215913325548172, 'rewards/total_composite/mean': 0.11442960798740387, 'rewards/total_composite/std': 0.09526680409908295, 'reward': 0.11442960798740387, 'reward_std': 0.09526679664850235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13939079642295837, 'sampling/sampling_logp_difference/max': 2.5638062953948975, 'sampling/importance_sampling_ratio/min': 0.07701105624437332, 'sampling/importance_sampling_ratio/mean': 1.029281497001648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7014232128858566, 'clip_ratio/low_mean': 0.034225694835186005, 'clip_ratio/low_min': 0.034225694835186005, 'clip_ratio/high_mean': 0.05332733877003193, 'clip_ratio/high_max': 0.05332733877003193, 'clip_ratio/region_mean': 0.08755303360521793, 'reward_total_mean': 0.11442960798740387, 'reward_meter_mean': 0.6980332136154175, 'reward_meter_std': 0.3264663815498352, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.2651650309562683, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.07954952120780945, 'reward_lexical_plausibility_mean': 0.9041027426719666, 'reward_lexical_plausibility_std': 0.1783706098794937, 'reward_repeat_penalty_mean': 0.6313503980636597, 'reward_repeat_penalty_std': 0.2825292646884918, 'reward_repeat_floor_mean': 0.9609284400939941, 'reward_repeat_floor_std': 0.0327136367559433, 'reward_near_duplicate_penalty_mean': 0.9083307385444641, 'reward_near_duplicate_penalty_std': 0.07224540412425995, 'reward_opening_diversity_mean': 0.8671875, 'reward_opening_diversity_std': 0.16173411905765533, 'reward_distinct_2_mean': 0.8071950078010559, 'reward_distinct_2_std': 0.215913325548172, 'reward_judge_quality_mean': 0.14375001192092896, 'reward_judge_quality_std': 0.10155048221349716, 'reward_total_composite_mean': 0.11442960798740387, 'reward_total_composite_std': 0.09526680409908295, 'epoch': 0.06}

 33%|███▎      | 1083/3300 [4:07:54<7:52:25, 12.79s/it]INFO 04-12 20:06:20 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1084/3300 [4:08:10<8:33:06, 13.89s/it]2026-04-12 20:06:36,930 | INFO | train_grpo_train | metrics_logged mode=train step=1084

                                                       
{'loss': -0.0553, 'grad_norm': 1.7104077339172363, 'learning_rate': 6.718181818181819e-06, 'num_tokens': 1933649.0, 'completions/mean_length': 166.75, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.66666793823242, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.8004549145698547, 'rewards/meter/std': 0.2955181896686554, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9153865575790405, 'rewards/lexical_plausibility/std': 0.1567201316356659, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.444543719291687, 'rewards/repeat_penalty/std': 0.3224257826805115, 'rewards/repeat_floor/mean': 0.9293628334999084, 'rewards/repeat_floor/std': 0.046146344393491745, 'rewards/near_duplicate_penalty/mean': 0.8347088098526001, 'rewards/near_duplicate_penalty/std': 0.10775653272867203, 'rewards/opening_diversity/mean': 0.6526786088943481, 'rewards/opening_diversity/std': 0.2602613568305969, 'rewards/distinct_2/mean': 0.7221752405166626, 'rewards/distinct_2/std': 0.2029658555984497, 'rewards/total_composite/mean': 0.26727473735809326, 'rewards/total_composite/std': 0.16430865228176117, 'reward': 0.26727473735809326, 'reward_std': 0.16430865228176117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10477927327156067, 'sampling/sampling_logp_difference/max': 1.3335072994232178, 'sampling/importance_sampling_ratio/min': 0.2635512948036194, 'sampling/importance_sampling_ratio/mean': 1.0224661827087402, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6080252602696419, 'clip_ratio/low_mean': 0.014880952425301075, 'clip_ratio/low_min': 0.014880952425301075, 'clip_ratio/high_mean': 0.09854944609105587, 'clip_ratio/high_max': 0.09854944609105587, 'clip_ratio/region_mean': 0.11343039851635695, 'reward_total_mean': 0.26727473735809326, 'reward_meter_mean': 0.8004549145698547, 'reward_meter_std': 0.2955181896686554, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9153865575790405, 'reward_lexical_plausibility_std': 0.1567201316356659, 'reward_repeat_penalty_mean': 0.444543719291687, 'reward_repeat_penalty_std': 0.3224257826805115, 'reward_repeat_floor_mean': 0.9293628334999084, 'reward_repeat_floor_std': 0.046146344393491745, 'reward_near_duplicate_penalty_mean': 0.8347088098526001, 'reward_near_duplicate_penalty_std': 0.10775653272867203, 'reward_opening_diversity_mean': 0.6526786088943481, 'reward_opening_diversity_std': 0.2602613568305969, 'reward_distinct_2_mean': 0.7221752405166626, 'reward_distinct_2_std': 0.2029658555984497, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.26727473735809326, 'reward_total_composite_std': 0.16430865228176117, 'epoch': 0.06}

 33%|███▎      | 1084/3300 [4:08:10<8:33:06, 13.89s/it]INFO 04-12 20:06:37 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1085/3300 [4:08:25<8:39:33, 14.07s/it]2026-04-12 20:06:51,379 | INFO | train_grpo_train | metrics_logged mode=train step=1085

                                                       
{'loss': -0.034, 'grad_norm': 5.871991157531738, 'learning_rate': 6.715151515151516e-06, 'num_tokens': 1935223.0, 'completions/mean_length': 105.75, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.71428680419922, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7791829109191895, 'rewards/meter/std': 0.33099138736724854, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9680292010307312, 'rewards/lexical_plausibility/std': 0.09042705595493317, 'rewards/judge_quality/mean': 0.7512499690055847, 'rewards/judge_quality/std': 0.349834144115448, 'rewards/repeat_penalty/mean': 0.9364655017852783, 'rewards/repeat_penalty/std': 0.1413634568452835, 'rewards/repeat_floor/mean': 0.9951254725456238, 'rewards/repeat_floor/std': 0.009855365380644798, 'rewards/near_duplicate_penalty/mean': 0.988947868347168, 'rewards/near_duplicate_penalty/std': 0.02277790755033493, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9642590284347534, 'rewards/distinct_2/std': 0.07363905012607574, 'rewards/total_composite/mean': 0.593836784362793, 'rewards/total_composite/std': 0.40051841735839844, 'reward': 0.593836784362793, 'reward_std': 0.40051838755607605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13935181498527527, 'sampling/sampling_logp_difference/max': 1.3706169128417969, 'sampling/importance_sampling_ratio/min': 0.25395023822784424, 'sampling/importance_sampling_ratio/mean': 1.0205938816070557, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7359886914491653, 'clip_ratio/low_mean': 0.036458333022892475, 'clip_ratio/low_min': 0.036458333022892475, 'clip_ratio/high_mean': 0.07284729508683085, 'clip_ratio/high_max': 0.07284729508683085, 'clip_ratio/region_mean': 0.10930562810972333, 'reward_total_mean': 0.593836784362793, 'reward_meter_mean': 0.7791829109191895, 'reward_meter_std': 0.33099138736724854, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9680292010307312, 'reward_lexical_plausibility_std': 0.09042705595493317, 'reward_repeat_penalty_mean': 0.9364655017852783, 'reward_repeat_penalty_std': 0.1413634568452835, 'reward_repeat_floor_mean': 0.9951254725456238, 'reward_repeat_floor_std': 0.009855365380644798, 'reward_near_duplicate_penalty_mean': 0.988947868347168, 'reward_near_duplicate_penalty_std': 0.02277790755033493, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9642590284347534, 'reward_distinct_2_std': 0.07363905012607574, 'reward_judge_quality_mean': 0.7512499690055847, 'reward_judge_quality_std': 0.349834144115448, 'reward_total_composite_mean': 0.593836784362793, 'reward_total_composite_std': 0.40051841735839844, 'epoch': 0.06}

 33%|███▎      | 1085/3300 [4:08:25<8:39:33, 14.07s/it]INFO 04-12 20:06:51 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1086/3300 [4:08:34<7:47:17, 12.66s/it]2026-04-12 20:07:00,748 | INFO | train_grpo_train | metrics_logged mode=train step=1086

                                                       
{'loss': 0.0465, 'grad_norm': 10.681161880493164, 'learning_rate': 6.712121212121213e-06, 'num_tokens': 1936944.0, 'completions/mean_length': 53.125, 'completions/min_length': 44.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9139430522918701, 'rewards/meter/std': 0.1263270080089569, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.24136148393154144, 'rewards/repeat_penalty/mean': 0.8210009932518005, 'rewards/repeat_penalty/std': 0.13047564029693604, 'rewards/repeat_floor/mean': 0.9796181917190552, 'rewards/repeat_floor/std': 0.012828967534005642, 'rewards/near_duplicate_penalty/mean': 0.9301998615264893, 'rewards/near_duplicate_penalty/std': 0.04048642888665199, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9107024669647217, 'rewards/distinct_2/std': 0.10493715852499008, 'rewards/total_composite/mean': 0.3343813121318817, 'rewards/total_composite/std': 0.2343132346868515, 'reward': 0.3343813121318817, 'reward_std': 0.2343132346868515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12319741398096085, 'sampling/sampling_logp_difference/max': 1.5332574844360352, 'sampling/importance_sampling_ratio/min': 0.215831458568573, 'sampling/importance_sampling_ratio/mean': 1.0099456310272217, 'sampling/importance_sampling_ratio/max': 1.9503875970840454, 'entropy': 0.7286128997802734, 'clip_ratio/low_mean': 0.07667247531935573, 'clip_ratio/low_min': 0.07667247531935573, 'clip_ratio/high_mean': 0.03490259870886803, 'clip_ratio/high_max': 0.03490259870886803, 'clip_ratio/region_mean': 0.11157507402822375, 'reward_total_mean': 0.3343813121318817, 'reward_meter_mean': 0.9139430522918701, 'reward_meter_std': 0.1263270080089569, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8210009932518005, 'reward_repeat_penalty_std': 0.13047564029693604, 'reward_repeat_floor_mean': 0.9796181917190552, 'reward_repeat_floor_std': 0.012828967534005642, 'reward_near_duplicate_penalty_mean': 0.9301998615264893, 'reward_near_duplicate_penalty_std': 0.04048642888665199, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9107024669647217, 'reward_distinct_2_std': 0.10493715852499008, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.24136148393154144, 'reward_total_composite_mean': 0.3343813121318817, 'reward_total_composite_std': 0.2343132346868515, 'epoch': 0.06}

 33%|███▎      | 1086/3300 [4:08:34<7:47:17, 12.66s/it]INFO 04-12 20:07:00 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1087/3300 [4:08:49<8:07:25, 13.22s/it]2026-04-12 20:07:15,251 | INFO | train_grpo_train | metrics_logged mode=train step=1087

                                                       
{'loss': -0.0599, 'grad_norm': 1.916584849357605, 'learning_rate': 6.709090909090909e-06, 'num_tokens': 1938556.0, 'completions/mean_length': 288.5, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.6596055030822754, 'rewards/meter/std': 0.37892818450927734, 'rewards/count_adherence/mean': 0.4583333730697632, 'rewards/count_adherence/std': 0.43415674567222595, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8374999761581421, 'rewards/count_floor/std': 0.13024702668190002, 'rewards/lexical_plausibility/mean': 0.8312267065048218, 'rewards/lexical_plausibility/std': 0.18548545241355896, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.7731016874313354, 'rewards/repeat_penalty/std': 0.07494602352380753, 'rewards/repeat_floor/mean': 0.9738824367523193, 'rewards/repeat_floor/std': 0.00962059572339058, 'rewards/near_duplicate_penalty/mean': 0.8970316648483276, 'rewards/near_duplicate_penalty/std': 0.03821737319231033, 'rewards/opening_diversity/mean': 0.9963235259056091, 'rewards/opening_diversity/std': 0.010398639366030693, 'rewards/distinct_2/mean': 0.8838698863983154, 'rewards/distinct_2/std': 0.07145090401172638, 'rewards/total_composite/mean': 0.16199465095996857, 'rewards/total_composite/std': 0.16370822489261627, 'reward': 0.16199465095996857, 'reward_std': 0.16370822489261627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1485721617937088, 'sampling/sampling_logp_difference/max': 2.2193174362182617, 'sampling/importance_sampling_ratio/min': 0.1086832657456398, 'sampling/importance_sampling_ratio/mean': 1.0044283866882324, 'sampling/importance_sampling_ratio/max': 1.9098730087280273, 'entropy': 0.4442877247929573, 'clip_ratio/low_mean': 0.013719512149691582, 'clip_ratio/low_min': 0.013719512149691582, 'clip_ratio/high_mean': 0.060597267001867294, 'clip_ratio/high_max': 0.060597267001867294, 'clip_ratio/region_mean': 0.07431677915155888, 'reward_total_mean': 0.16199465095996857, 'reward_meter_mean': 0.6596055030822754, 'reward_meter_std': 0.37892818450927734, 'reward_count_adherence_mean': 0.4583333730697632, 'reward_count_adherence_std': 0.43415674567222595, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8374999761581421, 'reward_count_floor_std': 0.13024702668190002, 'reward_lexical_plausibility_mean': 0.8312267065048218, 'reward_lexical_plausibility_std': 0.18548545241355896, 'reward_repeat_penalty_mean': 0.7731016874313354, 'reward_repeat_penalty_std': 0.07494602352380753, 'reward_repeat_floor_mean': 0.9738824367523193, 'reward_repeat_floor_std': 0.00962059572339058, 'reward_near_duplicate_penalty_mean': 0.8970316648483276, 'reward_near_duplicate_penalty_std': 0.03821737319231033, 'reward_opening_diversity_mean': 0.9963235259056091, 'reward_opening_diversity_std': 0.010398639366030693, 'reward_distinct_2_mean': 0.8838698863983154, 'reward_distinct_2_std': 0.07145090401172638, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.16199465095996857, 'reward_total_composite_std': 0.16370822489261627, 'epoch': 0.06}

 33%|███▎      | 1087/3300 [4:08:49<8:07:25, 13.22s/it]INFO 04-12 20:07:15 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1088/3300 [4:08:57<7:16:25, 11.84s/it]2026-04-12 20:07:23,879 | INFO | train_grpo_train | metrics_logged mode=train step=1088

                                                       
{'loss': 0.2228, 'grad_norm': 8.618753433227539, 'learning_rate': 6.706060606060607e-06, 'num_tokens': 1940104.0, 'completions/mean_length': 36.5, 'completions/min_length': 21.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.4850292205810547, 'rewards/meter/std': 0.3955121338367462, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.16035674512386322, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.7875106930732727, 'rewards/repeat_penalty/std': 0.16636258363723755, 'rewards/repeat_floor/mean': 0.9793968796730042, 'rewards/repeat_floor/std': 0.019574204459786415, 'rewards/near_duplicate_penalty/mean': 0.9421905279159546, 'rewards/near_duplicate_penalty/std': 0.09153043478727341, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9507588148117065, 'rewards/distinct_2/std': 0.05842531472444534, 'rewards/total_composite/mean': 0.2098846137523651, 'rewards/total_composite/std': 0.27752968668937683, 'reward': 0.2098846137523651, 'reward_std': 0.27752965688705444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1160566508769989, 'sampling/sampling_logp_difference/max': 1.4186387062072754, 'sampling/importance_sampling_ratio/min': 0.3461836874485016, 'sampling/importance_sampling_ratio/mean': 1.0321719646453857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8647604696452618, 'clip_ratio/low_mean': 0.09415545128285885, 'clip_ratio/low_min': 0.09415545128285885, 'clip_ratio/high_mean': 0.03977272845804691, 'clip_ratio/high_max': 0.03977272845804691, 'clip_ratio/region_mean': 0.13392817974090576, 'reward_total_mean': 0.2098846137523651, 'reward_meter_mean': 0.4850292205810547, 'reward_meter_std': 0.3955121338367462, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.16035674512386322, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7875106930732727, 'reward_repeat_penalty_std': 0.16636258363723755, 'reward_repeat_floor_mean': 0.9793968796730042, 'reward_repeat_floor_std': 0.019574204459786415, 'reward_near_duplicate_penalty_mean': 0.9421905279159546, 'reward_near_duplicate_penalty_std': 0.09153043478727341, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9507588148117065, 'reward_distinct_2_std': 0.05842531472444534, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.2098846137523651, 'reward_total_composite_std': 0.27752968668937683, 'epoch': 0.06}

 33%|███▎      | 1088/3300 [4:08:57<7:16:25, 11.84s/it]INFO 04-12 20:07:24 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1089/3300 [4:09:11<7:35:42, 12.37s/it]2026-04-12 20:07:37,476 | INFO | train_grpo_train | metrics_logged mode=train step=1089

                                                       
{'loss': -0.0509, 'grad_norm': 2.8513529300689697, 'learning_rate': 6.703030303030304e-06, 'num_tokens': 1941709.0, 'completions/mean_length': 161.625, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.833335876464844, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6778078079223633, 'rewards/meter/std': 0.3915116786956787, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9214867353439331, 'rewards/lexical_plausibility/std': 0.1500946283340454, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.27458736300468445, 'rewards/repeat_penalty/mean': 0.8548451662063599, 'rewards/repeat_penalty/std': 0.1594332903623581, 'rewards/repeat_floor/mean': 0.9812458157539368, 'rewards/repeat_floor/std': 0.01976749859750271, 'rewards/near_duplicate_penalty/mean': 0.9221019148826599, 'rewards/near_duplicate_penalty/std': 0.06468543410301208, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9214506149291992, 'rewards/distinct_2/std': 0.1300339698791504, 'rewards/total_composite/mean': 0.28842759132385254, 'rewards/total_composite/std': 0.28460586071014404, 'reward': 0.28842759132385254, 'reward_std': 0.28460586071014404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13578085601329803, 'sampling/sampling_logp_difference/max': 1.5256547927856445, 'sampling/importance_sampling_ratio/min': 0.21747861802577972, 'sampling/importance_sampling_ratio/mean': 1.033605694770813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7600734606385231, 'clip_ratio/low_mean': 0.0461454764008522, 'clip_ratio/low_min': 0.0461454764008522, 'clip_ratio/high_mean': 0.06600877456367016, 'clip_ratio/high_max': 0.06600877456367016, 'clip_ratio/region_mean': 0.11215425096452236, 'reward_total_mean': 0.28842759132385254, 'reward_meter_mean': 0.6778078079223633, 'reward_meter_std': 0.3915116786956787, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9214867353439331, 'reward_lexical_plausibility_std': 0.1500946283340454, 'reward_repeat_penalty_mean': 0.8548451662063599, 'reward_repeat_penalty_std': 0.1594332903623581, 'reward_repeat_floor_mean': 0.9812458157539368, 'reward_repeat_floor_std': 0.01976749859750271, 'reward_near_duplicate_penalty_mean': 0.9221019148826599, 'reward_near_duplicate_penalty_std': 0.06468543410301208, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9214506149291992, 'reward_distinct_2_std': 0.1300339698791504, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.27458736300468445, 'reward_total_composite_mean': 0.28842759132385254, 'reward_total_composite_std': 0.28460586071014404, 'epoch': 0.06}

 33%|███▎      | 1089/3300 [4:09:11<7:35:42, 12.37s/it]INFO 04-12 20:07:37 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1090/3300 [4:09:19<6:51:53, 11.18s/it]2026-04-12 20:07:45,896 | INFO | train_grpo_train | metrics_logged mode=train step=1090

                                                       
{'loss': 0.2498, 'grad_norm': 20.269132614135742, 'learning_rate': 6.700000000000001e-06, 'num_tokens': 1943162.0, 'completions/mean_length': 22.625, 'completions/min_length': 17.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7608385682106018, 'rewards/meter/std': 0.40729326009750366, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.7468118667602539, 'rewards/repeat_penalty/std': 0.09663929045200348, 'rewards/repeat_floor/mean': 0.9802461862564087, 'rewards/repeat_floor/std': 0.010517552495002747, 'rewards/near_duplicate_penalty/mean': 0.9558081030845642, 'rewards/near_duplicate_penalty/std': 0.059891778975725174, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24426288902759552, 'rewards/total_composite/std': 0.18034961819648743, 'reward': 0.24426288902759552, 'reward_std': 0.18034961819648743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11935863643884659, 'sampling/sampling_logp_difference/max': 1.0952835083007812, 'sampling/importance_sampling_ratio/min': 0.3344447612762451, 'sampling/importance_sampling_ratio/mean': 1.0266389846801758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.780202716588974, 'clip_ratio/low_mean': 0.07919938769191504, 'clip_ratio/low_min': 0.07919938769191504, 'clip_ratio/high_mean': 0.033010836225003004, 'clip_ratio/high_max': 0.033010836225003004, 'clip_ratio/region_mean': 0.11221022391691804, 'reward_total_mean': 0.24426288902759552, 'reward_meter_mean': 0.7608385682106018, 'reward_meter_std': 0.40729326009750366, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7468118667602539, 'reward_repeat_penalty_std': 0.09663929045200348, 'reward_repeat_floor_mean': 0.9802461862564087, 'reward_repeat_floor_std': 0.010517552495002747, 'reward_near_duplicate_penalty_mean': 0.9558081030845642, 'reward_near_duplicate_penalty_std': 0.059891778975725174, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.24426288902759552, 'reward_total_composite_std': 0.18034961819648743, 'epoch': 0.06}

 33%|███▎      | 1090/3300 [4:09:19<6:51:53, 11.18s/it]INFO 04-12 20:07:46 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1091/3300 [4:09:33<7:17:31, 11.88s/it]2026-04-12 20:07:59,437 | INFO | train_grpo_train | metrics_logged mode=train step=1091

                                                       
{'loss': -0.0313, 'grad_norm': 3.0998330116271973, 'learning_rate': 6.6969696969696975e-06, 'num_tokens': 1944676.0, 'completions/mean_length': 161.25, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.333335876464844, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.4616873562335968, 'rewards/meter/std': 0.38850751519203186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.919589102268219, 'rewards/lexical_plausibility/std': 0.15135924518108368, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.2295181304216385, 'rewards/repeat_penalty/mean': 0.7323417067527771, 'rewards/repeat_penalty/std': 0.1445988267660141, 'rewards/repeat_floor/mean': 0.9709863662719727, 'rewards/repeat_floor/std': 0.014883177354931831, 'rewards/near_duplicate_penalty/mean': 0.9120395183563232, 'rewards/near_duplicate_penalty/std': 0.04725426435470581, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8867272138595581, 'rewards/distinct_2/std': 0.10701995342969894, 'rewards/total_composite/mean': 0.08975566923618317, 'rewards/total_composite/std': 0.11220458894968033, 'reward': 0.08975566923618317, 'reward_std': 0.11220459640026093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14262858033180237, 'sampling/sampling_logp_difference/max': 1.5645205974578857, 'sampling/importance_sampling_ratio/min': 0.20918826758861542, 'sampling/importance_sampling_ratio/mean': 1.0364091396331787, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7283241897821426, 'clip_ratio/low_mean': 0.054470415227115154, 'clip_ratio/low_min': 0.054470415227115154, 'clip_ratio/high_mean': 0.04826454073190689, 'clip_ratio/high_max': 0.04826454073190689, 'clip_ratio/region_mean': 0.10273495595902205, 'reward_total_mean': 0.08975566923618317, 'reward_meter_mean': 0.4616873562335968, 'reward_meter_std': 0.38850751519203186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.919589102268219, 'reward_lexical_plausibility_std': 0.15135924518108368, 'reward_repeat_penalty_mean': 0.7323417067527771, 'reward_repeat_penalty_std': 0.1445988267660141, 'reward_repeat_floor_mean': 0.9709863662719727, 'reward_repeat_floor_std': 0.014883177354931831, 'reward_near_duplicate_penalty_mean': 0.9120395183563232, 'reward_near_duplicate_penalty_std': 0.04725426435470581, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8867272138595581, 'reward_distinct_2_std': 0.10701995342969894, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.2295181304216385, 'reward_total_composite_mean': 0.08975566923618317, 'reward_total_composite_std': 0.11220458894968033, 'epoch': 0.06}

 33%|███▎      | 1091/3300 [4:09:33<7:17:31, 11.88s/it]INFO 04-12 20:07:59 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1092/3300 [4:09:46<7:37:49, 12.44s/it]2026-04-12 20:08:13,160 | INFO | train_grpo_train | metrics_logged mode=train step=1092

                                                       
{'loss': -0.0939, 'grad_norm': 1.647580623626709, 'learning_rate': 6.693939393939395e-06, 'num_tokens': 1946178.0, 'completions/mean_length': 154.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.66666793823242, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7439446449279785, 'rewards/meter/std': 0.39357125759124756, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.8957195281982422, 'rewards/lexical_plausibility/std': 0.1949753612279892, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.1752549111843109, 'rewards/repeat_penalty/mean': 0.8643711805343628, 'rewards/repeat_penalty/std': 0.17102113366127014, 'rewards/repeat_floor/mean': 0.9866104125976562, 'rewards/repeat_floor/std': 0.017543351277709007, 'rewards/near_duplicate_penalty/mean': 0.9631621837615967, 'rewards/near_duplicate_penalty/std': 0.0577021948993206, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.9626235961914062, 'rewards/distinct_2/std': 0.06939946115016937, 'rewards/total_composite/mean': 0.2858754098415375, 'rewards/total_composite/std': 0.17618735134601593, 'reward': 0.2858754098415375, 'reward_std': 0.17618735134601593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13028892874717712, 'sampling/sampling_logp_difference/max': 0.8585249185562134, 'sampling/importance_sampling_ratio/min': 0.49698859453201294, 'sampling/importance_sampling_ratio/mean': 1.020411491394043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6634023115038872, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11563093308359385, 'clip_ratio/high_max': 0.11563093308359385, 'clip_ratio/region_mean': 0.11563093308359385, 'reward_total_mean': 0.2858754098415375, 'reward_meter_mean': 0.7439446449279785, 'reward_meter_std': 0.39357125759124756, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.8957195281982422, 'reward_lexical_plausibility_std': 0.1949753612279892, 'reward_repeat_penalty_mean': 0.8643711805343628, 'reward_repeat_penalty_std': 0.17102113366127014, 'reward_repeat_floor_mean': 0.9866104125976562, 'reward_repeat_floor_std': 0.017543351277709007, 'reward_near_duplicate_penalty_mean': 0.9631621837615967, 'reward_near_duplicate_penalty_std': 0.0577021948993206, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.9626235961914062, 'reward_distinct_2_std': 0.06939946115016937, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.1752549111843109, 'reward_total_composite_mean': 0.2858754098415375, 'reward_total_composite_std': 0.17618735134601593, 'epoch': 0.06}

 33%|███▎      | 1092/3300 [4:09:47<7:37:49, 12.44s/it]INFO 04-12 20:08:13 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1093/3300 [4:09:56<7:06:50, 11.60s/it]2026-04-12 20:08:22,807 | INFO | train_grpo_train | metrics_logged mode=train step=1093

                                                       
{'loss': 0.012, 'grad_norm': 9.187946319580078, 'learning_rate': 6.690909090909091e-06, 'num_tokens': 1948294.0, 'completions/mean_length': 70.5, 'completions/min_length': 62.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9771513938903809, 'rewards/meter/std': 0.03255423158407211, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.6503579616546631, 'rewards/repeat_penalty/std': 0.15134984254837036, 'rewards/repeat_floor/mean': 0.9553418159484863, 'rewards/repeat_floor/std': 0.02226443402469158, 'rewards/near_duplicate_penalty/mean': 0.8459196090698242, 'rewards/near_duplicate_penalty/std': 0.07903697341680527, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.7814317345619202, 'rewards/distinct_2/std': 0.08685024827718735, 'rewards/total_composite/mean': 0.19522690773010254, 'rewards/total_composite/std': 0.09244982898235321, 'reward': 0.19522690773010254, 'reward_std': 0.09244982153177261, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1326293647289276, 'sampling/sampling_logp_difference/max': 1.7948389053344727, 'sampling/importance_sampling_ratio/min': 0.16615422070026398, 'sampling/importance_sampling_ratio/mean': 1.0035054683685303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7150585576891899, 'clip_ratio/low_mean': 0.07776742801070213, 'clip_ratio/low_min': 0.07776742801070213, 'clip_ratio/high_mean': 0.05006090085953474, 'clip_ratio/high_max': 0.05006090085953474, 'clip_ratio/region_mean': 0.12782832887023687, 'reward_total_mean': 0.19522690773010254, 'reward_meter_mean': 0.9771513938903809, 'reward_meter_std': 0.03255423158407211, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6503579616546631, 'reward_repeat_penalty_std': 0.15134984254837036, 'reward_repeat_floor_mean': 0.9553418159484863, 'reward_repeat_floor_std': 0.02226443402469158, 'reward_near_duplicate_penalty_mean': 0.8459196090698242, 'reward_near_duplicate_penalty_std': 0.07903697341680527, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.7814317345619202, 'reward_distinct_2_std': 0.08685024827718735, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.19522690773010254, 'reward_total_composite_std': 0.09244982898235321, 'epoch': 0.06}

 33%|███▎      | 1093/3300 [4:09:56<7:06:50, 11.60s/it]INFO 04-12 20:08:23 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1094/3300 [4:10:05<6:33:28, 10.70s/it]2026-04-12 20:08:31,477 | INFO | train_grpo_train | metrics_logged mode=train step=1094

                                                       
{'loss': 0.0248, 'grad_norm': 21.760711669921875, 'learning_rate': 6.687878787878788e-06, 'num_tokens': 1949525.0, 'completions/mean_length': 22.875, 'completions/min_length': 20.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9305969476699829, 'rewards/meter/std': 0.12035281956195831, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5762500166893005, 'rewards/judge_quality/std': 0.30123260617256165, 'rewards/repeat_penalty/mean': 0.5536414980888367, 'rewards/repeat_penalty/std': 0.27665838599205017, 'rewards/repeat_floor/mean': 0.9421765208244324, 'rewards/repeat_floor/std': 0.0637635737657547, 'rewards/near_duplicate_penalty/mean': 0.802559494972229, 'rewards/near_duplicate_penalty/std': 0.2574443221092224, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.9157509207725525, 'rewards/distinct_2/std': 0.14504754543304443, 'rewards/total_composite/mean': 0.49066489934921265, 'rewards/total_composite/std': 0.24006322026252747, 'reward': 0.49066489934921265, 'reward_std': 0.24006320536136627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.194910928606987, 'sampling/sampling_logp_difference/max': 1.578312873840332, 'sampling/importance_sampling_ratio/min': 0.20632289350032806, 'sampling/importance_sampling_ratio/mean': 1.0179365873336792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0782008990645409, 'clip_ratio/low_mean': 0.09517195029184222, 'clip_ratio/low_min': 0.09517195029184222, 'clip_ratio/high_mean': 0.07516678236424923, 'clip_ratio/high_max': 0.07516678236424923, 'clip_ratio/region_mean': 0.17033873265609145, 'reward_total_mean': 0.49066489934921265, 'reward_meter_mean': 0.9305969476699829, 'reward_meter_std': 0.12035281956195831, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5536414980888367, 'reward_repeat_penalty_std': 0.27665838599205017, 'reward_repeat_floor_mean': 0.9421765208244324, 'reward_repeat_floor_std': 0.0637635737657547, 'reward_near_duplicate_penalty_mean': 0.802559494972229, 'reward_near_duplicate_penalty_std': 0.2574443221092224, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.9157509207725525, 'reward_distinct_2_std': 0.14504754543304443, 'reward_judge_quality_mean': 0.5762500166893005, 'reward_judge_quality_std': 0.30123260617256165, 'reward_total_composite_mean': 0.49066489934921265, 'reward_total_composite_std': 0.24006322026252747, 'epoch': 0.06}

 33%|███▎      | 1094/3300 [4:10:05<6:33:28, 10.70s/it]INFO 04-12 20:08:31 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1095/3300 [4:10:19<7:16:55, 11.89s/it]2026-04-12 20:08:46,122 | INFO | train_grpo_train | metrics_logged mode=train step=1095

                                                       
{'loss': -0.0376, 'grad_norm': 1.7080044746398926, 'learning_rate': 6.684848484848485e-06, 'num_tokens': 1951137.0, 'completions/mean_length': 277.5, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.5442564487457275, 'rewards/meter/std': 0.46114057302474976, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.8344399929046631, 'rewards/lexical_plausibility/std': 0.17873318493366241, 'rewards/judge_quality/mean': 0.2462500035762787, 'rewards/judge_quality/std': 0.3020850121974945, 'rewards/repeat_penalty/mean': 0.9330055713653564, 'rewards/repeat_penalty/std': 0.07860389351844788, 'rewards/repeat_floor/mean': 0.9916747212409973, 'rewards/repeat_floor/std': 0.00870470143854618, 'rewards/near_duplicate_penalty/mean': 0.964820146560669, 'rewards/near_duplicate_penalty/std': 0.026550035923719406, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.966130256652832, 'rewards/distinct_2/std': 0.06408049166202545, 'rewards/total_composite/mean': 0.18429481983184814, 'rewards/total_composite/std': 0.2996910810470581, 'reward': 0.18429481983184814, 'reward_std': 0.2996910512447357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15668749809265137, 'sampling/sampling_logp_difference/max': 0.9845449924468994, 'sampling/importance_sampling_ratio/min': 0.41279104351997375, 'sampling/importance_sampling_ratio/mean': 1.015380859375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5497327595949173, 'clip_ratio/low_mean': 0.037097953259944916, 'clip_ratio/low_min': 0.037097953259944916, 'clip_ratio/high_mean': 0.05080128461122513, 'clip_ratio/high_max': 0.05080128461122513, 'clip_ratio/region_mean': 0.08789923787117004, 'reward_total_mean': 0.18429481983184814, 'reward_meter_mean': 0.5442564487457275, 'reward_meter_std': 0.46114057302474976, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.8344399929046631, 'reward_lexical_plausibility_std': 0.17873318493366241, 'reward_repeat_penalty_mean': 0.9330055713653564, 'reward_repeat_penalty_std': 0.07860389351844788, 'reward_repeat_floor_mean': 0.9916747212409973, 'reward_repeat_floor_std': 0.00870470143854618, 'reward_near_duplicate_penalty_mean': 0.964820146560669, 'reward_near_duplicate_penalty_std': 0.026550035923719406, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.966130256652832, 'reward_distinct_2_std': 0.06408049166202545, 'reward_judge_quality_mean': 0.2462500035762787, 'reward_judge_quality_std': 0.3020850121974945, 'reward_total_composite_mean': 0.18429481983184814, 'reward_total_composite_std': 0.2996910810470581, 'epoch': 0.06}

 33%|███▎      | 1095/3300 [4:10:19<7:16:55, 11.89s/it]INFO 04-12 20:08:46 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1096/3300 [4:10:33<7:35:09, 12.39s/it]2026-04-12 20:08:59,645 | INFO | train_grpo_train | metrics_logged mode=train step=1096

                                                       
{'loss': -0.0479, 'grad_norm': 4.381524562835693, 'learning_rate': 6.681818181818183e-06, 'num_tokens': 1952584.0, 'completions/mean_length': 89.875, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 29.571430206298828, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6610127687454224, 'rewards/meter/std': 0.40588122606277466, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.947071373462677, 'rewards/lexical_plausibility/std': 0.1497047245502472, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.8847966194152832, 'rewards/repeat_penalty/std': 0.1087074875831604, 'rewards/repeat_floor/mean': 0.987322211265564, 'rewards/repeat_floor/std': 0.010964552871882915, 'rewards/near_duplicate_penalty/mean': 0.9510585069656372, 'rewards/near_duplicate_penalty/std': 0.04364753141999245, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9615384340286255, 'rewards/distinct_2/std': 0.07631013542413712, 'rewards/total_composite/mean': 0.20927679538726807, 'rewards/total_composite/std': 0.1652289777994156, 'reward': 0.20927679538726807, 'reward_std': 0.1652289777994156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17333891987800598, 'sampling/sampling_logp_difference/max': 1.238534927368164, 'sampling/importance_sampling_ratio/min': 0.2898085117340088, 'sampling/importance_sampling_ratio/mean': 1.0018415451049805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9342399910092354, 'clip_ratio/low_mean': 0.06966517865657806, 'clip_ratio/low_min': 0.06966517865657806, 'clip_ratio/high_mean': 0.08052733587101102, 'clip_ratio/high_max': 0.08052733587101102, 'clip_ratio/region_mean': 0.15019251452758908, 'reward_total_mean': 0.20927679538726807, 'reward_meter_mean': 0.6610127687454224, 'reward_meter_std': 0.40588122606277466, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.947071373462677, 'reward_lexical_plausibility_std': 0.1497047245502472, 'reward_repeat_penalty_mean': 0.8847966194152832, 'reward_repeat_penalty_std': 0.1087074875831604, 'reward_repeat_floor_mean': 0.987322211265564, 'reward_repeat_floor_std': 0.010964552871882915, 'reward_near_duplicate_penalty_mean': 0.9510585069656372, 'reward_near_duplicate_penalty_std': 0.04364753141999245, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9615384340286255, 'reward_distinct_2_std': 0.07631013542413712, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.20927679538726807, 'reward_total_composite_std': 0.1652289777994156, 'epoch': 0.06}

 33%|███▎      | 1096/3300 [4:10:33<7:35:09, 12.39s/it]INFO 04-12 20:08:59 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1097/3300 [4:10:41<6:47:56, 11.11s/it]2026-04-12 20:09:07,748 | INFO | train_grpo_train | metrics_logged mode=train step=1097

                                                       
{'loss': 0.0129, 'grad_norm': 12.897970199584961, 'learning_rate': 6.678787878787879e-06, 'num_tokens': 1954179.0, 'completions/mean_length': 42.375, 'completions/min_length': 36.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8010610342025757, 'rewards/meter/std': 0.337760865688324, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1685018092393875, 'rewards/repeat_penalty/mean': 0.6276440620422363, 'rewards/repeat_penalty/std': 0.34671539068222046, 'rewards/repeat_floor/mean': 0.9534791707992554, 'rewards/repeat_floor/std': 0.044727250933647156, 'rewards/near_duplicate_penalty/mean': 0.8738264441490173, 'rewards/near_duplicate_penalty/std': 0.11950135231018066, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.2734534442424774, 'rewards/distinct_2/mean': 0.8392242193222046, 'rewards/distinct_2/std': 0.13986997306346893, 'rewards/total_composite/mean': 0.3464347720146179, 'rewards/total_composite/std': 0.22131280601024628, 'reward': 0.3464347720146179, 'reward_std': 0.22131283581256866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1366889476776123, 'sampling/sampling_logp_difference/max': 1.9211750030517578, 'sampling/importance_sampling_ratio/min': 0.14643479883670807, 'sampling/importance_sampling_ratio/mean': 0.99560546875, 'sampling/importance_sampling_ratio/max': 1.992817759513855, 'entropy': 0.7572882398962975, 'clip_ratio/low_mean': 0.05509947193786502, 'clip_ratio/low_min': 0.05509947193786502, 'clip_ratio/high_mean': 0.06942121032625437, 'clip_ratio/high_max': 0.06942121032625437, 'clip_ratio/region_mean': 0.12452068226411939, 'reward_total_mean': 0.3464347720146179, 'reward_meter_mean': 0.8010610342025757, 'reward_meter_std': 0.337760865688324, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6276440620422363, 'reward_repeat_penalty_std': 0.34671539068222046, 'reward_repeat_floor_mean': 0.9534791707992554, 'reward_repeat_floor_std': 0.044727250933647156, 'reward_near_duplicate_penalty_mean': 0.8738264441490173, 'reward_near_duplicate_penalty_std': 0.11950135231018066, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.2734534442424774, 'reward_distinct_2_mean': 0.8392242193222046, 'reward_distinct_2_std': 0.13986997306346893, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1685018092393875, 'reward_total_composite_mean': 0.3464347720146179, 'reward_total_composite_std': 0.22131280601024628, 'epoch': 0.06}

 33%|███▎      | 1097/3300 [4:10:41<6:47:56, 11.11s/it]INFO 04-12 20:09:07 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1098/3300 [4:10:56<7:32:28, 12.33s/it]2026-04-12 20:09:22,996 | INFO | train_grpo_train | metrics_logged mode=train step=1098

                                                       
{'loss': -0.1113, 'grad_norm': 4.186286926269531, 'learning_rate': 6.6757575757575766e-06, 'num_tokens': 1956353.0, 'completions/mean_length': 139.75, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.5714340209961, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.8713853359222412, 'rewards/meter/std': 0.3166533410549164, 'rewards/count_adherence/mean': 0.675000011920929, 'rewards/count_adherence/std': 0.14880476891994476, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9024999737739563, 'rewards/count_floor/std': 0.04464143142104149, 'rewards/lexical_plausibility/mean': 0.9481337666511536, 'rewards/lexical_plausibility/std': 0.14669986069202423, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.6134224534034729, 'rewards/repeat_penalty/std': 0.21901676058769226, 'rewards/repeat_floor/mean': 0.9548375606536865, 'rewards/repeat_floor/std': 0.031802453100681305, 'rewards/near_duplicate_penalty/mean': 0.8885675668716431, 'rewards/near_duplicate_penalty/std': 0.08633626997470856, 'rewards/opening_diversity/mean': 0.8567708730697632, 'rewards/opening_diversity/std': 0.1726735681295395, 'rewards/distinct_2/mean': 0.7794021964073181, 'rewards/distinct_2/std': 0.15259352326393127, 'rewards/total_composite/mean': 0.1775781661272049, 'rewards/total_composite/std': 0.11254829168319702, 'reward': 0.1775781661272049, 'reward_std': 0.11254829168319702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1395624428987503, 'sampling/sampling_logp_difference/max': 3.1949408054351807, 'sampling/importance_sampling_ratio/min': 0.04096895083785057, 'sampling/importance_sampling_ratio/mean': 1.015233039855957, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8223958015441895, 'clip_ratio/low_mean': 0.056410242803394794, 'clip_ratio/low_min': 0.056410242803394794, 'clip_ratio/high_mean': 0.06409796886146069, 'clip_ratio/high_max': 0.06409796886146069, 'clip_ratio/region_mean': 0.12050821166485548, 'reward_total_mean': 0.1775781661272049, 'reward_meter_mean': 0.8713853359222412, 'reward_meter_std': 0.3166533410549164, 'reward_count_adherence_mean': 0.675000011920929, 'reward_count_adherence_std': 0.14880476891994476, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9024999737739563, 'reward_count_floor_std': 0.04464143142104149, 'reward_lexical_plausibility_mean': 0.9481337666511536, 'reward_lexical_plausibility_std': 0.14669986069202423, 'reward_repeat_penalty_mean': 0.6134224534034729, 'reward_repeat_penalty_std': 0.21901676058769226, 'reward_repeat_floor_mean': 0.9548375606536865, 'reward_repeat_floor_std': 0.031802453100681305, 'reward_near_duplicate_penalty_mean': 0.8885675668716431, 'reward_near_duplicate_penalty_std': 0.08633626997470856, 'reward_opening_diversity_mean': 0.8567708730697632, 'reward_opening_diversity_std': 0.1726735681295395, 'reward_distinct_2_mean': 0.7794021964073181, 'reward_distinct_2_std': 0.15259352326393127, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.1775781661272049, 'reward_total_composite_std': 0.11254829168319702, 'epoch': 0.06}

 33%|███▎      | 1098/3300 [4:10:56<7:32:28, 12.33s/it]INFO 04-12 20:09:23 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1099/3300 [4:11:05<6:54:00, 11.29s/it]2026-04-12 20:09:31,803 | INFO | train_grpo_train | metrics_logged mode=train step=1099

                                                       
{'loss': 0.1291, 'grad_norm': 12.249665260314941, 'learning_rate': 6.672727272727273e-06, 'num_tokens': 1958081.0, 'completions/mean_length': 41.0, 'completions/min_length': 32.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5705949664115906, 'rewards/meter/std': 0.36392638087272644, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.19776880741119385, 'rewards/repeat_penalty/mean': 0.8820878267288208, 'rewards/repeat_penalty/std': 0.10153698921203613, 'rewards/repeat_floor/mean': 0.985323965549469, 'rewards/repeat_floor/std': 0.01125027984380722, 'rewards/near_duplicate_penalty/mean': 0.9399466514587402, 'rewards/near_duplicate_penalty/std': 0.03149586170911789, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9370220303535461, 'rewards/distinct_2/std': 0.08818355947732925, 'rewards/total_composite/mean': 0.2716326117515564, 'rewards/total_composite/std': 0.21279367804527283, 'reward': 0.2716326117515564, 'reward_std': 0.21279366314411163, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15643452107906342, 'sampling/sampling_logp_difference/max': 4.454733848571777, 'sampling/importance_sampling_ratio/min': 0.011623413302004337, 'sampling/importance_sampling_ratio/mean': 1.0151728391647339, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8616652265191078, 'clip_ratio/low_mean': 0.09151867032051086, 'clip_ratio/low_min': 0.09151867032051086, 'clip_ratio/high_mean': 0.0582217276096344, 'clip_ratio/high_max': 0.0582217276096344, 'clip_ratio/region_mean': 0.14974039793014526, 'reward_total_mean': 0.2716326117515564, 'reward_meter_mean': 0.5705949664115906, 'reward_meter_std': 0.36392638087272644, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8820878267288208, 'reward_repeat_penalty_std': 0.10153698921203613, 'reward_repeat_floor_mean': 0.985323965549469, 'reward_repeat_floor_std': 0.01125027984380722, 'reward_near_duplicate_penalty_mean': 0.9399466514587402, 'reward_near_duplicate_penalty_std': 0.03149586170911789, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9370220303535461, 'reward_distinct_2_std': 0.08818355947732925, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.19776880741119385, 'reward_total_composite_mean': 0.2716326117515564, 'reward_total_composite_std': 0.21279367804527283, 'epoch': 0.06}

 33%|███▎      | 1099/3300 [4:11:05<6:54:00, 11.29s/it]INFO 04-12 20:09:32 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1100/3300 [4:11:15<6:41:31, 10.95s/it]2026-04-12 20:09:41,966 | INFO | train_grpo_train | metrics_logged mode=train step=1100

                                                       
{'loss': 0.0608, 'grad_norm': 14.894866943359375, 'learning_rate': 6.66969696969697e-06, 'num_tokens': 1960543.0, 'completions/mean_length': 109.75, 'completions/min_length': 94.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.75, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.5068317651748657, 'rewards/meter/std': 0.25957190990448, 'rewards/count_adherence/mean': 0.7857142686843872, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9357142448425293, 'rewards/count_floor/std': 0.02290808968245983, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.6283880472183228, 'rewards/repeat_penalty/std': 0.15834280848503113, 'rewards/repeat_floor/mean': 0.9582613706588745, 'rewards/repeat_floor/std': 0.01779051683843136, 'rewards/near_duplicate_penalty/mean': 0.9038777351379395, 'rewards/near_duplicate_penalty/std': 0.03803898021578789, 'rewards/opening_diversity/mean': 0.9094842672348022, 'rewards/opening_diversity/std': 0.10487770289182663, 'rewards/distinct_2/mean': 0.7567849159240723, 'rewards/distinct_2/std': 0.10809030383825302, 'rewards/total_composite/mean': 0.17452022433280945, 'rewards/total_composite/std': 0.08144509047269821, 'reward': 0.17452022433280945, 'reward_std': 0.08144509792327881, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1275382936000824, 'sampling/sampling_logp_difference/max': 2.9277753829956055, 'sampling/importance_sampling_ratio/min': 0.0535159558057785, 'sampling/importance_sampling_ratio/mean': 0.9980449080467224, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5731510818004608, 'clip_ratio/low_mean': 0.053869991563260555, 'clip_ratio/low_min': 0.053869991563260555, 'clip_ratio/high_mean': 0.05861018318682909, 'clip_ratio/high_max': 0.05861018318682909, 'clip_ratio/region_mean': 0.11248017475008965, 'reward_total_mean': 0.17452022433280945, 'reward_meter_mean': 0.5068317651748657, 'reward_meter_std': 0.25957190990448, 'reward_count_adherence_mean': 0.7857142686843872, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9357142448425293, 'reward_count_floor_std': 0.02290808968245983, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6283880472183228, 'reward_repeat_penalty_std': 0.15834280848503113, 'reward_repeat_floor_mean': 0.9582613706588745, 'reward_repeat_floor_std': 0.01779051683843136, 'reward_near_duplicate_penalty_mean': 0.9038777351379395, 'reward_near_duplicate_penalty_std': 0.03803898021578789, 'reward_opening_diversity_mean': 0.9094842672348022, 'reward_opening_diversity_std': 0.10487770289182663, 'reward_distinct_2_mean': 0.7567849159240723, 'reward_distinct_2_std': 0.10809030383825302, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.17452022433280945, 'reward_total_composite_std': 0.08144509047269821, 'epoch': 0.06}

 33%|███▎      | 1100/3300 [4:11:15<6:41:31, 10.95s/it]INFO 04-12 20:09:42 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:01,  5.59s/it]

 23%|██▎       | 3/13 [00:22<01:21,  8.11s/it]

 31%|███       | 4/13 [00:30<01:12,  8.06s/it]

 38%|███▊      | 5/13 [00:42<01:14,  9.37s/it]

 46%|████▌     | 6/13 [00:54<01:11, 10.23s/it]

 54%|█████▍    | 7/13 [01:06<01:04, 10.80s/it]

 62%|██████▏   | 8/13 [01:17<00:54, 10.86s/it]

 69%|██████▉   | 9/13 [01:27<00:41, 10.47s/it]

 77%|███████▋  | 10/13 [01:39<00:33, 11.09s/it]

 85%|████████▍ | 11/13 [01:50<00:22, 11.02s/it]

 92%|█████████▏| 12/13 [02:02<00:11, 11.33s/it]

100%|██████████| 13/13 [02:10<00:00, 10.23s/it]2026-04-12 20:12:05,336 | INFO | train_grpo_train | metrics_logged mode=eval step=1100

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 143.2477, 'eval_samples_per_second': 0.726, 'eval_steps_per_second': 0.091, 'eval_num_tokens': 1960543.0, 'eval_completions/mean_length': 161.42307692307693, 'eval_completions/min_length': 38.15384615384615, 'eval_completions/max_length': 459.15384615384613, 'eval_completions/clipped_ratio': 0.11538461538461539, 'eval_completions/mean_terminated_length': 114.97060805100661, 'eval_completions/min_terminated_length': 38.15384615384615, 'eval_completions/max_terminated_length': 229.76923076923077, 'eval_rewards/meter/mean': 0.5215150645146003, 'eval_rewards/meter/std': 0.38024141696783215, 'eval_rewards/count_adherence/mean': 0.8148601055145264, 'eval_rewards/count_adherence/std': 0.17769379340685332, 'eval_rewards/arabic_clean/mean': 0.8173076923076923, 'eval_rewards/arabic_clean/std': 0.3580388908202832, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9942307655627911, 'eval_rewards/arabic_floor/std': 0.016317846683355477, 'eval_rewards/count_floor/mean': 0.9444580399073087, 'eval_rewards/count_floor/std': 0.053308138193992466, 'eval_rewards/lexical_plausibility/mean': 0.9602617254624, 'eval_rewards/lexical_plausibility/std': 0.09390096194469012, 'eval_rewards/judge_quality/mean': 0.2883653835608409, 'eval_rewards/judge_quality/std': 0.1710561364889145, 'eval_rewards/repeat_penalty/mean': 0.6766370351497943, 'eval_rewards/repeat_penalty/std': 0.27264298727879155, 'eval_rewards/repeat_floor/mean': 0.9617866002596341, 'eval_rewards/repeat_floor/std': 0.03687848666539559, 'eval_rewards/near_duplicate_penalty/mean': 0.9054372677436242, 'eval_rewards/near_duplicate_penalty/std': 0.09143693057390359, 'eval_rewards/opening_diversity/mean': 0.9644795335256137, 'eval_rewards/opening_diversity/std': 0.08858410097085513, 'eval_rewards/distinct_2/mean': 0.7566915704653814, 'eval_rewards/distinct_2/std': 0.24649564176797867, 'eval_rewards/total_composite/mean': 0.1296087603729505, 'eval_rewards/total_composite/std': 0.127347124597201, 'eval_reward': 0.1296087603729505, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06633633776352955, 'eval_sampling/sampling_logp_difference/max': 1.2731393667367787, 'eval_sampling/importance_sampling_ratio/min': 0.30760770147809613, 'eval_sampling/importance_sampling_ratio/mean': 1.0148192185621996, 'eval_sampling/importance_sampling_ratio/max': 1.4713487441723163, 'eval_entropy': 0.7175642481217017, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1296087603729505, 'eval_reward_meter_mean': 0.5215150645146003, 'eval_reward_meter_std': 0.38024141696783215, 'eval_reward_count_adherence_mean': 0.8148601055145264, 'eval_reward_count_adherence_std': 0.17769379340685332, 'eval_reward_arabic_clean_mean': 0.8173076923076923, 'eval_reward_arabic_clean_std': 0.3580388908202832, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9942307655627911, 'eval_reward_arabic_floor_std': 0.016317846683355477, 'eval_reward_count_floor_mean': 0.9444580399073087, 'eval_reward_count_floor_std': 0.053308138193992466, 'eval_reward_lexical_plausibility_mean': 0.9602617254624, 'eval_reward_lexical_plausibility_std': 0.09390096194469012, 'eval_reward_repeat_penalty_mean': 0.6766370351497943, 'eval_reward_repeat_penalty_std': 0.27264298727879155, 'eval_reward_repeat_floor_mean': 0.9617866002596341, 'eval_reward_repeat_floor_std': 0.03687848666539559, 'eval_reward_near_duplicate_penalty_mean': 0.9054372677436242, 'eval_reward_near_duplicate_penalty_std': 0.09143693057390359, 'eval_reward_opening_diversity_mean': 0.9644795335256137, 'eval_reward_opening_diversity_std': 0.08858410097085513, 'eval_reward_distinct_2_mean': 0.7566915704653814, 'eval_reward_distinct_2_std': 0.24649564176797867, 'eval_reward_judge_quality_mean': 0.2883653835608409, 'eval_reward_judge_quality_std': 0.1710561364889145, 'eval_reward_total_composite_mean': 0.1296087603729505, 'eval_reward_total_composite_std': 0.127347124597201, 'epoch': 0.06}

 33%|███▎      | 1100/3300 [4:13:39<6:41:31, 10.95s/it]

100%|██████████| 13/13 [02:10<00:00, 10.23s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 33%|███▎      | 1101/3300 [4:13:55<33:54:10, 55.50s/it]2026-04-12 20:12:21,413 | INFO | train_grpo_train | metrics_logged mode=train step=1101

                                                        
{'loss': 0.081, 'grad_norm': 13.073052406311035, 'learning_rate': 6.666666666666667e-06, 'num_tokens': 1962881.0, 'completions/mean_length': 85.25, 'completions/min_length': 69.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.3148347735404968, 'rewards/meter/std': 0.33956053853034973, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9812500476837158, 'rewards/count_floor/std': 0.025877466425299644, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.854625940322876, 'rewards/repeat_penalty/std': 0.11572480201721191, 'rewards/repeat_floor/mean': 0.9845839738845825, 'rewards/repeat_floor/std': 0.01062406413257122, 'rewards/near_duplicate_penalty/mean': 0.9577128291130066, 'rewards/near_duplicate_penalty/std': 0.02396639995276928, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8991890549659729, 'rewards/distinct_2/std': 0.08846036344766617, 'rewards/total_composite/mean': 0.0951695591211319, 'rewards/total_composite/std': 0.12376604974269867, 'reward': 0.0951695591211319, 'reward_std': 0.12376604229211807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18066447973251343, 'sampling/sampling_logp_difference/max': 2.199256181716919, 'sampling/importance_sampling_ratio/min': 0.11088559776544571, 'sampling/importance_sampling_ratio/mean': 1.004384994506836, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7968593612313271, 'clip_ratio/low_mean': 0.08449057955294847, 'clip_ratio/low_min': 0.08449057955294847, 'clip_ratio/high_mean': 0.05091132130473852, 'clip_ratio/high_max': 0.05091132130473852, 'clip_ratio/region_mean': 0.135401900857687, 'reward_total_mean': 0.0951695591211319, 'reward_meter_mean': 0.3148347735404968, 'reward_meter_std': 0.33956053853034973, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9812500476837158, 'reward_count_floor_std': 0.025877466425299644, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.854625940322876, 'reward_repeat_penalty_std': 0.11572480201721191, 'reward_repeat_floor_mean': 0.9845839738845825, 'reward_repeat_floor_std': 0.01062406413257122, 'reward_near_duplicate_penalty_mean': 0.9577128291130066, 'reward_near_duplicate_penalty_std': 0.02396639995276928, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8991890549659729, 'reward_distinct_2_std': 0.08846036344766617, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.25670650601387024, 'reward_total_composite_mean': 0.0951695591211319, 'reward_total_composite_std': 0.12376604974269867, 'epoch': 0.06}

 33%|███▎      | 1101/3300 [4:13:55<33:54:10, 55.50s/it]INFO 04-12 20:12:21 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1102/3300 [4:14:04<25:24:35, 41.62s/it]2026-04-12 20:12:30,624 | INFO | train_grpo_train | metrics_logged mode=train step=1102

                                                        
{'loss': 0.0697, 'grad_norm': 16.103296279907227, 'learning_rate': 6.663636363636365e-06, 'num_tokens': 1964546.0, 'completions/mean_length': 37.125, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6071222424507141, 'rewards/meter/std': 0.37576961517333984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5, 'rewards/judge_quality/std': 0.22677868604660034, 'rewards/repeat_penalty/mean': 0.8731639981269836, 'rewards/repeat_penalty/std': 0.10080478340387344, 'rewards/repeat_floor/mean': 0.9834874868392944, 'rewards/repeat_floor/std': 0.012064182199537754, 'rewards/near_duplicate_penalty/mean': 0.9238909482955933, 'rewards/near_duplicate_penalty/std': 0.04522726312279701, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9433760643005371, 'rewards/distinct_2/std': 0.0789896696805954, 'rewards/total_composite/mean': 0.346088171005249, 'rewards/total_composite/std': 0.27740663290023804, 'reward': 0.346088171005249, 'reward_std': 0.27740663290023804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12328404933214188, 'sampling/sampling_logp_difference/max': 2.982288122177124, 'sampling/importance_sampling_ratio/min': 0.050676748156547546, 'sampling/importance_sampling_ratio/mean': 0.9920760989189148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6848688311874866, 'clip_ratio/low_mean': 0.028126374119892716, 'clip_ratio/low_min': 0.028126374119892716, 'clip_ratio/high_mean': 0.035058798966929317, 'clip_ratio/high_max': 0.035058798966929317, 'clip_ratio/region_mean': 0.06318517308682203, 'reward_total_mean': 0.346088171005249, 'reward_meter_mean': 0.6071222424507141, 'reward_meter_std': 0.37576961517333984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8731639981269836, 'reward_repeat_penalty_std': 0.10080478340387344, 'reward_repeat_floor_mean': 0.9834874868392944, 'reward_repeat_floor_std': 0.012064182199537754, 'reward_near_duplicate_penalty_mean': 0.9238909482955933, 'reward_near_duplicate_penalty_std': 0.04522726312279701, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9433760643005371, 'reward_distinct_2_std': 0.0789896696805954, 'reward_judge_quality_mean': 0.5, 'reward_judge_quality_std': 0.22677868604660034, 'reward_total_composite_mean': 0.346088171005249, 'reward_total_composite_std': 0.27740663290023804, 'epoch': 0.06}

 33%|███▎      | 1102/3300 [4:14:04<25:24:35, 41.62s/it]INFO 04-12 20:12:30 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1103/3300 [4:14:17<20:07:32, 32.98s/it]2026-04-12 20:12:43,447 | INFO | train_grpo_train | metrics_logged mode=train step=1103

                                                        
{'loss': -0.045, 'grad_norm': 7.793649196624756, 'learning_rate': 6.660606060606061e-06, 'num_tokens': 1966979.0, 'completions/mean_length': 117.125, 'completions/min_length': 103.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.125, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.9205447435379028, 'rewards/meter/std': 0.10239286720752716, 'rewards/count_adherence/mean': 0.734375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9203125238418579, 'rewards/count_floor/std': 0.013258260674774647, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.2935946583747864, 'rewards/repeat_penalty/std': 0.20078925788402557, 'rewards/repeat_floor/mean': 0.9082276821136475, 'rewards/repeat_floor/std': 0.03151927888393402, 'rewards/near_duplicate_penalty/mean': 0.7923674583435059, 'rewards/near_duplicate_penalty/std': 0.07853700965642929, 'rewards/opening_diversity/mean': 0.853365421295166, 'rewards/opening_diversity/std': 0.1527043879032135, 'rewards/distinct_2/mean': 0.42411795258522034, 'rewards/distinct_2/std': 0.15901106595993042, 'rewards/total_composite/mean': 0.16899530589580536, 'rewards/total_composite/std': 0.07099603861570358, 'reward': 0.16899530589580536, 'reward_std': 0.07099603861570358, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11038536578416824, 'sampling/sampling_logp_difference/max': 3.053380250930786, 'sampling/importance_sampling_ratio/min': 0.04719911143183708, 'sampling/importance_sampling_ratio/mean': 1.017039179801941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4790477119386196, 'clip_ratio/low_mean': 0.04311941610649228, 'clip_ratio/low_min': 0.04311941610649228, 'clip_ratio/high_mean': 0.03171086683869362, 'clip_ratio/high_max': 0.03171086683869362, 'clip_ratio/region_mean': 0.0748302829451859, 'reward_total_mean': 0.16899530589580536, 'reward_meter_mean': 0.9205447435379028, 'reward_meter_std': 0.10239286720752716, 'reward_count_adherence_mean': 0.734375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9203125238418579, 'reward_count_floor_std': 0.013258260674774647, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.2935946583747864, 'reward_repeat_penalty_std': 0.20078925788402557, 'reward_repeat_floor_mean': 0.9082276821136475, 'reward_repeat_floor_std': 0.03151927888393402, 'reward_near_duplicate_penalty_mean': 0.7923674583435059, 'reward_near_duplicate_penalty_std': 0.07853700965642929, 'reward_opening_diversity_mean': 0.853365421295166, 'reward_opening_diversity_std': 0.1527043879032135, 'reward_distinct_2_mean': 0.42411795258522034, 'reward_distinct_2_std': 0.15901106595993042, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.16899530589580536, 'reward_total_composite_std': 0.07099603861570358, 'epoch': 0.06}

 33%|███▎      | 1103/3300 [4:14:17<20:07:32, 32.98s/it]INFO 04-12 20:12:43 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1104/3300 [4:14:31<16:40:11, 27.33s/it]2026-04-12 20:12:57,591 | INFO | train_grpo_train | metrics_logged mode=train step=1104

                                                        
{'loss': -0.0347, 'grad_norm': 1.6550719738006592, 'learning_rate': 6.657575757575758e-06, 'num_tokens': 1968518.0, 'completions/mean_length': 282.375, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 52.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.529358983039856, 'rewards/meter/std': 0.44651857018470764, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.8022055625915527, 'rewards/lexical_plausibility/std': 0.21344909071922302, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9030297994613647, 'rewards/repeat_penalty/std': 0.21560585498809814, 'rewards/repeat_floor/mean': 0.9895449280738831, 'rewards/repeat_floor/std': 0.02082372084259987, 'rewards/near_duplicate_penalty/mean': 0.9677215218544006, 'rewards/near_duplicate_penalty/std': 0.03597903996706009, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9376299381256104, 'rewards/distinct_2/std': 0.17640918493270874, 'rewards/total_composite/mean': 0.0933130607008934, 'rewards/total_composite/std': 0.13378001749515533, 'reward': 0.0933130607008934, 'reward_std': 0.13378001749515533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1530848741531372, 'sampling/sampling_logp_difference/max': 1.1004223823547363, 'sampling/importance_sampling_ratio/min': 0.3327305018901825, 'sampling/importance_sampling_ratio/mean': 1.011047601699829, 'sampling/importance_sampling_ratio/max': 1.9163264036178589, 'entropy': 0.6772729009389877, 'clip_ratio/low_mean': 0.033603766933083534, 'clip_ratio/low_min': 0.033603766933083534, 'clip_ratio/high_mean': 0.043729523196816444, 'clip_ratio/high_max': 0.043729523196816444, 'clip_ratio/region_mean': 0.07733329012989998, 'reward_total_mean': 0.0933130607008934, 'reward_meter_mean': 0.529358983039856, 'reward_meter_std': 0.44651857018470764, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.8022055625915527, 'reward_lexical_plausibility_std': 0.21344909071922302, 'reward_repeat_penalty_mean': 0.9030297994613647, 'reward_repeat_penalty_std': 0.21560585498809814, 'reward_repeat_floor_mean': 0.9895449280738831, 'reward_repeat_floor_std': 0.02082372084259987, 'reward_near_duplicate_penalty_mean': 0.9677215218544006, 'reward_near_duplicate_penalty_std': 0.03597903996706009, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9376299381256104, 'reward_distinct_2_std': 0.17640918493270874, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.0933130607008934, 'reward_total_composite_std': 0.13378001749515533, 'epoch': 0.06}

 33%|███▎      | 1104/3300 [4:14:31<16:40:11, 27.33s/it]INFO 04-12 20:12:57 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1105/3300 [4:14:40<13:18:15, 21.82s/it]2026-04-12 20:13:06,566 | INFO | train_grpo_train | metrics_logged mode=train step=1105

                                                        
{'loss': 0.0401, 'grad_norm': 13.847715377807617, 'learning_rate': 6.654545454545455e-06, 'num_tokens': 1970547.0, 'completions/mean_length': 74.625, 'completions/min_length': 61.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.4378034770488739, 'rewards/meter/std': 0.34693700075149536, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.6074966192245483, 'rewards/repeat_penalty/std': 0.21280565857887268, 'rewards/repeat_floor/mean': 0.9571378827095032, 'rewards/repeat_floor/std': 0.02454056590795517, 'rewards/near_duplicate_penalty/mean': 0.9023202061653137, 'rewards/near_duplicate_penalty/std': 0.04660462960600853, 'rewards/opening_diversity/mean': 0.8095238208770752, 'rewards/opening_diversity/std': 0.16274075210094452, 'rewards/distinct_2/mean': 0.8135380744934082, 'rewards/distinct_2/std': 0.13743402063846588, 'rewards/total_composite/mean': 0.14529544115066528, 'rewards/total_composite/std': 0.13355346024036407, 'reward': 0.14529544115066528, 'reward_std': 0.13355346024036407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13097421824932098, 'sampling/sampling_logp_difference/max': 1.811705470085144, 'sampling/importance_sampling_ratio/min': 0.16337527334690094, 'sampling/importance_sampling_ratio/mean': 1.002448558807373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6577684804797173, 'clip_ratio/low_mean': 0.11290837451815605, 'clip_ratio/low_min': 0.11290837451815605, 'clip_ratio/high_mean': 0.03404602035880089, 'clip_ratio/high_max': 0.03404602035880089, 'clip_ratio/region_mean': 0.14695439487695694, 'reward_total_mean': 0.14529544115066528, 'reward_meter_mean': 0.4378034770488739, 'reward_meter_std': 0.34693700075149536, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6074966192245483, 'reward_repeat_penalty_std': 0.21280565857887268, 'reward_repeat_floor_mean': 0.9571378827095032, 'reward_repeat_floor_std': 0.02454056590795517, 'reward_near_duplicate_penalty_mean': 0.9023202061653137, 'reward_near_duplicate_penalty_std': 0.04660462960600853, 'reward_opening_diversity_mean': 0.8095238208770752, 'reward_opening_diversity_std': 0.16274075210094452, 'reward_distinct_2_mean': 0.8135380744934082, 'reward_distinct_2_std': 0.13743402063846588, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.14529544115066528, 'reward_total_composite_std': 0.13355346024036407, 'epoch': 0.06}

 33%|███▎      | 1105/3300 [4:14:40<13:18:15, 21.82s/it]INFO 04-12 20:13:06 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1106/3300 [4:14:54<11:55:17, 19.56s/it]2026-04-12 20:13:20,849 | INFO | train_grpo_train | metrics_logged mode=train step=1106

                                                        
{'loss': -0.0468, 'grad_norm': 2.429562568664551, 'learning_rate': 6.651515151515152e-06, 'num_tokens': 1972096.0, 'completions/mean_length': 154.625, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.5942372679710388, 'rewards/meter/std': 0.3154802918434143, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8606337308883667, 'rewards/lexical_plausibility/std': 0.17988498508930206, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.06408699601888657, 'rewards/repeat_penalty/mean': 0.5485076308250427, 'rewards/repeat_penalty/std': 0.28671008348464966, 'rewards/repeat_floor/mean': 0.9438349008560181, 'rewards/repeat_floor/std': 0.04892116039991379, 'rewards/near_duplicate_penalty/mean': 0.8530266880989075, 'rewards/near_duplicate_penalty/std': 0.14666646718978882, 'rewards/opening_diversity/mean': 0.71875, 'rewards/opening_diversity/std': 0.23857837915420532, 'rewards/distinct_2/mean': 0.8083994388580322, 'rewards/distinct_2/std': 0.17643962800502777, 'rewards/total_composite/mean': 0.08410358428955078, 'rewards/total_composite/std': 0.0748651921749115, 'reward': 0.08410358428955078, 'reward_std': 0.0748651847243309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15047846734523773, 'sampling/sampling_logp_difference/max': 1.846886157989502, 'sampling/importance_sampling_ratio/min': 0.27108946442604065, 'sampling/importance_sampling_ratio/mean': 1.0167381763458252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.640695109963417, 'clip_ratio/low_mean': 0.03189586754888296, 'clip_ratio/low_min': 0.03189586754888296, 'clip_ratio/high_mean': 0.06710030417889357, 'clip_ratio/high_max': 0.06710030417889357, 'clip_ratio/region_mean': 0.09899617172777653, 'reward_total_mean': 0.08410358428955078, 'reward_meter_mean': 0.5942372679710388, 'reward_meter_std': 0.3154802918434143, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8606337308883667, 'reward_lexical_plausibility_std': 0.17988498508930206, 'reward_repeat_penalty_mean': 0.5485076308250427, 'reward_repeat_penalty_std': 0.28671008348464966, 'reward_repeat_floor_mean': 0.9438349008560181, 'reward_repeat_floor_std': 0.04892116039991379, 'reward_near_duplicate_penalty_mean': 0.8530266880989075, 'reward_near_duplicate_penalty_std': 0.14666646718978882, 'reward_opening_diversity_mean': 0.71875, 'reward_opening_diversity_std': 0.23857837915420532, 'reward_distinct_2_mean': 0.8083994388580322, 'reward_distinct_2_std': 0.17643962800502777, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.06408699601888657, 'reward_total_composite_mean': 0.08410358428955078, 'reward_total_composite_std': 0.0748651921749115, 'epoch': 0.06}

 34%|███▎      | 1106/3300 [4:14:54<11:55:17, 19.56s/it]INFO 04-12 20:13:21 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1107/3300 [4:15:08<10:52:32, 17.85s/it]2026-04-12 20:13:34,722 | INFO | train_grpo_train | metrics_logged mode=train step=1107

                                                        
{'loss': -0.1244, 'grad_norm': 2.0914509296417236, 'learning_rate': 6.6484848484848485e-06, 'num_tokens': 1973872.0, 'completions/mean_length': 297.0, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 82.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.42121458053588867, 'rewards/meter/std': 0.33367618918418884, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.8283407092094421, 'rewards/lexical_plausibility/std': 0.18677712976932526, 'rewards/judge_quality/mean': 0.16249999403953552, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.828485369682312, 'rewards/repeat_penalty/std': 0.15440241992473602, 'rewards/repeat_floor/mean': 0.9807182550430298, 'rewards/repeat_floor/std': 0.0176449753344059, 'rewards/near_duplicate_penalty/mean': 0.9299378395080566, 'rewards/near_duplicate_penalty/std': 0.049431342631578445, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9077373743057251, 'rewards/distinct_2/std': 0.11320017278194427, 'rewards/total_composite/mean': 0.06908316910266876, 'rewards/total_composite/std': 0.06482606381177902, 'reward': 0.06908316910266876, 'reward_std': 0.06482606381177902, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14990414679050446, 'sampling/sampling_logp_difference/max': 1.4901032447814941, 'sampling/importance_sampling_ratio/min': 0.22534939646720886, 'sampling/importance_sampling_ratio/mean': 1.0324369668960571, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4034942016005516, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.05423634499311447, 'clip_ratio/high_max': 0.05423634499311447, 'clip_ratio/region_mean': 0.05423634499311447, 'reward_total_mean': 0.06908316910266876, 'reward_meter_mean': 0.42121458053588867, 'reward_meter_std': 0.33367618918418884, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.8283407092094421, 'reward_lexical_plausibility_std': 0.18677712976932526, 'reward_repeat_penalty_mean': 0.828485369682312, 'reward_repeat_penalty_std': 0.15440241992473602, 'reward_repeat_floor_mean': 0.9807182550430298, 'reward_repeat_floor_std': 0.0176449753344059, 'reward_near_duplicate_penalty_mean': 0.9299378395080566, 'reward_near_duplicate_penalty_std': 0.049431342631578445, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9077373743057251, 'reward_distinct_2_std': 0.11320017278194427, 'reward_judge_quality_mean': 0.16249999403953552, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.06908316910266876, 'reward_total_composite_std': 0.06482606381177902, 'epoch': 0.06}

 34%|███▎      | 1107/3300 [4:15:08<10:52:32, 17.85s/it]INFO 04-12 20:13:34 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1108/3300 [4:15:19<9:37:08, 15.80s/it] 2026-04-12 20:13:45,720 | INFO | train_grpo_train | metrics_logged mode=train step=1108

                                                       
{'loss': -0.0446, 'grad_norm': 16.983863830566406, 'learning_rate': 6.645454545454546e-06, 'num_tokens': 1975255.0, 'completions/mean_length': 18.875, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9133576154708862, 'rewards/meter/std': 0.15169987082481384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.2620250880718231, 'rewards/repeat_penalty/mean': 0.6188226342201233, 'rewards/repeat_penalty/std': 0.2549654245376587, 'rewards/repeat_floor/mean': 0.9648507833480835, 'rewards/repeat_floor/std': 0.03714272007346153, 'rewards/near_duplicate_penalty/mean': 0.9275311231613159, 'rewards/near_duplicate_penalty/std': 0.1134401336312294, 'rewards/opening_diversity/mean': 0.65625, 'rewards/opening_diversity/std': 0.2651650309562683, 'rewards/distinct_2/mean': 0.9594017267227173, 'rewards/distinct_2/std': 0.07531212270259857, 'rewards/total_composite/mean': 0.47816914319992065, 'rewards/total_composite/std': 0.2793678641319275, 'reward': 0.47816914319992065, 'reward_std': 0.2793678343296051, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10533981770277023, 'sampling/sampling_logp_difference/max': 1.3286073207855225, 'sampling/importance_sampling_ratio/min': 0.2648458480834961, 'sampling/importance_sampling_ratio/mean': 1.033332109451294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6206137649714947, 'clip_ratio/low_mean': 0.0996563108637929, 'clip_ratio/low_min': 0.0996563108637929, 'clip_ratio/high_mean': 0.017613636795431376, 'clip_ratio/high_max': 0.017613636795431376, 'clip_ratio/region_mean': 0.11726994765922427, 'reward_total_mean': 0.47816914319992065, 'reward_meter_mean': 0.9133576154708862, 'reward_meter_std': 0.15169987082481384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6188226342201233, 'reward_repeat_penalty_std': 0.2549654245376587, 'reward_repeat_floor_mean': 0.9648507833480835, 'reward_repeat_floor_std': 0.03714272007346153, 'reward_near_duplicate_penalty_mean': 0.9275311231613159, 'reward_near_duplicate_penalty_std': 0.1134401336312294, 'reward_opening_diversity_mean': 0.65625, 'reward_opening_diversity_std': 0.2651650309562683, 'reward_distinct_2_mean': 0.9594017267227173, 'reward_distinct_2_std': 0.07531212270259857, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.2620250880718231, 'reward_total_composite_mean': 0.47816914319992065, 'reward_total_composite_std': 0.2793678641319275, 'epoch': 0.06}

 34%|███▎      | 1108/3300 [4:15:19<9:37:08, 15.80s/it]INFO 04-12 20:13:45 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1109/3300 [4:15:32<9:09:32, 15.05s/it]2026-04-12 20:13:59,021 | INFO | train_grpo_train | metrics_logged mode=train step=1109

                                                       
{'loss': -0.0999, 'grad_norm': 2.0615975856781006, 'learning_rate': 6.642424242424242e-06, 'num_tokens': 1976866.0, 'completions/mean_length': 162.375, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.833335876464844, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8475849032402039, 'rewards/meter/std': 0.21688300371170044, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9197299480438232, 'rewards/lexical_plausibility/std': 0.15040437877178192, 'rewards/judge_quality/mean': 0.3687499761581421, 'rewards/judge_quality/std': 0.24192309379577637, 'rewards/repeat_penalty/mean': 0.9483822584152222, 'rewards/repeat_penalty/std': 0.07223664969205856, 'rewards/repeat_floor/mean': 0.9929223656654358, 'rewards/repeat_floor/std': 0.009071624837815762, 'rewards/near_duplicate_penalty/mean': 0.9620465040206909, 'rewards/near_duplicate_penalty/std': 0.03764668107032776, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9846153855323792, 'rewards/distinct_2/std': 0.04351425915956497, 'rewards/total_composite/mean': 0.28844931721687317, 'rewards/total_composite/std': 0.18522420525550842, 'reward': 0.28844931721687317, 'reward_std': 0.18522420525550842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12053454667329788, 'sampling/sampling_logp_difference/max': 1.2679967880249023, 'sampling/importance_sampling_ratio/min': 0.28139474987983704, 'sampling/importance_sampling_ratio/mean': 1.0279974937438965, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7419822961091995, 'clip_ratio/low_mean': 0.02445652149617672, 'clip_ratio/low_min': 0.02445652149617672, 'clip_ratio/high_mean': 0.09371843561530113, 'clip_ratio/high_max': 0.09371843561530113, 'clip_ratio/region_mean': 0.11817495711147785, 'reward_total_mean': 0.28844931721687317, 'reward_meter_mean': 0.8475849032402039, 'reward_meter_std': 0.21688300371170044, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9197299480438232, 'reward_lexical_plausibility_std': 0.15040437877178192, 'reward_repeat_penalty_mean': 0.9483822584152222, 'reward_repeat_penalty_std': 0.07223664969205856, 'reward_repeat_floor_mean': 0.9929223656654358, 'reward_repeat_floor_std': 0.009071624837815762, 'reward_near_duplicate_penalty_mean': 0.9620465040206909, 'reward_near_duplicate_penalty_std': 0.03764668107032776, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9846153855323792, 'reward_distinct_2_std': 0.04351425915956497, 'reward_judge_quality_mean': 0.3687499761581421, 'reward_judge_quality_std': 0.24192309379577637, 'reward_total_composite_mean': 0.28844931721687317, 'reward_total_composite_std': 0.18522420525550842, 'epoch': 0.06}

 34%|███▎      | 1109/3300 [4:15:32<9:09:32, 15.05s/it]INFO 04-12 20:13:59 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1110/3300 [4:15:47<9:02:23, 14.86s/it]2026-04-12 20:14:13,444 | INFO | train_grpo_train | metrics_logged mode=train step=1110

                                                       
{'loss': -0.0794, 'grad_norm': 2.9179773330688477, 'learning_rate': 6.63939393939394e-06, 'num_tokens': 1979022.0, 'completions/mean_length': 183.5, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 74.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.8318451642990112, 'rewards/meter/std': 0.20805364847183228, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.04008918255567551, 'rewards/lexical_plausibility/mean': 0.9160492420196533, 'rewards/lexical_plausibility/std': 0.15560375154018402, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.8809946179389954, 'rewards/repeat_penalty/std': 0.13774196803569794, 'rewards/repeat_floor/mean': 0.985791802406311, 'rewards/repeat_floor/std': 0.01610884815454483, 'rewards/near_duplicate_penalty/mean': 0.9527111053466797, 'rewards/near_duplicate_penalty/std': 0.04578179493546486, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9209458231925964, 'rewards/distinct_2/std': 0.11181499809026718, 'rewards/total_composite/mean': 0.19155310094356537, 'rewards/total_composite/std': 0.11219286173582077, 'reward': 0.19155310094356537, 'reward_std': 0.11219285428524017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13830317556858063, 'sampling/sampling_logp_difference/max': 3.227928638458252, 'sampling/importance_sampling_ratio/min': 0.03963952139019966, 'sampling/importance_sampling_ratio/mean': 1.014770746231079, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.59681037068367, 'clip_ratio/low_mean': 0.03618951514363289, 'clip_ratio/low_min': 0.03618951514363289, 'clip_ratio/high_mean': 0.06465556658804417, 'clip_ratio/high_max': 0.06465556658804417, 'clip_ratio/region_mean': 0.10084508173167706, 'reward_total_mean': 0.19155310094356537, 'reward_meter_mean': 0.8318451642990112, 'reward_meter_std': 0.20805364847183228, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.04008918255567551, 'reward_lexical_plausibility_mean': 0.9160492420196533, 'reward_lexical_plausibility_std': 0.15560375154018402, 'reward_repeat_penalty_mean': 0.8809946179389954, 'reward_repeat_penalty_std': 0.13774196803569794, 'reward_repeat_floor_mean': 0.985791802406311, 'reward_repeat_floor_std': 0.01610884815454483, 'reward_near_duplicate_penalty_mean': 0.9527111053466797, 'reward_near_duplicate_penalty_std': 0.04578179493546486, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9209458231925964, 'reward_distinct_2_std': 0.11181499809026718, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.19155310094356537, 'reward_total_composite_std': 0.11219286173582077, 'epoch': 0.06}

 34%|███▎      | 1110/3300 [4:15:47<9:02:23, 14.86s/it]INFO 04-12 20:14:13 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1111/3300 [4:16:01<8:52:02, 14.58s/it]2026-04-12 20:14:27,376 | INFO | train_grpo_train | metrics_logged mode=train step=1111

                                                       
{'loss': -0.0628, 'grad_norm': 5.7866668701171875, 'learning_rate': 6.6363636363636375e-06, 'num_tokens': 1981067.0, 'completions/mean_length': 119.625, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.57143020629883, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.3348644971847534, 'rewards/meter/std': 0.2584976553916931, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9496723413467407, 'rewards/lexical_plausibility/std': 0.14234818518161774, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.7817235589027405, 'rewards/repeat_penalty/std': 0.1948297768831253, 'rewards/repeat_floor/mean': 0.9792885184288025, 'rewards/repeat_floor/std': 0.01848774217069149, 'rewards/near_duplicate_penalty/mean': 0.9541678428649902, 'rewards/near_duplicate_penalty/std': 0.036467552185058594, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8514680862426758, 'rewards/distinct_2/std': 0.13727322220802307, 'rewards/total_composite/mean': 0.06865432113409042, 'rewards/total_composite/std': 0.05735473707318306, 'reward': 0.06865432113409042, 'reward_std': 0.05735473707318306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15230675041675568, 'sampling/sampling_logp_difference/max': 1.7144007682800293, 'sampling/importance_sampling_ratio/min': 0.18007159233093262, 'sampling/importance_sampling_ratio/mean': 1.015440583229065, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7344888225197792, 'clip_ratio/low_mean': 0.036706872284412384, 'clip_ratio/low_min': 0.036706872284412384, 'clip_ratio/high_mean': 0.06962484121322632, 'clip_ratio/high_max': 0.06962484121322632, 'clip_ratio/region_mean': 0.1063317134976387, 'reward_total_mean': 0.06865432113409042, 'reward_meter_mean': 0.3348644971847534, 'reward_meter_std': 0.2584976553916931, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9496723413467407, 'reward_lexical_plausibility_std': 0.14234818518161774, 'reward_repeat_penalty_mean': 0.7817235589027405, 'reward_repeat_penalty_std': 0.1948297768831253, 'reward_repeat_floor_mean': 0.9792885184288025, 'reward_repeat_floor_std': 0.01848774217069149, 'reward_near_duplicate_penalty_mean': 0.9541678428649902, 'reward_near_duplicate_penalty_std': 0.036467552185058594, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8514680862426758, 'reward_distinct_2_std': 0.13727322220802307, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.06865432113409042, 'reward_total_composite_std': 0.05735473707318306, 'epoch': 0.06}

 34%|███▎      | 1111/3300 [4:16:01<8:52:02, 14.58s/it]INFO 04-12 20:14:27 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1112/3300 [4:16:15<8:48:49, 14.50s/it]2026-04-12 20:14:41,687 | INFO | train_grpo_train | metrics_logged mode=train step=1112

                                                       
{'loss': -0.0805, 'grad_norm': 4.035440444946289, 'learning_rate': 6.633333333333334e-06, 'num_tokens': 1982990.0, 'completions/mean_length': 128.375, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.5714340209961, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8845691680908203, 'rewards/meter/std': 0.22545692324638367, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9690029621124268, 'rewards/lexical_plausibility/std': 0.08767286688089371, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9531344175338745, 'rewards/repeat_penalty/std': 0.06168410927057266, 'rewards/repeat_floor/mean': 0.9945938587188721, 'rewards/repeat_floor/std': 0.00690901093184948, 'rewards/near_duplicate_penalty/mean': 0.982026219367981, 'rewards/near_duplicate_penalty/std': 0.020426368340849876, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9698881506919861, 'rewards/distinct_2/std': 0.0463474802672863, 'rewards/total_composite/mean': 0.15288014709949493, 'rewards/total_composite/std': 0.08371905237436295, 'reward': 0.15288014709949493, 'reward_std': 0.08371905237436295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14734318852424622, 'sampling/sampling_logp_difference/max': 3.514888286590576, 'sampling/importance_sampling_ratio/min': 0.0297511275857687, 'sampling/importance_sampling_ratio/mean': 1.0045669078826904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7673340141773224, 'clip_ratio/low_mean': 0.07881138427183032, 'clip_ratio/low_min': 0.07881138427183032, 'clip_ratio/high_mean': 0.01809210516512394, 'clip_ratio/high_max': 0.01809210516512394, 'clip_ratio/region_mean': 0.09690348943695426, 'reward_total_mean': 0.15288014709949493, 'reward_meter_mean': 0.8845691680908203, 'reward_meter_std': 0.22545692324638367, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9690029621124268, 'reward_lexical_plausibility_std': 0.08767286688089371, 'reward_repeat_penalty_mean': 0.9531344175338745, 'reward_repeat_penalty_std': 0.06168410927057266, 'reward_repeat_floor_mean': 0.9945938587188721, 'reward_repeat_floor_std': 0.00690901093184948, 'reward_near_duplicate_penalty_mean': 0.982026219367981, 'reward_near_duplicate_penalty_std': 0.020426368340849876, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9698881506919861, 'reward_distinct_2_std': 0.0463474802672863, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.15288014709949493, 'reward_total_composite_std': 0.08371905237436295, 'epoch': 0.06}

 34%|███▎      | 1112/3300 [4:16:15<8:48:49, 14.50s/it]INFO 04-12 20:14:41 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1113/3300 [4:16:29<8:42:14, 14.33s/it]2026-04-12 20:14:55,608 | INFO | train_grpo_train | metrics_logged mode=train step=1113

                                                       
{'loss': -0.0504, 'grad_norm': 2.941549301147461, 'learning_rate': 6.630303030303031e-06, 'num_tokens': 1984514.0, 'completions/mean_length': 160.5, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 43.333335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.4371984004974365, 'rewards/meter/std': 0.45407575368881226, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9062854647636414, 'rewards/lexical_plausibility/std': 0.17460642755031586, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9275883436203003, 'rewards/repeat_penalty/std': 0.04702247679233551, 'rewards/repeat_floor/mean': 0.9904555082321167, 'rewards/repeat_floor/std': 0.006308635696768761, 'rewards/near_duplicate_penalty/mean': 0.9515511989593506, 'rewards/near_duplicate_penalty/std': 0.03600221872329712, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9746977686882019, 'rewards/distinct_2/std': 0.029676247388124466, 'rewards/total_composite/mean': 0.16044147312641144, 'rewards/total_composite/std': 0.18030868470668793, 'reward': 0.16044147312641144, 'reward_std': 0.18030866980552673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1451808214187622, 'sampling/sampling_logp_difference/max': 1.1717987060546875, 'sampling/importance_sampling_ratio/min': 0.325014591217041, 'sampling/importance_sampling_ratio/mean': 1.0261509418487549, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7441493943333626, 'clip_ratio/low_mean': 0.06391768157482147, 'clip_ratio/low_min': 0.06391768157482147, 'clip_ratio/high_mean': 0.0580859649926424, 'clip_ratio/high_max': 0.0580859649926424, 'clip_ratio/region_mean': 0.12200364656746387, 'reward_total_mean': 0.16044147312641144, 'reward_meter_mean': 0.4371984004974365, 'reward_meter_std': 0.45407575368881226, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9062854647636414, 'reward_lexical_plausibility_std': 0.17460642755031586, 'reward_repeat_penalty_mean': 0.9275883436203003, 'reward_repeat_penalty_std': 0.04702247679233551, 'reward_repeat_floor_mean': 0.9904555082321167, 'reward_repeat_floor_std': 0.006308635696768761, 'reward_near_duplicate_penalty_mean': 0.9515511989593506, 'reward_near_duplicate_penalty_std': 0.03600221872329712, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9746977686882019, 'reward_distinct_2_std': 0.029676247388124466, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.16044147312641144, 'reward_total_composite_std': 0.18030868470668793, 'epoch': 0.06}

 34%|███▎      | 1113/3300 [4:16:29<8:42:14, 14.33s/it]INFO 04-12 20:14:55 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1114/3300 [4:16:39<7:56:23, 13.08s/it]2026-04-12 20:15:05,764 | INFO | train_grpo_train | metrics_logged mode=train step=1114

                                                       
{'loss': 0.0633, 'grad_norm': 14.003056526184082, 'learning_rate': 6.627272727272728e-06, 'num_tokens': 1986178.0, 'completions/mean_length': 48.0, 'completions/min_length': 43.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4023730754852295, 'rewards/meter/std': 0.42862024903297424, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.8969141840934753, 'rewards/repeat_penalty/std': 0.14449533820152283, 'rewards/repeat_floor/mean': 0.9900321960449219, 'rewards/repeat_floor/std': 0.011697766371071339, 'rewards/near_duplicate_penalty/mean': 0.9690596461296082, 'rewards/near_duplicate_penalty/std': 0.02286618761718273, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9824812412261963, 'rewards/distinct_2/std': 0.03248300403356552, 'rewards/total_composite/mean': 0.17937220633029938, 'rewards/total_composite/std': 0.17441724240779877, 'reward': 0.17937220633029938, 'reward_std': 0.17441724240779877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1410696804523468, 'sampling/sampling_logp_difference/max': 1.538599967956543, 'sampling/importance_sampling_ratio/min': 0.2773173451423645, 'sampling/importance_sampling_ratio/mean': 1.0401309728622437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1267013922333717, 'clip_ratio/low_mean': 0.0674998564645648, 'clip_ratio/low_min': 0.0674998564645648, 'clip_ratio/high_mean': 0.07219608779996634, 'clip_ratio/high_max': 0.07219608779996634, 'clip_ratio/region_mean': 0.13969594426453114, 'reward_total_mean': 0.17937220633029938, 'reward_meter_mean': 0.4023730754852295, 'reward_meter_std': 0.42862024903297424, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8969141840934753, 'reward_repeat_penalty_std': 0.14449533820152283, 'reward_repeat_floor_mean': 0.9900321960449219, 'reward_repeat_floor_std': 0.011697766371071339, 'reward_near_duplicate_penalty_mean': 0.9690596461296082, 'reward_near_duplicate_penalty_std': 0.02286618761718273, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9824812412261963, 'reward_distinct_2_std': 0.03248300403356552, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.17937220633029938, 'reward_total_composite_std': 0.17441724240779877, 'epoch': 0.06}

 34%|███▍      | 1114/3300 [4:16:39<7:56:23, 13.08s/it]INFO 04-12 20:15:06 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1115/3300 [4:16:47<6:58:00, 11.48s/it]2026-04-12 20:15:13,520 | INFO | train_grpo_train | metrics_logged mode=train step=1115

                                                       
{'loss': 0.0144, 'grad_norm': 17.99871826171875, 'learning_rate': 6.624242424242425e-06, 'num_tokens': 1987665.0, 'completions/mean_length': 20.875, 'completions/min_length': 19.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.6114475727081299, 'rewards/meter/std': 0.33778610825538635, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.7137500047683716, 'rewards/judge_quality/std': 0.31536543369293213, 'rewards/repeat_penalty/mean': 0.6896699666976929, 'rewards/repeat_penalty/std': 0.08575751632452011, 'rewards/repeat_floor/mean': 0.973910927772522, 'rewards/repeat_floor/std': 0.016552619636058807, 'rewards/near_duplicate_penalty/mean': 0.9365619421005249, 'rewards/near_duplicate_penalty/std': 0.11026822775602341, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.46602147817611694, 'rewards/total_composite/std': 0.32216310501098633, 'reward': 0.46602147817611694, 'reward_std': 0.32216310501098633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08904021233320236, 'sampling/sampling_logp_difference/max': 1.1270697116851807, 'sampling/importance_sampling_ratio/min': 0.32398122549057007, 'sampling/importance_sampling_ratio/mean': 1.0219746828079224, 'sampling/importance_sampling_ratio/max': 1.7735116481781006, 'entropy': 0.6939110308885574, 'clip_ratio/low_mean': 0.06235047848895192, 'clip_ratio/low_min': 0.06235047848895192, 'clip_ratio/high_mean': 0.02949134260416031, 'clip_ratio/high_max': 0.02949134260416031, 'clip_ratio/region_mean': 0.09184182109311223, 'reward_total_mean': 0.46602147817611694, 'reward_meter_mean': 0.6114475727081299, 'reward_meter_std': 0.33778610825538635, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.6896699666976929, 'reward_repeat_penalty_std': 0.08575751632452011, 'reward_repeat_floor_mean': 0.973910927772522, 'reward_repeat_floor_std': 0.016552619636058807, 'reward_near_duplicate_penalty_mean': 0.9365619421005249, 'reward_near_duplicate_penalty_std': 0.11026822775602341, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.7137500047683716, 'reward_judge_quality_std': 0.31536543369293213, 'reward_total_composite_mean': 0.46602147817611694, 'reward_total_composite_std': 0.32216310501098633, 'epoch': 0.06}

 34%|███▍      | 1115/3300 [4:16:47<6:58:00, 11.48s/it]INFO 04-12 20:15:13 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1116/3300 [4:16:56<6:32:27, 10.78s/it]2026-04-12 20:15:22,672 | INFO | train_grpo_train | metrics_logged mode=train step=1116

                                                       
{'loss': 0.0012, 'grad_norm': 9.764896392822266, 'learning_rate': 6.621212121212121e-06, 'num_tokens': 1990171.0, 'completions/mean_length': 106.25, 'completions/min_length': 95.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.25, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.5923319458961487, 'rewards/meter/std': 0.1939777135848999, 'rewards/count_adherence/mean': 0.7777777910232544, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9333333373069763, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.7661954760551453, 'rewards/repeat_penalty/std': 0.12541718780994415, 'rewards/repeat_floor/mean': 0.9736922979354858, 'rewards/repeat_floor/std': 0.014339100569486618, 'rewards/near_duplicate_penalty/mean': 0.9339008331298828, 'rewards/near_duplicate_penalty/std': 0.03763081878423691, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8178571462631226, 'rewards/distinct_2/std': 0.11097404360771179, 'rewards/total_composite/mean': 0.11490198969841003, 'rewards/total_composite/std': 0.04570137336850166, 'reward': 0.11490198969841003, 'reward_std': 0.04570137336850166, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13520045578479767, 'sampling/sampling_logp_difference/max': 4.625446319580078, 'sampling/importance_sampling_ratio/min': 0.009799281135201454, 'sampling/importance_sampling_ratio/mean': 1.0275001525878906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5670370310544968, 'clip_ratio/low_mean': 0.07349181175231934, 'clip_ratio/low_min': 0.07349181175231934, 'clip_ratio/high_mean': 0.059430984780192375, 'clip_ratio/high_max': 0.059430984780192375, 'clip_ratio/region_mean': 0.1329227965325117, 'reward_total_mean': 0.11490198969841003, 'reward_meter_mean': 0.5923319458961487, 'reward_meter_std': 0.1939777135848999, 'reward_count_adherence_mean': 0.7777777910232544, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9333333373069763, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7661954760551453, 'reward_repeat_penalty_std': 0.12541718780994415, 'reward_repeat_floor_mean': 0.9736922979354858, 'reward_repeat_floor_std': 0.014339100569486618, 'reward_near_duplicate_penalty_mean': 0.9339008331298828, 'reward_near_duplicate_penalty_std': 0.03763081878423691, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8178571462631226, 'reward_distinct_2_std': 0.11097404360771179, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.11490198969841003, 'reward_total_composite_std': 0.04570137336850166, 'epoch': 0.06}

 34%|███▍      | 1116/3300 [4:16:56<6:32:27, 10.78s/it]INFO 04-12 20:15:22 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1117/3300 [4:17:05<6:11:46, 10.22s/it]2026-04-12 20:15:31,577 | INFO | train_grpo_train | metrics_logged mode=train step=1117

                                                       
{'loss': 0.0345, 'grad_norm': 14.916417121887207, 'learning_rate': 6.618181818181819e-06, 'num_tokens': 1991890.0, 'completions/mean_length': 37.875, 'completions/min_length': 32.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.3961948752403259, 'rewards/meter/std': 0.3561877906322479, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.923470139503479, 'rewards/repeat_penalty/std': 0.07474140077829361, 'rewards/repeat_floor/mean': 0.9925272464752197, 'rewards/repeat_floor/std': 0.0037442089524120092, 'rewards/near_duplicate_penalty/mean': 0.9748612642288208, 'rewards/near_duplicate_penalty/std': 0.020294705405831337, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9797008037567139, 'rewards/distinct_2/std': 0.037656061351299286, 'rewards/total_composite/mean': 0.09380505234003067, 'rewards/total_composite/std': 0.07677542418241501, 'reward': 0.09380505234003067, 'reward_std': 0.07677542418241501, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14778997004032135, 'sampling/sampling_logp_difference/max': 0.9533829689025879, 'sampling/importance_sampling_ratio/min': 0.38543492555618286, 'sampling/importance_sampling_ratio/mean': 1.0124022960662842, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1333774775266647, 'clip_ratio/low_mean': 0.0597513634711504, 'clip_ratio/low_min': 0.0597513634711504, 'clip_ratio/high_mean': 0.09166131168603897, 'clip_ratio/high_max': 0.09166131168603897, 'clip_ratio/region_mean': 0.15141267515718937, 'reward_total_mean': 0.09380505234003067, 'reward_meter_mean': 0.3961948752403259, 'reward_meter_std': 0.3561877906322479, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.923470139503479, 'reward_repeat_penalty_std': 0.07474140077829361, 'reward_repeat_floor_mean': 0.9925272464752197, 'reward_repeat_floor_std': 0.0037442089524120092, 'reward_near_duplicate_penalty_mean': 0.9748612642288208, 'reward_near_duplicate_penalty_std': 0.020294705405831337, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9797008037567139, 'reward_distinct_2_std': 0.037656061351299286, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.09380505234003067, 'reward_total_composite_std': 0.07677542418241501, 'epoch': 0.06}

 34%|███▍      | 1117/3300 [4:17:05<6:11:46, 10.22s/it]INFO 04-12 20:15:31 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1118/3300 [4:17:19<6:55:03, 11.41s/it]2026-04-12 20:15:45,835 | INFO | train_grpo_train | metrics_logged mode=train step=1118

                                                       
{'loss': -0.0989, 'grad_norm': 5.206904888153076, 'learning_rate': 6.615151515151516e-06, 'num_tokens': 1993986.0, 'completions/mean_length': 143.0, 'completions/min_length': 74.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 90.28572082519531, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.8414691090583801, 'rewards/meter/std': 0.21675491333007812, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.953125, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 0.9532358646392822, 'rewards/lexical_plausibility/std': 0.13226905465126038, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.8449242115020752, 'rewards/repeat_penalty/std': 0.14225640892982483, 'rewards/repeat_floor/mean': 0.9826798439025879, 'rewards/repeat_floor/std': 0.015020211227238178, 'rewards/near_duplicate_penalty/mean': 0.9494950175285339, 'rewards/near_duplicate_penalty/std': 0.030266793444752693, 'rewards/opening_diversity/mean': 0.9642857313156128, 'rewards/opening_diversity/std': 0.06613000482320786, 'rewards/distinct_2/mean': 0.9155377149581909, 'rewards/distinct_2/std': 0.07967370748519897, 'rewards/total_composite/mean': 0.19605238735675812, 'rewards/total_composite/std': 0.1010000929236412, 'reward': 0.19605238735675812, 'reward_std': 0.10100008547306061, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1536867618560791, 'sampling/sampling_logp_difference/max': 3.0424208641052246, 'sampling/importance_sampling_ratio/min': 0.04771922528743744, 'sampling/importance_sampling_ratio/mean': 1.0013548135757446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6599209234118462, 'clip_ratio/low_mean': 0.04460963048040867, 'clip_ratio/low_min': 0.04460963048040867, 'clip_ratio/high_mean': 0.07997776567935944, 'clip_ratio/high_max': 0.07997776567935944, 'clip_ratio/region_mean': 0.1245873961597681, 'reward_total_mean': 0.19605238735675812, 'reward_meter_mean': 0.8414691090583801, 'reward_meter_std': 0.21675491333007812, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.953125, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 0.9532358646392822, 'reward_lexical_plausibility_std': 0.13226905465126038, 'reward_repeat_penalty_mean': 0.8449242115020752, 'reward_repeat_penalty_std': 0.14225640892982483, 'reward_repeat_floor_mean': 0.9826798439025879, 'reward_repeat_floor_std': 0.015020211227238178, 'reward_near_duplicate_penalty_mean': 0.9494950175285339, 'reward_near_duplicate_penalty_std': 0.030266793444752693, 'reward_opening_diversity_mean': 0.9642857313156128, 'reward_opening_diversity_std': 0.06613000482320786, 'reward_distinct_2_mean': 0.9155377149581909, 'reward_distinct_2_std': 0.07967370748519897, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234977960587, 'reward_total_composite_mean': 0.19605238735675812, 'reward_total_composite_std': 0.1010000929236412, 'epoch': 0.06}

 34%|███▍      | 1118/3300 [4:17:19<6:55:03, 11.41s/it]INFO 04-12 20:15:46 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1119/3300 [4:17:28<6:29:51, 10.72s/it]2026-04-12 20:15:54,898 | INFO | train_grpo_train | metrics_logged mode=train step=1119

                                                       
{'loss': 0.0072, 'grad_norm': 15.530054092407227, 'learning_rate': 6.612121212121213e-06, 'num_tokens': 1995622.0, 'completions/mean_length': 40.5, 'completions/min_length': 38.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6489101052284241, 'rewards/meter/std': 0.32416361570358276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9871190190315247, 'rewards/repeat_penalty/std': 0.0182426068931818, 'rewards/repeat_floor/mean': 0.9980678558349609, 'rewards/repeat_floor/std': 0.0027363866101950407, 'rewards/near_duplicate_penalty/mean': 0.9871190190315247, 'rewards/near_duplicate_penalty/std': 0.0182426068931818, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.32625383138656616, 'rewards/total_composite/std': 0.19823554158210754, 'reward': 0.32625383138656616, 'reward_std': 0.19823555648326874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10378772020339966, 'sampling/sampling_logp_difference/max': 1.5988271236419678, 'sampling/importance_sampling_ratio/min': 0.20213346183300018, 'sampling/importance_sampling_ratio/mean': 1.0093050003051758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44540177285671234, 'clip_ratio/low_mean': 0.029111534357070923, 'clip_ratio/low_min': 0.029111534357070923, 'clip_ratio/high_mean': 0.045524826273322105, 'clip_ratio/high_max': 0.045524826273322105, 'clip_ratio/region_mean': 0.07463636063039303, 'reward_total_mean': 0.32625383138656616, 'reward_meter_mean': 0.6489101052284241, 'reward_meter_std': 0.32416361570358276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9871190190315247, 'reward_repeat_penalty_std': 0.0182426068931818, 'reward_repeat_floor_mean': 0.9980678558349609, 'reward_repeat_floor_std': 0.0027363866101950407, 'reward_near_duplicate_penalty_mean': 0.9871190190315247, 'reward_near_duplicate_penalty_std': 0.0182426068931818, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.32625383138656616, 'reward_total_composite_std': 0.19823554158210754, 'epoch': 0.06}

 34%|███▍      | 1119/3300 [4:17:28<6:29:51, 10.72s/it]INFO 04-12 20:15:55 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1120/3300 [4:17:42<7:05:45, 11.72s/it]2026-04-12 20:16:08,947 | INFO | train_grpo_train | metrics_logged mode=train step=1120

                                                       
{'loss': -0.0719, 'grad_norm': 3.9804394245147705, 'learning_rate': 6.609090909090909e-06, 'num_tokens': 1997655.0, 'completions/mean_length': 204.125, 'completions/min_length': 87.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 101.5, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.25961464643478394, 'rewards/meter/std': 0.34651073813438416, 'rewards/count_adherence/mean': 0.7916666269302368, 'rewards/count_adherence/std': 0.07715165615081787, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.023145509883761406, 'rewards/lexical_plausibility/mean': 0.9188123941421509, 'rewards/lexical_plausibility/std': 0.15209685266017914, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.8381876945495605, 'rewards/repeat_penalty/std': 0.09057272970676422, 'rewards/repeat_floor/mean': 0.9822417497634888, 'rewards/repeat_floor/std': 0.009547477588057518, 'rewards/near_duplicate_penalty/mean': 0.9560760259628296, 'rewards/near_duplicate_penalty/std': 0.02068331465125084, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8758928775787354, 'rewards/distinct_2/std': 0.08450737595558167, 'rewards/total_composite/mean': 0.05883817374706268, 'rewards/total_composite/std': 0.08443966507911682, 'reward': 0.05883817374706268, 'reward_std': 0.08443966507911682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15302841365337372, 'sampling/sampling_logp_difference/max': 2.613825798034668, 'sampling/importance_sampling_ratio/min': 0.07325375080108643, 'sampling/importance_sampling_ratio/mean': 0.9979686141014099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7585892826318741, 'clip_ratio/low_mean': 0.06769068352878094, 'clip_ratio/low_min': 0.06769068352878094, 'clip_ratio/high_mean': 0.05295042507350445, 'clip_ratio/high_max': 0.05295042507350445, 'clip_ratio/region_mean': 0.12064110860228539, 'reward_total_mean': 0.05883817374706268, 'reward_meter_mean': 0.25961464643478394, 'reward_meter_std': 0.34651073813438416, 'reward_count_adherence_mean': 0.7916666269302368, 'reward_count_adherence_std': 0.07715165615081787, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.023145509883761406, 'reward_lexical_plausibility_mean': 0.9188123941421509, 'reward_lexical_plausibility_std': 0.15209685266017914, 'reward_repeat_penalty_mean': 0.8381876945495605, 'reward_repeat_penalty_std': 0.09057272970676422, 'reward_repeat_floor_mean': 0.9822417497634888, 'reward_repeat_floor_std': 0.009547477588057518, 'reward_near_duplicate_penalty_mean': 0.9560760259628296, 'reward_near_duplicate_penalty_std': 0.02068331465125084, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8758928775787354, 'reward_distinct_2_std': 0.08450737595558167, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.05883817374706268, 'reward_total_composite_std': 0.08443966507911682, 'epoch': 0.06}

 34%|███▍      | 1120/3300 [4:17:42<7:05:45, 11.72s/it]INFO 04-12 20:16:09 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1121/3300 [4:17:57<7:43:13, 12.76s/it]2026-04-12 20:16:24,103 | INFO | train_grpo_train | metrics_logged mode=train step=1121

                                                       
{'loss': -0.1522, 'grad_norm': 4.552389621734619, 'learning_rate': 6.606060606060607e-06, 'num_tokens': 2000529.0, 'completions/mean_length': 184.25, 'completions/min_length': 129.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 137.42857360839844, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.7251021862030029, 'rewards/meter/std': 0.30393242835998535, 'rewards/count_adherence/mean': 0.671875, 'rewards/count_adherence/std': 0.13258251547813416, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9015624523162842, 'rewards/count_floor/std': 0.03977476432919502, 'rewards/lexical_plausibility/mean': 0.9584192037582397, 'rewards/lexical_plausibility/std': 0.11760833859443665, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.8525344133377075, 'rewards/repeat_penalty/std': 0.09830207377672195, 'rewards/repeat_floor/mean': 0.9834234714508057, 'rewards/repeat_floor/std': 0.010602456517517567, 'rewards/near_duplicate_penalty/mean': 0.9542155861854553, 'rewards/near_duplicate_penalty/std': 0.024001164361834526, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.892123818397522, 'rewards/distinct_2/std': 0.08656427264213562, 'rewards/total_composite/mean': 0.21732273697853088, 'rewards/total_composite/std': 0.1142578274011612, 'reward': 0.21732273697853088, 'reward_std': 0.1142578199505806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16900703310966492, 'sampling/sampling_logp_difference/max': 2.089836597442627, 'sampling/importance_sampling_ratio/min': 0.1237073540687561, 'sampling/importance_sampling_ratio/mean': 1.0127708911895752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7689536064863205, 'clip_ratio/low_mean': 0.03640069626271725, 'clip_ratio/low_min': 0.03640069626271725, 'clip_ratio/high_mean': 0.09764457121491432, 'clip_ratio/high_max': 0.09764457121491432, 'clip_ratio/region_mean': 0.13404526747763157, 'reward_total_mean': 0.21732273697853088, 'reward_meter_mean': 0.7251021862030029, 'reward_meter_std': 0.30393242835998535, 'reward_count_adherence_mean': 0.671875, 'reward_count_adherence_std': 0.13258251547813416, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9015624523162842, 'reward_count_floor_std': 0.03977476432919502, 'reward_lexical_plausibility_mean': 0.9584192037582397, 'reward_lexical_plausibility_std': 0.11760833859443665, 'reward_repeat_penalty_mean': 0.8525344133377075, 'reward_repeat_penalty_std': 0.09830207377672195, 'reward_repeat_floor_mean': 0.9834234714508057, 'reward_repeat_floor_std': 0.010602456517517567, 'reward_near_duplicate_penalty_mean': 0.9542155861854553, 'reward_near_duplicate_penalty_std': 0.024001164361834526, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.892123818397522, 'reward_distinct_2_std': 0.08656427264213562, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.21732273697853088, 'reward_total_composite_std': 0.1142578274011612, 'epoch': 0.06}

 34%|███▍      | 1121/3300 [4:17:57<7:43:13, 12.76s/it]INFO 04-12 20:16:24 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1122/3300 [4:18:11<7:55:15, 13.09s/it]2026-04-12 20:16:37,997 | INFO | train_grpo_train | metrics_logged mode=train step=1122

                                                       
{'loss': -0.0596, 'grad_norm': 3.564344882965088, 'learning_rate': 6.603030303030303e-06, 'num_tokens': 2002157.0, 'completions/mean_length': 166.5, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.333335876464844, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.41081592440605164, 'rewards/meter/std': 0.3736828565597534, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.2086307406425476, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9156250357627869, 'rewards/count_floor/std': 0.06258922815322876, 'rewards/lexical_plausibility/mean': 0.886710524559021, 'rewards/lexical_plausibility/std': 0.15059080719947815, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.2840491831302643, 'rewards/repeat_penalty/mean': 0.9600927233695984, 'rewards/repeat_penalty/std': 0.05911032482981682, 'rewards/repeat_floor/mean': 0.995184063911438, 'rewards/repeat_floor/std': 0.006072294898331165, 'rewards/near_duplicate_penalty/mean': 0.9801214933395386, 'rewards/near_duplicate_penalty/std': 0.024099119007587433, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9796204566955566, 'rewards/distinct_2/std': 0.05608870089054108, 'rewards/total_composite/mean': 0.12884840369224548, 'rewards/total_composite/std': 0.14484259486198425, 'reward': 0.12884840369224548, 'reward_std': 0.14484257996082306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14707811176776886, 'sampling/sampling_logp_difference/max': 2.2851123809814453, 'sampling/importance_sampling_ratio/min': 0.10176262259483337, 'sampling/importance_sampling_ratio/mean': 1.0025691986083984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4895203188061714, 'clip_ratio/low_mean': 0.06439827848225832, 'clip_ratio/low_min': 0.06439827848225832, 'clip_ratio/high_mean': 0.03513986058533192, 'clip_ratio/high_max': 0.03513986058533192, 'clip_ratio/region_mean': 0.09953813906759024, 'reward_total_mean': 0.12884840369224548, 'reward_meter_mean': 0.41081592440605164, 'reward_meter_std': 0.3736828565597534, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.2086307406425476, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9156250357627869, 'reward_count_floor_std': 0.06258922815322876, 'reward_lexical_plausibility_mean': 0.886710524559021, 'reward_lexical_plausibility_std': 0.15059080719947815, 'reward_repeat_penalty_mean': 0.9600927233695984, 'reward_repeat_penalty_std': 0.05911032482981682, 'reward_repeat_floor_mean': 0.995184063911438, 'reward_repeat_floor_std': 0.006072294898331165, 'reward_near_duplicate_penalty_mean': 0.9801214933395386, 'reward_near_duplicate_penalty_std': 0.024099119007587433, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9796204566955566, 'reward_distinct_2_std': 0.05608870089054108, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.2840491831302643, 'reward_total_composite_mean': 0.12884840369224548, 'reward_total_composite_std': 0.14484259486198425, 'epoch': 0.06}

 34%|███▍      | 1122/3300 [4:18:11<7:55:15, 13.09s/it]INFO 04-12 20:16:38 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1123/3300 [4:18:20<7:07:41, 11.79s/it]2026-04-12 20:16:46,738 | INFO | train_grpo_train | metrics_logged mode=train step=1123

                                                       
{'loss': 0.0776, 'grad_norm': 13.980622291564941, 'learning_rate': 6.600000000000001e-06, 'num_tokens': 2003871.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8525334000587463, 'rewards/meter/std': 0.3157230615615845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.21999594569206238, 'rewards/repeat_penalty/mean': 0.8316961526870728, 'rewards/repeat_penalty/std': 0.23539437353610992, 'rewards/repeat_floor/mean': 0.9764221906661987, 'rewards/repeat_floor/std': 0.034281615167856216, 'rewards/near_duplicate_penalty/mean': 0.9059597253799438, 'rewards/near_duplicate_penalty/std': 0.10149959474802017, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9155916571617126, 'rewards/distinct_2/std': 0.1588144749403, 'rewards/total_composite/mean': 0.31011462211608887, 'rewards/total_composite/std': 0.14059360325336456, 'reward': 0.31011462211608887, 'reward_std': 0.14059360325336456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11450143158435822, 'sampling/sampling_logp_difference/max': 2.1690657138824463, 'sampling/importance_sampling_ratio/min': 0.11428434401750565, 'sampling/importance_sampling_ratio/mean': 1.0173248052597046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6251378655433655, 'clip_ratio/low_mean': 0.012392900884151459, 'clip_ratio/low_min': 0.012392900884151459, 'clip_ratio/high_mean': 0.07436944916844368, 'clip_ratio/high_max': 0.07436944916844368, 'clip_ratio/region_mean': 0.08676235005259514, 'reward_total_mean': 0.31011462211608887, 'reward_meter_mean': 0.8525334000587463, 'reward_meter_std': 0.3157230615615845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8316961526870728, 'reward_repeat_penalty_std': 0.23539437353610992, 'reward_repeat_floor_mean': 0.9764221906661987, 'reward_repeat_floor_std': 0.034281615167856216, 'reward_near_duplicate_penalty_mean': 0.9059597253799438, 'reward_near_duplicate_penalty_std': 0.10149959474802017, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9155916571617126, 'reward_distinct_2_std': 0.1588144749403, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.21999594569206238, 'reward_total_composite_mean': 0.31011462211608887, 'reward_total_composite_std': 0.14059360325336456, 'epoch': 0.06}

 34%|███▍      | 1123/3300 [4:18:20<7:07:41, 11.79s/it]INFO 04-12 20:16:47 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1124/3300 [4:18:30<6:44:32, 11.15s/it]2026-04-12 20:16:56,412 | INFO | train_grpo_train | metrics_logged mode=train step=1124

                                                       
{'loss': 0.0846, 'grad_norm': 10.750117301940918, 'learning_rate': 6.596969696969698e-06, 'num_tokens': 2005713.0, 'completions/mean_length': 59.25, 'completions/min_length': 50.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7873080968856812, 'rewards/meter/std': 0.27957883477211, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.6102972030639648, 'rewards/repeat_penalty/std': 0.1749582439661026, 'rewards/repeat_floor/mean': 0.9612038731575012, 'rewards/repeat_floor/std': 0.015765490010380745, 'rewards/near_duplicate_penalty/mean': 0.9189068078994751, 'rewards/near_duplicate_penalty/std': 0.023575618863105774, 'rewards/opening_diversity/mean': 0.7633928656578064, 'rewards/opening_diversity/std': 0.19274136424064636, 'rewards/distinct_2/mean': 0.8618254661560059, 'rewards/distinct_2/std': 0.044812317937612534, 'rewards/total_composite/mean': 0.2615632116794586, 'rewards/total_composite/std': 0.14510861039161682, 'reward': 0.2615632116794586, 'reward_std': 0.14510861039161682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13254885375499725, 'sampling/sampling_logp_difference/max': 2.096240520477295, 'sampling/importance_sampling_ratio/min': 0.12291766703128815, 'sampling/importance_sampling_ratio/mean': 0.9984291195869446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6550969481468201, 'clip_ratio/low_mean': 0.0359583692625165, 'clip_ratio/low_min': 0.0359583692625165, 'clip_ratio/high_mean': 0.09929751511663198, 'clip_ratio/high_max': 0.09929751511663198, 'clip_ratio/region_mean': 0.13525588437914848, 'reward_total_mean': 0.2615632116794586, 'reward_meter_mean': 0.7873080968856812, 'reward_meter_std': 0.27957883477211, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6102972030639648, 'reward_repeat_penalty_std': 0.1749582439661026, 'reward_repeat_floor_mean': 0.9612038731575012, 'reward_repeat_floor_std': 0.015765490010380745, 'reward_near_duplicate_penalty_mean': 0.9189068078994751, 'reward_near_duplicate_penalty_std': 0.023575618863105774, 'reward_opening_diversity_mean': 0.7633928656578064, 'reward_opening_diversity_std': 0.19274136424064636, 'reward_distinct_2_mean': 0.8618254661560059, 'reward_distinct_2_std': 0.044812317937612534, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2615632116794586, 'reward_total_composite_std': 0.14510861039161682, 'epoch': 0.06}

 34%|███▍      | 1124/3300 [4:18:30<6:44:32, 11.15s/it]INFO 04-12 20:16:56 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1125/3300 [4:18:38<6:19:05, 10.46s/it]2026-04-12 20:17:05,244 | INFO | train_grpo_train | metrics_logged mode=train step=1125

                                                       
{'loss': 0.0717, 'grad_norm': 13.392807960510254, 'learning_rate': 6.593939393939395e-06, 'num_tokens': 2007246.0, 'completions/mean_length': 31.625, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.44038692116737366, 'rewards/meter/std': 0.3768443167209625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.1505940556526184, 'rewards/repeat_penalty/mean': 0.8650407195091248, 'rewards/repeat_penalty/std': 0.18009093403816223, 'rewards/repeat_floor/mean': 0.9842195510864258, 'rewards/repeat_floor/std': 0.020091760903596878, 'rewards/near_duplicate_penalty/mean': 0.9507880210876465, 'rewards/near_duplicate_penalty/std': 0.049686118960380554, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9275147914886475, 'rewards/distinct_2/std': 0.10676814615726471, 'rewards/total_composite/mean': 0.13445010781288147, 'rewards/total_composite/std': 0.14371103048324585, 'reward': 0.13445010781288147, 'reward_std': 0.14371101558208466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15485844016075134, 'sampling/sampling_logp_difference/max': 1.4093377590179443, 'sampling/importance_sampling_ratio/min': 0.24430501461029053, 'sampling/importance_sampling_ratio/mean': 0.9953176975250244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7089732997119427, 'clip_ratio/low_mean': 0.055113638285547495, 'clip_ratio/low_min': 0.055113638285547495, 'clip_ratio/high_mean': 0.07573551312088966, 'clip_ratio/high_max': 0.07573551312088966, 'clip_ratio/region_mean': 0.13084915140643716, 'reward_total_mean': 0.13445010781288147, 'reward_meter_mean': 0.44038692116737366, 'reward_meter_std': 0.3768443167209625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8650407195091248, 'reward_repeat_penalty_std': 0.18009093403816223, 'reward_repeat_floor_mean': 0.9842195510864258, 'reward_repeat_floor_std': 0.020091760903596878, 'reward_near_duplicate_penalty_mean': 0.9507880210876465, 'reward_near_duplicate_penalty_std': 0.049686118960380554, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9275147914886475, 'reward_distinct_2_std': 0.10676814615726471, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.1505940556526184, 'reward_total_composite_mean': 0.13445010781288147, 'reward_total_composite_std': 0.14371103048324585, 'epoch': 0.06}

 34%|███▍      | 1125/3300 [4:18:39<6:19:05, 10.46s/it]INFO 04-12 20:17:05 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1126/3300 [4:18:52<6:49:58, 11.32s/it]2026-04-12 20:17:18,562 | INFO | train_grpo_train | metrics_logged mode=train step=1126

                                                       
{'loss': -0.0392, 'grad_norm': 4.587419033050537, 'learning_rate': 6.590909090909091e-06, 'num_tokens': 2008715.0, 'completions/mean_length': 93.625, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.85714340209961, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7132863998413086, 'rewards/meter/std': 0.3601767122745514, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9137206077575684, 'rewards/lexical_plausibility/std': 0.18606792390346527, 'rewards/judge_quality/mean': 0.6875, 'rewards/judge_quality/std': 0.34964269399642944, 'rewards/repeat_penalty/mean': 0.775210440158844, 'rewards/repeat_penalty/std': 0.13890086114406586, 'rewards/repeat_floor/mean': 0.9822299480438232, 'rewards/repeat_floor/std': 0.015960371121764183, 'rewards/near_duplicate_penalty/mean': 0.9657285809516907, 'rewards/near_duplicate_penalty/std': 0.0695989653468132, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9846743941307068, 'rewards/distinct_2/std': 0.04165209084749222, 'rewards/total_composite/mean': 0.5038150548934937, 'rewards/total_composite/std': 0.3079272210597992, 'reward': 0.5038150548934937, 'reward_std': 0.3079272210597992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15248696506023407, 'sampling/sampling_logp_difference/max': 1.5049519538879395, 'sampling/importance_sampling_ratio/min': 0.2220279723405838, 'sampling/importance_sampling_ratio/mean': 1.0120183229446411, 'sampling/importance_sampling_ratio/max': 1.8857957124710083, 'entropy': 0.8756452202796936, 'clip_ratio/low_mean': 0.027657005470246077, 'clip_ratio/low_min': 0.027657005470246077, 'clip_ratio/high_mean': 0.05153837986290455, 'clip_ratio/high_max': 0.05153837986290455, 'clip_ratio/region_mean': 0.07919538533315063, 'reward_total_mean': 0.5038150548934937, 'reward_meter_mean': 0.7132863998413086, 'reward_meter_std': 0.3601767122745514, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9137206077575684, 'reward_lexical_plausibility_std': 0.18606792390346527, 'reward_repeat_penalty_mean': 0.775210440158844, 'reward_repeat_penalty_std': 0.13890086114406586, 'reward_repeat_floor_mean': 0.9822299480438232, 'reward_repeat_floor_std': 0.015960371121764183, 'reward_near_duplicate_penalty_mean': 0.9657285809516907, 'reward_near_duplicate_penalty_std': 0.0695989653468132, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9846743941307068, 'reward_distinct_2_std': 0.04165209084749222, 'reward_judge_quality_mean': 0.6875, 'reward_judge_quality_std': 0.34964269399642944, 'reward_total_composite_mean': 0.5038150548934937, 'reward_total_composite_std': 0.3079272210597992, 'epoch': 0.06}

 34%|███▍      | 1126/3300 [4:18:52<6:49:58, 11.32s/it]INFO 04-12 20:17:18 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1127/3300 [4:19:01<6:28:53, 10.74s/it]2026-04-12 20:17:27,962 | INFO | train_grpo_train | metrics_logged mode=train step=1127

                                                       
{'loss': 0.0048, 'grad_norm': 30.975505828857422, 'learning_rate': 6.5878787878787885e-06, 'num_tokens': 2010415.0, 'completions/mean_length': 45.5, 'completions/min_length': 31.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.517594575881958, 'rewards/meter/std': 0.351583868265152, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9349482655525208, 'rewards/repeat_penalty/std': 0.06766357272863388, 'rewards/repeat_floor/mean': 0.9918985366821289, 'rewards/repeat_floor/std': 0.007502663880586624, 'rewards/near_duplicate_penalty/mean': 0.9652293920516968, 'rewards/near_duplicate_penalty/std': 0.022758612409234047, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9679344296455383, 'rewards/distinct_2/std': 0.05440628528594971, 'rewards/total_composite/mean': 0.19004203379154205, 'rewards/total_composite/std': 0.1314927637577057, 'reward': 0.19004203379154205, 'reward_std': 0.1314927637577057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1117461547255516, 'sampling/sampling_logp_difference/max': 1.3882112503051758, 'sampling/importance_sampling_ratio/min': 0.3044358789920807, 'sampling/importance_sampling_ratio/mean': 1.021960735321045, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7055645324289799, 'clip_ratio/low_mean': 0.04885927261784673, 'clip_ratio/low_min': 0.04885927261784673, 'clip_ratio/high_mean': 0.03511530440300703, 'clip_ratio/high_max': 0.03511530440300703, 'clip_ratio/region_mean': 0.08397457702085376, 'reward_total_mean': 0.19004203379154205, 'reward_meter_mean': 0.517594575881958, 'reward_meter_std': 0.351583868265152, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9349482655525208, 'reward_repeat_penalty_std': 0.06766357272863388, 'reward_repeat_floor_mean': 0.9918985366821289, 'reward_repeat_floor_std': 0.007502663880586624, 'reward_near_duplicate_penalty_mean': 0.9652293920516968, 'reward_near_duplicate_penalty_std': 0.022758612409234047, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9679344296455383, 'reward_distinct_2_std': 0.05440628528594971, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.19004203379154205, 'reward_total_composite_std': 0.1314927637577057, 'epoch': 0.06}

 34%|███▍      | 1127/3300 [4:19:01<6:28:53, 10.74s/it]INFO 04-12 20:17:28 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1128/3300 [4:19:10<6:02:55, 10.03s/it]2026-04-12 20:17:36,308 | INFO | train_grpo_train | metrics_logged mode=train step=1128

                                                       
{'loss': -0.0051, 'grad_norm': 15.235906600952148, 'learning_rate': 6.584848484848485e-06, 'num_tokens': 2011804.0, 'completions/mean_length': 32.625, 'completions/min_length': 29.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.3482735753059387, 'rewards/meter/std': 0.3614812195301056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9653024673461914, 'rewards/lexical_plausibility/std': 0.06643702834844589, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.3384418189525604, 'rewards/repeat_penalty/mean': 0.9854830503463745, 'rewards/repeat_penalty/std': 0.017718393355607986, 'rewards/repeat_floor/mean': 0.9978224635124207, 'rewards/repeat_floor/std': 0.002657755510881543, 'rewards/near_duplicate_penalty/mean': 0.9854830503463745, 'rewards/near_duplicate_penalty/std': 0.017718393355607986, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29860031604766846, 'rewards/total_composite/std': 0.35767844319343567, 'reward': 0.29860031604766846, 'reward_std': 0.35767844319343567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14200299978256226, 'sampling/sampling_logp_difference/max': 1.2735671997070312, 'sampling/importance_sampling_ratio/min': 0.2798316180706024, 'sampling/importance_sampling_ratio/mean': 1.0043753385543823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9176896587014198, 'clip_ratio/low_mean': 0.08291233633644879, 'clip_ratio/low_min': 0.08291233633644879, 'clip_ratio/high_mean': 0.025649351067841053, 'clip_ratio/high_max': 0.025649351067841053, 'clip_ratio/region_mean': 0.10856168740428984, 'reward_total_mean': 0.29860031604766846, 'reward_meter_mean': 0.3482735753059387, 'reward_meter_std': 0.3614812195301056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9653024673461914, 'reward_lexical_plausibility_std': 0.06643702834844589, 'reward_repeat_penalty_mean': 0.9854830503463745, 'reward_repeat_penalty_std': 0.017718393355607986, 'reward_repeat_floor_mean': 0.9978224635124207, 'reward_repeat_floor_std': 0.002657755510881543, 'reward_near_duplicate_penalty_mean': 0.9854830503463745, 'reward_near_duplicate_penalty_std': 0.017718393355607986, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.3384418189525604, 'reward_total_composite_mean': 0.29860031604766846, 'reward_total_composite_std': 0.35767844319343567, 'epoch': 0.06}

 34%|███▍      | 1128/3300 [4:19:10<6:02:55, 10.03s/it]INFO 04-12 20:17:36 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1129/3300 [4:19:18<5:49:49,  9.67s/it]2026-04-12 20:17:45,151 | INFO | train_grpo_train | metrics_logged mode=train step=1129

                                                       
{'loss': -0.0055, 'grad_norm': 12.833215713500977, 'learning_rate': 6.581818181818182e-06, 'num_tokens': 2013296.0, 'completions/mean_length': 26.5, 'completions/min_length': 25.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.5, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.16604852676391602, 'rewards/meter/std': 0.3022329807281494, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8813630938529968, 'rewards/lexical_plausibility/std': 0.22464321553707123, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9446057081222534, 'rewards/repeat_penalty/std': 0.060503557324409485, 'rewards/repeat_floor/mean': 0.9923745393753052, 'rewards/repeat_floor/std': 0.007236011326313019, 'rewards/near_duplicate_penalty/mean': 0.958039402961731, 'rewards/near_duplicate_penalty/std': 0.0261548962444067, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.07083623111248016, 'rewards/total_composite/std': 0.1334410011768341, 'reward': 0.07083623111248016, 'reward_std': 0.1334410011768341, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09989334642887115, 'sampling/sampling_logp_difference/max': 3.296788215637207, 'sampling/importance_sampling_ratio/min': 0.03700181841850281, 'sampling/importance_sampling_ratio/mean': 1.0052670240402222, 'sampling/importance_sampling_ratio/max': 1.6640193462371826, 'entropy': 0.45605812780559063, 'clip_ratio/low_mean': 0.07714387588202953, 'clip_ratio/low_min': 0.07714387588202953, 'clip_ratio/high_mean': 0.008620689623057842, 'clip_ratio/high_max': 0.008620689623057842, 'clip_ratio/region_mean': 0.08576456550508738, 'reward_total_mean': 0.07083623111248016, 'reward_meter_mean': 0.16604852676391602, 'reward_meter_std': 0.3022329807281494, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8813630938529968, 'reward_lexical_plausibility_std': 0.22464321553707123, 'reward_repeat_penalty_mean': 0.9446057081222534, 'reward_repeat_penalty_std': 0.060503557324409485, 'reward_repeat_floor_mean': 0.9923745393753052, 'reward_repeat_floor_std': 0.007236011326313019, 'reward_near_duplicate_penalty_mean': 0.958039402961731, 'reward_near_duplicate_penalty_std': 0.0261548962444067, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.07083623111248016, 'reward_total_composite_std': 0.1334410011768341, 'epoch': 0.06}

 34%|███▍      | 1129/3300 [4:19:19<5:49:49,  9.67s/it]INFO 04-12 20:17:45 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1130/3300 [4:19:30<6:14:13, 10.35s/it]2026-04-12 20:17:57,076 | INFO | train_grpo_train | metrics_logged mode=train step=1130

                                                       
{'loss': 0.0621, 'grad_norm': 36.1773681640625, 'learning_rate': 6.578787878787879e-06, 'num_tokens': 2015752.0, 'completions/mean_length': 100.0, 'completions/min_length': 89.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.0, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9864547252655029, 'rewards/meter/std': 0.009229982271790504, 'rewards/count_adherence/mean': 0.5714285969734192, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8714285492897034, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.09161253273487091, 'rewards/repeat_penalty/mean': 0.6942178010940552, 'rewards/repeat_penalty/std': 0.16141824424266815, 'rewards/repeat_floor/mean': 0.9650284051895142, 'rewards/repeat_floor/std': 0.018470648676156998, 'rewards/near_duplicate_penalty/mean': 0.9118733406066895, 'rewards/near_duplicate_penalty/std': 0.04245678335428238, 'rewards/opening_diversity/mean': 0.953125, 'rewards/opening_diversity/std': 0.0867956355214119, 'rewards/distinct_2/mean': 0.789554238319397, 'rewards/distinct_2/std': 0.10843206197023392, 'rewards/total_composite/mean': 0.17613454163074493, 'rewards/total_composite/std': 0.07548650354146957, 'reward': 0.17613454163074493, 'reward_std': 0.07548649609088898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13595923781394958, 'sampling/sampling_logp_difference/max': 1.6110103130340576, 'sampling/importance_sampling_ratio/min': 0.19968575239181519, 'sampling/importance_sampling_ratio/mean': 1.0114250183105469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0017299130558968, 'clip_ratio/low_mean': 0.07044956088066101, 'clip_ratio/low_min': 0.07044956088066101, 'clip_ratio/high_mean': 0.05550385545939207, 'clip_ratio/high_max': 0.05550385545939207, 'clip_ratio/region_mean': 0.12595341634005308, 'reward_total_mean': 0.17613454163074493, 'reward_meter_mean': 0.9864547252655029, 'reward_meter_std': 0.009229982271790504, 'reward_count_adherence_mean': 0.5714285969734192, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8714285492897034, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6942178010940552, 'reward_repeat_penalty_std': 0.16141824424266815, 'reward_repeat_floor_mean': 0.9650284051895142, 'reward_repeat_floor_std': 0.018470648676156998, 'reward_near_duplicate_penalty_mean': 0.9118733406066895, 'reward_near_duplicate_penalty_std': 0.04245678335428238, 'reward_opening_diversity_mean': 0.953125, 'reward_opening_diversity_std': 0.0867956355214119, 'reward_distinct_2_mean': 0.789554238319397, 'reward_distinct_2_std': 0.10843206197023392, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.09161253273487091, 'reward_total_composite_mean': 0.17613454163074493, 'reward_total_composite_std': 0.07548650354146957, 'epoch': 0.06}

 34%|███▍      | 1130/3300 [4:19:30<6:14:13, 10.35s/it]INFO 04-12 20:17:57 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1131/3300 [4:19:40<6:08:06, 10.18s/it]2026-04-12 20:18:06,875 | INFO | train_grpo_train | metrics_logged mode=train step=1131

                                                       
{'loss': 0.0113, 'grad_norm': 13.86112117767334, 'learning_rate': 6.575757575757577e-06, 'num_tokens': 2017289.0, 'completions/mean_length': 39.125, 'completions/min_length': 37.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.48837825655937195, 'rewards/meter/std': 0.3425035774707794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.8703740239143372, 'rewards/repeat_penalty/std': 0.139456644654274, 'rewards/repeat_floor/mean': 0.9848619103431702, 'rewards/repeat_floor/std': 0.015189900994300842, 'rewards/near_duplicate_penalty/mean': 0.9423151016235352, 'rewards/near_duplicate_penalty/std': 0.05494488775730133, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9487741589546204, 'rewards/distinct_2/std': 0.06066346541047096, 'rewards/total_composite/mean': 0.17353501915931702, 'rewards/total_composite/std': 0.12023576349020004, 'reward': 0.17353501915931702, 'reward_std': 0.12023576349020004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11927077174186707, 'sampling/sampling_logp_difference/max': 1.4823565483093262, 'sampling/importance_sampling_ratio/min': 0.2271018922328949, 'sampling/importance_sampling_ratio/mean': 1.0048420429229736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7911979854106903, 'clip_ratio/low_mean': 0.04927143733948469, 'clip_ratio/low_min': 0.04927143733948469, 'clip_ratio/high_mean': 0.0989175345748663, 'clip_ratio/high_max': 0.0989175345748663, 'clip_ratio/region_mean': 0.148188971914351, 'reward_total_mean': 0.17353501915931702, 'reward_meter_mean': 0.48837825655937195, 'reward_meter_std': 0.3425035774707794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8703740239143372, 'reward_repeat_penalty_std': 0.139456644654274, 'reward_repeat_floor_mean': 0.9848619103431702, 'reward_repeat_floor_std': 0.015189900994300842, 'reward_near_duplicate_penalty_mean': 0.9423151016235352, 'reward_near_duplicate_penalty_std': 0.05494488775730133, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9487741589546204, 'reward_distinct_2_std': 0.06066346541047096, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.17353501915931702, 'reward_total_composite_std': 0.12023576349020004, 'epoch': 0.06}

 34%|███▍      | 1131/3300 [4:19:40<6:08:06, 10.18s/it]INFO 04-12 20:18:07 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1132/3300 [4:19:55<6:54:01, 11.46s/it]2026-04-12 20:18:21,307 | INFO | train_grpo_train | metrics_logged mode=train step=1132

                                                       
{'loss': -0.1145, 'grad_norm': 3.182318687438965, 'learning_rate': 6.572727272727273e-06, 'num_tokens': 2019220.0, 'completions/mean_length': 183.375, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 73.83333587646484, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.8254173994064331, 'rewards/meter/std': 0.11091367900371552, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9096426367759705, 'rewards/lexical_plausibility/std': 0.1676514446735382, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7916404008865356, 'rewards/repeat_penalty/std': 0.17198587954044342, 'rewards/repeat_floor/mean': 0.97715163230896, 'rewards/repeat_floor/std': 0.016961004585027695, 'rewards/near_duplicate_penalty/mean': 0.9342776536941528, 'rewards/near_duplicate_penalty/std': 0.03185384348034859, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8556665778160095, 'rewards/distinct_2/std': 0.14103658497333527, 'rewards/total_composite/mean': 0.17727932333946228, 'rewards/total_composite/std': 0.12055294960737228, 'reward': 0.17727932333946228, 'reward_std': 0.12055294215679169, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13627950847148895, 'sampling/sampling_logp_difference/max': 1.14845871925354, 'sampling/importance_sampling_ratio/min': 0.3628806173801422, 'sampling/importance_sampling_ratio/mean': 1.007847547531128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7004885524511337, 'clip_ratio/low_mean': 0.03209459409117699, 'clip_ratio/low_min': 0.03209459409117699, 'clip_ratio/high_mean': 0.07658103574067354, 'clip_ratio/high_max': 0.07658103574067354, 'clip_ratio/region_mean': 0.10867562983185053, 'reward_total_mean': 0.17727932333946228, 'reward_meter_mean': 0.8254173994064331, 'reward_meter_std': 0.11091367900371552, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9096426367759705, 'reward_lexical_plausibility_std': 0.1676514446735382, 'reward_repeat_penalty_mean': 0.7916404008865356, 'reward_repeat_penalty_std': 0.17198587954044342, 'reward_repeat_floor_mean': 0.97715163230896, 'reward_repeat_floor_std': 0.016961004585027695, 'reward_near_duplicate_penalty_mean': 0.9342776536941528, 'reward_near_duplicate_penalty_std': 0.03185384348034859, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8556665778160095, 'reward_distinct_2_std': 0.14103658497333527, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.17727932333946228, 'reward_total_composite_std': 0.12055294960737228, 'epoch': 0.06}

 34%|███▍      | 1132/3300 [4:19:55<6:54:01, 11.46s/it]INFO 04-12 20:18:21 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1133/3300 [4:20:08<7:19:54, 12.18s/it]2026-04-12 20:18:35,172 | INFO | train_grpo_train | metrics_logged mode=train step=1133

                                                       
{'loss': 0.0085, 'grad_norm': 15.432718276977539, 'learning_rate': 6.56969696969697e-06, 'num_tokens': 2020825.0, 'completions/mean_length': 37.625, 'completions/min_length': 35.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.6276660561561584, 'rewards/meter/std': 0.3171907067298889, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3137759864330292, 'rewards/repeat_penalty/mean': 0.8649664521217346, 'rewards/repeat_penalty/std': 0.19094587862491608, 'rewards/repeat_floor/mean': 0.983667254447937, 'rewards/repeat_floor/std': 0.023933250457048416, 'rewards/near_duplicate_penalty/mean': 0.9502384662628174, 'rewards/near_duplicate_penalty/std': 0.06433877348899841, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9222947955131531, 'rewards/distinct_2/std': 0.10613454133272171, 'rewards/total_composite/mean': 0.3853622078895569, 'rewards/total_composite/std': 0.31717705726623535, 'reward': 0.3853622078895569, 'reward_std': 0.31717702746391296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13214007019996643, 'sampling/sampling_logp_difference/max': 1.607330083847046, 'sampling/importance_sampling_ratio/min': 0.200422003865242, 'sampling/importance_sampling_ratio/mean': 1.0287736654281616, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.889851488173008, 'clip_ratio/low_mean': 0.08795452699996531, 'clip_ratio/low_min': 0.08795452699996531, 'clip_ratio/high_mean': 0.01883012824691832, 'clip_ratio/high_max': 0.01883012824691832, 'clip_ratio/region_mean': 0.10678465524688363, 'reward_total_mean': 0.3853622078895569, 'reward_meter_mean': 0.6276660561561584, 'reward_meter_std': 0.3171907067298889, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8649664521217346, 'reward_repeat_penalty_std': 0.19094587862491608, 'reward_repeat_floor_mean': 0.983667254447937, 'reward_repeat_floor_std': 0.023933250457048416, 'reward_near_duplicate_penalty_mean': 0.9502384662628174, 'reward_near_duplicate_penalty_std': 0.06433877348899841, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9222947955131531, 'reward_distinct_2_std': 0.10613454133272171, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.3137759864330292, 'reward_total_composite_mean': 0.3853622078895569, 'reward_total_composite_std': 0.31717705726623535, 'epoch': 0.06}

 34%|███▍      | 1133/3300 [4:20:09<7:19:54, 12.18s/it]INFO 04-12 20:18:35 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1134/3300 [4:20:24<7:55:25, 13.17s/it]2026-04-12 20:18:50,650 | INFO | train_grpo_train | metrics_logged mode=train step=1134

                                                       
{'loss': -0.12, 'grad_norm': 1.8836357593536377, 'learning_rate': 6.566666666666667e-06, 'num_tokens': 2022563.0, 'completions/mean_length': 172.25, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.818039059638977, 'rewards/meter/std': 0.32381972670555115, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.8922749757766724, 'rewards/lexical_plausibility/std': 0.22093923389911652, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/repeat_penalty/mean': 0.865650475025177, 'rewards/repeat_penalty/std': 0.08957969397306442, 'rewards/repeat_floor/mean': 0.9841786623001099, 'rewards/repeat_floor/std': 0.009905756451189518, 'rewards/near_duplicate_penalty/mean': 0.9467566609382629, 'rewards/near_duplicate_penalty/std': 0.024681156501173973, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9129465818405151, 'rewards/distinct_2/std': 0.07364129275083542, 'rewards/total_composite/mean': 0.2068629264831543, 'rewards/total_composite/std': 0.13120529055595398, 'reward': 0.2068629264831543, 'reward_std': 0.13120529055595398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10028151422739029, 'sampling/sampling_logp_difference/max': 2.1131720542907715, 'sampling/importance_sampling_ratio/min': 0.12085399776697159, 'sampling/importance_sampling_ratio/mean': 0.998281717300415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44601161032915115, 'clip_ratio/low_mean': 0.014830508269369602, 'clip_ratio/low_min': 0.014830508269369602, 'clip_ratio/high_mean': 0.0753679284825921, 'clip_ratio/high_max': 0.0753679284825921, 'clip_ratio/region_mean': 0.09019843675196171, 'reward_total_mean': 0.2068629264831543, 'reward_meter_mean': 0.818039059638977, 'reward_meter_std': 0.32381972670555115, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.8922749757766724, 'reward_lexical_plausibility_std': 0.22093923389911652, 'reward_repeat_penalty_mean': 0.865650475025177, 'reward_repeat_penalty_std': 0.08957969397306442, 'reward_repeat_floor_mean': 0.9841786623001099, 'reward_repeat_floor_std': 0.009905756451189518, 'reward_near_duplicate_penalty_mean': 0.9467566609382629, 'reward_near_duplicate_penalty_std': 0.024681156501173973, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9129465818405151, 'reward_distinct_2_std': 0.07364129275083542, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.2068629264831543, 'reward_total_composite_std': 0.13120529055595398, 'epoch': 0.06}

 34%|███▍      | 1134/3300 [4:20:24<7:55:25, 13.17s/it]INFO 04-12 20:18:50 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1135/3300 [4:20:37<7:59:14, 13.28s/it]2026-04-12 20:19:04,194 | INFO | train_grpo_train | metrics_logged mode=train step=1135

                                                       
{'loss': -0.0602, 'grad_norm': 4.82960319519043, 'learning_rate': 6.563636363636364e-06, 'num_tokens': 2024119.0, 'completions/mean_length': 101.5, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.85714340209961, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.28957587480545044, 'rewards/meter/std': 0.23467731475830078, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9404920339584351, 'rewards/lexical_plausibility/std': 0.16831403970718384, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8318454027175903, 'rewards/repeat_penalty/std': 0.33134788274765015, 'rewards/repeat_floor/mean': 0.9736530184745789, 'rewards/repeat_floor/std': 0.06063687801361084, 'rewards/near_duplicate_penalty/mean': 0.9209880232810974, 'rewards/near_duplicate_penalty/std': 0.17099036276340485, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.9118589758872986, 'rewards/distinct_2/std': 0.24930046498775482, 'rewards/total_composite/mean': 0.0776042640209198, 'rewards/total_composite/std': 0.05956317111849785, 'reward': 0.0776042640209198, 'reward_std': 0.05956317111849785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16488967835903168, 'sampling/sampling_logp_difference/max': 1.4712190628051758, 'sampling/importance_sampling_ratio/min': 0.22964537143707275, 'sampling/importance_sampling_ratio/mean': 1.0245412588119507, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.941789448261261, 'clip_ratio/low_mean': 0.041631230153143406, 'clip_ratio/low_min': 0.041631230153143406, 'clip_ratio/high_mean': 0.083386754617095, 'clip_ratio/high_max': 0.083386754617095, 'clip_ratio/region_mean': 0.1250179847702384, 'reward_total_mean': 0.0776042640209198, 'reward_meter_mean': 0.28957587480545044, 'reward_meter_std': 0.23467731475830078, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9404920339584351, 'reward_lexical_plausibility_std': 0.16831403970718384, 'reward_repeat_penalty_mean': 0.8318454027175903, 'reward_repeat_penalty_std': 0.33134788274765015, 'reward_repeat_floor_mean': 0.9736530184745789, 'reward_repeat_floor_std': 0.06063687801361084, 'reward_near_duplicate_penalty_mean': 0.9209880232810974, 'reward_near_duplicate_penalty_std': 0.17099036276340485, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.9118589758872986, 'reward_distinct_2_std': 0.24930046498775482, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.0776042640209198, 'reward_total_composite_std': 0.05956317111849785, 'epoch': 0.06}

 34%|███▍      | 1135/3300 [4:20:38<7:59:14, 13.28s/it]INFO 04-12 20:19:04 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1136/3300 [4:20:46<7:09:45, 11.92s/it]2026-04-12 20:19:12,923 | INFO | train_grpo_train | metrics_logged mode=train step=1136

                                                       
{'loss': -0.002, 'grad_norm': 11.741394996643066, 'learning_rate': 6.56060606060606e-06, 'num_tokens': 2025782.0, 'completions/mean_length': 42.875, 'completions/min_length': 40.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8009480237960815, 'rewards/meter/std': 0.22720018029212952, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32875001430511475, 'rewards/judge_quality/std': 0.24121344089508057, 'rewards/repeat_penalty/mean': 0.9277142286300659, 'rewards/repeat_penalty/std': 0.10197477787733078, 'rewards/repeat_floor/mean': 0.9906651973724365, 'rewards/repeat_floor/std': 0.012397919781506062, 'rewards/near_duplicate_penalty/mean': 0.9579403400421143, 'rewards/near_duplicate_penalty/std': 0.04623378440737724, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9663797616958618, 'rewards/distinct_2/std': 0.0741337239742279, 'rewards/total_composite/mean': 0.2769007682800293, 'rewards/total_composite/std': 0.2531771659851074, 'reward': 0.2769007682800293, 'reward_std': 0.25317713618278503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1619262546300888, 'sampling/sampling_logp_difference/max': 1.4930922985076904, 'sampling/importance_sampling_ratio/min': 0.22467681765556335, 'sampling/importance_sampling_ratio/mean': 1.004959225654602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1832910180091858, 'clip_ratio/low_mean': 0.07099359226413071, 'clip_ratio/low_min': 0.07099359226413071, 'clip_ratio/high_mean': 0.029980657622218132, 'clip_ratio/high_max': 0.029980657622218132, 'clip_ratio/region_mean': 0.10097424988634884, 'reward_total_mean': 0.2769007682800293, 'reward_meter_mean': 0.8009480237960815, 'reward_meter_std': 0.22720018029212952, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9277142286300659, 'reward_repeat_penalty_std': 0.10197477787733078, 'reward_repeat_floor_mean': 0.9906651973724365, 'reward_repeat_floor_std': 0.012397919781506062, 'reward_near_duplicate_penalty_mean': 0.9579403400421143, 'reward_near_duplicate_penalty_std': 0.04623378440737724, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9663797616958618, 'reward_distinct_2_std': 0.0741337239742279, 'reward_judge_quality_mean': 0.32875001430511475, 'reward_judge_quality_std': 0.24121344089508057, 'reward_total_composite_mean': 0.2769007682800293, 'reward_total_composite_std': 0.2531771659851074, 'epoch': 0.06}

 34%|███▍      | 1136/3300 [4:20:46<7:09:45, 11.92s/it]INFO 04-12 20:19:13 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1137/3300 [4:21:00<7:35:13, 12.63s/it]2026-04-12 20:19:27,216 | INFO | train_grpo_train | metrics_logged mode=train step=1137

                                                       
{'loss': -0.0953, 'grad_norm': 5.342452526092529, 'learning_rate': 6.5575757575757585e-06, 'num_tokens': 2027628.0, 'completions/mean_length': 111.75, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.57143020629883, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6841248869895935, 'rewards/meter/std': 0.3728896677494049, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9550659656524658, 'rewards/lexical_plausibility/std': 0.12709257006645203, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8799244165420532, 'rewards/repeat_penalty/std': 0.0797056257724762, 'rewards/repeat_floor/mean': 0.9848881363868713, 'rewards/repeat_floor/std': 0.009496498852968216, 'rewards/near_duplicate_penalty/mean': 0.9356658458709717, 'rewards/near_duplicate_penalty/std': 0.03665041923522949, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.939314067363739, 'rewards/distinct_2/std': 0.05864523723721504, 'rewards/total_composite/mean': 0.23774784803390503, 'rewards/total_composite/std': 0.14941243827342987, 'reward': 0.23774784803390503, 'reward_std': 0.14941245317459106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12418399751186371, 'sampling/sampling_logp_difference/max': 1.2790470123291016, 'sampling/importance_sampling_ratio/min': 0.278302401304245, 'sampling/importance_sampling_ratio/mean': 0.9947309494018555, 'sampling/importance_sampling_ratio/max': 1.963147521018982, 'entropy': 0.651507593691349, 'clip_ratio/low_mean': 0.033350951969623566, 'clip_ratio/low_min': 0.033350951969623566, 'clip_ratio/high_mean': 0.08394730091094971, 'clip_ratio/high_max': 0.08394730091094971, 'clip_ratio/region_mean': 0.11729825288057327, 'reward_total_mean': 0.23774784803390503, 'reward_meter_mean': 0.6841248869895935, 'reward_meter_std': 0.3728896677494049, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9550659656524658, 'reward_lexical_plausibility_std': 0.12709257006645203, 'reward_repeat_penalty_mean': 0.8799244165420532, 'reward_repeat_penalty_std': 0.0797056257724762, 'reward_repeat_floor_mean': 0.9848881363868713, 'reward_repeat_floor_std': 0.009496498852968216, 'reward_near_duplicate_penalty_mean': 0.9356658458709717, 'reward_near_duplicate_penalty_std': 0.03665041923522949, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.939314067363739, 'reward_distinct_2_std': 0.05864523723721504, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.23774784803390503, 'reward_total_composite_std': 0.14941243827342987, 'epoch': 0.06}

 34%|███▍      | 1137/3300 [4:21:01<7:35:13, 12.63s/it]INFO 04-12 20:19:27 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1138/3300 [4:21:10<7:01:04, 11.69s/it]2026-04-12 20:19:36,707 | INFO | train_grpo_train | metrics_logged mode=train step=1138

                                                       
{'loss': 0.0479, 'grad_norm': 18.135719299316406, 'learning_rate': 6.554545454545455e-06, 'num_tokens': 2029204.0, 'completions/mean_length': 41.0, 'completions/min_length': 39.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.2913917303085327, 'rewards/meter/std': 0.36782386898994446, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5499999523162842, 'rewards/judge_quality/std': 0.21902382373809814, 'rewards/repeat_penalty/mean': 0.9766601324081421, 'rewards/repeat_penalty/std': 0.025797616690397263, 'rewards/repeat_floor/mean': 0.9969615936279297, 'rewards/repeat_floor/std': 0.0028235320933163166, 'rewards/near_duplicate_penalty/mean': 0.9845515489578247, 'rewards/near_duplicate_penalty/std': 0.013140684925019741, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9919871687889099, 'rewards/distinct_2/std': 0.02266368828713894, 'rewards/total_composite/mean': 0.19136007130146027, 'rewards/total_composite/std': 0.2899673581123352, 'reward': 0.19136007130146027, 'reward_std': 0.2899673581123352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14844121038913727, 'sampling/sampling_logp_difference/max': 1.8613018989562988, 'sampling/importance_sampling_ratio/min': 0.1554700881242752, 'sampling/importance_sampling_ratio/mean': 0.9931487441062927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6197232231497765, 'clip_ratio/low_mean': 0.10148030612617731, 'clip_ratio/low_min': 0.10148030612617731, 'clip_ratio/high_mean': 0.04374999925494194, 'clip_ratio/high_max': 0.04374999925494194, 'clip_ratio/region_mean': 0.14523030538111925, 'reward_total_mean': 0.19136007130146027, 'reward_meter_mean': 0.2913917303085327, 'reward_meter_std': 0.36782386898994446, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9766601324081421, 'reward_repeat_penalty_std': 0.025797616690397263, 'reward_repeat_floor_mean': 0.9969615936279297, 'reward_repeat_floor_std': 0.0028235320933163166, 'reward_near_duplicate_penalty_mean': 0.9845515489578247, 'reward_near_duplicate_penalty_std': 0.013140684925019741, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9919871687889099, 'reward_distinct_2_std': 0.02266368828713894, 'reward_judge_quality_mean': 0.5499999523162842, 'reward_judge_quality_std': 0.21902382373809814, 'reward_total_composite_mean': 0.19136007130146027, 'reward_total_composite_std': 0.2899673581123352, 'epoch': 0.06}

 34%|███▍      | 1138/3300 [4:21:10<7:01:04, 11.69s/it]INFO 04-12 20:19:36 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1139/3300 [4:21:24<7:30:40, 12.51s/it]2026-04-12 20:19:51,141 | INFO | train_grpo_train | metrics_logged mode=train step=1139

                                                       
{'loss': -0.0205, 'grad_norm': 5.675747394561768, 'learning_rate': 6.551515151515152e-06, 'num_tokens': 2030825.0, 'completions/mean_length': 113.625, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.71428680419922, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.46092361211776733, 'rewards/meter/std': 0.38770899176597595, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9608503580093384, 'rewards/lexical_plausibility/std': 0.11073198169469833, 'rewards/judge_quality/mean': 0.20625001192092896, 'rewards/judge_quality/std': 0.1590990275144577, 'rewards/repeat_penalty/mean': 0.8663142323493958, 'rewards/repeat_penalty/std': 0.17410963773727417, 'rewards/repeat_floor/mean': 0.9843921065330505, 'rewards/repeat_floor/std': 0.01966203935444355, 'rewards/near_duplicate_penalty/mean': 0.9548225998878479, 'rewards/near_duplicate_penalty/std': 0.04271745681762695, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9018747806549072, 'rewards/distinct_2/std': 0.15136343240737915, 'rewards/total_composite/mean': 0.10493719577789307, 'rewards/total_composite/std': 0.12584581971168518, 'reward': 0.10493719577789307, 'reward_std': 0.12584581971168518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17330564558506012, 'sampling/sampling_logp_difference/max': 2.0196781158447266, 'sampling/importance_sampling_ratio/min': 0.1326981782913208, 'sampling/importance_sampling_ratio/mean': 1.0193663835525513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0731084495782852, 'clip_ratio/low_mean': 0.0699875932186842, 'clip_ratio/low_min': 0.0699875932186842, 'clip_ratio/high_mean': 0.06027951929718256, 'clip_ratio/high_max': 0.06027951929718256, 'clip_ratio/region_mean': 0.13026711251586676, 'reward_total_mean': 0.10493719577789307, 'reward_meter_mean': 0.46092361211776733, 'reward_meter_std': 0.38770899176597595, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9608503580093384, 'reward_lexical_plausibility_std': 0.11073198169469833, 'reward_repeat_penalty_mean': 0.8663142323493958, 'reward_repeat_penalty_std': 0.17410963773727417, 'reward_repeat_floor_mean': 0.9843921065330505, 'reward_repeat_floor_std': 0.01966203935444355, 'reward_near_duplicate_penalty_mean': 0.9548225998878479, 'reward_near_duplicate_penalty_std': 0.04271745681762695, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9018747806549072, 'reward_distinct_2_std': 0.15136343240737915, 'reward_judge_quality_mean': 0.20625001192092896, 'reward_judge_quality_std': 0.1590990275144577, 'reward_total_composite_mean': 0.10493719577789307, 'reward_total_composite_std': 0.12584581971168518, 'epoch': 0.06}

 35%|███▍      | 1139/3300 [4:21:24<7:30:40, 12.51s/it]INFO 04-12 20:19:51 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1140/3300 [4:21:38<7:40:02, 12.78s/it]2026-04-12 20:20:04,538 | INFO | train_grpo_train | metrics_logged mode=train step=1140

                                                       
{'loss': -0.0675, 'grad_norm': 3.6204586029052734, 'learning_rate': 6.5484848484848494e-06, 'num_tokens': 2032225.0, 'completions/mean_length': 98.0, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.85714340209961, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6756739616394043, 'rewards/meter/std': 0.35370534658432007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9530335664749146, 'rewards/lexical_plausibility/std': 0.1328410506248474, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.33161941170692444, 'rewards/repeat_penalty/mean': 0.9558984041213989, 'rewards/repeat_penalty/std': 0.04338403418660164, 'rewards/repeat_floor/mean': 0.9937871694564819, 'rewards/repeat_floor/std': 0.006231467239558697, 'rewards/near_duplicate_penalty/mean': 0.9628546237945557, 'rewards/near_duplicate_penalty/std': 0.04140779376029968, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9928774833679199, 'rewards/distinct_2/std': 0.02014549821615219, 'rewards/total_composite/mean': 0.4560588598251343, 'rewards/total_composite/std': 0.2655254900455475, 'reward': 0.4560588598251343, 'reward_std': 0.2655254602432251, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1267271339893341, 'sampling/sampling_logp_difference/max': 1.193772315979004, 'sampling/importance_sampling_ratio/min': 0.30307579040527344, 'sampling/importance_sampling_ratio/mean': 1.0059844255447388, 'sampling/importance_sampling_ratio/max': 1.6562862396240234, 'entropy': 0.7207122519612312, 'clip_ratio/low_mean': 0.06068417150527239, 'clip_ratio/low_min': 0.06068417150527239, 'clip_ratio/high_mean': 0.040540540125221014, 'clip_ratio/high_max': 0.040540540125221014, 'clip_ratio/region_mean': 0.1012247116304934, 'reward_total_mean': 0.4560588598251343, 'reward_meter_mean': 0.6756739616394043, 'reward_meter_std': 0.35370534658432007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9530335664749146, 'reward_lexical_plausibility_std': 0.1328410506248474, 'reward_repeat_penalty_mean': 0.9558984041213989, 'reward_repeat_penalty_std': 0.04338403418660164, 'reward_repeat_floor_mean': 0.9937871694564819, 'reward_repeat_floor_std': 0.006231467239558697, 'reward_near_duplicate_penalty_mean': 0.9628546237945557, 'reward_near_duplicate_penalty_std': 0.04140779376029968, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9928774833679199, 'reward_distinct_2_std': 0.02014549821615219, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.33161941170692444, 'reward_total_composite_mean': 0.4560588598251343, 'reward_total_composite_std': 0.2655254900455475, 'epoch': 0.06}

 35%|███▍      | 1140/3300 [4:21:38<7:40:02, 12.78s/it]INFO 04-12 20:20:04 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1141/3300 [4:21:46<6:47:53, 11.34s/it]2026-04-12 20:20:12,504 | INFO | train_grpo_train | metrics_logged mode=train step=1141

                                                       
{'loss': -0.0148, 'grad_norm': 26.83353614807129, 'learning_rate': 6.545454545454546e-06, 'num_tokens': 2033633.0, 'completions/mean_length': 19.0, 'completions/min_length': 17.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7602664828300476, 'rewards/meter/std': 0.39526283740997314, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6687500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.7529481649398804, 'rewards/repeat_penalty/std': 0.11471517384052277, 'rewards/repeat_floor/mean': 0.9812146425247192, 'rewards/repeat_floor/std': 0.013616316951811314, 'rewards/near_duplicate_penalty/mean': 0.9622641801834106, 'rewards/near_duplicate_penalty/std': 0.07688961178064346, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.44564753770828247, 'rewards/total_composite/std': 0.37537670135498047, 'reward': 0.44564753770828247, 'reward_std': 0.37537670135498047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1680663675069809, 'sampling/sampling_logp_difference/max': 0.8450136184692383, 'sampling/importance_sampling_ratio/min': 0.42955151200294495, 'sampling/importance_sampling_ratio/mean': 1.0307961702346802, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.063039354979992, 'clip_ratio/low_mean': 0.061185440979897976, 'clip_ratio/low_min': 0.061185440979897976, 'clip_ratio/high_mean': 0.04417293332517147, 'clip_ratio/high_max': 0.04417293332517147, 'clip_ratio/region_mean': 0.10535837430506945, 'reward_total_mean': 0.44564753770828247, 'reward_meter_mean': 0.7602664828300476, 'reward_meter_std': 0.39526283740997314, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7529481649398804, 'reward_repeat_penalty_std': 0.11471517384052277, 'reward_repeat_floor_mean': 0.9812146425247192, 'reward_repeat_floor_std': 0.013616316951811314, 'reward_near_duplicate_penalty_mean': 0.9622641801834106, 'reward_near_duplicate_penalty_std': 0.07688961178064346, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6687500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.44564753770828247, 'reward_total_composite_std': 0.37537670135498047, 'epoch': 0.06}

 35%|███▍      | 1141/3300 [4:21:46<6:47:53, 11.34s/it]INFO 04-12 20:20:12 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1142/3300 [4:22:01<7:27:06, 12.43s/it]2026-04-12 20:20:27,504 | INFO | train_grpo_train | metrics_logged mode=train step=1142

                                                       
{'loss': -0.0235, 'grad_norm': 5.856141090393066, 'learning_rate': 6.542424242424243e-06, 'num_tokens': 2035919.0, 'completions/mean_length': 148.75, 'completions/min_length': 90.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 96.85714721679688, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.7262279987335205, 'rewards/meter/std': 0.29544880986213684, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9142857193946838, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9561619758605957, 'rewards/lexical_plausibility/std': 0.12399259954690933, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.8861979842185974, 'rewards/repeat_penalty/std': 0.06992067396640778, 'rewards/repeat_floor/mean': 0.9877134561538696, 'rewards/repeat_floor/std': 0.007540726102888584, 'rewards/near_duplicate_penalty/mean': 0.9708187580108643, 'rewards/near_duplicate_penalty/std': 0.017901204526424408, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9121188521385193, 'rewards/distinct_2/std': 0.05898182839155197, 'rewards/total_composite/mean': 0.18572500348091125, 'rewards/total_composite/std': 0.08685710281133652, 'reward': 0.18572500348091125, 'reward_std': 0.08685710281133652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.136882483959198, 'sampling/sampling_logp_difference/max': 2.890528678894043, 'sampling/importance_sampling_ratio/min': 0.0555468387901783, 'sampling/importance_sampling_ratio/mean': 1.0035995244979858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6375368684530258, 'clip_ratio/low_mean': 0.05462704598903656, 'clip_ratio/low_min': 0.05462704598903656, 'clip_ratio/high_mean': 0.054704878479242325, 'clip_ratio/high_max': 0.054704878479242325, 'clip_ratio/region_mean': 0.10933192446827888, 'reward_total_mean': 0.18572500348091125, 'reward_meter_mean': 0.7262279987335205, 'reward_meter_std': 0.29544880986213684, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9142857193946838, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9561619758605957, 'reward_lexical_plausibility_std': 0.12399259954690933, 'reward_repeat_penalty_mean': 0.8861979842185974, 'reward_repeat_penalty_std': 0.06992067396640778, 'reward_repeat_floor_mean': 0.9877134561538696, 'reward_repeat_floor_std': 0.007540726102888584, 'reward_near_duplicate_penalty_mean': 0.9708187580108643, 'reward_near_duplicate_penalty_std': 0.017901204526424408, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9121188521385193, 'reward_distinct_2_std': 0.05898182839155197, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.18572500348091125, 'reward_total_composite_std': 0.08685710281133652, 'epoch': 0.06}

 35%|███▍      | 1142/3300 [4:22:01<7:27:06, 12.43s/it]INFO 04-12 20:20:27 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1143/3300 [4:22:15<7:46:17, 12.97s/it]2026-04-12 20:20:41,738 | INFO | train_grpo_train | metrics_logged mode=train step=1143

                                                       
{'loss': -0.0278, 'grad_norm': 6.398889064788818, 'learning_rate': 6.5393939393939395e-06, 'num_tokens': 2037622.0, 'completions/mean_length': 110.875, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.57143020629883, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.4308423399925232, 'rewards/meter/std': 0.33540067076683044, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.2086307406425476, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.06258922815322876, 'rewards/lexical_plausibility/mean': 0.9597708582878113, 'rewards/lexical_plausibility/std': 0.11378522217273712, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.33114519715309143, 'rewards/repeat_penalty/mean': 0.9547795057296753, 'rewards/repeat_penalty/std': 0.08326976746320724, 'rewards/repeat_floor/mean': 0.9960293769836426, 'rewards/repeat_floor/std': 0.004671937320381403, 'rewards/near_duplicate_penalty/mean': 0.9860295057296753, 'rewards/near_duplicate_penalty/std': 0.010920105502009392, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22858132421970367, 'rewards/total_composite/std': 0.23406212031841278, 'reward': 0.22858132421970367, 'reward_std': 0.2340621054172516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17115718126296997, 'sampling/sampling_logp_difference/max': 1.98124361038208, 'sampling/importance_sampling_ratio/min': 0.13789764046669006, 'sampling/importance_sampling_ratio/mean': 1.0093872547149658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6973089054226875, 'clip_ratio/low_mean': 0.04788055643439293, 'clip_ratio/low_min': 0.04788055643439293, 'clip_ratio/high_mean': 0.0812247795984149, 'clip_ratio/high_max': 0.0812247795984149, 'clip_ratio/region_mean': 0.12910533603280783, 'reward_total_mean': 0.22858132421970367, 'reward_meter_mean': 0.4308423399925232, 'reward_meter_std': 0.33540067076683044, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.2086307406425476, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.06258922815322876, 'reward_lexical_plausibility_mean': 0.9597708582878113, 'reward_lexical_plausibility_std': 0.11378522217273712, 'reward_repeat_penalty_mean': 0.9547795057296753, 'reward_repeat_penalty_std': 0.08326976746320724, 'reward_repeat_floor_mean': 0.9960293769836426, 'reward_repeat_floor_std': 0.004671937320381403, 'reward_near_duplicate_penalty_mean': 0.9860295057296753, 'reward_near_duplicate_penalty_std': 0.010920105502009392, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.33114519715309143, 'reward_total_composite_mean': 0.22858132421970367, 'reward_total_composite_std': 0.23406212031841278, 'epoch': 0.06}

 35%|███▍      | 1143/3300 [4:22:15<7:46:17, 12.97s/it]INFO 04-12 20:20:42 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1144/3300 [4:22:31<8:15:14, 13.78s/it]2026-04-12 20:20:57,402 | INFO | train_grpo_train | metrics_logged mode=train step=1144

                                                       
{'loss': -0.0509, 'grad_norm': 1.6195006370544434, 'learning_rate': 6.536363636363638e-06, 'num_tokens': 2038988.0, 'completions/mean_length': 335.75, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.25796008110046387, 'rewards/meter/std': 0.3918321430683136, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.11160357296466827, 'rewards/lexical_plausibility/mean': 0.7385250926017761, 'rewards/lexical_plausibility/std': 0.21849527955055237, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.18600594997406006, 'rewards/repeat_penalty/mean': 0.8957667350769043, 'rewards/repeat_penalty/std': 0.10599631071090698, 'rewards/repeat_floor/mean': 0.9909473657608032, 'rewards/repeat_floor/std': 0.00767924590036273, 'rewards/near_duplicate_penalty/mean': 0.976216197013855, 'rewards/near_duplicate_penalty/std': 0.029041754081845284, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9807209968566895, 'rewards/distinct_2/std': 0.0357571542263031, 'rewards/total_composite/mean': 0.07311014831066132, 'rewards/total_composite/std': 0.12083852291107178, 'reward': 0.07311014831066132, 'reward_std': 0.12083853036165237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2061946988105774, 'sampling/sampling_logp_difference/max': 1.6785106658935547, 'sampling/importance_sampling_ratio/min': 0.18665175139904022, 'sampling/importance_sampling_ratio/mean': 1.0507837533950806, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5393520146608353, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.04928728099912405, 'clip_ratio/high_max': 0.04928728099912405, 'clip_ratio/region_mean': 0.04928728099912405, 'reward_total_mean': 0.07311014831066132, 'reward_meter_mean': 0.25796008110046387, 'reward_meter_std': 0.3918321430683136, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.3720119297504425, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.11160357296466827, 'reward_lexical_plausibility_mean': 0.7385250926017761, 'reward_lexical_plausibility_std': 0.21849527955055237, 'reward_repeat_penalty_mean': 0.8957667350769043, 'reward_repeat_penalty_std': 0.10599631071090698, 'reward_repeat_floor_mean': 0.9909473657608032, 'reward_repeat_floor_std': 0.00767924590036273, 'reward_near_duplicate_penalty_mean': 0.976216197013855, 'reward_near_duplicate_penalty_std': 0.029041754081845284, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9807209968566895, 'reward_distinct_2_std': 0.0357571542263031, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.18600594997406006, 'reward_total_composite_mean': 0.07311014831066132, 'reward_total_composite_std': 0.12083852291107178, 'epoch': 0.06}

 35%|███▍      | 1144/3300 [4:22:31<8:15:14, 13.78s/it]INFO 04-12 20:20:57 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1145/3300 [4:22:39<7:15:10, 12.12s/it]2026-04-12 20:21:05,629 | INFO | train_grpo_train | metrics_logged mode=train step=1145

                                                       
{'loss': 0.058, 'grad_norm': 18.354543685913086, 'learning_rate': 6.533333333333334e-06, 'num_tokens': 2040303.0, 'completions/mean_length': 23.375, 'completions/min_length': 21.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.8853771686553955, 'rewards/meter/std': 0.17064696550369263, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.7260016798973083, 'rewards/repeat_penalty/std': 0.06787750124931335, 'rewards/repeat_floor/mean': 0.9808095693588257, 'rewards/repeat_floor/std': 0.011852389201521873, 'rewards/near_duplicate_penalty/mean': 0.9816790223121643, 'rewards/near_duplicate_penalty/std': 0.05181950703263283, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.6388901472091675, 'rewards/total_composite/std': 0.27662715315818787, 'reward': 0.6388901472091675, 'reward_std': 0.27662715315818787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14900989830493927, 'sampling/sampling_logp_difference/max': 1.2409031391143799, 'sampling/importance_sampling_ratio/min': 0.289122998714447, 'sampling/importance_sampling_ratio/mean': 1.0068740844726562, 'sampling/importance_sampling_ratio/max': 1.7706222534179688, 'entropy': 0.8659064024686813, 'clip_ratio/low_mean': 0.06323369592428207, 'clip_ratio/low_min': 0.06323369592428207, 'clip_ratio/high_mean': 0.06437229551374912, 'clip_ratio/high_max': 0.06437229551374912, 'clip_ratio/region_mean': 0.1276059914380312, 'reward_total_mean': 0.6388901472091675, 'reward_meter_mean': 0.8853771686553955, 'reward_meter_std': 0.17064696550369263, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7260016798973083, 'reward_repeat_penalty_std': 0.06787750124931335, 'reward_repeat_floor_mean': 0.9808095693588257, 'reward_repeat_floor_std': 0.011852389201521873, 'reward_near_duplicate_penalty_mean': 0.9816790223121643, 'reward_near_duplicate_penalty_std': 0.05181950703263283, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.6388901472091675, 'reward_total_composite_std': 0.27662715315818787, 'epoch': 0.06}

 35%|███▍      | 1145/3300 [4:22:39<7:15:10, 12.12s/it]INFO 04-12 20:21:05 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1146/3300 [4:22:53<7:37:46, 12.75s/it]2026-04-12 20:21:19,860 | INFO | train_grpo_train | metrics_logged mode=train step=1146

                                                       
{'loss': -0.0486, 'grad_norm': 3.964550495147705, 'learning_rate': 6.530303030303031e-06, 'num_tokens': 2041793.0, 'completions/mean_length': 158.25, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 40.333335876464844, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.20266953110694885, 'rewards/meter/std': 0.2659032344818115, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.921892523765564, 'rewards/lexical_plausibility/std': 0.1446504145860672, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1767766773700714, 'rewards/repeat_penalty/mean': 0.9146501421928406, 'rewards/repeat_penalty/std': 0.11435196548700333, 'rewards/repeat_floor/mean': 0.9908194541931152, 'rewards/repeat_floor/std': 0.010582190938293934, 'rewards/near_duplicate_penalty/mean': 0.9662138223648071, 'rewards/near_duplicate_penalty/std': 0.04272926598787308, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9751373529434204, 'rewards/distinct_2/std': 0.034471429884433746, 'rewards/total_composite/mean': 0.07987114042043686, 'rewards/total_composite/std': 0.12118276208639145, 'reward': 0.07987114042043686, 'reward_std': 0.12118275463581085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15076430141925812, 'sampling/sampling_logp_difference/max': 1.9134538173675537, 'sampling/importance_sampling_ratio/min': 0.14756983518600464, 'sampling/importance_sampling_ratio/mean': 1.0418813228607178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8976794555783272, 'clip_ratio/low_mean': 0.07418129313737154, 'clip_ratio/low_min': 0.07418129313737154, 'clip_ratio/high_mean': 0.02325581479817629, 'clip_ratio/high_max': 0.02325581479817629, 'clip_ratio/region_mean': 0.09743710793554783, 'reward_total_mean': 0.07987114042043686, 'reward_meter_mean': 0.20266953110694885, 'reward_meter_std': 0.2659032344818115, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.921892523765564, 'reward_lexical_plausibility_std': 0.1446504145860672, 'reward_repeat_penalty_mean': 0.9146501421928406, 'reward_repeat_penalty_std': 0.11435196548700333, 'reward_repeat_floor_mean': 0.9908194541931152, 'reward_repeat_floor_std': 0.010582190938293934, 'reward_near_duplicate_penalty_mean': 0.9662138223648071, 'reward_near_duplicate_penalty_std': 0.04272926598787308, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9751373529434204, 'reward_distinct_2_std': 0.034471429884433746, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1767766773700714, 'reward_total_composite_mean': 0.07987114042043686, 'reward_total_composite_std': 0.12118276208639145, 'epoch': 0.06}

 35%|███▍      | 1146/3300 [4:22:53<7:37:46, 12.75s/it]INFO 04-12 20:21:20 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1147/3300 [4:23:07<7:53:55, 13.21s/it]2026-04-12 20:21:34,141 | INFO | train_grpo_train | metrics_logged mode=train step=1147

                                                       
{'loss': -0.0959, 'grad_norm': 4.084112167358398, 'learning_rate': 6.527272727272728e-06, 'num_tokens': 2043591.0, 'completions/mean_length': 117.75, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.42857360839844, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7658863663673401, 'rewards/meter/std': 0.31261318922042847, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 0.9434326887130737, 'rewards/lexical_plausibility/std': 0.15999643504619598, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8328726291656494, 'rewards/repeat_penalty/std': 0.13319075107574463, 'rewards/repeat_floor/mean': 0.9800093173980713, 'rewards/repeat_floor/std': 0.015447062440216541, 'rewards/near_duplicate_penalty/mean': 0.9336251020431519, 'rewards/near_duplicate_penalty/std': 0.039976511150598526, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8885061144828796, 'rewards/distinct_2/std': 0.11086546629667282, 'rewards/total_composite/mean': 0.24992404878139496, 'rewards/total_composite/std': 0.14261551201343536, 'reward': 0.24992404878139496, 'reward_std': 0.14261551201343536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14335353672504425, 'sampling/sampling_logp_difference/max': 1.9305111169815063, 'sampling/importance_sampling_ratio/min': 0.1450740396976471, 'sampling/importance_sampling_ratio/mean': 1.0164743661880493, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7771314010024071, 'clip_ratio/low_mean': 0.033865710720419884, 'clip_ratio/low_min': 0.033865710720419884, 'clip_ratio/high_mean': 0.09077356196939945, 'clip_ratio/high_max': 0.09077356196939945, 'clip_ratio/region_mean': 0.12463927268981934, 'reward_total_mean': 0.24992404878139496, 'reward_meter_mean': 0.7658863663673401, 'reward_meter_std': 0.31261318922042847, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 0.9434326887130737, 'reward_lexical_plausibility_std': 0.15999643504619598, 'reward_repeat_penalty_mean': 0.8328726291656494, 'reward_repeat_penalty_std': 0.13319075107574463, 'reward_repeat_floor_mean': 0.9800093173980713, 'reward_repeat_floor_std': 0.015447062440216541, 'reward_near_duplicate_penalty_mean': 0.9336251020431519, 'reward_near_duplicate_penalty_std': 0.039976511150598526, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8885061144828796, 'reward_distinct_2_std': 0.11086546629667282, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.24992404878139496, 'reward_total_composite_std': 0.14261551201343536, 'epoch': 0.06}

 35%|███▍      | 1147/3300 [4:23:07<7:53:55, 13.21s/it]INFO 04-12 20:21:34 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1148/3300 [4:23:22<8:04:03, 13.50s/it]2026-04-12 20:21:48,310 | INFO | train_grpo_train | metrics_logged mode=train step=1148

                                                       
{'loss': 0.0164, 'grad_norm': 6.501523971557617, 'learning_rate': 6.524242424242425e-06, 'num_tokens': 2045448.0, 'completions/mean_length': 126.125, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.6120268702507019, 'rewards/meter/std': 0.290813684463501, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.9514014720916748, 'rewards/lexical_plausibility/std': 0.137457475066185, 'rewards/judge_quality/mean': 0.29999998211860657, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.8905959129333496, 'rewards/repeat_penalty/std': 0.0741792693734169, 'rewards/repeat_floor/mean': 0.9868966937065125, 'rewards/repeat_floor/std': 0.007907633669674397, 'rewards/near_duplicate_penalty/mean': 0.9514176845550537, 'rewards/near_duplicate_penalty/std': 0.017901966348290443, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9353775978088379, 'rewards/distinct_2/std': 0.06565867364406586, 'rewards/total_composite/mean': 0.16292856633663177, 'rewards/total_composite/std': 0.12048143148422241, 'reward': 0.16292856633663177, 'reward_std': 0.12048143893480301, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17229385673999786, 'sampling/sampling_logp_difference/max': 2.7195560932159424, 'sampling/importance_sampling_ratio/min': 0.06590400636196136, 'sampling/importance_sampling_ratio/mean': 1.000677466392517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7999205216765404, 'clip_ratio/low_mean': 0.07672486081719398, 'clip_ratio/low_min': 0.07672486081719398, 'clip_ratio/high_mean': 0.0744554940611124, 'clip_ratio/high_max': 0.0744554940611124, 'clip_ratio/region_mean': 0.1511803548783064, 'reward_total_mean': 0.16292856633663177, 'reward_meter_mean': 0.6120268702507019, 'reward_meter_std': 0.290813684463501, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.9514014720916748, 'reward_lexical_plausibility_std': 0.137457475066185, 'reward_repeat_penalty_mean': 0.8905959129333496, 'reward_repeat_penalty_std': 0.0741792693734169, 'reward_repeat_floor_mean': 0.9868966937065125, 'reward_repeat_floor_std': 0.007907633669674397, 'reward_near_duplicate_penalty_mean': 0.9514176845550537, 'reward_near_duplicate_penalty_std': 0.017901966348290443, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9353775978088379, 'reward_distinct_2_std': 0.06565867364406586, 'reward_judge_quality_mean': 0.29999998211860657, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.16292856633663177, 'reward_total_composite_std': 0.12048143148422241, 'epoch': 0.06}

 35%|███▍      | 1148/3300 [4:23:22<8:04:03, 13.50s/it]INFO 04-12 20:21:48 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1149/3300 [4:23:31<7:15:32, 12.15s/it]2026-04-12 20:21:57,318 | INFO | train_grpo_train | metrics_logged mode=train step=1149

                                                       
{'loss': 0.0612, 'grad_norm': 14.13011646270752, 'learning_rate': 6.521212121212121e-06, 'num_tokens': 2047024.0, 'completions/mean_length': 37.0, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.37118369340896606, 'rewards/meter/std': 0.35734549164772034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.889718770980835, 'rewards/repeat_penalty/std': 0.06981737911701202, 'rewards/repeat_floor/mean': 0.9864054322242737, 'rewards/repeat_floor/std': 0.007891464978456497, 'rewards/near_duplicate_penalty/mean': 0.9447933435440063, 'rewards/near_duplicate_penalty/std': 0.025792958214879036, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9409602880477905, 'rewards/distinct_2/std': 0.056140556931495667, 'rewards/total_composite/mean': 0.15181592106819153, 'rewards/total_composite/std': 0.13855406641960144, 'reward': 0.15181592106819153, 'reward_std': 0.13855406641960144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13481339812278748, 'sampling/sampling_logp_difference/max': 1.7303745746612549, 'sampling/importance_sampling_ratio/min': 0.1772180199623108, 'sampling/importance_sampling_ratio/mean': 1.0263570547103882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8146373778581619, 'clip_ratio/low_mean': 0.05393212754279375, 'clip_ratio/low_min': 0.05393212754279375, 'clip_ratio/high_mean': 0.05444540362805128, 'clip_ratio/high_max': 0.05444540362805128, 'clip_ratio/region_mean': 0.10837753117084503, 'reward_total_mean': 0.15181592106819153, 'reward_meter_mean': 0.37118369340896606, 'reward_meter_std': 0.35734549164772034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.889718770980835, 'reward_repeat_penalty_std': 0.06981737911701202, 'reward_repeat_floor_mean': 0.9864054322242737, 'reward_repeat_floor_std': 0.007891464978456497, 'reward_near_duplicate_penalty_mean': 0.9447933435440063, 'reward_near_duplicate_penalty_std': 0.025792958214879036, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9409602880477905, 'reward_distinct_2_std': 0.056140556931495667, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.15181592106819153, 'reward_total_composite_std': 0.13855406641960144, 'epoch': 0.06}

 35%|███▍      | 1149/3300 [4:23:31<7:15:32, 12.15s/it]INFO 04-12 20:21:57 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1150/3300 [4:23:39<6:33:54, 10.99s/it]2026-04-12 20:22:05,603 | INFO | train_grpo_train | metrics_logged mode=train step=1150

                                                       
{'loss': 0.0474, 'grad_norm': 21.6728458404541, 'learning_rate': 6.5181818181818195e-06, 'num_tokens': 2048574.0, 'completions/mean_length': 25.75, 'completions/min_length': 19.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8422211408615112, 'rewards/meter/std': 0.30025631189346313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.23279361426830292, 'rewards/repeat_penalty/mean': 0.8518966436386108, 'rewards/repeat_penalty/std': 0.2523607313632965, 'rewards/repeat_floor/mean': 0.9817835092544556, 'rewards/repeat_floor/std': 0.026638295501470566, 'rewards/near_duplicate_penalty/mean': 0.9225249886512756, 'rewards/near_duplicate_penalty/std': 0.05797421932220459, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9475524425506592, 'rewards/distinct_2/std': 0.14834408462047577, 'rewards/total_composite/mean': 0.45890873670578003, 'rewards/total_composite/std': 0.24174794554710388, 'reward': 0.45890873670578003, 'reward_std': 0.24174796044826508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1397968828678131, 'sampling/sampling_logp_difference/max': 1.3089745044708252, 'sampling/importance_sampling_ratio/min': 0.27009689807891846, 'sampling/importance_sampling_ratio/mean': 1.0170339345932007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5876163616776466, 'clip_ratio/low_mean': 0.10471547488123178, 'clip_ratio/low_min': 0.10471547488123178, 'clip_ratio/high_mean': 0.04418103490024805, 'clip_ratio/high_max': 0.04418103490024805, 'clip_ratio/region_mean': 0.14889650978147984, 'reward_total_mean': 0.45890873670578003, 'reward_meter_mean': 0.8422211408615112, 'reward_meter_std': 0.30025631189346313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8518966436386108, 'reward_repeat_penalty_std': 0.2523607313632965, 'reward_repeat_floor_mean': 0.9817835092544556, 'reward_repeat_floor_std': 0.026638295501470566, 'reward_near_duplicate_penalty_mean': 0.9225249886512756, 'reward_near_duplicate_penalty_std': 0.05797421932220459, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9475524425506592, 'reward_distinct_2_std': 0.14834408462047577, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.23279361426830292, 'reward_total_composite_mean': 0.45890873670578003, 'reward_total_composite_std': 0.24174794554710388, 'epoch': 0.06}

 35%|███▍      | 1150/3300 [4:23:39<6:33:54, 10.99s/it]INFO 04-12 20:22:05 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:11<01:05,  5.96s/it]

 23%|██▎       | 3/13 [00:17<00:58,  5.85s/it]

 31%|███       | 4/13 [00:29<01:12,  8.06s/it]

 38%|███▊      | 5/13 [00:41<01:15,  9.43s/it]

 46%|████▌     | 6/13 [00:52<01:10, 10.04s/it]

 54%|█████▍    | 7/13 [00:59<00:53,  8.97s/it]

 62%|██████▏   | 8/13 [01:10<00:49,  9.80s/it]

 69%|██████▉   | 9/13 [01:22<00:41, 10.39s/it]

 77%|███████▋  | 10/13 [01:34<00:32, 10.90s/it]

 85%|████████▍ | 11/13 [01:46<00:22, 11.02s/it]

 92%|█████████▏| 12/13 [01:53<00:10, 10.05s/it]

100%|██████████| 13/13 [02:01<00:00,  9.22s/it]2026-04-12 20:24:18,678 | INFO | train_grpo_train | metrics_logged mode=eval step=1150

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 132.9701, 'eval_samples_per_second': 0.782, 'eval_steps_per_second': 0.098, 'eval_num_tokens': 2048574.0, 'eval_completions/mean_length': 152.90384615384616, 'eval_completions/min_length': 33.46153846153846, 'eval_completions/max_length': 407.9230769230769, 'eval_completions/clipped_ratio': 0.14423076923076922, 'eval_completions/mean_terminated_length': 93.38425093430739, 'eval_completions/min_terminated_length': 33.46153846153846, 'eval_completions/max_terminated_length': 184.3846153846154, 'eval_rewards/meter/mean': 0.46407373593403745, 'eval_rewards/meter/std': 0.3627307415008545, 'eval_rewards/count_adherence/mean': 0.7267556465589083, 'eval_rewards/count_adherence/std': 0.17637487959403259, 'eval_rewards/arabic_clean/mean': 0.7596153846153846, 'eval_rewards/arabic_clean/std': 0.38037164165423465, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9951923076923077, 'eval_rewards/arabic_floor/std': 0.013598207670908708, 'eval_rewards/count_floor/mean': 0.918026694884667, 'eval_rewards/count_floor/std': 0.052912466801129855, 'eval_rewards/lexical_plausibility/mean': 0.9519930344361526, 'eval_rewards/lexical_plausibility/std': 0.09474431837980564, 'eval_rewards/judge_quality/mean': 0.29019230948044705, 'eval_rewards/judge_quality/std': 0.17068963727125755, 'eval_rewards/repeat_penalty/mean': 0.7603715383089505, 'eval_rewards/repeat_penalty/std': 0.2201156782416197, 'eval_rewards/repeat_floor/mean': 0.9729670607126676, 'eval_rewards/repeat_floor/std': 0.025789915440747373, 'eval_rewards/near_duplicate_penalty/mean': 0.9284641788556025, 'eval_rewards/near_duplicate_penalty/std': 0.06080597653411902, 'eval_rewards/opening_diversity/mean': 0.9599762604786799, 'eval_rewards/opening_diversity/std': 0.09110712345976096, 'eval_rewards/distinct_2/mean': 0.8455428251853356, 'eval_rewards/distinct_2/std': 0.18207372237856573, 'eval_rewards/total_composite/mean': 0.1142805413557933, 'eval_rewards/total_composite/std': 0.11712601465674546, 'eval_reward': 0.1142805413557933, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07252842159225391, 'eval_sampling/sampling_logp_difference/max': 1.0427780884962816, 'eval_sampling/importance_sampling_ratio/min': 0.3573741637743436, 'eval_sampling/importance_sampling_ratio/mean': 1.0154054898482103, 'eval_sampling/importance_sampling_ratio/max': 1.577569191272442, 'eval_entropy': 0.797071649478032, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1142805413557933, 'eval_reward_meter_mean': 0.46407373593403745, 'eval_reward_meter_std': 0.3627307415008545, 'eval_reward_count_adherence_mean': 0.7267556465589083, 'eval_reward_count_adherence_std': 0.17637487959403259, 'eval_reward_arabic_clean_mean': 0.7596153846153846, 'eval_reward_arabic_clean_std': 0.38037164165423465, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9951923076923077, 'eval_reward_arabic_floor_std': 0.013598207670908708, 'eval_reward_count_floor_mean': 0.918026694884667, 'eval_reward_count_floor_std': 0.052912466801129855, 'eval_reward_lexical_plausibility_mean': 0.9519930344361526, 'eval_reward_lexical_plausibility_std': 0.09474431837980564, 'eval_reward_repeat_penalty_mean': 0.7603715383089505, 'eval_reward_repeat_penalty_std': 0.2201156782416197, 'eval_reward_repeat_floor_mean': 0.9729670607126676, 'eval_reward_repeat_floor_std': 0.025789915440747373, 'eval_reward_near_duplicate_penalty_mean': 0.9284641788556025, 'eval_reward_near_duplicate_penalty_std': 0.06080597653411902, 'eval_reward_opening_diversity_mean': 0.9599762604786799, 'eval_reward_opening_diversity_std': 0.09110712345976096, 'eval_reward_distinct_2_mean': 0.8455428251853356, 'eval_reward_distinct_2_std': 0.18207372237856573, 'eval_reward_judge_quality_mean': 0.29019230948044705, 'eval_reward_judge_quality_std': 0.17068963727125755, 'eval_reward_total_composite_mean': 0.1142805413557933, 'eval_reward_total_composite_std': 0.11712601465674546, 'epoch': 0.06}

 35%|███▍      | 1150/3300 [4:25:52<6:33:54, 10.99s/it]

100%|██████████| 13/13 [02:01<00:00,  9.22s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 35%|███▍      | 1151/3300 [4:26:09<31:27:31, 52.70s/it]2026-04-12 20:24:35,622 | INFO | train_grpo_train | metrics_logged mode=train step=1151

                                                        
{'loss': -0.0616, 'grad_norm': 5.99368953704834, 'learning_rate': 6.515151515151516e-06, 'num_tokens': 2050940.0, 'completions/mean_length': 207.75, 'completions/min_length': 96.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 106.33333587646484, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.7844904661178589, 'rewards/meter/std': 0.3082217872142792, 'rewards/count_adherence/mean': 0.453125, 'rewards/count_adherence/std': 0.09300298243761063, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8359375, 'rewards/count_floor/std': 0.02790091000497341, 'rewards/lexical_plausibility/mean': 0.9240038990974426, 'rewards/lexical_plausibility/std': 0.14556916058063507, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.6493626236915588, 'rewards/repeat_penalty/std': 0.14069417119026184, 'rewards/repeat_floor/mean': 0.9627324938774109, 'rewards/repeat_floor/std': 0.015762509778141975, 'rewards/near_duplicate_penalty/mean': 0.901320219039917, 'rewards/near_duplicate_penalty/std': 0.04448346421122551, 'rewards/opening_diversity/mean': 0.8636904954910278, 'rewards/opening_diversity/std': 0.18898972868919373, 'rewards/distinct_2/mean': 0.8412556648254395, 'rewards/distinct_2/std': 0.10836755484342575, 'rewards/total_composite/mean': 0.09898905456066132, 'rewards/total_composite/std': 0.08719787746667862, 'reward': 0.09898905456066132, 'reward_std': 0.08719787746667862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15414954721927643, 'sampling/sampling_logp_difference/max': 2.5807323455810547, 'sampling/importance_sampling_ratio/min': 0.07571852952241898, 'sampling/importance_sampling_ratio/mean': 1.0037548542022705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5894298553466797, 'clip_ratio/low_mean': 0.03150651790201664, 'clip_ratio/low_min': 0.03150651790201664, 'clip_ratio/high_mean': 0.0673754969611764, 'clip_ratio/high_max': 0.0673754969611764, 'clip_ratio/region_mean': 0.09888201486319304, 'reward_total_mean': 0.09898905456066132, 'reward_meter_mean': 0.7844904661178589, 'reward_meter_std': 0.3082217872142792, 'reward_count_adherence_mean': 0.453125, 'reward_count_adherence_std': 0.09300298243761063, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8359375, 'reward_count_floor_std': 0.02790091000497341, 'reward_lexical_plausibility_mean': 0.9240038990974426, 'reward_lexical_plausibility_std': 0.14556916058063507, 'reward_repeat_penalty_mean': 0.6493626236915588, 'reward_repeat_penalty_std': 0.14069417119026184, 'reward_repeat_floor_mean': 0.9627324938774109, 'reward_repeat_floor_std': 0.015762509778141975, 'reward_near_duplicate_penalty_mean': 0.901320219039917, 'reward_near_duplicate_penalty_std': 0.04448346421122551, 'reward_opening_diversity_mean': 0.8636904954910278, 'reward_opening_diversity_std': 0.18898972868919373, 'reward_distinct_2_mean': 0.8412556648254395, 'reward_distinct_2_std': 0.10836755484342575, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.09898905456066132, 'reward_total_composite_std': 0.08719787746667862, 'epoch': 0.06}

 35%|███▍      | 1151/3300 [4:26:09<31:27:31, 52.70s/it]INFO 04-12 20:24:35 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1152/3300 [4:26:24<24:39:28, 41.33s/it]2026-04-12 20:24:50,414 | INFO | train_grpo_train | metrics_logged mode=train step=1152

                                                        
{'loss': -0.1078, 'grad_norm': 2.2056760787963867, 'learning_rate': 6.512121212121213e-06, 'num_tokens': 2053431.0, 'completions/mean_length': 415.375, 'completions/min_length': 121.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 125.5, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.630567729473114, 'rewards/meter/std': 0.3366723656654358, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0839921236038208, 'rewards/arabic_clean/mean': 0.25, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.02519763633608818, 'rewards/lexical_plausibility/mean': 0.7459694743156433, 'rewards/lexical_plausibility/std': 0.16210918128490448, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.8843338489532471, 'rewards/repeat_penalty/std': 0.08120060712099075, 'rewards/repeat_floor/mean': 0.9852336645126343, 'rewards/repeat_floor/std': 0.009063991717994213, 'rewards/near_duplicate_penalty/mean': 0.9262511730194092, 'rewards/near_duplicate_penalty/std': 0.033649466931819916, 'rewards/opening_diversity/mean': 0.979687511920929, 'rewards/opening_diversity/std': 0.03892387077212334, 'rewards/distinct_2/mean': 0.9723861217498779, 'rewards/distinct_2/std': 0.027408117428421974, 'rewards/total_composite/mean': 0.08911535143852234, 'rewards/total_composite/std': 0.09778188169002533, 'reward': 0.08911535143852234, 'reward_std': 0.09778188914060593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13808836042881012, 'sampling/sampling_logp_difference/max': 1.8025791645050049, 'sampling/importance_sampling_ratio/min': 0.16487310826778412, 'sampling/importance_sampling_ratio/mean': 1.0155705213546753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24934573471546173, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03575969487428665, 'clip_ratio/high_max': 0.03575969487428665, 'clip_ratio/region_mean': 0.03575969487428665, 'reward_total_mean': 0.08911535143852234, 'reward_meter_mean': 0.630567729473114, 'reward_meter_std': 0.3366723656654358, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0839921236038208, 'reward_arabic_clean_mean': 0.25, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.02519763633608818, 'reward_lexical_plausibility_mean': 0.7459694743156433, 'reward_lexical_plausibility_std': 0.16210918128490448, 'reward_repeat_penalty_mean': 0.8843338489532471, 'reward_repeat_penalty_std': 0.08120060712099075, 'reward_repeat_floor_mean': 0.9852336645126343, 'reward_repeat_floor_std': 0.009063991717994213, 'reward_near_duplicate_penalty_mean': 0.9262511730194092, 'reward_near_duplicate_penalty_std': 0.033649466931819916, 'reward_opening_diversity_mean': 0.979687511920929, 'reward_opening_diversity_std': 0.03892387077212334, 'reward_distinct_2_mean': 0.9723861217498779, 'reward_distinct_2_std': 0.027408117428421974, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.08911535143852234, 'reward_total_composite_std': 0.09778188169002533, 'epoch': 0.06}

 35%|███▍      | 1152/3300 [4:26:24<24:39:28, 41.33s/it]INFO 04-12 20:24:50 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1153/3300 [4:26:32<18:49:51, 31.58s/it]2026-04-12 20:24:59,237 | INFO | train_grpo_train | metrics_logged mode=train step=1153

                                                        
{'loss': 0.0266, 'grad_norm': 14.970558166503906, 'learning_rate': 6.5090909090909095e-06, 'num_tokens': 2054875.0, 'completions/mean_length': 33.5, 'completions/min_length': 29.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.3699997365474701, 'rewards/meter/std': 0.3258172869682312, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.7188122272491455, 'rewards/repeat_penalty/std': 0.19269207119941711, 'rewards/repeat_floor/mean': 0.968213677406311, 'rewards/repeat_floor/std': 0.020962757989764214, 'rewards/near_duplicate_penalty/mean': 0.8991570472717285, 'rewards/near_duplicate_penalty/std': 0.05692235380411148, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.8982231616973877, 'rewards/distinct_2/std': 0.083673395216465, 'rewards/total_composite/mean': 0.13423895835876465, 'rewards/total_composite/std': 0.12630689144134521, 'reward': 0.13423895835876465, 'reward_std': 0.12630687654018402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1432502269744873, 'sampling/sampling_logp_difference/max': 1.766525387763977, 'sampling/importance_sampling_ratio/min': 0.17092587053775787, 'sampling/importance_sampling_ratio/mean': 1.0145972967147827, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.848037950694561, 'clip_ratio/low_mean': 0.11832964140921831, 'clip_ratio/low_min': 0.11832964140921831, 'clip_ratio/high_mean': 0.022077922709286213, 'clip_ratio/high_max': 0.022077922709286213, 'clip_ratio/region_mean': 0.14040756411850452, 'reward_total_mean': 0.13423895835876465, 'reward_meter_mean': 0.3699997365474701, 'reward_meter_std': 0.3258172869682312, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7188122272491455, 'reward_repeat_penalty_std': 0.19269207119941711, 'reward_repeat_floor_mean': 0.968213677406311, 'reward_repeat_floor_std': 0.020962757989764214, 'reward_near_duplicate_penalty_mean': 0.8991570472717285, 'reward_near_duplicate_penalty_std': 0.05692235380411148, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.8982231616973877, 'reward_distinct_2_std': 0.083673395216465, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.13423895835876465, 'reward_total_composite_std': 0.12630689144134521, 'epoch': 0.06}

 35%|███▍      | 1153/3300 [4:26:33<18:49:51, 31.58s/it]INFO 04-12 20:24:59 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1154/3300 [4:26:42<14:50:12, 24.89s/it]2026-04-12 20:25:08,524 | INFO | train_grpo_train | metrics_logged mode=train step=1154

                                                        
{'loss': -0.0044, 'grad_norm': 17.46927833557129, 'learning_rate': 6.506060606060607e-06, 'num_tokens': 2056361.0, 'completions/mean_length': 28.75, 'completions/min_length': 25.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.635522186756134, 'rewards/meter/std': 0.3521137237548828, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9786953330039978, 'rewards/lexical_plausibility/std': 0.06025872007012367, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.9707440137863159, 'rewards/repeat_penalty/std': 0.04813366010785103, 'rewards/repeat_floor/mean': 0.9963599443435669, 'rewards/repeat_floor/std': 0.005245621781796217, 'rewards/near_duplicate_penalty/mean': 0.983974814414978, 'rewards/near_duplicate_penalty/std': 0.016977880150079727, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.2994198203086853, 'rewards/total_composite/std': 0.2327273041009903, 'reward': 0.2994198203086853, 'reward_std': 0.2327273041009903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16486014425754547, 'sampling/sampling_logp_difference/max': 2.0230321884155273, 'sampling/importance_sampling_ratio/min': 0.13225384056568146, 'sampling/importance_sampling_ratio/mean': 0.9832513928413391, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8267838284373283, 'clip_ratio/low_mean': 0.05542989354580641, 'clip_ratio/low_min': 0.05542989354580641, 'clip_ratio/high_mean': 0.08571589924395084, 'clip_ratio/high_max': 0.08571589924395084, 'clip_ratio/region_mean': 0.14114579278975725, 'reward_total_mean': 0.2994198203086853, 'reward_meter_mean': 0.635522186756134, 'reward_meter_std': 0.3521137237548828, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9786953330039978, 'reward_lexical_plausibility_std': 0.06025872007012367, 'reward_repeat_penalty_mean': 0.9707440137863159, 'reward_repeat_penalty_std': 0.04813366010785103, 'reward_repeat_floor_mean': 0.9963599443435669, 'reward_repeat_floor_std': 0.005245621781796217, 'reward_near_duplicate_penalty_mean': 0.983974814414978, 'reward_near_duplicate_penalty_std': 0.016977880150079727, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.2994198203086853, 'reward_total_composite_std': 0.2327273041009903, 'epoch': 0.06}

 35%|███▍      | 1154/3300 [4:26:42<14:50:12, 24.89s/it]INFO 04-12 20:25:08 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1155/3300 [4:26:52<12:14:00, 20.53s/it]2026-04-12 20:25:18,907 | INFO | train_grpo_train | metrics_logged mode=train step=1155

                                                        
{'loss': 0.005, 'grad_norm': 9.964399337768555, 'learning_rate': 6.503030303030303e-06, 'num_tokens': 2058618.0, 'completions/mean_length': 86.125, 'completions/min_length': 75.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.125, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.8563166856765747, 'rewards/meter/std': 0.2183181196451187, 'rewards/count_adherence/mean': 0.569444477558136, 'rewards/count_adherence/std': 0.03928370773792267, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8708333373069763, 'rewards/count_floor/std': 0.01178510207682848, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.6219333410263062, 'rewards/repeat_penalty/std': 0.20783403515815735, 'rewards/repeat_floor/mean': 0.9534505605697632, 'rewards/repeat_floor/std': 0.02509978972375393, 'rewards/near_duplicate_penalty/mean': 0.8502159118652344, 'rewards/near_duplicate_penalty/std': 0.07133748382329941, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.7532573938369751, 'rewards/distinct_2/std': 0.14269323647022247, 'rewards/total_composite/mean': 0.21906912326812744, 'rewards/total_composite/std': 0.09565562754869461, 'reward': 0.21906912326812744, 'reward_std': 0.09565562754869461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14156417548656464, 'sampling/sampling_logp_difference/max': 3.4508442878723145, 'sampling/importance_sampling_ratio/min': 0.031718842685222626, 'sampling/importance_sampling_ratio/mean': 1.0200921297073364, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7914101481437683, 'clip_ratio/low_mean': 0.05078322812914848, 'clip_ratio/low_min': 0.05078322812914848, 'clip_ratio/high_mean': 0.06147776357829571, 'clip_ratio/high_max': 0.06147776357829571, 'clip_ratio/region_mean': 0.11226099170744419, 'reward_total_mean': 0.21906912326812744, 'reward_meter_mean': 0.8563166856765747, 'reward_meter_std': 0.2183181196451187, 'reward_count_adherence_mean': 0.569444477558136, 'reward_count_adherence_std': 0.03928370773792267, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8708333373069763, 'reward_count_floor_std': 0.01178510207682848, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6219333410263062, 'reward_repeat_penalty_std': 0.20783403515815735, 'reward_repeat_floor_mean': 0.9534505605697632, 'reward_repeat_floor_std': 0.02509978972375393, 'reward_near_duplicate_penalty_mean': 0.8502159118652344, 'reward_near_duplicate_penalty_std': 0.07133748382329941, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.7532573938369751, 'reward_distinct_2_std': 0.14269323647022247, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.21906912326812744, 'reward_total_composite_std': 0.09565562754869461, 'epoch': 0.06}

 35%|███▌      | 1155/3300 [4:26:52<12:14:00, 20.53s/it]INFO 04-12 20:25:19 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1156/3300 [4:27:01<10:08:25, 17.03s/it]2026-04-12 20:25:27,738 | INFO | train_grpo_train | metrics_logged mode=train step=1156

                                                        
{'loss': 0.0643, 'grad_norm': 14.037107467651367, 'learning_rate': 6.5000000000000004e-06, 'num_tokens': 2060144.0, 'completions/mean_length': 32.75, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7854678630828857, 'rewards/meter/std': 0.33875739574432373, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8979510068893433, 'rewards/repeat_penalty/std': 0.1679760366678238, 'rewards/repeat_floor/mean': 0.9875774383544922, 'rewards/repeat_floor/std': 0.018704188987612724, 'rewards/near_duplicate_penalty/mean': 0.9534009695053101, 'rewards/near_duplicate_penalty/std': 0.043604057282209396, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9604700803756714, 'rewards/distinct_2/std': 0.08285844326019287, 'rewards/total_composite/mean': 0.2822146415710449, 'rewards/total_composite/std': 0.1538948267698288, 'reward': 0.2822146415710449, 'reward_std': 0.1538948267698288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13271188735961914, 'sampling/sampling_logp_difference/max': 1.9485602378845215, 'sampling/importance_sampling_ratio/min': 0.1424790620803833, 'sampling/importance_sampling_ratio/mean': 1.0143986940383911, 'sampling/importance_sampling_ratio/max': 1.859466552734375, 'entropy': 0.8529380187392235, 'clip_ratio/low_mean': 0.03912673145532608, 'clip_ratio/low_min': 0.03912673145532608, 'clip_ratio/high_mean': 0.06894915783777833, 'clip_ratio/high_max': 0.06894915783777833, 'clip_ratio/region_mean': 0.10807588929310441, 'reward_total_mean': 0.2822146415710449, 'reward_meter_mean': 0.7854678630828857, 'reward_meter_std': 0.33875739574432373, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.8979510068893433, 'reward_repeat_penalty_std': 0.1679760366678238, 'reward_repeat_floor_mean': 0.9875774383544922, 'reward_repeat_floor_std': 0.018704188987612724, 'reward_near_duplicate_penalty_mean': 0.9534009695053101, 'reward_near_duplicate_penalty_std': 0.043604057282209396, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9604700803756714, 'reward_distinct_2_std': 0.08285844326019287, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.2822146415710449, 'reward_total_composite_std': 0.1538948267698288, 'epoch': 0.06}

 35%|███▌      | 1156/3300 [4:27:01<10:08:25, 17.03s/it]INFO 04-12 20:25:28 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1157/3300 [4:27:16<9:42:29, 16.31s/it] 2026-04-12 20:25:42,376 | INFO | train_grpo_train | metrics_logged mode=train step=1157

                                                       
{'loss': -0.0319, 'grad_norm': 5.865137577056885, 'learning_rate': 6.496969696969697e-06, 'num_tokens': 2061950.0, 'completions/mean_length': 118.75, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 62.57143020629883, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.48764151334762573, 'rewards/meter/std': 0.3870513439178467, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9434781074523926, 'rewards/lexical_plausibility/std': 0.1598680317401886, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7771471738815308, 'rewards/repeat_penalty/std': 0.10980329662561417, 'rewards/repeat_floor/mean': 0.971560001373291, 'rewards/repeat_floor/std': 0.013160254806280136, 'rewards/near_duplicate_penalty/mean': 0.8849307298660278, 'rewards/near_duplicate_penalty/std': 0.0374838262796402, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8757820129394531, 'rewards/distinct_2/std': 0.09655886888504028, 'rewards/total_composite/mean': 0.13093577325344086, 'rewards/total_composite/std': 0.1346941590309143, 'reward': 0.13093577325344086, 'reward_std': 0.1346941739320755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15730227530002594, 'sampling/sampling_logp_difference/max': 2.12467098236084, 'sampling/importance_sampling_ratio/min': 0.11947227269411087, 'sampling/importance_sampling_ratio/mean': 0.9975865483283997, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7934385240077972, 'clip_ratio/low_mean': 0.07863768097013235, 'clip_ratio/low_min': 0.07863768097013235, 'clip_ratio/high_mean': 0.051933810114860535, 'clip_ratio/high_max': 0.051933810114860535, 'clip_ratio/region_mean': 0.13057149108499289, 'reward_total_mean': 0.13093577325344086, 'reward_meter_mean': 0.48764151334762573, 'reward_meter_std': 0.3870513439178467, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9434781074523926, 'reward_lexical_plausibility_std': 0.1598680317401886, 'reward_repeat_penalty_mean': 0.7771471738815308, 'reward_repeat_penalty_std': 0.10980329662561417, 'reward_repeat_floor_mean': 0.971560001373291, 'reward_repeat_floor_std': 0.013160254806280136, 'reward_near_duplicate_penalty_mean': 0.8849307298660278, 'reward_near_duplicate_penalty_std': 0.0374838262796402, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8757820129394531, 'reward_distinct_2_std': 0.09655886888504028, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.13093577325344086, 'reward_total_composite_std': 0.1346941590309143, 'epoch': 0.06}

 35%|███▌      | 1157/3300 [4:27:16<9:42:29, 16.31s/it]INFO 04-12 20:25:42 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1158/3300 [4:27:25<8:24:47, 14.14s/it]2026-04-12 20:25:51,453 | INFO | train_grpo_train | metrics_logged mode=train step=1158

                                                       
{'loss': 0.0459, 'grad_norm': 15.484068870544434, 'learning_rate': 6.493939393939395e-06, 'num_tokens': 2063529.0, 'completions/mean_length': 38.375, 'completions/min_length': 34.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6438114643096924, 'rewards/meter/std': 0.35187867283821106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8934659957885742, 'rewards/repeat_penalty/std': 0.08277325332164764, 'rewards/repeat_floor/mean': 0.9852427244186401, 'rewards/repeat_floor/std': 0.010353182442486286, 'rewards/near_duplicate_penalty/mean': 0.9192298650741577, 'rewards/near_duplicate_penalty/std': 0.04352216050028801, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9706477522850037, 'rewards/distinct_2/std': 0.05773414671421051, 'rewards/total_composite/mean': 0.22211618721485138, 'rewards/total_composite/std': 0.15434913337230682, 'reward': 0.22211618721485138, 'reward_std': 0.15434913337230682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14374874532222748, 'sampling/sampling_logp_difference/max': 2.1967883110046387, 'sampling/importance_sampling_ratio/min': 0.11115959286689758, 'sampling/importance_sampling_ratio/mean': 1.0207242965698242, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8655757308006287, 'clip_ratio/low_mean': 0.08804118540138006, 'clip_ratio/low_min': 0.08804118540138006, 'clip_ratio/high_mean': 0.05212550796568394, 'clip_ratio/high_max': 0.05212550796568394, 'clip_ratio/region_mean': 0.140166693367064, 'reward_total_mean': 0.22211618721485138, 'reward_meter_mean': 0.6438114643096924, 'reward_meter_std': 0.35187867283821106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8934659957885742, 'reward_repeat_penalty_std': 0.08277325332164764, 'reward_repeat_floor_mean': 0.9852427244186401, 'reward_repeat_floor_std': 0.010353182442486286, 'reward_near_duplicate_penalty_mean': 0.9192298650741577, 'reward_near_duplicate_penalty_std': 0.04352216050028801, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9706477522850037, 'reward_distinct_2_std': 0.05773414671421051, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.22211618721485138, 'reward_total_composite_std': 0.15434913337230682, 'epoch': 0.06}

 35%|███▌      | 1158/3300 [4:27:25<8:24:47, 14.14s/it]INFO 04-12 20:25:51 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1159/3300 [4:27:39<8:25:42, 14.17s/it]2026-04-12 20:26:05,696 | INFO | train_grpo_train | metrics_logged mode=train step=1159

                                                       
{'loss': -0.0157, 'grad_norm': 4.696249008178711, 'learning_rate': 6.490909090909091e-06, 'num_tokens': 2064964.0, 'completions/mean_length': 82.375, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.27648496627807617, 'rewards/meter/std': 0.32832154631614685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9272909164428711, 'rewards/lexical_plausibility/std': 0.14972831308841705, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.7200098037719727, 'rewards/repeat_penalty/std': 0.056112002581357956, 'rewards/repeat_floor/mean': 0.9800509214401245, 'rewards/repeat_floor/std': 0.009941146709024906, 'rewards/near_duplicate_penalty/mean': 0.977495551109314, 'rewards/near_duplicate_penalty/std': 0.06365221738815308, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.14794588088989258, 'rewards/total_composite/std': 0.2056187093257904, 'reward': 0.14794588088989258, 'reward_std': 0.2056187093257904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1781710684299469, 'sampling/sampling_logp_difference/max': 1.902235507965088, 'sampling/importance_sampling_ratio/min': 0.1492346227169037, 'sampling/importance_sampling_ratio/mean': 0.9933753609657288, 'sampling/importance_sampling_ratio/max': 1.8410460948944092, 'entropy': 0.8407760038971901, 'clip_ratio/low_mean': 0.031250000931322575, 'clip_ratio/low_min': 0.031250000931322575, 'clip_ratio/high_mean': 0.034090910106897354, 'clip_ratio/high_max': 0.034090910106897354, 'clip_ratio/region_mean': 0.06534091103821993, 'reward_total_mean': 0.14794588088989258, 'reward_meter_mean': 0.27648496627807617, 'reward_meter_std': 0.32832154631614685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9272909164428711, 'reward_lexical_plausibility_std': 0.14972831308841705, 'reward_repeat_penalty_mean': 0.7200098037719727, 'reward_repeat_penalty_std': 0.056112002581357956, 'reward_repeat_floor_mean': 0.9800509214401245, 'reward_repeat_floor_std': 0.009941146709024906, 'reward_near_duplicate_penalty_mean': 0.977495551109314, 'reward_near_duplicate_penalty_std': 0.06365221738815308, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.14794588088989258, 'reward_total_composite_std': 0.2056187093257904, 'epoch': 0.06}

 35%|███▌      | 1159/3300 [4:27:39<8:25:42, 14.17s/it]INFO 04-12 20:26:05 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1160/3300 [4:27:48<7:29:50, 12.61s/it]2026-04-12 20:26:14,672 | INFO | train_grpo_train | metrics_logged mode=train step=1160

                                                       
{'loss': 0.0581, 'grad_norm': 15.720664978027344, 'learning_rate': 6.487878787878789e-06, 'num_tokens': 2066467.0, 'completions/mean_length': 42.875, 'completions/min_length': 34.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4660980999469757, 'rewards/meter/std': 0.3450045585632324, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.8497967720031738, 'rewards/repeat_penalty/std': 0.10524605959653854, 'rewards/repeat_floor/mean': 0.9820445775985718, 'rewards/repeat_floor/std': 0.012370223179459572, 'rewards/near_duplicate_penalty/mean': 0.9378513693809509, 'rewards/near_duplicate_penalty/std': 0.04161651432514191, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9040762186050415, 'rewards/distinct_2/std': 0.08254225552082062, 'rewards/total_composite/mean': 0.09876013547182083, 'rewards/total_composite/std': 0.08257844299077988, 'reward': 0.09876013547182083, 'reward_std': 0.08257844299077988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13839685916900635, 'sampling/sampling_logp_difference/max': 2.001260280609131, 'sampling/importance_sampling_ratio/min': 0.13516482710838318, 'sampling/importance_sampling_ratio/mean': 1.0199694633483887, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9884352087974548, 'clip_ratio/low_mean': 0.06572796776890755, 'clip_ratio/low_min': 0.06572796776890755, 'clip_ratio/high_mean': 0.06660071201622486, 'clip_ratio/high_max': 0.06660071201622486, 'clip_ratio/region_mean': 0.1323286797851324, 'reward_total_mean': 0.09876013547182083, 'reward_meter_mean': 0.4660980999469757, 'reward_meter_std': 0.3450045585632324, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8497967720031738, 'reward_repeat_penalty_std': 0.10524605959653854, 'reward_repeat_floor_mean': 0.9820445775985718, 'reward_repeat_floor_std': 0.012370223179459572, 'reward_near_duplicate_penalty_mean': 0.9378513693809509, 'reward_near_duplicate_penalty_std': 0.04161651432514191, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9040762186050415, 'reward_distinct_2_std': 0.08254225552082062, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.09876013547182083, 'reward_total_composite_std': 0.08257844299077988, 'epoch': 0.06}

 35%|███▌      | 1160/3300 [4:27:48<7:29:50, 12.61s/it]INFO 04-12 20:26:14 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1161/3300 [4:28:03<8:01:15, 13.50s/it]2026-04-12 20:26:30,250 | INFO | train_grpo_train | metrics_logged mode=train step=1161

                                                       
{'loss': -0.008, 'grad_norm': 4.582535266876221, 'learning_rate': 6.484848484848485e-06, 'num_tokens': 2068065.0, 'completions/mean_length': 95.75, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.28571701049805, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.17230460047721863, 'rewards/meter/std': 0.16195982694625854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9570178985595703, 'rewards/lexical_plausibility/std': 0.12157180905342102, 'rewards/judge_quality/mean': 0.4337499737739563, 'rewards/judge_quality/std': 0.23868314921855927, 'rewards/repeat_penalty/mean': 0.9817172288894653, 'rewards/repeat_penalty/std': 0.02288021519780159, 'rewards/repeat_floor/mean': 0.9972575902938843, 'rewards/repeat_floor/std': 0.003432035446166992, 'rewards/near_duplicate_penalty/mean': 0.9817172288894653, 'rewards/near_duplicate_penalty/std': 0.02288021519780159, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08432656526565552, 'rewards/total_composite/std': 0.10472684353590012, 'reward': 0.08432656526565552, 'reward_std': 0.10472684353590012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1567433625459671, 'sampling/sampling_logp_difference/max': 1.515317440032959, 'sampling/importance_sampling_ratio/min': 0.2197384238243103, 'sampling/importance_sampling_ratio/mean': 0.9965807199478149, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8705822378396988, 'clip_ratio/low_mean': 0.07751431968063116, 'clip_ratio/low_min': 0.07751431968063116, 'clip_ratio/high_mean': 0.04292929358780384, 'clip_ratio/high_max': 0.04292929358780384, 'clip_ratio/region_mean': 0.120443613268435, 'reward_total_mean': 0.08432656526565552, 'reward_meter_mean': 0.17230460047721863, 'reward_meter_std': 0.16195982694625854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9570178985595703, 'reward_lexical_plausibility_std': 0.12157180905342102, 'reward_repeat_penalty_mean': 0.9817172288894653, 'reward_repeat_penalty_std': 0.02288021519780159, 'reward_repeat_floor_mean': 0.9972575902938843, 'reward_repeat_floor_std': 0.003432035446166992, 'reward_near_duplicate_penalty_mean': 0.9817172288894653, 'reward_near_duplicate_penalty_std': 0.02288021519780159, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4337499737739563, 'reward_judge_quality_std': 0.23868314921855927, 'reward_total_composite_mean': 0.08432656526565552, 'reward_total_composite_std': 0.10472684353590012, 'epoch': 0.06}

 35%|███▌      | 1161/3300 [4:28:04<8:01:15, 13.50s/it]INFO 04-12 20:26:30 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1162/3300 [4:28:19<8:18:01, 13.98s/it]2026-04-12 20:26:45,334 | INFO | train_grpo_train | metrics_logged mode=train step=1162

                                                       
{'loss': -0.079, 'grad_norm': 5.631038188934326, 'learning_rate': 6.481818181818182e-06, 'num_tokens': 2070181.0, 'completions/mean_length': 137.5, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 84.0, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.6919570565223694, 'rewards/meter/std': 0.3975428640842438, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9682310223579407, 'rewards/lexical_plausibility/std': 0.08985619992017746, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8089238405227661, 'rewards/repeat_penalty/std': 0.12179581820964813, 'rewards/repeat_floor/mean': 0.9776259660720825, 'rewards/repeat_floor/std': 0.01334670651704073, 'rewards/near_duplicate_penalty/mean': 0.9300089478492737, 'rewards/near_duplicate_penalty/std': 0.027919495478272438, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8732602000236511, 'rewards/distinct_2/std': 0.09980998933315277, 'rewards/total_composite/mean': 0.20294485986232758, 'rewards/total_composite/std': 0.1514461785554886, 'reward': 0.20294485986232758, 'reward_std': 0.1514461785554886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11978968977928162, 'sampling/sampling_logp_difference/max': 1.9871903657913208, 'sampling/importance_sampling_ratio/min': 0.13708002865314484, 'sampling/importance_sampling_ratio/mean': 1.0126415491104126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6159942448139191, 'clip_ratio/low_mean': 0.030828980961814523, 'clip_ratio/low_min': 0.030828980961814523, 'clip_ratio/high_mean': 0.06853400450199842, 'clip_ratio/high_max': 0.06853400450199842, 'clip_ratio/region_mean': 0.09936298546381295, 'reward_total_mean': 0.20294485986232758, 'reward_meter_mean': 0.6919570565223694, 'reward_meter_std': 0.3975428640842438, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9682310223579407, 'reward_lexical_plausibility_std': 0.08985619992017746, 'reward_repeat_penalty_mean': 0.8089238405227661, 'reward_repeat_penalty_std': 0.12179581820964813, 'reward_repeat_floor_mean': 0.9776259660720825, 'reward_repeat_floor_std': 0.01334670651704073, 'reward_near_duplicate_penalty_mean': 0.9300089478492737, 'reward_near_duplicate_penalty_std': 0.027919495478272438, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8732602000236511, 'reward_distinct_2_std': 0.09980998933315277, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.20294485986232758, 'reward_total_composite_std': 0.1514461785554886, 'epoch': 0.06}

 35%|███▌      | 1162/3300 [4:28:19<8:18:01, 13.98s/it]INFO 04-12 20:26:45 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1163/3300 [4:28:28<7:26:01, 12.52s/it]2026-04-12 20:26:54,464 | INFO | train_grpo_train | metrics_logged mode=train step=1163

                                                       
{'loss': 0.0485, 'grad_norm': 16.970054626464844, 'learning_rate': 6.478787878787879e-06, 'num_tokens': 2071664.0, 'completions/mean_length': 37.375, 'completions/min_length': 35.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8161476254463196, 'rewards/meter/std': 0.27444982528686523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.8676532506942749, 'rewards/repeat_penalty/std': 0.12680873274803162, 'rewards/repeat_floor/mean': 0.9854923486709595, 'rewards/repeat_floor/std': 0.012804005295038223, 'rewards/near_duplicate_penalty/mean': 0.9528676271438599, 'rewards/near_duplicate_penalty/std': 0.03966348245739937, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9381910562515259, 'rewards/distinct_2/std': 0.07818726450204849, 'rewards/total_composite/mean': 0.24345338344573975, 'rewards/total_composite/std': 0.11819523572921753, 'reward': 0.24345338344573975, 'reward_std': 0.11819523572921753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16171762347221375, 'sampling/sampling_logp_difference/max': 2.447770595550537, 'sampling/importance_sampling_ratio/min': 0.08648618310689926, 'sampling/importance_sampling_ratio/mean': 1.0307674407958984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7863077223300934, 'clip_ratio/low_mean': 0.06299201771616936, 'clip_ratio/low_min': 0.06299201771616936, 'clip_ratio/high_mean': 0.06924000778235495, 'clip_ratio/high_max': 0.06924000778235495, 'clip_ratio/region_mean': 0.1322320254985243, 'reward_total_mean': 0.24345338344573975, 'reward_meter_mean': 0.8161476254463196, 'reward_meter_std': 0.27444982528686523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8676532506942749, 'reward_repeat_penalty_std': 0.12680873274803162, 'reward_repeat_floor_mean': 0.9854923486709595, 'reward_repeat_floor_std': 0.012804005295038223, 'reward_near_duplicate_penalty_mean': 0.9528676271438599, 'reward_near_duplicate_penalty_std': 0.03966348245739937, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9381910562515259, 'reward_distinct_2_std': 0.07818726450204849, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.24345338344573975, 'reward_total_composite_std': 0.11819523572921753, 'epoch': 0.06}

 35%|███▌      | 1163/3300 [4:28:28<7:26:01, 12.52s/it]INFO 04-12 20:26:54 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1164/3300 [4:28:37<6:50:39, 11.54s/it]2026-04-12 20:27:03,695 | INFO | train_grpo_train | metrics_logged mode=train step=1164

                                                       
{'loss': 0.0649, 'grad_norm': 20.438199996948242, 'learning_rate': 6.475757575757576e-06, 'num_tokens': 2073318.0, 'completions/mean_length': 32.75, 'completions/min_length': 29.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9538679122924805, 'rewards/meter/std': 0.08358047157526016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5587499737739563, 'rewards/judge_quality/std': 0.1950412094593048, 'rewards/repeat_penalty/mean': 0.9687126874923706, 'rewards/repeat_penalty/std': 0.054679952561855316, 'rewards/repeat_floor/mean': 0.9959474802017212, 'rewards/repeat_floor/std': 0.006458973977714777, 'rewards/near_duplicate_penalty/mean': 0.9806758165359497, 'rewards/near_duplicate_penalty/std': 0.02333887480199337, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9871795177459717, 'rewards/distinct_2/std': 0.03626188635826111, 'rewards/total_composite/mean': 0.5294739007949829, 'rewards/total_composite/std': 0.21466690301895142, 'reward': 0.5294739007949829, 'reward_std': 0.21466688811779022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1903575360774994, 'sampling/sampling_logp_difference/max': 2.798466682434082, 'sampling/importance_sampling_ratio/min': 0.06090337410569191, 'sampling/importance_sampling_ratio/mean': 1.0329440832138062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7618510350584984, 'clip_ratio/low_mean': 0.12161163985729218, 'clip_ratio/low_min': 0.12161163985729218, 'clip_ratio/high_mean': 0.047407833859324455, 'clip_ratio/high_max': 0.047407833859324455, 'clip_ratio/region_mean': 0.16901947371661663, 'reward_total_mean': 0.5294739007949829, 'reward_meter_mean': 0.9538679122924805, 'reward_meter_std': 0.08358047157526016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9687126874923706, 'reward_repeat_penalty_std': 0.054679952561855316, 'reward_repeat_floor_mean': 0.9959474802017212, 'reward_repeat_floor_std': 0.006458973977714777, 'reward_near_duplicate_penalty_mean': 0.9806758165359497, 'reward_near_duplicate_penalty_std': 0.02333887480199337, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9871795177459717, 'reward_distinct_2_std': 0.03626188635826111, 'reward_judge_quality_mean': 0.5587499737739563, 'reward_judge_quality_std': 0.1950412094593048, 'reward_total_composite_mean': 0.5294739007949829, 'reward_total_composite_std': 0.21466690301895142, 'epoch': 0.06}

 35%|███▌      | 1164/3300 [4:28:37<6:50:39, 11.54s/it]INFO 04-12 20:27:03 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1165/3300 [4:28:45<6:17:09, 10.60s/it]2026-04-12 20:27:12,115 | INFO | train_grpo_train | metrics_logged mode=train step=1165

                                                       
{'loss': 0.0393, 'grad_norm': 15.200995445251465, 'learning_rate': 6.472727272727272e-06, 'num_tokens': 2074913.0, 'completions/mean_length': 35.375, 'completions/min_length': 32.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.42000287771224976, 'rewards/meter/std': 0.3212505877017975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.21249999105930328, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.7220735549926758, 'rewards/repeat_penalty/std': 0.2497500330209732, 'rewards/repeat_floor/mean': 0.9631826281547546, 'rewards/repeat_floor/std': 0.03353746607899666, 'rewards/near_duplicate_penalty/mean': 0.8581938743591309, 'rewards/near_duplicate_penalty/std': 0.11930513381958008, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8689281940460205, 'rewards/distinct_2/std': 0.14304015040397644, 'rewards/total_composite/mean': 0.100531205534935, 'rewards/total_composite/std': 0.10330844670534134, 'reward': 0.100531205534935, 'reward_std': 0.10330844670534134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11675301939249039, 'sampling/sampling_logp_difference/max': 1.809838056564331, 'sampling/importance_sampling_ratio/min': 0.16368064284324646, 'sampling/importance_sampling_ratio/mean': 1.0187861919403076, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.89546187967062, 'clip_ratio/low_mean': 0.03737593372352421, 'clip_ratio/low_min': 0.03737593372352421, 'clip_ratio/high_mean': 0.08466759137809277, 'clip_ratio/high_max': 0.08466759137809277, 'clip_ratio/region_mean': 0.12204352510161698, 'reward_total_mean': 0.100531205534935, 'reward_meter_mean': 0.42000287771224976, 'reward_meter_std': 0.3212505877017975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7220735549926758, 'reward_repeat_penalty_std': 0.2497500330209732, 'reward_repeat_floor_mean': 0.9631826281547546, 'reward_repeat_floor_std': 0.03353746607899666, 'reward_near_duplicate_penalty_mean': 0.8581938743591309, 'reward_near_duplicate_penalty_std': 0.11930513381958008, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8689281940460205, 'reward_distinct_2_std': 0.14304015040397644, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.100531205534935, 'reward_total_composite_std': 0.10330844670534134, 'epoch': 0.06}

 35%|███▌      | 1165/3300 [4:28:45<6:17:09, 10.60s/it]INFO 04-12 20:27:12 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1166/3300 [4:28:55<6:02:26, 10.19s/it]2026-04-12 20:27:21,348 | INFO | train_grpo_train | metrics_logged mode=train step=1166

                                                       
{'loss': 0.0533, 'grad_norm': 22.2900447845459, 'learning_rate': 6.4696969696969705e-06, 'num_tokens': 2076663.0, 'completions/mean_length': 46.75, 'completions/min_length': 41.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8413935303688049, 'rewards/meter/std': 0.26518484950065613, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4124999940395355, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9452246427536011, 'rewards/repeat_penalty/std': 0.039447903633117676, 'rewards/repeat_floor/mean': 0.9932481050491333, 'rewards/repeat_floor/std': 0.004163719248026609, 'rewards/near_duplicate_penalty/mean': 0.9708089232444763, 'rewards/near_duplicate_penalty/std': 0.016380518674850464, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9736305475234985, 'rewards/distinct_2/std': 0.03655470907688141, 'rewards/total_composite/mean': 0.31506264209747314, 'rewards/total_composite/std': 0.1310979276895523, 'reward': 0.31506264209747314, 'reward_std': 0.1310979276895523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14283476769924164, 'sampling/sampling_logp_difference/max': 1.736236572265625, 'sampling/importance_sampling_ratio/min': 0.1938168853521347, 'sampling/importance_sampling_ratio/mean': 1.003521203994751, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6892067454755306, 'clip_ratio/low_mean': 0.047656466253101826, 'clip_ratio/low_min': 0.047656466253101826, 'clip_ratio/high_mean': 0.06528578093275428, 'clip_ratio/high_max': 0.06528578093275428, 'clip_ratio/region_mean': 0.1129422471858561, 'reward_total_mean': 0.31506264209747314, 'reward_meter_mean': 0.8413935303688049, 'reward_meter_std': 0.26518484950065613, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9452246427536011, 'reward_repeat_penalty_std': 0.039447903633117676, 'reward_repeat_floor_mean': 0.9932481050491333, 'reward_repeat_floor_std': 0.004163719248026609, 'reward_near_duplicate_penalty_mean': 0.9708089232444763, 'reward_near_duplicate_penalty_std': 0.016380518674850464, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9736305475234985, 'reward_distinct_2_std': 0.03655470907688141, 'reward_judge_quality_mean': 0.4124999940395355, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.31506264209747314, 'reward_total_composite_std': 0.1310979276895523, 'epoch': 0.06}

 35%|███▌      | 1166/3300 [4:28:55<6:02:26, 10.19s/it]INFO 04-12 20:27:21 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1167/3300 [4:29:05<6:03:50, 10.23s/it]2026-04-12 20:27:31,693 | INFO | train_grpo_train | metrics_logged mode=train step=1167

                                                       
{'loss': -0.0126, 'grad_norm': 15.429891586303711, 'learning_rate': 6.466666666666667e-06, 'num_tokens': 2078270.0, 'completions/mean_length': 37.875, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.580768883228302, 'rewards/meter/std': 0.35080647468566895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5424999594688416, 'rewards/judge_quality/std': 0.23705033957958221, 'rewards/repeat_penalty/mean': 0.9361674785614014, 'rewards/repeat_penalty/std': 0.11030331999063492, 'rewards/repeat_floor/mean': 0.9932888150215149, 'rewards/repeat_floor/std': 0.008743004873394966, 'rewards/near_duplicate_penalty/mean': 0.9738315939903259, 'rewards/near_duplicate_penalty/std': 0.021498937159776688, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.36090075969696045, 'rewards/total_composite/std': 0.32910487055778503, 'reward': 0.36090075969696045, 'reward_std': 0.32910484075546265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15758608281612396, 'sampling/sampling_logp_difference/max': 1.8015403747558594, 'sampling/importance_sampling_ratio/min': 0.16504447162151337, 'sampling/importance_sampling_ratio/mean': 1.0100774765014648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.893947422504425, 'clip_ratio/low_mean': 0.0787806436419487, 'clip_ratio/low_min': 0.0787806436419487, 'clip_ratio/high_mean': 0.04233381152153015, 'clip_ratio/high_max': 0.04233381152153015, 'clip_ratio/region_mean': 0.12111445516347885, 'reward_total_mean': 0.36090075969696045, 'reward_meter_mean': 0.580768883228302, 'reward_meter_std': 0.35080647468566895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9361674785614014, 'reward_repeat_penalty_std': 0.11030331999063492, 'reward_repeat_floor_mean': 0.9932888150215149, 'reward_repeat_floor_std': 0.008743004873394966, 'reward_near_duplicate_penalty_mean': 0.9738315939903259, 'reward_near_duplicate_penalty_std': 0.021498937159776688, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.5424999594688416, 'reward_judge_quality_std': 0.23705033957958221, 'reward_total_composite_mean': 0.36090075969696045, 'reward_total_composite_std': 0.32910487055778503, 'epoch': 0.06}

 35%|███▌      | 1167/3300 [4:29:05<6:03:50, 10.23s/it]INFO 04-12 20:27:32 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1168/3300 [4:29:19<6:45:27, 11.41s/it]2026-04-12 20:27:45,844 | INFO | train_grpo_train | metrics_logged mode=train step=1168

                                                       
{'loss': -0.0721, 'grad_norm': 6.22426176071167, 'learning_rate': 6.463636363636364e-06, 'num_tokens': 2080252.0, 'completions/mean_length': 125.75, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.5714340209961, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.6669456958770752, 'rewards/meter/std': 0.35373440384864807, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.9652393460273743, 'rewards/lexical_plausibility/std': 0.0983179584145546, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.14500615000724792, 'rewards/repeat_penalty/mean': 0.897933304309845, 'rewards/repeat_penalty/std': 0.060476645827293396, 'rewards/repeat_floor/mean': 0.9872768521308899, 'rewards/repeat_floor/std': 0.00716354139149189, 'rewards/near_duplicate_penalty/mean': 0.9456610679626465, 'rewards/near_duplicate_penalty/std': 0.03142561390995979, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9491965174674988, 'rewards/distinct_2/std': 0.04890682175755501, 'rewards/total_composite/mean': 0.20631453394889832, 'rewards/total_composite/std': 0.14997436106204987, 'reward': 0.20631453394889832, 'reward_std': 0.14997434616088867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14736725389957428, 'sampling/sampling_logp_difference/max': 1.6134846210479736, 'sampling/importance_sampling_ratio/min': 0.19919230043888092, 'sampling/importance_sampling_ratio/mean': 1.0059081315994263, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8500893861055374, 'clip_ratio/low_mean': 0.027432807721197605, 'clip_ratio/low_min': 0.027432807721197605, 'clip_ratio/high_mean': 0.08002971019595861, 'clip_ratio/high_max': 0.08002971019595861, 'clip_ratio/region_mean': 0.10746251791715622, 'reward_total_mean': 0.20631453394889832, 'reward_meter_mean': 0.6669456958770752, 'reward_meter_std': 0.35373440384864807, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.9652393460273743, 'reward_lexical_plausibility_std': 0.0983179584145546, 'reward_repeat_penalty_mean': 0.897933304309845, 'reward_repeat_penalty_std': 0.060476645827293396, 'reward_repeat_floor_mean': 0.9872768521308899, 'reward_repeat_floor_std': 0.00716354139149189, 'reward_near_duplicate_penalty_mean': 0.9456610679626465, 'reward_near_duplicate_penalty_std': 0.03142561390995979, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9491965174674988, 'reward_distinct_2_std': 0.04890682175755501, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.14500615000724792, 'reward_total_composite_mean': 0.20631453394889832, 'reward_total_composite_std': 0.14997436106204987, 'epoch': 0.06}

 35%|███▌      | 1168/3300 [4:29:19<6:45:27, 11.41s/it]INFO 04-12 20:27:46 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1169/3300 [4:29:34<7:24:10, 12.51s/it]2026-04-12 20:28:00,904 | INFO | train_grpo_train | metrics_logged mode=train step=1169

                                                       
{'loss': -0.0233, 'grad_norm': 1.5329583883285522, 'learning_rate': 6.460606060606061e-06, 'num_tokens': 2081900.0, 'completions/mean_length': 287.0, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.16482141613960266, 'rewards/meter/std': 0.20791269838809967, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.05669468268752098, 'rewards/lexical_plausibility/mean': 0.8037311434745789, 'rewards/lexical_plausibility/std': 0.21235255897045135, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9125614762306213, 'rewards/repeat_penalty/std': 0.0802270919084549, 'rewards/repeat_floor/mean': 0.9909461736679077, 'rewards/repeat_floor/std': 0.00573686184361577, 'rewards/near_duplicate_penalty/mean': 0.9670474529266357, 'rewards/near_duplicate_penalty/std': 0.0214032381772995, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9751566052436829, 'rewards/distinct_2/std': 0.026700975373387337, 'rewards/total_composite/mean': 0.03795194625854492, 'rewards/total_composite/std': 0.06141568720340729, 'reward': 0.03795194625854492, 'reward_std': 0.06141568347811699, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1377536654472351, 'sampling/sampling_logp_difference/max': 1.4968687295913696, 'sampling/importance_sampling_ratio/min': 0.2238299548625946, 'sampling/importance_sampling_ratio/mean': 1.0231471061706543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46314381062984467, 'clip_ratio/low_mean': 0.03395562805235386, 'clip_ratio/low_min': 0.03395562805235386, 'clip_ratio/high_mean': 0.03532608784735203, 'clip_ratio/high_max': 0.03532608784735203, 'clip_ratio/region_mean': 0.06928171589970589, 'reward_total_mean': 0.03795194625854492, 'reward_meter_mean': 0.16482141613960266, 'reward_meter_std': 0.20791269838809967, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.18898223340511322, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.05669468268752098, 'reward_lexical_plausibility_mean': 0.8037311434745789, 'reward_lexical_plausibility_std': 0.21235255897045135, 'reward_repeat_penalty_mean': 0.9125614762306213, 'reward_repeat_penalty_std': 0.0802270919084549, 'reward_repeat_floor_mean': 0.9909461736679077, 'reward_repeat_floor_std': 0.00573686184361577, 'reward_near_duplicate_penalty_mean': 0.9670474529266357, 'reward_near_duplicate_penalty_std': 0.0214032381772995, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9751566052436829, 'reward_distinct_2_std': 0.026700975373387337, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.03795194625854492, 'reward_total_composite_std': 0.06141568720340729, 'epoch': 0.06}

 35%|███▌      | 1169/3300 [4:29:34<7:24:10, 12.51s/it]INFO 04-12 20:28:01 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1170/3300 [4:29:44<6:52:23, 11.62s/it]2026-04-12 20:28:10,464 | INFO | train_grpo_train | metrics_logged mode=train step=1170

                                                       
{'loss': 0.0431, 'grad_norm': 15.129555702209473, 'learning_rate': 6.457575757575758e-06, 'num_tokens': 2083562.0, 'completions/mean_length': 36.75, 'completions/min_length': 35.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5394048690795898, 'rewards/meter/std': 0.3251197338104248, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.7317183017730713, 'rewards/repeat_penalty/std': 0.2782125771045685, 'rewards/repeat_floor/mean': 0.965384840965271, 'rewards/repeat_floor/std': 0.03441338613629341, 'rewards/near_duplicate_penalty/mean': 0.8558469414710999, 'rewards/near_duplicate_penalty/std': 0.09792716056108475, 'rewards/opening_diversity/mean': 0.859375, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.9493926763534546, 'rewards/distinct_2/std': 0.08588341623544693, 'rewards/total_composite/mean': 0.20513157546520233, 'rewards/total_composite/std': 0.1099473237991333, 'reward': 0.20513157546520233, 'reward_std': 0.1099473237991333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1395585685968399, 'sampling/sampling_logp_difference/max': 1.2856788635253906, 'sampling/importance_sampling_ratio/min': 0.3174495995044708, 'sampling/importance_sampling_ratio/mean': 1.025923490524292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2124241590499878, 'clip_ratio/low_mean': 0.038241437170654535, 'clip_ratio/low_min': 0.038241437170654535, 'clip_ratio/high_mean': 0.08114373125135899, 'clip_ratio/high_max': 0.08114373125135899, 'clip_ratio/region_mean': 0.11938516842201352, 'reward_total_mean': 0.20513157546520233, 'reward_meter_mean': 0.5394048690795898, 'reward_meter_std': 0.3251197338104248, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7317183017730713, 'reward_repeat_penalty_std': 0.2782125771045685, 'reward_repeat_floor_mean': 0.965384840965271, 'reward_repeat_floor_std': 0.03441338613629341, 'reward_near_duplicate_penalty_mean': 0.8558469414710999, 'reward_near_duplicate_penalty_std': 0.09792716056108475, 'reward_opening_diversity_mean': 0.859375, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.9493926763534546, 'reward_distinct_2_std': 0.08588341623544693, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.20513157546520233, 'reward_total_composite_std': 0.1099473237991333, 'epoch': 0.06}

 35%|███▌      | 1170/3300 [4:29:44<6:52:23, 11.62s/it]INFO 04-12 20:28:10 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1171/3300 [4:29:59<7:33:50, 12.79s/it]2026-04-12 20:28:25,971 | INFO | train_grpo_train | metrics_logged mode=train step=1171

                                                       
{'loss': -0.1211, 'grad_norm': 2.693272829055786, 'learning_rate': 6.454545454545456e-06, 'num_tokens': 2085432.0, 'completions/mean_length': 180.75, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 70.33333587646484, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.6504790186882019, 'rewards/meter/std': 0.3757871985435486, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.9116208553314209, 'rewards/lexical_plausibility/std': 0.16364926099777222, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.8919795751571655, 'rewards/repeat_penalty/std': 0.07459060847759247, 'rewards/repeat_floor/mean': 0.9868773818016052, 'rewards/repeat_floor/std': 0.008385890163481236, 'rewards/near_duplicate_penalty/mean': 0.9481184482574463, 'rewards/near_duplicate_penalty/std': 0.03529313579201698, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9406622648239136, 'rewards/distinct_2/std': 0.06901639699935913, 'rewards/total_composite/mean': 0.18823757767677307, 'rewards/total_composite/std': 0.1282653957605362, 'reward': 0.18823757767677307, 'reward_std': 0.1282653957605362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15525995194911957, 'sampling/sampling_logp_difference/max': 1.603744626045227, 'sampling/importance_sampling_ratio/min': 0.20114189386367798, 'sampling/importance_sampling_ratio/mean': 1.0033159255981445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6417087316513062, 'clip_ratio/low_mean': 0.021604938432574272, 'clip_ratio/low_min': 0.021604938432574272, 'clip_ratio/high_mean': 0.09305895399302244, 'clip_ratio/high_max': 0.09305895399302244, 'clip_ratio/region_mean': 0.11466389242559671, 'reward_total_mean': 0.18823757767677307, 'reward_meter_mean': 0.6504790186882019, 'reward_meter_std': 0.3757871985435486, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.9116208553314209, 'reward_lexical_plausibility_std': 0.16364926099777222, 'reward_repeat_penalty_mean': 0.8919795751571655, 'reward_repeat_penalty_std': 0.07459060847759247, 'reward_repeat_floor_mean': 0.9868773818016052, 'reward_repeat_floor_std': 0.008385890163481236, 'reward_near_duplicate_penalty_mean': 0.9481184482574463, 'reward_near_duplicate_penalty_std': 0.03529313579201698, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9406622648239136, 'reward_distinct_2_std': 0.06901639699935913, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.18823757767677307, 'reward_total_composite_std': 0.1282653957605362, 'epoch': 0.06}

 35%|███▌      | 1171/3300 [4:29:59<7:33:50, 12.79s/it]INFO 04-12 20:28:26 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1172/3300 [4:30:14<7:51:48, 13.30s/it]2026-04-12 20:28:40,466 | INFO | train_grpo_train | metrics_logged mode=train step=1172

                                                       
{'loss': -0.0842, 'grad_norm': 4.343000888824463, 'learning_rate': 6.451515151515152e-06, 'num_tokens': 2087867.0, 'completions/mean_length': 206.375, 'completions/min_length': 89.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 104.5, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.25451478362083435, 'rewards/meter/std': 0.3005032539367676, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.020044591277837753, 'rewards/lexical_plausibility/mean': 0.8755115270614624, 'rewards/lexical_plausibility/std': 0.17141562700271606, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9441555738449097, 'rewards/repeat_penalty/std': 0.08863238245248795, 'rewards/repeat_floor/mean': 0.9929077625274658, 'rewards/repeat_floor/std': 0.009711802005767822, 'rewards/near_duplicate_penalty/mean': 0.9686335325241089, 'rewards/near_duplicate_penalty/std': 0.022336483001708984, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9734748005867004, 'rewards/distinct_2/std': 0.07502459734678268, 'rewards/total_composite/mean': 0.08731585741043091, 'rewards/total_composite/std': 0.11921091377735138, 'reward': 0.08731585741043091, 'reward_std': 0.11921091377735138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15906591713428497, 'sampling/sampling_logp_difference/max': 2.512744426727295, 'sampling/importance_sampling_ratio/min': 0.08104550838470459, 'sampling/importance_sampling_ratio/mean': 1.0241643190383911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6895046234130859, 'clip_ratio/low_mean': 0.08757777698338032, 'clip_ratio/low_min': 0.08757777698338032, 'clip_ratio/high_mean': 0.021990740671753883, 'clip_ratio/high_max': 0.021990740671753883, 'clip_ratio/region_mean': 0.1095685176551342, 'reward_total_mean': 0.08731585741043091, 'reward_meter_mean': 0.25451478362083435, 'reward_meter_std': 0.3005032539367676, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.020044591277837753, 'reward_lexical_plausibility_mean': 0.8755115270614624, 'reward_lexical_plausibility_std': 0.17141562700271606, 'reward_repeat_penalty_mean': 0.9441555738449097, 'reward_repeat_penalty_std': 0.08863238245248795, 'reward_repeat_floor_mean': 0.9929077625274658, 'reward_repeat_floor_std': 0.009711802005767822, 'reward_near_duplicate_penalty_mean': 0.9686335325241089, 'reward_near_duplicate_penalty_std': 0.022336483001708984, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9734748005867004, 'reward_distinct_2_std': 0.07502459734678268, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.08731585741043091, 'reward_total_composite_std': 0.11921091377735138, 'epoch': 0.06}

 36%|███▌      | 1172/3300 [4:30:14<7:51:48, 13.30s/it]INFO 04-12 20:28:40 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1173/3300 [4:30:28<7:57:18, 13.46s/it]2026-04-12 20:28:54,306 | INFO | train_grpo_train | metrics_logged mode=train step=1173

                                                       
{'loss': -0.0676, 'grad_norm': 3.6703689098358154, 'learning_rate': 6.4484848484848496e-06, 'num_tokens': 2089366.0, 'completions/mean_length': 94.375, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7640120983123779, 'rewards/meter/std': 0.3361947238445282, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9438018798828125, 'rewards/lexical_plausibility/std': 0.15895235538482666, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.14880476891994476, 'rewards/repeat_penalty/mean': 0.8772763013839722, 'rewards/repeat_penalty/std': 0.09475304186344147, 'rewards/repeat_floor/mean': 0.9869712591171265, 'rewards/repeat_floor/std': 0.009290582500398159, 'rewards/near_duplicate_penalty/mean': 0.9571102261543274, 'rewards/near_duplicate_penalty/std': 0.03337300941348076, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9475524425506592, 'rewards/distinct_2/std': 0.06198638677597046, 'rewards/total_composite/mean': 0.231920063495636, 'rewards/total_composite/std': 0.14521637558937073, 'reward': 0.231920063495636, 'reward_std': 0.14521637558937073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1311490833759308, 'sampling/sampling_logp_difference/max': 1.021317958831787, 'sampling/importance_sampling_ratio/min': 0.4232490062713623, 'sampling/importance_sampling_ratio/mean': 1.0339765548706055, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7758961468935013, 'clip_ratio/low_mean': 0.06531531736254692, 'clip_ratio/low_min': 0.06531531736254692, 'clip_ratio/high_mean': 0.07865659799426794, 'clip_ratio/high_max': 0.07865659799426794, 'clip_ratio/region_mean': 0.14397191535681486, 'reward_total_mean': 0.231920063495636, 'reward_meter_mean': 0.7640120983123779, 'reward_meter_std': 0.3361947238445282, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9438018798828125, 'reward_lexical_plausibility_std': 0.15895235538482666, 'reward_repeat_penalty_mean': 0.8772763013839722, 'reward_repeat_penalty_std': 0.09475304186344147, 'reward_repeat_floor_mean': 0.9869712591171265, 'reward_repeat_floor_std': 0.009290582500398159, 'reward_near_duplicate_penalty_mean': 0.9571102261543274, 'reward_near_duplicate_penalty_std': 0.03337300941348076, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9475524425506592, 'reward_distinct_2_std': 0.06198638677597046, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.14880476891994476, 'reward_total_composite_mean': 0.231920063495636, 'reward_total_composite_std': 0.14521637558937073, 'epoch': 0.06}

 36%|███▌      | 1173/3300 [4:30:28<7:57:18, 13.46s/it]INFO 04-12 20:28:54 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1174/3300 [4:30:36<7:07:09, 12.06s/it]2026-04-12 20:29:03,070 | INFO | train_grpo_train | metrics_logged mode=train step=1174

                                                       
{'loss': 0.0343, 'grad_norm': 12.081247329711914, 'learning_rate': 6.445454545454546e-06, 'num_tokens': 2091383.0, 'completions/mean_length': 73.125, 'completions/min_length': 69.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.2684102952480316, 'rewards/meter/std': 0.2224985808134079, 'rewards/count_adherence/mean': 0.7142857313156128, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9437500238418579, 'rewards/arabic_floor/std': 0.0776323676109314, 'rewards/count_floor/mean': 0.9142857193946838, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32124999165534973, 'rewards/judge_quality/std': 0.2613939344882965, 'rewards/repeat_penalty/mean': 0.9842795133590698, 'rewards/repeat_penalty/std': 0.011036689393222332, 'rewards/repeat_floor/mean': 0.997641921043396, 'rewards/repeat_floor/std': 0.0016554942121729255, 'rewards/near_duplicate_penalty/mean': 0.9842795133590698, 'rewards/near_duplicate_penalty/std': 0.011036689393222332, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11585226655006409, 'rewards/total_composite/std': 0.21081215143203735, 'reward': 0.11585226655006409, 'reward_std': 0.21081213653087616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1400773823261261, 'sampling/sampling_logp_difference/max': 3.6298511028289795, 'sampling/importance_sampling_ratio/min': 0.026520133018493652, 'sampling/importance_sampling_ratio/mean': 0.9967420101165771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4955849088728428, 'clip_ratio/low_mean': 0.07609959505498409, 'clip_ratio/low_min': 0.07609959505498409, 'clip_ratio/high_mean': 0.014999999664723873, 'clip_ratio/high_max': 0.014999999664723873, 'clip_ratio/region_mean': 0.09109959471970797, 'reward_total_mean': 0.11585226655006409, 'reward_meter_mean': 0.2684102952480316, 'reward_meter_std': 0.2224985808134079, 'reward_count_adherence_mean': 0.7142857313156128, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9437500238418579, 'reward_arabic_floor_std': 0.0776323676109314, 'reward_count_floor_mean': 0.9142857193946838, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9842795133590698, 'reward_repeat_penalty_std': 0.011036689393222332, 'reward_repeat_floor_mean': 0.997641921043396, 'reward_repeat_floor_std': 0.0016554942121729255, 'reward_near_duplicate_penalty_mean': 0.9842795133590698, 'reward_near_duplicate_penalty_std': 0.011036689393222332, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32124999165534973, 'reward_judge_quality_std': 0.2613939344882965, 'reward_total_composite_mean': 0.11585226655006409, 'reward_total_composite_std': 0.21081215143203735, 'epoch': 0.06}

 36%|███▌      | 1174/3300 [4:30:36<7:07:09, 12.06s/it]INFO 04-12 20:29:03 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1175/3300 [4:30:50<7:28:21, 12.66s/it]2026-04-12 20:29:17,149 | INFO | train_grpo_train | metrics_logged mode=train step=1175

                                                       
{'loss': -0.1229, 'grad_norm': 2.8570423126220703, 'learning_rate': 6.442424242424243e-06, 'num_tokens': 2093107.0, 'completions/mean_length': 169.5, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 55.333335876464844, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6233956813812256, 'rewards/meter/std': 0.34679114818573, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175493285059929, 'rewards/lexical_plausibility/mean': 0.9061822891235352, 'rewards/lexical_plausibility/std': 0.17565475404262543, 'rewards/judge_quality/mean': 0.24374999105930328, 'rewards/judge_quality/std': 0.17816023528575897, 'rewards/repeat_penalty/mean': 0.8485980033874512, 'rewards/repeat_penalty/std': 0.10092641413211823, 'rewards/repeat_floor/mean': 0.9814270734786987, 'rewards/repeat_floor/std': 0.013136054389178753, 'rewards/near_duplicate_penalty/mean': 0.9245766401290894, 'rewards/near_duplicate_penalty/std': 0.059332989156246185, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9276729226112366, 'rewards/distinct_2/std': 0.06419502943754196, 'rewards/total_composite/mean': 0.13856390118598938, 'rewards/total_composite/std': 0.10914583504199982, 'reward': 0.13856390118598938, 'reward_std': 0.10914584249258041, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14658518135547638, 'sampling/sampling_logp_difference/max': 1.446281909942627, 'sampling/importance_sampling_ratio/min': 0.2354440689086914, 'sampling/importance_sampling_ratio/mean': 1.021836757659912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7347785532474518, 'clip_ratio/low_mean': 0.023690078407526016, 'clip_ratio/low_min': 0.023690078407526016, 'clip_ratio/high_mean': 0.06728896172717214, 'clip_ratio/high_max': 0.06728896172717214, 'clip_ratio/region_mean': 0.09097904013469815, 'reward_total_mean': 0.13856390118598938, 'reward_meter_mean': 0.6233956813812256, 'reward_meter_std': 0.34679114818573, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175493285059929, 'reward_lexical_plausibility_mean': 0.9061822891235352, 'reward_lexical_plausibility_std': 0.17565475404262543, 'reward_repeat_penalty_mean': 0.8485980033874512, 'reward_repeat_penalty_std': 0.10092641413211823, 'reward_repeat_floor_mean': 0.9814270734786987, 'reward_repeat_floor_std': 0.013136054389178753, 'reward_near_duplicate_penalty_mean': 0.9245766401290894, 'reward_near_duplicate_penalty_std': 0.059332989156246185, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9276729226112366, 'reward_distinct_2_std': 0.06419502943754196, 'reward_judge_quality_mean': 0.24374999105930328, 'reward_judge_quality_std': 0.17816023528575897, 'reward_total_composite_mean': 0.13856390118598938, 'reward_total_composite_std': 0.10914583504199982, 'epoch': 0.06}

 36%|███▌      | 1175/3300 [4:30:51<7:28:21, 12.66s/it]INFO 04-12 20:29:17 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1176/3300 [4:31:00<6:59:54, 11.86s/it]2026-04-12 20:29:27,153 | INFO | train_grpo_train | metrics_logged mode=train step=1176

                                                       
{'loss': 0.0915, 'grad_norm': 10.135137557983398, 'learning_rate': 6.43939393939394e-06, 'num_tokens': 2095301.0, 'completions/mean_length': 83.25, 'completions/min_length': 73.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.25, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.646192193031311, 'rewards/meter/std': 0.37273940443992615, 'rewards/count_adherence/mean': 0.5892857313156128, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8767856955528259, 'rewards/count_floor/std': 0.015152297914028168, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.7674247026443481, 'rewards/repeat_penalty/std': 0.15474501252174377, 'rewards/repeat_floor/mean': 0.9736853837966919, 'rewards/repeat_floor/std': 0.01742369867861271, 'rewards/near_duplicate_penalty/mean': 0.9230344891548157, 'rewards/near_duplicate_penalty/std': 0.04531843215227127, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8358914852142334, 'rewards/distinct_2/std': 0.1191958338022232, 'rewards/total_composite/mean': 0.15712034702301025, 'rewards/total_composite/std': 0.09755794703960419, 'reward': 0.15712034702301025, 'reward_std': 0.09755793958902359, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13611240684986115, 'sampling/sampling_logp_difference/max': 2.0423898696899414, 'sampling/importance_sampling_ratio/min': 0.12971831858158112, 'sampling/importance_sampling_ratio/mean': 1.0118257999420166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8137337416410446, 'clip_ratio/low_mean': 0.06493366323411465, 'clip_ratio/low_min': 0.06493366323411465, 'clip_ratio/high_mean': 0.04521361459046602, 'clip_ratio/high_max': 0.04521361459046602, 'clip_ratio/region_mean': 0.11014727782458067, 'reward_total_mean': 0.15712034702301025, 'reward_meter_mean': 0.646192193031311, 'reward_meter_std': 0.37273940443992615, 'reward_count_adherence_mean': 0.5892857313156128, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8767856955528259, 'reward_count_floor_std': 0.015152297914028168, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7674247026443481, 'reward_repeat_penalty_std': 0.15474501252174377, 'reward_repeat_floor_mean': 0.9736853837966919, 'reward_repeat_floor_std': 0.01742369867861271, 'reward_near_duplicate_penalty_mean': 0.9230344891548157, 'reward_near_duplicate_penalty_std': 0.04531843215227127, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8358914852142334, 'reward_distinct_2_std': 0.1191958338022232, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.15712034702301025, 'reward_total_composite_std': 0.09755794703960419, 'epoch': 0.06}

 36%|███▌      | 1176/3300 [4:31:01<6:59:54, 11.86s/it]INFO 04-12 20:29:27 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1177/3300 [4:31:10<6:32:40, 11.10s/it]2026-04-12 20:29:36,456 | INFO | train_grpo_train | metrics_logged mode=train step=1177

                                                       
{'loss': 0.0172, 'grad_norm': 11.781129837036133, 'learning_rate': 6.436363636363637e-06, 'num_tokens': 2097376.0, 'completions/mean_length': 75.375, 'completions/min_length': 65.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.4562099575996399, 'rewards/meter/std': 0.30367332696914673, 'rewards/count_adherence/mean': 0.7321428656578064, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9196428060531616, 'rewards/count_floor/std': 0.015152277424931526, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.8536198139190674, 'rewards/repeat_penalty/std': 0.19675740599632263, 'rewards/repeat_floor/mean': 0.9849412441253662, 'rewards/repeat_floor/std': 0.017513787373900414, 'rewards/near_duplicate_penalty/mean': 0.957760214805603, 'rewards/near_duplicate_penalty/std': 0.032522544264793396, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9030799865722656, 'rewards/distinct_2/std': 0.14267270267009735, 'rewards/total_composite/mean': 0.14770402014255524, 'rewards/total_composite/std': 0.09274586290121078, 'reward': 0.14770402014255524, 'reward_std': 0.09274585545063019, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14335109293460846, 'sampling/sampling_logp_difference/max': 2.8070526123046875, 'sampling/importance_sampling_ratio/min': 0.06038270518183708, 'sampling/importance_sampling_ratio/mean': 1.0255144834518433, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7056219540536404, 'clip_ratio/low_mean': 0.08524096012115479, 'clip_ratio/low_min': 0.08524096012115479, 'clip_ratio/high_mean': 0.04571624659001827, 'clip_ratio/high_max': 0.04571624659001827, 'clip_ratio/region_mean': 0.13095720671117306, 'reward_total_mean': 0.14770402014255524, 'reward_meter_mean': 0.4562099575996399, 'reward_meter_std': 0.30367332696914673, 'reward_count_adherence_mean': 0.7321428656578064, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9196428060531616, 'reward_count_floor_std': 0.015152277424931526, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8536198139190674, 'reward_repeat_penalty_std': 0.19675740599632263, 'reward_repeat_floor_mean': 0.9849412441253662, 'reward_repeat_floor_std': 0.017513787373900414, 'reward_near_duplicate_penalty_mean': 0.957760214805603, 'reward_near_duplicate_penalty_std': 0.032522544264793396, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9030799865722656, 'reward_distinct_2_std': 0.14267270267009735, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.14770402014255524, 'reward_total_composite_std': 0.09274586290121078, 'epoch': 0.06}

 36%|███▌      | 1177/3300 [4:31:10<6:32:40, 11.10s/it]INFO 04-12 20:29:36 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1178/3300 [4:31:20<6:19:12, 10.72s/it]2026-04-12 20:29:46,302 | INFO | train_grpo_train | metrics_logged mode=train step=1178

                                                       
{'loss': 0.0706, 'grad_norm': 16.067766189575195, 'learning_rate': 6.433333333333333e-06, 'num_tokens': 2098961.0, 'completions/mean_length': 35.125, 'completions/min_length': 28.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.3932083249092102, 'rewards/meter/std': 0.31237199902534485, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.2620250880718231, 'rewards/repeat_penalty/mean': 0.9277840852737427, 'rewards/repeat_penalty/std': 0.12222915887832642, 'rewards/repeat_floor/mean': 0.9925423860549927, 'rewards/repeat_floor/std': 0.010928593575954437, 'rewards/near_duplicate_penalty/mean': 0.9765478372573853, 'rewards/near_duplicate_penalty/std': 0.022501083090901375, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9770580530166626, 'rewards/distinct_2/std': 0.04287045821547508, 'rewards/total_composite/mean': 0.25220540165901184, 'rewards/total_composite/std': 0.29897743463516235, 'reward': 0.25220540165901184, 'reward_std': 0.29897740483283997, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2015005648136139, 'sampling/sampling_logp_difference/max': 4.263955116271973, 'sampling/importance_sampling_ratio/min': 0.01406655739992857, 'sampling/importance_sampling_ratio/mean': 1.0184476375579834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0405635833740234, 'clip_ratio/low_mean': 0.08031512796878815, 'clip_ratio/low_min': 0.08031512796878815, 'clip_ratio/high_mean': 0.0418560616672039, 'clip_ratio/high_max': 0.0418560616672039, 'clip_ratio/region_mean': 0.12217118963599205, 'reward_total_mean': 0.25220540165901184, 'reward_meter_mean': 0.3932083249092102, 'reward_meter_std': 0.31237199902534485, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9277840852737427, 'reward_repeat_penalty_std': 0.12222915887832642, 'reward_repeat_floor_mean': 0.9925423860549927, 'reward_repeat_floor_std': 0.010928593575954437, 'reward_near_duplicate_penalty_mean': 0.9765478372573853, 'reward_near_duplicate_penalty_std': 0.022501083090901375, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9770580530166626, 'reward_distinct_2_std': 0.04287045821547508, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.2620250880718231, 'reward_total_composite_mean': 0.25220540165901184, 'reward_total_composite_std': 0.29897743463516235, 'epoch': 0.06}

 36%|███▌      | 1178/3300 [4:31:20<6:19:12, 10.72s/it]INFO 04-12 20:29:46 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1179/3300 [4:31:35<7:06:57, 12.08s/it]2026-04-12 20:30:01,549 | INFO | train_grpo_train | metrics_logged mode=train step=1179

                                                       
{'loss': -0.1342, 'grad_norm': 5.803771495819092, 'learning_rate': 6.430303030303031e-06, 'num_tokens': 2101726.0, 'completions/mean_length': 196.625, 'completions/min_length': 125.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 151.57144165039062, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 168.0, 'rewards/meter/mean': 0.7816752791404724, 'rewards/meter/std': 0.1955045461654663, 'rewards/count_adherence/mean': 0.6416666507720947, 'rewards/count_adherence/std': 0.1630561351776123, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8924999833106995, 'rewards/count_floor/std': 0.04891684651374817, 'rewards/lexical_plausibility/mean': 0.9546787142753601, 'rewards/lexical_plausibility/std': 0.12818799912929535, 'rewards/judge_quality/mean': 0.2875000238418579, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.7376455664634705, 'rewards/repeat_penalty/std': 0.19538435339927673, 'rewards/repeat_floor/mean': 0.9733842611312866, 'rewards/repeat_floor/std': 0.01891866698861122, 'rewards/near_duplicate_penalty/mean': 0.9357107281684875, 'rewards/near_duplicate_penalty/std': 0.03585259988903999, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.811418354511261, 'rewards/distinct_2/std': 0.15347079932689667, 'rewards/total_composite/mean': 0.20624026656150818, 'rewards/total_composite/std': 0.12254581600427628, 'reward': 0.20624026656150818, 'reward_std': 0.12254581600427628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1515682488679886, 'sampling/sampling_logp_difference/max': 2.480428695678711, 'sampling/importance_sampling_ratio/min': 0.08370733261108398, 'sampling/importance_sampling_ratio/mean': 1.0019123554229736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6952157318592072, 'clip_ratio/low_mean': 0.0440100422129035, 'clip_ratio/low_min': 0.0440100422129035, 'clip_ratio/high_mean': 0.07741628587245941, 'clip_ratio/high_max': 0.07741628587245941, 'clip_ratio/region_mean': 0.12142632808536291, 'reward_total_mean': 0.20624026656150818, 'reward_meter_mean': 0.7816752791404724, 'reward_meter_std': 0.1955045461654663, 'reward_count_adherence_mean': 0.6416666507720947, 'reward_count_adherence_std': 0.1630561351776123, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8924999833106995, 'reward_count_floor_std': 0.04891684651374817, 'reward_lexical_plausibility_mean': 0.9546787142753601, 'reward_lexical_plausibility_std': 0.12818799912929535, 'reward_repeat_penalty_mean': 0.7376455664634705, 'reward_repeat_penalty_std': 0.19538435339927673, 'reward_repeat_floor_mean': 0.9733842611312866, 'reward_repeat_floor_std': 0.01891866698861122, 'reward_near_duplicate_penalty_mean': 0.9357107281684875, 'reward_near_duplicate_penalty_std': 0.03585259988903999, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.811418354511261, 'reward_distinct_2_std': 0.15347079932689667, 'reward_judge_quality_mean': 0.2875000238418579, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.20624026656150818, 'reward_total_composite_std': 0.12254581600427628, 'epoch': 0.06}

 36%|███▌      | 1179/3300 [4:31:35<7:06:57, 12.08s/it]INFO 04-12 20:30:01 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1180/3300 [4:31:43<6:28:05, 10.98s/it]2026-04-12 20:30:09,975 | INFO | train_grpo_train | metrics_logged mode=train step=1180

                                                       
{'loss': 0.1742, 'grad_norm': 23.012863159179688, 'learning_rate': 6.427272727272728e-06, 'num_tokens': 2103134.0, 'completions/mean_length': 20.0, 'completions/min_length': 14.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.4516674876213074, 'rewards/meter/std': 0.4358198642730713, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.3700748085975647, 'rewards/repeat_penalty/mean': 0.7786733508110046, 'rewards/repeat_penalty/std': 0.08971989154815674, 'rewards/repeat_floor/mean': 0.986359715461731, 'rewards/repeat_floor/std': 0.005697190761566162, 'rewards/near_duplicate_penalty/mean': 0.996564507484436, 'rewards/near_duplicate_penalty/std': 0.009717104025185108, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.17987149953842163, 'rewards/total_composite/std': 0.1643475443124771, 'reward': 0.17987149953842163, 'reward_std': 0.16434752941131592, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1801695078611374, 'sampling/sampling_logp_difference/max': 1.1258525848388672, 'sampling/importance_sampling_ratio/min': 0.32437577843666077, 'sampling/importance_sampling_ratio/mean': 1.0048922300338745, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0632063671946526, 'clip_ratio/low_mean': 0.06390555063262582, 'clip_ratio/low_min': 0.06390555063262582, 'clip_ratio/high_mean': 0.054793234914541245, 'clip_ratio/high_max': 0.054793234914541245, 'clip_ratio/region_mean': 0.11869878554716706, 'reward_total_mean': 0.17987149953842163, 'reward_meter_mean': 0.4516674876213074, 'reward_meter_std': 0.4358198642730713, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7786733508110046, 'reward_repeat_penalty_std': 0.08971989154815674, 'reward_repeat_floor_mean': 0.986359715461731, 'reward_repeat_floor_std': 0.005697190761566162, 'reward_near_duplicate_penalty_mean': 0.996564507484436, 'reward_near_duplicate_penalty_std': 0.009717104025185108, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.3700748085975647, 'reward_total_composite_mean': 0.17987149953842163, 'reward_total_composite_std': 0.1643475443124771, 'epoch': 0.06}

 36%|███▌      | 1180/3300 [4:31:43<6:28:05, 10.98s/it]INFO 04-12 20:30:10 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1181/3300 [4:31:51<5:54:45, 10.05s/it]2026-04-12 20:30:17,830 | INFO | train_grpo_train | metrics_logged mode=train step=1181

                                                       
{'loss': 0.0329, 'grad_norm': 14.65428352355957, 'learning_rate': 6.424242424242425e-06, 'num_tokens': 2104503.0, 'completions/mean_length': 19.125, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.5089462995529175, 'rewards/meter/std': 0.46019771695137024, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9322916269302368, 'rewards/lexical_plausibility/std': 0.13717946410179138, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.73777174949646, 'rewards/repeat_penalty/std': 0.03458673879504204, 'rewards/repeat_floor/mean': 0.9825543761253357, 'rewards/repeat_floor/std': 0.006917351391166449, 'rewards/near_duplicate_penalty/mean': 0.9836956262588501, 'rewards/near_duplicate_penalty/std': 0.04611567035317421, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.09618495404720306, 'rewards/total_composite/std': 0.09962039440870285, 'reward': 0.09618495404720306, 'reward_std': 0.09962038695812225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1431964784860611, 'sampling/sampling_logp_difference/max': 1.3529396057128906, 'sampling/importance_sampling_ratio/min': 0.2584793269634247, 'sampling/importance_sampling_ratio/mean': 1.018261432647705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8322898298501968, 'clip_ratio/low_mean': 0.052010233979672194, 'clip_ratio/low_min': 0.052010233979672194, 'clip_ratio/high_mean': 0.07331349374726415, 'clip_ratio/high_max': 0.07331349374726415, 'clip_ratio/region_mean': 0.12532372772693634, 'reward_total_mean': 0.09618495404720306, 'reward_meter_mean': 0.5089462995529175, 'reward_meter_std': 0.46019771695137024, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9322916269302368, 'reward_lexical_plausibility_std': 0.13717946410179138, 'reward_repeat_penalty_mean': 0.73777174949646, 'reward_repeat_penalty_std': 0.03458673879504204, 'reward_repeat_floor_mean': 0.9825543761253357, 'reward_repeat_floor_std': 0.006917351391166449, 'reward_near_duplicate_penalty_mean': 0.9836956262588501, 'reward_near_duplicate_penalty_std': 0.04611567035317421, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.09618495404720306, 'reward_total_composite_std': 0.09962039440870285, 'epoch': 0.06}

 36%|███▌      | 1181/3300 [4:31:51<5:54:45, 10.05s/it]INFO 04-12 20:30:18 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1182/3300 [4:32:01<5:49:58,  9.91s/it]2026-04-12 20:30:27,549 | INFO | train_grpo_train | metrics_logged mode=train step=1182

                                                       
{'loss': 0.0381, 'grad_norm': 15.322174072265625, 'learning_rate': 6.4212121212121215e-06, 'num_tokens': 2106277.0, 'completions/mean_length': 47.75, 'completions/min_length': 36.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6455488204956055, 'rewards/meter/std': 0.34475696086883545, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.8652312755584717, 'rewards/repeat_penalty/std': 0.10077371448278427, 'rewards/repeat_floor/mean': 0.9828896522521973, 'rewards/repeat_floor/std': 0.012291517108678818, 'rewards/near_duplicate_penalty/mean': 0.9276446104049683, 'rewards/near_duplicate_penalty/std': 0.04552360624074936, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9304767847061157, 'rewards/distinct_2/std': 0.07052377611398697, 'rewards/total_composite/mean': 0.21648946404457092, 'rewards/total_composite/std': 0.11931361258029938, 'reward': 0.21648946404457092, 'reward_std': 0.11931361258029938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14189143478870392, 'sampling/sampling_logp_difference/max': 1.317641258239746, 'sampling/importance_sampling_ratio/min': 0.26776614785194397, 'sampling/importance_sampling_ratio/mean': 1.0085481405258179, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0084244906902313, 'clip_ratio/low_mean': 0.03993618651293218, 'clip_ratio/low_min': 0.03993618651293218, 'clip_ratio/high_mean': 0.0881309974938631, 'clip_ratio/high_max': 0.0881309974938631, 'clip_ratio/region_mean': 0.1280671840067953, 'reward_total_mean': 0.21648946404457092, 'reward_meter_mean': 0.6455488204956055, 'reward_meter_std': 0.34475696086883545, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8652312755584717, 'reward_repeat_penalty_std': 0.10077371448278427, 'reward_repeat_floor_mean': 0.9828896522521973, 'reward_repeat_floor_std': 0.012291517108678818, 'reward_near_duplicate_penalty_mean': 0.9276446104049683, 'reward_near_duplicate_penalty_std': 0.04552360624074936, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9304767847061157, 'reward_distinct_2_std': 0.07052377611398697, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.21648946404457092, 'reward_total_composite_std': 0.11931361258029938, 'epoch': 0.06}

 36%|███▌      | 1182/3300 [4:32:01<5:49:58,  9.91s/it]INFO 04-12 20:30:27 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1183/3300 [4:32:09<5:35:29,  9.51s/it]2026-04-12 20:30:35,995 | INFO | train_grpo_train | metrics_logged mode=train step=1183

                                                       
{'loss': 0.0112, 'grad_norm': 15.95504093170166, 'learning_rate': 6.418181818181819e-06, 'num_tokens': 2107836.0, 'completions/mean_length': 39.875, 'completions/min_length': 36.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7433016896247864, 'rewards/meter/std': 0.18175439536571503, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.5954976081848145, 'rewards/repeat_penalty/std': 0.21761220693588257, 'rewards/repeat_floor/mean': 0.9552149772644043, 'rewards/repeat_floor/std': 0.0264911986887455, 'rewards/near_duplicate_penalty/mean': 0.8680577874183655, 'rewards/near_duplicate_penalty/std': 0.08897467702627182, 'rewards/opening_diversity/mean': 0.765625, 'rewards/opening_diversity/std': 0.19408094882965088, 'rewards/distinct_2/mean': 0.8785424828529358, 'rewards/distinct_2/std': 0.09677964448928833, 'rewards/total_composite/mean': 0.31861215829849243, 'rewards/total_composite/std': 0.11181426048278809, 'reward': 0.31861215829849243, 'reward_std': 0.11181425303220749, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1395990252494812, 'sampling/sampling_logp_difference/max': 3.518556594848633, 'sampling/importance_sampling_ratio/min': 0.029642188921570778, 'sampling/importance_sampling_ratio/mean': 1.0078552961349487, 'sampling/importance_sampling_ratio/max': 1.81479811668396, 'entropy': 0.8162758573889732, 'clip_ratio/low_mean': 0.07800290104933083, 'clip_ratio/low_min': 0.07800290104933083, 'clip_ratio/high_mean': 0.02722902176901698, 'clip_ratio/high_max': 0.02722902176901698, 'clip_ratio/region_mean': 0.10523192281834781, 'reward_total_mean': 0.31861215829849243, 'reward_meter_mean': 0.7433016896247864, 'reward_meter_std': 0.18175439536571503, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5954976081848145, 'reward_repeat_penalty_std': 0.21761220693588257, 'reward_repeat_floor_mean': 0.9552149772644043, 'reward_repeat_floor_std': 0.0264911986887455, 'reward_near_duplicate_penalty_mean': 0.8680577874183655, 'reward_near_duplicate_penalty_std': 0.08897467702627182, 'reward_opening_diversity_mean': 0.765625, 'reward_opening_diversity_std': 0.19408094882965088, 'reward_distinct_2_mean': 0.8785424828529358, 'reward_distinct_2_std': 0.09677964448928833, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.31861215829849243, 'reward_total_composite_std': 0.11181426048278809, 'epoch': 0.06}

 36%|███▌      | 1183/3300 [4:32:09<5:35:29,  9.51s/it]INFO 04-12 20:30:36 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1184/3300 [4:32:19<5:32:51,  9.44s/it]2026-04-12 20:30:45,275 | INFO | train_grpo_train | metrics_logged mode=train step=1184

                                                       
{'loss': 0.0323, 'grad_norm': 9.809541702270508, 'learning_rate': 6.415151515151515e-06, 'num_tokens': 2110150.0, 'completions/mean_length': 94.25, 'completions/min_length': 77.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.25, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9119459390640259, 'rewards/meter/std': 0.15937098860740662, 'rewards/count_adherence/mean': 0.484375, 'rewards/count_adherence/std': 0.04419417306780815, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8453125357627869, 'rewards/count_floor/std': 0.013258260674774647, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.26250001788139343, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.8172634840011597, 'rewards/repeat_penalty/std': 0.11730685085058212, 'rewards/repeat_floor/mean': 0.9783135652542114, 'rewards/repeat_floor/std': 0.013516058214008808, 'rewards/near_duplicate_penalty/mean': 0.9289640784263611, 'rewards/near_duplicate_penalty/std': 0.03674700856208801, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8774330019950867, 'rewards/distinct_2/std': 0.10283441841602325, 'rewards/total_composite/mean': 0.20477306842803955, 'rewards/total_composite/std': 0.1126212552189827, 'reward': 0.20477306842803955, 'reward_std': 0.1126212552189827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16478417813777924, 'sampling/sampling_logp_difference/max': 4.918659687042236, 'sampling/importance_sampling_ratio/min': 0.007308920379728079, 'sampling/importance_sampling_ratio/mean': 1.0037118196487427, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9794083014130592, 'clip_ratio/low_mean': 0.05482036620378494, 'clip_ratio/low_min': 0.05482036620378494, 'clip_ratio/high_mean': 0.07793444767594337, 'clip_ratio/high_max': 0.07793444767594337, 'clip_ratio/region_mean': 0.13275481387972832, 'reward_total_mean': 0.20477306842803955, 'reward_meter_mean': 0.9119459390640259, 'reward_meter_std': 0.15937098860740662, 'reward_count_adherence_mean': 0.484375, 'reward_count_adherence_std': 0.04419417306780815, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8453125357627869, 'reward_count_floor_std': 0.013258260674774647, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8172634840011597, 'reward_repeat_penalty_std': 0.11730685085058212, 'reward_repeat_floor_mean': 0.9783135652542114, 'reward_repeat_floor_std': 0.013516058214008808, 'reward_near_duplicate_penalty_mean': 0.9289640784263611, 'reward_near_duplicate_penalty_std': 0.03674700856208801, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8774330019950867, 'reward_distinct_2_std': 0.10283441841602325, 'reward_judge_quality_mean': 0.26250001788139343, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.20477306842803955, 'reward_total_composite_std': 0.1126212552189827, 'epoch': 0.06}

 36%|███▌      | 1184/3300 [4:32:19<5:32:51,  9.44s/it]INFO 04-12 20:30:45 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1185/3300 [4:32:28<5:36:07,  9.54s/it]2026-04-12 20:30:55,037 | INFO | train_grpo_train | metrics_logged mode=train step=1185

                                                       
{'loss': 0.0198, 'grad_norm': 18.055950164794922, 'learning_rate': 6.412121212121213e-06, 'num_tokens': 2112116.0, 'completions/mean_length': 73.75, 'completions/min_length': 62.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.75, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.7976605892181396, 'rewards/meter/std': 0.30140194296836853, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9134676456451416, 'rewards/repeat_penalty/std': 0.037574149668216705, 'rewards/repeat_floor/mean': 0.9891688823699951, 'rewards/repeat_floor/std': 0.0046572377905249596, 'rewards/near_duplicate_penalty/mean': 0.9521973729133606, 'rewards/near_duplicate_penalty/std': 0.025508565828204155, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.959325909614563, 'rewards/distinct_2/std': 0.02949613146483898, 'rewards/total_composite/mean': 0.26701101660728455, 'rewards/total_composite/std': 0.11397452652454376, 'reward': 0.26701101660728455, 'reward_std': 0.11397451907396317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14282502233982086, 'sampling/sampling_logp_difference/max': 2.9267032146453857, 'sampling/importance_sampling_ratio/min': 0.0535733699798584, 'sampling/importance_sampling_ratio/mean': 1.0160162448883057, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.930775836110115, 'clip_ratio/low_mean': 0.039340363815426826, 'clip_ratio/low_min': 0.039340363815426826, 'clip_ratio/high_mean': 0.09138065297156572, 'clip_ratio/high_max': 0.09138065297156572, 'clip_ratio/region_mean': 0.13072101678699255, 'reward_total_mean': 0.26701101660728455, 'reward_meter_mean': 0.7976605892181396, 'reward_meter_std': 0.30140194296836853, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9134676456451416, 'reward_repeat_penalty_std': 0.037574149668216705, 'reward_repeat_floor_mean': 0.9891688823699951, 'reward_repeat_floor_std': 0.0046572377905249596, 'reward_near_duplicate_penalty_mean': 0.9521973729133606, 'reward_near_duplicate_penalty_std': 0.025508565828204155, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.959325909614563, 'reward_distinct_2_std': 0.02949613146483898, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.26701101660728455, 'reward_total_composite_std': 0.11397452652454376, 'epoch': 0.06}

 36%|███▌      | 1185/3300 [4:32:28<5:36:07,  9.54s/it]INFO 04-12 20:30:55 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1186/3300 [4:32:38<5:34:00,  9.48s/it]2026-04-12 20:31:04,387 | INFO | train_grpo_train | metrics_logged mode=train step=1186

                                                       
{'loss': -0.0545, 'grad_norm': 11.506623268127441, 'learning_rate': 6.40909090909091e-06, 'num_tokens': 2113774.0, 'completions/mean_length': 53.25, 'completions/min_length': 43.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8872829079627991, 'rewards/meter/std': 0.16512970626354218, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5037499666213989, 'rewards/judge_quality/std': 0.177999809384346, 'rewards/repeat_penalty/mean': 0.9021210670471191, 'rewards/repeat_penalty/std': 0.0825454369187355, 'rewards/repeat_floor/mean': 0.9884834289550781, 'rewards/repeat_floor/std': 0.009969102218747139, 'rewards/near_duplicate_penalty/mean': 0.9608796834945679, 'rewards/near_duplicate_penalty/std': 0.03545333445072174, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9372386932373047, 'rewards/distinct_2/std': 0.05369313806295395, 'rewards/total_composite/mean': 0.44249606132507324, 'rewards/total_composite/std': 0.21247518062591553, 'reward': 0.44249606132507324, 'reward_std': 0.21247516572475433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14859214425086975, 'sampling/sampling_logp_difference/max': 1.6988357305526733, 'sampling/importance_sampling_ratio/min': 0.18289633095264435, 'sampling/importance_sampling_ratio/mean': 0.9840831160545349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7777009159326553, 'clip_ratio/low_mean': 0.07531295251101255, 'clip_ratio/low_min': 0.07531295251101255, 'clip_ratio/high_mean': 0.030569007620215416, 'clip_ratio/high_max': 0.030569007620215416, 'clip_ratio/region_mean': 0.10588196013122797, 'reward_total_mean': 0.44249606132507324, 'reward_meter_mean': 0.8872829079627991, 'reward_meter_std': 0.16512970626354218, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9021210670471191, 'reward_repeat_penalty_std': 0.0825454369187355, 'reward_repeat_floor_mean': 0.9884834289550781, 'reward_repeat_floor_std': 0.009969102218747139, 'reward_near_duplicate_penalty_mean': 0.9608796834945679, 'reward_near_duplicate_penalty_std': 0.03545333445072174, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9372386932373047, 'reward_distinct_2_std': 0.05369313806295395, 'reward_judge_quality_mean': 0.5037499666213989, 'reward_judge_quality_std': 0.177999809384346, 'reward_total_composite_mean': 0.44249606132507324, 'reward_total_composite_std': 0.21247518062591553, 'epoch': 0.06}

 36%|███▌      | 1186/3300 [4:32:38<5:34:00,  9.48s/it]INFO 04-12 20:31:04 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1187/3300 [4:32:45<5:16:06,  8.98s/it]2026-04-12 20:31:12,187 | INFO | train_grpo_train | metrics_logged mode=train step=1187

                                                       
{'loss': -0.0121, 'grad_norm': 17.133201599121094, 'learning_rate': 6.406060606060607e-06, 'num_tokens': 2115060.0, 'completions/mean_length': 16.75, 'completions/min_length': 15.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.8836097717285156, 'rewards/meter/std': 0.1592770516872406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/repeat_penalty/mean': 0.705302357673645, 'rewards/repeat_penalty/std': 0.04968933388590813, 'rewards/repeat_floor/mean': 0.9760604500770569, 'rewards/repeat_floor/std': 0.009937874972820282, 'rewards/near_duplicate_penalty/mean': 0.9404031038284302, 'rewards/near_duplicate_penalty/std': 0.0662524476647377, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7368852496147156, 'rewards/total_composite/std': 0.1850358545780182, 'reward': 0.7368852496147156, 'reward_std': 0.185035839676857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08635969460010529, 'sampling/sampling_logp_difference/max': 0.9684031009674072, 'sampling/importance_sampling_ratio/min': 0.4997761845588684, 'sampling/importance_sampling_ratio/mean': 1.043078064918518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49771571159362793, 'clip_ratio/low_mean': 0.014705882407724857, 'clip_ratio/low_min': 0.014705882407724857, 'clip_ratio/high_mean': 0.058986930176615715, 'clip_ratio/high_max': 0.058986930176615715, 'clip_ratio/region_mean': 0.07369281258434057, 'reward_total_mean': 0.7368852496147156, 'reward_meter_mean': 0.8836097717285156, 'reward_meter_std': 0.1592770516872406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.705302357673645, 'reward_repeat_penalty_std': 0.04968933388590813, 'reward_repeat_floor_mean': 0.9760604500770569, 'reward_repeat_floor_std': 0.009937874972820282, 'reward_near_duplicate_penalty_mean': 0.9404031038284302, 'reward_near_duplicate_penalty_std': 0.0662524476647377, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.7368852496147156, 'reward_total_composite_std': 0.1850358545780182, 'epoch': 0.06}

 36%|███▌      | 1187/3300 [4:32:46<5:16:06,  8.98s/it]INFO 04-12 20:31:12 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1188/3300 [4:32:53<5:01:13,  8.56s/it]2026-04-12 20:31:19,768 | INFO | train_grpo_train | metrics_logged mode=train step=1188

                                                       
{'loss': 0.0157, 'grad_norm': 19.255817413330078, 'learning_rate': 6.403030303030303e-06, 'num_tokens': 2116357.0, 'completions/mean_length': 17.125, 'completions/min_length': 16.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.125, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.6553815007209778, 'rewards/meter/std': 0.4051918387413025, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7725000381469727, 'rewards/judge_quality/std': 0.3027375340461731, 'rewards/repeat_penalty/mean': 0.7455357313156128, 'rewards/repeat_penalty/std': 0.012626901268959045, 'rewards/repeat_floor/mean': 0.9841071367263794, 'rewards/repeat_floor/std': 0.002525375923141837, 'rewards/near_duplicate_penalty/mean': 0.9940476417541504, 'rewards/near_duplicate_penalty/std': 0.016835875809192657, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5080605745315552, 'rewards/total_composite/std': 0.42490366101264954, 'reward': 0.5080605745315552, 'reward_std': 0.4249036908149719, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1546800136566162, 'sampling/sampling_logp_difference/max': 2.8401095867156982, 'sampling/importance_sampling_ratio/min': 0.05841926112771034, 'sampling/importance_sampling_ratio/mean': 0.9928350448608398, 'sampling/importance_sampling_ratio/max': 1.7932406663894653, 'entropy': 0.8976735547184944, 'clip_ratio/low_mean': 0.058799341320991516, 'clip_ratio/low_min': 0.058799341320991516, 'clip_ratio/high_mean': 0.054871323984116316, 'clip_ratio/high_max': 0.054871323984116316, 'clip_ratio/region_mean': 0.11367066530510783, 'reward_total_mean': 0.5080605745315552, 'reward_meter_mean': 0.6553815007209778, 'reward_meter_std': 0.4051918387413025, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7455357313156128, 'reward_repeat_penalty_std': 0.012626901268959045, 'reward_repeat_floor_mean': 0.9841071367263794, 'reward_repeat_floor_std': 0.002525375923141837, 'reward_near_duplicate_penalty_mean': 0.9940476417541504, 'reward_near_duplicate_penalty_std': 0.016835875809192657, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7725000381469727, 'reward_judge_quality_std': 0.3027375340461731, 'reward_total_composite_mean': 0.5080605745315552, 'reward_total_composite_std': 0.42490366101264954, 'epoch': 0.06}

 36%|███▌      | 1188/3300 [4:32:53<5:01:13,  8.56s/it]INFO 04-12 20:31:20 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1189/3300 [4:33:02<5:07:10,  8.73s/it]2026-04-12 20:31:28,899 | INFO | train_grpo_train | metrics_logged mode=train step=1189

                                                       
{'loss': 0.1288, 'grad_norm': 12.094244956970215, 'learning_rate': 6.4000000000000006e-06, 'num_tokens': 2117869.0, 'completions/mean_length': 39.0, 'completions/min_length': 28.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.2448902726173401, 'rewards/meter/std': 0.2321670800447464, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.7840290665626526, 'rewards/repeat_penalty/std': 0.22433924674987793, 'rewards/repeat_floor/mean': 0.9759631156921387, 'rewards/repeat_floor/std': 0.0252972561866045, 'rewards/near_duplicate_penalty/mean': 0.9217918515205383, 'rewards/near_duplicate_penalty/std': 0.07592284679412842, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.9132702350616455, 'rewards/distinct_2/std': 0.10772211104631424, 'rewards/total_composite/mean': 0.09393828362226486, 'rewards/total_composite/std': 0.08279355615377426, 'reward': 0.09393828362226486, 'reward_std': 0.08279355615377426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16901198029518127, 'sampling/sampling_logp_difference/max': 4.9106645584106445, 'sampling/importance_sampling_ratio/min': 0.007367590442299843, 'sampling/importance_sampling_ratio/mean': 1.0038750171661377, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9866219609975815, 'clip_ratio/low_mean': 0.09241046849638224, 'clip_ratio/low_min': 0.09241046849638224, 'clip_ratio/high_mean': 0.061492979526519775, 'clip_ratio/high_max': 0.061492979526519775, 'clip_ratio/region_mean': 0.153903448022902, 'reward_total_mean': 0.09393828362226486, 'reward_meter_mean': 0.2448902726173401, 'reward_meter_std': 0.2321670800447464, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7840290665626526, 'reward_repeat_penalty_std': 0.22433924674987793, 'reward_repeat_floor_mean': 0.9759631156921387, 'reward_repeat_floor_std': 0.0252972561866045, 'reward_near_duplicate_penalty_mean': 0.9217918515205383, 'reward_near_duplicate_penalty_std': 0.07592284679412842, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.9132702350616455, 'reward_distinct_2_std': 0.10772211104631424, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.09393828362226486, 'reward_total_composite_std': 0.08279355615377426, 'epoch': 0.06}

 36%|███▌      | 1189/3300 [4:33:02<5:07:10,  8.73s/it]INFO 04-12 20:31:29 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1190/3300 [4:33:16<6:04:11, 10.36s/it]2026-04-12 20:31:43,053 | INFO | train_grpo_train | metrics_logged mode=train step=1190

                                                       
{'loss': -0.1151, 'grad_norm': 3.064037561416626, 'learning_rate': 6.396969696969697e-06, 'num_tokens': 2120149.0, 'completions/mean_length': 251.0, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 94.4000015258789, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.5597401857376099, 'rewards/meter/std': 0.3874618411064148, 'rewards/count_adherence/mean': 0.535714328289032, 'rewards/count_adherence/std': 0.22587698698043823, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.860714316368103, 'rewards/count_floor/std': 0.06776309758424759, 'rewards/lexical_plausibility/mean': 0.8444586992263794, 'rewards/lexical_plausibility/std': 0.22641272842884064, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.17512750625610352, 'rewards/repeat_penalty/mean': 0.9328254461288452, 'rewards/repeat_penalty/std': 0.07999370992183685, 'rewards/repeat_floor/mean': 0.9921799898147583, 'rewards/repeat_floor/std': 0.009029122069478035, 'rewards/near_duplicate_penalty/mean': 0.9739542007446289, 'rewards/near_duplicate_penalty/std': 0.025501254945993423, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9565209746360779, 'rewards/distinct_2/std': 0.06015944480895996, 'rewards/total_composite/mean': 0.09549776464700699, 'rewards/total_composite/std': 0.10370214283466339, 'reward': 0.09549776464700699, 'reward_std': 0.1037021353840828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1545894593000412, 'sampling/sampling_logp_difference/max': 1.3918174505233765, 'sampling/importance_sampling_ratio/min': 0.24862302839756012, 'sampling/importance_sampling_ratio/mean': 1.0047316551208496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48921435326337814, 'clip_ratio/low_mean': 0.03175813052803278, 'clip_ratio/low_min': 0.03175813052803278, 'clip_ratio/high_mean': 0.055323295295238495, 'clip_ratio/high_max': 0.055323295295238495, 'clip_ratio/region_mean': 0.08708142582327127, 'reward_total_mean': 0.09549776464700699, 'reward_meter_mean': 0.5597401857376099, 'reward_meter_std': 0.3874618411064148, 'reward_count_adherence_mean': 0.535714328289032, 'reward_count_adherence_std': 0.22587698698043823, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.860714316368103, 'reward_count_floor_std': 0.06776309758424759, 'reward_lexical_plausibility_mean': 0.8444586992263794, 'reward_lexical_plausibility_std': 0.22641272842884064, 'reward_repeat_penalty_mean': 0.9328254461288452, 'reward_repeat_penalty_std': 0.07999370992183685, 'reward_repeat_floor_mean': 0.9921799898147583, 'reward_repeat_floor_std': 0.009029122069478035, 'reward_near_duplicate_penalty_mean': 0.9739542007446289, 'reward_near_duplicate_penalty_std': 0.025501254945993423, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9565209746360779, 'reward_distinct_2_std': 0.06015944480895996, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.17512750625610352, 'reward_total_composite_mean': 0.09549776464700699, 'reward_total_composite_std': 0.10370214283466339, 'epoch': 0.06}

 36%|███▌      | 1190/3300 [4:33:16<6:04:11, 10.36s/it]INFO 04-12 20:31:43 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1191/3300 [4:33:25<5:44:49,  9.81s/it]2026-04-12 20:31:51,592 | INFO | train_grpo_train | metrics_logged mode=train step=1191

                                                       
{'loss': 0.0521, 'grad_norm': 15.420903205871582, 'learning_rate': 6.393939393939394e-06, 'num_tokens': 2121809.0, 'completions/mean_length': 39.5, 'completions/min_length': 22.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.33362430334091187, 'rewards/meter/std': 0.34837689995765686, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.8976688385009766, 'rewards/repeat_penalty/std': 0.12704874575138092, 'rewards/repeat_floor/mean': 0.9910262823104858, 'rewards/repeat_floor/std': 0.00865508895367384, 'rewards/near_duplicate_penalty/mean': 0.9794062376022339, 'rewards/near_duplicate_penalty/std': 0.016195926815271378, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9762820601463318, 'rewards/distinct_2/std': 0.0473938025534153, 'rewards/total_composite/mean': 0.08406940847635269, 'rewards/total_composite/std': 0.08071156591176987, 'reward': 0.08406940847635269, 'reward_std': 0.08071156591176987, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15563789010047913, 'sampling/sampling_logp_difference/max': 1.7960224151611328, 'sampling/importance_sampling_ratio/min': 0.16595768928527832, 'sampling/importance_sampling_ratio/mean': 1.0279182195663452, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.25249882042408, 'clip_ratio/low_mean': 0.08641419466584921, 'clip_ratio/low_min': 0.08641419466584921, 'clip_ratio/high_mean': 0.052471186965703964, 'clip_ratio/high_max': 0.052471186965703964, 'clip_ratio/region_mean': 0.13888538163155317, 'reward_total_mean': 0.08406940847635269, 'reward_meter_mean': 0.33362430334091187, 'reward_meter_std': 0.34837689995765686, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.8976688385009766, 'reward_repeat_penalty_std': 0.12704874575138092, 'reward_repeat_floor_mean': 0.9910262823104858, 'reward_repeat_floor_std': 0.00865508895367384, 'reward_near_duplicate_penalty_mean': 0.9794062376022339, 'reward_near_duplicate_penalty_std': 0.016195926815271378, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9762820601463318, 'reward_distinct_2_std': 0.0473938025534153, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.08406940847635269, 'reward_total_composite_std': 0.08071156591176987, 'epoch': 0.06}

 36%|███▌      | 1191/3300 [4:33:25<5:44:49,  9.81s/it]INFO 04-12 20:31:51 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1192/3300 [4:33:33<5:29:11,  9.37s/it]2026-04-12 20:31:59,932 | INFO | train_grpo_train | metrics_logged mode=train step=1192

                                                       
{'loss': 0.0462, 'grad_norm': 13.260112762451172, 'learning_rate': 6.390909090909091e-06, 'num_tokens': 2123400.0, 'completions/mean_length': 35.875, 'completions/min_length': 21.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7292162179946899, 'rewards/meter/std': 0.33589762449264526, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.22273865342140198, 'rewards/repeat_penalty/mean': 0.7868608236312866, 'rewards/repeat_penalty/std': 0.12876349687576294, 'rewards/repeat_floor/mean': 0.9787724614143372, 'rewards/repeat_floor/std': 0.014359559863805771, 'rewards/near_duplicate_penalty/mean': 0.9416112899780273, 'rewards/near_duplicate_penalty/std': 0.05764366313815117, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9239528179168701, 'rewards/distinct_2/std': 0.07524978369474411, 'rewards/total_composite/mean': 0.30526676774024963, 'rewards/total_composite/std': 0.25670892000198364, 'reward': 0.30526676774024963, 'reward_std': 0.25670892000198364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1421644538640976, 'sampling/sampling_logp_difference/max': 1.4621171951293945, 'sampling/importance_sampling_ratio/min': 0.23174512386322021, 'sampling/importance_sampling_ratio/mean': 1.0219980478286743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0340564623475075, 'clip_ratio/low_mean': 0.1301058055832982, 'clip_ratio/low_min': 0.1301058055832982, 'clip_ratio/high_mean': 0.03627622500061989, 'clip_ratio/high_max': 0.03627622500061989, 'clip_ratio/region_mean': 0.1663820305839181, 'reward_total_mean': 0.30526676774024963, 'reward_meter_mean': 0.7292162179946899, 'reward_meter_std': 0.33589762449264526, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7868608236312866, 'reward_repeat_penalty_std': 0.12876349687576294, 'reward_repeat_floor_mean': 0.9787724614143372, 'reward_repeat_floor_std': 0.014359559863805771, 'reward_near_duplicate_penalty_mean': 0.9416112899780273, 'reward_near_duplicate_penalty_std': 0.05764366313815117, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9239528179168701, 'reward_distinct_2_std': 0.07524978369474411, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.22273865342140198, 'reward_total_composite_mean': 0.30526676774024963, 'reward_total_composite_std': 0.25670892000198364, 'epoch': 0.06}

 36%|███▌      | 1192/3300 [4:33:33<5:29:11,  9.37s/it]INFO 04-12 20:32:00 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1193/3300 [4:33:43<5:29:35,  9.39s/it]2026-04-12 20:32:09,353 | INFO | train_grpo_train | metrics_logged mode=train step=1193

                                                       
{'loss': -0.0015, 'grad_norm': 9.142109870910645, 'learning_rate': 6.387878787878789e-06, 'num_tokens': 2125750.0, 'completions/mean_length': 94.75, 'completions/min_length': 75.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.75, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9359992742538452, 'rewards/meter/std': 0.05971302092075348, 'rewards/count_adherence/mean': 0.46875, 'rewards/count_adherence/std': 0.043129101395606995, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.840624988079071, 'rewards/count_floor/std': 0.012938748113811016, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.5473600625991821, 'rewards/repeat_penalty/std': 0.16508932411670685, 'rewards/repeat_floor/mean': 0.9428822994232178, 'rewards/repeat_floor/std': 0.024883070960640907, 'rewards/near_duplicate_penalty/mean': 0.8338198661804199, 'rewards/near_duplicate_penalty/std': 0.07837218791246414, 'rewards/opening_diversity/mean': 0.9593750238418579, 'rewards/opening_diversity/std': 0.056596167385578156, 'rewards/distinct_2/mean': 0.6694088578224182, 'rewards/distinct_2/std': 0.12782509624958038, 'rewards/total_composite/mean': 0.2599286735057831, 'rewards/total_composite/std': 0.022021885961294174, 'reward': 0.2599286735057831, 'reward_std': 0.022021880373358727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13182973861694336, 'sampling/sampling_logp_difference/max': 1.8844695091247559, 'sampling/importance_sampling_ratio/min': 0.15190961956977844, 'sampling/importance_sampling_ratio/mean': 1.0104258060455322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6066080182790756, 'clip_ratio/low_mean': 0.031279259361326694, 'clip_ratio/low_min': 0.031279259361326694, 'clip_ratio/high_mean': 0.0918187159113586, 'clip_ratio/high_max': 0.0918187159113586, 'clip_ratio/region_mean': 0.12309797527268529, 'reward_total_mean': 0.2599286735057831, 'reward_meter_mean': 0.9359992742538452, 'reward_meter_std': 0.05971302092075348, 'reward_count_adherence_mean': 0.46875, 'reward_count_adherence_std': 0.043129101395606995, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.840624988079071, 'reward_count_floor_std': 0.012938748113811016, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5473600625991821, 'reward_repeat_penalty_std': 0.16508932411670685, 'reward_repeat_floor_mean': 0.9428822994232178, 'reward_repeat_floor_std': 0.024883070960640907, 'reward_near_duplicate_penalty_mean': 0.8338198661804199, 'reward_near_duplicate_penalty_std': 0.07837218791246414, 'reward_opening_diversity_mean': 0.9593750238418579, 'reward_opening_diversity_std': 0.056596167385578156, 'reward_distinct_2_mean': 0.6694088578224182, 'reward_distinct_2_std': 0.12782509624958038, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.2599286735057831, 'reward_total_composite_std': 0.022021885961294174, 'epoch': 0.06}

 36%|███▌      | 1193/3300 [4:33:43<5:29:35,  9.39s/it]INFO 04-12 20:32:09 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1194/3300 [4:33:53<5:36:25,  9.58s/it]2026-04-12 20:32:19,402 | INFO | train_grpo_train | metrics_logged mode=train step=1194

                                                       
{'loss': 0.0208, 'grad_norm': 11.12405776977539, 'learning_rate': 6.384848484848485e-06, 'num_tokens': 2128102.0, 'completions/mean_length': 103.0, 'completions/min_length': 89.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.0, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.740294337272644, 'rewards/meter/std': 0.29995906352996826, 'rewards/count_adherence/mean': 0.4861111044883728, 'rewards/count_adherence/std': 0.057505473494529724, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8458333611488342, 'rewards/count_floor/std': 0.017251653596758842, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.7783148288726807, 'rewards/repeat_penalty/std': 0.10792271047830582, 'rewards/repeat_floor/mean': 0.9739859104156494, 'rewards/repeat_floor/std': 0.013446926139295101, 'rewards/near_duplicate_penalty/mean': 0.9214696884155273, 'rewards/near_duplicate_penalty/std': 0.04053705930709839, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8418387770652771, 'rewards/distinct_2/std': 0.08892444521188736, 'rewards/total_composite/mean': 0.22537961602210999, 'rewards/total_composite/std': 0.0901152640581131, 'reward': 0.22537961602210999, 'reward_std': 0.0901152640581131, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14989331364631653, 'sampling/sampling_logp_difference/max': 2.6498146057128906, 'sampling/importance_sampling_ratio/min': 0.07066430896520615, 'sampling/importance_sampling_ratio/mean': 1.0311756134033203, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0670210123062134, 'clip_ratio/low_mean': 0.05114095564931631, 'clip_ratio/low_min': 0.05114095564931631, 'clip_ratio/high_mean': 0.08290205523371696, 'clip_ratio/high_max': 0.08290205523371696, 'clip_ratio/region_mean': 0.13404301088303328, 'reward_total_mean': 0.22537961602210999, 'reward_meter_mean': 0.740294337272644, 'reward_meter_std': 0.29995906352996826, 'reward_count_adherence_mean': 0.4861111044883728, 'reward_count_adherence_std': 0.057505473494529724, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8458333611488342, 'reward_count_floor_std': 0.017251653596758842, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7783148288726807, 'reward_repeat_penalty_std': 0.10792271047830582, 'reward_repeat_floor_mean': 0.9739859104156494, 'reward_repeat_floor_std': 0.013446926139295101, 'reward_near_duplicate_penalty_mean': 0.9214696884155273, 'reward_near_duplicate_penalty_std': 0.04053705930709839, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8418387770652771, 'reward_distinct_2_std': 0.08892444521188736, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.22537961602210999, 'reward_total_composite_std': 0.0901152640581131, 'epoch': 0.06}

 36%|███▌      | 1194/3300 [4:33:53<5:36:25,  9.58s/it]INFO 04-12 20:32:19 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1195/3300 [4:34:02<5:34:20,  9.53s/it]2026-04-12 20:32:28,830 | INFO | train_grpo_train | metrics_logged mode=train step=1195

                                                       
{'loss': -0.0036, 'grad_norm': 12.172980308532715, 'learning_rate': 6.381818181818182e-06, 'num_tokens': 2129708.0, 'completions/mean_length': 33.75, 'completions/min_length': 31.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9084851741790771, 'rewards/meter/std': 0.1901261806488037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9177228212356567, 'rewards/repeat_penalty/std': 0.046730946749448776, 'rewards/repeat_floor/mean': 0.989124059677124, 'rewards/repeat_floor/std': 0.005586832296103239, 'rewards/near_duplicate_penalty/mean': 0.9445270299911499, 'rewards/near_duplicate_penalty/std': 0.028661437332630157, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.971611738204956, 'rewards/distinct_2/std': 0.03919585049152374, 'rewards/total_composite/mean': 0.3621888756752014, 'rewards/total_composite/std': 0.09562387317419052, 'reward': 0.3621888756752014, 'reward_std': 0.09562388062477112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14849862456321716, 'sampling/sampling_logp_difference/max': 2.1155805587768555, 'sampling/importance_sampling_ratio/min': 0.12056327611207962, 'sampling/importance_sampling_ratio/mean': 1.0559884309768677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.100086785852909, 'clip_ratio/low_mean': 0.04571143467910588, 'clip_ratio/low_min': 0.04571143467910588, 'clip_ratio/high_mean': 0.059406036511063576, 'clip_ratio/high_max': 0.059406036511063576, 'clip_ratio/region_mean': 0.10511747119016945, 'reward_total_mean': 0.3621888756752014, 'reward_meter_mean': 0.9084851741790771, 'reward_meter_std': 0.1901261806488037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9177228212356567, 'reward_repeat_penalty_std': 0.046730946749448776, 'reward_repeat_floor_mean': 0.989124059677124, 'reward_repeat_floor_std': 0.005586832296103239, 'reward_near_duplicate_penalty_mean': 0.9445270299911499, 'reward_near_duplicate_penalty_std': 0.028661437332630157, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.971611738204956, 'reward_distinct_2_std': 0.03919585049152374, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.3621888756752014, 'reward_total_composite_std': 0.09562387317419052, 'epoch': 0.06}

 36%|███▌      | 1195/3300 [4:34:02<5:34:20,  9.53s/it]INFO 04-12 20:32:29 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1196/3300 [4:34:11<5:25:19,  9.28s/it]2026-04-12 20:32:37,507 | INFO | train_grpo_train | metrics_logged mode=train step=1196

                                                       
{'loss': 0.0354, 'grad_norm': 19.802976608276367, 'learning_rate': 6.37878787878788e-06, 'num_tokens': 2131321.0, 'completions/mean_length': 21.625, 'completions/min_length': 19.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.3409562110900879, 'rewards/meter/std': 0.4013894498348236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9322916269302368, 'rewards/lexical_plausibility/std': 0.13717946410179138, 'rewards/judge_quality/mean': 0.4050000309944153, 'rewards/judge_quality/std': 0.33713075518608093, 'rewards/repeat_penalty/mean': 0.6805267333984375, 'rewards/repeat_penalty/std': 0.1280427724123001, 'rewards/repeat_floor/mean': 0.9711188077926636, 'rewards/repeat_floor/std': 0.025574974715709686, 'rewards/near_duplicate_penalty/mean': 0.9163343906402588, 'rewards/near_duplicate_penalty/std': 0.14881302416324615, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.14375658333301544, 'rewards/total_composite/std': 0.2785193622112274, 'reward': 0.14375658333301544, 'reward_std': 0.27851933240890503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14784274995326996, 'sampling/sampling_logp_difference/max': 1.161886215209961, 'sampling/importance_sampling_ratio/min': 0.31289544701576233, 'sampling/importance_sampling_ratio/mean': 0.9942523241043091, 'sampling/importance_sampling_ratio/max': 1.9750866889953613, 'entropy': 1.2131140232086182, 'clip_ratio/low_mean': 0.07700497936457396, 'clip_ratio/low_min': 0.07700497936457396, 'clip_ratio/high_mean': 0.019078947603702545, 'clip_ratio/high_max': 0.019078947603702545, 'clip_ratio/region_mean': 0.0960839269682765, 'reward_total_mean': 0.14375658333301544, 'reward_meter_mean': 0.3409562110900879, 'reward_meter_std': 0.4013894498348236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9322916269302368, 'reward_lexical_plausibility_std': 0.13717946410179138, 'reward_repeat_penalty_mean': 0.6805267333984375, 'reward_repeat_penalty_std': 0.1280427724123001, 'reward_repeat_floor_mean': 0.9711188077926636, 'reward_repeat_floor_std': 0.025574974715709686, 'reward_near_duplicate_penalty_mean': 0.9163343906402588, 'reward_near_duplicate_penalty_std': 0.14881302416324615, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.4050000309944153, 'reward_judge_quality_std': 0.33713075518608093, 'reward_total_composite_mean': 0.14375658333301544, 'reward_total_composite_std': 0.2785193622112274, 'epoch': 0.06}

 36%|███▌      | 1196/3300 [4:34:11<5:25:19,  9.28s/it]INFO 04-12 20:32:37 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1197/3300 [4:34:25<6:13:49, 10.67s/it]2026-04-12 20:32:51,399 | INFO | train_grpo_train | metrics_logged mode=train step=1197

                                                       
{'loss': -0.0318, 'grad_norm': 4.761552810668945, 'learning_rate': 6.375757575757576e-06, 'num_tokens': 2132795.0, 'completions/mean_length': 97.25, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6433270573616028, 'rewards/meter/std': 0.3524598479270935, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9619851112365723, 'rewards/lexical_plausibility/std': 0.10752230137586594, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.25753986835479736, 'rewards/repeat_penalty/mean': 0.8302223086357117, 'rewards/repeat_penalty/std': 0.1848599910736084, 'rewards/repeat_floor/mean': 0.9805862307548523, 'rewards/repeat_floor/std': 0.019555971026420593, 'rewards/near_duplicate_penalty/mean': 0.9340364933013916, 'rewards/near_duplicate_penalty/std': 0.04922031611204147, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9358974695205688, 'rewards/distinct_2/std': 0.0824902281165123, 'rewards/total_composite/mean': 0.26143789291381836, 'rewards/total_composite/std': 0.18207389116287231, 'reward': 0.26143789291381836, 'reward_std': 0.18207387626171112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12284072488546371, 'sampling/sampling_logp_difference/max': 1.630645751953125, 'sampling/importance_sampling_ratio/min': 0.19580307602882385, 'sampling/importance_sampling_ratio/mean': 1.01539945602417, 'sampling/importance_sampling_ratio/max': 1.8028939962387085, 'entropy': 0.7139197960495949, 'clip_ratio/low_mean': 0.014003378339111805, 'clip_ratio/low_min': 0.014003378339111805, 'clip_ratio/high_mean': 0.06070714630186558, 'clip_ratio/high_max': 0.06070714630186558, 'clip_ratio/region_mean': 0.07471052464097738, 'reward_total_mean': 0.26143789291381836, 'reward_meter_mean': 0.6433270573616028, 'reward_meter_std': 0.3524598479270935, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9619851112365723, 'reward_lexical_plausibility_std': 0.10752230137586594, 'reward_repeat_penalty_mean': 0.8302223086357117, 'reward_repeat_penalty_std': 0.1848599910736084, 'reward_repeat_floor_mean': 0.9805862307548523, 'reward_repeat_floor_std': 0.019555971026420593, 'reward_near_duplicate_penalty_mean': 0.9340364933013916, 'reward_near_duplicate_penalty_std': 0.04922031611204147, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9358974695205688, 'reward_distinct_2_std': 0.0824902281165123, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.25753986835479736, 'reward_total_composite_mean': 0.26143789291381836, 'reward_total_composite_std': 0.18207389116287231, 'epoch': 0.06}

 36%|███▋      | 1197/3300 [4:34:25<6:13:49, 10.67s/it]INFO 04-12 20:32:51 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1198/3300 [4:34:34<5:59:36, 10.26s/it]2026-04-12 20:33:00,755 | INFO | train_grpo_train | metrics_logged mode=train step=1198

                                                       
{'loss': 0.0037, 'grad_norm': 16.240053176879883, 'learning_rate': 6.372727272727274e-06, 'num_tokens': 2134598.0, 'completions/mean_length': 47.375, 'completions/min_length': 29.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.19367700815200806, 'rewards/meter/std': 0.1618591994047165, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.8449052572250366, 'rewards/repeat_penalty/std': 0.10780539363622665, 'rewards/repeat_floor/mean': 0.9817018508911133, 'rewards/repeat_floor/std': 0.013330195099115372, 'rewards/near_duplicate_penalty/mean': 0.9405525326728821, 'rewards/near_duplicate_penalty/std': 0.05005362257361412, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.895766019821167, 'rewards/distinct_2/std': 0.07663242518901825, 'rewards/total_composite/mean': 0.05933813005685806, 'rewards/total_composite/std': 0.052925460040569305, 'reward': 0.05933813005685806, 'reward_std': 0.052925460040569305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1752455085515976, 'sampling/sampling_logp_difference/max': 2.4137229919433594, 'sampling/importance_sampling_ratio/min': 0.08948154002428055, 'sampling/importance_sampling_ratio/mean': 1.030357003211975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1912523359060287, 'clip_ratio/low_mean': 0.07709004636853933, 'clip_ratio/low_min': 0.07709004636853933, 'clip_ratio/high_mean': 0.07225293572992086, 'clip_ratio/high_max': 0.07225293572992086, 'clip_ratio/region_mean': 0.1493429820984602, 'reward_total_mean': 0.05933813005685806, 'reward_meter_mean': 0.19367700815200806, 'reward_meter_std': 0.1618591994047165, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8449052572250366, 'reward_repeat_penalty_std': 0.10780539363622665, 'reward_repeat_floor_mean': 0.9817018508911133, 'reward_repeat_floor_std': 0.013330195099115372, 'reward_near_duplicate_penalty_mean': 0.9405525326728821, 'reward_near_duplicate_penalty_std': 0.05005362257361412, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.895766019821167, 'reward_distinct_2_std': 0.07663242518901825, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.05933813005685806, 'reward_total_composite_std': 0.052925460040569305, 'epoch': 0.06}

 36%|███▋      | 1198/3300 [4:34:34<5:59:36, 10.26s/it]INFO 04-12 20:33:01 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1199/3300 [4:34:49<6:46:03, 11.60s/it]2026-04-12 20:33:15,450 | INFO | train_grpo_train | metrics_logged mode=train step=1199

                                                       
{'loss': -0.0192, 'grad_norm': 4.026417255401611, 'learning_rate': 6.3696969696969706e-06, 'num_tokens': 2136143.0, 'completions/mean_length': 85.125, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.142858505249023, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.6765033006668091, 'rewards/meter/std': 0.45151782035827637, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.8960034847259521, 'rewards/lexical_plausibility/std': 0.17476549744606018, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.7788149118423462, 'rewards/repeat_penalty/std': 0.08963186293840408, 'rewards/repeat_floor/mean': 0.9863879680633545, 'rewards/repeat_floor/std': 0.005666534882038832, 'rewards/near_duplicate_penalty/mean': 0.9967532157897949, 'rewards/near_duplicate_penalty/std': 0.009183208458125591, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21465714275836945, 'rewards/total_composite/std': 0.1751587986946106, 'reward': 0.21465714275836945, 'reward_std': 0.1751587986946106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14708156883716583, 'sampling/sampling_logp_difference/max': 1.144773006439209, 'sampling/importance_sampling_ratio/min': 0.3182961642742157, 'sampling/importance_sampling_ratio/mean': 1.0248851776123047, 'sampling/importance_sampling_ratio/max': 1.794689416885376, 'entropy': 1.215759925544262, 'clip_ratio/low_mean': 0.045703339856117964, 'clip_ratio/low_min': 0.045703339856117964, 'clip_ratio/high_mean': 0.09492244198918343, 'clip_ratio/high_max': 0.09492244198918343, 'clip_ratio/region_mean': 0.1406257818453014, 'reward_total_mean': 0.21465714275836945, 'reward_meter_mean': 0.6765033006668091, 'reward_meter_std': 0.45151782035827637, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.8960034847259521, 'reward_lexical_plausibility_std': 0.17476549744606018, 'reward_repeat_penalty_mean': 0.7788149118423462, 'reward_repeat_penalty_std': 0.08963186293840408, 'reward_repeat_floor_mean': 0.9863879680633545, 'reward_repeat_floor_std': 0.005666534882038832, 'reward_near_duplicate_penalty_mean': 0.9967532157897949, 'reward_near_duplicate_penalty_std': 0.009183208458125591, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.21465714275836945, 'reward_total_composite_std': 0.1751587986946106, 'epoch': 0.06}

 36%|███▋      | 1199/3300 [4:34:49<6:46:03, 11.60s/it]INFO 04-12 20:33:15 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1200/3300 [4:35:00<6:45:50, 11.60s/it]2026-04-12 20:33:27,035 | INFO | train_grpo_train | metrics_logged mode=train step=1200

                                                       
{'loss': 0.0866, 'grad_norm': 7.650173187255859, 'learning_rate': 6.366666666666668e-06, 'num_tokens': 2138897.0, 'completions/mean_length': 148.25, 'completions/min_length': 127.0, 'completions/max_length': 200.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.25, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 200.0, 'rewards/meter/mean': 0.8345358371734619, 'rewards/meter/std': 0.16472771763801575, 'rewards/count_adherence/mean': 0.598214328289032, 'rewards/count_adherence/std': 0.06543752551078796, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8794642686843872, 'rewards/count_floor/std': 0.01963125728070736, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.11456432938575745, 'rewards/repeat_penalty/std': 0.07127604633569717, 'rewards/repeat_floor/mean': 0.8602797389030457, 'rewards/repeat_floor/std': 0.040996063500642776, 'rewards/near_duplicate_penalty/mean': 0.6809592247009277, 'rewards/near_duplicate_penalty/std': 0.1290561556816101, 'rewards/opening_diversity/mean': 0.47280287742614746, 'rewards/opening_diversity/std': 0.24958428740501404, 'rewards/distinct_2/mean': 0.33075234293937683, 'rewards/distinct_2/std': 0.15208888053894043, 'rewards/total_composite/mean': 0.21884483098983765, 'rewards/total_composite/std': 0.07116051018238068, 'reward': 0.21884483098983765, 'reward_std': 0.07116050273180008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09359678626060486, 'sampling/sampling_logp_difference/max': 2.6280698776245117, 'sampling/importance_sampling_ratio/min': 0.07221771776676178, 'sampling/importance_sampling_ratio/mean': 1.0027965307235718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6755676865577698, 'clip_ratio/low_mean': 0.019725318998098373, 'clip_ratio/low_min': 0.019725318998098373, 'clip_ratio/high_mean': 0.08952105836942792, 'clip_ratio/high_max': 0.08952105836942792, 'clip_ratio/region_mean': 0.10924637736752629, 'reward_total_mean': 0.21884483098983765, 'reward_meter_mean': 0.8345358371734619, 'reward_meter_std': 0.16472771763801575, 'reward_count_adherence_mean': 0.598214328289032, 'reward_count_adherence_std': 0.06543752551078796, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8794642686843872, 'reward_count_floor_std': 0.01963125728070736, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.11456432938575745, 'reward_repeat_penalty_std': 0.07127604633569717, 'reward_repeat_floor_mean': 0.8602797389030457, 'reward_repeat_floor_std': 0.040996063500642776, 'reward_near_duplicate_penalty_mean': 0.6809592247009277, 'reward_near_duplicate_penalty_std': 0.1290561556816101, 'reward_opening_diversity_mean': 0.47280287742614746, 'reward_opening_diversity_std': 0.24958428740501404, 'reward_distinct_2_mean': 0.33075234293937683, 'reward_distinct_2_std': 0.15208888053894043, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.21884483098983765, 'reward_total_composite_std': 0.07116051018238068, 'epoch': 0.06}

 36%|███▋      | 1200/3300 [4:35:00<6:45:50, 11.60s/it]INFO 04-12 20:33:27 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:41,  3.77s/it]

 23%|██▎       | 3/13 [00:13<00:48,  4.86s/it]

 31%|███       | 4/13 [00:25<01:07,  7.52s/it]

 38%|███▊      | 5/13 [00:33<01:01,  7.63s/it]

 46%|████▌     | 6/13 [00:45<01:03,  9.01s/it]

 54%|█████▍    | 7/13 [00:52<00:49,  8.33s/it]

 62%|██████▏   | 8/13 [01:01<00:42,  8.45s/it]

 69%|██████▉   | 9/13 [01:09<00:34,  8.57s/it]

 77%|███████▋  | 10/13 [01:17<00:24,  8.19s/it]

 85%|████████▍ | 11/13 [01:24<00:15,  7.98s/it]WARNING 04-12 20:35:04 [detokenizer.py:246] Encountered invalid prefix detokenization error for request 12085, resetting decode stream.


 92%|█████████▏| 12/13 [01:32<00:07,  7.91s/it]

100%|██████████| 13/13 [01:39<00:00,  7.67s/it]2026-04-12 20:35:18,831 | INFO | train_grpo_train | metrics_logged mode=eval step=1200

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 111.625, 'eval_samples_per_second': 0.932, 'eval_steps_per_second': 0.116, 'eval_num_tokens': 2138897.0, 'eval_completions/mean_length': 99.53846153846153, 'eval_completions/min_length': 28.23076923076923, 'eval_completions/max_length': 258.6923076923077, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 87.33379246638371, 'eval_completions/min_terminated_length': 28.23076923076923, 'eval_completions/max_terminated_length': 190.53846153846155, 'eval_rewards/meter/mean': 0.5399709687783167, 'eval_rewards/meter/std': 0.403679588666329, 'eval_rewards/count_adherence/mean': 0.6893736857634324, 'eval_rewards/count_adherence/std': 0.17183029307768896, 'eval_rewards/arabic_clean/mean': 0.9423076923076923, 'eval_rewards/arabic_clean/std': 0.12140072079805228, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.906812099310068, 'eval_rewards/count_floor/std': 0.05154908906954985, 'eval_rewards/lexical_plausibility/mean': 0.9854171963838431, 'eval_rewards/lexical_plausibility/std': 0.041246407880232885, 'eval_rewards/judge_quality/mean': 0.35634615329595715, 'eval_rewards/judge_quality/std': 0.17094641694655785, 'eval_rewards/repeat_penalty/mean': 0.7649480929741492, 'eval_rewards/repeat_penalty/std': 0.22522002802445337, 'eval_rewards/repeat_floor/mean': 0.9734325546484727, 'eval_rewards/repeat_floor/std': 0.026794853405310556, 'eval_rewards/near_duplicate_penalty/mean': 0.9295695607478802, 'eval_rewards/near_duplicate_penalty/std': 0.0630534474666302, 'eval_rewards/opening_diversity/mean': 0.9619390826958877, 'eval_rewards/opening_diversity/std': 0.0756664781902845, 'eval_rewards/distinct_2/mean': 0.8475641654087946, 'eval_rewards/distinct_2/std': 0.1847701961031327, 'eval_rewards/total_composite/mean': 0.16604607437665647, 'eval_rewards/total_composite/std': 0.15247203266391388, 'eval_reward': 0.16604607437665647, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07127173588826106, 'eval_sampling/sampling_logp_difference/max': 1.1292851521418645, 'eval_sampling/importance_sampling_ratio/min': 0.33061610276882464, 'eval_sampling/importance_sampling_ratio/mean': 1.0207775556124175, 'eval_sampling/importance_sampling_ratio/max': 1.5605387045786931, 'eval_entropy': 0.8297604001485385, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.16604607437665647, 'eval_reward_meter_mean': 0.5399709687783167, 'eval_reward_meter_std': 0.403679588666329, 'eval_reward_count_adherence_mean': 0.6893736857634324, 'eval_reward_count_adherence_std': 0.17183029307768896, 'eval_reward_arabic_clean_mean': 0.9423076923076923, 'eval_reward_arabic_clean_std': 0.12140072079805228, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.906812099310068, 'eval_reward_count_floor_std': 0.05154908906954985, 'eval_reward_lexical_plausibility_mean': 0.9854171963838431, 'eval_reward_lexical_plausibility_std': 0.041246407880232885, 'eval_reward_repeat_penalty_mean': 0.7649480929741492, 'eval_reward_repeat_penalty_std': 0.22522002802445337, 'eval_reward_repeat_floor_mean': 0.9734325546484727, 'eval_reward_repeat_floor_std': 0.026794853405310556, 'eval_reward_near_duplicate_penalty_mean': 0.9295695607478802, 'eval_reward_near_duplicate_penalty_std': 0.0630534474666302, 'eval_reward_opening_diversity_mean': 0.9619390826958877, 'eval_reward_opening_diversity_std': 0.0756664781902845, 'eval_reward_distinct_2_mean': 0.8475641654087946, 'eval_reward_distinct_2_std': 0.1847701961031327, 'eval_reward_judge_quality_mean': 0.35634615329595715, 'eval_reward_judge_quality_std': 0.17094641694655785, 'eval_reward_total_composite_mean': 0.16604607437665647, 'eval_reward_total_composite_std': 0.15247203266391388, 'epoch': 0.06}

 36%|███▋      | 1200/3300 [4:36:52<6:45:50, 11.60s/it]

100%|██████████| 13/13 [01:39<00:00,  7.67s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 36%|███▋      | 1201/3300 [4:37:10<27:21:50, 46.93s/it]2026-04-12 20:35:36,441 | INFO | train_grpo_train | metrics_logged mode=train step=1201

                                                        
{'loss': -0.072, 'grad_norm': 7.008788108825684, 'learning_rate': 6.363636363636364e-06, 'num_tokens': 2140397.0, 'completions/mean_length': 97.5, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.28571701049805, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8029114603996277, 'rewards/meter/std': 0.35101836919784546, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9611736536026001, 'rewards/lexical_plausibility/std': 0.10981755703687668, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.3071993589401245, 'rewards/repeat_penalty/mean': 0.9936205744743347, 'rewards/repeat_penalty/std': 0.009501545690000057, 'rewards/repeat_floor/mean': 0.9990431070327759, 'rewards/repeat_floor/std': 0.0014252327382564545, 'rewards/near_duplicate_penalty/mean': 0.9936205744743347, 'rewards/near_duplicate_penalty/std': 0.009501545690000057, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.48663556575775146, 'rewards/total_composite/std': 0.33136990666389465, 'reward': 0.48663556575775146, 'reward_std': 0.33136990666389465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14729870855808258, 'sampling/sampling_logp_difference/max': 1.2603862285614014, 'sampling/importance_sampling_ratio/min': 0.28354448080062866, 'sampling/importance_sampling_ratio/mean': 1.0013937950134277, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7681055665016174, 'clip_ratio/low_mean': 0.08650779351592064, 'clip_ratio/low_min': 0.08650779351592064, 'clip_ratio/high_mean': 0.03159098140895367, 'clip_ratio/high_max': 0.03159098140895367, 'clip_ratio/region_mean': 0.1180987749248743, 'reward_total_mean': 0.48663556575775146, 'reward_meter_mean': 0.8029114603996277, 'reward_meter_std': 0.35101836919784546, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9611736536026001, 'reward_lexical_plausibility_std': 0.10981755703687668, 'reward_repeat_penalty_mean': 0.9936205744743347, 'reward_repeat_penalty_std': 0.009501545690000057, 'reward_repeat_floor_mean': 0.9990431070327759, 'reward_repeat_floor_std': 0.0014252327382564545, 'reward_near_duplicate_penalty_mean': 0.9936205744743347, 'reward_near_duplicate_penalty_std': 0.009501545690000057, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.3071993589401245, 'reward_total_composite_mean': 0.48663556575775146, 'reward_total_composite_std': 0.33136990666389465, 'epoch': 0.06}

 36%|███▋      | 1201/3300 [4:37:10<27:21:50, 46.93s/it]INFO 04-12 20:35:36 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1202/3300 [4:37:18<20:40:45, 35.48s/it]2026-04-12 20:35:45,229 | INFO | train_grpo_train | metrics_logged mode=train step=1202

                                                        
{'loss': 0.0206, 'grad_norm': 16.42057228088379, 'learning_rate': 6.3606060606060615e-06, 'num_tokens': 2141842.0, 'completions/mean_length': 37.625, 'completions/min_length': 34.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9312632083892822, 'rewards/meter/std': 0.06470198929309845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.7945879697799683, 'rewards/repeat_penalty/std': 0.169195756316185, 'rewards/repeat_floor/mean': 0.9748214483261108, 'rewards/repeat_floor/std': 0.019403159618377686, 'rewards/near_duplicate_penalty/mean': 0.9054898619651794, 'rewards/near_duplicate_penalty/std': 0.05964653939008713, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.898588240146637, 'rewards/distinct_2/std': 0.10835838317871094, 'rewards/total_composite/mean': 0.3089689016342163, 'rewards/total_composite/std': 0.1473286747932434, 'reward': 0.3089689016342163, 'reward_std': 0.14732865989208221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12740515172481537, 'sampling/sampling_logp_difference/max': 1.2794170379638672, 'sampling/importance_sampling_ratio/min': 0.2781994342803955, 'sampling/importance_sampling_ratio/mean': 1.0375691652297974, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9828255251049995, 'clip_ratio/low_mean': 0.04012907110154629, 'clip_ratio/low_min': 0.04012907110154629, 'clip_ratio/high_mean': 0.05993717163801193, 'clip_ratio/high_max': 0.05993717163801193, 'clip_ratio/region_mean': 0.10006624273955822, 'reward_total_mean': 0.3089689016342163, 'reward_meter_mean': 0.9312632083892822, 'reward_meter_std': 0.06470198929309845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7945879697799683, 'reward_repeat_penalty_std': 0.169195756316185, 'reward_repeat_floor_mean': 0.9748214483261108, 'reward_repeat_floor_std': 0.019403159618377686, 'reward_near_duplicate_penalty_mean': 0.9054898619651794, 'reward_near_duplicate_penalty_std': 0.05964653939008713, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.898588240146637, 'reward_distinct_2_std': 0.10835838317871094, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.3089689016342163, 'reward_total_composite_std': 0.1473286747932434, 'epoch': 0.06}

 36%|███▋      | 1202/3300 [4:37:19<20:40:45, 35.48s/it]INFO 04-12 20:35:45 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1203/3300 [4:37:27<15:54:05, 27.30s/it]2026-04-12 20:35:53,395 | INFO | train_grpo_train | metrics_logged mode=train step=1203

                                                        
{'loss': 0.0639, 'grad_norm': 21.74378204345703, 'learning_rate': 6.357575757575758e-06, 'num_tokens': 2143520.0, 'completions/mean_length': 34.75, 'completions/min_length': 32.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6564161777496338, 'rewards/meter/std': 0.37329354882240295, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.9125479459762573, 'rewards/repeat_penalty/std': 0.09410092979669571, 'rewards/repeat_floor/mean': 0.9882010817527771, 'rewards/repeat_floor/std': 0.01300717331469059, 'rewards/near_duplicate_penalty/mean': 0.9395591020584106, 'rewards/near_duplicate_penalty/std': 0.06932149082422256, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9696356058120728, 'rewards/distinct_2/std': 0.04190681502223015, 'rewards/total_composite/mean': 0.4698362946510315, 'rewards/total_composite/std': 0.3359878361225128, 'reward': 0.4698362946510315, 'reward_std': 0.3359878361225128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12448786944150925, 'sampling/sampling_logp_difference/max': 1.6460838317871094, 'sampling/importance_sampling_ratio/min': 0.1928034871816635, 'sampling/importance_sampling_ratio/mean': 0.9984982013702393, 'sampling/importance_sampling_ratio/max': 1.8500926494598389, 'entropy': 0.6688754260540009, 'clip_ratio/low_mean': 0.05910828988999128, 'clip_ratio/low_min': 0.05910828988999128, 'clip_ratio/high_mean': 0.03399342764168978, 'clip_ratio/high_max': 0.03399342764168978, 'clip_ratio/region_mean': 0.09310171753168106, 'reward_total_mean': 0.4698362946510315, 'reward_meter_mean': 0.6564161777496338, 'reward_meter_std': 0.37329354882240295, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9125479459762573, 'reward_repeat_penalty_std': 0.09410092979669571, 'reward_repeat_floor_mean': 0.9882010817527771, 'reward_repeat_floor_std': 0.01300717331469059, 'reward_near_duplicate_penalty_mean': 0.9395591020584106, 'reward_near_duplicate_penalty_std': 0.06932149082422256, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9696356058120728, 'reward_distinct_2_std': 0.04190681502223015, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.4698362946510315, 'reward_total_composite_std': 0.3359878361225128, 'epoch': 0.06}

 36%|███▋      | 1203/3300 [4:37:27<15:54:05, 27.30s/it]INFO 04-12 20:35:53 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1204/3300 [4:37:36<12:50:23, 22.05s/it]2026-04-12 20:36:03,205 | INFO | train_grpo_train | metrics_logged mode=train step=1204

                                                        
{'loss': 0.1669, 'grad_norm': 10.4877347946167, 'learning_rate': 6.354545454545455e-06, 'num_tokens': 2145527.0, 'completions/mean_length': 63.875, 'completions/min_length': 41.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.733081042766571, 'rewards/meter/std': 0.33875375986099243, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8968750238418579, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9397873878479004, 'rewards/repeat_penalty/std': 0.05229237675666809, 'rewards/repeat_floor/mean': 0.991862416267395, 'rewards/repeat_floor/std': 0.00680643180385232, 'rewards/near_duplicate_penalty/mean': 0.9566202163696289, 'rewards/near_duplicate_penalty/std': 0.03369787707924843, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9818822145462036, 'rewards/distinct_2/std': 0.025771545246243477, 'rewards/total_composite/mean': 0.16590189933776855, 'rewards/total_composite/std': 0.09892627596855164, 'reward': 0.16590189933776855, 'reward_std': 0.09892627596855164, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14697673916816711, 'sampling/sampling_logp_difference/max': 1.9397528171539307, 'sampling/importance_sampling_ratio/min': 0.1437394767999649, 'sampling/importance_sampling_ratio/mean': 1.0007494688034058, 'sampling/importance_sampling_ratio/max': 1.9181885719299316, 'entropy': 0.9829893261194229, 'clip_ratio/low_mean': 0.05677100457251072, 'clip_ratio/low_min': 0.05677100457251072, 'clip_ratio/high_mean': 0.06569111905992031, 'clip_ratio/high_max': 0.06569111905992031, 'clip_ratio/region_mean': 0.12246212363243103, 'reward_total_mean': 0.16590189933776855, 'reward_meter_mean': 0.733081042766571, 'reward_meter_std': 0.33875375986099243, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8968750238418579, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9397873878479004, 'reward_repeat_penalty_std': 0.05229237675666809, 'reward_repeat_floor_mean': 0.991862416267395, 'reward_repeat_floor_std': 0.00680643180385232, 'reward_near_duplicate_penalty_mean': 0.9566202163696289, 'reward_near_duplicate_penalty_std': 0.03369787707924843, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9818822145462036, 'reward_distinct_2_std': 0.025771545246243477, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.16590189933776855, 'reward_total_composite_std': 0.09892627596855164, 'epoch': 0.06}

 36%|███▋      | 1204/3300 [4:37:37<12:50:23, 22.05s/it]INFO 04-12 20:36:03 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1205/3300 [4:37:45<10:31:33, 18.09s/it]2026-04-12 20:36:12,049 | INFO | train_grpo_train | metrics_logged mode=train step=1205

                                                        
{'loss': 0.0233, 'grad_norm': 21.569499969482422, 'learning_rate': 6.3515151515151516e-06, 'num_tokens': 2146853.0, 'completions/mean_length': 18.75, 'completions/min_length': 17.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7382795810699463, 'rewards/meter/std': 0.4383016526699066, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.36286312341690063, 'rewards/repeat_penalty/mean': 0.7296368479728699, 'rewards/repeat_penalty/std': 0.14529210329055786, 'rewards/repeat_floor/mean': 0.9770223498344421, 'rewards/repeat_floor/std': 0.020359160378575325, 'rewards/near_duplicate_penalty/mean': 0.9448052048683167, 'rewards/near_duplicate_penalty/std': 0.1022009328007698, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.35865330696105957, 'rewards/total_composite/std': 0.3571521043777466, 'reward': 0.35865330696105957, 'reward_std': 0.35715213418006897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17650271952152252, 'sampling/sampling_logp_difference/max': 1.5873017311096191, 'sampling/importance_sampling_ratio/min': 0.20447658002376556, 'sampling/importance_sampling_ratio/mean': 0.9966835975646973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0803226828575134, 'clip_ratio/low_mean': 0.09669086989015341, 'clip_ratio/low_min': 0.09669086989015341, 'clip_ratio/high_mean': 0.03947368264198303, 'clip_ratio/high_max': 0.03947368264198303, 'clip_ratio/region_mean': 0.13616455253213644, 'reward_total_mean': 0.35865330696105957, 'reward_meter_mean': 0.7382795810699463, 'reward_meter_std': 0.4383016526699066, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7296368479728699, 'reward_repeat_penalty_std': 0.14529210329055786, 'reward_repeat_floor_mean': 0.9770223498344421, 'reward_repeat_floor_std': 0.020359160378575325, 'reward_near_duplicate_penalty_mean': 0.9448052048683167, 'reward_near_duplicate_penalty_std': 0.1022009328007698, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.36286312341690063, 'reward_total_composite_mean': 0.35865330696105957, 'reward_total_composite_std': 0.3571521043777466, 'epoch': 0.06}

 37%|███▋      | 1205/3300 [4:37:45<10:31:33, 18.09s/it]INFO 04-12 20:36:12 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1206/3300 [4:37:54<8:52:19, 15.25s/it] 2026-04-12 20:36:20,673 | INFO | train_grpo_train | metrics_logged mode=train step=1206

                                                       
{'loss': -0.0109, 'grad_norm': 15.907942771911621, 'learning_rate': 6.34848484848485e-06, 'num_tokens': 2148488.0, 'completions/mean_length': 28.375, 'completions/min_length': 24.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.5188062191009521, 'rewards/meter/std': 0.39765051007270813, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9719680547714233, 'rewards/repeat_penalty/std': 0.03793061897158623, 'rewards/repeat_floor/mean': 0.9957951903343201, 'rewards/repeat_floor/std': 0.005689589306712151, 'rewards/near_duplicate_penalty/mean': 0.9719680547714233, 'rewards/near_duplicate_penalty/std': 0.03793061897158623, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.234795480966568, 'rewards/total_composite/std': 0.17182940244674683, 'reward': 0.234795480966568, 'reward_std': 0.17182940244674683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13703396916389465, 'sampling/sampling_logp_difference/max': 1.294355869293213, 'sampling/importance_sampling_ratio/min': 0.27407434582710266, 'sampling/importance_sampling_ratio/mean': 0.9839217662811279, 'sampling/importance_sampling_ratio/max': 1.9476102590560913, 'entropy': 0.9434890896081924, 'clip_ratio/low_mean': 0.0903625562787056, 'clip_ratio/low_min': 0.0903625562787056, 'clip_ratio/high_mean': 0.060693761333823204, 'clip_ratio/high_max': 0.060693761333823204, 'clip_ratio/region_mean': 0.1510563176125288, 'reward_total_mean': 0.234795480966568, 'reward_meter_mean': 0.5188062191009521, 'reward_meter_std': 0.39765051007270813, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9719680547714233, 'reward_repeat_penalty_std': 0.03793061897158623, 'reward_repeat_floor_mean': 0.9957951903343201, 'reward_repeat_floor_std': 0.005689589306712151, 'reward_near_duplicate_penalty_mean': 0.9719680547714233, 'reward_near_duplicate_penalty_std': 0.03793061897158623, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.234795480966568, 'reward_total_composite_std': 0.17182940244674683, 'epoch': 0.06}

 37%|███▋      | 1206/3300 [4:37:54<8:52:19, 15.25s/it]INFO 04-12 20:36:20 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1207/3300 [4:38:08<8:35:47, 14.79s/it]2026-04-12 20:36:34,366 | INFO | train_grpo_train | metrics_logged mode=train step=1207

                                                       
{'loss': -0.0949, 'grad_norm': 3.60880708694458, 'learning_rate': 6.345454545454546e-06, 'num_tokens': 2150185.0, 'completions/mean_length': 107.125, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.28571701049805, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7138907313346863, 'rewards/meter/std': 0.39355406165122986, 'rewards/count_adherence/mean': 0.59375, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8781250715255737, 'rewards/count_floor/std': 0.05580179765820503, 'rewards/lexical_plausibility/mean': 0.9553365111351013, 'rewards/lexical_plausibility/std': 0.12632741034030914, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.7574905157089233, 'rewards/repeat_penalty/std': 0.1678810715675354, 'rewards/repeat_floor/mean': 0.9705373048782349, 'rewards/repeat_floor/std': 0.02032845839858055, 'rewards/near_duplicate_penalty/mean': 0.8933776617050171, 'rewards/near_duplicate_penalty/std': 0.08141732960939407, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.8795071840286255, 'rewards/distinct_2/std': 0.10103638470172882, 'rewards/total_composite/mean': 0.2213066965341568, 'rewards/total_composite/std': 0.13975903391838074, 'reward': 0.2213066965341568, 'reward_std': 0.13975903391838074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18231259286403656, 'sampling/sampling_logp_difference/max': 4.9658894538879395, 'sampling/importance_sampling_ratio/min': 0.006971747148782015, 'sampling/importance_sampling_ratio/mean': 0.9963338375091553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8248138725757599, 'clip_ratio/low_mean': 0.02142857201397419, 'clip_ratio/low_min': 0.02142857201397419, 'clip_ratio/high_mean': 0.12161261029541492, 'clip_ratio/high_max': 0.12161261029541492, 'clip_ratio/region_mean': 0.14304118230938911, 'reward_total_mean': 0.2213066965341568, 'reward_meter_mean': 0.7138907313346863, 'reward_meter_std': 0.39355406165122986, 'reward_count_adherence_mean': 0.59375, 'reward_count_adherence_std': 0.18600596487522125, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8781250715255737, 'reward_count_floor_std': 0.05580179765820503, 'reward_lexical_plausibility_mean': 0.9553365111351013, 'reward_lexical_plausibility_std': 0.12632741034030914, 'reward_repeat_penalty_mean': 0.7574905157089233, 'reward_repeat_penalty_std': 0.1678810715675354, 'reward_repeat_floor_mean': 0.9705373048782349, 'reward_repeat_floor_std': 0.02032845839858055, 'reward_near_duplicate_penalty_mean': 0.8933776617050171, 'reward_near_duplicate_penalty_std': 0.08141732960939407, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.8795071840286255, 'reward_distinct_2_std': 0.10103638470172882, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2213066965341568, 'reward_total_composite_std': 0.13975903391838074, 'epoch': 0.06}

 37%|███▋      | 1207/3300 [4:38:08<8:35:47, 14.79s/it]INFO 04-12 20:36:34 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1208/3300 [4:38:16<7:30:03, 12.91s/it]2026-04-12 20:36:42,890 | INFO | train_grpo_train | metrics_logged mode=train step=1208

                                                       
{'loss': 0.0455, 'grad_norm': 23.564870834350586, 'learning_rate': 6.342424242424243e-06, 'num_tokens': 2151509.0, 'completions/mean_length': 17.5, 'completions/min_length': 14.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7611930966377258, 'rewards/meter/std': 0.3555090129375458, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.7329545617103577, 'rewards/repeat_penalty/std': 0.04821182042360306, 'rewards/repeat_floor/mean': 0.9815909266471863, 'rewards/repeat_floor/std': 0.009642376564443111, 'rewards/near_duplicate_penalty/mean': 0.9772727489471436, 'rewards/near_duplicate_penalty/std': 0.06428243964910507, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2852109372615814, 'rewards/total_composite/std': 0.13652414083480835, 'reward': 0.2852109372615814, 'reward_std': 0.13652415573596954, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1368054747581482, 'sampling/sampling_logp_difference/max': 1.1629854440689087, 'sampling/importance_sampling_ratio/min': 0.31255167722702026, 'sampling/importance_sampling_ratio/mean': 1.0286072492599487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.042983703315258, 'clip_ratio/low_mean': 0.031746032647788525, 'clip_ratio/low_min': 0.031746032647788525, 'clip_ratio/high_mean': 0.07504438143223524, 'clip_ratio/high_max': 0.07504438143223524, 'clip_ratio/region_mean': 0.10679041408002377, 'reward_total_mean': 0.2852109372615814, 'reward_meter_mean': 0.7611930966377258, 'reward_meter_std': 0.3555090129375458, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7329545617103577, 'reward_repeat_penalty_std': 0.04821182042360306, 'reward_repeat_floor_mean': 0.9815909266471863, 'reward_repeat_floor_std': 0.009642376564443111, 'reward_near_duplicate_penalty_mean': 0.9772727489471436, 'reward_near_duplicate_penalty_std': 0.06428243964910507, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.2173829823732376, 'reward_total_composite_mean': 0.2852109372615814, 'reward_total_composite_std': 0.13652414083480835, 'epoch': 0.06}

 37%|███▋      | 1208/3300 [4:38:16<7:30:03, 12.91s/it]INFO 04-12 20:36:43 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1209/3300 [4:38:30<7:40:14, 13.21s/it]2026-04-12 20:36:56,803 | INFO | train_grpo_train | metrics_logged mode=train step=1209

                                                       
{'loss': -0.0248, 'grad_norm': 5.25899600982666, 'learning_rate': 6.33939393939394e-06, 'num_tokens': 2153559.0, 'completions/mean_length': 193.25, 'completions/min_length': 83.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.20541584491729736, 'rewards/meter/std': 0.2512703239917755, 'rewards/count_adherence/mean': 0.6458333730697632, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8937499523162842, 'rewards/count_floor/std': 0.017677653580904007, 'rewards/lexical_plausibility/mean': 0.9278985261917114, 'rewards/lexical_plausibility/std': 0.13665519654750824, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1505940705537796, 'rewards/repeat_penalty/mean': 0.9035552740097046, 'rewards/repeat_penalty/std': 0.053205035626888275, 'rewards/repeat_floor/mean': 0.9888391494750977, 'rewards/repeat_floor/std': 0.005328211933374405, 'rewards/near_duplicate_penalty/mean': 0.9597481489181519, 'rewards/near_duplicate_penalty/std': 0.014031111262738705, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9482852220535278, 'rewards/distinct_2/std': 0.035729460418224335, 'rewards/total_composite/mean': 0.030430592596530914, 'rewards/total_composite/std': 0.02682132087647915, 'reward': 0.030430592596530914, 'reward_std': 0.026821322739124298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20513655245304108, 'sampling/sampling_logp_difference/max': 6.901684761047363, 'sampling/importance_sampling_ratio/min': 0.0010060889180749655, 'sampling/importance_sampling_ratio/mean': 1.0060534477233887, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.75008674710989, 'clip_ratio/low_mean': 0.020650184247642756, 'clip_ratio/low_min': 0.020650184247642756, 'clip_ratio/high_mean': 0.07043274398893118, 'clip_ratio/high_max': 0.07043274398893118, 'clip_ratio/region_mean': 0.09108292823657393, 'reward_total_mean': 0.030430592596530914, 'reward_meter_mean': 0.20541584491729736, 'reward_meter_std': 0.2512703239917755, 'reward_count_adherence_mean': 0.6458333730697632, 'reward_count_adherence_std': 0.0589255727827549, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8937499523162842, 'reward_count_floor_std': 0.017677653580904007, 'reward_lexical_plausibility_mean': 0.9278985261917114, 'reward_lexical_plausibility_std': 0.13665519654750824, 'reward_repeat_penalty_mean': 0.9035552740097046, 'reward_repeat_penalty_std': 0.053205035626888275, 'reward_repeat_floor_mean': 0.9888391494750977, 'reward_repeat_floor_std': 0.005328211933374405, 'reward_near_duplicate_penalty_mean': 0.9597481489181519, 'reward_near_duplicate_penalty_std': 0.014031111262738705, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9482852220535278, 'reward_distinct_2_std': 0.035729460418224335, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1505940705537796, 'reward_total_composite_mean': 0.030430592596530914, 'reward_total_composite_std': 0.02682132087647915, 'epoch': 0.06}

 37%|███▋      | 1209/3300 [4:38:30<7:40:14, 13.21s/it]INFO 04-12 20:36:57 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1210/3300 [4:38:38<6:47:07, 11.69s/it]2026-04-12 20:37:04,937 | INFO | train_grpo_train | metrics_logged mode=train step=1210

                                                       
{'loss': 0.0911, 'grad_norm': 21.43655776977539, 'learning_rate': 6.336363636363637e-06, 'num_tokens': 2155006.0, 'completions/mean_length': 18.875, 'completions/min_length': 17.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.752964437007904, 'rewards/meter/std': 0.3473961055278778, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27183228731155396, 'rewards/total_composite/std': 0.16105426847934723, 'reward': 0.27183228731155396, 'reward_std': 0.16105428338050842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15867464244365692, 'sampling/sampling_logp_difference/max': 2.0039024353027344, 'sampling/importance_sampling_ratio/min': 0.13480816781520844, 'sampling/importance_sampling_ratio/mean': 1.0121946334838867, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1300195679068565, 'clip_ratio/low_mean': 0.05700188176706433, 'clip_ratio/low_min': 0.05700188176706433, 'clip_ratio/high_mean': 0.06751867523416877, 'clip_ratio/high_max': 0.06751867523416877, 'clip_ratio/region_mean': 0.1245205570012331, 'reward_total_mean': 0.27183228731155396, 'reward_meter_mean': 0.752964437007904, 'reward_meter_std': 0.3473961055278778, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.27183228731155396, 'reward_total_composite_std': 0.16105426847934723, 'epoch': 0.06}

 37%|███▋      | 1210/3300 [4:38:38<6:47:07, 11.69s/it]INFO 04-12 20:37:05 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1211/3300 [4:38:52<7:09:31, 12.34s/it]2026-04-12 20:37:18,802 | INFO | train_grpo_train | metrics_logged mode=train step=1211

                                                       
{'loss': 0.0038, 'grad_norm': 14.992835998535156, 'learning_rate': 6.333333333333333e-06, 'num_tokens': 2156592.0, 'completions/mean_length': 92.25, 'completions/min_length': 19.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.28571701049805, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.4348967969417572, 'rewards/meter/std': 0.4210846424102783, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 0.9319894313812256, 'rewards/lexical_plausibility/std': 0.1379537731409073, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.8651514053344727, 'rewards/repeat_penalty/std': 0.0973212793469429, 'rewards/repeat_floor/mean': 0.9892520904541016, 'rewards/repeat_floor/std': 0.004238321911543608, 'rewards/near_duplicate_penalty/mean': 0.976586639881134, 'rewards/near_duplicate_penalty/std': 0.030550716444849968, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9821012020111084, 'rewards/distinct_2/std': 0.03315402567386627, 'rewards/total_composite/mean': 0.09467208385467529, 'rewards/total_composite/std': 0.1028226688504219, 'reward': 0.09467208385467529, 'reward_std': 0.1028226688504219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1479664295911789, 'sampling/sampling_logp_difference/max': 1.4311161041259766, 'sampling/importance_sampling_ratio/min': 0.2390419840812683, 'sampling/importance_sampling_ratio/mean': 1.0245029926300049, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0029484778642654, 'clip_ratio/low_mean': 0.0645554643124342, 'clip_ratio/low_min': 0.0645554643124342, 'clip_ratio/high_mean': 0.03510458953678608, 'clip_ratio/high_max': 0.03510458953678608, 'clip_ratio/region_mean': 0.09966005384922028, 'reward_total_mean': 0.09467208385467529, 'reward_meter_mean': 0.4348967969417572, 'reward_meter_std': 0.4210846424102783, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 0.9319894313812256, 'reward_lexical_plausibility_std': 0.1379537731409073, 'reward_repeat_penalty_mean': 0.8651514053344727, 'reward_repeat_penalty_std': 0.0973212793469429, 'reward_repeat_floor_mean': 0.9892520904541016, 'reward_repeat_floor_std': 0.004238321911543608, 'reward_near_duplicate_penalty_mean': 0.976586639881134, 'reward_near_duplicate_penalty_std': 0.030550716444849968, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9821012020111084, 'reward_distinct_2_std': 0.03315402567386627, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.09467208385467529, 'reward_total_composite_std': 0.1028226688504219, 'epoch': 0.06}

 37%|███▋      | 1211/3300 [4:38:52<7:09:31, 12.34s/it]INFO 04-12 20:37:19 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1212/3300 [4:39:06<7:22:50, 12.73s/it]2026-04-12 20:37:32,435 | INFO | train_grpo_train | metrics_logged mode=train step=1212

                                                       
{'loss': -0.1068, 'grad_norm': 4.133994102478027, 'learning_rate': 6.330303030303031e-06, 'num_tokens': 2158403.0, 'completions/mean_length': 180.375, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 69.83333587646484, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7337206602096558, 'rewards/meter/std': 0.3387097120285034, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8799999952316284, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.935899555683136, 'rewards/lexical_plausibility/std': 0.12873761355876923, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.8237954378128052, 'rewards/repeat_penalty/std': 0.13474205136299133, 'rewards/repeat_floor/mean': 0.9798691272735596, 'rewards/repeat_floor/std': 0.014733619056642056, 'rewards/near_duplicate_penalty/mean': 0.9424534440040588, 'rewards/near_duplicate_penalty/std': 0.030256416648626328, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.8841394186019897, 'rewards/distinct_2/std': 0.09345018118619919, 'rewards/total_composite/mean': 0.1660452038049698, 'rewards/total_composite/std': 0.13699889183044434, 'reward': 0.1660452038049698, 'reward_std': 0.13699889183044434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16011005640029907, 'sampling/sampling_logp_difference/max': 1.9461746215820312, 'sampling/importance_sampling_ratio/min': 0.1428193598985672, 'sampling/importance_sampling_ratio/mean': 1.0197577476501465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7529733255505562, 'clip_ratio/low_mean': 0.05727875977754593, 'clip_ratio/low_min': 0.05727875977754593, 'clip_ratio/high_mean': 0.050101449713110924, 'clip_ratio/high_max': 0.050101449713110924, 'clip_ratio/region_mean': 0.10738020949065685, 'reward_total_mean': 0.1660452038049698, 'reward_meter_mean': 0.7337206602096558, 'reward_meter_std': 0.3387097120285034, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8799999952316284, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.935899555683136, 'reward_lexical_plausibility_std': 0.12873761355876923, 'reward_repeat_penalty_mean': 0.8237954378128052, 'reward_repeat_penalty_std': 0.13474205136299133, 'reward_repeat_floor_mean': 0.9798691272735596, 'reward_repeat_floor_std': 0.014733619056642056, 'reward_near_duplicate_penalty_mean': 0.9424534440040588, 'reward_near_duplicate_penalty_std': 0.030256416648626328, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.8841394186019897, 'reward_distinct_2_std': 0.09345018118619919, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.1660452038049698, 'reward_total_composite_std': 0.13699889183044434, 'epoch': 0.06}

 37%|███▋      | 1212/3300 [4:39:06<7:22:50, 12.73s/it]INFO 04-12 20:37:32 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1213/3300 [4:39:15<6:43:01, 11.59s/it]2026-04-12 20:37:41,359 | INFO | train_grpo_train | metrics_logged mode=train step=1213

                                                       
{'loss': 0.1538, 'grad_norm': 21.867694854736328, 'learning_rate': 6.327272727272727e-06, 'num_tokens': 2159802.0, 'completions/mean_length': 20.875, 'completions/min_length': 17.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8007805943489075, 'rewards/meter/std': 0.3548671007156372, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.25088417530059814, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4320274293422699, 'rewards/total_composite/std': 0.31167376041412354, 'reward': 0.4320274293422699, 'reward_std': 0.31167376041412354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15956537425518036, 'sampling/sampling_logp_difference/max': 1.5569047927856445, 'sampling/importance_sampling_ratio/min': 0.21078747510910034, 'sampling/importance_sampling_ratio/mean': 1.0307022333145142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9438512027263641, 'clip_ratio/low_mean': 0.07086055586114526, 'clip_ratio/low_min': 0.07086055586114526, 'clip_ratio/high_mean': 0.040204678662121296, 'clip_ratio/high_max': 0.040204678662121296, 'clip_ratio/region_mean': 0.11106523452326655, 'reward_total_mean': 0.4320274293422699, 'reward_meter_mean': 0.8007805943489075, 'reward_meter_std': 0.3548671007156372, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.25088417530059814, 'reward_total_composite_mean': 0.4320274293422699, 'reward_total_composite_std': 0.31167376041412354, 'epoch': 0.06}

 37%|███▋      | 1213/3300 [4:39:15<6:43:01, 11.59s/it]INFO 04-12 20:37:41 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1214/3300 [4:39:23<6:09:40, 10.63s/it]2026-04-12 20:37:49,781 | INFO | train_grpo_train | metrics_logged mode=train step=1214

                                                       
{'loss': -0.001, 'grad_norm': 21.382360458374023, 'learning_rate': 6.324242424242425e-06, 'num_tokens': 2161232.0, 'completions/mean_length': 19.75, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.020442329347133636, 'rewards/meter/std': 0.017357273027300835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.35860592126846313, 'rewards/repeat_penalty/mean': 0.6933351755142212, 'rewards/repeat_penalty/std': 0.06524933129549026, 'rewards/repeat_floor/mean': 0.9736670255661011, 'rewards/repeat_floor/std': 0.013049869798123837, 'rewards/near_duplicate_penalty/mean': 0.9244468212127686, 'rewards/near_duplicate_penalty/std': 0.08699910342693329, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.01031399704515934, 'rewards/total_composite/std': 0.014790808781981468, 'reward': 0.01031399704515934, 'reward_std': 0.014790808781981468, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12007023394107819, 'sampling/sampling_logp_difference/max': 1.2201991081237793, 'sampling/importance_sampling_ratio/min': 0.2951713800430298, 'sampling/importance_sampling_ratio/mean': 0.9926290512084961, 'sampling/importance_sampling_ratio/max': 1.4620743989944458, 'entropy': 0.9448820129036903, 'clip_ratio/low_mean': 0.054750101175159216, 'clip_ratio/low_min': 0.054750101175159216, 'clip_ratio/high_mean': 0.014705882407724857, 'clip_ratio/high_max': 0.014705882407724857, 'clip_ratio/region_mean': 0.06945598358288407, 'reward_total_mean': 0.01031399704515934, 'reward_meter_mean': 0.020442329347133636, 'reward_meter_std': 0.017357273027300835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.6933351755142212, 'reward_repeat_penalty_std': 0.06524933129549026, 'reward_repeat_floor_mean': 0.9736670255661011, 'reward_repeat_floor_std': 0.013049869798123837, 'reward_near_duplicate_penalty_mean': 0.9244468212127686, 'reward_near_duplicate_penalty_std': 0.08699910342693329, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.35860592126846313, 'reward_total_composite_mean': 0.01031399704515934, 'reward_total_composite_std': 0.014790808781981468, 'epoch': 0.06}

 37%|███▋      | 1214/3300 [4:39:23<6:09:40, 10.63s/it]INFO 04-12 20:37:50 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1215/3300 [4:39:37<6:42:57, 11.60s/it]2026-04-12 20:38:03,600 | INFO | train_grpo_train | metrics_logged mode=train step=1215

                                                       
{'loss': -0.0453, 'grad_norm': 3.159048557281494, 'learning_rate': 6.3212121212121216e-06, 'num_tokens': 2163032.0, 'completions/mean_length': 290.0, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6043152213096619, 'rewards/meter/std': 0.3914482891559601, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.05669468268752098, 'rewards/lexical_plausibility/mean': 0.8393545150756836, 'rewards/lexical_plausibility/std': 0.17289237678050995, 'rewards/judge_quality/mean': 0.17499999701976776, 'rewards/judge_quality/std': 0.14880476891994476, 'rewards/repeat_penalty/mean': 0.9636474847793579, 'rewards/repeat_penalty/std': 0.03607404604554176, 'rewards/repeat_floor/mean': 0.9955552816390991, 'rewards/repeat_floor/std': 0.00433861929923296, 'rewards/near_duplicate_penalty/mean': 0.981377363204956, 'rewards/near_duplicate_penalty/std': 0.01754741743206978, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9816523194313049, 'rewards/distinct_2/std': 0.020408710464835167, 'rewards/total_composite/mean': 0.07501804828643799, 'rewards/total_composite/std': 0.09521345794200897, 'reward': 0.07501804828643799, 'reward_std': 0.09521345049142838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17560173571109772, 'sampling/sampling_logp_difference/max': 1.2245416641235352, 'sampling/importance_sampling_ratio/min': 0.32175910472869873, 'sampling/importance_sampling_ratio/mean': 1.0271246433258057, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.757672131061554, 'clip_ratio/low_mean': 0.019275929778814316, 'clip_ratio/low_min': 0.019275929778814316, 'clip_ratio/high_mean': 0.056129809468984604, 'clip_ratio/high_max': 0.056129809468984604, 'clip_ratio/region_mean': 0.07540573924779892, 'reward_total_mean': 0.07501804828643799, 'reward_meter_mean': 0.6043152213096619, 'reward_meter_std': 0.3914482891559601, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.18898223340511322, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.05669468268752098, 'reward_lexical_plausibility_mean': 0.8393545150756836, 'reward_lexical_plausibility_std': 0.17289237678050995, 'reward_repeat_penalty_mean': 0.9636474847793579, 'reward_repeat_penalty_std': 0.03607404604554176, 'reward_repeat_floor_mean': 0.9955552816390991, 'reward_repeat_floor_std': 0.00433861929923296, 'reward_near_duplicate_penalty_mean': 0.981377363204956, 'reward_near_duplicate_penalty_std': 0.01754741743206978, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9816523194313049, 'reward_distinct_2_std': 0.020408710464835167, 'reward_judge_quality_mean': 0.17499999701976776, 'reward_judge_quality_std': 0.14880476891994476, 'reward_total_composite_mean': 0.07501804828643799, 'reward_total_composite_std': 0.09521345794200897, 'epoch': 0.06}

 37%|███▋      | 1215/3300 [4:39:37<6:42:57, 11.60s/it]INFO 04-12 20:38:03 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1216/3300 [4:39:45<6:06:04, 10.54s/it]2026-04-12 20:38:11,680 | INFO | train_grpo_train | metrics_logged mode=train step=1216

                                                       
{'loss': 0.0119, 'grad_norm': 13.02386474609375, 'learning_rate': 6.318181818181819e-06, 'num_tokens': 2164507.0, 'completions/mean_length': 40.375, 'completions/min_length': 36.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5762792825698853, 'rewards/meter/std': 0.39428186416625977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.2507951855659485, 'rewards/repeat_penalty/mean': 0.6602959036827087, 'rewards/repeat_penalty/std': 0.26908808946609497, 'rewards/repeat_floor/mean': 0.9588524103164673, 'rewards/repeat_floor/std': 0.03361457958817482, 'rewards/near_duplicate_penalty/mean': 0.8791007995605469, 'rewards/near_duplicate_penalty/std': 0.08795655518770218, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.8068033456802368, 'rewards/distinct_2/std': 0.15945789217948914, 'rewards/total_composite/mean': 0.19602055847644806, 'rewards/total_composite/std': 0.1265619546175003, 'reward': 0.19602055847644806, 'reward_std': 0.1265619397163391, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13155457377433777, 'sampling/sampling_logp_difference/max': 1.6768778562545776, 'sampling/importance_sampling_ratio/min': 0.1869567632675171, 'sampling/importance_sampling_ratio/mean': 1.0167096853256226, 'sampling/importance_sampling_ratio/max': 1.6693665981292725, 'entropy': 0.9068821221590042, 'clip_ratio/low_mean': 0.061954366974532604, 'clip_ratio/low_min': 0.061954366974532604, 'clip_ratio/high_mean': 0.07152792531996965, 'clip_ratio/high_max': 0.07152792531996965, 'clip_ratio/region_mean': 0.13348229229450226, 'reward_total_mean': 0.19602055847644806, 'reward_meter_mean': 0.5762792825698853, 'reward_meter_std': 0.39428186416625977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6602959036827087, 'reward_repeat_penalty_std': 0.26908808946609497, 'reward_repeat_floor_mean': 0.9588524103164673, 'reward_repeat_floor_std': 0.03361457958817482, 'reward_near_duplicate_penalty_mean': 0.8791007995605469, 'reward_near_duplicate_penalty_std': 0.08795655518770218, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.8068033456802368, 'reward_distinct_2_std': 0.15945789217948914, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.2507951855659485, 'reward_total_composite_mean': 0.19602055847644806, 'reward_total_composite_std': 0.1265619546175003, 'epoch': 0.06}

 37%|███▋      | 1216/3300 [4:39:45<6:06:04, 10.54s/it]INFO 04-12 20:38:11 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1217/3300 [4:40:00<6:50:56, 11.84s/it]2026-04-12 20:38:26,558 | INFO | train_grpo_train | metrics_logged mode=train step=1217

                                                       
{'loss': -0.0791, 'grad_norm': 5.531728267669678, 'learning_rate': 6.315151515151515e-06, 'num_tokens': 2166367.0, 'completions/mean_length': 115.5, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.857147216796875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5791007280349731, 'rewards/meter/std': 0.35110318660736084, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9581830501556396, 'rewards/lexical_plausibility/std': 0.11827609688043594, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8939722776412964, 'rewards/repeat_penalty/std': 0.06201153248548508, 'rewards/repeat_floor/mean': 0.9865633249282837, 'rewards/repeat_floor/std': 0.007288097403943539, 'rewards/near_duplicate_penalty/mean': 0.9402076005935669, 'rewards/near_duplicate_penalty/std': 0.030698206275701523, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9503576755523682, 'rewards/distinct_2/std': 0.04762725532054901, 'rewards/total_composite/mean': 0.20109564065933228, 'rewards/total_composite/std': 0.13192442059516907, 'reward': 0.20109564065933228, 'reward_std': 0.13192442059516907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14121000468730927, 'sampling/sampling_logp_difference/max': 1.369049072265625, 'sampling/importance_sampling_ratio/min': 0.26712557673454285, 'sampling/importance_sampling_ratio/mean': 1.0121617317199707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7656353935599327, 'clip_ratio/low_mean': 0.04952818900346756, 'clip_ratio/low_min': 0.04952818900346756, 'clip_ratio/high_mean': 0.054745656438171864, 'clip_ratio/high_max': 0.054745656438171864, 'clip_ratio/region_mean': 0.10427384544163942, 'reward_total_mean': 0.20109564065933228, 'reward_meter_mean': 0.5791007280349731, 'reward_meter_std': 0.35110318660736084, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9581830501556396, 'reward_lexical_plausibility_std': 0.11827609688043594, 'reward_repeat_penalty_mean': 0.8939722776412964, 'reward_repeat_penalty_std': 0.06201153248548508, 'reward_repeat_floor_mean': 0.9865633249282837, 'reward_repeat_floor_std': 0.007288097403943539, 'reward_near_duplicate_penalty_mean': 0.9402076005935669, 'reward_near_duplicate_penalty_std': 0.030698206275701523, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9503576755523682, 'reward_distinct_2_std': 0.04762725532054901, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.20109564065933228, 'reward_total_composite_std': 0.13192442059516907, 'epoch': 0.06}

 37%|███▋      | 1217/3300 [4:40:00<6:50:56, 11.84s/it]INFO 04-12 20:38:26 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1218/3300 [4:40:14<7:12:09, 12.45s/it]2026-04-12 20:38:40,450 | INFO | train_grpo_train | metrics_logged mode=train step=1218

                                                       
{'loss': -0.037, 'grad_norm': 2.199390172958374, 'learning_rate': 6.3121212121212125e-06, 'num_tokens': 2168057.0, 'completions/mean_length': 225.25, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 53.20000076293945, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.2198893427848816, 'rewards/meter/std': 0.22230005264282227, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8457233905792236, 'rewards/lexical_plausibility/std': 0.21446913480758667, 'rewards/judge_quality/mean': 0.19374999403953552, 'rewards/judge_quality/std': 0.16783814132213593, 'rewards/repeat_penalty/mean': 0.9438612461090088, 'rewards/repeat_penalty/std': 0.065047986805439, 'rewards/repeat_floor/mean': 0.9930291175842285, 'rewards/repeat_floor/std': 0.007196676451712847, 'rewards/near_duplicate_penalty/mean': 0.9699330925941467, 'rewards/near_duplicate_penalty/std': 0.025620054453611374, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9726577997207642, 'rewards/distinct_2/std': 0.05350784212350845, 'rewards/total_composite/mean': 0.05306670069694519, 'rewards/total_composite/std': 0.08914055675268173, 'reward': 0.05306670069694519, 'reward_std': 0.08914055675268173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16026510298252106, 'sampling/sampling_logp_difference/max': 1.733330488204956, 'sampling/importance_sampling_ratio/min': 0.22949977219104767, 'sampling/importance_sampling_ratio/mean': 1.0221580266952515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7499352991580963, 'clip_ratio/low_mean': 0.05001780670136213, 'clip_ratio/low_min': 0.05001780670136213, 'clip_ratio/high_mean': 0.03986378200352192, 'clip_ratio/high_max': 0.03986378200352192, 'clip_ratio/region_mean': 0.08988158870488405, 'reward_total_mean': 0.05306670069694519, 'reward_meter_mean': 0.2198893427848816, 'reward_meter_std': 0.22230005264282227, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8457233905792236, 'reward_lexical_plausibility_std': 0.21446913480758667, 'reward_repeat_penalty_mean': 0.9438612461090088, 'reward_repeat_penalty_std': 0.065047986805439, 'reward_repeat_floor_mean': 0.9930291175842285, 'reward_repeat_floor_std': 0.007196676451712847, 'reward_near_duplicate_penalty_mean': 0.9699330925941467, 'reward_near_duplicate_penalty_std': 0.025620054453611374, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9726577997207642, 'reward_distinct_2_std': 0.05350784212350845, 'reward_judge_quality_mean': 0.19374999403953552, 'reward_judge_quality_std': 0.16783814132213593, 'reward_total_composite_mean': 0.05306670069694519, 'reward_total_composite_std': 0.08914055675268173, 'epoch': 0.06}

 37%|███▋      | 1218/3300 [4:40:14<7:12:09, 12.45s/it]INFO 04-12 20:38:40 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1219/3300 [4:40:24<6:46:44, 11.73s/it]2026-04-12 20:38:50,472 | INFO | train_grpo_train | metrics_logged mode=train step=1219

                                                       
{'loss': 0.0296, 'grad_norm': 8.804862976074219, 'learning_rate': 6.309090909090909e-06, 'num_tokens': 2170582.0, 'completions/mean_length': 113.625, 'completions/min_length': 102.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.625, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.44080227613449097, 'rewards/meter/std': 0.24647030234336853, 'rewards/count_adherence/mean': 0.6136363744735718, 'rewards/count_adherence/std': 0.04208271950483322, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8840909004211426, 'rewards/count_floor/std': 0.01262480765581131, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.6801637411117554, 'rewards/repeat_penalty/std': 0.11364476382732391, 'rewards/repeat_floor/mean': 0.9634428024291992, 'rewards/repeat_floor/std': 0.012377992272377014, 'rewards/near_duplicate_penalty/mean': 0.900574803352356, 'rewards/near_duplicate_penalty/std': 0.02994886040687561, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.7595179677009583, 'rewards/distinct_2/std': 0.09867044538259506, 'rewards/total_composite/mean': 0.1381712555885315, 'rewards/total_composite/std': 0.08209511637687683, 'reward': 0.1381712555885315, 'reward_std': 0.08209511637687683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12140367925167084, 'sampling/sampling_logp_difference/max': 3.9601335525512695, 'sampling/importance_sampling_ratio/min': 0.019060568884015083, 'sampling/importance_sampling_ratio/mean': 1.017128348350525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.70613794028759, 'clip_ratio/low_mean': 0.07515633245930076, 'clip_ratio/low_min': 0.07515633245930076, 'clip_ratio/high_mean': 0.027110234834253788, 'clip_ratio/high_max': 0.027110234834253788, 'clip_ratio/region_mean': 0.10226656729355454, 'reward_total_mean': 0.1381712555885315, 'reward_meter_mean': 0.44080227613449097, 'reward_meter_std': 0.24647030234336853, 'reward_count_adherence_mean': 0.6136363744735718, 'reward_count_adherence_std': 0.04208271950483322, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8840909004211426, 'reward_count_floor_std': 0.01262480765581131, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6801637411117554, 'reward_repeat_penalty_std': 0.11364476382732391, 'reward_repeat_floor_mean': 0.9634428024291992, 'reward_repeat_floor_std': 0.012377992272377014, 'reward_near_duplicate_penalty_mean': 0.900574803352356, 'reward_near_duplicate_penalty_std': 0.02994886040687561, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.7595179677009583, 'reward_distinct_2_std': 0.09867044538259506, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.1381712555885315, 'reward_total_composite_std': 0.08209511637687683, 'epoch': 0.06}

 37%|███▋      | 1219/3300 [4:40:24<6:46:44, 11.73s/it]INFO 04-12 20:38:50 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1220/3300 [4:40:33<6:20:18, 10.97s/it]2026-04-12 20:38:59,684 | INFO | train_grpo_train | metrics_logged mode=train step=1220

                                                       
{'loss': -0.0296, 'grad_norm': 9.735910415649414, 'learning_rate': 6.306060606060607e-06, 'num_tokens': 2172902.0, 'completions/mean_length': 95.0, 'completions/min_length': 77.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.7763108015060425, 'rewards/meter/std': 0.20298224687576294, 'rewards/count_adherence/mean': 0.5454545617103577, 'rewards/count_adherence/std': 0.06872080266475677, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8636363744735718, 'rewards/count_floor/std': 0.02061624825000763, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.4445405602455139, 'rewards/repeat_penalty/std': 0.21346114575862885, 'rewards/repeat_floor/mean': 0.9284533262252808, 'rewards/repeat_floor/std': 0.038077108561992645, 'rewards/near_duplicate_penalty/mean': 0.8144187331199646, 'rewards/near_duplicate_penalty/std': 0.10507257282733917, 'rewards/opening_diversity/mean': 0.7781655788421631, 'rewards/opening_diversity/std': 0.18842655420303345, 'rewards/distinct_2/mean': 0.6622289419174194, 'rewards/distinct_2/std': 0.14104324579238892, 'rewards/total_composite/mean': 0.22396959364414215, 'rewards/total_composite/std': 0.053816284984350204, 'reward': 0.22396959364414215, 'reward_std': 0.053816284984350204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13207414746284485, 'sampling/sampling_logp_difference/max': 2.4983208179473877, 'sampling/importance_sampling_ratio/min': 0.11782732605934143, 'sampling/importance_sampling_ratio/mean': 1.0074642896652222, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5844785496592522, 'clip_ratio/low_mean': 0.04841542150825262, 'clip_ratio/low_min': 0.04841542150825262, 'clip_ratio/high_mean': 0.07083963975310326, 'clip_ratio/high_max': 0.07083963975310326, 'clip_ratio/region_mean': 0.11925506126135588, 'reward_total_mean': 0.22396959364414215, 'reward_meter_mean': 0.7763108015060425, 'reward_meter_std': 0.20298224687576294, 'reward_count_adherence_mean': 0.5454545617103577, 'reward_count_adherence_std': 0.06872080266475677, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8636363744735718, 'reward_count_floor_std': 0.02061624825000763, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.4445405602455139, 'reward_repeat_penalty_std': 0.21346114575862885, 'reward_repeat_floor_mean': 0.9284533262252808, 'reward_repeat_floor_std': 0.038077108561992645, 'reward_near_duplicate_penalty_mean': 0.8144187331199646, 'reward_near_duplicate_penalty_std': 0.10507257282733917, 'reward_opening_diversity_mean': 0.7781655788421631, 'reward_opening_diversity_std': 0.18842655420303345, 'reward_distinct_2_mean': 0.6622289419174194, 'reward_distinct_2_std': 0.14104324579238892, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.22396959364414215, 'reward_total_composite_std': 0.053816284984350204, 'epoch': 0.06}

 37%|███▋      | 1220/3300 [4:40:33<6:20:18, 10.97s/it]INFO 04-12 20:38:59 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1221/3300 [4:40:41<5:55:01, 10.25s/it]2026-04-12 20:39:08,231 | INFO | train_grpo_train | metrics_logged mode=train step=1221

                                                       
{'loss': 0.0349, 'grad_norm': 14.995492935180664, 'learning_rate': 6.303030303030303e-06, 'num_tokens': 2174459.0, 'completions/mean_length': 33.625, 'completions/min_length': 30.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.37095850706100464, 'rewards/meter/std': 0.29680106043815613, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9696150422096252, 'rewards/lexical_plausibility/std': 0.0859416201710701, 'rewards/judge_quality/mean': 0.45374998450279236, 'rewards/judge_quality/std': 0.20170257985591888, 'rewards/repeat_penalty/mean': 0.8760749101638794, 'rewards/repeat_penalty/std': 0.17035168409347534, 'rewards/repeat_floor/mean': 0.9844744205474854, 'rewards/repeat_floor/std': 0.019319143146276474, 'rewards/near_duplicate_penalty/mean': 0.9355652332305908, 'rewards/near_duplicate_penalty/std': 0.05685482174158096, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9557186365127563, 'rewards/distinct_2/std': 0.08239380270242691, 'rewards/total_composite/mean': 0.1714765429496765, 'rewards/total_composite/std': 0.1576114445924759, 'reward': 0.1714765429496765, 'reward_std': 0.1576114445924759, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14743992686271667, 'sampling/sampling_logp_difference/max': 1.28440523147583, 'sampling/importance_sampling_ratio/min': 0.27681517601013184, 'sampling/importance_sampling_ratio/mean': 1.021804928779602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0102483034133911, 'clip_ratio/low_mean': 0.059101749677211046, 'clip_ratio/low_min': 0.059101749677211046, 'clip_ratio/high_mean': 0.06423611240461469, 'clip_ratio/high_max': 0.06423611240461469, 'clip_ratio/region_mean': 0.12333786208182573, 'reward_total_mean': 0.1714765429496765, 'reward_meter_mean': 0.37095850706100464, 'reward_meter_std': 0.29680106043815613, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9696150422096252, 'reward_lexical_plausibility_std': 0.0859416201710701, 'reward_repeat_penalty_mean': 0.8760749101638794, 'reward_repeat_penalty_std': 0.17035168409347534, 'reward_repeat_floor_mean': 0.9844744205474854, 'reward_repeat_floor_std': 0.019319143146276474, 'reward_near_duplicate_penalty_mean': 0.9355652332305908, 'reward_near_duplicate_penalty_std': 0.05685482174158096, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9557186365127563, 'reward_distinct_2_std': 0.08239380270242691, 'reward_judge_quality_mean': 0.45374998450279236, 'reward_judge_quality_std': 0.20170257985591888, 'reward_total_composite_mean': 0.1714765429496765, 'reward_total_composite_std': 0.1576114445924759, 'epoch': 0.06}

 37%|███▋      | 1221/3300 [4:40:42<5:55:01, 10.25s/it]INFO 04-12 20:39:08 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1222/3300 [4:40:56<6:41:18, 11.59s/it]2026-04-12 20:39:22,954 | INFO | train_grpo_train | metrics_logged mode=train step=1222

                                                       
{'loss': -0.1187, 'grad_norm': 4.0689239501953125, 'learning_rate': 6.300000000000001e-06, 'num_tokens': 2176344.0, 'completions/mean_length': 116.625, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.142860412597656, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7610733509063721, 'rewards/meter/std': 0.3392793536186218, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9665437340736389, 'rewards/lexical_plausibility/std': 0.09462856501340866, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.7470518350601196, 'rewards/repeat_penalty/std': 0.15076427161693573, 'rewards/repeat_floor/mean': 0.9713449478149414, 'rewards/repeat_floor/std': 0.018709450960159302, 'rewards/near_duplicate_penalty/mean': 0.912311851978302, 'rewards/near_duplicate_penalty/std': 0.053679209202528, 'rewards/opening_diversity/mean': 0.9791666269302368, 'rewards/opening_diversity/std': 0.0589255727827549, 'rewards/distinct_2/mean': 0.8416461944580078, 'rewards/distinct_2/std': 0.09909360110759735, 'rewards/total_composite/mean': 0.23799125850200653, 'rewards/total_composite/std': 0.120130255818367, 'reward': 0.23799125850200653, 'reward_std': 0.1201302632689476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13118958473205566, 'sampling/sampling_logp_difference/max': 1.665986180305481, 'sampling/importance_sampling_ratio/min': 0.18900416791439056, 'sampling/importance_sampling_ratio/mean': 0.9999014735221863, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6908783726394176, 'clip_ratio/low_mean': 0.012711863964796066, 'clip_ratio/low_min': 0.012711863964796066, 'clip_ratio/high_mean': 0.10025918297469616, 'clip_ratio/high_max': 0.10025918297469616, 'clip_ratio/region_mean': 0.11297104693949223, 'reward_total_mean': 0.23799125850200653, 'reward_meter_mean': 0.7610733509063721, 'reward_meter_std': 0.3392793536186218, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9665437340736389, 'reward_lexical_plausibility_std': 0.09462856501340866, 'reward_repeat_penalty_mean': 0.7470518350601196, 'reward_repeat_penalty_std': 0.15076427161693573, 'reward_repeat_floor_mean': 0.9713449478149414, 'reward_repeat_floor_std': 0.018709450960159302, 'reward_near_duplicate_penalty_mean': 0.912311851978302, 'reward_near_duplicate_penalty_std': 0.053679209202528, 'reward_opening_diversity_mean': 0.9791666269302368, 'reward_opening_diversity_std': 0.0589255727827549, 'reward_distinct_2_mean': 0.8416461944580078, 'reward_distinct_2_std': 0.09909360110759735, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.23799125850200653, 'reward_total_composite_std': 0.120130255818367, 'epoch': 0.06}

 37%|███▋      | 1222/3300 [4:40:56<6:41:18, 11.59s/it]INFO 04-12 20:39:23 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1223/3300 [4:41:11<7:10:57, 12.45s/it]2026-04-12 20:39:37,428 | INFO | train_grpo_train | metrics_logged mode=train step=1223

                                                       
{'loss': -0.0311, 'grad_norm': 4.6294050216674805, 'learning_rate': 6.296969696969697e-06, 'num_tokens': 2177848.0, 'completions/mean_length': 92.0, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6059252023696899, 'rewards/meter/std': 0.3446142077445984, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9573439955711365, 'rewards/lexical_plausibility/std': 0.12064942717552185, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.23627692461013794, 'rewards/repeat_penalty/mean': 0.9202711582183838, 'rewards/repeat_penalty/std': 0.13714632391929626, 'rewards/repeat_floor/mean': 0.9909429550170898, 'rewards/repeat_floor/std': 0.014037278480827808, 'rewards/near_duplicate_penalty/mean': 0.9653171300888062, 'rewards/near_duplicate_penalty/std': 0.04035671800374985, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.978003978729248, 'rewards/distinct_2/std': 0.04075079411268234, 'rewards/total_composite/mean': 0.28693416714668274, 'rewards/total_composite/std': 0.24088135361671448, 'reward': 0.28693416714668274, 'reward_std': 0.24088133871555328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1436370313167572, 'sampling/sampling_logp_difference/max': 1.0054512023925781, 'sampling/importance_sampling_ratio/min': 0.4024932086467743, 'sampling/importance_sampling_ratio/mean': 1.0243269205093384, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8930256366729736, 'clip_ratio/low_mean': 0.043560607358813286, 'clip_ratio/low_min': 0.043560607358813286, 'clip_ratio/high_mean': 0.055880406871438026, 'clip_ratio/high_max': 0.055880406871438026, 'clip_ratio/region_mean': 0.09944101423025131, 'reward_total_mean': 0.28693416714668274, 'reward_meter_mean': 0.6059252023696899, 'reward_meter_std': 0.3446142077445984, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9573439955711365, 'reward_lexical_plausibility_std': 0.12064942717552185, 'reward_repeat_penalty_mean': 0.9202711582183838, 'reward_repeat_penalty_std': 0.13714632391929626, 'reward_repeat_floor_mean': 0.9909429550170898, 'reward_repeat_floor_std': 0.014037278480827808, 'reward_near_duplicate_penalty_mean': 0.9653171300888062, 'reward_near_duplicate_penalty_std': 0.04035671800374985, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.978003978729248, 'reward_distinct_2_std': 0.04075079411268234, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.23627692461013794, 'reward_total_composite_mean': 0.28693416714668274, 'reward_total_composite_std': 0.24088135361671448, 'epoch': 0.06}

 37%|███▋      | 1223/3300 [4:41:11<7:10:57, 12.45s/it]INFO 04-12 20:39:37 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1224/3300 [4:41:25<7:26:04, 12.89s/it]2026-04-12 20:39:51,342 | INFO | train_grpo_train | metrics_logged mode=train step=1224

                                                       
{'loss': -0.0283, 'grad_norm': 5.391702651977539, 'learning_rate': 6.293939393939394e-06, 'num_tokens': 2179387.0, 'completions/mean_length': 93.375, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.57143020629883, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.4266401529312134, 'rewards/meter/std': 0.3757961392402649, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9573965668678284, 'rewards/lexical_plausibility/std': 0.12050070613622665, 'rewards/judge_quality/mean': 0.45000001788139343, 'rewards/judge_quality/std': 0.2828426957130432, 'rewards/repeat_penalty/mean': 0.9389135837554932, 'rewards/repeat_penalty/std': 0.08198247849941254, 'rewards/repeat_floor/mean': 0.9936495423316956, 'rewards/repeat_floor/std': 0.005686817225068808, 'rewards/near_duplicate_penalty/mean': 0.9701635837554932, 'rewards/near_duplicate_penalty/std': 0.03224588558077812, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18385660648345947, 'rewards/total_composite/std': 0.1858883798122406, 'reward': 0.18385660648345947, 'reward_std': 0.1858883798122406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17873063683509827, 'sampling/sampling_logp_difference/max': 1.8777542114257812, 'sampling/importance_sampling_ratio/min': 0.15293318033218384, 'sampling/importance_sampling_ratio/mean': 1.0017623901367188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1278115212917328, 'clip_ratio/low_mean': 0.07725170347839594, 'clip_ratio/low_min': 0.07725170347839594, 'clip_ratio/high_mean': 0.07338709570467472, 'clip_ratio/high_max': 0.07338709570467472, 'clip_ratio/region_mean': 0.15063879918307066, 'reward_total_mean': 0.18385660648345947, 'reward_meter_mean': 0.4266401529312134, 'reward_meter_std': 0.3757961392402649, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9573965668678284, 'reward_lexical_plausibility_std': 0.12050070613622665, 'reward_repeat_penalty_mean': 0.9389135837554932, 'reward_repeat_penalty_std': 0.08198247849941254, 'reward_repeat_floor_mean': 0.9936495423316956, 'reward_repeat_floor_std': 0.005686817225068808, 'reward_near_duplicate_penalty_mean': 0.9701635837554932, 'reward_near_duplicate_penalty_std': 0.03224588558077812, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45000001788139343, 'reward_judge_quality_std': 0.2828426957130432, 'reward_total_composite_mean': 0.18385660648345947, 'reward_total_composite_std': 0.1858883798122406, 'epoch': 0.06}

 37%|███▋      | 1224/3300 [4:41:25<7:26:04, 12.89s/it]INFO 04-12 20:39:51 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1225/3300 [4:41:33<6:43:44, 11.67s/it]2026-04-12 20:40:00,182 | INFO | train_grpo_train | metrics_logged mode=train step=1225

                                                       
{'loss': 0.0078, 'grad_norm': 14.657776832580566, 'learning_rate': 6.290909090909092e-06, 'num_tokens': 2181043.0, 'completions/mean_length': 44.0, 'completions/min_length': 29.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8110363483428955, 'rewards/meter/std': 0.32047566771507263, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.9276831150054932, 'rewards/repeat_penalty/std': 0.06526709347963333, 'rewards/repeat_floor/mean': 0.9911633729934692, 'rewards/repeat_floor/std': 0.00710683036595583, 'rewards/near_duplicate_penalty/mean': 0.9638071060180664, 'rewards/near_duplicate_penalty/std': 0.02484997548162937, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9621370434761047, 'rewards/distinct_2/std': 0.05582091957330704, 'rewards/total_composite/mean': 0.23005390167236328, 'rewards/total_composite/std': 0.14171624183654785, 'reward': 0.23005390167236328, 'reward_std': 0.14171624183654785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15819786489009857, 'sampling/sampling_logp_difference/max': 1.872924566268921, 'sampling/importance_sampling_ratio/min': 0.15367357432842255, 'sampling/importance_sampling_ratio/mean': 0.9857171177864075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8455542996525764, 'clip_ratio/low_mean': 0.0755122834816575, 'clip_ratio/low_min': 0.0755122834816575, 'clip_ratio/high_mean': 0.06759650958701968, 'clip_ratio/high_max': 0.06759650958701968, 'clip_ratio/region_mean': 0.1431087930686772, 'reward_total_mean': 0.23005390167236328, 'reward_meter_mean': 0.8110363483428955, 'reward_meter_std': 0.32047566771507263, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9276831150054932, 'reward_repeat_penalty_std': 0.06526709347963333, 'reward_repeat_floor_mean': 0.9911633729934692, 'reward_repeat_floor_std': 0.00710683036595583, 'reward_near_duplicate_penalty_mean': 0.9638071060180664, 'reward_near_duplicate_penalty_std': 0.02484997548162937, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9621370434761047, 'reward_distinct_2_std': 0.05582091957330704, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.23005390167236328, 'reward_total_composite_std': 0.14171624183654785, 'epoch': 0.06}

 37%|███▋      | 1225/3300 [4:41:34<6:43:44, 11.67s/it]INFO 04-12 20:40:00 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1226/3300 [4:41:42<6:07:32, 10.63s/it]2026-04-12 20:40:08,378 | INFO | train_grpo_train | metrics_logged mode=train step=1226

                                                       
{'loss': 0.0385, 'grad_norm': 13.133723258972168, 'learning_rate': 6.287878787878788e-06, 'num_tokens': 2182718.0, 'completions/mean_length': 29.375, 'completions/min_length': 26.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.7395759224891663, 'rewards/meter/std': 0.28121909499168396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2639264166355133, 'rewards/repeat_penalty/mean': 0.9634692668914795, 'rewards/repeat_penalty/std': 0.049862924963235855, 'rewards/repeat_floor/mean': 0.994520366191864, 'rewards/repeat_floor/std': 0.007479443214833736, 'rewards/near_duplicate_penalty/mean': 0.9634692668914795, 'rewards/near_duplicate_penalty/std': 0.049862924963235855, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3715677559375763, 'rewards/total_composite/std': 0.24364785850048065, 'reward': 0.3715677559375763, 'reward_std': 0.24364784359931946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1478663980960846, 'sampling/sampling_logp_difference/max': 1.7367868423461914, 'sampling/importance_sampling_ratio/min': 0.17608527839183807, 'sampling/importance_sampling_ratio/mean': 0.9975112676620483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8369600102305412, 'clip_ratio/low_mean': 0.05892255902290344, 'clip_ratio/low_min': 0.05892255902290344, 'clip_ratio/high_mean': 0.09431419614702463, 'clip_ratio/high_max': 0.09431419614702463, 'clip_ratio/region_mean': 0.15323675516992807, 'reward_total_mean': 0.3715677559375763, 'reward_meter_mean': 0.7395759224891663, 'reward_meter_std': 0.28121909499168396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9634692668914795, 'reward_repeat_penalty_std': 0.049862924963235855, 'reward_repeat_floor_mean': 0.994520366191864, 'reward_repeat_floor_std': 0.007479443214833736, 'reward_near_duplicate_penalty_mean': 0.9634692668914795, 'reward_near_duplicate_penalty_std': 0.049862924963235855, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.2639264166355133, 'reward_total_composite_mean': 0.3715677559375763, 'reward_total_composite_std': 0.24364785850048065, 'epoch': 0.06}

 37%|███▋      | 1226/3300 [4:41:42<6:07:32, 10.63s/it]INFO 04-12 20:40:08 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1227/3300 [4:41:50<5:43:49,  9.95s/it]2026-04-12 20:40:16,734 | INFO | train_grpo_train | metrics_logged mode=train step=1227

                                                       
{'loss': -0.024, 'grad_norm': 19.35190773010254, 'learning_rate': 6.284848484848486e-06, 'num_tokens': 2184057.0, 'completions/mean_length': 31.375, 'completions/min_length': 28.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5555839538574219, 'rewards/meter/std': 0.3538241684436798, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.8659398555755615, 'rewards/repeat_penalty/std': 0.14649713039398193, 'rewards/repeat_floor/mean': 0.9835029244422913, 'rewards/repeat_floor/std': 0.017463911324739456, 'rewards/near_duplicate_penalty/mean': 0.9395708441734314, 'rewards/near_duplicate_penalty/std': 0.048510950058698654, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9174147844314575, 'rewards/distinct_2/std': 0.1265544444322586, 'rewards/total_composite/mean': 0.1759277582168579, 'rewards/total_composite/std': 0.1372385323047638, 'reward': 0.1759277582168579, 'reward_std': 0.1372385323047638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15816184878349304, 'sampling/sampling_logp_difference/max': 1.1287744045257568, 'sampling/importance_sampling_ratio/min': 0.3234294056892395, 'sampling/importance_sampling_ratio/mean': 1.018064022064209, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1887256801128387, 'clip_ratio/low_mean': 0.08529490092769265, 'clip_ratio/low_min': 0.08529490092769265, 'clip_ratio/high_mean': 0.06788240931928158, 'clip_ratio/high_max': 0.06788240931928158, 'clip_ratio/region_mean': 0.15317731024697423, 'reward_total_mean': 0.1759277582168579, 'reward_meter_mean': 0.5555839538574219, 'reward_meter_std': 0.3538241684436798, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8659398555755615, 'reward_repeat_penalty_std': 0.14649713039398193, 'reward_repeat_floor_mean': 0.9835029244422913, 'reward_repeat_floor_std': 0.017463911324739456, 'reward_near_duplicate_penalty_mean': 0.9395708441734314, 'reward_near_duplicate_penalty_std': 0.048510950058698654, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9174147844314575, 'reward_distinct_2_std': 0.1265544444322586, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.1759277582168579, 'reward_total_composite_std': 0.1372385323047638, 'epoch': 0.06}

 37%|███▋      | 1227/3300 [4:41:50<5:43:49,  9.95s/it]INFO 04-12 20:40:16 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1228/3300 [4:41:59<5:33:00,  9.64s/it]2026-04-12 20:40:25,684 | INFO | train_grpo_train | metrics_logged mode=train step=1228

                                                       
{'loss': 0.0674, 'grad_norm': 14.310054779052734, 'learning_rate': 6.2818181818181825e-06, 'num_tokens': 2185508.0, 'completions/mean_length': 31.375, 'completions/min_length': 27.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.26273313164711, 'rewards/meter/std': 0.3580339252948761, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.18137866258621216, 'rewards/repeat_penalty/mean': 0.984682023525238, 'rewards/repeat_penalty/std': 0.021919764578342438, 'rewards/repeat_floor/mean': 0.9977023005485535, 'rewards/repeat_floor/std': 0.003287959611043334, 'rewards/near_duplicate_penalty/mean': 0.984682023525238, 'rewards/near_duplicate_penalty/std': 0.021919764578342438, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13236719369888306, 'rewards/total_composite/std': 0.17755210399627686, 'reward': 0.13236719369888306, 'reward_std': 0.17755210399627686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14894165098667145, 'sampling/sampling_logp_difference/max': 1.5280179977416992, 'sampling/importance_sampling_ratio/min': 0.21696527302265167, 'sampling/importance_sampling_ratio/mean': 1.0312650203704834, 'sampling/importance_sampling_ratio/max': 1.9037361145019531, 'entropy': 1.311284214258194, 'clip_ratio/low_mean': 0.10541710071265697, 'clip_ratio/low_min': 0.10541710071265697, 'clip_ratio/high_mean': 0.057189542800188065, 'clip_ratio/high_max': 0.057189542800188065, 'clip_ratio/region_mean': 0.16260664351284504, 'reward_total_mean': 0.13236719369888306, 'reward_meter_mean': 0.26273313164711, 'reward_meter_std': 0.3580339252948761, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.984682023525238, 'reward_repeat_penalty_std': 0.021919764578342438, 'reward_repeat_floor_mean': 0.9977023005485535, 'reward_repeat_floor_std': 0.003287959611043334, 'reward_near_duplicate_penalty_mean': 0.984682023525238, 'reward_near_duplicate_penalty_std': 0.021919764578342438, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.18137866258621216, 'reward_total_composite_mean': 0.13236719369888306, 'reward_total_composite_std': 0.17755210399627686, 'epoch': 0.06}

 37%|███▋      | 1228/3300 [4:41:59<5:33:00,  9.64s/it]INFO 04-12 20:40:25 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1229/3300 [4:42:13<6:22:27, 11.08s/it]2026-04-12 20:40:40,095 | INFO | train_grpo_train | metrics_logged mode=train step=1229

                                                       
{'loss': -0.026, 'grad_norm': 6.57251501083374, 'learning_rate': 6.27878787878788e-06, 'num_tokens': 2186797.0, 'completions/mean_length': 80.125, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 18.428571701049805, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.35416775941848755, 'rewards/meter/std': 0.44328710436820984, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9538531303405762, 'rewards/lexical_plausibility/std': 0.13052314519882202, 'rewards/judge_quality/mean': 0.476250022649765, 'rewards/judge_quality/std': 0.38119879364967346, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24739056825637817, 'rewards/total_composite/std': 0.3496955931186676, 'reward': 0.24739056825637817, 'reward_std': 0.3496955931186676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12106632441282272, 'sampling/sampling_logp_difference/max': 1.3868143558502197, 'sampling/importance_sampling_ratio/min': 0.24987004697322845, 'sampling/importance_sampling_ratio/mean': 1.0496184825897217, 'sampling/importance_sampling_ratio/max': 1.8097660541534424, 'entropy': 0.8599603101611137, 'clip_ratio/low_mean': 0.04983660206198692, 'clip_ratio/low_min': 0.04983660206198692, 'clip_ratio/high_mean': 0.045138888992369175, 'clip_ratio/high_max': 0.045138888992369175, 'clip_ratio/region_mean': 0.0949754910543561, 'reward_total_mean': 0.24739056825637817, 'reward_meter_mean': 0.35416775941848755, 'reward_meter_std': 0.44328710436820984, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9538531303405762, 'reward_lexical_plausibility_std': 0.13052314519882202, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.476250022649765, 'reward_judge_quality_std': 0.38119879364967346, 'reward_total_composite_mean': 0.24739056825637817, 'reward_total_composite_std': 0.3496955931186676, 'epoch': 0.06}

 37%|███▋      | 1229/3300 [4:42:13<6:22:27, 11.08s/it]INFO 04-12 20:40:40 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1230/3300 [4:42:22<6:01:38, 10.48s/it]2026-04-12 20:40:49,193 | INFO | train_grpo_train | metrics_logged mode=train step=1230

                                                       
{'loss': 0.0823, 'grad_norm': 16.972959518432617, 'learning_rate': 6.275757575757576e-06, 'num_tokens': 2188420.0, 'completions/mean_length': 39.875, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8223479390144348, 'rewards/meter/std': 0.15021897852420807, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.11877348273992538, 'rewards/repeat_penalty/mean': 0.9021636843681335, 'rewards/repeat_penalty/std': 0.12987715005874634, 'rewards/repeat_floor/mean': 0.9905874133110046, 'rewards/repeat_floor/std': 0.009797558188438416, 'rewards/near_duplicate_penalty/mean': 0.965046763420105, 'rewards/near_duplicate_penalty/std': 0.025503626093268394, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9536713361740112, 'rewards/distinct_2/std': 0.06758687645196915, 'rewards/total_composite/mean': 0.25501781702041626, 'rewards/total_composite/std': 0.11399103701114655, 'reward': 0.25501781702041626, 'reward_std': 0.11399103701114655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18526025116443634, 'sampling/sampling_logp_difference/max': 1.9111475944519043, 'sampling/importance_sampling_ratio/min': 0.14791055023670197, 'sampling/importance_sampling_ratio/mean': 0.9953362345695496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2463163137435913, 'clip_ratio/low_mean': 0.05903615802526474, 'clip_ratio/low_min': 0.05903615802526474, 'clip_ratio/high_mean': 0.11196553893387318, 'clip_ratio/high_max': 0.11196553893387318, 'clip_ratio/region_mean': 0.17100169695913792, 'reward_total_mean': 0.25501781702041626, 'reward_meter_mean': 0.8223479390144348, 'reward_meter_std': 0.15021897852420807, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9021636843681335, 'reward_repeat_penalty_std': 0.12987715005874634, 'reward_repeat_floor_mean': 0.9905874133110046, 'reward_repeat_floor_std': 0.009797558188438416, 'reward_near_duplicate_penalty_mean': 0.965046763420105, 'reward_near_duplicate_penalty_std': 0.025503626093268394, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9536713361740112, 'reward_distinct_2_std': 0.06758687645196915, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.11877348273992538, 'reward_total_composite_mean': 0.25501781702041626, 'reward_total_composite_std': 0.11399103701114655, 'epoch': 0.06}

 37%|███▋      | 1230/3300 [4:42:23<6:01:38, 10.48s/it]INFO 04-12 20:40:49 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1231/3300 [4:42:31<5:46:30, 10.05s/it]2026-04-12 20:40:58,224 | INFO | train_grpo_train | metrics_logged mode=train step=1231

                                                       
{'loss': 0.095, 'grad_norm': 15.801769256591797, 'learning_rate': 6.2727272727272734e-06, 'num_tokens': 2190329.0, 'completions/mean_length': 41.625, 'completions/min_length': 38.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7712146043777466, 'rewards/meter/std': 0.2963913083076477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9583333134651184, 'rewards/lexical_plausibility/std': 0.07715168595314026, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9823122620582581, 'rewards/repeat_penalty/std': 0.014481368474662304, 'rewards/repeat_floor/mean': 0.9973468780517578, 'rewards/repeat_floor/std': 0.002172210020944476, 'rewards/near_duplicate_penalty/mean': 0.9823122620582581, 'rewards/near_duplicate_penalty/std': 0.014481368474662304, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20297935605049133, 'rewards/total_composite/std': 0.11905229836702347, 'reward': 0.20297935605049133, 'reward_std': 0.11905229091644287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1693575382232666, 'sampling/sampling_logp_difference/max': 1.5651897192001343, 'sampling/importance_sampling_ratio/min': 0.2090483456850052, 'sampling/importance_sampling_ratio/mean': 1.0169260501861572, 'sampling/importance_sampling_ratio/max': 1.8917797803878784, 'entropy': 1.4480042457580566, 'clip_ratio/low_mean': 0.055440512485802174, 'clip_ratio/low_min': 0.055440512485802174, 'clip_ratio/high_mean': 0.06758436374366283, 'clip_ratio/high_max': 0.06758436374366283, 'clip_ratio/region_mean': 0.12302487622946501, 'reward_total_mean': 0.20297935605049133, 'reward_meter_mean': 0.7712146043777466, 'reward_meter_std': 0.2963913083076477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9583333134651184, 'reward_lexical_plausibility_std': 0.07715168595314026, 'reward_repeat_penalty_mean': 0.9823122620582581, 'reward_repeat_penalty_std': 0.014481368474662304, 'reward_repeat_floor_mean': 0.9973468780517578, 'reward_repeat_floor_std': 0.002172210020944476, 'reward_near_duplicate_penalty_mean': 0.9823122620582581, 'reward_near_duplicate_penalty_std': 0.014481368474662304, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.20297935605049133, 'reward_total_composite_std': 0.11905229836702347, 'epoch': 0.06}

 37%|███▋      | 1231/3300 [4:42:32<5:46:30, 10.05s/it]INFO 04-12 20:40:58 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1232/3300 [4:42:46<6:34:11, 11.44s/it]2026-04-12 20:41:12,901 | INFO | train_grpo_train | metrics_logged mode=train step=1232

                                                       
{'loss': -0.0944, 'grad_norm': 1.9240204095840454, 'learning_rate': 6.26969696969697e-06, 'num_tokens': 2192110.0, 'completions/mean_length': 230.625, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 61.79999923706055, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.6351822018623352, 'rewards/meter/std': 0.3960917890071869, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.22903135418891907, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8968750238418579, 'rewards/count_floor/std': 0.06870941072702408, 'rewards/lexical_plausibility/mean': 0.8640150427818298, 'rewards/lexical_plausibility/std': 0.19077719748020172, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.8195990920066833, 'rewards/repeat_penalty/std': 0.19482024013996124, 'rewards/repeat_floor/mean': 0.9790264368057251, 'rewards/repeat_floor/std': 0.023111246526241302, 'rewards/near_duplicate_penalty/mean': 0.9283485412597656, 'rewards/near_duplicate_penalty/std': 0.06932410597801208, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9280460476875305, 'rewards/distinct_2/std': 0.10154540836811066, 'rewards/total_composite/mean': 0.1878122240304947, 'rewards/total_composite/std': 0.16088247299194336, 'reward': 0.1878122240304947, 'reward_std': 0.16088247299194336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16117610037326813, 'sampling/sampling_logp_difference/max': 2.038297653198242, 'sampling/importance_sampling_ratio/min': 0.1302502602338791, 'sampling/importance_sampling_ratio/mean': 1.0264315605163574, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6506724879145622, 'clip_ratio/low_mean': 0.010416666977107525, 'clip_ratio/low_min': 0.010416666977107525, 'clip_ratio/high_mean': 0.05322717409580946, 'clip_ratio/high_max': 0.05322717409580946, 'clip_ratio/region_mean': 0.06364384107291698, 'reward_total_mean': 0.1878122240304947, 'reward_meter_mean': 0.6351822018623352, 'reward_meter_std': 0.3960917890071869, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.22903135418891907, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8968750238418579, 'reward_count_floor_std': 0.06870941072702408, 'reward_lexical_plausibility_mean': 0.8640150427818298, 'reward_lexical_plausibility_std': 0.19077719748020172, 'reward_repeat_penalty_mean': 0.8195990920066833, 'reward_repeat_penalty_std': 0.19482024013996124, 'reward_repeat_floor_mean': 0.9790264368057251, 'reward_repeat_floor_std': 0.023111246526241302, 'reward_near_duplicate_penalty_mean': 0.9283485412597656, 'reward_near_duplicate_penalty_std': 0.06932410597801208, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9280460476875305, 'reward_distinct_2_std': 0.10154540836811066, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.1878122240304947, 'reward_total_composite_std': 0.16088247299194336, 'epoch': 0.06}

 37%|███▋      | 1232/3300 [4:42:46<6:34:11, 11.44s/it]INFO 04-12 20:41:13 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1233/3300 [4:43:01<7:06:39, 12.38s/it]2026-04-12 20:41:27,483 | INFO | train_grpo_train | metrics_logged mode=train step=1233

                                                       
{'loss': -0.0365, 'grad_norm': 5.7209553718566895, 'learning_rate': 6.266666666666668e-06, 'num_tokens': 2193670.0, 'completions/mean_length': 97.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.71428680419922, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.345956027507782, 'rewards/meter/std': 0.392540842294693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.964140772819519, 'rewards/lexical_plausibility/std': 0.10142520815134048, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9224766492843628, 'rewards/repeat_penalty/std': 0.13886435329914093, 'rewards/repeat_floor/mean': 0.9912518262863159, 'rewards/repeat_floor/std': 0.014125322923064232, 'rewards/near_duplicate_penalty/mean': 0.9680249691009521, 'rewards/near_duplicate_penalty/std': 0.036666907370090485, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9769231081008911, 'rewards/distinct_2/std': 0.04578607901930809, 'rewards/total_composite/mean': 0.11356526613235474, 'rewards/total_composite/std': 0.14179803431034088, 'reward': 0.11356526613235474, 'reward_std': 0.14179803431034088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16457846760749817, 'sampling/sampling_logp_difference/max': 1.6693463325500488, 'sampling/importance_sampling_ratio/min': 0.18837015330791473, 'sampling/importance_sampling_ratio/mean': 1.0209168195724487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3274514079093933, 'clip_ratio/low_mean': 0.0551587319932878, 'clip_ratio/low_min': 0.0551587319932878, 'clip_ratio/high_mean': 0.06817394308745861, 'clip_ratio/high_max': 0.06817394308745861, 'clip_ratio/region_mean': 0.12333267508074641, 'reward_total_mean': 0.11356526613235474, 'reward_meter_mean': 0.345956027507782, 'reward_meter_std': 0.392540842294693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.964140772819519, 'reward_lexical_plausibility_std': 0.10142520815134048, 'reward_repeat_penalty_mean': 0.9224766492843628, 'reward_repeat_penalty_std': 0.13886435329914093, 'reward_repeat_floor_mean': 0.9912518262863159, 'reward_repeat_floor_std': 0.014125322923064232, 'reward_near_duplicate_penalty_mean': 0.9680249691009521, 'reward_near_duplicate_penalty_std': 0.036666907370090485, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9769231081008911, 'reward_distinct_2_std': 0.04578607901930809, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.11356526613235474, 'reward_total_composite_std': 0.14179803431034088, 'epoch': 0.06}

 37%|███▋      | 1233/3300 [4:43:01<7:06:39, 12.38s/it]INFO 04-12 20:41:27 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1234/3300 [4:43:15<7:24:05, 12.90s/it]2026-04-12 20:41:41,583 | INFO | train_grpo_train | metrics_logged mode=train step=1234

                                                       
{'loss': -0.0196, 'grad_norm': 1.247450590133667, 'learning_rate': 6.263636363636364e-06, 'num_tokens': 2194934.0, 'completions/mean_length': 265.0, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 18.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.39558348059654236, 'rewards/meter/std': 0.38997212052345276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7501569390296936, 'rewards/lexical_plausibility/std': 0.26882272958755493, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.10350984334945679, 'rewards/repeat_penalty/mean': 0.7495925426483154, 'rewards/repeat_penalty/std': 0.0011524002766236663, 'rewards/repeat_floor/mean': 0.9849511384963989, 'rewards/repeat_floor/std': 0.0001382838236168027, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9994567632675171, 'rewards/distinct_2/std': 0.001536526600830257, 'rewards/total_composite/mean': 0.04846624284982681, 'rewards/total_composite/std': 0.06216733530163765, 'reward': 0.04846624284982681, 'reward_std': 0.06216733530163765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16079159080982208, 'sampling/sampling_logp_difference/max': 0.8369097709655762, 'sampling/importance_sampling_ratio/min': 0.43304669857025146, 'sampling/importance_sampling_ratio/mean': 1.0477958917617798, 'sampling/importance_sampling_ratio/max': 1.7438623905181885, 'entropy': 0.8355869203805923, 'clip_ratio/low_mean': 0.014705882407724857, 'clip_ratio/low_min': 0.014705882407724857, 'clip_ratio/high_mean': 0.06423611240461469, 'clip_ratio/high_max': 0.06423611240461469, 'clip_ratio/region_mean': 0.07894199481233954, 'reward_total_mean': 0.04846624284982681, 'reward_meter_mean': 0.39558348059654236, 'reward_meter_std': 0.38997212052345276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.7501569390296936, 'reward_lexical_plausibility_std': 0.26882272958755493, 'reward_repeat_penalty_mean': 0.7495925426483154, 'reward_repeat_penalty_std': 0.0011524002766236663, 'reward_repeat_floor_mean': 0.9849511384963989, 'reward_repeat_floor_std': 0.0001382838236168027, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9994567632675171, 'reward_distinct_2_std': 0.001536526600830257, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.10350984334945679, 'reward_total_composite_mean': 0.04846624284982681, 'reward_total_composite_std': 0.06216733530163765, 'epoch': 0.06}

 37%|███▋      | 1234/3300 [4:43:15<7:24:05, 12.90s/it]INFO 04-12 20:41:41 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1235/3300 [4:43:29<7:36:55, 13.28s/it]2026-04-12 20:41:55,739 | INFO | train_grpo_train | metrics_logged mode=train step=1235

                                                       
{'loss': -0.0914, 'grad_norm': 1.8325289487838745, 'learning_rate': 6.260606060606062e-06, 'num_tokens': 2196427.0, 'completions/mean_length': 154.625, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7336145043373108, 'rewards/meter/std': 0.4413461983203888, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8891605138778687, 'rewards/lexical_plausibility/std': 0.20523492991924286, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.6898847222328186, 'rewards/repeat_penalty/std': 0.32939156889915466, 'rewards/repeat_floor/mean': 0.9750527143478394, 'rewards/repeat_floor/std': 0.021731629967689514, 'rewards/near_duplicate_penalty/mean': 0.957435131072998, 'rewards/near_duplicate_penalty/std': 0.041575632989406586, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.32732683420181274, 'rewards/distinct_2/mean': 0.9604164361953735, 'rewards/distinct_2/std': 0.056162718683481216, 'rewards/total_composite/mean': 0.2856692373752594, 'rewards/total_composite/std': 0.18015047907829285, 'reward': 0.2856692373752594, 'reward_std': 0.18015047907829285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1669953167438507, 'sampling/sampling_logp_difference/max': 1.4733266830444336, 'sampling/importance_sampling_ratio/min': 0.22916185855865479, 'sampling/importance_sampling_ratio/mean': 1.0710846185684204, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0276648104190826, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10665093921124935, 'clip_ratio/high_max': 0.10665093921124935, 'clip_ratio/region_mean': 0.10665093921124935, 'reward_total_mean': 0.2856692373752594, 'reward_meter_mean': 0.7336145043373108, 'reward_meter_std': 0.4413461983203888, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8891605138778687, 'reward_lexical_plausibility_std': 0.20523492991924286, 'reward_repeat_penalty_mean': 0.6898847222328186, 'reward_repeat_penalty_std': 0.32939156889915466, 'reward_repeat_floor_mean': 0.9750527143478394, 'reward_repeat_floor_std': 0.021731629967689514, 'reward_near_duplicate_penalty_mean': 0.957435131072998, 'reward_near_duplicate_penalty_std': 0.041575632989406586, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.32732683420181274, 'reward_distinct_2_mean': 0.9604164361953735, 'reward_distinct_2_std': 0.056162718683481216, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.2856692373752594, 'reward_total_composite_std': 0.18015047907829285, 'epoch': 0.06}

 37%|███▋      | 1235/3300 [4:43:29<7:36:55, 13.28s/it]INFO 04-12 20:41:55 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1236/3300 [4:43:43<7:46:46, 13.57s/it]2026-04-12 20:42:09,988 | INFO | train_grpo_train | metrics_logged mode=train step=1236

                                                       
{'loss': -0.1221, 'grad_norm': 2.4707868099212646, 'learning_rate': 6.257575757575758e-06, 'num_tokens': 2198380.0, 'completions/mean_length': 231.125, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 62.60000228881836, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.9382025003433228, 'rewards/meter/std': 0.08106688410043716, 'rewards/count_adherence/mean': 0.4166666865348816, 'rewards/count_adherence/std': 0.1653386503458023, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8250000476837158, 'rewards/count_floor/std': 0.04960159957408905, 'rewards/lexical_plausibility/mean': 0.8655129671096802, 'rewards/lexical_plausibility/std': 0.1479320228099823, 'rewards/judge_quality/mean': 0.23125000298023224, 'rewards/judge_quality/std': 0.13611313700675964, 'rewards/repeat_penalty/mean': 0.901749849319458, 'rewards/repeat_penalty/std': 0.0676775798201561, 'rewards/repeat_floor/mean': 0.9873590469360352, 'rewards/repeat_floor/std': 0.008352715522050858, 'rewards/near_duplicate_penalty/mean': 0.9418801665306091, 'rewards/near_duplicate_penalty/std': 0.03421046957373619, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9564115405082703, 'rewards/distinct_2/std': 0.04288877546787262, 'rewards/total_composite/mean': 0.1718834787607193, 'rewards/total_composite/std': 0.10159119963645935, 'reward': 0.1718834787607193, 'reward_std': 0.10159119963645935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17375414073467255, 'sampling/sampling_logp_difference/max': 2.4321956634521484, 'sampling/importance_sampling_ratio/min': 0.2225916087627411, 'sampling/importance_sampling_ratio/mean': 1.0492405891418457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9741256237030029, 'clip_ratio/low_mean': 0.01744186133146286, 'clip_ratio/low_min': 0.01744186133146286, 'clip_ratio/high_mean': 0.08449536841362715, 'clip_ratio/high_max': 0.08449536841362715, 'clip_ratio/region_mean': 0.10193722974509001, 'reward_total_mean': 0.1718834787607193, 'reward_meter_mean': 0.9382025003433228, 'reward_meter_std': 0.08106688410043716, 'reward_count_adherence_mean': 0.4166666865348816, 'reward_count_adherence_std': 0.1653386503458023, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8250000476837158, 'reward_count_floor_std': 0.04960159957408905, 'reward_lexical_plausibility_mean': 0.8655129671096802, 'reward_lexical_plausibility_std': 0.1479320228099823, 'reward_repeat_penalty_mean': 0.901749849319458, 'reward_repeat_penalty_std': 0.0676775798201561, 'reward_repeat_floor_mean': 0.9873590469360352, 'reward_repeat_floor_std': 0.008352715522050858, 'reward_near_duplicate_penalty_mean': 0.9418801665306091, 'reward_near_duplicate_penalty_std': 0.03421046957373619, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9564115405082703, 'reward_distinct_2_std': 0.04288877546787262, 'reward_judge_quality_mean': 0.23125000298023224, 'reward_judge_quality_std': 0.13611313700675964, 'reward_total_composite_mean': 0.1718834787607193, 'reward_total_composite_std': 0.10159119963645935, 'epoch': 0.07}

 37%|███▋      | 1236/3300 [4:43:43<7:46:46, 13.57s/it]INFO 04-12 20:42:10 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1237/3300 [4:43:57<7:53:24, 13.77s/it]2026-04-12 20:42:24,228 | INFO | train_grpo_train | metrics_logged mode=train step=1237

                                                       
{'loss': -0.0557, 'grad_norm': 5.74350118637085, 'learning_rate': 6.254545454545455e-06, 'num_tokens': 2200015.0, 'completions/mean_length': 105.375, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.28571701049805, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6636344194412231, 'rewards/meter/std': 0.3934630751609802, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.9641571044921875, 'rewards/lexical_plausibility/std': 0.10137906670570374, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9089483022689819, 'rewards/repeat_penalty/std': 0.06338323652744293, 'rewards/repeat_floor/mean': 0.9879863262176514, 'rewards/repeat_floor/std': 0.008179265074431896, 'rewards/near_duplicate_penalty/mean': 0.9401991367340088, 'rewards/near_duplicate_penalty/std': 0.04138028994202614, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9661828279495239, 'rewards/distinct_2/std': 0.03781760856509209, 'rewards/total_composite/mean': 0.2201250195503235, 'rewards/total_composite/std': 0.15626908838748932, 'reward': 0.2201250195503235, 'reward_std': 0.15626908838748932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18314969539642334, 'sampling/sampling_logp_difference/max': 1.3584260940551758, 'sampling/importance_sampling_ratio/min': 0.2570650577545166, 'sampling/importance_sampling_ratio/mean': 1.0382224321365356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3279864937067032, 'clip_ratio/low_mean': 0.023518981877714396, 'clip_ratio/low_min': 0.023518981877714396, 'clip_ratio/high_mean': 0.08756113331764936, 'clip_ratio/high_max': 0.08756113331764936, 'clip_ratio/region_mean': 0.11108011519536376, 'reward_total_mean': 0.2201250195503235, 'reward_meter_mean': 0.6636344194412231, 'reward_meter_std': 0.3934630751609802, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.9641571044921875, 'reward_lexical_plausibility_std': 0.10137906670570374, 'reward_repeat_penalty_mean': 0.9089483022689819, 'reward_repeat_penalty_std': 0.06338323652744293, 'reward_repeat_floor_mean': 0.9879863262176514, 'reward_repeat_floor_std': 0.008179265074431896, 'reward_near_duplicate_penalty_mean': 0.9401991367340088, 'reward_near_duplicate_penalty_std': 0.04138028994202614, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9661828279495239, 'reward_distinct_2_std': 0.03781760856509209, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2201250195503235, 'reward_total_composite_std': 0.15626908838748932, 'epoch': 0.07}

 37%|███▋      | 1237/3300 [4:43:58<7:53:24, 13.77s/it]INFO 04-12 20:42:24 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1238/3300 [4:44:06<6:55:25, 12.09s/it]2026-04-12 20:42:32,394 | INFO | train_grpo_train | metrics_logged mode=train step=1238

                                                       
{'loss': -0.0297, 'grad_norm': 14.210705757141113, 'learning_rate': 6.251515151515152e-06, 'num_tokens': 2201544.0, 'completions/mean_length': 34.125, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5090712904930115, 'rewards/meter/std': 0.3570987582206726, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9781991839408875, 'rewards/lexical_plausibility/std': 0.06166204437613487, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101401686668396, 'rewards/repeat_penalty/mean': 0.9412117600440979, 'rewards/repeat_penalty/std': 0.058571357280015945, 'rewards/repeat_floor/mean': 0.9920860528945923, 'rewards/repeat_floor/std': 0.007197629660367966, 'rewards/near_duplicate_penalty/mean': 0.9585583209991455, 'rewards/near_duplicate_penalty/std': 0.02920488268136978, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9811372756958008, 'rewards/distinct_2/std': 0.035051409155130386, 'rewards/total_composite/mean': 0.304751992225647, 'rewards/total_composite/std': 0.3495134115219116, 'reward': 0.304751992225647, 'reward_std': 0.3495134115219116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1359197348356247, 'sampling/sampling_logp_difference/max': 1.2594246864318848, 'sampling/importance_sampling_ratio/min': 0.28381726145744324, 'sampling/importance_sampling_ratio/mean': 1.0325759649276733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2088068053126335, 'clip_ratio/low_mean': 0.09459223784506321, 'clip_ratio/low_min': 0.09459223784506321, 'clip_ratio/high_mean': 0.05419511627405882, 'clip_ratio/high_max': 0.05419511627405882, 'clip_ratio/region_mean': 0.14878735411912203, 'reward_total_mean': 0.304751992225647, 'reward_meter_mean': 0.5090712904930115, 'reward_meter_std': 0.3570987582206726, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9781991839408875, 'reward_lexical_plausibility_std': 0.06166204437613487, 'reward_repeat_penalty_mean': 0.9412117600440979, 'reward_repeat_penalty_std': 0.058571357280015945, 'reward_repeat_floor_mean': 0.9920860528945923, 'reward_repeat_floor_std': 0.007197629660367966, 'reward_near_duplicate_penalty_mean': 0.9585583209991455, 'reward_near_duplicate_penalty_std': 0.02920488268136978, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9811372756958008, 'reward_distinct_2_std': 0.035051409155130386, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101401686668396, 'reward_total_composite_mean': 0.304751992225647, 'reward_total_composite_std': 0.3495134115219116, 'epoch': 0.07}

 38%|███▊      | 1238/3300 [4:44:06<6:55:25, 12.09s/it]INFO 04-12 20:42:32 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1239/3300 [4:44:14<6:13:47, 10.88s/it]2026-04-12 20:42:40,455 | INFO | train_grpo_train | metrics_logged mode=train step=1239

                                                       
{'loss': -0.0679, 'grad_norm': 14.023561477661133, 'learning_rate': 6.248484848484849e-06, 'num_tokens': 2203072.0, 'completions/mean_length': 30.0, 'completions/min_length': 5.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 5.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5566118955612183, 'rewards/meter/std': 0.4028037488460541, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2781743109226227, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8874999284744263, 'rewards/count_floor/std': 0.08345229923725128, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.768750011920929, 'rewards/judge_quality/std': 0.3004015386104584, 'rewards/repeat_penalty/mean': 0.9690935611724854, 'rewards/repeat_penalty/std': 0.01671244017779827, 'rewards/repeat_floor/mean': 0.9953640699386597, 'rewards/repeat_floor/std': 0.0025068672839552164, 'rewards/near_duplicate_penalty/mean': 0.9690935611724854, 'rewards/near_duplicate_penalty/std': 0.01671244017779827, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.45701491832733154, 'rewards/total_composite/std': 0.3411944806575775, 'reward': 0.45701491832733154, 'reward_std': 0.3411944508552551, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12629644572734833, 'sampling/sampling_logp_difference/max': 1.4598636627197266, 'sampling/importance_sampling_ratio/min': 0.23226793110370636, 'sampling/importance_sampling_ratio/mean': 1.0134378671646118, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45294495671987534, 'clip_ratio/low_mean': 0.06294171139597893, 'clip_ratio/low_min': 0.06294171139597893, 'clip_ratio/high_mean': 0.04567307699471712, 'clip_ratio/high_max': 0.04567307699471712, 'clip_ratio/region_mean': 0.10861478839069605, 'reward_total_mean': 0.45701491832733154, 'reward_meter_mean': 0.5566118955612183, 'reward_meter_std': 0.4028037488460541, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2781743109226227, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8874999284744263, 'reward_count_floor_std': 0.08345229923725128, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9690935611724854, 'reward_repeat_penalty_std': 0.01671244017779827, 'reward_repeat_floor_mean': 0.9953640699386597, 'reward_repeat_floor_std': 0.0025068672839552164, 'reward_near_duplicate_penalty_mean': 0.9690935611724854, 'reward_near_duplicate_penalty_std': 0.01671244017779827, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.768750011920929, 'reward_judge_quality_std': 0.3004015386104584, 'reward_total_composite_mean': 0.45701491832733154, 'reward_total_composite_std': 0.3411944806575775, 'epoch': 0.07}

 38%|███▊      | 1239/3300 [4:44:14<6:13:47, 10.88s/it]INFO 04-12 20:42:40 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1240/3300 [4:44:21<5:38:48,  9.87s/it]2026-04-12 20:42:47,963 | INFO | train_grpo_train | metrics_logged mode=train step=1240

                                                       
{'loss': 0.0112, 'grad_norm': 17.732521057128906, 'learning_rate': 6.245454545454545e-06, 'num_tokens': 2204286.0, 'completions/mean_length': 14.75, 'completions/min_length': 13.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 14.75, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.9761737585067749, 'rewards/meter/std': 0.01978938840329647, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.24032345414161682, 'rewards/repeat_penalty/mean': 0.7171111106872559, 'rewards/repeat_penalty/std': 0.052729666233062744, 'rewards/repeat_floor/mean': 0.978422224521637, 'rewards/repeat_floor/std': 0.010545928031206131, 'rewards/near_duplicate_penalty/mean': 0.9561481475830078, 'rewards/near_duplicate_penalty/std': 0.0703062191605568, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.40208137035369873, 'rewards/total_composite/std': 0.227709099650383, 'reward': 0.40208137035369873, 'reward_std': 0.227709099650383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.139068141579628, 'sampling/sampling_logp_difference/max': 0.9917736053466797, 'sampling/importance_sampling_ratio/min': 0.37091824412345886, 'sampling/importance_sampling_ratio/mean': 1.010571837425232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.176104299724102, 'clip_ratio/low_mean': 0.03708791360259056, 'clip_ratio/low_min': 0.03708791360259056, 'clip_ratio/high_mean': 0.08474702574312687, 'clip_ratio/high_max': 0.08474702574312687, 'clip_ratio/region_mean': 0.12183493934571743, 'reward_total_mean': 0.40208137035369873, 'reward_meter_mean': 0.9761737585067749, 'reward_meter_std': 0.01978938840329647, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7171111106872559, 'reward_repeat_penalty_std': 0.052729666233062744, 'reward_repeat_floor_mean': 0.978422224521637, 'reward_repeat_floor_std': 0.010545928031206131, 'reward_near_duplicate_penalty_mean': 0.9561481475830078, 'reward_near_duplicate_penalty_std': 0.0703062191605568, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.24032345414161682, 'reward_total_composite_mean': 0.40208137035369873, 'reward_total_composite_std': 0.227709099650383, 'epoch': 0.07}

 38%|███▊      | 1240/3300 [4:44:21<5:38:48,  9.87s/it]INFO 04-12 20:42:48 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1241/3300 [4:44:29<5:13:32,  9.14s/it]2026-04-12 20:42:55,389 | INFO | train_grpo_train | metrics_logged mode=train step=1241

                                                       
{'loss': 0.0515, 'grad_norm': 28.220304489135742, 'learning_rate': 6.2424242424242434e-06, 'num_tokens': 2205802.0, 'completions/mean_length': 17.5, 'completions/min_length': 16.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.7593625783920288, 'rewards/meter/std': 0.25426849722862244, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7887499928474426, 'rewards/judge_quality/std': 0.27115824818611145, 'rewards/repeat_penalty/mean': 0.7366601228713989, 'rewards/repeat_penalty/std': 0.025940246880054474, 'rewards/repeat_floor/mean': 0.9823319911956787, 'rewards/repeat_floor/std': 0.005188049282878637, 'rewards/near_duplicate_penalty/mean': 0.9822134375572205, 'rewards/near_duplicate_penalty/std': 0.034586984664201736, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5923237800598145, 'rewards/total_composite/std': 0.2872130274772644, 'reward': 0.5923237800598145, 'reward_std': 0.2872130274772644, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14005941152572632, 'sampling/sampling_logp_difference/max': 1.0842657089233398, 'sampling/importance_sampling_ratio/min': 0.4309065043926239, 'sampling/importance_sampling_ratio/mean': 0.9893203973770142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9171800911426544, 'clip_ratio/low_mean': 0.0390625, 'clip_ratio/low_min': 0.0390625, 'clip_ratio/high_mean': 0.10661764768883586, 'clip_ratio/high_max': 0.10661764768883586, 'clip_ratio/region_mean': 0.14568014768883586, 'reward_total_mean': 0.5923237800598145, 'reward_meter_mean': 0.7593625783920288, 'reward_meter_std': 0.25426849722862244, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7366601228713989, 'reward_repeat_penalty_std': 0.025940246880054474, 'reward_repeat_floor_mean': 0.9823319911956787, 'reward_repeat_floor_std': 0.005188049282878637, 'reward_near_duplicate_penalty_mean': 0.9822134375572205, 'reward_near_duplicate_penalty_std': 0.034586984664201736, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7887499928474426, 'reward_judge_quality_std': 0.27115824818611145, 'reward_total_composite_mean': 0.5923237800598145, 'reward_total_composite_std': 0.2872130274772644, 'epoch': 0.07}

 38%|███▊      | 1241/3300 [4:44:29<5:13:32,  9.14s/it]INFO 04-12 20:42:55 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1242/3300 [4:44:43<6:04:30, 10.63s/it]2026-04-12 20:43:09,496 | INFO | train_grpo_train | metrics_logged mode=train step=1242

                                                       
{'loss': -0.0633, 'grad_norm': 4.929039001464844, 'learning_rate': 6.23939393939394e-06, 'num_tokens': 2207455.0, 'completions/mean_length': 101.625, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.000003814697266, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9189140796661377, 'rewards/meter/std': 0.10611801594495773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9609698057174683, 'rewards/lexical_plausibility/std': 0.11039415001869202, 'rewards/judge_quality/mean': 0.20625001192092896, 'rewards/judge_quality/std': 0.1293872892856598, 'rewards/repeat_penalty/mean': 0.9624922871589661, 'rewards/repeat_penalty/std': 0.053655955940485, 'rewards/repeat_floor/mean': 0.9946796298027039, 'rewards/repeat_floor/std': 0.007227218244224787, 'rewards/near_duplicate_penalty/mean': 0.9688042402267456, 'rewards/near_duplicate_penalty/std': 0.03699269890785217, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9928774833679199, 'rewards/distinct_2/std': 0.02014549821615219, 'rewards/total_composite/mean': 0.1921328902244568, 'rewards/total_composite/std': 0.12995624542236328, 'reward': 0.1921328902244568, 'reward_std': 0.1299562305212021, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1773170381784439, 'sampling/sampling_logp_difference/max': 1.926767349243164, 'sampling/importance_sampling_ratio/min': 0.14561817049980164, 'sampling/importance_sampling_ratio/mean': 1.056402564048767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3925556987524033, 'clip_ratio/low_mean': 0.07583586126565933, 'clip_ratio/low_min': 0.07583586126565933, 'clip_ratio/high_mean': 0.04462869931012392, 'clip_ratio/high_max': 0.04462869931012392, 'clip_ratio/region_mean': 0.12046456057578325, 'reward_total_mean': 0.1921328902244568, 'reward_meter_mean': 0.9189140796661377, 'reward_meter_std': 0.10611801594495773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9609698057174683, 'reward_lexical_plausibility_std': 0.11039415001869202, 'reward_repeat_penalty_mean': 0.9624922871589661, 'reward_repeat_penalty_std': 0.053655955940485, 'reward_repeat_floor_mean': 0.9946796298027039, 'reward_repeat_floor_std': 0.007227218244224787, 'reward_near_duplicate_penalty_mean': 0.9688042402267456, 'reward_near_duplicate_penalty_std': 0.03699269890785217, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9928774833679199, 'reward_distinct_2_std': 0.02014549821615219, 'reward_judge_quality_mean': 0.20625001192092896, 'reward_judge_quality_std': 0.1293872892856598, 'reward_total_composite_mean': 0.1921328902244568, 'reward_total_composite_std': 0.12995624542236328, 'epoch': 0.07}

 38%|███▊      | 1242/3300 [4:44:43<6:04:30, 10.63s/it]INFO 04-12 20:43:09 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1243/3300 [4:44:57<6:39:39, 11.66s/it]2026-04-12 20:43:23,550 | INFO | train_grpo_train | metrics_logged mode=train step=1243

                                                       
{'loss': -0.1326, 'grad_norm': 1.853240728378296, 'learning_rate': 6.236363636363637e-06, 'num_tokens': 2209666.0, 'completions/mean_length': 249.375, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 91.80000305175781, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.8503066897392273, 'rewards/meter/std': 0.21898460388183594, 'rewards/count_adherence/mean': 0.4107142984867096, 'rewards/count_adherence/std': 0.1937432438135147, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8232142925262451, 'rewards/count_floor/std': 0.05812295898795128, 'rewards/lexical_plausibility/mean': 0.883672833442688, 'rewards/lexical_plausibility/std': 0.17210981249809265, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.17728105187416077, 'rewards/repeat_penalty/mean': 0.8924397826194763, 'rewards/repeat_penalty/std': 0.12868241965770721, 'rewards/repeat_floor/mean': 0.9905014634132385, 'rewards/repeat_floor/std': 0.010197503492236137, 'rewards/near_duplicate_penalty/mean': 0.9795410633087158, 'rewards/near_duplicate_penalty/std': 0.02328227460384369, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9702256917953491, 'rewards/distinct_2/std': 0.039875928312540054, 'rewards/total_composite/mean': 0.18152840435504913, 'rewards/total_composite/std': 0.1370563805103302, 'reward': 0.18152840435504913, 'reward_std': 0.1370563805103302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14794373512268066, 'sampling/sampling_logp_difference/max': 1.7479491233825684, 'sampling/importance_sampling_ratio/min': 0.17413069307804108, 'sampling/importance_sampling_ratio/mean': 1.0459083318710327, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8835892453789711, 'clip_ratio/low_mean': 0.02604166604578495, 'clip_ratio/low_min': 0.02604166604578495, 'clip_ratio/high_mean': 0.05065748514607549, 'clip_ratio/high_max': 0.05065748514607549, 'clip_ratio/region_mean': 0.07669915119186044, 'reward_total_mean': 0.18152840435504913, 'reward_meter_mean': 0.8503066897392273, 'reward_meter_std': 0.21898460388183594, 'reward_count_adherence_mean': 0.4107142984867096, 'reward_count_adherence_std': 0.1937432438135147, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8232142925262451, 'reward_count_floor_std': 0.05812295898795128, 'reward_lexical_plausibility_mean': 0.883672833442688, 'reward_lexical_plausibility_std': 0.17210981249809265, 'reward_repeat_penalty_mean': 0.8924397826194763, 'reward_repeat_penalty_std': 0.12868241965770721, 'reward_repeat_floor_mean': 0.9905014634132385, 'reward_repeat_floor_std': 0.010197503492236137, 'reward_near_duplicate_penalty_mean': 0.9795410633087158, 'reward_near_duplicate_penalty_std': 0.02328227460384369, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9702256917953491, 'reward_distinct_2_std': 0.039875928312540054, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.17728105187416077, 'reward_total_composite_mean': 0.18152840435504913, 'reward_total_composite_std': 0.1370563805103302, 'epoch': 0.07}

 38%|███▊      | 1243/3300 [4:44:57<6:39:39, 11.66s/it]INFO 04-12 20:43:23 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1244/3300 [4:45:10<6:59:12, 12.23s/it]2026-04-12 20:43:37,137 | INFO | train_grpo_train | metrics_logged mode=train step=1244

                                                       
{'loss': -0.1541, 'grad_norm': 1.9030165672302246, 'learning_rate': 6.2333333333333335e-06, 'num_tokens': 2211666.0, 'completions/mean_length': 181.0, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 70.66667175292969, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9835989475250244, 'rewards/meter/std': 0.013030430302023888, 'rewards/count_adherence/mean': 0.7000000476837158, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9099999666213989, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.8965067863464355, 'rewards/lexical_plausibility/std': 0.193262979388237, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.28822651505470276, 'rewards/repeat_penalty/std': 0.2483576536178589, 'rewards/repeat_floor/mean': 0.9058868885040283, 'rewards/repeat_floor/std': 0.0566122941672802, 'rewards/near_duplicate_penalty/mean': 0.7477419376373291, 'rewards/near_duplicate_penalty/std': 0.1612212359905243, 'rewards/opening_diversity/mean': 0.4352678656578064, 'rewards/opening_diversity/std': 0.2844659388065338, 'rewards/distinct_2/mean': 0.7512165904045105, 'rewards/distinct_2/std': 0.25841039419174194, 'rewards/total_composite/mean': 0.22412312030792236, 'rewards/total_composite/std': 0.11611055582761765, 'reward': 0.22412312030792236, 'reward_std': 0.11611056327819824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13060210645198822, 'sampling/sampling_logp_difference/max': 1.560983657836914, 'sampling/importance_sampling_ratio/min': 0.2099294662475586, 'sampling/importance_sampling_ratio/mean': 1.010269284248352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5903434529900551, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07258011866360903, 'clip_ratio/high_max': 0.07258011866360903, 'clip_ratio/region_mean': 0.07258011866360903, 'reward_total_mean': 0.22412312030792236, 'reward_meter_mean': 0.9835989475250244, 'reward_meter_std': 0.013030430302023888, 'reward_count_adherence_mean': 0.7000000476837158, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9099999666213989, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.8965067863464355, 'reward_lexical_plausibility_std': 0.193262979388237, 'reward_repeat_penalty_mean': 0.28822651505470276, 'reward_repeat_penalty_std': 0.2483576536178589, 'reward_repeat_floor_mean': 0.9058868885040283, 'reward_repeat_floor_std': 0.0566122941672802, 'reward_near_duplicate_penalty_mean': 0.7477419376373291, 'reward_near_duplicate_penalty_std': 0.1612212359905243, 'reward_opening_diversity_mean': 0.4352678656578064, 'reward_opening_diversity_std': 0.2844659388065338, 'reward_distinct_2_mean': 0.7512165904045105, 'reward_distinct_2_std': 0.25841039419174194, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.22412312030792236, 'reward_total_composite_std': 0.11611055582761765, 'epoch': 0.07}

 38%|███▊      | 1244/3300 [4:45:10<6:59:12, 12.23s/it]INFO 04-12 20:43:37 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1245/3300 [4:45:20<6:28:15, 11.34s/it]2026-04-12 20:43:46,377 | INFO | train_grpo_train | metrics_logged mode=train step=1245

                                                       
{'loss': 0.0629, 'grad_norm': 15.939921379089355, 'learning_rate': 6.230303030303031e-06, 'num_tokens': 2213499.0, 'completions/mean_length': 46.125, 'completions/min_length': 31.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8733986616134644, 'rewards/meter/std': 0.17946261167526245, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 0.9156249761581421, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9535636901855469, 'rewards/repeat_penalty/std': 0.04536581039428711, 'rewards/repeat_floor/mean': 0.9941201210021973, 'rewards/repeat_floor/std': 0.004832054954022169, 'rewards/near_duplicate_penalty/mean': 0.9727669954299927, 'rewards/near_duplicate_penalty/std': 0.014959770254790783, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9800570011138916, 'rewards/distinct_2/std': 0.038885053247213364, 'rewards/total_composite/mean': 0.3151639699935913, 'rewards/total_composite/std': 0.08295875042676926, 'reward': 0.3151639699935913, 'reward_std': 0.08295875042676926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16736873984336853, 'sampling/sampling_logp_difference/max': 2.2824487686157227, 'sampling/importance_sampling_ratio/min': 0.10203403979539871, 'sampling/importance_sampling_ratio/mean': 1.0098350048065186, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9740136489272118, 'clip_ratio/low_mean': 0.06517662992700934, 'clip_ratio/low_min': 0.06517662992700934, 'clip_ratio/high_mean': 0.08775888569653034, 'clip_ratio/high_max': 0.08775888569653034, 'clip_ratio/region_mean': 0.1529355156235397, 'reward_total_mean': 0.3151639699935913, 'reward_meter_mean': 0.8733986616134644, 'reward_meter_std': 0.17946261167526245, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 0.9156249761581421, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9535636901855469, 'reward_repeat_penalty_std': 0.04536581039428711, 'reward_repeat_floor_mean': 0.9941201210021973, 'reward_repeat_floor_std': 0.004832054954022169, 'reward_near_duplicate_penalty_mean': 0.9727669954299927, 'reward_near_duplicate_penalty_std': 0.014959770254790783, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9800570011138916, 'reward_distinct_2_std': 0.038885053247213364, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3151639699935913, 'reward_total_composite_std': 0.08295875042676926, 'epoch': 0.07}

 38%|███▊      | 1245/3300 [4:45:20<6:28:15, 11.34s/it]INFO 04-12 20:43:46 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1246/3300 [4:45:34<6:59:18, 12.25s/it]2026-04-12 20:44:00,767 | INFO | train_grpo_train | metrics_logged mode=train step=1246

                                                       
{'loss': -0.1205, 'grad_norm': 2.6888742446899414, 'learning_rate': 6.227272727272727e-06, 'num_tokens': 2215517.0, 'completions/mean_length': 176.25, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 64.33333587646484, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.513663113117218, 'rewards/meter/std': 0.2534431219100952, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.9237847328186035, 'rewards/lexical_plausibility/std': 0.14155985414981842, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.18077214062213898, 'rewards/repeat_penalty/mean': 0.9332049489021301, 'rewards/repeat_penalty/std': 0.04546697065234184, 'rewards/repeat_floor/mean': 0.9919899702072144, 'rewards/repeat_floor/std': 0.005299018695950508, 'rewards/near_duplicate_penalty/mean': 0.9688516855239868, 'rewards/near_duplicate_penalty/std': 0.021502871066331863, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.962913990020752, 'rewards/distinct_2/std': 0.03316647931933403, 'rewards/total_composite/mean': 0.17804402112960815, 'rewards/total_composite/std': 0.11710407584905624, 'reward': 0.17804402112960815, 'reward_std': 0.11710407584905624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1626768261194229, 'sampling/sampling_logp_difference/max': 1.681884765625, 'sampling/importance_sampling_ratio/min': 0.1860230267047882, 'sampling/importance_sampling_ratio/mean': 1.0211745500564575, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7871479243040085, 'clip_ratio/low_mean': 0.019480518996715546, 'clip_ratio/low_min': 0.019480518996715546, 'clip_ratio/high_mean': 0.07765456382185221, 'clip_ratio/high_max': 0.07765456382185221, 'clip_ratio/region_mean': 0.09713508281856775, 'reward_total_mean': 0.17804402112960815, 'reward_meter_mean': 0.513663113117218, 'reward_meter_std': 0.2534431219100952, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.9237847328186035, 'reward_lexical_plausibility_std': 0.14155985414981842, 'reward_repeat_penalty_mean': 0.9332049489021301, 'reward_repeat_penalty_std': 0.04546697065234184, 'reward_repeat_floor_mean': 0.9919899702072144, 'reward_repeat_floor_std': 0.005299018695950508, 'reward_near_duplicate_penalty_mean': 0.9688516855239868, 'reward_near_duplicate_penalty_std': 0.021502871066331863, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.962913990020752, 'reward_distinct_2_std': 0.03316647931933403, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.18077214062213898, 'reward_total_composite_mean': 0.17804402112960815, 'reward_total_composite_std': 0.11710407584905624, 'epoch': 0.07}

 38%|███▊      | 1246/3300 [4:45:34<6:59:18, 12.25s/it]INFO 04-12 20:44:01 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1247/3300 [4:45:42<6:20:13, 11.11s/it]2026-04-12 20:44:09,232 | INFO | train_grpo_train | metrics_logged mode=train step=1247

                                                       
{'loss': 0.023, 'grad_norm': 20.79667091369629, 'learning_rate': 6.224242424242425e-06, 'num_tokens': 2217002.0, 'completions/mean_length': 20.625, 'completions/min_length': 18.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.7085248827934265, 'rewards/meter/std': 0.38873299956321716, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.15059404075145721, 'rewards/repeat_penalty/mean': 0.7411185503005981, 'rewards/repeat_penalty/std': 0.019554078578948975, 'rewards/repeat_floor/mean': 0.9832237362861633, 'rewards/repeat_floor/std': 0.003910813480615616, 'rewards/near_duplicate_penalty/mean': 0.9881580471992493, 'rewards/near_duplicate_penalty/std': 0.026072107255458832, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19634845852851868, 'rewards/total_composite/std': 0.14513981342315674, 'reward': 0.19634845852851868, 'reward_std': 0.14513979852199554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17221178114414215, 'sampling/sampling_logp_difference/max': 1.073185920715332, 'sampling/importance_sampling_ratio/min': 0.341917484998703, 'sampling/importance_sampling_ratio/mean': 1.0430773496627808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5565488040447235, 'clip_ratio/low_mean': 0.10535714589059353, 'clip_ratio/low_min': 0.10535714589059353, 'clip_ratio/high_mean': 0.05227272864431143, 'clip_ratio/high_max': 0.05227272864431143, 'clip_ratio/region_mean': 0.15762987453490496, 'reward_total_mean': 0.19634845852851868, 'reward_meter_mean': 0.7085248827934265, 'reward_meter_std': 0.38873299956321716, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7411185503005981, 'reward_repeat_penalty_std': 0.019554078578948975, 'reward_repeat_floor_mean': 0.9832237362861633, 'reward_repeat_floor_std': 0.003910813480615616, 'reward_near_duplicate_penalty_mean': 0.9881580471992493, 'reward_near_duplicate_penalty_std': 0.026072107255458832, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.15059404075145721, 'reward_total_composite_mean': 0.19634845852851868, 'reward_total_composite_std': 0.14513981342315674, 'epoch': 0.07}

 38%|███▊      | 1247/3300 [4:45:43<6:20:13, 11.11s/it]INFO 04-12 20:44:09 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1248/3300 [4:45:52<5:59:25, 10.51s/it]2026-04-12 20:44:18,331 | INFO | train_grpo_train | metrics_logged mode=train step=1248

                                                       
{'loss': 0.0507, 'grad_norm': 15.76718521118164, 'learning_rate': 6.221212121212121e-06, 'num_tokens': 2218935.0, 'completions/mean_length': 59.625, 'completions/min_length': 57.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5233222246170044, 'rewards/meter/std': 0.38666650652885437, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.9243854284286499, 'rewards/repeat_penalty/std': 0.0598582923412323, 'rewards/repeat_floor/mean': 0.9905862212181091, 'rewards/repeat_floor/std': 0.0066826920956373215, 'rewards/near_duplicate_penalty/mean': 0.9595956802368164, 'rewards/near_duplicate_penalty/std': 0.02047358825802803, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9627429246902466, 'rewards/distinct_2/std': 0.04781792685389519, 'rewards/total_composite/mean': 0.17514003813266754, 'rewards/total_composite/std': 0.14418698847293854, 'reward': 0.17514003813266754, 'reward_std': 0.14418698847293854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.152029350399971, 'sampling/sampling_logp_difference/max': 2.5474061965942383, 'sampling/importance_sampling_ratio/min': 0.07828445732593536, 'sampling/importance_sampling_ratio/mean': 1.0235058069229126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1948922351002693, 'clip_ratio/low_mean': 0.09359662793576717, 'clip_ratio/low_min': 0.09359662793576717, 'clip_ratio/high_mean': 0.05369666777551174, 'clip_ratio/high_max': 0.05369666777551174, 'clip_ratio/region_mean': 0.14729329571127892, 'reward_total_mean': 0.17514003813266754, 'reward_meter_mean': 0.5233222246170044, 'reward_meter_std': 0.38666650652885437, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9243854284286499, 'reward_repeat_penalty_std': 0.0598582923412323, 'reward_repeat_floor_mean': 0.9905862212181091, 'reward_repeat_floor_std': 0.0066826920956373215, 'reward_near_duplicate_penalty_mean': 0.9595956802368164, 'reward_near_duplicate_penalty_std': 0.02047358825802803, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9627429246902466, 'reward_distinct_2_std': 0.04781792685389519, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.17514003813266754, 'reward_total_composite_std': 0.14418698847293854, 'epoch': 0.07}

 38%|███▊      | 1248/3300 [4:45:52<5:59:25, 10.51s/it]INFO 04-12 20:44:18 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1249/3300 [4:46:01<5:46:36, 10.14s/it]2026-04-12 20:44:27,603 | INFO | train_grpo_train | metrics_logged mode=train step=1249

                                                       
{'loss': 0.0949, 'grad_norm': 11.137635231018066, 'learning_rate': 6.218181818181819e-06, 'num_tokens': 2220625.0, 'completions/mean_length': 43.25, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6899032592773438, 'rewards/meter/std': 0.39812588691711426, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.3258587718009949, 'rewards/repeat_penalty/mean': 0.9699445962905884, 'rewards/repeat_penalty/std': 0.02828710526227951, 'rewards/repeat_floor/mean': 0.9959218502044678, 'rewards/repeat_floor/std': 0.0036720731295645237, 'rewards/near_duplicate_penalty/mean': 0.9772500991821289, 'rewards/near_duplicate_penalty/std': 0.02340538427233696, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.2737385630607605, 'rewards/total_composite/std': 0.2619189918041229, 'reward': 0.2737385630607605, 'reward_std': 0.2619189918041229, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1389586329460144, 'sampling/sampling_logp_difference/max': 1.3593130111694336, 'sampling/importance_sampling_ratio/min': 0.2568371593952179, 'sampling/importance_sampling_ratio/mean': 1.0159103870391846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9785799756646156, 'clip_ratio/low_mean': 0.07975180726498365, 'clip_ratio/low_min': 0.07975180726498365, 'clip_ratio/high_mean': 0.050666118040680885, 'clip_ratio/high_max': 0.050666118040680885, 'clip_ratio/region_mean': 0.13041792530566454, 'reward_total_mean': 0.2737385630607605, 'reward_meter_mean': 0.6899032592773438, 'reward_meter_std': 0.39812588691711426, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9699445962905884, 'reward_repeat_penalty_std': 0.02828710526227951, 'reward_repeat_floor_mean': 0.9959218502044678, 'reward_repeat_floor_std': 0.0036720731295645237, 'reward_near_duplicate_penalty_mean': 0.9772500991821289, 'reward_near_duplicate_penalty_std': 0.02340538427233696, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.3258587718009949, 'reward_total_composite_mean': 0.2737385630607605, 'reward_total_composite_std': 0.2619189918041229, 'epoch': 0.07}

 38%|███▊      | 1249/3300 [4:46:01<5:46:36, 10.14s/it]INFO 04-12 20:44:27 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1250/3300 [4:46:10<5:36:15,  9.84s/it]2026-04-12 20:44:36,749 | INFO | train_grpo_train | metrics_logged mode=train step=1250

                                                       
{'loss': -0.0456, 'grad_norm': 18.66234016418457, 'learning_rate': 6.215151515151515e-06, 'num_tokens': 2221926.0, 'completions/mean_length': 20.625, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.5020642280578613, 'rewards/meter/std': 0.48159053921699524, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8541666269302368, 'rewards/lexical_plausibility/std': 0.3500283658504486, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.7149804830551147, 'rewards/repeat_penalty/std': 0.08950236439704895, 'rewards/repeat_floor/mean': 0.9785131216049194, 'rewards/repeat_floor/std': 0.016445640474557877, 'rewards/near_duplicate_penalty/mean': 0.9682923555374146, 'rewards/near_duplicate_penalty/std': 0.07725809514522552, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.26162445545196533, 'rewards/total_composite/std': 0.3064093589782715, 'reward': 0.26162445545196533, 'reward_std': 0.3064093589782715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15142585337162018, 'sampling/sampling_logp_difference/max': 1.40028715133667, 'sampling/importance_sampling_ratio/min': 0.24652616679668427, 'sampling/importance_sampling_ratio/mean': 1.0279133319854736, 'sampling/importance_sampling_ratio/max': 1.9877774715423584, 'entropy': 1.3119773119688034, 'clip_ratio/low_mean': 0.057565790601074696, 'clip_ratio/low_min': 0.057565790601074696, 'clip_ratio/high_mean': 0.08418062329292297, 'clip_ratio/high_max': 0.08418062329292297, 'clip_ratio/region_mean': 0.14174641389399767, 'reward_total_mean': 0.26162445545196533, 'reward_meter_mean': 0.5020642280578613, 'reward_meter_std': 0.48159053921699524, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8541666269302368, 'reward_lexical_plausibility_std': 0.3500283658504486, 'reward_repeat_penalty_mean': 0.7149804830551147, 'reward_repeat_penalty_std': 0.08950236439704895, 'reward_repeat_floor_mean': 0.9785131216049194, 'reward_repeat_floor_std': 0.016445640474557877, 'reward_near_duplicate_penalty_mean': 0.9682923555374146, 'reward_near_duplicate_penalty_std': 0.07725809514522552, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.26162445545196533, 'reward_total_composite_std': 0.3064093589782715, 'epoch': 0.07}

 38%|███▊      | 1250/3300 [4:46:10<5:36:15,  9.84s/it]INFO 04-12 20:44:37 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:06<00:37,  3.42s/it]

 23%|██▎       | 3/13 [00:18<01:08,  6.80s/it]

 31%|███       | 4/13 [00:24<01:00,  6.72s/it]

 38%|███▊      | 5/13 [00:32<00:55,  6.88s/it]

 46%|████▌     | 6/13 [00:39<00:48,  6.89s/it]

 54%|█████▍    | 7/13 [00:46<00:41,  6.94s/it]

 62%|██████▏   | 8/13 [00:53<00:35,  7.04s/it]

 69%|██████▉   | 9/13 [01:03<00:31,  7.99s/it]

 77%|███████▋  | 10/13 [01:09<00:22,  7.51s/it]

 85%|████████▍ | 11/13 [01:21<00:17,  8.65s/it]

 92%|█████████▏| 12/13 [01:33<00:09,  9.83s/it]

100%|██████████| 13/13 [01:45<00:00, 10.58s/it]2026-04-12 20:46:30,724 | INFO | train_grpo_train | metrics_logged mode=eval step=1250

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 113.8564, 'eval_samples_per_second': 0.913, 'eval_steps_per_second': 0.114, 'eval_num_tokens': 2221926.0, 'eval_completions/mean_length': 97.08653846153847, 'eval_completions/min_length': 31.46153846153846, 'eval_completions/max_length': 261.15384615384613, 'eval_completions/clipped_ratio': 0.028846153846153848, 'eval_completions/mean_terminated_length': 84.32829724825345, 'eval_completions/min_terminated_length': 31.46153846153846, 'eval_completions/max_terminated_length': 170.76923076923077, 'eval_rewards/meter/mean': 0.5966296516931974, 'eval_rewards/meter/std': 0.3781478554010391, 'eval_rewards/count_adherence/mean': 0.6978326898354751, 'eval_rewards/count_adherence/std': 0.17388536379887506, 'eval_rewards/arabic_clean/mean': 0.9230769230769231, 'eval_rewards/arabic_clean/std': 0.1800025884921734, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.909349799156189, 'eval_rewards/count_floor/std': 0.052165612005270444, 'eval_rewards/lexical_plausibility/mean': 0.9841088423362145, 'eval_rewards/lexical_plausibility/std': 0.04021226242184639, 'eval_rewards/judge_quality/mean': 0.4044230649104485, 'eval_rewards/judge_quality/std': 0.14548370414055312, 'eval_rewards/repeat_penalty/mean': 0.8389853605857263, 'eval_rewards/repeat_penalty/std': 0.1801053428879151, 'eval_rewards/repeat_floor/mean': 0.9808639471347516, 'eval_rewards/repeat_floor/std': 0.02278821413906721, 'eval_rewards/near_duplicate_penalty/mean': 0.9447789742396429, 'eval_rewards/near_duplicate_penalty/std': 0.05819114068379769, 'eval_rewards/opening_diversity/mean': 0.9653127101751474, 'eval_rewards/opening_diversity/std': 0.093134797966251, 'eval_rewards/distinct_2/mean': 0.9025370341080886, 'eval_rewards/distinct_2/std': 0.12902091615475142, 'eval_rewards/total_composite/mean': 0.20754941839438218, 'eval_rewards/total_composite/std': 0.15265308664395258, 'eval_reward': 0.20754941839438218, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07792179229167792, 'eval_sampling/sampling_logp_difference/max': 1.0831471223097582, 'eval_sampling/importance_sampling_ratio/min': 0.3470337184575888, 'eval_sampling/importance_sampling_ratio/mean': 1.0167787350141084, 'eval_sampling/importance_sampling_ratio/max': 1.570376992225647, 'eval_entropy': 0.8809040509737455, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.20754941839438218, 'eval_reward_meter_mean': 0.5966296516931974, 'eval_reward_meter_std': 0.3781478554010391, 'eval_reward_count_adherence_mean': 0.6978326898354751, 'eval_reward_count_adherence_std': 0.17388536379887506, 'eval_reward_arabic_clean_mean': 0.9230769230769231, 'eval_reward_arabic_clean_std': 0.1800025884921734, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.909349799156189, 'eval_reward_count_floor_std': 0.052165612005270444, 'eval_reward_lexical_plausibility_mean': 0.9841088423362145, 'eval_reward_lexical_plausibility_std': 0.04021226242184639, 'eval_reward_repeat_penalty_mean': 0.8389853605857263, 'eval_reward_repeat_penalty_std': 0.1801053428879151, 'eval_reward_repeat_floor_mean': 0.9808639471347516, 'eval_reward_repeat_floor_std': 0.02278821413906721, 'eval_reward_near_duplicate_penalty_mean': 0.9447789742396429, 'eval_reward_near_duplicate_penalty_std': 0.05819114068379769, 'eval_reward_opening_diversity_mean': 0.9653127101751474, 'eval_reward_opening_diversity_std': 0.093134797966251, 'eval_reward_distinct_2_mean': 0.9025370341080886, 'eval_reward_distinct_2_std': 0.12902091615475142, 'eval_reward_judge_quality_mean': 0.4044230649104485, 'eval_reward_judge_quality_std': 0.14548370414055312, 'eval_reward_total_composite_mean': 0.20754941839438218, 'eval_reward_total_composite_std': 0.15265308664395258, 'epoch': 0.07}

 38%|███▊      | 1250/3300 [4:48:04<5:36:15,  9.84s/it]

100%|██████████| 13/13 [01:46<00:00, 10.58s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 38%|███▊      | 1251/3300 [4:48:18<25:47:28, 45.31s/it]2026-04-12 20:46:44,835 | INFO | train_grpo_train | metrics_logged mode=train step=1251

                                                        
{'loss': 0.034, 'grad_norm': 13.278849601745605, 'learning_rate': 6.212121212121213e-06, 'num_tokens': 2223525.0, 'completions/mean_length': 44.875, 'completions/min_length': 39.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7909760475158691, 'rewards/meter/std': 0.3239612877368927, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9940759539604187, 'rewards/lexical_plausibility/std': 0.016755690798163414, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.20364098250865936, 'rewards/repeat_penalty/mean': 0.99138343334198, 'rewards/repeat_penalty/std': 0.019528528675436974, 'rewards/repeat_floor/mean': 0.9987075328826904, 'rewards/repeat_floor/std': 0.00292928796261549, 'rewards/near_duplicate_penalty/mean': 0.99138343334198, 'rewards/near_duplicate_penalty/std': 0.019528528675436974, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.42991146445274353, 'rewards/total_composite/std': 0.26296699047088623, 'reward': 0.42991146445274353, 'reward_std': 0.26296699047088623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19633692502975464, 'sampling/sampling_logp_difference/max': 3.283474922180176, 'sampling/importance_sampling_ratio/min': 0.03749772906303406, 'sampling/importance_sampling_ratio/mean': 1.0168532133102417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.360746517777443, 'clip_ratio/low_mean': 0.0833424273878336, 'clip_ratio/low_min': 0.0833424273878336, 'clip_ratio/high_mean': 0.08187267184257507, 'clip_ratio/high_max': 0.08187267184257507, 'clip_ratio/region_mean': 0.16521509923040867, 'reward_total_mean': 0.42991146445274353, 'reward_meter_mean': 0.7909760475158691, 'reward_meter_std': 0.3239612877368927, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9940759539604187, 'reward_lexical_plausibility_std': 0.016755690798163414, 'reward_repeat_penalty_mean': 0.99138343334198, 'reward_repeat_penalty_std': 0.019528528675436974, 'reward_repeat_floor_mean': 0.9987075328826904, 'reward_repeat_floor_std': 0.00292928796261549, 'reward_near_duplicate_penalty_mean': 0.99138343334198, 'reward_near_duplicate_penalty_std': 0.019528528675436974, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.20364098250865936, 'reward_total_composite_mean': 0.42991146445274353, 'reward_total_composite_std': 0.26296699047088623, 'epoch': 0.07}

 38%|███▊      | 1251/3300 [4:48:18<25:47:28, 45.31s/it]INFO 04-12 20:46:45 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1252/3300 [4:48:36<21:03:20, 37.01s/it]2026-04-12 20:47:02,538 | INFO | train_grpo_train | metrics_logged mode=train step=1252

                                                        
{'loss': -0.0322, 'grad_norm': 5.458512783050537, 'learning_rate': 6.209090909090909e-06, 'num_tokens': 2225286.0, 'completions/mean_length': 106.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.142860412597656, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6572843194007874, 'rewards/meter/std': 0.39681798219680786, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.10690449923276901, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8799999952316284, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.9629882574081421, 'rewards/lexical_plausibility/std': 0.10468495637178421, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.7683491110801697, 'rewards/repeat_penalty/std': 0.22019585967063904, 'rewards/repeat_floor/mean': 0.975979208946228, 'rewards/repeat_floor/std': 0.024363616481423378, 'rewards/near_duplicate_penalty/mean': 0.9293384552001953, 'rewards/near_duplicate_penalty/std': 0.06565709412097931, 'rewards/opening_diversity/mean': 0.9116071462631226, 'rewards/opening_diversity/std': 0.1401757299900055, 'rewards/distinct_2/mean': 0.9098001718521118, 'rewards/distinct_2/std': 0.08037779480218887, 'rewards/total_composite/mean': 0.13883529603481293, 'rewards/total_composite/std': 0.11671510338783264, 'reward': 0.13883529603481293, 'reward_std': 0.11671510338783264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16912361979484558, 'sampling/sampling_logp_difference/max': 2.014023780822754, 'sampling/importance_sampling_ratio/min': 0.13345061242580414, 'sampling/importance_sampling_ratio/mean': 0.9999454617500305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7984330877661705, 'clip_ratio/low_mean': 0.061143194790929556, 'clip_ratio/low_min': 0.061143194790929556, 'clip_ratio/high_mean': 0.06245866417884827, 'clip_ratio/high_max': 0.06245866417884827, 'clip_ratio/region_mean': 0.12360185896977782, 'reward_total_mean': 0.13883529603481293, 'reward_meter_mean': 0.6572843194007874, 'reward_meter_std': 0.39681798219680786, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.10690449923276901, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8799999952316284, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.9629882574081421, 'reward_lexical_plausibility_std': 0.10468495637178421, 'reward_repeat_penalty_mean': 0.7683491110801697, 'reward_repeat_penalty_std': 0.22019585967063904, 'reward_repeat_floor_mean': 0.975979208946228, 'reward_repeat_floor_std': 0.024363616481423378, 'reward_near_duplicate_penalty_mean': 0.9293384552001953, 'reward_near_duplicate_penalty_std': 0.06565709412097931, 'reward_opening_diversity_mean': 0.9116071462631226, 'reward_opening_diversity_std': 0.1401757299900055, 'reward_distinct_2_mean': 0.9098001718521118, 'reward_distinct_2_std': 0.08037779480218887, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.13883529603481293, 'reward_total_composite_std': 0.11671510338783264, 'epoch': 0.07}

 38%|███▊      | 1252/3300 [4:48:36<21:03:20, 37.01s/it]INFO 04-12 20:47:02 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1253/3300 [4:48:51<17:18:16, 30.43s/it]2026-04-12 20:47:17,574 | INFO | train_grpo_train | metrics_logged mode=train step=1253

                                                        
{'loss': -0.109, 'grad_norm': 4.875716686248779, 'learning_rate': 6.206060606060606e-06, 'num_tokens': 2227751.0, 'completions/mean_length': 160.125, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 109.85714721679688, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.6825902462005615, 'rewards/meter/std': 0.26857298612594604, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.046290989965200424, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.01388731598854065, 'rewards/lexical_plausibility/mean': 0.9683085083961487, 'rewards/lexical_plausibility/std': 0.08963703364133835, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.7191486954689026, 'rewards/repeat_penalty/std': 0.1966438591480255, 'rewards/repeat_floor/mean': 0.968305230140686, 'rewards/repeat_floor/std': 0.02208957076072693, 'rewards/near_duplicate_penalty/mean': 0.9228141903877258, 'rewards/near_duplicate_penalty/std': 0.04181179031729698, 'rewards/opening_diversity/mean': 0.9714972972869873, 'rewards/opening_diversity/std': 0.06735674291849136, 'rewards/distinct_2/mean': 0.7954806089401245, 'rewards/distinct_2/std': 0.15416808426380157, 'rewards/total_composite/mean': 0.19127149879932404, 'rewards/total_composite/std': 0.11537996679544449, 'reward': 0.19127149879932404, 'reward_std': 0.11537995934486389, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14006738364696503, 'sampling/sampling_logp_difference/max': 2.62046480178833, 'sampling/importance_sampling_ratio/min': 0.07276903092861176, 'sampling/importance_sampling_ratio/mean': 1.012315273284912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7036051526665688, 'clip_ratio/low_mean': 0.026867925189435482, 'clip_ratio/low_min': 0.026867925189435482, 'clip_ratio/high_mean': 0.10295883566141129, 'clip_ratio/high_max': 0.10295883566141129, 'clip_ratio/region_mean': 0.12982676085084677, 'reward_total_mean': 0.19127149879932404, 'reward_meter_mean': 0.6825902462005615, 'reward_meter_std': 0.26857298612594604, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.046290989965200424, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.01388731598854065, 'reward_lexical_plausibility_mean': 0.9683085083961487, 'reward_lexical_plausibility_std': 0.08963703364133835, 'reward_repeat_penalty_mean': 0.7191486954689026, 'reward_repeat_penalty_std': 0.1966438591480255, 'reward_repeat_floor_mean': 0.968305230140686, 'reward_repeat_floor_std': 0.02208957076072693, 'reward_near_duplicate_penalty_mean': 0.9228141903877258, 'reward_near_duplicate_penalty_std': 0.04181179031729698, 'reward_opening_diversity_mean': 0.9714972972869873, 'reward_opening_diversity_std': 0.06735674291849136, 'reward_distinct_2_mean': 0.7954806089401245, 'reward_distinct_2_std': 0.15416808426380157, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.19127149879932404, 'reward_total_composite_std': 0.11537996679544449, 'epoch': 0.07}

 38%|███▊      | 1253/3300 [4:48:51<17:18:16, 30.43s/it]INFO 04-12 20:47:17 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1254/3300 [4:48:59<13:35:28, 23.91s/it]2026-04-12 20:47:26,287 | INFO | train_grpo_train | metrics_logged mode=train step=1254

                                                        
{'loss': -0.0029, 'grad_norm': 17.301715850830078, 'learning_rate': 6.203030303030304e-06, 'num_tokens': 2229346.0, 'completions/mean_length': 34.375, 'completions/min_length': 29.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6378952860832214, 'rewards/meter/std': 0.48884040117263794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.20310096442699432, 'rewards/repeat_penalty/mean': 0.9229481220245361, 'rewards/repeat_penalty/std': 0.12945304811000824, 'rewards/repeat_floor/mean': 0.98879075050354, 'rewards/repeat_floor/std': 0.018455978482961655, 'rewards/near_duplicate_penalty/mean': 0.9368101358413696, 'rewards/near_duplicate_penalty/std': 0.09152856469154358, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.2965428829193115, 'rewards/total_composite/std': 0.25965338945388794, 'reward': 0.2965428829193115, 'reward_std': 0.25965335965156555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16391807794570923, 'sampling/sampling_logp_difference/max': 1.4643560647964478, 'sampling/importance_sampling_ratio/min': 0.2312268316745758, 'sampling/importance_sampling_ratio/mean': 1.0039533376693726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8995163664221764, 'clip_ratio/low_mean': 0.04930837359279394, 'clip_ratio/low_min': 0.04930837359279394, 'clip_ratio/high_mean': 0.1144637931138277, 'clip_ratio/high_max': 0.1144637931138277, 'clip_ratio/region_mean': 0.16377216670662165, 'reward_total_mean': 0.2965428829193115, 'reward_meter_mean': 0.6378952860832214, 'reward_meter_std': 0.48884040117263794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9229481220245361, 'reward_repeat_penalty_std': 0.12945304811000824, 'reward_repeat_floor_mean': 0.98879075050354, 'reward_repeat_floor_std': 0.018455978482961655, 'reward_near_duplicate_penalty_mean': 0.9368101358413696, 'reward_near_duplicate_penalty_std': 0.09152856469154358, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.20310096442699432, 'reward_total_composite_mean': 0.2965428829193115, 'reward_total_composite_std': 0.25965338945388794, 'epoch': 0.07}

 38%|███▊      | 1254/3300 [4:49:00<13:35:28, 23.91s/it]INFO 04-12 20:47:26 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1255/3300 [4:49:13<11:47:18, 20.75s/it]2026-04-12 20:47:39,629 | INFO | train_grpo_train | metrics_logged mode=train step=1255

                                                        
{'loss': -0.0326, 'grad_norm': 4.500360012054443, 'learning_rate': 6.200000000000001e-06, 'num_tokens': 2230788.0, 'completions/mean_length': 81.25, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.71428680419922, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7399331331253052, 'rewards/meter/std': 0.35489702224731445, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8943848609924316, 'rewards/lexical_plausibility/std': 0.17779874801635742, 'rewards/judge_quality/mean': 0.4712500274181366, 'rewards/judge_quality/std': 0.40484344959259033, 'rewards/repeat_penalty/mean': 0.716228723526001, 'rewards/repeat_penalty/std': 0.09551949054002762, 'rewards/repeat_floor/mean': 0.9786228537559509, 'rewards/repeat_floor/std': 0.01803729124367237, 'rewards/near_duplicate_penalty/mean': 0.9679751992225647, 'rewards/near_duplicate_penalty/std': 0.0905797928571701, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.41331934928894043, 'rewards/total_composite/std': 0.38725849986076355, 'reward': 0.41331934928894043, 'reward_std': 0.38725847005844116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17487557232379913, 'sampling/sampling_logp_difference/max': 2.0627710819244385, 'sampling/importance_sampling_ratio/min': 0.25391247868537903, 'sampling/importance_sampling_ratio/mean': 1.0444165468215942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1215150207281113, 'clip_ratio/low_mean': 0.049753288738429546, 'clip_ratio/low_min': 0.049753288738429546, 'clip_ratio/high_mean': 0.09697420801967382, 'clip_ratio/high_max': 0.09697420801967382, 'clip_ratio/region_mean': 0.14672749675810337, 'reward_total_mean': 0.41331934928894043, 'reward_meter_mean': 0.7399331331253052, 'reward_meter_std': 0.35489702224731445, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8943848609924316, 'reward_lexical_plausibility_std': 0.17779874801635742, 'reward_repeat_penalty_mean': 0.716228723526001, 'reward_repeat_penalty_std': 0.09551949054002762, 'reward_repeat_floor_mean': 0.9786228537559509, 'reward_repeat_floor_std': 0.01803729124367237, 'reward_near_duplicate_penalty_mean': 0.9679751992225647, 'reward_near_duplicate_penalty_std': 0.0905797928571701, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.4712500274181366, 'reward_judge_quality_std': 0.40484344959259033, 'reward_total_composite_mean': 0.41331934928894043, 'reward_total_composite_std': 0.38725849986076355, 'epoch': 0.07}

 38%|███▊      | 1255/3300 [4:49:13<11:47:18, 20.75s/it]INFO 04-12 20:47:39 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1256/3300 [4:49:23<9:54:02, 17.44s/it] 2026-04-12 20:47:49,359 | INFO | train_grpo_train | metrics_logged mode=train step=1256

                                                       
{'loss': 0.0226, 'grad_norm': 16.340438842773438, 'learning_rate': 6.196969696969698e-06, 'num_tokens': 2232139.0, 'completions/mean_length': 18.875, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.9844534397125244, 'rewards/meter/std': 0.009304544888436794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9360517263412476, 'rewards/lexical_plausibility/std': 0.12239138036966324, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.7193916440010071, 'rewards/repeat_penalty/std': 0.08657350391149521, 'rewards/repeat_floor/mean': 0.979559063911438, 'rewards/repeat_floor/std': 0.015389395877718925, 'rewards/near_duplicate_penalty/mean': 0.976547360420227, 'rewards/near_duplicate_penalty/std': 0.06633399426937103, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.6836011409759521, 'rewards/total_composite/std': 0.3004305064678192, 'reward': 0.6836011409759521, 'reward_std': 0.3004305064678192, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1253630816936493, 'sampling/sampling_logp_difference/max': 1.2026491165161133, 'sampling/importance_sampling_ratio/min': 0.3003973662853241, 'sampling/importance_sampling_ratio/mean': 0.9946272969245911, 'sampling/importance_sampling_ratio/max': 1.794089913368225, 'entropy': 0.7536482810974121, 'clip_ratio/low_mean': 0.03289473615586758, 'clip_ratio/low_min': 0.03289473615586758, 'clip_ratio/high_mean': 0.09863199945539236, 'clip_ratio/high_max': 0.09863199945539236, 'clip_ratio/region_mean': 0.13152673561125994, 'reward_total_mean': 0.6836011409759521, 'reward_meter_mean': 0.9844534397125244, 'reward_meter_std': 0.009304544888436794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9360517263412476, 'reward_lexical_plausibility_std': 0.12239138036966324, 'reward_repeat_penalty_mean': 0.7193916440010071, 'reward_repeat_penalty_std': 0.08657350391149521, 'reward_repeat_floor_mean': 0.979559063911438, 'reward_repeat_floor_std': 0.015389395877718925, 'reward_near_duplicate_penalty_mean': 0.976547360420227, 'reward_near_duplicate_penalty_std': 0.06633399426937103, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.6836011409759521, 'reward_total_composite_std': 0.3004305064678192, 'epoch': 0.07}

 38%|███▊      | 1256/3300 [4:49:23<9:54:02, 17.44s/it]INFO 04-12 20:47:49 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1257/3300 [4:49:37<9:21:34, 16.49s/it]2026-04-12 20:48:03,622 | INFO | train_grpo_train | metrics_logged mode=train step=1257

                                                       
{'loss': -0.0412, 'grad_norm': 4.738894939422607, 'learning_rate': 6.1939393939393944e-06, 'num_tokens': 2233588.0, 'completions/mean_length': 91.125, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.000001907348633, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.48015978932380676, 'rewards/meter/std': 0.3897162675857544, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9451248645782471, 'rewards/lexical_plausibility/std': 0.15521031618118286, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9197390079498291, 'rewards/repeat_penalty/std': 0.10586664080619812, 'rewards/repeat_floor/mean': 0.9886679649353027, 'rewards/repeat_floor/std': 0.014973809011280537, 'rewards/near_duplicate_penalty/mean': 0.9365989565849304, 'rewards/near_duplicate_penalty/std': 0.08111708611249924, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9797570705413818, 'rewards/distinct_2/std': 0.03748259320855141, 'rewards/total_composite/mean': 0.15227575600147247, 'rewards/total_composite/std': 0.12180215120315552, 'reward': 0.15227575600147247, 'reward_std': 0.12180214375257492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15141573548316956, 'sampling/sampling_logp_difference/max': 1.4547791481018066, 'sampling/importance_sampling_ratio/min': 0.23345191776752472, 'sampling/importance_sampling_ratio/mean': 1.0041520595550537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.003646194934845, 'clip_ratio/low_mean': 0.07050189469009638, 'clip_ratio/low_min': 0.07050189469009638, 'clip_ratio/high_mean': 0.024245689623057842, 'clip_ratio/high_max': 0.024245689623057842, 'clip_ratio/region_mean': 0.09474758431315422, 'reward_total_mean': 0.15227575600147247, 'reward_meter_mean': 0.48015978932380676, 'reward_meter_std': 0.3897162675857544, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9451248645782471, 'reward_lexical_plausibility_std': 0.15521031618118286, 'reward_repeat_penalty_mean': 0.9197390079498291, 'reward_repeat_penalty_std': 0.10586664080619812, 'reward_repeat_floor_mean': 0.9886679649353027, 'reward_repeat_floor_std': 0.014973809011280537, 'reward_near_duplicate_penalty_mean': 0.9365989565849304, 'reward_near_duplicate_penalty_std': 0.08111708611249924, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9797570705413818, 'reward_distinct_2_std': 0.03748259320855141, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.15227575600147247, 'reward_total_composite_std': 0.12180215120315552, 'epoch': 0.07}

 38%|███▊      | 1257/3300 [4:49:37<9:21:34, 16.49s/it]INFO 04-12 20:48:03 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1258/3300 [4:49:47<8:15:53, 14.57s/it]2026-04-12 20:48:13,709 | INFO | train_grpo_train | metrics_logged mode=train step=1258

                                                       
{'loss': 0.0921, 'grad_norm': 13.316137313842773, 'learning_rate': 6.190909090909092e-06, 'num_tokens': 2235700.0, 'completions/mean_length': 81.0, 'completions/min_length': 67.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.685248613357544, 'rewards/meter/std': 0.30939343571662903, 'rewards/count_adherence/mean': 0.6750000715255737, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9024999737739563, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9204732775688171, 'rewards/repeat_penalty/std': 0.037678152322769165, 'rewards/repeat_floor/mean': 0.9904094934463501, 'rewards/repeat_floor/std': 0.0037440124433487654, 'rewards/near_duplicate_penalty/mean': 0.9617751836776733, 'rewards/near_duplicate_penalty/std': 0.013675026595592499, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9571467638015747, 'rewards/distinct_2/std': 0.03926120698451996, 'rewards/total_composite/mean': 0.2565568685531616, 'rewards/total_composite/std': 0.11464803665876389, 'reward': 0.2565568685531616, 'reward_std': 0.11464803665876389, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14566437900066376, 'sampling/sampling_logp_difference/max': 2.5363893508911133, 'sampling/importance_sampling_ratio/min': 0.07915167510509491, 'sampling/importance_sampling_ratio/mean': 1.0065324306488037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8205195590853691, 'clip_ratio/low_mean': 0.049017183016985655, 'clip_ratio/low_min': 0.049017183016985655, 'clip_ratio/high_mean': 0.06952247768640518, 'clip_ratio/high_max': 0.06952247768640518, 'clip_ratio/region_mean': 0.11853966070339084, 'reward_total_mean': 0.2565568685531616, 'reward_meter_mean': 0.685248613357544, 'reward_meter_std': 0.30939343571662903, 'reward_count_adherence_mean': 0.6750000715255737, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9024999737739563, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9204732775688171, 'reward_repeat_penalty_std': 0.037678152322769165, 'reward_repeat_floor_mean': 0.9904094934463501, 'reward_repeat_floor_std': 0.0037440124433487654, 'reward_near_duplicate_penalty_mean': 0.9617751836776733, 'reward_near_duplicate_penalty_std': 0.013675026595592499, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9571467638015747, 'reward_distinct_2_std': 0.03926120698451996, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.2565568685531616, 'reward_total_composite_std': 0.11464803665876389, 'epoch': 0.07}

 38%|███▊      | 1258/3300 [4:49:47<8:15:53, 14.57s/it]INFO 04-12 20:48:13 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1259/3300 [4:49:58<7:39:21, 13.50s/it]2026-04-12 20:48:24,713 | INFO | train_grpo_train | metrics_logged mode=train step=1259

                                                       
{'loss': 0.0388, 'grad_norm': 16.644067764282227, 'learning_rate': 6.187878787878788e-06, 'num_tokens': 2237502.0, 'completions/mean_length': 47.25, 'completions/min_length': 41.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4338234066963196, 'rewards/meter/std': 0.3957770764827728, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9787946343421936, 'rewards/lexical_plausibility/std': 0.04103745147585869, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9020435810089111, 'rewards/repeat_penalty/std': 0.18300504982471466, 'rewards/repeat_floor/mean': 0.9880900979042053, 'rewards/repeat_floor/std': 0.02130706235766411, 'rewards/near_duplicate_penalty/mean': 0.9589366912841797, 'rewards/near_duplicate_penalty/std': 0.0589282251894474, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9569398164749146, 'rewards/distinct_2/std': 0.09462050348520279, 'rewards/total_composite/mean': 0.19060809910297394, 'rewards/total_composite/std': 0.18203598260879517, 'reward': 0.19060809910297394, 'reward_std': 0.18203599750995636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1600101739168167, 'sampling/sampling_logp_difference/max': 2.0403048992156982, 'sampling/importance_sampling_ratio/min': 0.12998907268047333, 'sampling/importance_sampling_ratio/mean': 1.0351078510284424, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4189226180315018, 'clip_ratio/low_mean': 0.045179912354797125, 'clip_ratio/low_min': 0.045179912354797125, 'clip_ratio/high_mean': 0.0699779293499887, 'clip_ratio/high_max': 0.0699779293499887, 'clip_ratio/region_mean': 0.11515784170478582, 'reward_total_mean': 0.19060809910297394, 'reward_meter_mean': 0.4338234066963196, 'reward_meter_std': 0.3957770764827728, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9787946343421936, 'reward_lexical_plausibility_std': 0.04103745147585869, 'reward_repeat_penalty_mean': 0.9020435810089111, 'reward_repeat_penalty_std': 0.18300504982471466, 'reward_repeat_floor_mean': 0.9880900979042053, 'reward_repeat_floor_std': 0.02130706235766411, 'reward_near_duplicate_penalty_mean': 0.9589366912841797, 'reward_near_duplicate_penalty_std': 0.0589282251894474, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9569398164749146, 'reward_distinct_2_std': 0.09462050348520279, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.19060809910297394, 'reward_total_composite_std': 0.18203598260879517, 'epoch': 0.07}

 38%|███▊      | 1259/3300 [4:49:58<7:39:21, 13.50s/it]INFO 04-12 20:48:24 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1260/3300 [4:50:09<7:18:43, 12.90s/it]2026-04-12 20:48:36,315 | INFO | train_grpo_train | metrics_logged mode=train step=1260

                                                       
{'loss': 0.0373, 'grad_norm': 51.46892547607422, 'learning_rate': 6.184848484848485e-06, 'num_tokens': 2239677.0, 'completions/mean_length': 71.875, 'completions/min_length': 64.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.683354914188385, 'rewards/meter/std': 0.3178428113460541, 'rewards/count_adherence/mean': 0.6428571939468384, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8928571343421936, 'rewards/count_floor/std': 0.022908121347427368, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8680921792984009, 'rewards/repeat_penalty/std': 0.1462443768978119, 'rewards/repeat_floor/mean': 0.9849512577056885, 'rewards/repeat_floor/std': 0.015877826139330864, 'rewards/near_duplicate_penalty/mean': 0.9536528587341309, 'rewards/near_duplicate_penalty/std': 0.03723161295056343, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9308700561523438, 'rewards/distinct_2/std': 0.0646088495850563, 'rewards/total_composite/mean': 0.2129134237766266, 'rewards/total_composite/std': 0.1447014957666397, 'reward': 0.2129134237766266, 'reward_std': 0.1447015106678009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1904776692390442, 'sampling/sampling_logp_difference/max': 3.6638755798339844, 'sampling/importance_sampling_ratio/min': 0.02563297748565674, 'sampling/importance_sampling_ratio/mean': 1.0279250144958496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2319042310118675, 'clip_ratio/low_mean': 0.10676970705389977, 'clip_ratio/low_min': 0.10676970705389977, 'clip_ratio/high_mean': 0.06631868332624435, 'clip_ratio/high_max': 0.06631868332624435, 'clip_ratio/region_mean': 0.17308839038014412, 'reward_total_mean': 0.2129134237766266, 'reward_meter_mean': 0.683354914188385, 'reward_meter_std': 0.3178428113460541, 'reward_count_adherence_mean': 0.6428571939468384, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8928571343421936, 'reward_count_floor_std': 0.022908121347427368, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8680921792984009, 'reward_repeat_penalty_std': 0.1462443768978119, 'reward_repeat_floor_mean': 0.9849512577056885, 'reward_repeat_floor_std': 0.015877826139330864, 'reward_near_duplicate_penalty_mean': 0.9536528587341309, 'reward_near_duplicate_penalty_std': 0.03723161295056343, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9308700561523438, 'reward_distinct_2_std': 0.0646088495850563, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.2129134237766266, 'reward_total_composite_std': 0.1447014957666397, 'epoch': 0.07}

 38%|███▊      | 1260/3300 [4:50:10<7:18:43, 12.90s/it]INFO 04-12 20:48:36 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1261/3300 [4:50:20<6:50:18, 12.07s/it]2026-04-12 20:48:46,378 | INFO | train_grpo_train | metrics_logged mode=train step=1261

                                                       
{'loss': 0.0451, 'grad_norm': 23.265932083129883, 'learning_rate': 6.181818181818182e-06, 'num_tokens': 2241033.0, 'completions/mean_length': 17.5, 'completions/min_length': 15.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.5, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.7343837022781372, 'rewards/meter/std': 0.36674001812934875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9420955777168274, 'rewards/lexical_plausibility/std': 0.11673673987388611, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.7020621299743652, 'rewards/repeat_penalty/std': 0.09208667278289795, 'rewards/repeat_floor/mean': 0.9754124283790588, 'rewards/repeat_floor/std': 0.018417349085211754, 'rewards/near_duplicate_penalty/mean': 0.9360828399658203, 'rewards/near_duplicate_penalty/std': 0.12278221547603607, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16880950331687927, 'rewards/total_composite/std': 0.09944184869527817, 'reward': 0.16880950331687927, 'reward_std': 0.09944184869527817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14555129408836365, 'sampling/sampling_logp_difference/max': 1.1002131700515747, 'sampling/importance_sampling_ratio/min': 0.33280012011528015, 'sampling/importance_sampling_ratio/mean': 1.0176080465316772, 'sampling/importance_sampling_ratio/max': 1.652567744255066, 'entropy': 1.1419166252017021, 'clip_ratio/low_mean': 0.054502063896507025, 'clip_ratio/low_min': 0.054502063896507025, 'clip_ratio/high_mean': 0.0564236119389534, 'clip_ratio/high_max': 0.0564236119389534, 'clip_ratio/region_mean': 0.11092567583546042, 'reward_total_mean': 0.16880950331687927, 'reward_meter_mean': 0.7343837022781372, 'reward_meter_std': 0.36674001812934875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9420955777168274, 'reward_lexical_plausibility_std': 0.11673673987388611, 'reward_repeat_penalty_mean': 0.7020621299743652, 'reward_repeat_penalty_std': 0.09208667278289795, 'reward_repeat_floor_mean': 0.9754124283790588, 'reward_repeat_floor_std': 0.018417349085211754, 'reward_near_duplicate_penalty_mean': 0.9360828399658203, 'reward_near_duplicate_penalty_std': 0.12278221547603607, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.16880950331687927, 'reward_total_composite_std': 0.09944184869527817, 'epoch': 0.07}

 38%|███▊      | 1261/3300 [4:50:20<6:50:18, 12.07s/it]INFO 04-12 20:48:46 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1262/3300 [4:50:32<6:54:39, 12.21s/it]2026-04-12 20:48:58,882 | INFO | train_grpo_train | metrics_logged mode=train step=1262

                                                       
{'loss': 0.0699, 'grad_norm': 24.21959114074707, 'learning_rate': 6.17878787878788e-06, 'num_tokens': 2242677.0, 'completions/mean_length': 38.5, 'completions/min_length': 35.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5172942876815796, 'rewards/meter/std': 0.35494574904441833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9384489059448242, 'rewards/repeat_penalty/std': 0.06010139361023903, 'rewards/repeat_floor/mean': 0.9916393756866455, 'rewards/repeat_floor/std': 0.0077572292648255825, 'rewards/near_duplicate_penalty/mean': 0.9550015926361084, 'rewards/near_duplicate_penalty/std': 0.03920619562268257, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9821012020111084, 'rewards/distinct_2/std': 0.03315402567386627, 'rewards/total_composite/mean': 0.1482156366109848, 'rewards/total_composite/std': 0.09594321250915527, 'reward': 0.1482156366109848, 'reward_std': 0.09594321250915527, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17619386315345764, 'sampling/sampling_logp_difference/max': 2.140575408935547, 'sampling/importance_sampling_ratio/min': 0.11758716404438019, 'sampling/importance_sampling_ratio/mean': 1.006761074066162, 'sampling/importance_sampling_ratio/max': 1.964910864830017, 'entropy': 1.307152271270752, 'clip_ratio/low_mean': 0.10503984801471233, 'clip_ratio/low_min': 0.10503984801471233, 'clip_ratio/high_mean': 0.07996433600783348, 'clip_ratio/high_max': 0.07996433600783348, 'clip_ratio/region_mean': 0.18500418402254581, 'reward_total_mean': 0.1482156366109848, 'reward_meter_mean': 0.5172942876815796, 'reward_meter_std': 0.35494574904441833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9384489059448242, 'reward_repeat_penalty_std': 0.06010139361023903, 'reward_repeat_floor_mean': 0.9916393756866455, 'reward_repeat_floor_std': 0.0077572292648255825, 'reward_near_duplicate_penalty_mean': 0.9550015926361084, 'reward_near_duplicate_penalty_std': 0.03920619562268257, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9821012020111084, 'reward_distinct_2_std': 0.03315402567386627, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.1482156366109848, 'reward_total_composite_std': 0.09594321250915527, 'epoch': 0.07}

 38%|███▊      | 1262/3300 [4:50:32<6:54:39, 12.21s/it]INFO 04-12 20:48:59 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1263/3300 [4:50:42<6:26:09, 11.37s/it]2026-04-12 20:49:08,314 | INFO | train_grpo_train | metrics_logged mode=train step=1263

                                                       
{'loss': 0.057, 'grad_norm': 16.291786193847656, 'learning_rate': 6.175757575757576e-06, 'num_tokens': 2244506.0, 'completions/mean_length': 54.625, 'completions/min_length': 38.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9612852334976196, 'rewards/meter/std': 0.06804504990577698, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175493285059929, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.8350206017494202, 'rewards/repeat_penalty/std': 0.10663489997386932, 'rewards/repeat_floor/mean': 0.9820359349250793, 'rewards/repeat_floor/std': 0.010365995578467846, 'rewards/near_duplicate_penalty/mean': 0.9458608627319336, 'rewards/near_duplicate_penalty/std': 0.03843100741505623, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9114644527435303, 'rewards/distinct_2/std': 0.06882385909557343, 'rewards/total_composite/mean': 0.36352118849754333, 'rewards/total_composite/std': 0.05873783677816391, 'reward': 0.36352118849754333, 'reward_std': 0.05873784050345421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16212032735347748, 'sampling/sampling_logp_difference/max': 1.8741812705993652, 'sampling/importance_sampling_ratio/min': 0.15348057448863983, 'sampling/importance_sampling_ratio/mean': 1.0290857553482056, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0166698470711708, 'clip_ratio/low_mean': 0.08290879055857658, 'clip_ratio/low_min': 0.08290879055857658, 'clip_ratio/high_mean': 0.062196952290832996, 'clip_ratio/high_max': 0.062196952290832996, 'clip_ratio/region_mean': 0.14510574284940958, 'reward_total_mean': 0.36352118849754333, 'reward_meter_mean': 0.9612852334976196, 'reward_meter_std': 0.06804504990577698, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175493285059929, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8350206017494202, 'reward_repeat_penalty_std': 0.10663489997386932, 'reward_repeat_floor_mean': 0.9820359349250793, 'reward_repeat_floor_std': 0.010365995578467846, 'reward_near_duplicate_penalty_mean': 0.9458608627319336, 'reward_near_duplicate_penalty_std': 0.03843100741505623, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9114644527435303, 'reward_distinct_2_std': 0.06882385909557343, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.36352118849754333, 'reward_total_composite_std': 0.05873783677816391, 'epoch': 0.07}

 38%|███▊      | 1263/3300 [4:50:42<6:26:09, 11.37s/it]INFO 04-12 20:49:08 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1264/3300 [4:50:52<6:20:38, 11.22s/it]2026-04-12 20:49:19,163 | INFO | train_grpo_train | metrics_logged mode=train step=1264

                                                       
{'loss': 0.0202, 'grad_norm': 15.28254508972168, 'learning_rate': 6.1727272727272735e-06, 'num_tokens': 2245840.0, 'completions/mean_length': 22.75, 'completions/min_length': 21.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.75, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6410651206970215, 'rewards/meter/std': 0.3981606364250183, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9505208134651184, 'rewards/lexical_plausibility/std': 0.07547593116760254, 'rewards/judge_quality/mean': 0.7900000214576721, 'rewards/judge_quality/std': 0.24219238758087158, 'rewards/repeat_penalty/mean': 0.7227855920791626, 'rewards/repeat_penalty/std': 0.0522967204451561, 'rewards/repeat_floor/mean': 0.9795571565628052, 'rewards/repeat_floor/std': 0.01045935321599245, 'rewards/near_duplicate_penalty/mean': 0.9637141227722168, 'rewards/near_duplicate_penalty/std': 0.06972895562648773, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4546905755996704, 'rewards/total_composite/std': 0.3120134472846985, 'reward': 0.4546905755996704, 'reward_std': 0.3120134174823761, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12129475921392441, 'sampling/sampling_logp_difference/max': 1.448277473449707, 'sampling/importance_sampling_ratio/min': 0.2349746972322464, 'sampling/importance_sampling_ratio/mean': 1.0157819986343384, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8585151359438896, 'clip_ratio/low_mean': 0.05714122951030731, 'clip_ratio/low_min': 0.05714122951030731, 'clip_ratio/high_mean': 0.061264822259545326, 'clip_ratio/high_max': 0.061264822259545326, 'clip_ratio/region_mean': 0.11840605176985264, 'reward_total_mean': 0.4546905755996704, 'reward_meter_mean': 0.6410651206970215, 'reward_meter_std': 0.3981606364250183, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9505208134651184, 'reward_lexical_plausibility_std': 0.07547593116760254, 'reward_repeat_penalty_mean': 0.7227855920791626, 'reward_repeat_penalty_std': 0.0522967204451561, 'reward_repeat_floor_mean': 0.9795571565628052, 'reward_repeat_floor_std': 0.01045935321599245, 'reward_near_duplicate_penalty_mean': 0.9637141227722168, 'reward_near_duplicate_penalty_std': 0.06972895562648773, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7900000214576721, 'reward_judge_quality_std': 0.24219238758087158, 'reward_total_composite_mean': 0.4546905755996704, 'reward_total_composite_std': 0.3120134472846985, 'epoch': 0.07}

 38%|███▊      | 1264/3300 [4:50:53<6:20:38, 11.22s/it]INFO 04-12 20:49:19 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1265/3300 [4:51:05<6:34:02, 11.62s/it]2026-04-12 20:49:31,784 | INFO | train_grpo_train | metrics_logged mode=train step=1265

                                                       
{'loss': 0.018, 'grad_norm': 18.336753845214844, 'learning_rate': 6.16969696969697e-06, 'num_tokens': 2247450.0, 'completions/mean_length': 37.25, 'completions/min_length': 34.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7474797964096069, 'rewards/meter/std': 0.32458820939064026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.7601014375686646, 'rewards/repeat_penalty/std': 0.20466993749141693, 'rewards/repeat_floor/mean': 0.9752048850059509, 'rewards/repeat_floor/std': 0.0214733574539423, 'rewards/near_duplicate_penalty/mean': 0.9393946528434753, 'rewards/near_duplicate_penalty/std': 0.04691839963197708, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9088411331176758, 'rewards/distinct_2/std': 0.09400047361850739, 'rewards/total_composite/mean': 0.316353976726532, 'rewards/total_composite/std': 0.13710325956344604, 'reward': 0.316353976726532, 'reward_std': 0.13710327446460724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14864195883274078, 'sampling/sampling_logp_difference/max': 2.556056499481201, 'sampling/importance_sampling_ratio/min': 0.07761019468307495, 'sampling/importance_sampling_ratio/mean': 1.0261150598526, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0611946359276772, 'clip_ratio/low_mean': 0.03016718663275242, 'clip_ratio/low_min': 0.03016718663275242, 'clip_ratio/high_mean': 0.08687753137201071, 'clip_ratio/high_max': 0.08687753137201071, 'clip_ratio/region_mean': 0.11704471800476313, 'reward_total_mean': 0.316353976726532, 'reward_meter_mean': 0.7474797964096069, 'reward_meter_std': 0.32458820939064026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.7601014375686646, 'reward_repeat_penalty_std': 0.20466993749141693, 'reward_repeat_floor_mean': 0.9752048850059509, 'reward_repeat_floor_std': 0.0214733574539423, 'reward_near_duplicate_penalty_mean': 0.9393946528434753, 'reward_near_duplicate_penalty_std': 0.04691839963197708, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9088411331176758, 'reward_distinct_2_std': 0.09400047361850739, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.316353976726532, 'reward_total_composite_std': 0.13710325956344604, 'epoch': 0.07}

 38%|███▊      | 1265/3300 [4:51:05<6:34:02, 11.62s/it]INFO 04-12 20:49:32 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1266/3300 [4:51:20<7:11:28, 12.73s/it]2026-04-12 20:49:47,044 | INFO | train_grpo_train | metrics_logged mode=train step=1266

                                                       
{'loss': -0.0092, 'grad_norm': 13.861244201660156, 'learning_rate': 6.166666666666667e-06, 'num_tokens': 2249670.0, 'completions/mean_length': 62.5, 'completions/min_length': 55.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6740539073944092, 'rewards/meter/std': 0.2800249755382538, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9260563850402832, 'rewards/repeat_penalty/std': 0.052895672619342804, 'rewards/repeat_floor/mean': 0.9907950162887573, 'rewards/repeat_floor/std': 0.005818056873977184, 'rewards/near_duplicate_penalty/mean': 0.9600675106048584, 'rewards/near_duplicate_penalty/std': 0.020493807271122932, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9642767310142517, 'rewards/distinct_2/std': 0.043864890933036804, 'rewards/total_composite/mean': 0.22967740893363953, 'rewards/total_composite/std': 0.10760796070098877, 'reward': 0.22967740893363953, 'reward_std': 0.10760795325040817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16858060657978058, 'sampling/sampling_logp_difference/max': 2.4223482608795166, 'sampling/importance_sampling_ratio/min': 0.08871304988861084, 'sampling/importance_sampling_ratio/mean': 1.0314624309539795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.123613402247429, 'clip_ratio/low_mean': 0.08123089745640755, 'clip_ratio/low_min': 0.08123089745640755, 'clip_ratio/high_mean': 0.080549038015306, 'clip_ratio/high_max': 0.080549038015306, 'clip_ratio/region_mean': 0.16177993547171354, 'reward_total_mean': 0.22967740893363953, 'reward_meter_mean': 0.6740539073944092, 'reward_meter_std': 0.2800249755382538, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9260563850402832, 'reward_repeat_penalty_std': 0.052895672619342804, 'reward_repeat_floor_mean': 0.9907950162887573, 'reward_repeat_floor_std': 0.005818056873977184, 'reward_near_duplicate_penalty_mean': 0.9600675106048584, 'reward_near_duplicate_penalty_std': 0.020493807271122932, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9642767310142517, 'reward_distinct_2_std': 0.043864890933036804, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.22967740893363953, 'reward_total_composite_std': 0.10760796070098877, 'epoch': 0.07}

 38%|███▊      | 1266/3300 [4:51:20<7:11:28, 12.73s/it]INFO 04-12 20:49:47 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1267/3300 [4:51:32<6:58:33, 12.35s/it]2026-04-12 20:49:58,585 | INFO | train_grpo_train | metrics_logged mode=train step=1267

                                                       
{'loss': -0.023, 'grad_norm': 20.50056266784668, 'learning_rate': 6.163636363636364e-06, 'num_tokens': 2250984.0, 'completions/mean_length': 21.25, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9926935434341431, 'rewards/meter/std': 0.012431460432708263, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.3123185336589813, 'rewards/repeat_penalty/mean': 0.7480332255363464, 'rewards/repeat_penalty/std': 0.005562899634242058, 'rewards/repeat_floor/mean': 0.9846066236495972, 'rewards/repeat_floor/std': 0.0011125925229862332, 'rewards/near_duplicate_penalty/mean': 0.9973776340484619, 'rewards/near_duplicate_penalty/std': 0.007417213171720505, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.6952759027481079, 'rewards/total_composite/std': 0.30866101384162903, 'reward': 0.6952759027481079, 'reward_std': 0.30866101384162903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1120215505361557, 'sampling/sampling_logp_difference/max': 0.973543643951416, 'sampling/importance_sampling_ratio/min': 0.37774208188056946, 'sampling/importance_sampling_ratio/mean': 1.0003728866577148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7908177301287651, 'clip_ratio/low_mean': 0.044407895766198635, 'clip_ratio/low_min': 0.044407895766198635, 'clip_ratio/high_mean': 0.05077757500112057, 'clip_ratio/high_max': 0.05077757500112057, 'clip_ratio/region_mean': 0.0951854707673192, 'reward_total_mean': 0.6952759027481079, 'reward_meter_mean': 0.9926935434341431, 'reward_meter_std': 0.012431460432708263, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7480332255363464, 'reward_repeat_penalty_std': 0.005562899634242058, 'reward_repeat_floor_mean': 0.9846066236495972, 'reward_repeat_floor_std': 0.0011125925229862332, 'reward_near_duplicate_penalty_mean': 0.9973776340484619, 'reward_near_duplicate_penalty_std': 0.007417213171720505, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.3123185336589813, 'reward_total_composite_mean': 0.6952759027481079, 'reward_total_composite_std': 0.30866101384162903, 'epoch': 0.07}

 38%|███▊      | 1267/3300 [4:51:32<6:58:33, 12.35s/it]INFO 04-12 20:49:58 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1268/3300 [4:51:41<6:23:07, 11.31s/it]2026-04-12 20:50:07,409 | INFO | train_grpo_train | metrics_logged mode=train step=1268

                                                       
{'loss': 0.0299, 'grad_norm': 12.037437438964844, 'learning_rate': 6.160606060606062e-06, 'num_tokens': 2252548.0, 'completions/mean_length': 46.5, 'completions/min_length': 42.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9866486191749573, 'rewards/meter/std': 0.005395363084971905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8592847585678101, 'rewards/repeat_penalty/std': 0.14945675432682037, 'rewards/repeat_floor/mean': 0.9841450452804565, 'rewards/repeat_floor/std': 0.015937577933073044, 'rewards/near_duplicate_penalty/mean': 0.9527406096458435, 'rewards/near_duplicate_penalty/std': 0.041233573108911514, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9234330654144287, 'rewards/distinct_2/std': 0.07231061160564423, 'rewards/total_composite/mean': 0.40093937516212463, 'rewards/total_composite/std': 0.05397419258952141, 'reward': 0.40093937516212463, 'reward_std': 0.05397419631481171, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1417992264032364, 'sampling/sampling_logp_difference/max': 1.3404064178466797, 'sampling/importance_sampling_ratio/min': 0.2617392838001251, 'sampling/importance_sampling_ratio/mean': 1.014907956123352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1005871817469597, 'clip_ratio/low_mean': 0.045286303386092186, 'clip_ratio/low_min': 0.045286303386092186, 'clip_ratio/high_mean': 0.09812999051064253, 'clip_ratio/high_max': 0.09812999051064253, 'clip_ratio/region_mean': 0.14341629389673471, 'reward_total_mean': 0.40093937516212463, 'reward_meter_mean': 0.9866486191749573, 'reward_meter_std': 0.005395363084971905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8592847585678101, 'reward_repeat_penalty_std': 0.14945675432682037, 'reward_repeat_floor_mean': 0.9841450452804565, 'reward_repeat_floor_std': 0.015937577933073044, 'reward_near_duplicate_penalty_mean': 0.9527406096458435, 'reward_near_duplicate_penalty_std': 0.041233573108911514, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9234330654144287, 'reward_distinct_2_std': 0.07231061160564423, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.40093937516212463, 'reward_total_composite_std': 0.05397419258952141, 'epoch': 0.07}

 38%|███▊      | 1268/3300 [4:51:41<6:23:07, 11.31s/it]INFO 04-12 20:50:07 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1269/3300 [4:51:49<5:54:54, 10.48s/it]2026-04-12 20:50:16,191 | INFO | train_grpo_train | metrics_logged mode=train step=1269

                                                       
{'loss': 0.0632, 'grad_norm': 15.516514778137207, 'learning_rate': 6.157575757575758e-06, 'num_tokens': 2254052.0, 'completions/mean_length': 32.0, 'completions/min_length': 28.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.890855073928833, 'rewards/meter/std': 0.15631577372550964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5799999833106995, 'rewards/judge_quality/std': 0.2919393479824066, 'rewards/repeat_penalty/mean': 0.9839155673980713, 'rewards/repeat_penalty/std': 0.01513652689754963, 'rewards/repeat_floor/mean': 0.9975873231887817, 'rewards/repeat_floor/std': 0.0022704889997839928, 'rewards/near_duplicate_penalty/mean': 0.9839155673980713, 'rewards/near_duplicate_penalty/std': 0.01513652689754963, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5272724628448486, 'rewards/total_composite/std': 0.29654330015182495, 'reward': 0.5272724628448486, 'reward_std': 0.29654330015182495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.132646843791008, 'sampling/sampling_logp_difference/max': 2.2651548385620117, 'sampling/importance_sampling_ratio/min': 0.10381396114826202, 'sampling/importance_sampling_ratio/mean': 0.9975581169128418, 'sampling/importance_sampling_ratio/max': 1.9516104459762573, 'entropy': 0.5118381641805172, 'clip_ratio/low_mean': 0.06889673369005322, 'clip_ratio/low_min': 0.06889673369005322, 'clip_ratio/high_mean': 0.04442902561277151, 'clip_ratio/high_max': 0.04442902561277151, 'clip_ratio/region_mean': 0.11332575930282474, 'reward_total_mean': 0.5272724628448486, 'reward_meter_mean': 0.890855073928833, 'reward_meter_std': 0.15631577372550964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9839155673980713, 'reward_repeat_penalty_std': 0.01513652689754963, 'reward_repeat_floor_mean': 0.9975873231887817, 'reward_repeat_floor_std': 0.0022704889997839928, 'reward_near_duplicate_penalty_mean': 0.9839155673980713, 'reward_near_duplicate_penalty_std': 0.01513652689754963, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5799999833106995, 'reward_judge_quality_std': 0.2919393479824066, 'reward_total_composite_mean': 0.5272724628448486, 'reward_total_composite_std': 0.29654330015182495, 'epoch': 0.07}

 38%|███▊      | 1269/3300 [4:51:50<5:54:54, 10.48s/it]INFO 04-12 20:50:16 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1270/3300 [4:52:00<6:01:56, 10.70s/it]2026-04-12 20:50:27,194 | INFO | train_grpo_train | metrics_logged mode=train step=1270

                                                       
{'loss': 0.0658, 'grad_norm': 19.7850399017334, 'learning_rate': 6.154545454545455e-06, 'num_tokens': 2255664.0, 'completions/mean_length': 21.5, 'completions/min_length': 19.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.7712091207504272, 'rewards/meter/std': 0.4052234888076782, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.984375, 'rewards/lexical_plausibility/std': 0.0289318785071373, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.36130717396736145, 'rewards/repeat_penalty/mean': 0.7275916337966919, 'rewards/repeat_penalty/std': 0.045595794916152954, 'rewards/repeat_floor/mean': 0.9805183410644531, 'rewards/repeat_floor/std': 0.00911917444318533, 'rewards/near_duplicate_penalty/mean': 0.9701220989227295, 'rewards/near_duplicate_penalty/std': 0.0607944019138813, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5143482089042664, 'rewards/total_composite/std': 0.4180837571620941, 'reward': 0.5143482089042664, 'reward_std': 0.41808369755744934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15848523378372192, 'sampling/sampling_logp_difference/max': 1.434493064880371, 'sampling/importance_sampling_ratio/min': 0.23823609948158264, 'sampling/importance_sampling_ratio/mean': 0.9839390516281128, 'sampling/importance_sampling_ratio/max': 1.670961856842041, 'entropy': 1.060697577893734, 'clip_ratio/low_mean': 0.05887050414457917, 'clip_ratio/low_min': 0.05887050414457917, 'clip_ratio/high_mean': 0.07446741918101907, 'clip_ratio/high_max': 0.07446741918101907, 'clip_ratio/region_mean': 0.13333792332559824, 'reward_total_mean': 0.5143482089042664, 'reward_meter_mean': 0.7712091207504272, 'reward_meter_std': 0.4052234888076782, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.984375, 'reward_lexical_plausibility_std': 0.0289318785071373, 'reward_repeat_penalty_mean': 0.7275916337966919, 'reward_repeat_penalty_std': 0.045595794916152954, 'reward_repeat_floor_mean': 0.9805183410644531, 'reward_repeat_floor_std': 0.00911917444318533, 'reward_near_duplicate_penalty_mean': 0.9701220989227295, 'reward_near_duplicate_penalty_std': 0.0607944019138813, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.36130717396736145, 'reward_total_composite_mean': 0.5143482089042664, 'reward_total_composite_std': 0.4180837571620941, 'epoch': 0.07}

 38%|███▊      | 1270/3300 [4:52:01<6:01:56, 10.70s/it]INFO 04-12 20:50:27 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1271/3300 [4:52:15<6:44:32, 11.96s/it]2026-04-12 20:50:42,139 | INFO | train_grpo_train | metrics_logged mode=train step=1271

                                                       
{'loss': -0.0548, 'grad_norm': 4.0297932624816895, 'learning_rate': 6.151515151515152e-06, 'num_tokens': 2257160.0, 'completions/mean_length': 97.0, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.71428680419922, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7833063006401062, 'rewards/meter/std': 0.38132694363594055, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9453494548797607, 'rewards/lexical_plausibility/std': 0.15457506477832794, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.7664239406585693, 'rewards/repeat_penalty/std': 0.23794934153556824, 'rewards/repeat_floor/mean': 0.974159300327301, 'rewards/repeat_floor/std': 0.023759540170431137, 'rewards/near_duplicate_penalty/mean': 0.9218281507492065, 'rewards/near_duplicate_penalty/std': 0.04649970680475235, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.9160839319229126, 'rewards/distinct_2/std': 0.08180836588144302, 'rewards/total_composite/mean': 0.268919438123703, 'rewards/total_composite/std': 0.18038040399551392, 'reward': 0.268919438123703, 'reward_std': 0.18038038909435272, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1051422506570816, 'sampling/sampling_logp_difference/max': 1.1019303798675537, 'sampling/importance_sampling_ratio/min': 0.3322291374206543, 'sampling/importance_sampling_ratio/mean': 1.0184539556503296, 'sampling/importance_sampling_ratio/max': 1.8097755908966064, 'entropy': 0.7051658853888512, 'clip_ratio/low_mean': 0.025760134682059288, 'clip_ratio/low_min': 0.025760134682059288, 'clip_ratio/high_mean': 0.05704365251585841, 'clip_ratio/high_max': 0.05704365251585841, 'clip_ratio/region_mean': 0.0828037871979177, 'reward_total_mean': 0.268919438123703, 'reward_meter_mean': 0.7833063006401062, 'reward_meter_std': 0.38132694363594055, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9453494548797607, 'reward_lexical_plausibility_std': 0.15457506477832794, 'reward_repeat_penalty_mean': 0.7664239406585693, 'reward_repeat_penalty_std': 0.23794934153556824, 'reward_repeat_floor_mean': 0.974159300327301, 'reward_repeat_floor_std': 0.023759540170431137, 'reward_near_duplicate_penalty_mean': 0.9218281507492065, 'reward_near_duplicate_penalty_std': 0.04649970680475235, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.9160839319229126, 'reward_distinct_2_std': 0.08180836588144302, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.268919438123703, 'reward_total_composite_std': 0.18038040399551392, 'epoch': 0.07}

 39%|███▊      | 1271/3300 [4:52:15<6:44:32, 11.96s/it]INFO 04-12 20:50:42 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1272/3300 [4:52:31<7:26:30, 13.21s/it]2026-04-12 20:50:58,269 | INFO | train_grpo_train | metrics_logged mode=train step=1272

                                                       
{'loss': -0.185, 'grad_norm': 1.8744791746139526, 'learning_rate': 6.148484848484849e-06, 'num_tokens': 2259287.0, 'completions/mean_length': 197.875, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 93.16667175292969, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.8875970840454102, 'rewards/meter/std': 0.2766491770744324, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.023145480081439018, 'rewards/lexical_plausibility/mean': 0.8949588537216187, 'rewards/lexical_plausibility/std': 0.195562481880188, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8824868202209473, 'rewards/repeat_penalty/std': 0.13397713005542755, 'rewards/repeat_floor/mean': 0.9877086877822876, 'rewards/repeat_floor/std': 0.012082344852387905, 'rewards/near_duplicate_penalty/mean': 0.9649714827537537, 'rewards/near_duplicate_penalty/std': 0.01671692356467247, 'rewards/opening_diversity/mean': 0.9609375, 'rewards/opening_diversity/std': 0.08799287676811218, 'rewards/distinct_2/mean': 0.9478525519371033, 'rewards/distinct_2/std': 0.074729323387146, 'rewards/total_composite/mean': 0.23223958909511566, 'rewards/total_composite/std': 0.12774422764778137, 'reward': 0.23223958909511566, 'reward_std': 0.12774422764778137, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15499474108219147, 'sampling/sampling_logp_difference/max': 1.9885585308074951, 'sampling/importance_sampling_ratio/min': 0.13689260184764862, 'sampling/importance_sampling_ratio/mean': 1.0061284303665161, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8128736615180969, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.12715011462569237, 'clip_ratio/high_max': 0.12715011462569237, 'clip_ratio/region_mean': 0.12715011462569237, 'reward_total_mean': 0.23223958909511566, 'reward_meter_mean': 0.8875970840454102, 'reward_meter_std': 0.2766491770744324, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.023145480081439018, 'reward_lexical_plausibility_mean': 0.8949588537216187, 'reward_lexical_plausibility_std': 0.195562481880188, 'reward_repeat_penalty_mean': 0.8824868202209473, 'reward_repeat_penalty_std': 0.13397713005542755, 'reward_repeat_floor_mean': 0.9877086877822876, 'reward_repeat_floor_std': 0.012082344852387905, 'reward_near_duplicate_penalty_mean': 0.9649714827537537, 'reward_near_duplicate_penalty_std': 0.01671692356467247, 'reward_opening_diversity_mean': 0.9609375, 'reward_opening_diversity_std': 0.08799287676811218, 'reward_distinct_2_mean': 0.9478525519371033, 'reward_distinct_2_std': 0.074729323387146, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.23223958909511566, 'reward_total_composite_std': 0.12774422764778137, 'epoch': 0.07}

 39%|███▊      | 1272/3300 [4:52:32<7:26:30, 13.21s/it]INFO 04-12 20:50:58 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1273/3300 [4:52:43<7:05:17, 12.59s/it]2026-04-12 20:51:09,385 | INFO | train_grpo_train | metrics_logged mode=train step=1273

                                                       
{'loss': -0.0014, 'grad_norm': 19.351198196411133, 'learning_rate': 6.1454545454545454e-06, 'num_tokens': 2260517.0, 'completions/mean_length': 18.75, 'completions/min_length': 16.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.889035701751709, 'rewards/meter/std': 0.2521689236164093, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.11259915679693222, 'rewards/repeat_penalty/mean': 0.7478047609329224, 'rewards/repeat_penalty/std': 0.004292828496545553, 'rewards/repeat_floor/mean': 0.9845609664916992, 'rewards/repeat_floor/std': 0.0008585641044192016, 'rewards/near_duplicate_penalty/mean': 0.9970729947090149, 'rewards/near_duplicate_penalty/std': 0.005723773967474699, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3275934159755707, 'rewards/total_composite/std': 0.14551202952861786, 'reward': 0.3275934159755707, 'reward_std': 0.14551201462745667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16604115068912506, 'sampling/sampling_logp_difference/max': 2.1093358993530273, 'sampling/importance_sampling_ratio/min': 0.1213185042142868, 'sampling/importance_sampling_ratio/mean': 1.019057273864746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8111205697059631, 'clip_ratio/low_mean': 0.04062500037252903, 'clip_ratio/low_min': 0.04062500037252903, 'clip_ratio/high_mean': 0.09177355095744133, 'clip_ratio/high_max': 0.09177355095744133, 'clip_ratio/region_mean': 0.13239855132997036, 'reward_total_mean': 0.3275934159755707, 'reward_meter_mean': 0.889035701751709, 'reward_meter_std': 0.2521689236164093, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7478047609329224, 'reward_repeat_penalty_std': 0.004292828496545553, 'reward_repeat_floor_mean': 0.9845609664916992, 'reward_repeat_floor_std': 0.0008585641044192016, 'reward_near_duplicate_penalty_mean': 0.9970729947090149, 'reward_near_duplicate_penalty_std': 0.005723773967474699, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.11259915679693222, 'reward_total_composite_mean': 0.3275934159755707, 'reward_total_composite_std': 0.14551202952861786, 'epoch': 0.07}

 39%|███▊      | 1273/3300 [4:52:43<7:05:17, 12.59s/it]INFO 04-12 20:51:09 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1274/3300 [4:52:51<6:24:11, 11.38s/it]2026-04-12 20:51:17,934 | INFO | train_grpo_train | metrics_logged mode=train step=1274

                                                       
{'loss': 0.0317, 'grad_norm': 20.619544982910156, 'learning_rate': 6.142424242424243e-06, 'num_tokens': 2262087.0, 'completions/mean_length': 20.25, 'completions/min_length': 19.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.5778706073760986, 'rewards/meter/std': 0.4129282534122467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6687500476837158, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.6929779052734375, 'rewards/repeat_penalty/std': 0.0843266025185585, 'rewards/repeat_floor/mean': 0.9735955595970154, 'rewards/repeat_floor/std': 0.016865329816937447, 'rewards/near_duplicate_penalty/mean': 0.9239704608917236, 'rewards/near_duplicate_penalty/std': 0.1124354749917984, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.42597997188568115, 'rewards/total_composite/std': 0.40709230303764343, 'reward': 0.42597997188568115, 'reward_std': 0.40709230303764343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13488778471946716, 'sampling/sampling_logp_difference/max': 1.7053031921386719, 'sampling/importance_sampling_ratio/min': 0.18171727657318115, 'sampling/importance_sampling_ratio/mean': 1.046378493309021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9697450995445251, 'clip_ratio/low_mean': 0.06265664286911488, 'clip_ratio/low_min': 0.06265664286911488, 'clip_ratio/high_mean': 0.04302876768633723, 'clip_ratio/high_max': 0.04302876768633723, 'clip_ratio/region_mean': 0.10568541055545211, 'reward_total_mean': 0.42597997188568115, 'reward_meter_mean': 0.5778706073760986, 'reward_meter_std': 0.4129282534122467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.6929779052734375, 'reward_repeat_penalty_std': 0.0843266025185585, 'reward_repeat_floor_mean': 0.9735955595970154, 'reward_repeat_floor_std': 0.016865329816937447, 'reward_near_duplicate_penalty_mean': 0.9239704608917236, 'reward_near_duplicate_penalty_std': 0.1124354749917984, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6687500476837158, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.42597997188568115, 'reward_total_composite_std': 0.40709230303764343, 'epoch': 0.07}

 39%|███▊      | 1274/3300 [4:52:51<6:24:11, 11.38s/it]INFO 04-12 20:51:18 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1275/3300 [4:52:59<5:52:34, 10.45s/it]2026-04-12 20:51:26,159 | INFO | train_grpo_train | metrics_logged mode=train step=1275

                                                       
{'loss': -0.0046, 'grad_norm': 15.704371452331543, 'learning_rate': 6.139393939393939e-06, 'num_tokens': 2263567.0, 'completions/mean_length': 23.0, 'completions/min_length': 20.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5792652368545532, 'rewards/meter/std': 0.4077938199043274, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8049999475479126, 'rewards/judge_quality/std': 0.2517368495464325, 'rewards/repeat_penalty/mean': 0.7330512404441833, 'rewards/repeat_penalty/std': 0.032561879605054855, 'rewards/repeat_floor/mean': 0.9816102981567383, 'rewards/repeat_floor/std': 0.006512374151498079, 'rewards/near_duplicate_penalty/mean': 0.977401614189148, 'rewards/near_duplicate_penalty/std': 0.043415818363428116, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.43998605012893677, 'rewards/total_composite/std': 0.3601395785808563, 'reward': 0.43998605012893677, 'reward_std': 0.3601395785808563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14472268521785736, 'sampling/sampling_logp_difference/max': 1.3146476745605469, 'sampling/importance_sampling_ratio/min': 0.2685689330101013, 'sampling/importance_sampling_ratio/mean': 0.992941677570343, 'sampling/importance_sampling_ratio/max': 1.663733720779419, 'entropy': 1.2278536260128021, 'clip_ratio/low_mean': 0.05238742334768176, 'clip_ratio/low_min': 0.05238742334768176, 'clip_ratio/high_mean': 0.09493485651910305, 'clip_ratio/high_max': 0.09493485651910305, 'clip_ratio/region_mean': 0.1473222798667848, 'reward_total_mean': 0.43998605012893677, 'reward_meter_mean': 0.5792652368545532, 'reward_meter_std': 0.4077938199043274, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7330512404441833, 'reward_repeat_penalty_std': 0.032561879605054855, 'reward_repeat_floor_mean': 0.9816102981567383, 'reward_repeat_floor_std': 0.006512374151498079, 'reward_near_duplicate_penalty_mean': 0.977401614189148, 'reward_near_duplicate_penalty_std': 0.043415818363428116, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8049999475479126, 'reward_judge_quality_std': 0.2517368495464325, 'reward_total_composite_mean': 0.43998605012893677, 'reward_total_composite_std': 0.3601395785808563, 'epoch': 0.07}

 39%|███▊      | 1275/3300 [4:53:00<5:52:34, 10.45s/it]INFO 04-12 20:51:26 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1276/3300 [4:53:09<5:48:09, 10.32s/it]2026-04-12 20:51:36,258 | INFO | train_grpo_train | metrics_logged mode=train step=1276

                                                       
{'loss': 0.0295, 'grad_norm': 9.942580223083496, 'learning_rate': 6.136363636363637e-06, 'num_tokens': 2266485.0, 'completions/mean_length': 125.75, 'completions/min_length': 115.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.75, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.9674736261367798, 'rewards/meter/std': 0.07034432142972946, 'rewards/count_adherence/mean': 0.5555555820465088, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8666666746139526, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.7272279858589172, 'rewards/repeat_penalty/std': 0.16049474477767944, 'rewards/repeat_floor/mean': 0.9704926013946533, 'rewards/repeat_floor/std': 0.015507549047470093, 'rewards/near_duplicate_penalty/mean': 0.9288506507873535, 'rewards/near_duplicate_penalty/std': 0.023925166577100754, 'rewards/opening_diversity/mean': 0.9147727489471436, 'rewards/opening_diversity/std': 0.14389893412590027, 'rewards/distinct_2/mean': 0.8475406169891357, 'rewards/distinct_2/std': 0.060464322566986084, 'rewards/total_composite/mean': 0.24789929389953613, 'rewards/total_composite/std': 0.08466501533985138, 'reward': 0.24789929389953613, 'reward_std': 0.08466501533985138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16390083730220795, 'sampling/sampling_logp_difference/max': 2.1465890407562256, 'sampling/importance_sampling_ratio/min': 0.11688216030597687, 'sampling/importance_sampling_ratio/mean': 1.0131659507751465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0201633274555206, 'clip_ratio/low_mean': 0.031063650734722614, 'clip_ratio/low_min': 0.031063650734722614, 'clip_ratio/high_mean': 0.12579221092164516, 'clip_ratio/high_max': 0.12579221092164516, 'clip_ratio/region_mean': 0.15685586165636778, 'reward_total_mean': 0.24789929389953613, 'reward_meter_mean': 0.9674736261367798, 'reward_meter_std': 0.07034432142972946, 'reward_count_adherence_mean': 0.5555555820465088, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8666666746139526, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7272279858589172, 'reward_repeat_penalty_std': 0.16049474477767944, 'reward_repeat_floor_mean': 0.9704926013946533, 'reward_repeat_floor_std': 0.015507549047470093, 'reward_near_duplicate_penalty_mean': 0.9288506507873535, 'reward_near_duplicate_penalty_std': 0.023925166577100754, 'reward_opening_diversity_mean': 0.9147727489471436, 'reward_opening_diversity_std': 0.14389893412590027, 'reward_distinct_2_mean': 0.8475406169891357, 'reward_distinct_2_std': 0.060464322566986084, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.24789929389953613, 'reward_total_composite_std': 0.08466501533985138, 'epoch': 0.07}

 39%|███▊      | 1276/3300 [4:53:10<5:48:09, 10.32s/it]INFO 04-12 20:51:36 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1277/3300 [4:53:19<5:38:48, 10.05s/it]2026-04-12 20:51:45,584 | INFO | train_grpo_train | metrics_logged mode=train step=1277

                                                       
{'loss': -0.0896, 'grad_norm': 26.013608932495117, 'learning_rate': 6.133333333333334e-06, 'num_tokens': 2267761.0, 'completions/mean_length': 17.5, 'completions/min_length': 13.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.5, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7073818445205688, 'rewards/meter/std': 0.36538729071617126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6475000381469727, 'rewards/judge_quality/std': 0.30742016434669495, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4580766558647156, 'rewards/total_composite/std': 0.3642033636569977, 'reward': 0.4580766558647156, 'reward_std': 0.3642033040523529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21130113303661346, 'sampling/sampling_logp_difference/max': 1.5415663719177246, 'sampling/importance_sampling_ratio/min': 0.21404556930065155, 'sampling/importance_sampling_ratio/mean': 1.0190093517303467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3722561374306679, 'clip_ratio/low_mean': 0.12367457337677479, 'clip_ratio/low_min': 0.12367457337677479, 'clip_ratio/high_mean': 0.07932423241436481, 'clip_ratio/high_max': 0.07932423241436481, 'clip_ratio/region_mean': 0.2029988057911396, 'reward_total_mean': 0.4580766558647156, 'reward_meter_mean': 0.7073818445205688, 'reward_meter_std': 0.36538729071617126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6475000381469727, 'reward_judge_quality_std': 0.30742016434669495, 'reward_total_composite_mean': 0.4580766558647156, 'reward_total_composite_std': 0.3642033636569977, 'epoch': 0.07}

 39%|███▊      | 1277/3300 [4:53:19<5:38:48, 10.05s/it]INFO 04-12 20:51:45 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1278/3300 [4:53:31<5:55:12, 10.54s/it]2026-04-12 20:51:57,376 | INFO | train_grpo_train | metrics_logged mode=train step=1278

                                                       
{'loss': 0.0498, 'grad_norm': 15.023978233337402, 'learning_rate': 6.130303030303031e-06, 'num_tokens': 2269458.0, 'completions/mean_length': 44.125, 'completions/min_length': 39.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6327494382858276, 'rewards/meter/std': 0.3527475595474243, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.988175630569458, 'rewards/repeat_penalty/std': 0.025807036086916924, 'rewards/repeat_floor/mean': 0.9986739754676819, 'rewards/repeat_floor/std': 0.0026872691232711077, 'rewards/near_duplicate_penalty/mean': 0.9957752227783203, 'rewards/near_duplicate_penalty/std': 0.0082395039498806, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.24861422181129456, 'rewards/total_composite/std': 0.14773215353488922, 'reward': 0.24861422181129456, 'reward_std': 0.14773216843605042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19556066393852234, 'sampling/sampling_logp_difference/max': 1.4899392127990723, 'sampling/importance_sampling_ratio/min': 0.2705949544906616, 'sampling/importance_sampling_ratio/mean': 1.0542553663253784, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8790571242570877, 'clip_ratio/low_mean': 0.06543927919119596, 'clip_ratio/low_min': 0.06543927919119596, 'clip_ratio/high_mean': 0.0822699274867773, 'clip_ratio/high_max': 0.0822699274867773, 'clip_ratio/region_mean': 0.14770920667797327, 'reward_total_mean': 0.24861422181129456, 'reward_meter_mean': 0.6327494382858276, 'reward_meter_std': 0.3527475595474243, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.988175630569458, 'reward_repeat_penalty_std': 0.025807036086916924, 'reward_repeat_floor_mean': 0.9986739754676819, 'reward_repeat_floor_std': 0.0026872691232711077, 'reward_near_duplicate_penalty_mean': 0.9957752227783203, 'reward_near_duplicate_penalty_std': 0.0082395039498806, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.24861422181129456, 'reward_total_composite_std': 0.14773215353488922, 'epoch': 0.07}

 39%|███▊      | 1278/3300 [4:53:31<5:55:12, 10.54s/it]INFO 04-12 20:51:57 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1279/3300 [4:53:44<6:25:32, 11.45s/it]2026-04-12 20:52:10,830 | INFO | train_grpo_train | metrics_logged mode=train step=1279

                                                       
{'loss': 0.0127, 'grad_norm': 10.804234504699707, 'learning_rate': 6.127272727272727e-06, 'num_tokens': 2271055.0, 'completions/mean_length': 51.625, 'completions/min_length': 44.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9879533052444458, 'rewards/meter/std': 0.005108564160764217, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121345579624176, 'rewards/repeat_penalty/mean': 0.6141194105148315, 'rewards/repeat_penalty/std': 0.06002319976687431, 'rewards/repeat_floor/mean': 0.9644373655319214, 'rewards/repeat_floor/std': 0.009719708934426308, 'rewards/near_duplicate_penalty/mean': 0.9412378668785095, 'rewards/near_duplicate_penalty/std': 0.04470231384038925, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8694632053375244, 'rewards/distinct_2/std': 0.06619791686534882, 'rewards/total_composite/mean': 0.3886520564556122, 'rewards/total_composite/std': 0.22585095465183258, 'reward': 0.3886520564556122, 'reward_std': 0.2258509397506714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13005231320858002, 'sampling/sampling_logp_difference/max': 1.3732376098632812, 'sampling/importance_sampling_ratio/min': 0.25328558683395386, 'sampling/importance_sampling_ratio/mean': 1.0239617824554443, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9145958125591278, 'clip_ratio/low_mean': 0.06817550025880337, 'clip_ratio/low_min': 0.06817550025880337, 'clip_ratio/high_mean': 0.0639254841953516, 'clip_ratio/high_max': 0.0639254841953516, 'clip_ratio/region_mean': 0.13210098445415497, 'reward_total_mean': 0.3886520564556122, 'reward_meter_mean': 0.9879533052444458, 'reward_meter_std': 0.005108564160764217, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6141194105148315, 'reward_repeat_penalty_std': 0.06002319976687431, 'reward_repeat_floor_mean': 0.9644373655319214, 'reward_repeat_floor_std': 0.009719708934426308, 'reward_near_duplicate_penalty_mean': 0.9412378668785095, 'reward_near_duplicate_penalty_std': 0.04470231384038925, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8694632053375244, 'reward_distinct_2_std': 0.06619791686534882, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.24121345579624176, 'reward_total_composite_mean': 0.3886520564556122, 'reward_total_composite_std': 0.22585095465183258, 'epoch': 0.07}

 39%|███▉      | 1279/3300 [4:53:44<6:25:32, 11.45s/it]INFO 04-12 20:52:11 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1280/3300 [4:53:52<5:53:34, 10.50s/it]2026-04-12 20:52:19,139 | INFO | train_grpo_train | metrics_logged mode=train step=1280

                                                       
{'loss': 0.0655, 'grad_norm': 19.38922691345215, 'learning_rate': 6.1242424242424245e-06, 'num_tokens': 2272399.0, 'completions/mean_length': 22.0, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9677039980888367, 'rewards/meter/std': 0.04641558229923248, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.7378586530685425, 'rewards/repeat_penalty/std': 0.031095918267965317, 'rewards/repeat_floor/mean': 0.9834593534469604, 'rewards/repeat_floor/std': 0.0037274223286658525, 'rewards/near_duplicate_penalty/mean': 0.9986044764518738, 'rewards/near_duplicate_penalty/std': 0.003947157878428698, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.8240794539451599, 'rewards/total_composite/std': 0.1687736064195633, 'reward': 0.8240794539451599, 'reward_std': 0.1687736064195633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.134673610329628, 'sampling/sampling_logp_difference/max': 0.9066610336303711, 'sampling/importance_sampling_ratio/min': 0.40387049317359924, 'sampling/importance_sampling_ratio/mean': 1.0463017225265503, 'sampling/importance_sampling_ratio/max': 1.8804807662963867, 'entropy': 1.0864523872733116, 'clip_ratio/low_mean': 0.04166666604578495, 'clip_ratio/low_min': 0.04166666604578495, 'clip_ratio/high_mean': 0.09955168422311544, 'clip_ratio/high_max': 0.09955168422311544, 'clip_ratio/region_mean': 0.1412183502689004, 'reward_total_mean': 0.8240794539451599, 'reward_meter_mean': 0.9677039980888367, 'reward_meter_std': 0.04641558229923248, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7378586530685425, 'reward_repeat_penalty_std': 0.031095918267965317, 'reward_repeat_floor_mean': 0.9834593534469604, 'reward_repeat_floor_std': 0.0037274223286658525, 'reward_near_duplicate_penalty_mean': 0.9986044764518738, 'reward_near_duplicate_penalty_std': 0.003947157878428698, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.8240794539451599, 'reward_total_composite_std': 0.1687736064195633, 'epoch': 0.07}

 39%|███▉      | 1280/3300 [4:53:52<5:53:34, 10.50s/it]INFO 04-12 20:52:19 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1281/3300 [4:54:08<6:41:54, 11.94s/it]2026-04-12 20:52:34,461 | INFO | train_grpo_train | metrics_logged mode=train step=1281

                                                       
{'loss': -0.1159, 'grad_norm': 4.049616813659668, 'learning_rate': 6.121212121212121e-06, 'num_tokens': 2274800.0, 'completions/mean_length': 158.125, 'completions/min_length': 96.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 107.5714340209961, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.932712197303772, 'rewards/meter/std': 0.153072789311409, 'rewards/count_adherence/mean': 0.5714285969734192, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8714285492897034, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9623285531997681, 'rewards/lexical_plausibility/std': 0.1065509244799614, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.6622878313064575, 'rewards/repeat_penalty/std': 0.24458257853984833, 'rewards/repeat_floor/mean': 0.9626510143280029, 'rewards/repeat_floor/std': 0.030960487201809883, 'rewards/near_duplicate_penalty/mean': 0.9089532494544983, 'rewards/near_duplicate_penalty/std': 0.059056829661130905, 'rewards/opening_diversity/mean': 0.8671875, 'rewards/opening_diversity/std': 0.21505165100097656, 'rewards/distinct_2/mean': 0.8252971768379211, 'rewards/distinct_2/std': 0.14173123240470886, 'rewards/total_composite/mean': 0.2526596188545227, 'rewards/total_composite/std': 0.10356839001178741, 'reward': 0.2526596188545227, 'reward_std': 0.10356838256120682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14163656532764435, 'sampling/sampling_logp_difference/max': 1.6069011688232422, 'sampling/importance_sampling_ratio/min': 0.2005079984664917, 'sampling/importance_sampling_ratio/mean': 1.005495309829712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9161988273262978, 'clip_ratio/low_mean': 0.016447369009256363, 'clip_ratio/low_min': 0.016447369009256363, 'clip_ratio/high_mean': 0.10982071235775948, 'clip_ratio/high_max': 0.10982071235775948, 'clip_ratio/region_mean': 0.12626808136701584, 'reward_total_mean': 0.2526596188545227, 'reward_meter_mean': 0.932712197303772, 'reward_meter_std': 0.153072789311409, 'reward_count_adherence_mean': 0.5714285969734192, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8714285492897034, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9623285531997681, 'reward_lexical_plausibility_std': 0.1065509244799614, 'reward_repeat_penalty_mean': 0.6622878313064575, 'reward_repeat_penalty_std': 0.24458257853984833, 'reward_repeat_floor_mean': 0.9626510143280029, 'reward_repeat_floor_std': 0.030960487201809883, 'reward_near_duplicate_penalty_mean': 0.9089532494544983, 'reward_near_duplicate_penalty_std': 0.059056829661130905, 'reward_opening_diversity_mean': 0.8671875, 'reward_opening_diversity_std': 0.21505165100097656, 'reward_distinct_2_mean': 0.8252971768379211, 'reward_distinct_2_std': 0.14173123240470886, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.2526596188545227, 'reward_total_composite_std': 0.10356839001178741, 'epoch': 0.07}

 39%|███▉      | 1281/3300 [4:54:08<6:41:54, 11.94s/it]INFO 04-12 20:52:34 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1282/3300 [4:54:22<7:09:01, 12.76s/it]2026-04-12 20:52:49,094 | INFO | train_grpo_train | metrics_logged mode=train step=1282

                                                       
{'loss': -0.0367, 'grad_norm': 4.579404354095459, 'learning_rate': 6.118181818181819e-06, 'num_tokens': 2276398.0, 'completions/mean_length': 99.75, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.85714340209961, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7003982067108154, 'rewards/meter/std': 0.421455442905426, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9506106376647949, 'rewards/lexical_plausibility/std': 0.13969433307647705, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.8909257650375366, 'rewards/repeat_penalty/std': 0.146341472864151, 'rewards/repeat_floor/mean': 0.9866161346435547, 'rewards/repeat_floor/std': 0.01638607122004032, 'rewards/near_duplicate_penalty/mean': 0.9431406259536743, 'rewards/near_duplicate_penalty/std': 0.05240917205810547, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9668891429901123, 'rewards/distinct_2/std': 0.04862732067704201, 'rewards/total_composite/mean': 0.2612893879413605, 'rewards/total_composite/std': 0.1929890662431717, 'reward': 0.2612893879413605, 'reward_std': 0.19298908114433289, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16385674476623535, 'sampling/sampling_logp_difference/max': 2.9079105854034424, 'sampling/importance_sampling_ratio/min': 0.05458967387676239, 'sampling/importance_sampling_ratio/mean': 1.0130107402801514, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.126872956752777, 'clip_ratio/low_mean': 0.034722222946584225, 'clip_ratio/low_min': 0.034722222946584225, 'clip_ratio/high_mean': 0.08194526610895991, 'clip_ratio/high_max': 0.08194526610895991, 'clip_ratio/region_mean': 0.11666748905554414, 'reward_total_mean': 0.2612893879413605, 'reward_meter_mean': 0.7003982067108154, 'reward_meter_std': 0.421455442905426, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9506106376647949, 'reward_lexical_plausibility_std': 0.13969433307647705, 'reward_repeat_penalty_mean': 0.8909257650375366, 'reward_repeat_penalty_std': 0.146341472864151, 'reward_repeat_floor_mean': 0.9866161346435547, 'reward_repeat_floor_std': 0.01638607122004032, 'reward_near_duplicate_penalty_mean': 0.9431406259536743, 'reward_near_duplicate_penalty_std': 0.05240917205810547, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9668891429901123, 'reward_distinct_2_std': 0.04862732067704201, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.2612893879413605, 'reward_total_composite_std': 0.1929890662431717, 'epoch': 0.07}

 39%|███▉      | 1282/3300 [4:54:22<7:09:01, 12.76s/it]INFO 04-12 20:52:49 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1283/3300 [4:54:31<6:29:30, 11.59s/it]2026-04-12 20:52:57,963 | INFO | train_grpo_train | metrics_logged mode=train step=1283

                                                       
{'loss': 0.0156, 'grad_norm': 18.195934295654297, 'learning_rate': 6.115151515151516e-06, 'num_tokens': 2277797.0, 'completions/mean_length': 26.875, 'completions/min_length': 23.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.3963731825351715, 'rewards/meter/std': 0.4179953634738922, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9705680012702942, 'rewards/repeat_penalty/std': 0.02303503267467022, 'rewards/repeat_floor/mean': 0.9955852031707764, 'rewards/repeat_floor/std': 0.003455251455307007, 'rewards/near_duplicate_penalty/mean': 0.9705680012702942, 'rewards/near_duplicate_penalty/std': 0.02303503267467022, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24516846239566803, 'rewards/total_composite/std': 0.2693071961402893, 'reward': 0.24516846239566803, 'reward_std': 0.2693071961402893, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16867254674434662, 'sampling/sampling_logp_difference/max': 1.825056552886963, 'sampling/importance_sampling_ratio/min': 0.16120852530002594, 'sampling/importance_sampling_ratio/mean': 1.0346688032150269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0673444122076035, 'clip_ratio/low_mean': 0.09447878506034613, 'clip_ratio/low_min': 0.09447878506034613, 'clip_ratio/high_mean': 0.0480658458545804, 'clip_ratio/high_max': 0.0480658458545804, 'clip_ratio/region_mean': 0.14254463091492653, 'reward_total_mean': 0.24516846239566803, 'reward_meter_mean': 0.3963731825351715, 'reward_meter_std': 0.4179953634738922, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9705680012702942, 'reward_repeat_penalty_std': 0.02303503267467022, 'reward_repeat_floor_mean': 0.9955852031707764, 'reward_repeat_floor_std': 0.003455251455307007, 'reward_near_duplicate_penalty_mean': 0.9705680012702942, 'reward_near_duplicate_penalty_std': 0.02303503267467022, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.24516846239566803, 'reward_total_composite_std': 0.2693071961402893, 'epoch': 0.07}

 39%|███▉      | 1283/3300 [4:54:31<6:29:30, 11.59s/it]INFO 04-12 20:52:58 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1284/3300 [4:54:41<6:10:02, 11.01s/it]2026-04-12 20:53:07,654 | INFO | train_grpo_train | metrics_logged mode=train step=1284

                                                       
{'loss': 0.0247, 'grad_norm': 20.05014419555664, 'learning_rate': 6.112121212121213e-06, 'num_tokens': 2279296.0, 'completions/mean_length': 34.375, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.629314661026001, 'rewards/meter/std': 0.44294917583465576, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9662244915962219, 'rewards/repeat_penalty/std': 0.03941755369305611, 'rewards/repeat_floor/mean': 0.9954374432563782, 'rewards/repeat_floor/std': 0.0049843802116811275, 'rewards/near_duplicate_penalty/mean': 0.9750772714614868, 'rewards/near_duplicate_penalty/std': 0.027527790516614914, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.25952810049057007, 'rewards/total_composite/std': 0.22622127830982208, 'reward': 0.25952810049057007, 'reward_std': 0.2262212634086609, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1595010608434677, 'sampling/sampling_logp_difference/max': 1.266634464263916, 'sampling/importance_sampling_ratio/min': 0.28177836537361145, 'sampling/importance_sampling_ratio/mean': 1.003468632698059, 'sampling/importance_sampling_ratio/max': 1.7857407331466675, 'entropy': 1.1013183817267418, 'clip_ratio/low_mean': 0.08061044104397297, 'clip_ratio/low_min': 0.08061044104397297, 'clip_ratio/high_mean': 0.09057654067873955, 'clip_ratio/high_max': 0.09057654067873955, 'clip_ratio/region_mean': 0.17118698172271252, 'reward_total_mean': 0.25952810049057007, 'reward_meter_mean': 0.629314661026001, 'reward_meter_std': 0.44294917583465576, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9662244915962219, 'reward_repeat_penalty_std': 0.03941755369305611, 'reward_repeat_floor_mean': 0.9954374432563782, 'reward_repeat_floor_std': 0.0049843802116811275, 'reward_near_duplicate_penalty_mean': 0.9750772714614868, 'reward_near_duplicate_penalty_std': 0.027527790516614914, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.25952810049057007, 'reward_total_composite_std': 0.22622127830982208, 'epoch': 0.07}

 39%|███▉      | 1284/3300 [4:54:41<6:10:02, 11.01s/it]INFO 04-12 20:53:07 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1285/3300 [4:54:55<6:45:58, 12.09s/it]2026-04-12 20:53:22,244 | INFO | train_grpo_train | metrics_logged mode=train step=1285

                                                       
{'loss': -0.1397, 'grad_norm': 5.8364434242248535, 'learning_rate': 6.10909090909091e-06, 'num_tokens': 2281144.0, 'completions/mean_length': 121.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.14286041259766, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.7154441475868225, 'rewards/meter/std': 0.3799416124820709, 'rewards/count_adherence/mean': 0.6750000715255737, 'rewards/count_adherence/std': 0.14880476891994476, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 0.9024999737739563, 'rewards/count_floor/std': 0.04464143142104149, 'rewards/lexical_plausibility/mean': 0.9469554424285889, 'rewards/lexical_plausibility/std': 0.15003259479999542, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.8863404989242554, 'rewards/repeat_penalty/std': 0.07637932151556015, 'rewards/repeat_floor/mean': 0.9869344234466553, 'rewards/repeat_floor/std': 0.008240737952291965, 'rewards/near_duplicate_penalty/mean': 0.9557437300682068, 'rewards/near_duplicate_penalty/std': 0.025231685489416122, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9337952136993408, 'rewards/distinct_2/std': 0.05908522382378578, 'rewards/total_composite/mean': 0.20943179726600647, 'rewards/total_composite/std': 0.16537350416183472, 'reward': 0.20943179726600647, 'reward_std': 0.16537350416183472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15690886974334717, 'sampling/sampling_logp_difference/max': 1.7209768295288086, 'sampling/importance_sampling_ratio/min': 0.17889131605625153, 'sampling/importance_sampling_ratio/mean': 1.0197601318359375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8006609156727791, 'clip_ratio/low_mean': 0.07394255138933659, 'clip_ratio/low_min': 0.07394255138933659, 'clip_ratio/high_mean': 0.08503968454897404, 'clip_ratio/high_max': 0.08503968454897404, 'clip_ratio/region_mean': 0.15898223593831062, 'reward_total_mean': 0.20943179726600647, 'reward_meter_mean': 0.7154441475868225, 'reward_meter_std': 0.3799416124820709, 'reward_count_adherence_mean': 0.6750000715255737, 'reward_count_adherence_std': 0.14880476891994476, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 0.9024999737739563, 'reward_count_floor_std': 0.04464143142104149, 'reward_lexical_plausibility_mean': 0.9469554424285889, 'reward_lexical_plausibility_std': 0.15003259479999542, 'reward_repeat_penalty_mean': 0.8863404989242554, 'reward_repeat_penalty_std': 0.07637932151556015, 'reward_repeat_floor_mean': 0.9869344234466553, 'reward_repeat_floor_std': 0.008240737952291965, 'reward_near_duplicate_penalty_mean': 0.9557437300682068, 'reward_near_duplicate_penalty_std': 0.025231685489416122, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9337952136993408, 'reward_distinct_2_std': 0.05908522382378578, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.20943179726600647, 'reward_total_composite_std': 0.16537350416183472, 'epoch': 0.07}

 39%|███▉      | 1285/3300 [4:54:56<6:45:58, 12.09s/it]INFO 04-12 20:53:22 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1286/3300 [4:55:06<6:27:39, 11.55s/it]2026-04-12 20:53:32,522 | INFO | train_grpo_train | metrics_logged mode=train step=1286

                                                       
{'loss': 0.0628, 'grad_norm': 15.974264144897461, 'learning_rate': 6.106060606060606e-06, 'num_tokens': 2282649.0, 'completions/mean_length': 28.125, 'completions/min_length': 24.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.125, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.4469379782676697, 'rewards/meter/std': 0.21926090121269226, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.9348220229148865, 'rewards/repeat_penalty/std': 0.04650967940688133, 'rewards/repeat_floor/mean': 0.9902232885360718, 'rewards/repeat_floor/std': 0.0069764466024935246, 'rewards/near_duplicate_penalty/mean': 0.9348220229148865, 'rewards/near_duplicate_penalty/std': 0.04650967940688133, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20960259437561035, 'rewards/total_composite/std': 0.1295873373746872, 'reward': 0.20960259437561035, 'reward_std': 0.1295873373746872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14299459755420685, 'sampling/sampling_logp_difference/max': 1.3126544952392578, 'sampling/importance_sampling_ratio/min': 0.2691047787666321, 'sampling/importance_sampling_ratio/mean': 1.0156716108322144, 'sampling/importance_sampling_ratio/max': 1.7582467794418335, 'entropy': 0.7529254257678986, 'clip_ratio/low_mean': 0.07921245554462075, 'clip_ratio/low_min': 0.07921245554462075, 'clip_ratio/high_mean': 0.07598907081410289, 'clip_ratio/high_max': 0.07598907081410289, 'clip_ratio/region_mean': 0.15520152635872364, 'reward_total_mean': 0.20960259437561035, 'reward_meter_mean': 0.4469379782676697, 'reward_meter_std': 0.21926090121269226, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9348220229148865, 'reward_repeat_penalty_std': 0.04650967940688133, 'reward_repeat_floor_mean': 0.9902232885360718, 'reward_repeat_floor_std': 0.0069764466024935246, 'reward_near_duplicate_penalty_mean': 0.9348220229148865, 'reward_near_duplicate_penalty_std': 0.04650967940688133, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.20960259437561035, 'reward_total_composite_std': 0.1295873373746872, 'epoch': 0.07}

 39%|███▉      | 1286/3300 [4:55:06<6:27:39, 11.55s/it]INFO 04-12 20:53:32 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1287/3300 [4:55:15<6:07:49, 10.96s/it]2026-04-12 20:53:42,128 | INFO | train_grpo_train | metrics_logged mode=train step=1287

                                                       
{'loss': 0.0199, 'grad_norm': 25.172632217407227, 'learning_rate': 6.103030303030304e-06, 'num_tokens': 2284070.0, 'completions/mean_length': 20.625, 'completions/min_length': 19.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8185949325561523, 'rewards/meter/std': 0.3426293730735779, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9345755577087402, 'rewards/lexical_plausibility/std': 0.13747143745422363, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.7785084247589111, 'rewards/repeat_penalty/std': 0.08966934680938721, 'rewards/repeat_floor/mean': 0.9863266944885254, 'rewards/repeat_floor/std': 0.005636090412735939, 'rewards/near_duplicate_penalty/mean': 0.9963445663452148, 'rewards/near_duplicate_penalty/std': 0.007573677226901054, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.462772011756897, 'rewards/total_composite/std': 0.29279395937919617, 'reward': 0.462772011756897, 'reward_std': 0.29279395937919617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22643554210662842, 'sampling/sampling_logp_difference/max': 1.3879175186157227, 'sampling/importance_sampling_ratio/min': 0.2495945394039154, 'sampling/importance_sampling_ratio/mean': 1.034777283668518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.004843458533287, 'clip_ratio/low_mean': 0.11690094834193587, 'clip_ratio/low_min': 0.11690094834193587, 'clip_ratio/high_mean': 0.04204545542597771, 'clip_ratio/high_max': 0.04204545542597771, 'clip_ratio/region_mean': 0.15894640376791358, 'reward_total_mean': 0.462772011756897, 'reward_meter_mean': 0.8185949325561523, 'reward_meter_std': 0.3426293730735779, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9345755577087402, 'reward_lexical_plausibility_std': 0.13747143745422363, 'reward_repeat_penalty_mean': 0.7785084247589111, 'reward_repeat_penalty_std': 0.08966934680938721, 'reward_repeat_floor_mean': 0.9863266944885254, 'reward_repeat_floor_std': 0.005636090412735939, 'reward_near_duplicate_penalty_mean': 0.9963445663452148, 'reward_near_duplicate_penalty_std': 0.007573677226901054, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.462772011756897, 'reward_total_composite_std': 0.29279395937919617, 'epoch': 0.07}

 39%|███▉      | 1287/3300 [4:55:15<6:07:49, 10.96s/it]INFO 04-12 20:53:42 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1288/3300 [4:55:25<5:49:50, 10.43s/it]2026-04-12 20:53:51,466 | INFO | train_grpo_train | metrics_logged mode=train step=1288

                                                       
{'loss': 0.0692, 'grad_norm': 14.786998748779297, 'learning_rate': 6.1e-06, 'num_tokens': 2285553.0, 'completions/mean_length': 34.375, 'completions/min_length': 30.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.551730751991272, 'rewards/meter/std': 0.4425676763057709, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9871809482574463, 'rewards/repeat_penalty/std': 0.01605856604874134, 'rewards/repeat_floor/mean': 0.9980771541595459, 'rewards/repeat_floor/std': 0.0024087761994451284, 'rewards/near_duplicate_penalty/mean': 0.9871809482574463, 'rewards/near_duplicate_penalty/std': 0.01605856604874134, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22984516620635986, 'rewards/total_composite/std': 0.19475726783275604, 'reward': 0.22984516620635986, 'reward_std': 0.19475726783275604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21011680364608765, 'sampling/sampling_logp_difference/max': 2.3008906841278076, 'sampling/importance_sampling_ratio/min': 0.1001695841550827, 'sampling/importance_sampling_ratio/mean': 1.0125607252120972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.554593026638031, 'clip_ratio/low_mean': 0.08579343743622303, 'clip_ratio/low_min': 0.08579343743622303, 'clip_ratio/high_mean': 0.09757401421666145, 'clip_ratio/high_max': 0.09757401421666145, 'clip_ratio/region_mean': 0.18336745165288448, 'reward_total_mean': 0.22984516620635986, 'reward_meter_mean': 0.551730751991272, 'reward_meter_std': 0.4425676763057709, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9871809482574463, 'reward_repeat_penalty_std': 0.01605856604874134, 'reward_repeat_floor_mean': 0.9980771541595459, 'reward_repeat_floor_std': 0.0024087761994451284, 'reward_near_duplicate_penalty_mean': 0.9871809482574463, 'reward_near_duplicate_penalty_std': 0.01605856604874134, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.22984516620635986, 'reward_total_composite_std': 0.19475726783275604, 'epoch': 0.07}

 39%|███▉      | 1288/3300 [4:55:25<5:49:50, 10.43s/it]INFO 04-12 20:53:51 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1289/3300 [4:55:39<6:27:55, 11.57s/it]2026-04-12 20:54:05,536 | INFO | train_grpo_train | metrics_logged mode=train step=1289

                                                       
{'loss': -0.0897, 'grad_norm': 3.2224481105804443, 'learning_rate': 6.096969696969698e-06, 'num_tokens': 2286896.0, 'completions/mean_length': 95.875, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.42857360839844, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8115239143371582, 'rewards/meter/std': 0.15003734827041626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.936459481716156, 'rewards/lexical_plausibility/std': 0.1797197312116623, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8086724877357483, 'rewards/repeat_penalty/std': 0.2693471908569336, 'rewards/repeat_floor/mean': 0.9807487726211548, 'rewards/repeat_floor/std': 0.027442872524261475, 'rewards/near_duplicate_penalty/mean': 0.9522339105606079, 'rewards/near_duplicate_penalty/std': 0.058630239218473434, 'rewards/opening_diversity/mean': 0.859375, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.9594579339027405, 'rewards/distinct_2/std': 0.06129398196935654, 'rewards/total_composite/mean': 0.28309640288352966, 'rewards/total_composite/std': 0.12222478538751602, 'reward': 0.28309640288352966, 'reward_std': 0.12222477793693542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14320380985736847, 'sampling/sampling_logp_difference/max': 1.214773178100586, 'sampling/importance_sampling_ratio/min': 0.2967773377895355, 'sampling/importance_sampling_ratio/mean': 1.0355345010757446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0828647539019585, 'clip_ratio/low_mean': 0.04904306307435036, 'clip_ratio/low_min': 0.04904306307435036, 'clip_ratio/high_mean': 0.08503476483747363, 'clip_ratio/high_max': 0.08503476483747363, 'clip_ratio/region_mean': 0.134077827911824, 'reward_total_mean': 0.28309640288352966, 'reward_meter_mean': 0.8115239143371582, 'reward_meter_std': 0.15003734827041626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.936459481716156, 'reward_lexical_plausibility_std': 0.1797197312116623, 'reward_repeat_penalty_mean': 0.8086724877357483, 'reward_repeat_penalty_std': 0.2693471908569336, 'reward_repeat_floor_mean': 0.9807487726211548, 'reward_repeat_floor_std': 0.027442872524261475, 'reward_near_duplicate_penalty_mean': 0.9522339105606079, 'reward_near_duplicate_penalty_std': 0.058630239218473434, 'reward_opening_diversity_mean': 0.859375, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.9594579339027405, 'reward_distinct_2_std': 0.06129398196935654, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.28309640288352966, 'reward_total_composite_std': 0.12222478538751602, 'epoch': 0.07}

 39%|███▉      | 1289/3300 [4:55:39<6:27:55, 11.57s/it]INFO 04-12 20:54:05 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1290/3300 [4:55:49<6:16:46, 11.25s/it]2026-04-12 20:54:16,047 | INFO | train_grpo_train | metrics_logged mode=train step=1290

                                                       
{'loss': 0.0747, 'grad_norm': 15.839608192443848, 'learning_rate': 6.0939393939393946e-06, 'num_tokens': 2288413.0, 'completions/mean_length': 35.625, 'completions/min_length': 31.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6005629301071167, 'rewards/meter/std': 0.44296953082084656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9887223839759827, 'rewards/repeat_penalty/std': 0.01224371138960123, 'rewards/repeat_floor/mean': 0.9983083605766296, 'rewards/repeat_floor/std': 0.0018365527503192425, 'rewards/near_duplicate_penalty/mean': 0.9887223839759827, 'rewards/near_duplicate_penalty/std': 0.01224371138960123, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2462134212255478, 'rewards/total_composite/std': 0.1885007917881012, 'reward': 0.2462134212255478, 'reward_std': 0.18850077688694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17666497826576233, 'sampling/sampling_logp_difference/max': 1.56201171875, 'sampling/importance_sampling_ratio/min': 0.20971375703811646, 'sampling/importance_sampling_ratio/mean': 1.0257396697998047, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5847289711236954, 'clip_ratio/low_mean': 0.05239400640130043, 'clip_ratio/low_min': 0.05239400640130043, 'clip_ratio/high_mean': 0.11932100262492895, 'clip_ratio/high_max': 0.11932100262492895, 'clip_ratio/region_mean': 0.17171500902622938, 'reward_total_mean': 0.2462134212255478, 'reward_meter_mean': 0.6005629301071167, 'reward_meter_std': 0.44296953082084656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9887223839759827, 'reward_repeat_penalty_std': 0.01224371138960123, 'reward_repeat_floor_mean': 0.9983083605766296, 'reward_repeat_floor_std': 0.0018365527503192425, 'reward_near_duplicate_penalty_mean': 0.9887223839759827, 'reward_near_duplicate_penalty_std': 0.01224371138960123, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.2462134212255478, 'reward_total_composite_std': 0.1885007917881012, 'epoch': 0.07}

 39%|███▉      | 1290/3300 [4:55:49<6:16:46, 11.25s/it]INFO 04-12 20:54:16 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1291/3300 [4:56:03<6:42:28, 12.02s/it]2026-04-12 20:54:30,021 | INFO | train_grpo_train | metrics_logged mode=train step=1291

                                                       
{'loss': -0.0949, 'grad_norm': 3.454296350479126, 'learning_rate': 6.090909090909092e-06, 'num_tokens': 2289912.0, 'completions/mean_length': 163.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 47.16666793823242, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.538061797618866, 'rewards/meter/std': 0.4162462651729584, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 0.9052866101264954, 'rewards/lexical_plausibility/std': 0.17761211097240448, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.5908900499343872, 'rewards/repeat_penalty/std': 0.19897331297397614, 'rewards/repeat_floor/mean': 0.9611257314682007, 'rewards/repeat_floor/std': 0.019320053979754448, 'rewards/near_duplicate_penalty/mean': 0.9115623235702515, 'rewards/near_duplicate_penalty/std': 0.05392604321241379, 'rewards/opening_diversity/mean': 0.703125, 'rewards/opening_diversity/std': 0.1882425844669342, 'rewards/distinct_2/mean': 0.9133763909339905, 'rewards/distinct_2/std': 0.059238214045763016, 'rewards/total_composite/mean': 0.19371573626995087, 'rewards/total_composite/std': 0.17422915995121002, 'reward': 0.19371573626995087, 'reward_std': 0.17422915995121002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11189450323581696, 'sampling/sampling_logp_difference/max': 1.0207560062408447, 'sampling/importance_sampling_ratio/min': 0.3603224456310272, 'sampling/importance_sampling_ratio/mean': 1.028931736946106, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.761118933558464, 'clip_ratio/low_mean': 0.02691798936575651, 'clip_ratio/low_min': 0.02691798936575651, 'clip_ratio/high_mean': 0.04535717656835914, 'clip_ratio/high_max': 0.04535717656835914, 'clip_ratio/region_mean': 0.07227516593411565, 'reward_total_mean': 0.19371573626995087, 'reward_meter_mean': 0.538061797618866, 'reward_meter_std': 0.4162462651729584, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 0.9052866101264954, 'reward_lexical_plausibility_std': 0.17761211097240448, 'reward_repeat_penalty_mean': 0.5908900499343872, 'reward_repeat_penalty_std': 0.19897331297397614, 'reward_repeat_floor_mean': 0.9611257314682007, 'reward_repeat_floor_std': 0.019320053979754448, 'reward_near_duplicate_penalty_mean': 0.9115623235702515, 'reward_near_duplicate_penalty_std': 0.05392604321241379, 'reward_opening_diversity_mean': 0.703125, 'reward_opening_diversity_std': 0.1882425844669342, 'reward_distinct_2_mean': 0.9133763909339905, 'reward_distinct_2_std': 0.059238214045763016, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.19371573626995087, 'reward_total_composite_std': 0.17422915995121002, 'epoch': 0.07}

 39%|███▉      | 1291/3300 [4:56:03<6:42:28, 12.02s/it]INFO 04-12 20:54:30 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1292/3300 [4:56:17<7:06:01, 12.73s/it]2026-04-12 20:54:44,232 | INFO | train_grpo_train | metrics_logged mode=train step=1292

                                                       
{'loss': -0.0531, 'grad_norm': 5.128190517425537, 'learning_rate': 6.087878787878788e-06, 'num_tokens': 2291342.0, 'completions/mean_length': 97.75, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.57143020629883, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.43891412019729614, 'rewards/meter/std': 0.3075217604637146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9566828608512878, 'rewards/lexical_plausibility/std': 0.12251932919025421, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9951488971710205, 'rewards/repeat_penalty/std': 0.006459453608840704, 'rewards/repeat_floor/mean': 0.999272346496582, 'rewards/repeat_floor/std': 0.0009689141297712922, 'rewards/near_duplicate_penalty/mean': 0.9951488971710205, 'rewards/near_duplicate_penalty/std': 0.006459453608840704, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12514343857765198, 'rewards/total_composite/std': 0.09997636824846268, 'reward': 0.12514343857765198, 'reward_std': 0.09997636079788208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1624237596988678, 'sampling/sampling_logp_difference/max': 1.6810216903686523, 'sampling/importance_sampling_ratio/min': 0.18618366122245789, 'sampling/importance_sampling_ratio/mean': 1.036087155342102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3929873704910278, 'clip_ratio/low_mean': 0.09315388463437557, 'clip_ratio/low_min': 0.09315388463437557, 'clip_ratio/high_mean': 0.06837544031441212, 'clip_ratio/high_max': 0.06837544031441212, 'clip_ratio/region_mean': 0.1615293249487877, 'reward_total_mean': 0.12514343857765198, 'reward_meter_mean': 0.43891412019729614, 'reward_meter_std': 0.3075217604637146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9566828608512878, 'reward_lexical_plausibility_std': 0.12251932919025421, 'reward_repeat_penalty_mean': 0.9951488971710205, 'reward_repeat_penalty_std': 0.006459453608840704, 'reward_repeat_floor_mean': 0.999272346496582, 'reward_repeat_floor_std': 0.0009689141297712922, 'reward_near_duplicate_penalty_mean': 0.9951488971710205, 'reward_near_duplicate_penalty_std': 0.006459453608840704, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.12514343857765198, 'reward_total_composite_std': 0.09997636824846268, 'epoch': 0.07}

 39%|███▉      | 1292/3300 [4:56:18<7:06:01, 12.73s/it]INFO 04-12 20:54:44 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1293/3300 [4:56:32<7:20:09, 13.16s/it]2026-04-12 20:54:58,390 | INFO | train_grpo_train | metrics_logged mode=train step=1293

                                                       
{'loss': -0.1105, 'grad_norm': 2.3477559089660645, 'learning_rate': 6.0848484848484855e-06, 'num_tokens': 2292845.0, 'completions/mean_length': 97.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.71428680419922, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.8492938280105591, 'rewards/meter/std': 0.2699936628341675, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.938922643661499, 'rewards/lexical_plausibility/std': 0.17275281250476837, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9123156070709229, 'rewards/repeat_penalty/std': 0.05167490988969803, 'rewards/repeat_floor/mean': 0.9868473410606384, 'rewards/repeat_floor/std': 0.00775123480707407, 'rewards/near_duplicate_penalty/mean': 0.9123156070709229, 'rewards/near_duplicate_penalty/std': 0.05167490988969803, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33280426263809204, 'rewards/total_composite/std': 0.13991817831993103, 'reward': 0.33280426263809204, 'reward_std': 0.13991817831993103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14128628373146057, 'sampling/sampling_logp_difference/max': 1.5668621063232422, 'sampling/importance_sampling_ratio/min': 0.208699032664299, 'sampling/importance_sampling_ratio/mean': 1.0547375679016113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0181243270635605, 'clip_ratio/low_mean': 0.038194444961845875, 'clip_ratio/low_min': 0.038194444961845875, 'clip_ratio/high_mean': 0.0634755166247487, 'clip_ratio/high_max': 0.0634755166247487, 'clip_ratio/region_mean': 0.10166996158659458, 'reward_total_mean': 0.33280426263809204, 'reward_meter_mean': 0.8492938280105591, 'reward_meter_std': 0.2699936628341675, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.938922643661499, 'reward_lexical_plausibility_std': 0.17275281250476837, 'reward_repeat_penalty_mean': 0.9123156070709229, 'reward_repeat_penalty_std': 0.05167490988969803, 'reward_repeat_floor_mean': 0.9868473410606384, 'reward_repeat_floor_std': 0.00775123480707407, 'reward_near_duplicate_penalty_mean': 0.9123156070709229, 'reward_near_duplicate_penalty_std': 0.05167490988969803, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.33280426263809204, 'reward_total_composite_std': 0.13991817831993103, 'epoch': 0.07}

 39%|███▉      | 1293/3300 [4:56:32<7:20:09, 13.16s/it]INFO 04-12 20:54:58 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1294/3300 [4:56:46<7:27:12, 13.38s/it]2026-04-12 20:55:12,320 | INFO | train_grpo_train | metrics_logged mode=train step=1294

                                                       
{'loss': -0.1314, 'grad_norm': 2.8284456729888916, 'learning_rate': 6.081818181818182e-06, 'num_tokens': 2295149.0, 'completions/mean_length': 202.0, 'completions/min_length': 86.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 98.66667175292969, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.7316447496414185, 'rewards/meter/std': 0.21830470860004425, 'rewards/count_adherence/mean': 0.5833333730697632, 'rewards/count_adherence/std': 0.15430335700511932, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.875, 'rewards/count_floor/std': 0.046290989965200424, 'rewards/lexical_plausibility/mean': 0.9123651385307312, 'rewards/lexical_plausibility/std': 0.16310109198093414, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.9727830290794373, 'rewards/repeat_penalty/std': 0.03844749927520752, 'rewards/repeat_floor/mean': 0.9968199729919434, 'rewards/repeat_floor/std': 0.004119279328733683, 'rewards/near_duplicate_penalty/mean': 0.9884527325630188, 'rewards/near_duplicate_penalty/std': 0.010653919540345669, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839115142822266, 'rewards/distinct_2/std': 0.029867565259337425, 'rewards/total_composite/mean': 0.2000296264886856, 'rewards/total_composite/std': 0.1373065710067749, 'reward': 0.2000296264886856, 'reward_std': 0.1373065710067749, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13939042389392853, 'sampling/sampling_logp_difference/max': 2.2049474716186523, 'sampling/importance_sampling_ratio/min': 0.11025631427764893, 'sampling/importance_sampling_ratio/mean': 1.0134644508361816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7392666414380074, 'clip_ratio/low_mean': 0.030182072892785072, 'clip_ratio/low_min': 0.030182072892785072, 'clip_ratio/high_mean': 0.06121049169450998, 'clip_ratio/high_max': 0.06121049169450998, 'clip_ratio/region_mean': 0.09139256458729506, 'reward_total_mean': 0.2000296264886856, 'reward_meter_mean': 0.7316447496414185, 'reward_meter_std': 0.21830470860004425, 'reward_count_adherence_mean': 0.5833333730697632, 'reward_count_adherence_std': 0.15430335700511932, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.875, 'reward_count_floor_std': 0.046290989965200424, 'reward_lexical_plausibility_mean': 0.9123651385307312, 'reward_lexical_plausibility_std': 0.16310109198093414, 'reward_repeat_penalty_mean': 0.9727830290794373, 'reward_repeat_penalty_std': 0.03844749927520752, 'reward_repeat_floor_mean': 0.9968199729919434, 'reward_repeat_floor_std': 0.004119279328733683, 'reward_near_duplicate_penalty_mean': 0.9884527325630188, 'reward_near_duplicate_penalty_std': 0.010653919540345669, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839115142822266, 'reward_distinct_2_std': 0.029867565259337425, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.2000296264886856, 'reward_total_composite_std': 0.1373065710067749, 'epoch': 0.07}

 39%|███▉      | 1294/3300 [4:56:46<7:27:12, 13.38s/it]INFO 04-12 20:55:12 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1295/3300 [4:57:00<7:35:14, 13.62s/it]2026-04-12 20:55:26,517 | INFO | train_grpo_train | metrics_logged mode=train step=1295

                                                       
{'loss': -0.0312, 'grad_norm': 6.965628147125244, 'learning_rate': 6.07878787878788e-06, 'num_tokens': 2296697.0, 'completions/mean_length': 91.5, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.428573608398438, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.3820682168006897, 'rewards/meter/std': 0.43224993348121643, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9436441659927368, 'rewards/lexical_plausibility/std': 0.1593983769416809, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9394865036010742, 'rewards/repeat_penalty/std': 0.0780569463968277, 'rewards/repeat_floor/mean': 0.9913161396980286, 'rewards/repeat_floor/std': 0.010765504091978073, 'rewards/near_duplicate_penalty/mean': 0.948517918586731, 'rewards/near_duplicate_penalty/std': 0.05753665417432785, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9893162250518799, 'rewards/distinct_2/std': 0.03021823801100254, 'rewards/total_composite/mean': 0.19535186886787415, 'rewards/total_composite/std': 0.25788185000419617, 'reward': 0.19535186886787415, 'reward_std': 0.2578818202018738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16937421262264252, 'sampling/sampling_logp_difference/max': 1.2371962070465088, 'sampling/importance_sampling_ratio/min': 0.3047093152999878, 'sampling/importance_sampling_ratio/mean': 1.0371733903884888, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1494939476251602, 'clip_ratio/low_mean': 0.04092689557000995, 'clip_ratio/low_min': 0.04092689557000995, 'clip_ratio/high_mean': 0.05595238134264946, 'clip_ratio/high_max': 0.05595238134264946, 'clip_ratio/region_mean': 0.0968792769126594, 'reward_total_mean': 0.19535186886787415, 'reward_meter_mean': 0.3820682168006897, 'reward_meter_std': 0.43224993348121643, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9436441659927368, 'reward_lexical_plausibility_std': 0.1593983769416809, 'reward_repeat_penalty_mean': 0.9394865036010742, 'reward_repeat_penalty_std': 0.0780569463968277, 'reward_repeat_floor_mean': 0.9913161396980286, 'reward_repeat_floor_std': 0.010765504091978073, 'reward_near_duplicate_penalty_mean': 0.948517918586731, 'reward_near_duplicate_penalty_std': 0.05753665417432785, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9893162250518799, 'reward_distinct_2_std': 0.03021823801100254, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.19535186886787415, 'reward_total_composite_std': 0.25788185000419617, 'epoch': 0.07}

 39%|███▉      | 1295/3300 [4:57:00<7:35:14, 13.62s/it]INFO 04-12 20:55:26 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1296/3300 [4:57:18<8:18:47, 14.93s/it]2026-04-12 20:55:44,640 | INFO | train_grpo_train | metrics_logged mode=train step=1296

                                                       
{'loss': -0.0967, 'grad_norm': 2.3545382022857666, 'learning_rate': 6.0757575757575755e-06, 'num_tokens': 2298451.0, 'completions/mean_length': 231.25, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 62.79999923706055, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6328991651535034, 'rewards/meter/std': 0.46825525164604187, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.8044552803039551, 'rewards/lexical_plausibility/std': 0.2723645269870758, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.8559375405311584, 'rewards/repeat_penalty/std': 0.17017002403736115, 'rewards/repeat_floor/mean': 0.9816082715988159, 'rewards/repeat_floor/std': 0.01958322897553444, 'rewards/near_duplicate_penalty/mean': 0.9228945970535278, 'rewards/near_duplicate_penalty/std': 0.048808202147483826, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9449899196624756, 'rewards/distinct_2/std': 0.09677664190530777, 'rewards/total_composite/mean': 0.19450627267360687, 'rewards/total_composite/std': 0.19062136113643646, 'reward': 0.19450627267360687, 'reward_std': 0.19062136113643646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1399243324995041, 'sampling/sampling_logp_difference/max': 1.9881031513214111, 'sampling/importance_sampling_ratio/min': 0.13695494830608368, 'sampling/importance_sampling_ratio/mean': 1.0335744619369507, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6328702121973038, 'clip_ratio/low_mean': 0.02330508455634117, 'clip_ratio/low_min': 0.02330508455634117, 'clip_ratio/high_mean': 0.07017860561609268, 'clip_ratio/high_max': 0.07017860561609268, 'clip_ratio/region_mean': 0.09348369017243385, 'reward_total_mean': 0.19450627267360687, 'reward_meter_mean': 0.6328991651535034, 'reward_meter_std': 0.46825525164604187, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.8044552803039551, 'reward_lexical_plausibility_std': 0.2723645269870758, 'reward_repeat_penalty_mean': 0.8559375405311584, 'reward_repeat_penalty_std': 0.17017002403736115, 'reward_repeat_floor_mean': 0.9816082715988159, 'reward_repeat_floor_std': 0.01958322897553444, 'reward_near_duplicate_penalty_mean': 0.9228945970535278, 'reward_near_duplicate_penalty_std': 0.048808202147483826, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9449899196624756, 'reward_distinct_2_std': 0.09677664190530777, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.19450627267360687, 'reward_total_composite_std': 0.19062136113643646, 'epoch': 0.07}

 39%|███▉      | 1296/3300 [4:57:18<8:18:47, 14.93s/it]INFO 04-12 20:55:44 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1297/3300 [4:57:27<7:18:19, 13.13s/it]2026-04-12 20:55:53,392 | INFO | train_grpo_train | metrics_logged mode=train step=1297

                                                       
{'loss': -0.0416, 'grad_norm': 18.243976593017578, 'learning_rate': 6.072727272727274e-06, 'num_tokens': 2299935.0, 'completions/mean_length': 35.5, 'completions/min_length': 19.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.8618534803390503, 'rewards/meter/std': 0.3101614713668823, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.864165186882019, 'rewards/repeat_penalty/std': 0.17676688730716705, 'rewards/repeat_floor/mean': 0.9873805046081543, 'rewards/repeat_floor/std': 0.017130037769675255, 'rewards/near_duplicate_penalty/mean': 0.9669445753097534, 'rewards/near_duplicate_penalty/std': 0.04131477326154709, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9773755669593811, 'rewards/distinct_2/std': 0.06399156153202057, 'rewards/total_composite/mean': 0.39117056131362915, 'rewards/total_composite/std': 0.24060793220996857, 'reward': 0.39117056131362915, 'reward_std': 0.24060791730880737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16465415060520172, 'sampling/sampling_logp_difference/max': 1.4147634506225586, 'sampling/importance_sampling_ratio/min': 0.2746078073978424, 'sampling/importance_sampling_ratio/mean': 1.0316237211227417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2209869027137756, 'clip_ratio/low_mean': 0.08628104254603386, 'clip_ratio/low_min': 0.08628104254603386, 'clip_ratio/high_mean': 0.09904278069734573, 'clip_ratio/high_max': 0.09904278069734573, 'clip_ratio/region_mean': 0.1853238232433796, 'reward_total_mean': 0.39117056131362915, 'reward_meter_mean': 0.8618534803390503, 'reward_meter_std': 0.3101614713668823, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.864165186882019, 'reward_repeat_penalty_std': 0.17676688730716705, 'reward_repeat_floor_mean': 0.9873805046081543, 'reward_repeat_floor_std': 0.017130037769675255, 'reward_near_duplicate_penalty_mean': 0.9669445753097534, 'reward_near_duplicate_penalty_std': 0.04131477326154709, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9773755669593811, 'reward_distinct_2_std': 0.06399156153202057, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.39117056131362915, 'reward_total_composite_std': 0.24060793220996857, 'epoch': 0.07}

 39%|███▉      | 1297/3300 [4:57:27<7:18:19, 13.13s/it]INFO 04-12 20:55:53 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1298/3300 [4:57:36<6:42:17, 12.06s/it]2026-04-12 20:56:02,948 | INFO | train_grpo_train | metrics_logged mode=train step=1298

                                                       
{'loss': 0.0695, 'grad_norm': 11.957763671875, 'learning_rate': 6.06969696969697e-06, 'num_tokens': 2301799.0, 'completions/mean_length': 70.0, 'completions/min_length': 56.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.9721131920814514, 'rewards/meter/std': 0.036159951239824295, 'rewards/count_adherence/mean': 0.6500000357627869, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8949999809265137, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5249999761581421, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9271843433380127, 'rewards/repeat_penalty/std': 0.06422819942235947, 'rewards/repeat_floor/mean': 0.9918217658996582, 'rewards/repeat_floor/std': 0.007283197250217199, 'rewards/near_duplicate_penalty/mean': 0.9757203459739685, 'rewards/near_duplicate_penalty/std': 0.025583982467651367, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9495973587036133, 'rewards/distinct_2/std': 0.04776235297322273, 'rewards/total_composite/mean': 0.4547930359840393, 'rewards/total_composite/std': 0.10195016860961914, 'reward': 0.4547930359840393, 'reward_std': 0.10195016115903854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16802051663398743, 'sampling/sampling_logp_difference/max': 2.2862086296081543, 'sampling/importance_sampling_ratio/min': 0.10165112465620041, 'sampling/importance_sampling_ratio/mean': 1.0065144300460815, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9691305160522461, 'clip_ratio/low_mean': 0.0830877535045147, 'clip_ratio/low_min': 0.0830877535045147, 'clip_ratio/high_mean': 0.0645708478987217, 'clip_ratio/high_max': 0.0645708478987217, 'clip_ratio/region_mean': 0.1476586014032364, 'reward_total_mean': 0.4547930359840393, 'reward_meter_mean': 0.9721131920814514, 'reward_meter_std': 0.036159951239824295, 'reward_count_adherence_mean': 0.6500000357627869, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8949999809265137, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9271843433380127, 'reward_repeat_penalty_std': 0.06422819942235947, 'reward_repeat_floor_mean': 0.9918217658996582, 'reward_repeat_floor_std': 0.007283197250217199, 'reward_near_duplicate_penalty_mean': 0.9757203459739685, 'reward_near_duplicate_penalty_std': 0.025583982467651367, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9495973587036133, 'reward_distinct_2_std': 0.04776235297322273, 'reward_judge_quality_mean': 0.5249999761581421, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.4547930359840393, 'reward_total_composite_std': 0.10195016860961914, 'epoch': 0.07}

 39%|███▉      | 1298/3300 [4:57:36<6:42:17, 12.06s/it]INFO 04-12 20:56:03 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1299/3300 [4:57:46<6:23:47, 11.51s/it]2026-04-12 20:56:13,212 | INFO | train_grpo_train | metrics_logged mode=train step=1299

                                                       
{'loss': -0.0045, 'grad_norm': 9.149087905883789, 'learning_rate': 6.066666666666667e-06, 'num_tokens': 2304051.0, 'completions/mean_length': 94.5, 'completions/min_length': 82.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.5, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.0871487706899643, 'rewards/meter/std': 0.05589567869901657, 'rewards/count_adherence/mean': 0.59375, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8781249523162842, 'rewards/count_floor/std': 0.017359111458063126, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8439352512359619, 'rewards/repeat_penalty/std': 0.11250229924917221, 'rewards/repeat_floor/mean': 0.9819709062576294, 'rewards/repeat_floor/std': 0.014264138415455818, 'rewards/near_duplicate_penalty/mean': 0.9465708136558533, 'rewards/near_duplicate_penalty/std': 0.04803643375635147, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.8939341306686401, 'rewards/distinct_2/std': 0.06730176508426666, 'rewards/total_composite/mean': 0.02984056994318962, 'rewards/total_composite/std': 0.018717072904109955, 'reward': 0.02984056994318962, 'reward_std': 0.018717072904109955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1701647937297821, 'sampling/sampling_logp_difference/max': 4.698284149169922, 'sampling/importance_sampling_ratio/min': 0.009110896848142147, 'sampling/importance_sampling_ratio/mean': 1.019084095954895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.194883018732071, 'clip_ratio/low_mean': 0.04696781188249588, 'clip_ratio/low_min': 0.04696781188249588, 'clip_ratio/high_mean': 0.09871867671608925, 'clip_ratio/high_max': 0.09871867671608925, 'clip_ratio/region_mean': 0.14568648859858513, 'reward_total_mean': 0.02984056994318962, 'reward_meter_mean': 0.0871487706899643, 'reward_meter_std': 0.05589567869901657, 'reward_count_adherence_mean': 0.59375, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8781249523162842, 'reward_count_floor_std': 0.017359111458063126, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8439352512359619, 'reward_repeat_penalty_std': 0.11250229924917221, 'reward_repeat_floor_mean': 0.9819709062576294, 'reward_repeat_floor_std': 0.014264138415455818, 'reward_near_duplicate_penalty_mean': 0.9465708136558533, 'reward_near_duplicate_penalty_std': 0.04803643375635147, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.8939341306686401, 'reward_distinct_2_std': 0.06730176508426666, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.02984056994318962, 'reward_total_composite_std': 0.018717072904109955, 'epoch': 0.07}

 39%|███▉      | 1299/3300 [4:57:47<6:23:47, 11.51s/it]INFO 04-12 20:56:13 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1300/3300 [4:58:01<6:57:31, 12.53s/it]2026-04-12 20:56:28,078 | INFO | train_grpo_train | metrics_logged mode=train step=1300

                                                       
{'loss': -0.114, 'grad_norm': 2.764055013656616, 'learning_rate': 6.063636363636364e-06, 'num_tokens': 2305478.0, 'completions/mean_length': 99.375, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.42857360839844, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8953641653060913, 'rewards/meter/std': 0.1911994367837906, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9293834567070007, 'rewards/lexical_plausibility/std': 0.19973376393318176, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9048247933387756, 'rewards/repeat_penalty/std': 0.12128104269504547, 'rewards/repeat_floor/mean': 0.9914360046386719, 'rewards/repeat_floor/std': 0.008178388699889183, 'rewards/near_duplicate_penalty/mean': 0.9736760854721069, 'rewards/near_duplicate_penalty/std': 0.02217986062169075, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.990384578704834, 'rewards/distinct_2/std': 0.027196412906050682, 'rewards/total_composite/mean': 0.34310245513916016, 'rewards/total_composite/std': 0.14191284775733948, 'reward': 0.34310245513916016, 'reward_std': 0.14191284775733948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14120209217071533, 'sampling/sampling_logp_difference/max': 1.3661503791809082, 'sampling/importance_sampling_ratio/min': 0.2550870478153229, 'sampling/importance_sampling_ratio/mean': 1.0422509908676147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9768808931112289, 'clip_ratio/low_mean': 0.04039634112268686, 'clip_ratio/low_min': 0.04039634112268686, 'clip_ratio/high_mean': 0.05309934401884675, 'clip_ratio/high_max': 0.05309934401884675, 'clip_ratio/region_mean': 0.09349568514153361, 'reward_total_mean': 0.34310245513916016, 'reward_meter_mean': 0.8953641653060913, 'reward_meter_std': 0.1911994367837906, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9293834567070007, 'reward_lexical_plausibility_std': 0.19973376393318176, 'reward_repeat_penalty_mean': 0.9048247933387756, 'reward_repeat_penalty_std': 0.12128104269504547, 'reward_repeat_floor_mean': 0.9914360046386719, 'reward_repeat_floor_std': 0.008178388699889183, 'reward_near_duplicate_penalty_mean': 0.9736760854721069, 'reward_near_duplicate_penalty_std': 0.02217986062169075, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.990384578704834, 'reward_distinct_2_std': 0.027196412906050682, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.34310245513916016, 'reward_total_composite_std': 0.14191284775733948, 'epoch': 0.07}

 39%|███▉      | 1300/3300 [4:58:01<6:57:31, 12.53s/it]INFO 04-12 20:56:28 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:10<00:57,  5.19s/it]

 23%|██▎       | 3/13 [00:16<00:56,  5.68s/it]

 31%|███       | 4/13 [00:28<01:10,  7.80s/it]

 38%|███▊      | 5/13 [00:36<01:04,  8.06s/it]

 46%|████▌     | 6/13 [00:44<00:55,  7.98s/it]

 54%|█████▍    | 7/13 [00:51<00:46,  7.81s/it]

 62%|██████▏   | 8/13 [01:00<00:39,  7.92s/it]

 69%|██████▉   | 9/13 [01:09<00:33,  8.37s/it]

 77%|███████▋  | 10/13 [01:21<00:28,  9.50s/it]

 85%|████████▍ | 11/13 [01:29<00:18,  9.05s/it]

 92%|█████████▏| 12/13 [01:37<00:08,  8.69s/it]

100%|██████████| 13/13 [01:49<00:00,  9.63s/it]2026-04-12 20:58:29,179 | INFO | train_grpo_train | metrics_logged mode=eval step=1300

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 120.9789, 'eval_samples_per_second': 0.86, 'eval_steps_per_second': 0.107, 'eval_num_tokens': 2305478.0, 'eval_completions/mean_length': 109.5576923076923, 'eval_completions/min_length': 33.92307692307692, 'eval_completions/max_length': 296.2307692307692, 'eval_completions/clipped_ratio': 0.038461538461538464, 'eval_completions/mean_terminated_length': 93.45741858849159, 'eval_completions/min_terminated_length': 33.92307692307692, 'eval_completions/max_terminated_length': 196.0, 'eval_rewards/meter/mean': 0.6069046258926392, 'eval_rewards/meter/std': 0.361237400999436, 'eval_rewards/count_adherence/mean': 0.7424028653364915, 'eval_rewards/count_adherence/std': 0.16372321718014204, 'eval_rewards/arabic_clean/mean': 0.9326923076923077, 'eval_rewards/arabic_clean/std': 0.15280617429659918, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9971153827813956, 'eval_rewards/arabic_floor/std': 0.00534126850274893, 'eval_rewards/count_floor/mean': 0.9227208632689256, 'eval_rewards/count_floor/std': 0.04911696968170313, 'eval_rewards/lexical_plausibility/mean': 0.9813547226098868, 'eval_rewards/lexical_plausibility/std': 0.047964142492184274, 'eval_rewards/judge_quality/mean': 0.4107692241668701, 'eval_rewards/judge_quality/std': 0.17742446007636878, 'eval_rewards/repeat_penalty/mean': 0.8255389011823214, 'eval_rewards/repeat_penalty/std': 0.1907717648607034, 'eval_rewards/repeat_floor/mean': 0.9801494341630203, 'eval_rewards/repeat_floor/std': 0.022458405162279423, 'eval_rewards/near_duplicate_penalty/mean': 0.9443351626396179, 'eval_rewards/near_duplicate_penalty/std': 0.05648960918188095, 'eval_rewards/opening_diversity/mean': 0.9528421988854041, 'eval_rewards/opening_diversity/std': 0.087744112771291, 'eval_rewards/distinct_2/mean': 0.9036515125861535, 'eval_rewards/distinct_2/std': 0.1262516829256828, 'eval_rewards/total_composite/mean': 0.23303316189692572, 'eval_rewards/total_composite/std': 0.18125356791111139, 'eval_reward': 0.23303316189692572, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07943717848796111, 'eval_sampling/sampling_logp_difference/max': 1.0771342057448168, 'eval_sampling/importance_sampling_ratio/min': 0.3445100234105037, 'eval_sampling/importance_sampling_ratio/mean': 1.0226388894594634, 'eval_sampling/importance_sampling_ratio/max': 1.5075985926848192, 'eval_entropy': 0.9149978986153235, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.23303316189692572, 'eval_reward_meter_mean': 0.6069046258926392, 'eval_reward_meter_std': 0.361237400999436, 'eval_reward_count_adherence_mean': 0.7424028653364915, 'eval_reward_count_adherence_std': 0.16372321718014204, 'eval_reward_arabic_clean_mean': 0.9326923076923077, 'eval_reward_arabic_clean_std': 0.15280617429659918, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9971153827813956, 'eval_reward_arabic_floor_std': 0.00534126850274893, 'eval_reward_count_floor_mean': 0.9227208632689256, 'eval_reward_count_floor_std': 0.04911696968170313, 'eval_reward_lexical_plausibility_mean': 0.9813547226098868, 'eval_reward_lexical_plausibility_std': 0.047964142492184274, 'eval_reward_repeat_penalty_mean': 0.8255389011823214, 'eval_reward_repeat_penalty_std': 0.1907717648607034, 'eval_reward_repeat_floor_mean': 0.9801494341630203, 'eval_reward_repeat_floor_std': 0.022458405162279423, 'eval_reward_near_duplicate_penalty_mean': 0.9443351626396179, 'eval_reward_near_duplicate_penalty_std': 0.05648960918188095, 'eval_reward_opening_diversity_mean': 0.9528421988854041, 'eval_reward_opening_diversity_std': 0.087744112771291, 'eval_reward_distinct_2_mean': 0.9036515125861535, 'eval_reward_distinct_2_std': 0.1262516829256828, 'eval_reward_judge_quality_mean': 0.4107692241668701, 'eval_reward_judge_quality_std': 0.17742446007636878, 'eval_reward_total_composite_mean': 0.23303316189692572, 'eval_reward_total_composite_std': 0.18125356791111139, 'epoch': 0.07}

 39%|███▉      | 1300/3300 [5:00:03<6:57:31, 12.53s/it]

100%|██████████| 13/13 [01:49<00:00,  9.63s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 39%|███▉      | 1301/3300 [5:00:20<27:57:16, 50.34s/it]2026-04-12 20:58:46,680 | INFO | train_grpo_train | metrics_logged mode=train step=1301

                                                        
{'loss': -0.0484, 'grad_norm': 2.3586647510528564, 'learning_rate': 6.060606060606061e-06, 'num_tokens': 2307009.0, 'completions/mean_length': 221.375, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.056619659066200256, 'rewards/meter/std': 0.06666263937950134, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.8429890871047974, 'rewards/lexical_plausibility/std': 0.21044902503490448, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.1767766922712326, 'rewards/repeat_penalty/mean': 0.7565028071403503, 'rewards/repeat_penalty/std': 0.17413675785064697, 'rewards/repeat_floor/mean': 0.9719423651695251, 'rewards/repeat_floor/std': 0.022083863615989685, 'rewards/near_duplicate_penalty/mean': 0.8957297801971436, 'rewards/near_duplicate_penalty/std': 0.07852613180875778, 'rewards/opening_diversity/mean': 0.893750011920929, 'rewards/opening_diversity/std': 0.1237436905503273, 'rewards/distinct_2/mean': 0.9328654408454895, 'rewards/distinct_2/std': 0.07793726772069931, 'rewards/total_composite/mean': 0.01084198523312807, 'rewards/total_composite/std': 0.015057297423481941, 'reward': 0.01084198523312807, 'reward_std': 0.015057297423481941, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13505558669567108, 'sampling/sampling_logp_difference/max': 1.7290325164794922, 'sampling/importance_sampling_ratio/min': 0.17745602130889893, 'sampling/importance_sampling_ratio/mean': 1.0444562435150146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.941662922501564, 'clip_ratio/low_mean': 0.033517041709274054, 'clip_ratio/low_min': 0.033517041709274054, 'clip_ratio/high_mean': 0.03996829688549042, 'clip_ratio/high_max': 0.03996829688549042, 'clip_ratio/region_mean': 0.07348533859476447, 'reward_total_mean': 0.01084198523312807, 'reward_meter_mean': 0.056619659066200256, 'reward_meter_std': 0.06666263937950134, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.8429890871047974, 'reward_lexical_plausibility_std': 0.21044902503490448, 'reward_repeat_penalty_mean': 0.7565028071403503, 'reward_repeat_penalty_std': 0.17413675785064697, 'reward_repeat_floor_mean': 0.9719423651695251, 'reward_repeat_floor_std': 0.022083863615989685, 'reward_near_duplicate_penalty_mean': 0.8957297801971436, 'reward_near_duplicate_penalty_std': 0.07852613180875778, 'reward_opening_diversity_mean': 0.893750011920929, 'reward_opening_diversity_std': 0.1237436905503273, 'reward_distinct_2_mean': 0.9328654408454895, 'reward_distinct_2_std': 0.07793726772069931, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.1767766922712326, 'reward_total_composite_mean': 0.01084198523312807, 'reward_total_composite_std': 0.015057297423481941, 'epoch': 0.07}

 39%|███▉      | 1301/3300 [5:00:20<27:57:16, 50.34s/it]INFO 04-12 20:58:46 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1302/3300 [5:00:30<21:11:33, 38.19s/it]2026-04-12 20:58:56,486 | INFO | train_grpo_train | metrics_logged mode=train step=1302

                                                        
{'loss': 0.0079, 'grad_norm': 13.128995895385742, 'learning_rate': 6.057575757575757e-06, 'num_tokens': 2308613.0, 'completions/mean_length': 37.5, 'completions/min_length': 26.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.947344958782196, 'rewards/meter/std': 0.06553370505571365, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.7058526277542114, 'rewards/repeat_penalty/std': 0.3451331853866577, 'rewards/repeat_floor/mean': 0.9632654190063477, 'rewards/repeat_floor/std': 0.04565392807126045, 'rewards/near_duplicate_penalty/mean': 0.9015648365020752, 'rewards/near_duplicate_penalty/std': 0.10774897783994675, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.2834733724594116, 'rewards/distinct_2/mean': 0.8808964490890503, 'rewards/distinct_2/std': 0.14804908633232117, 'rewards/total_composite/mean': 0.2578604817390442, 'rewards/total_composite/std': 0.10757430642843246, 'reward': 0.2578604817390442, 'reward_std': 0.10757429152727127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14275914430618286, 'sampling/sampling_logp_difference/max': 1.5708026885986328, 'sampling/importance_sampling_ratio/min': 0.2078782469034195, 'sampling/importance_sampling_ratio/mean': 1.0205974578857422, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0866042152047157, 'clip_ratio/low_mean': 0.02970750373788178, 'clip_ratio/low_min': 0.02970750373788178, 'clip_ratio/high_mean': 0.08215924631804228, 'clip_ratio/high_max': 0.08215924631804228, 'clip_ratio/region_mean': 0.11186675005592406, 'reward_total_mean': 0.2578604817390442, 'reward_meter_mean': 0.947344958782196, 'reward_meter_std': 0.06553370505571365, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7058526277542114, 'reward_repeat_penalty_std': 0.3451331853866577, 'reward_repeat_floor_mean': 0.9632654190063477, 'reward_repeat_floor_std': 0.04565392807126045, 'reward_near_duplicate_penalty_mean': 0.9015648365020752, 'reward_near_duplicate_penalty_std': 0.10774897783994675, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.2834733724594116, 'reward_distinct_2_mean': 0.8808964490890503, 'reward_distinct_2_std': 0.14804908633232117, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.2578604817390442, 'reward_total_composite_std': 0.10757430642843246, 'epoch': 0.07}

 39%|███▉      | 1302/3300 [5:00:30<21:11:33, 38.19s/it]INFO 04-12 20:58:56 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1303/3300 [5:00:40<16:28:13, 29.69s/it]2026-04-12 20:59:06,366 | INFO | train_grpo_train | metrics_logged mode=train step=1303

                                                        
{'loss': -0.0018, 'grad_norm': 15.61508560180664, 'learning_rate': 6.0545454545454555e-06, 'num_tokens': 2310201.0, 'completions/mean_length': 31.5, 'completions/min_length': 23.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.3252526521682739, 'rewards/meter/std': 0.33772918581962585, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.956250011920929, 'rewards/arabic_floor/std': 0.1237436980009079, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.3142679035663605, 'rewards/repeat_penalty/mean': 0.9827235341072083, 'rewards/repeat_penalty/std': 0.014099577441811562, 'rewards/repeat_floor/mean': 0.9974085688591003, 'rewards/repeat_floor/std': 0.0021149360109120607, 'rewards/near_duplicate_penalty/mean': 0.9827235341072083, 'rewards/near_duplicate_penalty/std': 0.014099577441811562, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.10949956625699997, 'rewards/total_composite/std': 0.10760539025068283, 'reward': 0.10949956625699997, 'reward_std': 0.10760539025068283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1416730284690857, 'sampling/sampling_logp_difference/max': 1.2041034698486328, 'sampling/importance_sampling_ratio/min': 0.345889151096344, 'sampling/importance_sampling_ratio/mean': 1.0523443222045898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9178810492157936, 'clip_ratio/low_mean': 0.09333518287166953, 'clip_ratio/low_min': 0.09333518287166953, 'clip_ratio/high_mean': 0.06143673602491617, 'clip_ratio/high_max': 0.06143673602491617, 'clip_ratio/region_mean': 0.1547719188965857, 'reward_total_mean': 0.10949956625699997, 'reward_meter_mean': 0.3252526521682739, 'reward_meter_std': 0.33772918581962585, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.956250011920929, 'reward_arabic_floor_std': 0.1237436980009079, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9827235341072083, 'reward_repeat_penalty_std': 0.014099577441811562, 'reward_repeat_floor_mean': 0.9974085688591003, 'reward_repeat_floor_std': 0.0021149360109120607, 'reward_near_duplicate_penalty_mean': 0.9827235341072083, 'reward_near_duplicate_penalty_std': 0.014099577441811562, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.3142679035663605, 'reward_total_composite_mean': 0.10949956625699997, 'reward_total_composite_std': 0.10760539025068283, 'epoch': 0.07}

 39%|███▉      | 1303/3300 [5:00:40<16:28:13, 29.69s/it]INFO 04-12 20:59:06 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1304/3300 [5:00:48<12:50:53, 23.17s/it]2026-04-12 20:59:14,310 | INFO | train_grpo_train | metrics_logged mode=train step=1304

                                                        
{'loss': 0.0429, 'grad_norm': 24.225555419921875, 'learning_rate': 6.051515151515152e-06, 'num_tokens': 2311558.0, 'completions/mean_length': 16.625, 'completions/min_length': 15.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.9575055837631226, 'rewards/meter/std': 0.05491168797016144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.831250011920929, 'rewards/judge_quality/std': 0.27559998631477356, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7807838320732117, 'rewards/total_composite/std': 0.26158440113067627, 'reward': 0.7807838320732117, 'reward_std': 0.2615843713283539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13659630715847015, 'sampling/sampling_logp_difference/max': 1.3137197494506836, 'sampling/importance_sampling_ratio/min': 0.2688182592391968, 'sampling/importance_sampling_ratio/mean': 1.015814185142517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8701995760202408, 'clip_ratio/low_mean': 0.0363562093116343, 'clip_ratio/low_min': 0.0363562093116343, 'clip_ratio/high_mean': 0.1007046583108604, 'clip_ratio/high_max': 0.1007046583108604, 'clip_ratio/region_mean': 0.1370608676224947, 'reward_total_mean': 0.7807838320732117, 'reward_meter_mean': 0.9575055837631226, 'reward_meter_std': 0.05491168797016144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.831250011920929, 'reward_judge_quality_std': 0.27559998631477356, 'reward_total_composite_mean': 0.7807838320732117, 'reward_total_composite_std': 0.26158440113067627, 'epoch': 0.07}

 40%|███▉      | 1304/3300 [5:00:48<12:50:53, 23.17s/it]INFO 04-12 20:59:14 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1305/3300 [5:00:55<10:15:48, 18.52s/it]2026-04-12 20:59:21,978 | INFO | train_grpo_train | metrics_logged mode=train step=1305

                                                        
{'loss': 0.108, 'grad_norm': 20.345901489257812, 'learning_rate': 6.048484848484849e-06, 'num_tokens': 2312988.0, 'completions/mean_length': 21.75, 'completions/min_length': 19.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.6204198598861694, 'rewards/meter/std': 0.4102761447429657, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6762499809265137, 'rewards/judge_quality/std': 0.3651198446750641, 'rewards/repeat_penalty/mean': 0.7637537717819214, 'rewards/repeat_penalty/std': 0.09457574039697647, 'rewards/repeat_floor/mean': 0.9846832752227783, 'rewards/repeat_floor/std': 0.006749381311237812, 'rewards/near_duplicate_penalty/mean': 0.9969267845153809, 'rewards/near_duplicate_penalty/std': 0.008692282252013683, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9807692170143127, 'rewards/distinct_2/std': 0.054392825812101364, 'rewards/total_composite/mean': 0.4669201970100403, 'rewards/total_composite/std': 0.4180513024330139, 'reward': 0.4669201970100403, 'reward_std': 0.41805127263069153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13564196228981018, 'sampling/sampling_logp_difference/max': 1.8703927993774414, 'sampling/importance_sampling_ratio/min': 0.1540631353855133, 'sampling/importance_sampling_ratio/mean': 1.0397413969039917, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.033564269542694, 'clip_ratio/low_mean': 0.08855090383440256, 'clip_ratio/low_min': 0.08855090383440256, 'clip_ratio/high_mean': 0.025689222384244204, 'clip_ratio/high_max': 0.025689222384244204, 'clip_ratio/region_mean': 0.11424012621864676, 'reward_total_mean': 0.4669201970100403, 'reward_meter_mean': 0.6204198598861694, 'reward_meter_std': 0.4102761447429657, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7637537717819214, 'reward_repeat_penalty_std': 0.09457574039697647, 'reward_repeat_floor_mean': 0.9846832752227783, 'reward_repeat_floor_std': 0.006749381311237812, 'reward_near_duplicate_penalty_mean': 0.9969267845153809, 'reward_near_duplicate_penalty_std': 0.008692282252013683, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9807692170143127, 'reward_distinct_2_std': 0.054392825812101364, 'reward_judge_quality_mean': 0.6762499809265137, 'reward_judge_quality_std': 0.3651198446750641, 'reward_total_composite_mean': 0.4669201970100403, 'reward_total_composite_std': 0.4180513024330139, 'epoch': 0.07}

 40%|███▉      | 1305/3300 [5:00:55<10:15:48, 18.52s/it]INFO 04-12 20:59:22 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1306/3300 [5:01:09<9:31:30, 17.20s/it] 2026-04-12 20:59:36,082 | INFO | train_grpo_train | metrics_logged mode=train step=1306

                                                       
{'loss': -0.1344, 'grad_norm': 2.2445077896118164, 'learning_rate': 6.0454545454545456e-06, 'num_tokens': 2315013.0, 'completions/mean_length': 245.125, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 85.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.6065161228179932, 'rewards/meter/std': 0.27965378761291504, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 0.8701736927032471, 'rewards/lexical_plausibility/std': 0.18445897102355957, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.16569657623767853, 'rewards/repeat_penalty/mean': 0.950911819934845, 'rewards/repeat_penalty/std': 0.015403605066239834, 'rewards/repeat_floor/mean': 0.9935417771339417, 'rewards/repeat_floor/std': 0.0018675546161830425, 'rewards/near_duplicate_penalty/mean': 0.9666640758514404, 'rewards/near_duplicate_penalty/std': 0.013528765179216862, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9838022589683533, 'rewards/distinct_2/std': 0.01680346392095089, 'rewards/total_composite/mean': 0.1783428192138672, 'rewards/total_composite/std': 0.12931343913078308, 'reward': 0.1783428192138672, 'reward_std': 0.1293134242296219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1359337866306305, 'sampling/sampling_logp_difference/max': 1.2560145854949951, 'sampling/importance_sampling_ratio/min': 0.28478676080703735, 'sampling/importance_sampling_ratio/mean': 1.013946294784546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6520774513483047, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0884337117895484, 'clip_ratio/high_max': 0.0884337117895484, 'clip_ratio/region_mean': 0.0884337117895484, 'reward_total_mean': 0.1783428192138672, 'reward_meter_mean': 0.6065161228179932, 'reward_meter_std': 0.27965378761291504, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 0.8701736927032471, 'reward_lexical_plausibility_std': 0.18445897102355957, 'reward_repeat_penalty_mean': 0.950911819934845, 'reward_repeat_penalty_std': 0.015403605066239834, 'reward_repeat_floor_mean': 0.9935417771339417, 'reward_repeat_floor_std': 0.0018675546161830425, 'reward_near_duplicate_penalty_mean': 0.9666640758514404, 'reward_near_duplicate_penalty_std': 0.013528765179216862, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9838022589683533, 'reward_distinct_2_std': 0.01680346392095089, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.16569657623767853, 'reward_total_composite_mean': 0.1783428192138672, 'reward_total_composite_std': 0.12931343913078308, 'epoch': 0.07}

 40%|███▉      | 1306/3300 [5:01:09<9:31:30, 17.20s/it]INFO 04-12 20:59:36 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1307/3300 [5:01:17<7:55:54, 14.33s/it]2026-04-12 20:59:43,717 | INFO | train_grpo_train | metrics_logged mode=train step=1307

                                                       
{'loss': -0.0239, 'grad_norm': 16.9303035736084, 'learning_rate': 6.042424242424243e-06, 'num_tokens': 2316525.0, 'completions/mean_length': 21.0, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9783234000205994, 'rewards/meter/std': 0.021103810518980026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9800419807434082, 'rewards/lexical_plausibility/std': 0.05644970014691353, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/repeat_penalty/mean': 0.7399276494979858, 'rewards/repeat_penalty/std': 0.028488805517554283, 'rewards/repeat_floor/mean': 0.9829855561256409, 'rewards/repeat_floor/std': 0.005697769112884998, 'rewards/near_duplicate_penalty/mean': 0.9865702390670776, 'rewards/near_duplicate_penalty/std': 0.03798507899045944, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.8284487724304199, 'rewards/total_composite/std': 0.16166505217552185, 'reward': 0.8284487724304199, 'reward_std': 0.16166505217552185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14091189205646515, 'sampling/sampling_logp_difference/max': 1.6892740726470947, 'sampling/importance_sampling_ratio/min': 0.18465352058410645, 'sampling/importance_sampling_ratio/mean': 0.9874743819236755, 'sampling/importance_sampling_ratio/max': 1.9850788116455078, 'entropy': 0.8596550449728966, 'clip_ratio/low_mean': 0.01315789483487606, 'clip_ratio/low_min': 0.01315789483487606, 'clip_ratio/high_mean': 0.14212603773921728, 'clip_ratio/high_max': 0.14212603773921728, 'clip_ratio/region_mean': 0.15528393257409334, 'reward_total_mean': 0.8284487724304199, 'reward_meter_mean': 0.9783234000205994, 'reward_meter_std': 0.021103810518980026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9800419807434082, 'reward_lexical_plausibility_std': 0.05644970014691353, 'reward_repeat_penalty_mean': 0.7399276494979858, 'reward_repeat_penalty_std': 0.028488805517554283, 'reward_repeat_floor_mean': 0.9829855561256409, 'reward_repeat_floor_std': 0.005697769112884998, 'reward_near_duplicate_penalty_mean': 0.9865702390670776, 'reward_near_duplicate_penalty_std': 0.03798507899045944, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.8284487724304199, 'reward_total_composite_std': 0.16166505217552185, 'epoch': 0.07}

 40%|███▉      | 1307/3300 [5:01:17<7:55:54, 14.33s/it]INFO 04-12 20:59:43 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1308/3300 [5:01:31<7:51:15, 14.19s/it]2026-04-12 20:59:57,601 | INFO | train_grpo_train | metrics_logged mode=train step=1308

                                                       
{'loss': -0.1655, 'grad_norm': 2.0397679805755615, 'learning_rate': 6.039393939393939e-06, 'num_tokens': 2318476.0, 'completions/mean_length': 190.875, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 83.83333587646484, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9645935297012329, 'rewards/meter/std': 0.02735067345201969, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.971875011920929, 'rewards/count_floor/std': 0.0388161838054657, 'rewards/lexical_plausibility/mean': 0.9290376305580139, 'rewards/lexical_plausibility/std': 0.1357519030570984, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.8682023286819458, 'rewards/repeat_penalty/std': 0.07635921239852905, 'rewards/repeat_floor/mean': 0.9851706027984619, 'rewards/repeat_floor/std': 0.00832895003259182, 'rewards/near_duplicate_penalty/mean': 0.9575604200363159, 'rewards/near_duplicate_penalty/std': 0.024139340966939926, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9059618711471558, 'rewards/distinct_2/std': 0.06573186069726944, 'rewards/total_composite/mean': 0.32616159319877625, 'rewards/total_composite/std': 0.15014615654945374, 'reward': 0.32616159319877625, 'reward_std': 0.15014615654945374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16500526666641235, 'sampling/sampling_logp_difference/max': 1.8273468017578125, 'sampling/importance_sampling_ratio/min': 0.1608397364616394, 'sampling/importance_sampling_ratio/mean': 1.0042880773544312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6830124631524086, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11945360247045755, 'clip_ratio/high_max': 0.11945360247045755, 'clip_ratio/region_mean': 0.11945360247045755, 'reward_total_mean': 0.32616159319877625, 'reward_meter_mean': 0.9645935297012329, 'reward_meter_std': 0.02735067345201969, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.971875011920929, 'reward_count_floor_std': 0.0388161838054657, 'reward_lexical_plausibility_mean': 0.9290376305580139, 'reward_lexical_plausibility_std': 0.1357519030570984, 'reward_repeat_penalty_mean': 0.8682023286819458, 'reward_repeat_penalty_std': 0.07635921239852905, 'reward_repeat_floor_mean': 0.9851706027984619, 'reward_repeat_floor_std': 0.00832895003259182, 'reward_near_duplicate_penalty_mean': 0.9575604200363159, 'reward_near_duplicate_penalty_std': 0.024139340966939926, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9059618711471558, 'reward_distinct_2_std': 0.06573186069726944, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.32616159319877625, 'reward_total_composite_std': 0.15014615654945374, 'epoch': 0.07}

 40%|███▉      | 1308/3300 [5:01:31<7:51:15, 14.19s/it]INFO 04-12 20:59:57 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1309/3300 [5:01:40<7:04:01, 12.78s/it]2026-04-12 21:00:07,212 | INFO | train_grpo_train | metrics_logged mode=train step=1309

                                                       
{'loss': -0.0161, 'grad_norm': 10.594659805297852, 'learning_rate': 6.0363636363636365e-06, 'num_tokens': 2320583.0, 'completions/mean_length': 80.375, 'completions/min_length': 68.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.6064520478248596, 'rewards/meter/std': 0.43351876735687256, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9555724263191223, 'rewards/repeat_penalty/std': 0.058471377938985825, 'rewards/repeat_floor/mean': 0.9941740036010742, 'rewards/repeat_floor/std': 0.00724637508392334, 'rewards/near_duplicate_penalty/mean': 0.9713509678840637, 'rewards/near_duplicate_penalty/std': 0.029811298474669456, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9830142855644226, 'rewards/distinct_2/std': 0.03562766686081886, 'rewards/total_composite/mean': 0.21657785773277283, 'rewards/total_composite/std': 0.15629920363426208, 'reward': 0.21657785773277283, 'reward_std': 0.15629920363426208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15427684783935547, 'sampling/sampling_logp_difference/max': 1.7791357040405273, 'sampling/importance_sampling_ratio/min': 0.1687839776277542, 'sampling/importance_sampling_ratio/mean': 1.006324052810669, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0275729671120644, 'clip_ratio/low_mean': 0.05930265784263611, 'clip_ratio/low_min': 0.05930265784263611, 'clip_ratio/high_mean': 0.08229111041873693, 'clip_ratio/high_max': 0.08229111041873693, 'clip_ratio/region_mean': 0.14159376826137304, 'reward_total_mean': 0.21657785773277283, 'reward_meter_mean': 0.6064520478248596, 'reward_meter_std': 0.43351876735687256, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9555724263191223, 'reward_repeat_penalty_std': 0.058471377938985825, 'reward_repeat_floor_mean': 0.9941740036010742, 'reward_repeat_floor_std': 0.00724637508392334, 'reward_near_duplicate_penalty_mean': 0.9713509678840637, 'reward_near_duplicate_penalty_std': 0.029811298474669456, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9830142855644226, 'reward_distinct_2_std': 0.03562766686081886, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.21657785773277283, 'reward_total_composite_std': 0.15629920363426208, 'epoch': 0.07}

 40%|███▉      | 1309/3300 [5:01:41<7:04:01, 12.78s/it]INFO 04-12 21:00:07 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1310/3300 [5:01:50<6:29:49, 11.75s/it]2026-04-12 21:00:16,439 | INFO | train_grpo_train | metrics_logged mode=train step=1310

                                                       
{'loss': 0.0243, 'grad_norm': 12.423072814941406, 'learning_rate': 6.033333333333335e-06, 'num_tokens': 2322872.0, 'completions/mean_length': 83.125, 'completions/min_length': 74.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.125, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.7969707250595093, 'rewards/meter/std': 0.2704329192638397, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9143716096878052, 'rewards/repeat_penalty/std': 0.055609311908483505, 'rewards/repeat_floor/mean': 0.9902158379554749, 'rewards/repeat_floor/std': 0.00613121222704649, 'rewards/near_duplicate_penalty/mean': 0.9686281085014343, 'rewards/near_duplicate_penalty/std': 0.019577756524086, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9435734748840332, 'rewards/distinct_2/std': 0.04531228914856911, 'rewards/total_composite/mean': 0.29415443539619446, 'rewards/total_composite/std': 0.10011166334152222, 'reward': 0.29415443539619446, 'reward_std': 0.10011165589094162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18388620018959045, 'sampling/sampling_logp_difference/max': 3.9115958213806152, 'sampling/importance_sampling_ratio/min': 0.020008545368909836, 'sampling/importance_sampling_ratio/mean': 1.0047262907028198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9870238676667213, 'clip_ratio/low_mean': 0.05580186937004328, 'clip_ratio/low_min': 0.05580186937004328, 'clip_ratio/high_mean': 0.10272555612027645, 'clip_ratio/high_max': 0.10272555612027645, 'clip_ratio/region_mean': 0.15852742549031973, 'reward_total_mean': 0.29415443539619446, 'reward_meter_mean': 0.7969707250595093, 'reward_meter_std': 0.2704329192638397, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9143716096878052, 'reward_repeat_penalty_std': 0.055609311908483505, 'reward_repeat_floor_mean': 0.9902158379554749, 'reward_repeat_floor_std': 0.00613121222704649, 'reward_near_duplicate_penalty_mean': 0.9686281085014343, 'reward_near_duplicate_penalty_std': 0.019577756524086, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9435734748840332, 'reward_distinct_2_std': 0.04531228914856911, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.29415443539619446, 'reward_total_composite_std': 0.10011166334152222, 'epoch': 0.07}

 40%|███▉      | 1310/3300 [5:01:50<6:29:49, 11.75s/it]INFO 04-12 21:00:16 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1311/3300 [5:01:59<6:08:46, 11.12s/it]2026-04-12 21:00:26,093 | INFO | train_grpo_train | metrics_logged mode=train step=1311

                                                       
{'loss': 0.0398, 'grad_norm': 12.791458129882812, 'learning_rate': 6.030303030303031e-06, 'num_tokens': 2325039.0, 'completions/mean_length': 93.875, 'completions/min_length': 77.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9819897413253784, 'rewards/meter/std': 0.03369574621319771, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8999999761581421, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.718949019908905, 'rewards/repeat_penalty/std': 0.1386510133743286, 'rewards/repeat_floor/mean': 0.969192624092102, 'rewards/repeat_floor/std': 0.015712536871433258, 'rewards/near_duplicate_penalty/mean': 0.9355074167251587, 'rewards/near_duplicate_penalty/std': 0.03240552917122841, 'rewards/opening_diversity/mean': 0.984375, 'rewards/opening_diversity/std': 0.0289318785071373, 'rewards/distinct_2/mean': 0.7755997180938721, 'rewards/distinct_2/std': 0.11071112751960754, 'rewards/total_composite/mean': 0.3096751570701599, 'rewards/total_composite/std': 0.08193899691104889, 'reward': 0.3096751570701599, 'reward_std': 0.08193900436162949, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14177894592285156, 'sampling/sampling_logp_difference/max': 2.0536086559295654, 'sampling/importance_sampling_ratio/min': 0.1282711774110794, 'sampling/importance_sampling_ratio/mean': 1.015848994255066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9214318841695786, 'clip_ratio/low_mean': 0.07762840390205383, 'clip_ratio/low_min': 0.07762840390205383, 'clip_ratio/high_mean': 0.0579518536105752, 'clip_ratio/high_max': 0.0579518536105752, 'clip_ratio/region_mean': 0.13558025751262903, 'reward_total_mean': 0.3096751570701599, 'reward_meter_mean': 0.9819897413253784, 'reward_meter_std': 0.03369574621319771, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8999999761581421, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.718949019908905, 'reward_repeat_penalty_std': 0.1386510133743286, 'reward_repeat_floor_mean': 0.969192624092102, 'reward_repeat_floor_std': 0.015712536871433258, 'reward_near_duplicate_penalty_mean': 0.9355074167251587, 'reward_near_duplicate_penalty_std': 0.03240552917122841, 'reward_opening_diversity_mean': 0.984375, 'reward_opening_diversity_std': 0.0289318785071373, 'reward_distinct_2_mean': 0.7755997180938721, 'reward_distinct_2_std': 0.11071112751960754, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.3096751570701599, 'reward_total_composite_std': 0.08193899691104889, 'epoch': 0.07}

 40%|███▉      | 1311/3300 [5:01:59<6:08:46, 11.12s/it]INFO 04-12 21:00:26 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1312/3300 [5:02:09<5:55:31, 10.73s/it]2026-04-12 21:00:35,915 | INFO | train_grpo_train | metrics_logged mode=train step=1312

                                                       
{'loss': 0.0195, 'grad_norm': 12.449087142944336, 'learning_rate': 6.027272727272728e-06, 'num_tokens': 2327132.0, 'completions/mean_length': 87.625, 'completions/min_length': 72.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.8854688405990601, 'rewards/meter/std': 0.16350716352462769, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.08908706158399582, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.02672613225877285, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.7740010619163513, 'rewards/repeat_penalty/std': 0.23590870201587677, 'rewards/repeat_floor/mean': 0.9746969938278198, 'rewards/repeat_floor/std': 0.02535582147538662, 'rewards/near_duplicate_penalty/mean': 0.9258442521095276, 'rewards/near_duplicate_penalty/std': 0.06613500416278839, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.847656786441803, 'rewards/distinct_2/std': 0.16402672231197357, 'rewards/total_composite/mean': 0.3287997841835022, 'rewards/total_composite/std': 0.07662955671548843, 'reward': 0.3287997841835022, 'reward_std': 0.07662956416606903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13598813116550446, 'sampling/sampling_logp_difference/max': 2.357959747314453, 'sampling/importance_sampling_ratio/min': 0.09461306035518646, 'sampling/importance_sampling_ratio/mean': 1.0301519632339478, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9698840007185936, 'clip_ratio/low_mean': 0.060020530596375465, 'clip_ratio/low_min': 0.060020530596375465, 'clip_ratio/high_mean': 0.050905884243547916, 'clip_ratio/high_max': 0.050905884243547916, 'clip_ratio/region_mean': 0.11092641483992338, 'reward_total_mean': 0.3287997841835022, 'reward_meter_mean': 0.8854688405990601, 'reward_meter_std': 0.16350716352462769, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.08908706158399582, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.02672613225877285, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7740010619163513, 'reward_repeat_penalty_std': 0.23590870201587677, 'reward_repeat_floor_mean': 0.9746969938278198, 'reward_repeat_floor_std': 0.02535582147538662, 'reward_near_duplicate_penalty_mean': 0.9258442521095276, 'reward_near_duplicate_penalty_std': 0.06613500416278839, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.847656786441803, 'reward_distinct_2_std': 0.16402672231197357, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.3287997841835022, 'reward_total_composite_std': 0.07662955671548843, 'epoch': 0.07}

 40%|███▉      | 1312/3300 [5:02:09<5:55:31, 10.73s/it]INFO 04-12 21:00:36 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1313/3300 [5:02:18<5:32:26, 10.04s/it]2026-04-12 21:00:44,331 | INFO | train_grpo_train | metrics_logged mode=train step=1313

                                                       
{'loss': 0.006, 'grad_norm': 14.321449279785156, 'learning_rate': 6.024242424242425e-06, 'num_tokens': 2328826.0, 'completions/mean_length': 37.75, 'completions/min_length': 33.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8645392656326294, 'rewards/meter/std': 0.29772910475730896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.28768783807754517, 'rewards/repeat_penalty/mean': 0.8131360411643982, 'rewards/repeat_penalty/std': 0.2588255703449249, 'rewards/repeat_floor/mean': 0.9789941310882568, 'rewards/repeat_floor/std': 0.02537693828344345, 'rewards/near_duplicate_penalty/mean': 0.9252829551696777, 'rewards/near_duplicate_penalty/std': 0.06041572988033295, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.964046835899353, 'rewards/distinct_2/std': 0.06678803265094757, 'rewards/total_composite/mean': 0.4043280780315399, 'rewards/total_composite/std': 0.29879361391067505, 'reward': 0.4043280780315399, 'reward_std': 0.29879361391067505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17448312044143677, 'sampling/sampling_logp_difference/max': 1.6411969661712646, 'sampling/importance_sampling_ratio/min': 0.19374798238277435, 'sampling/importance_sampling_ratio/mean': 1.0407819747924805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.324341043829918, 'clip_ratio/low_mean': 0.0916575170122087, 'clip_ratio/low_min': 0.0916575170122087, 'clip_ratio/high_mean': 0.024187306873500347, 'clip_ratio/high_max': 0.024187306873500347, 'clip_ratio/region_mean': 0.11584482388570905, 'reward_total_mean': 0.4043280780315399, 'reward_meter_mean': 0.8645392656326294, 'reward_meter_std': 0.29772910475730896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.8131360411643982, 'reward_repeat_penalty_std': 0.2588255703449249, 'reward_repeat_floor_mean': 0.9789941310882568, 'reward_repeat_floor_std': 0.02537693828344345, 'reward_near_duplicate_penalty_mean': 0.9252829551696777, 'reward_near_duplicate_penalty_std': 0.06041572988033295, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.964046835899353, 'reward_distinct_2_std': 0.06678803265094757, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.28768783807754517, 'reward_total_composite_mean': 0.4043280780315399, 'reward_total_composite_std': 0.29879361391067505, 'epoch': 0.07}

 40%|███▉      | 1313/3300 [5:02:18<5:32:26, 10.04s/it]INFO 04-12 21:00:44 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1314/3300 [5:02:28<5:37:33, 10.20s/it]2026-04-12 21:00:54,893 | INFO | train_grpo_train | metrics_logged mode=train step=1314

                                                       
{'loss': 0.0849, 'grad_norm': 11.951732635498047, 'learning_rate': 6.021212121212122e-06, 'num_tokens': 2330931.0, 'completions/mean_length': 73.125, 'completions/min_length': 60.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.05381731688976288, 'rewards/meter/std': 0.07238277792930603, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9381395578384399, 'rewards/repeat_penalty/std': 0.03527694568037987, 'rewards/repeat_floor/mean': 0.9920425415039062, 'rewards/repeat_floor/std': 0.003405706724151969, 'rewards/near_duplicate_penalty/mean': 0.9612358212471008, 'rewards/near_duplicate_penalty/std': 0.01571418158710003, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.976190447807312, 'rewards/distinct_2/std': 0.03930572792887688, 'rewards/total_composite/mean': 0.015479287132620811, 'rewards/total_composite/std': 0.02345142886042595, 'reward': 0.015479287132620811, 'reward_std': 0.02345142886042595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14509731531143188, 'sampling/sampling_logp_difference/max': 3.974477529525757, 'sampling/importance_sampling_ratio/min': 0.01878911629319191, 'sampling/importance_sampling_ratio/mean': 1.0299352407455444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8261949345469475, 'clip_ratio/low_mean': 0.07739626802504063, 'clip_ratio/low_min': 0.07739626802504063, 'clip_ratio/high_mean': 0.024249551817774773, 'clip_ratio/high_max': 0.024249551817774773, 'clip_ratio/region_mean': 0.1016458198428154, 'reward_total_mean': 0.015479287132620811, 'reward_meter_mean': 0.05381731688976288, 'reward_meter_std': 0.07238277792930603, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9381395578384399, 'reward_repeat_penalty_std': 0.03527694568037987, 'reward_repeat_floor_mean': 0.9920425415039062, 'reward_repeat_floor_std': 0.003405706724151969, 'reward_near_duplicate_penalty_mean': 0.9612358212471008, 'reward_near_duplicate_penalty_std': 0.01571418158710003, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.976190447807312, 'reward_distinct_2_std': 0.03930572792887688, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.015479287132620811, 'reward_total_composite_std': 0.02345142886042595, 'epoch': 0.07}

 40%|███▉      | 1314/3300 [5:02:28<5:37:33, 10.20s/it]INFO 04-12 21:00:55 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1315/3300 [5:02:42<6:12:55, 11.27s/it]2026-04-12 21:01:08,687 | INFO | train_grpo_train | metrics_logged mode=train step=1315

                                                       
{'loss': -0.0505, 'grad_norm': 3.983793020248413, 'learning_rate': 6.018181818181818e-06, 'num_tokens': 2332539.0, 'completions/mean_length': 100.0, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.142860412597656, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8745691776275635, 'rewards/meter/std': 0.31833603978157043, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9495176076889038, 'rewards/lexical_plausibility/std': 0.14278587698936462, 'rewards/judge_quality/mean': 0.3712499737739563, 'rewards/judge_quality/std': 0.2622124254703522, 'rewards/repeat_penalty/mean': 0.9335485696792603, 'rewards/repeat_penalty/std': 0.06857750564813614, 'rewards/repeat_floor/mean': 0.9913874268531799, 'rewards/repeat_floor/std': 0.008522238582372665, 'rewards/near_duplicate_penalty/mean': 0.959395170211792, 'rewards/near_duplicate_penalty/std': 0.035574499517679214, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9719797372817993, 'rewards/distinct_2/std': 0.038789909332990646, 'rewards/total_composite/mean': 0.35731732845306396, 'rewards/total_composite/std': 0.26430076360702515, 'reward': 0.35731732845306396, 'reward_std': 0.26430076360702515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14089587330818176, 'sampling/sampling_logp_difference/max': 1.3785748481750488, 'sampling/importance_sampling_ratio/min': 0.2519373595714569, 'sampling/importance_sampling_ratio/mean': 1.0218290090560913, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7774863839149475, 'clip_ratio/low_mean': 0.06027312437072396, 'clip_ratio/low_min': 0.06027312437072396, 'clip_ratio/high_mean': 0.06789530161768198, 'clip_ratio/high_max': 0.06789530161768198, 'clip_ratio/region_mean': 0.12816842598840594, 'reward_total_mean': 0.35731732845306396, 'reward_meter_mean': 0.8745691776275635, 'reward_meter_std': 0.31833603978157043, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9495176076889038, 'reward_lexical_plausibility_std': 0.14278587698936462, 'reward_repeat_penalty_mean': 0.9335485696792603, 'reward_repeat_penalty_std': 0.06857750564813614, 'reward_repeat_floor_mean': 0.9913874268531799, 'reward_repeat_floor_std': 0.008522238582372665, 'reward_near_duplicate_penalty_mean': 0.959395170211792, 'reward_near_duplicate_penalty_std': 0.035574499517679214, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9719797372817993, 'reward_distinct_2_std': 0.038789909332990646, 'reward_judge_quality_mean': 0.3712499737739563, 'reward_judge_quality_std': 0.2622124254703522, 'reward_total_composite_mean': 0.35731732845306396, 'reward_total_composite_std': 0.26430076360702515, 'epoch': 0.07}

 40%|███▉      | 1315/3300 [5:02:42<6:12:55, 11.27s/it]INFO 04-12 21:01:08 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1316/3300 [5:02:50<5:41:14, 10.32s/it]2026-04-12 21:01:16,775 | INFO | train_grpo_train | metrics_logged mode=train step=1316

                                                       
{'loss': -0.0924, 'grad_norm': 13.727326393127441, 'learning_rate': 6.015151515151516e-06, 'num_tokens': 2333996.0, 'completions/mean_length': 35.125, 'completions/min_length': 30.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5148382186889648, 'rewards/meter/std': 0.4152170419692993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9445988535881042, 'rewards/repeat_penalty/std': 0.06967490911483765, 'rewards/repeat_floor/mean': 0.993800699710846, 'rewards/repeat_floor/std': 0.007518926169723272, 'rewards/near_duplicate_penalty/mean': 0.9818398356437683, 'rewards/near_duplicate_penalty/std': 0.020505281165242195, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9613858461380005, 'rewards/distinct_2/std': 0.057023826986551285, 'rewards/total_composite/mean': 0.2743109464645386, 'rewards/total_composite/std': 0.2878408432006836, 'reward': 0.2743109464645386, 'reward_std': 0.2878408432006836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13995783030986786, 'sampling/sampling_logp_difference/max': 1.7937779426574707, 'sampling/importance_sampling_ratio/min': 0.16633060574531555, 'sampling/importance_sampling_ratio/mean': 1.020034670829773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9805162474513054, 'clip_ratio/low_mean': 0.08769271802157164, 'clip_ratio/low_min': 0.08769271802157164, 'clip_ratio/high_mean': 0.05447733961045742, 'clip_ratio/high_max': 0.05447733961045742, 'clip_ratio/region_mean': 0.14217005763202906, 'reward_total_mean': 0.2743109464645386, 'reward_meter_mean': 0.5148382186889648, 'reward_meter_std': 0.4152170419692993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9445988535881042, 'reward_repeat_penalty_std': 0.06967490911483765, 'reward_repeat_floor_mean': 0.993800699710846, 'reward_repeat_floor_std': 0.007518926169723272, 'reward_near_duplicate_penalty_mean': 0.9818398356437683, 'reward_near_duplicate_penalty_std': 0.020505281165242195, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9613858461380005, 'reward_distinct_2_std': 0.057023826986551285, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.2743109464645386, 'reward_total_composite_std': 0.2878408432006836, 'epoch': 0.07}

 40%|███▉      | 1316/3300 [5:02:50<5:41:14, 10.32s/it]INFO 04-12 21:01:17 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1317/3300 [5:03:04<6:15:47, 11.37s/it]2026-04-12 21:01:30,595 | INFO | train_grpo_train | metrics_logged mode=train step=1317

                                                       
{'loss': -0.1153, 'grad_norm': 5.6942853927612305, 'learning_rate': 6.012121212121213e-06, 'num_tokens': 2335949.0, 'completions/mean_length': 118.125, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.857147216796875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.49686211347579956, 'rewards/meter/std': 0.3207494616508484, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9501303434371948, 'rewards/lexical_plausibility/std': 0.14105276763439178, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.24076886475086212, 'rewards/repeat_penalty/mean': 0.922383189201355, 'rewards/repeat_penalty/std': 0.06103139743208885, 'rewards/repeat_floor/mean': 0.9903059005737305, 'rewards/repeat_floor/std': 0.007146673742681742, 'rewards/near_duplicate_penalty/mean': 0.9581903219223022, 'rewards/near_duplicate_penalty/std': 0.026926938444375992, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.961970865726471, 'rewards/distinct_2/std': 0.043403610587120056, 'rewards/total_composite/mean': 0.15692131221294403, 'rewards/total_composite/std': 0.10595479607582092, 'reward': 0.15692131221294403, 'reward_std': 0.10595478862524033, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14740301668643951, 'sampling/sampling_logp_difference/max': 1.5146105289459229, 'sampling/importance_sampling_ratio/min': 0.32130512595176697, 'sampling/importance_sampling_ratio/mean': 1.0112230777740479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6793710291385651, 'clip_ratio/low_mean': 0.05092772003263235, 'clip_ratio/low_min': 0.05092772003263235, 'clip_ratio/high_mean': 0.07314252480864525, 'clip_ratio/high_max': 0.07314252480864525, 'clip_ratio/region_mean': 0.1240702448412776, 'reward_total_mean': 0.15692131221294403, 'reward_meter_mean': 0.49686211347579956, 'reward_meter_std': 0.3207494616508484, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9501303434371948, 'reward_lexical_plausibility_std': 0.14105276763439178, 'reward_repeat_penalty_mean': 0.922383189201355, 'reward_repeat_penalty_std': 0.06103139743208885, 'reward_repeat_floor_mean': 0.9903059005737305, 'reward_repeat_floor_std': 0.007146673742681742, 'reward_near_duplicate_penalty_mean': 0.9581903219223022, 'reward_near_duplicate_penalty_std': 0.026926938444375992, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.961970865726471, 'reward_distinct_2_std': 0.043403610587120056, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.24076886475086212, 'reward_total_composite_mean': 0.15692131221294403, 'reward_total_composite_std': 0.10595479607582092, 'epoch': 0.07}

 40%|███▉      | 1317/3300 [5:03:04<6:15:47, 11.37s/it]INFO 04-12 21:01:30 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1318/3300 [5:03:18<6:41:48, 12.16s/it]2026-04-12 21:01:44,606 | INFO | train_grpo_train | metrics_logged mode=train step=1318

                                                       
{'loss': -0.038, 'grad_norm': 4.971512317657471, 'learning_rate': 6.00909090909091e-06, 'num_tokens': 2337457.0, 'completions/mean_length': 94.5, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.85714340209961, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7518837451934814, 'rewards/meter/std': 0.3673200011253357, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9454420804977417, 'rewards/lexical_plausibility/std': 0.15431299805641174, 'rewards/judge_quality/mean': 0.4462500214576721, 'rewards/judge_quality/std': 0.2855539917945862, 'rewards/repeat_penalty/mean': 0.8880783915519714, 'rewards/repeat_penalty/std': 0.15090906620025635, 'rewards/repeat_floor/mean': 0.9858551025390625, 'rewards/repeat_floor/std': 0.01885424740612507, 'rewards/near_duplicate_penalty/mean': 0.9459102153778076, 'rewards/near_duplicate_penalty/std': 0.06418748199939728, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9329836964607239, 'rewards/distinct_2/std': 0.10612636059522629, 'rewards/total_composite/mean': 0.30501455068588257, 'rewards/total_composite/std': 0.2405243217945099, 'reward': 0.30501455068588257, 'reward_std': 0.2405243217945099, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16135302186012268, 'sampling/sampling_logp_difference/max': 1.5767261981964111, 'sampling/importance_sampling_ratio/min': 0.2066505253314972, 'sampling/importance_sampling_ratio/mean': 1.0379185676574707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9826783165335655, 'clip_ratio/low_mean': 0.04921874962747097, 'clip_ratio/low_min': 0.04921874962747097, 'clip_ratio/high_mean': 0.09913973836228251, 'clip_ratio/high_max': 0.09913973836228251, 'clip_ratio/region_mean': 0.14835848798975348, 'reward_total_mean': 0.30501455068588257, 'reward_meter_mean': 0.7518837451934814, 'reward_meter_std': 0.3673200011253357, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9454420804977417, 'reward_lexical_plausibility_std': 0.15431299805641174, 'reward_repeat_penalty_mean': 0.8880783915519714, 'reward_repeat_penalty_std': 0.15090906620025635, 'reward_repeat_floor_mean': 0.9858551025390625, 'reward_repeat_floor_std': 0.01885424740612507, 'reward_near_duplicate_penalty_mean': 0.9459102153778076, 'reward_near_duplicate_penalty_std': 0.06418748199939728, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9329836964607239, 'reward_distinct_2_std': 0.10612636059522629, 'reward_judge_quality_mean': 0.4462500214576721, 'reward_judge_quality_std': 0.2855539917945862, 'reward_total_composite_mean': 0.30501455068588257, 'reward_total_composite_std': 0.2405243217945099, 'epoch': 0.07}

 40%|███▉      | 1318/3300 [5:03:18<6:41:48, 12.16s/it]INFO 04-12 21:01:44 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1319/3300 [5:03:26<6:05:00, 11.06s/it]2026-04-12 21:01:53,079 | INFO | train_grpo_train | metrics_logged mode=train step=1319

                                                       
{'loss': 0.0385, 'grad_norm': 9.739234924316406, 'learning_rate': 6.0060606060606065e-06, 'num_tokens': 2339079.0, 'completions/mean_length': 46.75, 'completions/min_length': 42.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9708136320114136, 'rewards/meter/std': 0.02970542199909687, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9489890336990356, 'rewards/repeat_penalty/std': 0.0525766983628273, 'rewards/repeat_floor/mean': 0.9932128190994263, 'rewards/repeat_floor/std': 0.006497122347354889, 'rewards/near_duplicate_penalty/mean': 0.965013325214386, 'rewards/near_duplicate_penalty/std': 0.029271913692355156, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9828975200653076, 'rewards/distinct_2/std': 0.031678035855293274, 'rewards/total_composite/mean': 0.4109557867050171, 'rewards/total_composite/std': 0.1414686143398285, 'reward': 0.4109557867050171, 'reward_std': 0.1414686143398285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16364572942256927, 'sampling/sampling_logp_difference/max': 1.5967597961425781, 'sampling/importance_sampling_ratio/min': 0.20255176723003387, 'sampling/importance_sampling_ratio/mean': 1.0076313018798828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9042903259396553, 'clip_ratio/low_mean': 0.07417727634310722, 'clip_ratio/low_min': 0.07417727634310722, 'clip_ratio/high_mean': 0.035674708895385265, 'clip_ratio/high_max': 0.035674708895385265, 'clip_ratio/region_mean': 0.10985198523849249, 'reward_total_mean': 0.4109557867050171, 'reward_meter_mean': 0.9708136320114136, 'reward_meter_std': 0.02970542199909687, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9489890336990356, 'reward_repeat_penalty_std': 0.0525766983628273, 'reward_repeat_floor_mean': 0.9932128190994263, 'reward_repeat_floor_std': 0.006497122347354889, 'reward_near_duplicate_penalty_mean': 0.965013325214386, 'reward_near_duplicate_penalty_std': 0.029271913692355156, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9828975200653076, 'reward_distinct_2_std': 0.031678035855293274, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.4109557867050171, 'reward_total_composite_std': 0.1414686143398285, 'epoch': 0.07}

 40%|███▉      | 1319/3300 [5:03:26<6:05:00, 11.06s/it]INFO 04-12 21:01:53 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1320/3300 [5:03:35<5:39:10, 10.28s/it]2026-04-12 21:02:01,532 | INFO | train_grpo_train | metrics_logged mode=train step=1320

                                                       
{'loss': 0.2271, 'grad_norm': 14.736226081848145, 'learning_rate': 6.003030303030304e-06, 'num_tokens': 2340326.0, 'completions/mean_length': 36.875, 'completions/min_length': 21.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9866637587547302, 'rewards/meter/std': 0.023029297590255737, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5299999713897705, 'rewards/judge_quality/std': 0.24512389302253723, 'rewards/repeat_penalty/mean': 0.8360481262207031, 'rewards/repeat_penalty/std': 0.11681586503982544, 'rewards/repeat_floor/mean': 0.9855867624282837, 'rewards/repeat_floor/std': 0.008560135960578918, 'rewards/near_duplicate_penalty/mean': 0.9689322710037231, 'rewards/near_duplicate_penalty/std': 0.02249053306877613, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.954132080078125, 'rewards/distinct_2/std': 0.054448146373033524, 'rewards/total_composite/mean': 0.47399526834487915, 'rewards/total_composite/std': 0.1828477829694748, 'reward': 0.47399526834487915, 'reward_std': 0.1828477829694748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16722162067890167, 'sampling/sampling_logp_difference/max': 2.1273062229156494, 'sampling/importance_sampling_ratio/min': 0.11915784329175949, 'sampling/importance_sampling_ratio/mean': 1.0354115962982178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5829506516456604, 'clip_ratio/low_mean': 0.09206017386168242, 'clip_ratio/low_min': 0.09206017386168242, 'clip_ratio/high_mean': 0.04840909130871296, 'clip_ratio/high_max': 0.04840909130871296, 'clip_ratio/region_mean': 0.14046926517039537, 'reward_total_mean': 0.47399526834487915, 'reward_meter_mean': 0.9866637587547302, 'reward_meter_std': 0.023029297590255737, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8360481262207031, 'reward_repeat_penalty_std': 0.11681586503982544, 'reward_repeat_floor_mean': 0.9855867624282837, 'reward_repeat_floor_std': 0.008560135960578918, 'reward_near_duplicate_penalty_mean': 0.9689322710037231, 'reward_near_duplicate_penalty_std': 0.02249053306877613, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.954132080078125, 'reward_distinct_2_std': 0.054448146373033524, 'reward_judge_quality_mean': 0.5299999713897705, 'reward_judge_quality_std': 0.24512389302253723, 'reward_total_composite_mean': 0.47399526834487915, 'reward_total_composite_std': 0.1828477829694748, 'epoch': 0.07}

 40%|████      | 1320/3300 [5:03:35<5:39:10, 10.28s/it]INFO 04-12 21:02:01 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1321/3300 [5:03:43<5:20:46,  9.73s/it]2026-04-12 21:02:09,980 | INFO | train_grpo_train | metrics_logged mode=train step=1321

                                                       
{'loss': -0.0097, 'grad_norm': 14.679240226745605, 'learning_rate': 6e-06, 'num_tokens': 2341754.0, 'completions/mean_length': 34.5, 'completions/min_length': 31.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9376527070999146, 'rewards/meter/std': 0.09049329161643982, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.1060660257935524, 'rewards/repeat_penalty/mean': 0.93580162525177, 'rewards/repeat_penalty/std': 0.024152051657438278, 'rewards/repeat_floor/mean': 0.9903702735900879, 'rewards/repeat_floor/std': 0.0036228084936738014, 'rewards/near_duplicate_penalty/mean': 0.93580162525177, 'rewards/near_duplicate_penalty/std': 0.024152051657438278, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4537506103515625, 'rewards/total_composite/std': 0.11392330378293991, 'reward': 0.4537506103515625, 'reward_std': 0.11392328888177872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1428253948688507, 'sampling/sampling_logp_difference/max': 1.610318660736084, 'sampling/importance_sampling_ratio/min': 0.19982393085956573, 'sampling/importance_sampling_ratio/mean': 1.0028667449951172, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6616759896278381, 'clip_ratio/low_mean': 0.12112255487591028, 'clip_ratio/low_min': 0.12112255487591028, 'clip_ratio/high_mean': 0.0032051282469183207, 'clip_ratio/high_max': 0.0032051282469183207, 'clip_ratio/region_mean': 0.1243276831228286, 'reward_total_mean': 0.4537506103515625, 'reward_meter_mean': 0.9376527070999146, 'reward_meter_std': 0.09049329161643982, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.93580162525177, 'reward_repeat_penalty_std': 0.024152051657438278, 'reward_repeat_floor_mean': 0.9903702735900879, 'reward_repeat_floor_std': 0.0036228084936738014, 'reward_near_duplicate_penalty_mean': 0.93580162525177, 'reward_near_duplicate_penalty_std': 0.024152051657438278, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.1060660257935524, 'reward_total_composite_mean': 0.4537506103515625, 'reward_total_composite_std': 0.11392330378293991, 'epoch': 0.07}

 40%|████      | 1321/3300 [5:03:43<5:20:46,  9.73s/it]INFO 04-12 21:02:10 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1322/3300 [5:03:57<5:58:25, 10.87s/it]2026-04-12 21:02:23,522 | INFO | train_grpo_train | metrics_logged mode=train step=1322

                                                       
{'loss': -0.0318, 'grad_norm': 6.467280387878418, 'learning_rate': 5.996969696969697e-06, 'num_tokens': 2343272.0, 'completions/mean_length': 92.75, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.85714340209961, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5087072253227234, 'rewards/meter/std': 0.43609893321990967, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9415919780731201, 'rewards/lexical_plausibility/std': 0.14164769649505615, 'rewards/judge_quality/mean': 0.3712499737739563, 'rewards/judge_quality/std': 0.25670647621154785, 'rewards/repeat_penalty/mean': 0.9200747013092041, 'rewards/repeat_penalty/std': 0.1338854283094406, 'rewards/repeat_floor/mean': 0.9911704063415527, 'rewards/repeat_floor/std': 0.013205713592469692, 'rewards/near_duplicate_penalty/mean': 0.9681156873703003, 'rewards/near_duplicate_penalty/std': 0.03609072044491768, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9758672714233398, 'rewards/distinct_2/std': 0.04480134695768356, 'rewards/total_composite/mean': 0.16547256708145142, 'rewards/total_composite/std': 0.13278472423553467, 'reward': 0.16547256708145142, 'reward_std': 0.13278470933437347, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17435140907764435, 'sampling/sampling_logp_difference/max': 2.1308696269989014, 'sampling/importance_sampling_ratio/min': 0.11873398721218109, 'sampling/importance_sampling_ratio/mean': 0.9970294237136841, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8414795845746994, 'clip_ratio/low_mean': 0.061873641330748796, 'clip_ratio/low_min': 0.061873641330748796, 'clip_ratio/high_mean': 0.06802262924611568, 'clip_ratio/high_max': 0.06802262924611568, 'clip_ratio/region_mean': 0.12989627057686448, 'reward_total_mean': 0.16547256708145142, 'reward_meter_mean': 0.5087072253227234, 'reward_meter_std': 0.43609893321990967, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9415919780731201, 'reward_lexical_plausibility_std': 0.14164769649505615, 'reward_repeat_penalty_mean': 0.9200747013092041, 'reward_repeat_penalty_std': 0.1338854283094406, 'reward_repeat_floor_mean': 0.9911704063415527, 'reward_repeat_floor_std': 0.013205713592469692, 'reward_near_duplicate_penalty_mean': 0.9681156873703003, 'reward_near_duplicate_penalty_std': 0.03609072044491768, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9758672714233398, 'reward_distinct_2_std': 0.04480134695768356, 'reward_judge_quality_mean': 0.3712499737739563, 'reward_judge_quality_std': 0.25670647621154785, 'reward_total_composite_mean': 0.16547256708145142, 'reward_total_composite_std': 0.13278472423553467, 'epoch': 0.07}

 40%|████      | 1322/3300 [5:03:57<5:58:25, 10.87s/it]INFO 04-12 21:02:23 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1323/3300 [5:04:05<5:34:15, 10.14s/it]2026-04-12 21:02:31,977 | INFO | train_grpo_train | metrics_logged mode=train step=1323

                                                       
{'loss': 0.0436, 'grad_norm': 16.48020362854004, 'learning_rate': 5.993939393939394e-06, 'num_tokens': 2344847.0, 'completions/mean_length': 27.875, 'completions/min_length': 26.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.875, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.25278496742248535, 'rewards/meter/std': 0.22370514273643494, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.49125000834465027, 'rewards/judge_quality/std': 0.16092921793460846, 'rewards/repeat_penalty/mean': 0.9452831745147705, 'rewards/repeat_penalty/std': 0.07141169905662537, 'rewards/repeat_floor/mean': 0.9928609132766724, 'rewards/repeat_floor/std': 0.007813921198248863, 'rewards/near_duplicate_penalty/mean': 0.9652265310287476, 'rewards/near_duplicate_penalty/std': 0.021553587168455124, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.12173649668693542, 'rewards/total_composite/std': 0.10125062614679337, 'reward': 0.12173649668693542, 'reward_std': 0.10125061869621277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1733459234237671, 'sampling/sampling_logp_difference/max': 1.4115428924560547, 'sampling/importance_sampling_ratio/min': 0.2437668889760971, 'sampling/importance_sampling_ratio/mean': 1.004873514175415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1955547407269478, 'clip_ratio/low_mean': 0.09232541639357805, 'clip_ratio/low_min': 0.09232541639357805, 'clip_ratio/high_mean': 0.08802655804902315, 'clip_ratio/high_max': 0.08802655804902315, 'clip_ratio/region_mean': 0.1803519744426012, 'reward_total_mean': 0.12173649668693542, 'reward_meter_mean': 0.25278496742248535, 'reward_meter_std': 0.22370514273643494, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9452831745147705, 'reward_repeat_penalty_std': 0.07141169905662537, 'reward_repeat_floor_mean': 0.9928609132766724, 'reward_repeat_floor_std': 0.007813921198248863, 'reward_near_duplicate_penalty_mean': 0.9652265310287476, 'reward_near_duplicate_penalty_std': 0.021553587168455124, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.49125000834465027, 'reward_judge_quality_std': 0.16092921793460846, 'reward_total_composite_mean': 0.12173649668693542, 'reward_total_composite_std': 0.10125062614679337, 'epoch': 0.07}

 40%|████      | 1323/3300 [5:04:05<5:34:15, 10.14s/it]INFO 04-12 21:02:32 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1324/3300 [5:04:14<5:24:30,  9.85s/it]2026-04-12 21:02:41,150 | INFO | train_grpo_train | metrics_logged mode=train step=1324

                                                       
{'loss': 0.0327, 'grad_norm': 14.842118263244629, 'learning_rate': 5.990909090909092e-06, 'num_tokens': 2346405.0, 'completions/mean_length': 36.75, 'completions/min_length': 32.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6028118133544922, 'rewards/meter/std': 0.41412031650543213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.983134925365448, 'rewards/lexical_plausibility/std': 0.037750594317913055, 'rewards/judge_quality/mean': 0.5249999761581421, 'rewards/judge_quality/std': 0.13887302577495575, 'rewards/repeat_penalty/mean': 0.8834500312805176, 'rewards/repeat_penalty/std': 0.14494600892066956, 'rewards/repeat_floor/mean': 0.9892170429229736, 'rewards/repeat_floor/std': 0.012261439114809036, 'rewards/near_duplicate_penalty/mean': 0.9768266677856445, 'rewards/near_duplicate_penalty/std': 0.027350183576345444, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9604780673980713, 'rewards/distinct_2/std': 0.042326003313064575, 'rewards/total_composite/mean': 0.307438462972641, 'rewards/total_composite/std': 0.24083061516284943, 'reward': 0.307438462972641, 'reward_std': 0.24083060026168823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16083471477031708, 'sampling/sampling_logp_difference/max': 2.03851318359375, 'sampling/importance_sampling_ratio/min': 0.13022218644618988, 'sampling/importance_sampling_ratio/mean': 0.9952160120010376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7971574068069458, 'clip_ratio/low_mean': 0.05712006613612175, 'clip_ratio/low_min': 0.05712006613612175, 'clip_ratio/high_mean': 0.08258928637951612, 'clip_ratio/high_max': 0.08258928637951612, 'clip_ratio/region_mean': 0.13970935251563787, 'reward_total_mean': 0.307438462972641, 'reward_meter_mean': 0.6028118133544922, 'reward_meter_std': 0.41412031650543213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.983134925365448, 'reward_lexical_plausibility_std': 0.037750594317913055, 'reward_repeat_penalty_mean': 0.8834500312805176, 'reward_repeat_penalty_std': 0.14494600892066956, 'reward_repeat_floor_mean': 0.9892170429229736, 'reward_repeat_floor_std': 0.012261439114809036, 'reward_near_duplicate_penalty_mean': 0.9768266677856445, 'reward_near_duplicate_penalty_std': 0.027350183576345444, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9604780673980713, 'reward_distinct_2_std': 0.042326003313064575, 'reward_judge_quality_mean': 0.5249999761581421, 'reward_judge_quality_std': 0.13887302577495575, 'reward_total_composite_mean': 0.307438462972641, 'reward_total_composite_std': 0.24083061516284943, 'epoch': 0.07}

 40%|████      | 1324/3300 [5:04:15<5:24:30,  9.85s/it]INFO 04-12 21:02:41 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1325/3300 [5:04:22<5:06:41,  9.32s/it]2026-04-12 21:02:49,215 | INFO | train_grpo_train | metrics_logged mode=train step=1325

                                                       
{'loss': 0.0546, 'grad_norm': 29.41624641418457, 'learning_rate': 5.987878787878788e-06, 'num_tokens': 2347687.0, 'completions/mean_length': 22.25, 'completions/min_length': 19.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9083833694458008, 'rewards/meter/std': 0.16666342318058014, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9574281573295593, 'rewards/lexical_plausibility/std': 0.09763853251934052, 'rewards/judge_quality/mean': 0.8525000214576721, 'rewards/judge_quality/std': 0.20331189036369324, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.7594344615936279, 'rewards/total_composite/std': 0.23033544421195984, 'reward': 0.7594344615936279, 'reward_std': 0.23033544421195984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.136161670088768, 'sampling/sampling_logp_difference/max': 1.8659148216247559, 'sampling/importance_sampling_ratio/min': 0.15475456416606903, 'sampling/importance_sampling_ratio/mean': 0.993350088596344, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8273940756917, 'clip_ratio/low_mean': 0.049761904403567314, 'clip_ratio/low_min': 0.049761904403567314, 'clip_ratio/high_mean': 0.112219940405339, 'clip_ratio/high_max': 0.112219940405339, 'clip_ratio/region_mean': 0.16198184480890632, 'reward_total_mean': 0.7594344615936279, 'reward_meter_mean': 0.9083833694458008, 'reward_meter_std': 0.16666342318058014, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9574281573295593, 'reward_lexical_plausibility_std': 0.09763853251934052, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8525000214576721, 'reward_judge_quality_std': 0.20331189036369324, 'reward_total_composite_mean': 0.7594344615936279, 'reward_total_composite_std': 0.23033544421195984, 'epoch': 0.07}

 40%|████      | 1325/3300 [5:04:23<5:06:41,  9.32s/it]INFO 04-12 21:02:49 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1326/3300 [5:04:36<5:46:21, 10.53s/it]2026-04-12 21:03:02,566 | INFO | train_grpo_train | metrics_logged mode=train step=1326

                                                       
{'loss': -0.0678, 'grad_norm': 5.055856227874756, 'learning_rate': 5.984848484848486e-06, 'num_tokens': 2349200.0, 'completions/mean_length': 99.125, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.142860412597656, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5484107732772827, 'rewards/meter/std': 0.41628241539001465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.942851185798645, 'rewards/lexical_plausibility/std': 0.1616411954164505, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.29364702105522156, 'rewards/repeat_penalty/mean': 0.8851054906845093, 'rewards/repeat_penalty/std': 0.11812806874513626, 'rewards/repeat_floor/mean': 0.9843192100524902, 'rewards/repeat_floor/std': 0.015226820483803749, 'rewards/near_duplicate_penalty/mean': 0.9205446839332581, 'rewards/near_duplicate_penalty/std': 0.062098801136016846, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9581940174102783, 'rewards/distinct_2/std': 0.07944715023040771, 'rewards/total_composite/mean': 0.30883216857910156, 'rewards/total_composite/std': 0.24022093415260315, 'reward': 0.30883216857910156, 'reward_std': 0.24022091925144196, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11366228759288788, 'sampling/sampling_logp_difference/max': 1.4165725708007812, 'sampling/importance_sampling_ratio/min': 0.24254390597343445, 'sampling/importance_sampling_ratio/mean': 1.0237001180648804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6658881455659866, 'clip_ratio/low_mean': 0.028819444123655558, 'clip_ratio/low_min': 0.028819444123655558, 'clip_ratio/high_mean': 0.0543975536711514, 'clip_ratio/high_max': 0.0543975536711514, 'clip_ratio/region_mean': 0.08321699779480696, 'reward_total_mean': 0.30883216857910156, 'reward_meter_mean': 0.5484107732772827, 'reward_meter_std': 0.41628241539001465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.942851185798645, 'reward_lexical_plausibility_std': 0.1616411954164505, 'reward_repeat_penalty_mean': 0.8851054906845093, 'reward_repeat_penalty_std': 0.11812806874513626, 'reward_repeat_floor_mean': 0.9843192100524902, 'reward_repeat_floor_std': 0.015226820483803749, 'reward_near_duplicate_penalty_mean': 0.9205446839332581, 'reward_near_duplicate_penalty_std': 0.062098801136016846, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9581940174102783, 'reward_distinct_2_std': 0.07944715023040771, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.29364702105522156, 'reward_total_composite_mean': 0.30883216857910156, 'reward_total_composite_std': 0.24022093415260315, 'epoch': 0.07}

 40%|████      | 1326/3300 [5:04:36<5:46:21, 10.53s/it]INFO 04-12 21:03:02 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1327/3300 [5:04:45<5:35:23, 10.20s/it]2026-04-12 21:03:12,007 | INFO | train_grpo_train | metrics_logged mode=train step=1327

                                                       
{'loss': -0.0074, 'grad_norm': 11.903730392456055, 'learning_rate': 5.981818181818182e-06, 'num_tokens': 2350940.0, 'completions/mean_length': 39.5, 'completions/min_length': 37.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8935865163803101, 'rewards/meter/std': 0.16214394569396973, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.6039083003997803, 'rewards/repeat_penalty/std': 0.32538071274757385, 'rewards/repeat_floor/mean': 0.9577351212501526, 'rewards/repeat_floor/std': 0.03346175327897072, 'rewards/near_duplicate_penalty/mean': 0.8908302783966064, 'rewards/near_duplicate_penalty/std': 0.06586231291294098, 'rewards/opening_diversity/mean': 0.703125, 'rewards/opening_diversity/std': 0.290761798620224, 'rewards/distinct_2/mean': 0.9102563858032227, 'rewards/distinct_2/std': 0.07952458411455154, 'rewards/total_composite/mean': 0.31969985365867615, 'rewards/total_composite/std': 0.06430788338184357, 'reward': 0.31969985365867615, 'reward_std': 0.06430788338184357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13153253495693207, 'sampling/sampling_logp_difference/max': 1.143632411956787, 'sampling/importance_sampling_ratio/min': 0.3186594247817993, 'sampling/importance_sampling_ratio/mean': 1.0315687656402588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.03305284678936, 'clip_ratio/low_mean': 0.03309078235179186, 'clip_ratio/low_min': 0.03309078235179186, 'clip_ratio/high_mean': 0.08435847330838442, 'clip_ratio/high_max': 0.08435847330838442, 'clip_ratio/region_mean': 0.11744925566017628, 'reward_total_mean': 0.31969985365867615, 'reward_meter_mean': 0.8935865163803101, 'reward_meter_std': 0.16214394569396973, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6039083003997803, 'reward_repeat_penalty_std': 0.32538071274757385, 'reward_repeat_floor_mean': 0.9577351212501526, 'reward_repeat_floor_std': 0.03346175327897072, 'reward_near_duplicate_penalty_mean': 0.8908302783966064, 'reward_near_duplicate_penalty_std': 0.06586231291294098, 'reward_opening_diversity_mean': 0.703125, 'reward_opening_diversity_std': 0.290761798620224, 'reward_distinct_2_mean': 0.9102563858032227, 'reward_distinct_2_std': 0.07952458411455154, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.31969985365867615, 'reward_total_composite_std': 0.06430788338184357, 'epoch': 0.07}

 40%|████      | 1327/3300 [5:04:45<5:35:23, 10.20s/it]INFO 04-12 21:03:12 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1328/3300 [5:04:53<5:09:36,  9.42s/it]2026-04-12 21:03:19,603 | INFO | train_grpo_train | metrics_logged mode=train step=1328

                                                       
{'loss': 0.0451, 'grad_norm': 23.10528564453125, 'learning_rate': 5.978787878787879e-06, 'num_tokens': 2352330.0, 'completions/mean_length': 17.75, 'completions/min_length': 15.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.8451917171478271, 'rewards/meter/std': 0.26264089345932007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.8487499952316284, 'rewards/judge_quality/std': 0.20152544975280762, 'rewards/repeat_penalty/mean': 0.7129127979278564, 'rewards/repeat_penalty/std': 0.05004237964749336, 'rewards/repeat_floor/mean': 0.9787973761558533, 'rewards/repeat_floor/std': 0.007357305381447077, 'rewards/near_duplicate_penalty/mean': 0.9714695811271667, 'rewards/near_duplicate_penalty/std': 0.036223381757736206, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.6928799748420715, 'rewards/total_composite/std': 0.27362367510795593, 'reward': 0.6928799748420715, 'reward_std': 0.27362367510795593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1491914689540863, 'sampling/sampling_logp_difference/max': 1.5200119018554688, 'sampling/importance_sampling_ratio/min': 0.2187092900276184, 'sampling/importance_sampling_ratio/mean': 1.0318156480789185, 'sampling/importance_sampling_ratio/max': 1.6486778259277344, 'entropy': 1.2375606149435043, 'clip_ratio/low_mean': 0.027820777613669634, 'clip_ratio/low_min': 0.027820777613669634, 'clip_ratio/high_mean': 0.08621162362396717, 'clip_ratio/high_max': 0.08621162362396717, 'clip_ratio/region_mean': 0.1140324012376368, 'reward_total_mean': 0.6928799748420715, 'reward_meter_mean': 0.8451917171478271, 'reward_meter_std': 0.26264089345932007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7129127979278564, 'reward_repeat_penalty_std': 0.05004237964749336, 'reward_repeat_floor_mean': 0.9787973761558533, 'reward_repeat_floor_std': 0.007357305381447077, 'reward_near_duplicate_penalty_mean': 0.9714695811271667, 'reward_near_duplicate_penalty_std': 0.036223381757736206, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.8487499952316284, 'reward_judge_quality_std': 0.20152544975280762, 'reward_total_composite_mean': 0.6928799748420715, 'reward_total_composite_std': 0.27362367510795593, 'epoch': 0.07}

 40%|████      | 1328/3300 [5:04:53<5:09:36,  9.42s/it]INFO 04-12 21:03:19 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1329/3300 [5:05:02<5:06:35,  9.33s/it]2026-04-12 21:03:28,731 | INFO | train_grpo_train | metrics_logged mode=train step=1329

                                                       
{'loss': 0.0199, 'grad_norm': 12.892824172973633, 'learning_rate': 5.975757575757576e-06, 'num_tokens': 2354196.0, 'completions/mean_length': 52.25, 'completions/min_length': 47.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5871219635009766, 'rewards/meter/std': 0.3183690905570984, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.9443455338478088, 'rewards/repeat_penalty/std': 0.06944360584020615, 'rewards/repeat_floor/mean': 0.992408275604248, 'rewards/repeat_floor/std': 0.008408610709011555, 'rewards/near_duplicate_penalty/mean': 0.9594216346740723, 'rewards/near_duplicate_penalty/std': 0.03177790343761444, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9832775592803955, 'rewards/distinct_2/std': 0.0472981221973896, 'rewards/total_composite/mean': 0.1908721923828125, 'rewards/total_composite/std': 0.12809525430202484, 'reward': 0.1908721923828125, 'reward_std': 0.12809525430202484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12676312029361725, 'sampling/sampling_logp_difference/max': 1.7793502807617188, 'sampling/importance_sampling_ratio/min': 0.16874775290489197, 'sampling/importance_sampling_ratio/mean': 1.0191175937652588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6296042837202549, 'clip_ratio/low_mean': 0.08223074488341808, 'clip_ratio/low_min': 0.08223074488341808, 'clip_ratio/high_mean': 0.04419934703037143, 'clip_ratio/high_max': 0.04419934703037143, 'clip_ratio/region_mean': 0.1264300919137895, 'reward_total_mean': 0.1908721923828125, 'reward_meter_mean': 0.5871219635009766, 'reward_meter_std': 0.3183690905570984, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9443455338478088, 'reward_repeat_penalty_std': 0.06944360584020615, 'reward_repeat_floor_mean': 0.992408275604248, 'reward_repeat_floor_std': 0.008408610709011555, 'reward_near_duplicate_penalty_mean': 0.9594216346740723, 'reward_near_duplicate_penalty_std': 0.03177790343761444, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9832775592803955, 'reward_distinct_2_std': 0.0472981221973896, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.1908721923828125, 'reward_total_composite_std': 0.12809525430202484, 'epoch': 0.07}

 40%|████      | 1329/3300 [5:05:02<5:06:35,  9.33s/it]INFO 04-12 21:03:28 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1330/3300 [5:05:11<5:02:58,  9.23s/it]2026-04-12 21:03:37,728 | INFO | train_grpo_train | metrics_logged mode=train step=1330

                                                       
{'loss': 0.0165, 'grad_norm': 13.198930740356445, 'learning_rate': 5.972727272727274e-06, 'num_tokens': 2355877.0, 'completions/mean_length': 40.125, 'completions/min_length': 32.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5731260776519775, 'rewards/meter/std': 0.2905130088329315, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18196842074394226, 'rewards/repeat_penalty/mean': 0.9875781536102295, 'rewards/repeat_penalty/std': 0.01274846214801073, 'rewards/repeat_floor/mean': 0.9981367588043213, 'rewards/repeat_floor/std': 0.0019122709054499865, 'rewards/near_duplicate_penalty/mean': 0.9875781536102295, 'rewards/near_duplicate_penalty/std': 0.01274846214801073, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.27391618490219116, 'rewards/total_composite/std': 0.16317319869995117, 'reward': 0.27391618490219116, 'reward_std': 0.16317319869995117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14458508789539337, 'sampling/sampling_logp_difference/max': 1.3310105800628662, 'sampling/importance_sampling_ratio/min': 0.2642101049423218, 'sampling/importance_sampling_ratio/mean': 1.0001013278961182, 'sampling/importance_sampling_ratio/max': 1.8657915592193604, 'entropy': 0.8650990799069405, 'clip_ratio/low_mean': 0.06866229884326458, 'clip_ratio/low_min': 0.06866229884326458, 'clip_ratio/high_mean': 0.09004537668079138, 'clip_ratio/high_max': 0.09004537668079138, 'clip_ratio/region_mean': 0.15870767552405596, 'reward_total_mean': 0.27391618490219116, 'reward_meter_mean': 0.5731260776519775, 'reward_meter_std': 0.2905130088329315, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9875781536102295, 'reward_repeat_penalty_std': 0.01274846214801073, 'reward_repeat_floor_mean': 0.9981367588043213, 'reward_repeat_floor_std': 0.0019122709054499865, 'reward_near_duplicate_penalty_mean': 0.9875781536102295, 'reward_near_duplicate_penalty_std': 0.01274846214801073, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18196842074394226, 'reward_total_composite_mean': 0.27391618490219116, 'reward_total_composite_std': 0.16317319869995117, 'epoch': 0.07}

 40%|████      | 1330/3300 [5:05:11<5:02:58,  9.23s/it]INFO 04-12 21:03:37 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1331/3300 [5:05:25<5:45:22, 10.52s/it]2026-04-12 21:03:51,267 | INFO | train_grpo_train | metrics_logged mode=train step=1331

                                                       
{'loss': -0.0359, 'grad_norm': 5.450880527496338, 'learning_rate': 5.96969696969697e-06, 'num_tokens': 2357417.0, 'completions/mean_length': 99.5, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.57143020629883, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.41248220205307007, 'rewards/meter/std': 0.3949904441833496, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9381199479103088, 'rewards/lexical_plausibility/std': 0.17502321302890778, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.16035674512386322, 'rewards/repeat_penalty/mean': 0.8331549763679504, 'rewards/repeat_penalty/std': 0.2011079639196396, 'rewards/repeat_floor/mean': 0.980330765247345, 'rewards/repeat_floor/std': 0.023782573640346527, 'rewards/near_duplicate_penalty/mean': 0.93040531873703, 'rewards/near_duplicate_penalty/std': 0.077924445271492, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9391106367111206, 'rewards/distinct_2/std': 0.08394287526607513, 'rewards/total_composite/mean': 0.12054440379142761, 'rewards/total_composite/std': 0.14209884405136108, 'reward': 0.12054440379142761, 'reward_std': 0.14209885895252228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15003983676433563, 'sampling/sampling_logp_difference/max': 2.1318397521972656, 'sampling/importance_sampling_ratio/min': 0.11861886084079742, 'sampling/importance_sampling_ratio/mean': 1.0325063467025757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2800438106060028, 'clip_ratio/low_mean': 0.07968073710799217, 'clip_ratio/low_min': 0.07968073710799217, 'clip_ratio/high_mean': 0.04280869476497173, 'clip_ratio/high_max': 0.04280869476497173, 'clip_ratio/region_mean': 0.1224894318729639, 'reward_total_mean': 0.12054440379142761, 'reward_meter_mean': 0.41248220205307007, 'reward_meter_std': 0.3949904441833496, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9381199479103088, 'reward_lexical_plausibility_std': 0.17502321302890778, 'reward_repeat_penalty_mean': 0.8331549763679504, 'reward_repeat_penalty_std': 0.2011079639196396, 'reward_repeat_floor_mean': 0.980330765247345, 'reward_repeat_floor_std': 0.023782573640346527, 'reward_near_duplicate_penalty_mean': 0.93040531873703, 'reward_near_duplicate_penalty_std': 0.077924445271492, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9391106367111206, 'reward_distinct_2_std': 0.08394287526607513, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.16035674512386322, 'reward_total_composite_mean': 0.12054440379142761, 'reward_total_composite_std': 0.14209884405136108, 'epoch': 0.07}

 40%|████      | 1331/3300 [5:05:25<5:45:22, 10.52s/it]INFO 04-12 21:03:51 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1332/3300 [5:05:33<5:29:37, 10.05s/it]2026-04-12 21:04:00,205 | INFO | train_grpo_train | metrics_logged mode=train step=1332

                                                       
{'loss': 0.0256, 'grad_norm': 12.972583770751953, 'learning_rate': 5.966666666666667e-06, 'num_tokens': 2359148.0, 'completions/mean_length': 47.375, 'completions/min_length': 42.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5087057948112488, 'rewards/meter/std': 0.2856754660606384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.35875001549720764, 'rewards/judge_quality/std': 0.2547512948513031, 'rewards/repeat_penalty/mean': 0.9703654050827026, 'rewards/repeat_penalty/std': 0.030687464401125908, 'rewards/repeat_floor/mean': 0.9955548048019409, 'rewards/repeat_floor/std': 0.004603128414601088, 'rewards/near_duplicate_penalty/mean': 0.9703654050827026, 'rewards/near_duplicate_penalty/std': 0.030687464401125908, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19225703179836273, 'rewards/total_composite/std': 0.23409301042556763, 'reward': 0.19225703179836273, 'reward_std': 0.23409299552440643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12546521425247192, 'sampling/sampling_logp_difference/max': 1.9128966331481934, 'sampling/importance_sampling_ratio/min': 0.14765207469463348, 'sampling/importance_sampling_ratio/mean': 1.0062079429626465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7576346695423126, 'clip_ratio/low_mean': 0.06588046811521053, 'clip_ratio/low_min': 0.06588046811521053, 'clip_ratio/high_mean': 0.03418661281466484, 'clip_ratio/high_max': 0.03418661281466484, 'clip_ratio/region_mean': 0.10006708092987537, 'reward_total_mean': 0.19225703179836273, 'reward_meter_mean': 0.5087057948112488, 'reward_meter_std': 0.2856754660606384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9703654050827026, 'reward_repeat_penalty_std': 0.030687464401125908, 'reward_repeat_floor_mean': 0.9955548048019409, 'reward_repeat_floor_std': 0.004603128414601088, 'reward_near_duplicate_penalty_mean': 0.9703654050827026, 'reward_near_duplicate_penalty_std': 0.030687464401125908, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.35875001549720764, 'reward_judge_quality_std': 0.2547512948513031, 'reward_total_composite_mean': 0.19225703179836273, 'reward_total_composite_std': 0.23409301042556763, 'epoch': 0.07}

 40%|████      | 1332/3300 [5:05:34<5:29:37, 10.05s/it]INFO 04-12 21:04:00 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1333/3300 [5:05:41<5:06:30,  9.35s/it]2026-04-12 21:04:07,921 | INFO | train_grpo_train | metrics_logged mode=train step=1333

                                                       
{'loss': 0.0849, 'grad_norm': 19.595592498779297, 'learning_rate': 5.963636363636364e-06, 'num_tokens': 2360587.0, 'completions/mean_length': 19.875, 'completions/min_length': 17.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.631788969039917, 'rewards/meter/std': 0.30732953548431396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8274999856948853, 'rewards/judge_quality/std': 0.27395257353782654, 'rewards/repeat_penalty/mean': 0.7358227968215942, 'rewards/repeat_penalty/std': 0.021537406370043755, 'rewards/repeat_floor/mean': 0.9821645021438599, 'rewards/repeat_floor/std': 0.004307494033128023, 'rewards/near_duplicate_penalty/mean': 0.9810969829559326, 'rewards/near_duplicate_penalty/std': 0.02871653251349926, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5233801603317261, 'rewards/total_composite/std': 0.33242127299308777, 'reward': 0.5233801603317261, 'reward_std': 0.33242127299308777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12777064740657806, 'sampling/sampling_logp_difference/max': 0.9108784198760986, 'sampling/importance_sampling_ratio/min': 0.4021708071231842, 'sampling/importance_sampling_ratio/mean': 0.9895033836364746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0007993504405022, 'clip_ratio/low_mean': 0.016612200532108545, 'clip_ratio/low_min': 0.016612200532108545, 'clip_ratio/high_mean': 0.10523990355432034, 'clip_ratio/high_max': 0.10523990355432034, 'clip_ratio/region_mean': 0.12185210408642888, 'reward_total_mean': 0.5233801603317261, 'reward_meter_mean': 0.631788969039917, 'reward_meter_std': 0.30732953548431396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7358227968215942, 'reward_repeat_penalty_std': 0.021537406370043755, 'reward_repeat_floor_mean': 0.9821645021438599, 'reward_repeat_floor_std': 0.004307494033128023, 'reward_near_duplicate_penalty_mean': 0.9810969829559326, 'reward_near_duplicate_penalty_std': 0.02871653251349926, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8274999856948853, 'reward_judge_quality_std': 0.27395257353782654, 'reward_total_composite_mean': 0.5233801603317261, 'reward_total_composite_std': 0.33242127299308777, 'epoch': 0.07}

 40%|████      | 1333/3300 [5:05:41<5:06:30,  9.35s/it]INFO 04-12 21:04:08 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1334/3300 [5:05:55<5:51:53, 10.74s/it]2026-04-12 21:04:21,907 | INFO | train_grpo_train | metrics_logged mode=train step=1334

                                                       
{'loss': -0.0093, 'grad_norm': 4.876037120819092, 'learning_rate': 5.960606060606061e-06, 'num_tokens': 2362005.0, 'completions/mean_length': 90.25, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 30.000001907348633, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.31388479471206665, 'rewards/meter/std': 0.31333035230636597, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.928798258304596, 'rewards/lexical_plausibility/std': 0.14532391726970673, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.32965078949928284, 'rewards/repeat_penalty/mean': 0.9824309349060059, 'rewards/repeat_penalty/std': 0.024257978424429893, 'rewards/repeat_floor/mean': 0.9973646402359009, 'rewards/repeat_floor/std': 0.003638692433014512, 'rewards/near_duplicate_penalty/mean': 0.9824309349060059, 'rewards/near_duplicate_penalty/std': 0.024257978424429893, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.16911378502845764, 'rewards/total_composite/std': 0.21125462651252747, 'reward': 0.16911378502845764, 'reward_std': 0.21125461161136627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15908081829547882, 'sampling/sampling_logp_difference/max': 2.7745375633239746, 'sampling/importance_sampling_ratio/min': 0.06237831339240074, 'sampling/importance_sampling_ratio/mean': 1.030200481414795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6044005341827869, 'clip_ratio/low_mean': 0.06565580330789089, 'clip_ratio/low_min': 0.06565580330789089, 'clip_ratio/high_mean': 0.04785714251920581, 'clip_ratio/high_max': 0.04785714251920581, 'clip_ratio/region_mean': 0.1135129458270967, 'reward_total_mean': 0.16911378502845764, 'reward_meter_mean': 0.31388479471206665, 'reward_meter_std': 0.31333035230636597, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.928798258304596, 'reward_lexical_plausibility_std': 0.14532391726970673, 'reward_repeat_penalty_mean': 0.9824309349060059, 'reward_repeat_penalty_std': 0.024257978424429893, 'reward_repeat_floor_mean': 0.9973646402359009, 'reward_repeat_floor_std': 0.003638692433014512, 'reward_near_duplicate_penalty_mean': 0.9824309349060059, 'reward_near_duplicate_penalty_std': 0.024257978424429893, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.32965078949928284, 'reward_total_composite_mean': 0.16911378502845764, 'reward_total_composite_std': 0.21125462651252747, 'epoch': 0.07}

 40%|████      | 1334/3300 [5:05:55<5:51:53, 10.74s/it]INFO 04-12 21:04:22 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1335/3300 [5:06:04<5:29:39, 10.07s/it]2026-04-12 21:04:30,400 | INFO | train_grpo_train | metrics_logged mode=train step=1335

                                                       
{'loss': 0.0229, 'grad_norm': 15.11474609375, 'learning_rate': 5.9575757575757575e-06, 'num_tokens': 2363454.0, 'completions/mean_length': 31.125, 'completions/min_length': 30.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.7449922561645508, 'rewards/meter/std': 0.345890611410141, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.8962662220001221, 'rewards/repeat_penalty/std': 0.12483160197734833, 'rewards/repeat_floor/mean': 0.9873902797698975, 'rewards/repeat_floor/std': 0.012200391851365566, 'rewards/near_duplicate_penalty/mean': 0.9420095682144165, 'rewards/near_duplicate_penalty/std': 0.031333182007074356, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9773755669593811, 'rewards/distinct_2/std': 0.041892312467098236, 'rewards/total_composite/mean': 0.2827579975128174, 'rewards/total_composite/std': 0.141471728682518, 'reward': 0.2827579975128174, 'reward_std': 0.1414717137813568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13726504147052765, 'sampling/sampling_logp_difference/max': 1.2310478687286377, 'sampling/importance_sampling_ratio/min': 0.3327532708644867, 'sampling/importance_sampling_ratio/mean': 1.0139583349227905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8659693971276283, 'clip_ratio/low_mean': 0.03214885760098696, 'clip_ratio/low_min': 0.03214885760098696, 'clip_ratio/high_mean': 0.07657701056450605, 'clip_ratio/high_max': 0.07657701056450605, 'clip_ratio/region_mean': 0.10872586816549301, 'reward_total_mean': 0.2827579975128174, 'reward_meter_mean': 0.7449922561645508, 'reward_meter_std': 0.345890611410141, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8962662220001221, 'reward_repeat_penalty_std': 0.12483160197734833, 'reward_repeat_floor_mean': 0.9873902797698975, 'reward_repeat_floor_std': 0.012200391851365566, 'reward_near_duplicate_penalty_mean': 0.9420095682144165, 'reward_near_duplicate_penalty_std': 0.031333182007074356, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9773755669593811, 'reward_distinct_2_std': 0.041892312467098236, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.2827579975128174, 'reward_total_composite_std': 0.141471728682518, 'epoch': 0.07}

 40%|████      | 1335/3300 [5:06:04<5:29:39, 10.07s/it]INFO 04-12 21:04:30 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1336/3300 [5:06:12<5:09:20,  9.45s/it]2026-04-12 21:04:38,412 | INFO | train_grpo_train | metrics_logged mode=train step=1336

                                                       
{'loss': 0.0494, 'grad_norm': 18.25756072998047, 'learning_rate': 5.954545454545455e-06, 'num_tokens': 2364797.0, 'completions/mean_length': 18.875, 'completions/min_length': 17.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.4056404232978821, 'rewards/meter/std': 0.4821144640445709, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9583333134651184, 'rewards/lexical_plausibility/std': 0.07715168595314026, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.3985128402709961, 'rewards/repeat_penalty/mean': 0.7222431898117065, 'rewards/repeat_penalty/std': 0.043818067759275436, 'rewards/repeat_floor/mean': 0.9804101586341858, 'rewards/repeat_floor/std': 0.007357505150139332, 'rewards/near_duplicate_penalty/mean': 0.9790165424346924, 'rewards/near_duplicate_penalty/std': 0.04612409323453903, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9839743375778198, 'rewards/distinct_2/std': 0.045327357947826385, 'rewards/total_composite/mean': 0.27239811420440674, 'rewards/total_composite/std': 0.38402673602104187, 'reward': 0.27239811420440674, 'reward_std': 0.38402673602104187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1581733375787735, 'sampling/sampling_logp_difference/max': 1.3176019191741943, 'sampling/importance_sampling_ratio/min': 0.2677766680717468, 'sampling/importance_sampling_ratio/mean': 1.024709701538086, 'sampling/importance_sampling_ratio/max': 1.932660698890686, 'entropy': 1.444681629538536, 'clip_ratio/low_mean': 0.0983396009542048, 'clip_ratio/low_min': 0.0983396009542048, 'clip_ratio/high_mean': 0.020955882500857115, 'clip_ratio/high_max': 0.020955882500857115, 'clip_ratio/region_mean': 0.11929548345506191, 'reward_total_mean': 0.27239811420440674, 'reward_meter_mean': 0.4056404232978821, 'reward_meter_std': 0.4821144640445709, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9583333134651184, 'reward_lexical_plausibility_std': 0.07715168595314026, 'reward_repeat_penalty_mean': 0.7222431898117065, 'reward_repeat_penalty_std': 0.043818067759275436, 'reward_repeat_floor_mean': 0.9804101586341858, 'reward_repeat_floor_std': 0.007357505150139332, 'reward_near_duplicate_penalty_mean': 0.9790165424346924, 'reward_near_duplicate_penalty_std': 0.04612409323453903, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9839743375778198, 'reward_distinct_2_std': 0.045327357947826385, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.3985128402709961, 'reward_total_composite_mean': 0.27239811420440674, 'reward_total_composite_std': 0.38402673602104187, 'epoch': 0.07}

 40%|████      | 1336/3300 [5:06:12<5:09:20,  9.45s/it]INFO 04-12 21:04:38 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1337/3300 [5:06:20<4:56:10,  9.05s/it]2026-04-12 21:04:46,536 | INFO | train_grpo_train | metrics_logged mode=train step=1337

                                                       
{'loss': 0.0489, 'grad_norm': 19.248550415039062, 'learning_rate': 5.951515151515151e-06, 'num_tokens': 2366363.0, 'completions/mean_length': 29.75, 'completions/min_length': 26.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7595674991607666, 'rewards/meter/std': 0.31441497802734375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9774717092514038, 'rewards/repeat_penalty/std': 0.02531072497367859, 'rewards/repeat_floor/mean': 0.996620774269104, 'rewards/repeat_floor/std': 0.0037966088857501745, 'rewards/near_duplicate_penalty/mean': 0.9774717092514038, 'rewards/near_duplicate_penalty/std': 0.02531072497367859, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.49833065271377563, 'rewards/total_composite/std': 0.3029429018497467, 'reward': 0.49833065271377563, 'reward_std': 0.3029429018497467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13983580470085144, 'sampling/sampling_logp_difference/max': 2.5629851818084717, 'sampling/importance_sampling_ratio/min': 0.0770743116736412, 'sampling/importance_sampling_ratio/mean': 1.0089911222457886, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.654912032186985, 'clip_ratio/low_mean': 0.06771845556795597, 'clip_ratio/low_min': 0.06771845556795597, 'clip_ratio/high_mean': 0.051996528171002865, 'clip_ratio/high_max': 0.051996528171002865, 'clip_ratio/region_mean': 0.11971498373895884, 'reward_total_mean': 0.49833065271377563, 'reward_meter_mean': 0.7595674991607666, 'reward_meter_std': 0.31441497802734375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9774717092514038, 'reward_repeat_penalty_std': 0.02531072497367859, 'reward_repeat_floor_mean': 0.996620774269104, 'reward_repeat_floor_std': 0.0037966088857501745, 'reward_near_duplicate_penalty_mean': 0.9774717092514038, 'reward_near_duplicate_penalty_std': 0.02531072497367859, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.49833065271377563, 'reward_total_composite_std': 0.3029429018497467, 'epoch': 0.07}

 41%|████      | 1337/3300 [5:06:20<4:56:10,  9.05s/it]INFO 04-12 21:04:46 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1338/3300 [5:06:29<4:53:24,  8.97s/it]2026-04-12 21:04:55,337 | INFO | train_grpo_train | metrics_logged mode=train step=1338

                                                       
{'loss': 0.0239, 'grad_norm': 16.78009605407715, 'learning_rate': 5.948484848484849e-06, 'num_tokens': 2367850.0, 'completions/mean_length': 22.875, 'completions/min_length': 21.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6153329610824585, 'rewards/meter/std': 0.4986436665058136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8966874480247498, 'rewards/lexical_plausibility/std': 0.19146236777305603, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.31664496660232544, 'rewards/total_composite/std': 0.303794801235199, 'reward': 0.31664496660232544, 'reward_std': 0.30379483103752136, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.132952019572258, 'sampling/sampling_logp_difference/max': 1.6494522094726562, 'sampling/importance_sampling_ratio/min': 0.1921551376581192, 'sampling/importance_sampling_ratio/mean': 1.018856167793274, 'sampling/importance_sampling_ratio/max': 1.979230523109436, 'entropy': 0.7560481987893581, 'clip_ratio/low_mean': 0.022727273404598236, 'clip_ratio/low_min': 0.022727273404598236, 'clip_ratio/high_mean': 0.0716314953751862, 'clip_ratio/high_max': 0.0716314953751862, 'clip_ratio/region_mean': 0.09435876877978444, 'reward_total_mean': 0.31664496660232544, 'reward_meter_mean': 0.6153329610824585, 'reward_meter_std': 0.4986436665058136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8966874480247498, 'reward_lexical_plausibility_std': 0.19146236777305603, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.31664496660232544, 'reward_total_composite_std': 0.303794801235199, 'epoch': 0.07}

 41%|████      | 1338/3300 [5:06:29<4:53:24,  8.97s/it]INFO 04-12 21:04:55 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1339/3300 [5:06:37<4:52:06,  8.94s/it]2026-04-12 21:05:04,187 | INFO | train_grpo_train | metrics_logged mode=train step=1339

                                                       
{'loss': 0.027, 'grad_norm': 12.262737274169922, 'learning_rate': 5.9454545454545465e-06, 'num_tokens': 2369434.0, 'completions/mean_length': 43.0, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.2623853385448456, 'rewards/meter/std': 0.20935364067554474, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.19272480905056, 'rewards/repeat_penalty/mean': 0.8575524091720581, 'rewards/repeat_penalty/std': 0.14713476598262787, 'rewards/repeat_floor/mean': 0.9816502928733826, 'rewards/repeat_floor/std': 0.01966172643005848, 'rewards/near_duplicate_penalty/mean': 0.9254277944564819, 'rewards/near_duplicate_penalty/std': 0.07353144884109497, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9204013347625732, 'rewards/distinct_2/std': 0.1032467931509018, 'rewards/total_composite/mean': 0.10366801917552948, 'rewards/total_composite/std': 0.11016611754894257, 'reward': 0.10366801917552948, 'reward_std': 0.11016611009836197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14227178692817688, 'sampling/sampling_logp_difference/max': 1.1470770835876465, 'sampling/importance_sampling_ratio/min': 0.3489612936973572, 'sampling/importance_sampling_ratio/mean': 1.0451797246932983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0176433250308037, 'clip_ratio/low_mean': 0.05705008003860712, 'clip_ratio/low_min': 0.05705008003860712, 'clip_ratio/high_mean': 0.04405931103974581, 'clip_ratio/high_max': 0.04405931103974581, 'clip_ratio/region_mean': 0.10110939107835293, 'reward_total_mean': 0.10366801917552948, 'reward_meter_mean': 0.2623853385448456, 'reward_meter_std': 0.20935364067554474, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.8575524091720581, 'reward_repeat_penalty_std': 0.14713476598262787, 'reward_repeat_floor_mean': 0.9816502928733826, 'reward_repeat_floor_std': 0.01966172643005848, 'reward_near_duplicate_penalty_mean': 0.9254277944564819, 'reward_near_duplicate_penalty_std': 0.07353144884109497, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9204013347625732, 'reward_distinct_2_std': 0.1032467931509018, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.19272480905056, 'reward_total_composite_mean': 0.10366801917552948, 'reward_total_composite_std': 0.11016611754894257, 'epoch': 0.07}

 41%|████      | 1339/3300 [5:06:38<4:52:06,  8.94s/it]INFO 04-12 21:05:04 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1340/3300 [5:06:47<4:57:59,  9.12s/it]2026-04-12 21:05:13,742 | INFO | train_grpo_train | metrics_logged mode=train step=1340

                                                       
{'loss': 0.0322, 'grad_norm': 16.645898818969727, 'learning_rate': 5.942424242424243e-06, 'num_tokens': 2370831.0, 'completions/mean_length': 21.625, 'completions/min_length': 20.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8770867586135864, 'rewards/meter/std': 0.15601903200149536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.7244318127632141, 'rewards/repeat_penalty/std': 0.07231773436069489, 'rewards/repeat_floor/mean': 0.9798864126205444, 'rewards/repeat_floor/std': 0.014463555067777634, 'rewards/near_duplicate_penalty/mean': 0.9659091234207153, 'rewards/near_duplicate_penalty/std': 0.09642364084720612, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2758499085903168, 'rewards/total_composite/std': 0.12271127849817276, 'reward': 0.2758499085903168, 'reward_std': 0.12271127104759216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11157049983739853, 'sampling/sampling_logp_difference/max': 1.4268903732299805, 'sampling/importance_sampling_ratio/min': 0.35402795672416687, 'sampling/importance_sampling_ratio/mean': 1.0266693830490112, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0060660168528557, 'clip_ratio/low_mean': 0.03436147281900048, 'clip_ratio/low_min': 0.03436147281900048, 'clip_ratio/high_mean': 0.029051383957266808, 'clip_ratio/high_max': 0.029051383957266808, 'clip_ratio/region_mean': 0.06341285677626729, 'reward_total_mean': 0.2758499085903168, 'reward_meter_mean': 0.8770867586135864, 'reward_meter_std': 0.15601903200149536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7244318127632141, 'reward_repeat_penalty_std': 0.07231773436069489, 'reward_repeat_floor_mean': 0.9798864126205444, 'reward_repeat_floor_std': 0.014463555067777634, 'reward_near_duplicate_penalty_mean': 0.9659091234207153, 'reward_near_duplicate_penalty_std': 0.09642364084720612, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.2758499085903168, 'reward_total_composite_std': 0.12271127849817276, 'epoch': 0.07}

 41%|████      | 1340/3300 [5:06:47<4:57:59,  9.12s/it]INFO 04-12 21:05:14 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1341/3300 [5:06:56<4:54:19,  9.01s/it]2026-04-12 21:05:22,511 | INFO | train_grpo_train | metrics_logged mode=train step=1341

                                                       
{'loss': -0.0235, 'grad_norm': 18.747194290161133, 'learning_rate': 5.93939393939394e-06, 'num_tokens': 2372133.0, 'completions/mean_length': 19.75, 'completions/min_length': 16.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.69395911693573, 'rewards/meter/std': 0.4084942936897278, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.32101404666900635, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2757655084133148, 'rewards/total_composite/std': 0.283772349357605, 'reward': 0.2757655084133148, 'reward_std': 0.2837723195552826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1769968569278717, 'sampling/sampling_logp_difference/max': 1.8524599075317383, 'sampling/importance_sampling_ratio/min': 0.15685085952281952, 'sampling/importance_sampling_ratio/mean': 1.0295871496200562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4551164731383324, 'clip_ratio/low_mean': 0.10098127275705338, 'clip_ratio/low_min': 0.10098127275705338, 'clip_ratio/high_mean': 0.08438108954578638, 'clip_ratio/high_max': 0.08438108954578638, 'clip_ratio/region_mean': 0.18536236230283976, 'reward_total_mean': 0.2757655084133148, 'reward_meter_mean': 0.69395911693573, 'reward_meter_std': 0.4084942936897278, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.32101404666900635, 'reward_total_composite_mean': 0.2757655084133148, 'reward_total_composite_std': 0.283772349357605, 'epoch': 0.07}

 41%|████      | 1341/3300 [5:06:56<4:54:19,  9.01s/it]INFO 04-12 21:05:22 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1342/3300 [5:07:04<4:46:58,  8.79s/it]2026-04-12 21:05:30,784 | INFO | train_grpo_train | metrics_logged mode=train step=1342

                                                       
{'loss': 0.0421, 'grad_norm': 22.94560432434082, 'learning_rate': 5.936363636363637e-06, 'num_tokens': 2373481.0, 'completions/mean_length': 20.5, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.7192360162734985, 'rewards/meter/std': 0.40795576572418213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8970077037811279, 'rewards/lexical_plausibility/std': 0.15017901360988617, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.28400203585624695, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3278992176055908, 'rewards/total_composite/std': 0.26390430331230164, 'reward': 0.3278992176055908, 'reward_std': 0.26390430331230164, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13400386273860931, 'sampling/sampling_logp_difference/max': 1.407149314880371, 'sampling/importance_sampling_ratio/min': 0.24484024941921234, 'sampling/importance_sampling_ratio/mean': 0.990134060382843, 'sampling/importance_sampling_ratio/max': 1.6668946743011475, 'entropy': 0.8336775004863739, 'clip_ratio/low_mean': 0.029051383957266808, 'clip_ratio/low_min': 0.029051383957266808, 'clip_ratio/high_mean': 0.08135981857776642, 'clip_ratio/high_max': 0.08135981857776642, 'clip_ratio/region_mean': 0.11041120253503323, 'reward_total_mean': 0.3278992176055908, 'reward_meter_mean': 0.7192360162734985, 'reward_meter_std': 0.40795576572418213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8970077037811279, 'reward_lexical_plausibility_std': 0.15017901360988617, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.28400203585624695, 'reward_total_composite_mean': 0.3278992176055908, 'reward_total_composite_std': 0.26390430331230164, 'epoch': 0.07}

 41%|████      | 1342/3300 [5:07:04<4:46:58,  8.79s/it]INFO 04-12 21:05:31 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1343/3300 [5:07:13<4:51:47,  8.95s/it]2026-04-12 21:05:40,091 | INFO | train_grpo_train | metrics_logged mode=train step=1343

                                                       
{'loss': -0.2433, 'grad_norm': 28.78751564025879, 'learning_rate': 5.933333333333335e-06, 'num_tokens': 2375007.0, 'completions/mean_length': 23.75, 'completions/min_length': 20.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.1911235898733139, 'rewards/meter/std': 0.31535324454307556, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8687500357627869, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5425000190734863, 'rewards/judge_quality/std': 0.24300207197666168, 'rewards/repeat_penalty/mean': 0.6611138582229614, 'rewards/repeat_penalty/std': 0.1448705792427063, 'rewards/repeat_floor/mean': 0.9642741084098816, 'rewards/repeat_floor/std': 0.02458568662405014, 'rewards/near_duplicate_penalty/mean': 0.8653113842010498, 'rewards/near_duplicate_penalty/std': 0.13877561688423157, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9733599424362183, 'rewards/distinct_2/std': 0.05244094878435135, 'rewards/total_composite/mean': 0.08030793815851212, 'rewards/total_composite/std': 0.13778293132781982, 'reward': 0.08030793815851212, 'reward_std': 0.13778293132781982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11889927834272385, 'sampling/sampling_logp_difference/max': 1.9986169338226318, 'sampling/importance_sampling_ratio/min': 0.13552258908748627, 'sampling/importance_sampling_ratio/mean': 1.0226681232452393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5903017409145832, 'clip_ratio/low_mean': 0.059188548009842634, 'clip_ratio/low_min': 0.059188548009842634, 'clip_ratio/high_mean': 0.048096624203026295, 'clip_ratio/high_max': 0.048096624203026295, 'clip_ratio/region_mean': 0.10728517221286893, 'reward_total_mean': 0.08030793815851212, 'reward_meter_mean': 0.1911235898733139, 'reward_meter_std': 0.31535324454307556, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8687500357627869, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6611138582229614, 'reward_repeat_penalty_std': 0.1448705792427063, 'reward_repeat_floor_mean': 0.9642741084098816, 'reward_repeat_floor_std': 0.02458568662405014, 'reward_near_duplicate_penalty_mean': 0.8653113842010498, 'reward_near_duplicate_penalty_std': 0.13877561688423157, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9733599424362183, 'reward_distinct_2_std': 0.05244094878435135, 'reward_judge_quality_mean': 0.5425000190734863, 'reward_judge_quality_std': 0.24300207197666168, 'reward_total_composite_mean': 0.08030793815851212, 'reward_total_composite_std': 0.13778293132781982, 'epoch': 0.07}

 41%|████      | 1343/3300 [5:07:13<4:51:47,  8.95s/it]INFO 04-12 21:05:40 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1344/3300 [5:07:27<5:35:56, 10.31s/it]2026-04-12 21:05:53,556 | INFO | train_grpo_train | metrics_logged mode=train step=1344

                                                       
{'loss': -0.0507, 'grad_norm': 4.89669132232666, 'learning_rate': 5.93030303030303e-06, 'num_tokens': 2376668.0, 'completions/mean_length': 95.625, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.142860412597656, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.762590765953064, 'rewards/meter/std': 0.39094144105911255, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9368090629577637, 'rewards/lexical_plausibility/std': 0.17873097956180573, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.26221245527267456, 'rewards/repeat_penalty/mean': 0.9478107690811157, 'rewards/repeat_penalty/std': 0.06123568117618561, 'rewards/repeat_floor/mean': 0.9921716451644897, 'rewards/repeat_floor/std': 0.009185356087982655, 'rewards/near_duplicate_penalty/mean': 0.9478107690811157, 'rewards/near_duplicate_penalty/std': 0.06123568117618561, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33013030886650085, 'rewards/total_composite/std': 0.2820541262626648, 'reward': 0.33013030886650085, 'reward_std': 0.2820541262626648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1354329138994217, 'sampling/sampling_logp_difference/max': 1.396352767944336, 'sampling/importance_sampling_ratio/min': 0.24749800562858582, 'sampling/importance_sampling_ratio/mean': 1.0293217897415161, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.849513866007328, 'clip_ratio/low_mean': 0.024766900576651096, 'clip_ratio/low_min': 0.024766900576651096, 'clip_ratio/high_mean': 0.07912216754630208, 'clip_ratio/high_max': 0.07912216754630208, 'clip_ratio/region_mean': 0.10388906812295318, 'reward_total_mean': 0.33013030886650085, 'reward_meter_mean': 0.762590765953064, 'reward_meter_std': 0.39094144105911255, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9368090629577637, 'reward_lexical_plausibility_std': 0.17873097956180573, 'reward_repeat_penalty_mean': 0.9478107690811157, 'reward_repeat_penalty_std': 0.06123568117618561, 'reward_repeat_floor_mean': 0.9921716451644897, 'reward_repeat_floor_std': 0.009185356087982655, 'reward_near_duplicate_penalty_mean': 0.9478107690811157, 'reward_near_duplicate_penalty_std': 0.06123568117618561, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.26221245527267456, 'reward_total_composite_mean': 0.33013030886650085, 'reward_total_composite_std': 0.2820541262626648, 'epoch': 0.07}

 41%|████      | 1344/3300 [5:07:27<5:35:56, 10.31s/it]INFO 04-12 21:05:53 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1345/3300 [5:07:37<5:30:39, 10.15s/it]2026-04-12 21:06:03,350 | INFO | train_grpo_train | metrics_logged mode=train step=1345

                                                       
{'loss': 0.0159, 'grad_norm': 11.656355857849121, 'learning_rate': 5.927272727272728e-06, 'num_tokens': 2378719.0, 'completions/mean_length': 69.375, 'completions/min_length': 50.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.23912623524665833, 'rewards/meter/std': 0.10712283104658127, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9455697536468506, 'rewards/repeat_penalty/std': 0.04058167710900307, 'rewards/repeat_floor/mean': 0.99339759349823, 'rewards/repeat_floor/std': 0.004497841466218233, 'rewards/near_duplicate_penalty/mean': 0.9730124473571777, 'rewards/near_duplicate_penalty/std': 0.016804536804556847, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9716193675994873, 'rewards/distinct_2/std': 0.03252477943897247, 'rewards/total_composite/mean': 0.12076690047979355, 'rewards/total_composite/std': 0.08739064633846283, 'reward': 0.12076690047979355, 'reward_std': 0.08739064633846283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12272132933139801, 'sampling/sampling_logp_difference/max': 2.1556904315948486, 'sampling/importance_sampling_ratio/min': 0.11582319438457489, 'sampling/importance_sampling_ratio/mean': 1.021725058555603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8600974380970001, 'clip_ratio/low_mean': 0.08951093722134829, 'clip_ratio/low_min': 0.08951093722134829, 'clip_ratio/high_mean': 0.03435477614402771, 'clip_ratio/high_max': 0.03435477614402771, 'clip_ratio/region_mean': 0.123865713365376, 'reward_total_mean': 0.12076690047979355, 'reward_meter_mean': 0.23912623524665833, 'reward_meter_std': 0.10712283104658127, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9455697536468506, 'reward_repeat_penalty_std': 0.04058167710900307, 'reward_repeat_floor_mean': 0.99339759349823, 'reward_repeat_floor_std': 0.004497841466218233, 'reward_near_duplicate_penalty_mean': 0.9730124473571777, 'reward_near_duplicate_penalty_std': 0.016804536804556847, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9716193675994873, 'reward_distinct_2_std': 0.03252477943897247, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.12076690047979355, 'reward_total_composite_std': 0.08739064633846283, 'epoch': 0.07}

 41%|████      | 1345/3300 [5:07:37<5:30:39, 10.15s/it]INFO 04-12 21:06:03 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1346/3300 [5:07:46<5:23:13,  9.92s/it]2026-04-12 21:06:12,747 | INFO | train_grpo_train | metrics_logged mode=train step=1346

                                                       
{'loss': 0.1058, 'grad_norm': 11.502700805664062, 'learning_rate': 5.924242424242425e-06, 'num_tokens': 2380523.0, 'completions/mean_length': 52.5, 'completions/min_length': 47.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5352987051010132, 'rewards/meter/std': 0.39597681164741516, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5037499666213989, 'rewards/judge_quality/std': 0.20063738524913788, 'rewards/repeat_penalty/mean': 0.9632002115249634, 'rewards/repeat_penalty/std': 0.035562869161367416, 'rewards/repeat_floor/mean': 0.9944800138473511, 'rewards/repeat_floor/std': 0.005334425717592239, 'rewards/near_duplicate_penalty/mean': 0.9632002115249634, 'rewards/near_duplicate_penalty/std': 0.035562869161367416, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24508020281791687, 'rewards/total_composite/std': 0.24223540723323822, 'reward': 0.24508020281791687, 'reward_std': 0.24223540723323822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1574449986219406, 'sampling/sampling_logp_difference/max': 1.6950253248214722, 'sampling/importance_sampling_ratio/min': 0.18359458446502686, 'sampling/importance_sampling_ratio/mean': 1.0045570135116577, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9134382382035255, 'clip_ratio/low_mean': 0.06826232187449932, 'clip_ratio/low_min': 0.06826232187449932, 'clip_ratio/high_mean': 0.06593528389930725, 'clip_ratio/high_max': 0.06593528389930725, 'clip_ratio/region_mean': 0.13419760577380657, 'reward_total_mean': 0.24508020281791687, 'reward_meter_mean': 0.5352987051010132, 'reward_meter_std': 0.39597681164741516, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9632002115249634, 'reward_repeat_penalty_std': 0.035562869161367416, 'reward_repeat_floor_mean': 0.9944800138473511, 'reward_repeat_floor_std': 0.005334425717592239, 'reward_near_duplicate_penalty_mean': 0.9632002115249634, 'reward_near_duplicate_penalty_std': 0.035562869161367416, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5037499666213989, 'reward_judge_quality_std': 0.20063738524913788, 'reward_total_composite_mean': 0.24508020281791687, 'reward_total_composite_std': 0.24223540723323822, 'epoch': 0.07}

 41%|████      | 1346/3300 [5:07:46<5:23:13,  9.92s/it]INFO 04-12 21:06:13 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1347/3300 [5:07:56<5:23:26,  9.94s/it]2026-04-12 21:06:22,718 | INFO | train_grpo_train | metrics_logged mode=train step=1347

                                                       
{'loss': -0.0308, 'grad_norm': 15.962719917297363, 'learning_rate': 5.921212121212122e-06, 'num_tokens': 2382024.0, 'completions/mean_length': 34.625, 'completions/min_length': 30.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.34192168712615967, 'rewards/meter/std': 0.4041561782360077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.8611999154090881, 'rewards/repeat_penalty/std': 0.1596219688653946, 'rewards/repeat_floor/mean': 0.9850544929504395, 'rewards/repeat_floor/std': 0.016116490587592125, 'rewards/near_duplicate_penalty/mean': 0.9493511319160461, 'rewards/near_duplicate_penalty/std': 0.051018115133047104, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9600202441215515, 'rewards/distinct_2/std': 0.0427606962621212, 'rewards/total_composite/mean': 0.15181896090507507, 'rewards/total_composite/std': 0.17981407046318054, 'reward': 0.15181896090507507, 'reward_std': 0.17981407046318054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13988859951496124, 'sampling/sampling_logp_difference/max': 1.478764533996582, 'sampling/importance_sampling_ratio/min': 0.2279191017150879, 'sampling/importance_sampling_ratio/mean': 1.0210132598876953, 'sampling/importance_sampling_ratio/max': 1.8616483211517334, 'entropy': 0.812345951795578, 'clip_ratio/low_mean': 0.06189163215458393, 'clip_ratio/low_min': 0.06189163215458393, 'clip_ratio/high_mean': 0.05940402578562498, 'clip_ratio/high_max': 0.05940402578562498, 'clip_ratio/region_mean': 0.12129565794020891, 'reward_total_mean': 0.15181896090507507, 'reward_meter_mean': 0.34192168712615967, 'reward_meter_std': 0.4041561782360077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8611999154090881, 'reward_repeat_penalty_std': 0.1596219688653946, 'reward_repeat_floor_mean': 0.9850544929504395, 'reward_repeat_floor_std': 0.016116490587592125, 'reward_near_duplicate_penalty_mean': 0.9493511319160461, 'reward_near_duplicate_penalty_std': 0.051018115133047104, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9600202441215515, 'reward_distinct_2_std': 0.0427606962621212, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.15181896090507507, 'reward_total_composite_std': 0.17981407046318054, 'epoch': 0.07}

 41%|████      | 1347/3300 [5:07:56<5:23:26,  9.94s/it]INFO 04-12 21:06:22 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1348/3300 [5:08:06<5:20:05,  9.84s/it]2026-04-12 21:06:32,328 | INFO | train_grpo_train | metrics_logged mode=train step=1348

                                                       
{'loss': -0.0222, 'grad_norm': 8.237163543701172, 'learning_rate': 5.9181818181818184e-06, 'num_tokens': 2384064.0, 'completions/mean_length': 81.0, 'completions/min_length': 71.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.4922606945037842, 'rewards/meter/std': 0.28914427757263184, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.6950436234474182, 'rewards/repeat_penalty/std': 0.15762493014335632, 'rewards/repeat_floor/mean': 0.9687133431434631, 'rewards/repeat_floor/std': 0.014708437956869602, 'rewards/near_duplicate_penalty/mean': 0.9244886636734009, 'rewards/near_duplicate_penalty/std': 0.026336971670389175, 'rewards/opening_diversity/mean': 0.8459821343421936, 'rewards/opening_diversity/std': 0.13788791000843048, 'rewards/distinct_2/mean': 0.8809012174606323, 'rewards/distinct_2/std': 0.056934669613838196, 'rewards/total_composite/mean': 0.15719106793403625, 'rewards/total_composite/std': 0.09392759203910828, 'reward': 0.15719106793403625, 'reward_std': 0.09392759203910828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09304961562156677, 'sampling/sampling_logp_difference/max': 1.6117572784423828, 'sampling/importance_sampling_ratio/min': 0.19953665137290955, 'sampling/importance_sampling_ratio/mean': 1.0139282941818237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5484872162342072, 'clip_ratio/low_mean': 0.04864370170980692, 'clip_ratio/low_min': 0.04864370170980692, 'clip_ratio/high_mean': 0.04072190914303064, 'clip_ratio/high_max': 0.04072190914303064, 'clip_ratio/region_mean': 0.08936561085283756, 'reward_total_mean': 0.15719106793403625, 'reward_meter_mean': 0.4922606945037842, 'reward_meter_std': 0.28914427757263184, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6950436234474182, 'reward_repeat_penalty_std': 0.15762493014335632, 'reward_repeat_floor_mean': 0.9687133431434631, 'reward_repeat_floor_std': 0.014708437956869602, 'reward_near_duplicate_penalty_mean': 0.9244886636734009, 'reward_near_duplicate_penalty_std': 0.026336971670389175, 'reward_opening_diversity_mean': 0.8459821343421936, 'reward_opening_diversity_std': 0.13788791000843048, 'reward_distinct_2_mean': 0.8809012174606323, 'reward_distinct_2_std': 0.056934669613838196, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.15719106793403625, 'reward_total_composite_std': 0.09392759203910828, 'epoch': 0.07}

 41%|████      | 1348/3300 [5:08:06<5:20:05,  9.84s/it]INFO 04-12 21:06:32 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1349/3300 [5:08:15<5:20:20,  9.85s/it]2026-04-12 21:06:42,207 | INFO | train_grpo_train | metrics_logged mode=train step=1349

                                                       
{'loss': 0.0481, 'grad_norm': 18.330005645751953, 'learning_rate': 5.915151515151516e-06, 'num_tokens': 2386218.0, 'completions/mean_length': 77.25, 'completions/min_length': 57.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.25, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.7401382923126221, 'rewards/meter/std': 0.29889345169067383, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.0862581729888916, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.918749988079071, 'rewards/count_floor/std': 0.025877466425299644, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.8779439926147461, 'rewards/repeat_penalty/std': 0.14439068734645844, 'rewards/repeat_floor/mean': 0.9890173673629761, 'rewards/repeat_floor/std': 0.01079222746193409, 'rewards/near_duplicate_penalty/mean': 0.9661540389060974, 'rewards/near_duplicate_penalty/std': 0.018562816083431244, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9395886659622192, 'rewards/distinct_2/std': 0.09003165364265442, 'rewards/total_composite/mean': 0.23585084080696106, 'rewards/total_composite/std': 0.09582413733005524, 'reward': 0.23585084080696106, 'reward_std': 0.09582413733005524, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1618015468120575, 'sampling/sampling_logp_difference/max': 2.6795687675476074, 'sampling/importance_sampling_ratio/min': 0.06859272718429565, 'sampling/importance_sampling_ratio/mean': 0.9965869784355164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7429374642670155, 'clip_ratio/low_mean': 0.036150677129626274, 'clip_ratio/low_min': 0.036150677129626274, 'clip_ratio/high_mean': 0.0889024892821908, 'clip_ratio/high_max': 0.0889024892821908, 'clip_ratio/region_mean': 0.12505316641181707, 'reward_total_mean': 0.23585084080696106, 'reward_meter_mean': 0.7401382923126221, 'reward_meter_std': 0.29889345169067383, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.0862581729888916, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.918749988079071, 'reward_count_floor_std': 0.025877466425299644, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8779439926147461, 'reward_repeat_penalty_std': 0.14439068734645844, 'reward_repeat_floor_mean': 0.9890173673629761, 'reward_repeat_floor_std': 0.01079222746193409, 'reward_near_duplicate_penalty_mean': 0.9661540389060974, 'reward_near_duplicate_penalty_std': 0.018562816083431244, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9395886659622192, 'reward_distinct_2_std': 0.09003165364265442, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.23585084080696106, 'reward_total_composite_std': 0.09582413733005524, 'epoch': 0.07}

 41%|████      | 1349/3300 [5:08:16<5:20:20,  9.85s/it]INFO 04-12 21:06:42 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1350/3300 [5:08:29<5:55:54, 10.95s/it]2026-04-12 21:06:55,731 | INFO | train_grpo_train | metrics_logged mode=train step=1350

                                                       
{'loss': -0.1168, 'grad_norm': 4.394704341888428, 'learning_rate': 5.912121212121212e-06, 'num_tokens': 2387843.0, 'completions/mean_length': 133.125, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.7351717352867126, 'rewards/meter/std': 0.2567211985588074, 'rewards/count_adherence/mean': 0.7000000476837158, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9099999666213989, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.9479739665985107, 'rewards/lexical_plausibility/std': 0.14715181291103363, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.16850179433822632, 'rewards/repeat_penalty/mean': 0.6231215000152588, 'rewards/repeat_penalty/std': 0.1780499666929245, 'rewards/repeat_floor/mean': 0.9563164114952087, 'rewards/repeat_floor/std': 0.021974630653858185, 'rewards/near_duplicate_penalty/mean': 0.8819572925567627, 'rewards/near_duplicate_penalty/std': 0.06823018938302994, 'rewards/opening_diversity/mean': 0.9821428656578064, 'rewards/opening_diversity/std': 0.05050762742757797, 'rewards/distinct_2/mean': 0.723268985748291, 'rewards/distinct_2/std': 0.1439959555864334, 'rewards/total_composite/mean': 0.19434238970279694, 'rewards/total_composite/std': 0.1278955191373825, 'reward': 0.19434238970279694, 'reward_std': 0.1278955042362213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15208876132965088, 'sampling/sampling_logp_difference/max': 1.3241920471191406, 'sampling/importance_sampling_ratio/min': 0.2660177946090698, 'sampling/importance_sampling_ratio/mean': 1.0494314432144165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2887413948774338, 'clip_ratio/low_mean': 0.0809740275144577, 'clip_ratio/low_min': 0.0809740275144577, 'clip_ratio/high_mean': 0.031374077312648296, 'clip_ratio/high_max': 0.031374077312648296, 'clip_ratio/region_mean': 0.112348104827106, 'reward_total_mean': 0.19434238970279694, 'reward_meter_mean': 0.7351717352867126, 'reward_meter_std': 0.2567211985588074, 'reward_count_adherence_mean': 0.7000000476837158, 'reward_count_adherence_std': 0.10690449178218842, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9099999666213989, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.9479739665985107, 'reward_lexical_plausibility_std': 0.14715181291103363, 'reward_repeat_penalty_mean': 0.6231215000152588, 'reward_repeat_penalty_std': 0.1780499666929245, 'reward_repeat_floor_mean': 0.9563164114952087, 'reward_repeat_floor_std': 0.021974630653858185, 'reward_near_duplicate_penalty_mean': 0.8819572925567627, 'reward_near_duplicate_penalty_std': 0.06823018938302994, 'reward_opening_diversity_mean': 0.9821428656578064, 'reward_opening_diversity_std': 0.05050762742757797, 'reward_distinct_2_mean': 0.723268985748291, 'reward_distinct_2_std': 0.1439959555864334, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.16850179433822632, 'reward_total_composite_mean': 0.19434238970279694, 'reward_total_composite_std': 0.1278955191373825, 'epoch': 0.07}

 41%|████      | 1350/3300 [5:08:29<5:55:54, 10.95s/it]INFO 04-12 21:06:55 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:10<00:59,  5.39s/it]

 23%|██▎       | 3/13 [00:16<00:54,  5.47s/it]

 31%|███       | 4/13 [00:24<00:56,  6.30s/it]

 38%|███▊      | 5/13 [00:31<00:53,  6.66s/it]

 46%|████▌     | 6/13 [00:39<00:50,  7.18s/it]

 54%|█████▍    | 7/13 [00:46<00:42,  7.07s/it]

 62%|██████▏   | 8/13 [00:54<00:36,  7.36s/it]

 69%|██████▉   | 9/13 [01:06<00:34,  8.73s/it]

 77%|███████▋  | 10/13 [01:13<00:24,  8.21s/it]

 85%|████████▍ | 11/13 [01:25<00:18,  9.29s/it]

 92%|█████████▏| 12/13 [01:33<00:08,  8.95s/it]

100%|██████████| 13/13 [01:44<00:00,  9.69s/it]2026-04-12 21:08:49,745 | INFO | train_grpo_train | metrics_logged mode=eval step=1350

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 113.8978, 'eval_samples_per_second': 0.913, 'eval_steps_per_second': 0.114, 'eval_num_tokens': 2387843.0, 'eval_completions/mean_length': 123.10576923076923, 'eval_completions/min_length': 34.38461538461539, 'eval_completions/max_length': 296.84615384615387, 'eval_completions/clipped_ratio': 0.057692307692307696, 'eval_completions/mean_terminated_length': 98.96456263615535, 'eval_completions/min_terminated_length': 34.38461538461539, 'eval_completions/max_terminated_length': 201.07692307692307, 'eval_rewards/meter/mean': 0.3967132041087517, 'eval_rewards/meter/std': 0.323750662115904, 'eval_rewards/count_adherence/mean': 0.77610732958867, 'eval_rewards/count_adherence/std': 0.1695853414443823, 'eval_rewards/arabic_clean/mean': 0.9134615384615384, 'eval_rewards/arabic_clean/std': 0.202989963384775, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9328321997935956, 'eval_rewards/count_floor/std': 0.05087560472580103, 'eval_rewards/lexical_plausibility/mean': 0.9669988705561712, 'eval_rewards/lexical_plausibility/std': 0.06795883981081155, 'eval_rewards/judge_quality/mean': 0.36019229888916016, 'eval_rewards/judge_quality/std': 0.1346744207235483, 'eval_rewards/repeat_penalty/mean': 0.8318327757028433, 'eval_rewards/repeat_penalty/std': 0.17920942375293145, 'eval_rewards/repeat_floor/mean': 0.9802892116399912, 'eval_rewards/repeat_floor/std': 0.021804944038964234, 'eval_rewards/near_duplicate_penalty/mean': 0.9425610762376052, 'eval_rewards/near_duplicate_penalty/std': 0.05716199628435648, 'eval_rewards/opening_diversity/mean': 0.9599059866024897, 'eval_rewards/opening_diversity/std': 0.10513016638847497, 'eval_rewards/distinct_2/mean': 0.9034522496736966, 'eval_rewards/distinct_2/std': 0.11727129410092647, 'eval_rewards/total_composite/mean': 0.1378111976843614, 'eval_rewards/total_composite/std': 0.13167950023825353, 'eval_reward': 0.1378111976843614, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08142225730877656, 'eval_sampling/sampling_logp_difference/max': 1.2555783711946928, 'eval_sampling/importance_sampling_ratio/min': 0.2909806794845141, 'eval_sampling/importance_sampling_ratio/mean': 1.0189277667265673, 'eval_sampling/importance_sampling_ratio/max': 1.6015074253082275, 'eval_entropy': 0.9267999208890475, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.1378111976843614, 'eval_reward_meter_mean': 0.3967132041087517, 'eval_reward_meter_std': 0.323750662115904, 'eval_reward_count_adherence_mean': 0.77610732958867, 'eval_reward_count_adherence_std': 0.1695853414443823, 'eval_reward_arabic_clean_mean': 0.9134615384615384, 'eval_reward_arabic_clean_std': 0.202989963384775, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9328321997935956, 'eval_reward_count_floor_std': 0.05087560472580103, 'eval_reward_lexical_plausibility_mean': 0.9669988705561712, 'eval_reward_lexical_plausibility_std': 0.06795883981081155, 'eval_reward_repeat_penalty_mean': 0.8318327757028433, 'eval_reward_repeat_penalty_std': 0.17920942375293145, 'eval_reward_repeat_floor_mean': 0.9802892116399912, 'eval_reward_repeat_floor_std': 0.021804944038964234, 'eval_reward_near_duplicate_penalty_mean': 0.9425610762376052, 'eval_reward_near_duplicate_penalty_std': 0.05716199628435648, 'eval_reward_opening_diversity_mean': 0.9599059866024897, 'eval_reward_opening_diversity_std': 0.10513016638847497, 'eval_reward_distinct_2_mean': 0.9034522496736966, 'eval_reward_distinct_2_std': 0.11727129410092647, 'eval_reward_judge_quality_mean': 0.36019229888916016, 'eval_reward_judge_quality_std': 0.1346744207235483, 'eval_reward_total_composite_mean': 0.1378111976843614, 'eval_reward_total_composite_std': 0.13167950023825353, 'epoch': 0.07}

 41%|████      | 1350/3300 [5:10:23<5:55:54, 10.95s/it]

100%|██████████| 13/13 [01:44<00:00,  9.69s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 41%|████      | 1351/3300 [5:10:38<25:03:19, 46.28s/it]2026-04-12 21:09:04,424 | INFO | train_grpo_train | metrics_logged mode=train step=1351

                                                        
{'loss': 0.1002, 'grad_norm': 8.832219123840332, 'learning_rate': 5.90909090909091e-06, 'num_tokens': 2391010.0, 'completions/mean_length': 166.875, 'completions/min_length': 129.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.875, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.41311126947402954, 'rewards/meter/std': 0.36640501022338867, 'rewards/count_adherence/mean': 0.7019230723381042, 'rewards/count_adherence/std': 0.10432328283786774, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9105769395828247, 'rewards/count_floor/std': 0.0312969833612442, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.29687660932540894, 'rewards/repeat_penalty/std': 0.2861458957195282, 'rewards/repeat_floor/mean': 0.9037836194038391, 'rewards/repeat_floor/std': 0.05622662603855133, 'rewards/near_duplicate_penalty/mean': 0.7750030159950256, 'rewards/near_duplicate_penalty/std': 0.13784529268741608, 'rewards/opening_diversity/mean': 0.7093137502670288, 'rewards/opening_diversity/std': 0.2800697386264801, 'rewards/distinct_2/mean': 0.4997152090072632, 'rewards/distinct_2/std': 0.23885813355445862, 'rewards/total_composite/mean': 0.11222703009843826, 'rewards/total_composite/std': 0.11665745824575424, 'reward': 0.11222703009843826, 'reward_std': 0.11665745079517365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11122331768274307, 'sampling/sampling_logp_difference/max': 2.2513864040374756, 'sampling/importance_sampling_ratio/min': 0.1052531972527504, 'sampling/importance_sampling_ratio/mean': 1.008052945137024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5516439266502857, 'clip_ratio/low_mean': 0.07277052337303758, 'clip_ratio/low_min': 0.07277052337303758, 'clip_ratio/high_mean': 0.028003876097500324, 'clip_ratio/high_max': 0.028003876097500324, 'clip_ratio/region_mean': 0.1007743994705379, 'reward_total_mean': 0.11222703009843826, 'reward_meter_mean': 0.41311126947402954, 'reward_meter_std': 0.36640501022338867, 'reward_count_adherence_mean': 0.7019230723381042, 'reward_count_adherence_std': 0.10432328283786774, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9105769395828247, 'reward_count_floor_std': 0.0312969833612442, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.29687660932540894, 'reward_repeat_penalty_std': 0.2861458957195282, 'reward_repeat_floor_mean': 0.9037836194038391, 'reward_repeat_floor_std': 0.05622662603855133, 'reward_near_duplicate_penalty_mean': 0.7750030159950256, 'reward_near_duplicate_penalty_std': 0.13784529268741608, 'reward_opening_diversity_mean': 0.7093137502670288, 'reward_opening_diversity_std': 0.2800697386264801, 'reward_distinct_2_mean': 0.4997152090072632, 'reward_distinct_2_std': 0.23885813355445862, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.11222703009843826, 'reward_total_composite_std': 0.11665745824575424, 'epoch': 0.07}

 41%|████      | 1351/3300 [5:10:38<25:03:19, 46.28s/it]INFO 04-12 21:09:04 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1352/3300 [5:10:47<18:59:42, 35.10s/it]2026-04-12 21:09:13,452 | INFO | train_grpo_train | metrics_logged mode=train step=1352

                                                        
{'loss': 0.0392, 'grad_norm': 11.583563804626465, 'learning_rate': 5.906060606060607e-06, 'num_tokens': 2392905.0, 'completions/mean_length': 61.875, 'completions/min_length': 43.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.3075094223022461, 'rewards/meter/std': 0.3981509804725647, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9286723136901855, 'rewards/repeat_penalty/std': 0.0527019128203392, 'rewards/repeat_floor/mean': 0.9909531474113464, 'rewards/repeat_floor/std': 0.005987527780234814, 'rewards/near_duplicate_penalty/mean': 0.9588726162910461, 'rewards/near_duplicate_penalty/std': 0.019831784069538116, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.968024730682373, 'rewards/distinct_2/std': 0.03976638615131378, 'rewards/total_composite/mean': 0.12538199126720428, 'rewards/total_composite/std': 0.15767979621887207, 'reward': 0.12538199126720428, 'reward_std': 0.15767979621887207, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14805741608142853, 'sampling/sampling_logp_difference/max': 1.8405616283416748, 'sampling/importance_sampling_ratio/min': 0.15872825682163239, 'sampling/importance_sampling_ratio/mean': 0.9996080994606018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8243724331259727, 'clip_ratio/low_mean': 0.05267860344611108, 'clip_ratio/low_min': 0.05267860344611108, 'clip_ratio/high_mean': 0.042106867767870426, 'clip_ratio/high_max': 0.042106867767870426, 'clip_ratio/region_mean': 0.09478547121398151, 'reward_total_mean': 0.12538199126720428, 'reward_meter_mean': 0.3075094223022461, 'reward_meter_std': 0.3981509804725647, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9286723136901855, 'reward_repeat_penalty_std': 0.0527019128203392, 'reward_repeat_floor_mean': 0.9909531474113464, 'reward_repeat_floor_std': 0.005987527780234814, 'reward_near_duplicate_penalty_mean': 0.9588726162910461, 'reward_near_duplicate_penalty_std': 0.019831784069538116, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.968024730682373, 'reward_distinct_2_std': 0.03976638615131378, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.12538199126720428, 'reward_total_composite_std': 0.15767979621887207, 'epoch': 0.07}

 41%|████      | 1352/3300 [5:10:47<18:59:42, 35.10s/it]INFO 04-12 21:09:13 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1353/3300 [5:10:55<14:39:13, 27.09s/it]2026-04-12 21:09:21,863 | INFO | train_grpo_train | metrics_logged mode=train step=1353

                                                        
{'loss': 0.1589, 'grad_norm': 16.21390151977539, 'learning_rate': 5.903030303030304e-06, 'num_tokens': 2394530.0, 'completions/mean_length': 38.125, 'completions/min_length': 26.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5851883292198181, 'rewards/meter/std': 0.28261780738830566, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.08017836511135101, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9855026602745056, 'rewards/repeat_penalty/std': 0.014037727378308773, 'rewards/repeat_floor/mean': 0.9978253841400146, 'rewards/repeat_floor/std': 0.0021056621335446835, 'rewards/near_duplicate_penalty/mean': 0.9855026602745056, 'rewards/near_duplicate_penalty/std': 0.014037727378308773, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2699908912181854, 'rewards/total_composite/std': 0.17090578377246857, 'reward': 0.2699908912181854, 'reward_std': 0.17090578377246857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1833714246749878, 'sampling/sampling_logp_difference/max': 1.799229621887207, 'sampling/importance_sampling_ratio/min': 0.16542628407478333, 'sampling/importance_sampling_ratio/mean': 1.009809970855713, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.16598941385746, 'clip_ratio/low_mean': 0.06562512181699276, 'clip_ratio/low_min': 0.06562512181699276, 'clip_ratio/high_mean': 0.04002100881189108, 'clip_ratio/high_max': 0.04002100881189108, 'clip_ratio/region_mean': 0.10564613062888384, 'reward_total_mean': 0.2699908912181854, 'reward_meter_mean': 0.5851883292198181, 'reward_meter_std': 0.28261780738830566, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.08017836511135101, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9855026602745056, 'reward_repeat_penalty_std': 0.014037727378308773, 'reward_repeat_floor_mean': 0.9978253841400146, 'reward_repeat_floor_std': 0.0021056621335446835, 'reward_near_duplicate_penalty_mean': 0.9855026602745056, 'reward_near_duplicate_penalty_std': 0.014037727378308773, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.2699908912181854, 'reward_total_composite_std': 0.17090578377246857, 'epoch': 0.07}

 41%|████      | 1353/3300 [5:10:55<14:39:13, 27.09s/it]INFO 04-12 21:09:22 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1354/3300 [5:11:09<12:31:58, 23.19s/it]2026-04-12 21:09:35,924 | INFO | train_grpo_train | metrics_logged mode=train step=1354

                                                        
{'loss': -0.0749, 'grad_norm': 2.1872777938842773, 'learning_rate': 5.9e-06, 'num_tokens': 2395954.0, 'completions/mean_length': 156.0, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 37.333335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5997096300125122, 'rewards/meter/std': 0.38438868522644043, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.8902959227561951, 'rewards/lexical_plausibility/std': 0.20633329451084137, 'rewards/judge_quality/mean': 0.29624998569488525, 'rewards/judge_quality/std': 0.19115717709064484, 'rewards/repeat_penalty/mean': 0.9697509407997131, 'rewards/repeat_penalty/std': 0.033267371356487274, 'rewards/repeat_floor/mean': 0.9954626560211182, 'rewards/repeat_floor/std': 0.004990102257579565, 'rewards/near_duplicate_penalty/mean': 0.9697509407997131, 'rewards/near_duplicate_penalty/std': 0.033267371356487274, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22220458090305328, 'rewards/total_composite/std': 0.15941786766052246, 'reward': 0.22220458090305328, 'reward_std': 0.15941786766052246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16308002173900604, 'sampling/sampling_logp_difference/max': 1.8037405014038086, 'sampling/importance_sampling_ratio/min': 0.16468174755573273, 'sampling/importance_sampling_ratio/mean': 1.0090183019638062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7480884790420532, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.08500408753752708, 'clip_ratio/high_max': 0.08500408753752708, 'clip_ratio/region_mean': 0.10062908753752708, 'reward_total_mean': 0.22220458090305328, 'reward_meter_mean': 0.5997096300125122, 'reward_meter_std': 0.38438868522644043, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.8902959227561951, 'reward_lexical_plausibility_std': 0.20633329451084137, 'reward_repeat_penalty_mean': 0.9697509407997131, 'reward_repeat_penalty_std': 0.033267371356487274, 'reward_repeat_floor_mean': 0.9954626560211182, 'reward_repeat_floor_std': 0.004990102257579565, 'reward_near_duplicate_penalty_mean': 0.9697509407997131, 'reward_near_duplicate_penalty_std': 0.033267371356487274, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.29624998569488525, 'reward_judge_quality_std': 0.19115717709064484, 'reward_total_composite_mean': 0.22220458090305328, 'reward_total_composite_std': 0.15941786766052246, 'epoch': 0.07}

 41%|████      | 1354/3300 [5:11:09<12:31:58, 23.19s/it]INFO 04-12 21:09:36 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1355/3300 [5:11:18<10:10:35, 18.84s/it]2026-04-12 21:09:44,616 | INFO | train_grpo_train | metrics_logged mode=train step=1355

                                                        
{'loss': -0.003, 'grad_norm': 11.956860542297363, 'learning_rate': 5.8969696969696975e-06, 'num_tokens': 2397766.0, 'completions/mean_length': 64.5, 'completions/min_length': 55.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.3731825053691864, 'rewards/meter/std': 0.31172871589660645, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9521870613098145, 'rewards/repeat_penalty/std': 0.04388941824436188, 'rewards/repeat_floor/mean': 0.9940490126609802, 'rewards/repeat_floor/std': 0.004772914573550224, 'rewards/near_duplicate_penalty/mean': 0.9739071726799011, 'rewards/near_duplicate_penalty/std': 0.012133183889091015, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9773657321929932, 'rewards/distinct_2/std': 0.03421974554657936, 'rewards/total_composite/mean': 0.1522723287343979, 'rewards/total_composite/std': 0.13348764181137085, 'reward': 0.1522723287343979, 'reward_std': 0.13348765671253204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15042823553085327, 'sampling/sampling_logp_difference/max': 1.8057926893234253, 'sampling/importance_sampling_ratio/min': 0.16434413194656372, 'sampling/importance_sampling_ratio/mean': 1.0341154336929321, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1172214671969414, 'clip_ratio/low_mean': 0.08428987581282854, 'clip_ratio/low_min': 0.08428987581282854, 'clip_ratio/high_mean': 0.04304391983896494, 'clip_ratio/high_max': 0.04304391983896494, 'clip_ratio/region_mean': 0.12733379565179348, 'reward_total_mean': 0.1522723287343979, 'reward_meter_mean': 0.3731825053691864, 'reward_meter_std': 0.31172871589660645, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9521870613098145, 'reward_repeat_penalty_std': 0.04388941824436188, 'reward_repeat_floor_mean': 0.9940490126609802, 'reward_repeat_floor_std': 0.004772914573550224, 'reward_near_duplicate_penalty_mean': 0.9739071726799011, 'reward_near_duplicate_penalty_std': 0.012133183889091015, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9773657321929932, 'reward_distinct_2_std': 0.03421974554657936, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.1522723287343979, 'reward_total_composite_std': 0.13348764181137085, 'epoch': 0.07}

 41%|████      | 1355/3300 [5:11:18<10:10:35, 18.84s/it]INFO 04-12 21:09:44 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1356/3300 [5:11:32<9:23:22, 17.39s/it] 2026-04-12 21:09:58,620 | INFO | train_grpo_train | metrics_logged mode=train step=1356

                                                       
{'loss': -0.0346, 'grad_norm': 5.269680500030518, 'learning_rate': 5.893939393939394e-06, 'num_tokens': 2399313.0, 'completions/mean_length': 90.375, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 30.142858505249023, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5154728889465332, 'rewards/meter/std': 0.4530288577079773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9640141725540161, 'rewards/lexical_plausibility/std': 0.10178320854902267, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.335769385099411, 'rewards/repeat_penalty/mean': 0.9006885886192322, 'rewards/repeat_penalty/std': 0.1884777545928955, 'rewards/repeat_floor/mean': 0.9870600700378418, 'rewards/repeat_floor/std': 0.022856641560792923, 'rewards/near_duplicate_penalty/mean': 0.9465957880020142, 'rewards/near_duplicate_penalty/std': 0.06550285965204239, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.966063380241394, 'rewards/distinct_2/std': 0.09598734229803085, 'rewards/total_composite/mean': 0.35103386640548706, 'rewards/total_composite/std': 0.3306849002838135, 'reward': 0.35103386640548706, 'reward_std': 0.3306848704814911, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16407063603401184, 'sampling/sampling_logp_difference/max': 2.409271240234375, 'sampling/importance_sampling_ratio/min': 0.08988077938556671, 'sampling/importance_sampling_ratio/mean': 1.0390325784683228, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1144169569015503, 'clip_ratio/low_mean': 0.022288603708148003, 'clip_ratio/low_min': 0.022288603708148003, 'clip_ratio/high_mean': 0.07925925962626934, 'clip_ratio/high_max': 0.07925925962626934, 'clip_ratio/region_mean': 0.10154786333441734, 'reward_total_mean': 0.35103386640548706, 'reward_meter_mean': 0.5154728889465332, 'reward_meter_std': 0.4530288577079773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9640141725540161, 'reward_lexical_plausibility_std': 0.10178320854902267, 'reward_repeat_penalty_mean': 0.9006885886192322, 'reward_repeat_penalty_std': 0.1884777545928955, 'reward_repeat_floor_mean': 0.9870600700378418, 'reward_repeat_floor_std': 0.022856641560792923, 'reward_near_duplicate_penalty_mean': 0.9465957880020142, 'reward_near_duplicate_penalty_std': 0.06550285965204239, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.966063380241394, 'reward_distinct_2_std': 0.09598734229803085, 'reward_judge_quality_mean': 0.6937500238418579, 'reward_judge_quality_std': 0.335769385099411, 'reward_total_composite_mean': 0.35103386640548706, 'reward_total_composite_std': 0.3306849002838135, 'epoch': 0.07}

 41%|████      | 1356/3300 [5:11:32<9:23:22, 17.39s/it]INFO 04-12 21:09:58 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1357/3300 [5:11:41<7:58:37, 14.78s/it]2026-04-12 21:10:07,312 | INFO | train_grpo_train | metrics_logged mode=train step=1357

                                                       
{'loss': 0.0673, 'grad_norm': 16.956844329833984, 'learning_rate': 5.890909090909091e-06, 'num_tokens': 2400868.0, 'completions/mean_length': 39.375, 'completions/min_length': 36.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.335527241230011, 'rewards/meter/std': 0.26939159631729126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9629907608032227, 'rewards/repeat_penalty/std': 0.04661880061030388, 'rewards/repeat_floor/mean': 0.9949173331260681, 'rewards/repeat_floor/std': 0.006006376817822456, 'rewards/near_duplicate_penalty/mean': 0.9716099500656128, 'rewards/near_duplicate_penalty/std': 0.031306542456150055, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.1155606210231781, 'rewards/total_composite/std': 0.10034933686256409, 'reward': 0.1155606210231781, 'reward_std': 0.10034932941198349, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17751607298851013, 'sampling/sampling_logp_difference/max': 1.6329107284545898, 'sampling/importance_sampling_ratio/min': 0.19536009430885315, 'sampling/importance_sampling_ratio/mean': 1.037315011024475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3439089804887772, 'clip_ratio/low_mean': 0.07852805173024535, 'clip_ratio/low_min': 0.07852805173024535, 'clip_ratio/high_mean': 0.06803776789456606, 'clip_ratio/high_max': 0.06803776789456606, 'clip_ratio/region_mean': 0.1465658196248114, 'reward_total_mean': 0.1155606210231781, 'reward_meter_mean': 0.335527241230011, 'reward_meter_std': 0.26939159631729126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9629907608032227, 'reward_repeat_penalty_std': 0.04661880061030388, 'reward_repeat_floor_mean': 0.9949173331260681, 'reward_repeat_floor_std': 0.006006376817822456, 'reward_near_duplicate_penalty_mean': 0.9716099500656128, 'reward_near_duplicate_penalty_std': 0.031306542456150055, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.1155606210231781, 'reward_total_composite_std': 0.10034933686256409, 'epoch': 0.07}

 41%|████      | 1357/3300 [5:11:41<7:58:37, 14.78s/it]INFO 04-12 21:10:07 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1358/3300 [5:11:54<7:48:20, 14.47s/it]2026-04-12 21:10:21,065 | INFO | train_grpo_train | metrics_logged mode=train step=1358

                                                       
{'loss': 0.0105, 'grad_norm': 6.21705961227417, 'learning_rate': 5.887878787878788e-06, 'num_tokens': 2402892.0, 'completions/mean_length': 125.0, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.71428680419922, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.26811739802360535, 'rewards/meter/std': 0.31841468811035156, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9584158658981323, 'rewards/lexical_plausibility/std': 0.11761771887540817, 'rewards/judge_quality/mean': 0.34999996423721313, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9397847056388855, 'rewards/repeat_penalty/std': 0.04290353134274483, 'rewards/repeat_floor/mean': 0.9927159547805786, 'rewards/repeat_floor/std': 0.0050594499334692955, 'rewards/near_duplicate_penalty/mean': 0.9707650542259216, 'rewards/near_duplicate_penalty/std': 0.02095148153603077, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9677910804748535, 'rewards/distinct_2/std': 0.029113205149769783, 'rewards/total_composite/mean': 0.09462591260671616, 'rewards/total_composite/std': 0.1390860378742218, 'reward': 0.09462591260671616, 'reward_std': 0.1390860378742218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17054177820682526, 'sampling/sampling_logp_difference/max': 1.2233905792236328, 'sampling/importance_sampling_ratio/min': 0.2942308485507965, 'sampling/importance_sampling_ratio/mean': 1.0266481637954712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1378167420625687, 'clip_ratio/low_mean': 0.1034215921536088, 'clip_ratio/low_min': 0.1034215921536088, 'clip_ratio/high_mean': 0.051551453769207, 'clip_ratio/high_max': 0.051551453769207, 'clip_ratio/region_mean': 0.1549730459228158, 'reward_total_mean': 0.09462591260671616, 'reward_meter_mean': 0.26811739802360535, 'reward_meter_std': 0.31841468811035156, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9584158658981323, 'reward_lexical_plausibility_std': 0.11761771887540817, 'reward_repeat_penalty_mean': 0.9397847056388855, 'reward_repeat_penalty_std': 0.04290353134274483, 'reward_repeat_floor_mean': 0.9927159547805786, 'reward_repeat_floor_std': 0.0050594499334692955, 'reward_near_duplicate_penalty_mean': 0.9707650542259216, 'reward_near_duplicate_penalty_std': 0.02095148153603077, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9677910804748535, 'reward_distinct_2_std': 0.029113205149769783, 'reward_judge_quality_mean': 0.34999996423721313, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.09462591260671616, 'reward_total_composite_std': 0.1390860378742218, 'epoch': 0.07}

 41%|████      | 1358/3300 [5:11:54<7:48:20, 14.47s/it]INFO 04-12 21:10:21 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1359/3300 [5:12:02<6:46:52, 12.58s/it]2026-04-12 21:10:29,218 | INFO | train_grpo_train | metrics_logged mode=train step=1359

                                                       
{'loss': 0.0241, 'grad_norm': 11.658727645874023, 'learning_rate': 5.884848484848486e-06, 'num_tokens': 2404701.0, 'completions/mean_length': 60.125, 'completions/min_length': 56.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7138677835464478, 'rewards/meter/std': 0.23756182193756104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.2794605791568756, 'rewards/repeat_penalty/mean': 0.9727405905723572, 'rewards/repeat_penalty/std': 0.026929466053843498, 'rewards/repeat_floor/mean': 0.9962838888168335, 'rewards/repeat_floor/std': 0.003047413658350706, 'rewards/near_duplicate_penalty/mean': 0.9793470501899719, 'rewards/near_duplicate_penalty/std': 0.010606988333165646, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9931318759918213, 'rewards/distinct_2/std': 0.01942601054906845, 'rewards/total_composite/mean': 0.5407896637916565, 'rewards/total_composite/std': 0.2931385934352875, 'reward': 0.5407896637916565, 'reward_std': 0.2931385934352875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12463036179542542, 'sampling/sampling_logp_difference/max': 2.492025136947632, 'sampling/importance_sampling_ratio/min': 0.08274222910404205, 'sampling/importance_sampling_ratio/mean': 0.9909713864326477, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6973786652088165, 'clip_ratio/low_mean': 0.03901098994538188, 'clip_ratio/low_min': 0.03901098994538188, 'clip_ratio/high_mean': 0.07961014751344919, 'clip_ratio/high_max': 0.07961014751344919, 'clip_ratio/region_mean': 0.11862113745883107, 'reward_total_mean': 0.5407896637916565, 'reward_meter_mean': 0.7138677835464478, 'reward_meter_std': 0.23756182193756104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9727405905723572, 'reward_repeat_penalty_std': 0.026929466053843498, 'reward_repeat_floor_mean': 0.9962838888168335, 'reward_repeat_floor_std': 0.003047413658350706, 'reward_near_duplicate_penalty_mean': 0.9793470501899719, 'reward_near_duplicate_penalty_std': 0.010606988333165646, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9931318759918213, 'reward_distinct_2_std': 0.01942601054906845, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.2794605791568756, 'reward_total_composite_mean': 0.5407896637916565, 'reward_total_composite_std': 0.2931385934352875, 'epoch': 0.07}

 41%|████      | 1359/3300 [5:12:03<6:46:52, 12.58s/it]INFO 04-12 21:10:29 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1360/3300 [5:12:16<6:59:00, 12.96s/it]2026-04-12 21:10:43,072 | INFO | train_grpo_train | metrics_logged mode=train step=1360

                                                       
{'loss': -0.0461, 'grad_norm': 4.810898780822754, 'learning_rate': 5.881818181818182e-06, 'num_tokens': 2406301.0, 'completions/mean_length': 99.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.3990902304649353, 'rewards/meter/std': 0.4627956748008728, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8402870893478394, 'rewards/lexical_plausibility/std': 0.18085414171218872, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8204519748687744, 'rewards/repeat_penalty/std': 0.27600109577178955, 'rewards/repeat_floor/mean': 0.9796364307403564, 'rewards/repeat_floor/std': 0.030433200299739838, 'rewards/near_duplicate_penalty/mean': 0.9380507469177246, 'rewards/near_duplicate_penalty/std': 0.0795394778251648, 'rewards/opening_diversity/mean': 0.890625, 'rewards/opening_diversity/std': 0.2259652018547058, 'rewards/distinct_2/mean': 0.9499030113220215, 'rewards/distinct_2/std': 0.07513502985239029, 'rewards/total_composite/mean': 0.11561507731676102, 'rewards/total_composite/std': 0.14266031980514526, 'reward': 0.11561507731676102, 'reward_std': 0.14266031980514526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12486492097377777, 'sampling/sampling_logp_difference/max': 1.2106590270996094, 'sampling/importance_sampling_ratio/min': 0.2980008125305176, 'sampling/importance_sampling_ratio/mean': 1.0246108770370483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0419716835021973, 'clip_ratio/low_mean': 0.04115276248194277, 'clip_ratio/low_min': 0.04115276248194277, 'clip_ratio/high_mean': 0.03902312647551298, 'clip_ratio/high_max': 0.03902312647551298, 'clip_ratio/region_mean': 0.08017588895745575, 'reward_total_mean': 0.11561507731676102, 'reward_meter_mean': 0.3990902304649353, 'reward_meter_std': 0.4627956748008728, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8402870893478394, 'reward_lexical_plausibility_std': 0.18085414171218872, 'reward_repeat_penalty_mean': 0.8204519748687744, 'reward_repeat_penalty_std': 0.27600109577178955, 'reward_repeat_floor_mean': 0.9796364307403564, 'reward_repeat_floor_std': 0.030433200299739838, 'reward_near_duplicate_penalty_mean': 0.9380507469177246, 'reward_near_duplicate_penalty_std': 0.0795394778251648, 'reward_opening_diversity_mean': 0.890625, 'reward_opening_diversity_std': 0.2259652018547058, 'reward_distinct_2_mean': 0.9499030113220215, 'reward_distinct_2_std': 0.07513502985239029, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.11561507731676102, 'reward_total_composite_std': 0.14266031980514526, 'epoch': 0.07}

 41%|████      | 1360/3300 [5:12:16<6:59:00, 12.96s/it]INFO 04-12 21:10:43 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1361/3300 [5:12:30<7:03:38, 13.11s/it]2026-04-12 21:10:56,529 | INFO | train_grpo_train | metrics_logged mode=train step=1361

                                                       
{'loss': -0.1028, 'grad_norm': 3.076615571975708, 'learning_rate': 5.878787878787879e-06, 'num_tokens': 2408021.0, 'completions/mean_length': 172.0, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 58.66666793823242, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.2548764646053314, 'rewards/meter/std': 0.2447020262479782, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9249999523162842, 'rewards/count_floor/std': 0.0707106813788414, 'rewards/lexical_plausibility/mean': 0.9021851420402527, 'rewards/lexical_plausibility/std': 0.18599532544612885, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8881406784057617, 'rewards/repeat_penalty/std': 0.1280057728290558, 'rewards/repeat_floor/mean': 0.9888988733291626, 'rewards/repeat_floor/std': 0.010923240333795547, 'rewards/near_duplicate_penalty/mean': 0.9595867991447449, 'rewards/near_duplicate_penalty/std': 0.03943898528814316, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9856764078140259, 'rewards/distinct_2/std': 0.02661900594830513, 'rewards/total_composite/mean': 0.052175745368003845, 'rewards/total_composite/std': 0.04424585774540901, 'reward': 0.052175745368003845, 'reward_std': 0.04424585774540901, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1724514365196228, 'sampling/sampling_logp_difference/max': 1.8498058319091797, 'sampling/importance_sampling_ratio/min': 0.1572677046060562, 'sampling/importance_sampling_ratio/mean': 1.0251553058624268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2266998514533043, 'clip_ratio/low_mean': 0.03442029049620032, 'clip_ratio/low_min': 0.03442029049620032, 'clip_ratio/high_mean': 0.09937055967748165, 'clip_ratio/high_max': 0.09937055967748165, 'clip_ratio/region_mean': 0.13379085017368197, 'reward_total_mean': 0.052175745368003845, 'reward_meter_mean': 0.2548764646053314, 'reward_meter_std': 0.2447020262479782, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9249999523162842, 'reward_count_floor_std': 0.0707106813788414, 'reward_lexical_plausibility_mean': 0.9021851420402527, 'reward_lexical_plausibility_std': 0.18599532544612885, 'reward_repeat_penalty_mean': 0.8881406784057617, 'reward_repeat_penalty_std': 0.1280057728290558, 'reward_repeat_floor_mean': 0.9888988733291626, 'reward_repeat_floor_std': 0.010923240333795547, 'reward_near_duplicate_penalty_mean': 0.9595867991447449, 'reward_near_duplicate_penalty_std': 0.03943898528814316, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9856764078140259, 'reward_distinct_2_std': 0.02661900594830513, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.052175745368003845, 'reward_total_composite_std': 0.04424585774540901, 'epoch': 0.07}

 41%|████      | 1361/3300 [5:12:30<7:03:38, 13.11s/it]INFO 04-12 21:10:56 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1362/3300 [5:12:45<7:22:25, 13.70s/it]2026-04-12 21:11:11,601 | INFO | train_grpo_train | metrics_logged mode=train step=1362

                                                       
{'loss': -0.0958, 'grad_norm': 4.198287010192871, 'learning_rate': 5.875757575757576e-06, 'num_tokens': 2410815.0, 'completions/mean_length': 181.25, 'completions/min_length': 109.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 134.0, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.33927029371261597, 'rewards/meter/std': 0.23474349081516266, 'rewards/count_adherence/mean': 0.8214285373687744, 'rewards/count_adherence/std': 0.10101525485515594, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9464285373687744, 'rewards/count_floor/std': 0.030304571613669395, 'rewards/lexical_plausibility/mean': 0.9662079811096191, 'rewards/lexical_plausibility/std': 0.09557818621397018, 'rewards/judge_quality/mean': 0.38124996423721313, 'rewards/judge_quality/std': 0.15797263383865356, 'rewards/repeat_penalty/mean': 0.9583768248558044, 'rewards/repeat_penalty/std': 0.03497450798749924, 'rewards/repeat_floor/mean': 0.9950957298278809, 'rewards/repeat_floor/std': 0.003586851991713047, 'rewards/near_duplicate_penalty/mean': 0.981342077255249, 'rewards/near_duplicate_penalty/std': 0.013582875952124596, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9766049385070801, 'rewards/distinct_2/std': 0.03318921476602554, 'rewards/total_composite/mean': 0.12686169147491455, 'rewards/total_composite/std': 0.10618890076875687, 'reward': 0.12686169147491455, 'reward_std': 0.10618889331817627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12347649037837982, 'sampling/sampling_logp_difference/max': 1.9963264465332031, 'sampling/importance_sampling_ratio/min': 0.13583335280418396, 'sampling/importance_sampling_ratio/mean': 1.0084747076034546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7365588247776031, 'clip_ratio/low_mean': 0.058945417404174805, 'clip_ratio/low_min': 0.058945417404174805, 'clip_ratio/high_mean': 0.04070075694471598, 'clip_ratio/high_max': 0.04070075694471598, 'clip_ratio/region_mean': 0.09964617434889078, 'reward_total_mean': 0.12686169147491455, 'reward_meter_mean': 0.33927029371261597, 'reward_meter_std': 0.23474349081516266, 'reward_count_adherence_mean': 0.8214285373687744, 'reward_count_adherence_std': 0.10101525485515594, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9464285373687744, 'reward_count_floor_std': 0.030304571613669395, 'reward_lexical_plausibility_mean': 0.9662079811096191, 'reward_lexical_plausibility_std': 0.09557818621397018, 'reward_repeat_penalty_mean': 0.9583768248558044, 'reward_repeat_penalty_std': 0.03497450798749924, 'reward_repeat_floor_mean': 0.9950957298278809, 'reward_repeat_floor_std': 0.003586851991713047, 'reward_near_duplicate_penalty_mean': 0.981342077255249, 'reward_near_duplicate_penalty_std': 0.013582875952124596, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9766049385070801, 'reward_distinct_2_std': 0.03318921476602554, 'reward_judge_quality_mean': 0.38124996423721313, 'reward_judge_quality_std': 0.15797263383865356, 'reward_total_composite_mean': 0.12686169147491455, 'reward_total_composite_std': 0.10618890076875687, 'epoch': 0.07}

 41%|████▏     | 1362/3300 [5:12:45<7:22:25, 13.70s/it]INFO 04-12 21:11:11 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1363/3300 [5:12:58<7:19:29, 13.61s/it]2026-04-12 21:11:25,020 | INFO | train_grpo_train | metrics_logged mode=train step=1363

                                                       
{'loss': -0.0695, 'grad_norm': 4.720059871673584, 'learning_rate': 5.872727272727273e-06, 'num_tokens': 2412419.0, 'completions/mean_length': 111.5, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.28571701049805, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.13672298192977905, 'rewards/meter/std': 0.14602231979370117, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9417310357093811, 'rewards/lexical_plausibility/std': 0.16480956971645355, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.11649647355079651, 'rewards/repeat_penalty/mean': 0.9243472218513489, 'rewards/repeat_penalty/std': 0.05871216952800751, 'rewards/repeat_floor/mean': 0.9911385774612427, 'rewards/repeat_floor/std': 0.006358617916703224, 'rewards/near_duplicate_penalty/mean': 0.9686267375946045, 'rewards/near_duplicate_penalty/std': 0.01730871945619583, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9538290500640869, 'rewards/distinct_2/std': 0.04971982166171074, 'rewards/total_composite/mean': 0.02115466445684433, 'rewards/total_composite/std': 0.020929690450429916, 'reward': 0.02115466445684433, 'reward_std': 0.020929690450429916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1577281653881073, 'sampling/sampling_logp_difference/max': 1.5699477195739746, 'sampling/importance_sampling_ratio/min': 0.2652134597301483, 'sampling/importance_sampling_ratio/mean': 1.0405060052871704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2455516830086708, 'clip_ratio/low_mean': 0.07088014390319586, 'clip_ratio/low_min': 0.07088014390319586, 'clip_ratio/high_mean': 0.04233583714812994, 'clip_ratio/high_max': 0.04233583714812994, 'clip_ratio/region_mean': 0.1132159810513258, 'reward_total_mean': 0.02115466445684433, 'reward_meter_mean': 0.13672298192977905, 'reward_meter_std': 0.14602231979370117, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9417310357093811, 'reward_lexical_plausibility_std': 0.16480956971645355, 'reward_repeat_penalty_mean': 0.9243472218513489, 'reward_repeat_penalty_std': 0.05871216952800751, 'reward_repeat_floor_mean': 0.9911385774612427, 'reward_repeat_floor_std': 0.006358617916703224, 'reward_near_duplicate_penalty_mean': 0.9686267375946045, 'reward_near_duplicate_penalty_std': 0.01730871945619583, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9538290500640869, 'reward_distinct_2_std': 0.04971982166171074, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.11649647355079651, 'reward_total_composite_mean': 0.02115466445684433, 'reward_total_composite_std': 0.020929690450429916, 'epoch': 0.07}

 41%|████▏     | 1363/3300 [5:12:58<7:19:29, 13.61s/it]INFO 04-12 21:11:25 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1364/3300 [5:13:07<6:35:24, 12.25s/it]2026-04-12 21:11:34,106 | INFO | train_grpo_train | metrics_logged mode=train step=1364

                                                       
{'loss': 0.0257, 'grad_norm': 8.75416374206543, 'learning_rate': 5.8696969696969694e-06, 'num_tokens': 2414589.0, 'completions/mean_length': 84.25, 'completions/min_length': 74.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.25, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.7619160413742065, 'rewards/meter/std': 0.2968056797981262, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.8796122074127197, 'rewards/repeat_penalty/std': 0.062147945165634155, 'rewards/repeat_floor/mean': 0.9859489798545837, 'rewards/repeat_floor/std': 0.007579122669994831, 'rewards/near_duplicate_penalty/mean': 0.9530795812606812, 'rewards/near_duplicate_penalty/std': 0.028114236891269684, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9220781326293945, 'rewards/distinct_2/std': 0.04306622967123985, 'rewards/total_composite/mean': 0.19920063018798828, 'rewards/total_composite/std': 0.10604803264141083, 'reward': 0.19920063018798828, 'reward_std': 0.10604802519083023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1493125855922699, 'sampling/sampling_logp_difference/max': 4.442129611968994, 'sampling/importance_sampling_ratio/min': 0.011770844459533691, 'sampling/importance_sampling_ratio/mean': 1.0320141315460205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1193640530109406, 'clip_ratio/low_mean': 0.0734433950856328, 'clip_ratio/low_min': 0.0734433950856328, 'clip_ratio/high_mean': 0.047249577939510345, 'clip_ratio/high_max': 0.047249577939510345, 'clip_ratio/region_mean': 0.12069297302514315, 'reward_total_mean': 0.19920063018798828, 'reward_meter_mean': 0.7619160413742065, 'reward_meter_std': 0.2968056797981262, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8796122074127197, 'reward_repeat_penalty_std': 0.062147945165634155, 'reward_repeat_floor_mean': 0.9859489798545837, 'reward_repeat_floor_std': 0.007579122669994831, 'reward_near_duplicate_penalty_mean': 0.9530795812606812, 'reward_near_duplicate_penalty_std': 0.028114236891269684, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9220781326293945, 'reward_distinct_2_std': 0.04306622967123985, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.19920063018798828, 'reward_total_composite_std': 0.10604803264141083, 'epoch': 0.07}

 41%|████▏     | 1364/3300 [5:13:07<6:35:24, 12.25s/it]INFO 04-12 21:11:34 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1365/3300 [5:13:18<6:16:33, 11.68s/it]2026-04-12 21:11:44,427 | INFO | train_grpo_train | metrics_logged mode=train step=1365

                                                       
{'loss': 0.0532, 'grad_norm': 7.535861492156982, 'learning_rate': 5.8666666666666675e-06, 'num_tokens': 2417230.0, 'completions/mean_length': 107.125, 'completions/min_length': 100.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.125, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.31687265634536743, 'rewards/meter/std': 0.2848235070705414, 'rewards/count_adherence/mean': 0.6428571939468384, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8928571343421936, 'rewards/count_floor/std': 0.022908121347427368, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9028449058532715, 'rewards/repeat_penalty/std': 0.037171151489019394, 'rewards/repeat_floor/mean': 0.9894146919250488, 'rewards/repeat_floor/std': 0.0042105065658688545, 'rewards/near_duplicate_penalty/mean': 0.9724725484848022, 'rewards/near_duplicate_penalty/std': 0.01640428975224495, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9282649755477905, 'rewards/distinct_2/std': 0.030583176761865616, 'rewards/total_composite/mean': 0.11097601056098938, 'rewards/total_composite/std': 0.11250045150518417, 'reward': 0.11097601056098938, 'reward_std': 0.11250045150518417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14996173977851868, 'sampling/sampling_logp_difference/max': 3.206545829772949, 'sampling/importance_sampling_ratio/min': 0.040496256202459335, 'sampling/importance_sampling_ratio/mean': 1.0081321001052856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9402342960238457, 'clip_ratio/low_mean': 0.0886575486510992, 'clip_ratio/low_min': 0.0886575486510992, 'clip_ratio/high_mean': 0.030782858841121197, 'clip_ratio/high_max': 0.030782858841121197, 'clip_ratio/region_mean': 0.1194404074922204, 'reward_total_mean': 0.11097601056098938, 'reward_meter_mean': 0.31687265634536743, 'reward_meter_std': 0.2848235070705414, 'reward_count_adherence_mean': 0.6428571939468384, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8928571343421936, 'reward_count_floor_std': 0.022908121347427368, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9028449058532715, 'reward_repeat_penalty_std': 0.037171151489019394, 'reward_repeat_floor_mean': 0.9894146919250488, 'reward_repeat_floor_std': 0.0042105065658688545, 'reward_near_duplicate_penalty_mean': 0.9724725484848022, 'reward_near_duplicate_penalty_std': 0.01640428975224495, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9282649755477905, 'reward_distinct_2_std': 0.030583176761865616, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.11097601056098938, 'reward_total_composite_std': 0.11250045150518417, 'epoch': 0.07}

 41%|████▏     | 1365/3300 [5:13:18<6:16:33, 11.68s/it]INFO 04-12 21:11:44 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1366/3300 [5:13:31<6:35:52, 12.28s/it]2026-04-12 21:11:58,120 | INFO | train_grpo_train | metrics_logged mode=train step=1366

                                                       
{'loss': 0.0196, 'grad_norm': 14.12635326385498, 'learning_rate': 5.863636363636364e-06, 'num_tokens': 2418917.0, 'completions/mean_length': 32.875, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.4733375906944275, 'rewards/meter/std': 0.3433862030506134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6425000429153442, 'rewards/judge_quality/std': 0.21558891236782074, 'rewards/repeat_penalty/mean': 0.9666086435317993, 'rewards/repeat_penalty/std': 0.05108451470732689, 'rewards/repeat_floor/mean': 0.9956244230270386, 'rewards/repeat_floor/std': 0.006222573108971119, 'rewards/near_duplicate_penalty/mean': 0.9780410528182983, 'rewards/near_duplicate_penalty/std': 0.02936101146042347, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9879807829856873, 'rewards/distinct_2/std': 0.03399552404880524, 'rewards/total_composite/mean': 0.2751588225364685, 'rewards/total_composite/std': 0.21421192586421967, 'reward': 0.2751588225364685, 'reward_std': 0.21421192586421967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13260090351104736, 'sampling/sampling_logp_difference/max': 1.8856416940689087, 'sampling/importance_sampling_ratio/min': 0.1517316699028015, 'sampling/importance_sampling_ratio/mean': 1.02219557762146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8842187747359276, 'clip_ratio/low_mean': 0.05132781062275171, 'clip_ratio/low_min': 0.05132781062275171, 'clip_ratio/high_mean': 0.07079154346138239, 'clip_ratio/high_max': 0.07079154346138239, 'clip_ratio/region_mean': 0.1221193540841341, 'reward_total_mean': 0.2751588225364685, 'reward_meter_mean': 0.4733375906944275, 'reward_meter_std': 0.3433862030506134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9666086435317993, 'reward_repeat_penalty_std': 0.05108451470732689, 'reward_repeat_floor_mean': 0.9956244230270386, 'reward_repeat_floor_std': 0.006222573108971119, 'reward_near_duplicate_penalty_mean': 0.9780410528182983, 'reward_near_duplicate_penalty_std': 0.02936101146042347, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9879807829856873, 'reward_distinct_2_std': 0.03399552404880524, 'reward_judge_quality_mean': 0.6425000429153442, 'reward_judge_quality_std': 0.21558891236782074, 'reward_total_composite_mean': 0.2751588225364685, 'reward_total_composite_std': 0.21421192586421967, 'epoch': 0.07}

 41%|████▏     | 1366/3300 [5:13:31<6:35:52, 12.28s/it]INFO 04-12 21:11:58 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1367/3300 [5:13:45<6:45:47, 12.60s/it]2026-04-12 21:12:11,468 | INFO | train_grpo_train | metrics_logged mode=train step=1367

                                                       
{'loss': -0.0389, 'grad_norm': 4.24796199798584, 'learning_rate': 5.860606060606061e-06, 'num_tokens': 2420577.0, 'completions/mean_length': 100.5, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.71428680419922, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.18226322531700134, 'rewards/meter/std': 0.18298299610614777, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9473918676376343, 'rewards/lexical_plausibility/std': 0.1269720047712326, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9929934144020081, 'rewards/repeat_penalty/std': 0.00786297582089901, 'rewards/repeat_floor/mean': 0.9989490509033203, 'rewards/repeat_floor/std': 0.0011794412275776267, 'rewards/near_duplicate_penalty/mean': 0.9929934144020081, 'rewards/near_duplicate_penalty/std': 0.00786297582089901, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.050226546823978424, 'rewards/total_composite/std': 0.06132219359278679, 'reward': 0.050226546823978424, 'reward_std': 0.06132219359278679, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14662757515907288, 'sampling/sampling_logp_difference/max': 1.2821555137634277, 'sampling/importance_sampling_ratio/min': 0.2774386405944824, 'sampling/importance_sampling_ratio/mean': 1.0189090967178345, 'sampling/importance_sampling_ratio/max': 1.8976069688796997, 'entropy': 1.428486630320549, 'clip_ratio/low_mean': 0.07684226986020803, 'clip_ratio/low_min': 0.07684226986020803, 'clip_ratio/high_mean': 0.04945652186870575, 'clip_ratio/high_max': 0.04945652186870575, 'clip_ratio/region_mean': 0.12629879172891378, 'reward_total_mean': 0.050226546823978424, 'reward_meter_mean': 0.18226322531700134, 'reward_meter_std': 0.18298299610614777, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9473918676376343, 'reward_lexical_plausibility_std': 0.1269720047712326, 'reward_repeat_penalty_mean': 0.9929934144020081, 'reward_repeat_penalty_std': 0.00786297582089901, 'reward_repeat_floor_mean': 0.9989490509033203, 'reward_repeat_floor_std': 0.0011794412275776267, 'reward_near_duplicate_penalty_mean': 0.9929934144020081, 'reward_near_duplicate_penalty_std': 0.00786297582089901, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.050226546823978424, 'reward_total_composite_std': 0.06132219359278679, 'epoch': 0.07}

 41%|████▏     | 1367/3300 [5:13:45<6:45:47, 12.60s/it]INFO 04-12 21:12:11 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1368/3300 [5:13:54<6:13:02, 11.59s/it]2026-04-12 21:12:20,694 | INFO | train_grpo_train | metrics_logged mode=train step=1368

                                                       
{'loss': 0.0017, 'grad_norm': 18.617734909057617, 'learning_rate': 5.8575757575757584e-06, 'num_tokens': 2421965.0, 'completions/mean_length': 35.5, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.228774756193161, 'rewards/meter/std': 0.2565818428993225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5087499618530273, 'rewards/judge_quality/std': 0.2811424434185028, 'rewards/repeat_penalty/mean': 0.9844053983688354, 'rewards/repeat_penalty/std': 0.017846377566456795, 'rewards/repeat_floor/mean': 0.9976608157157898, 'rewards/repeat_floor/std': 0.0026769612450152636, 'rewards/near_duplicate_penalty/mean': 0.9844053983688354, 'rewards/near_duplicate_penalty/std': 0.017846377566456795, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.13428552448749542, 'rewards/total_composite/std': 0.21110010147094727, 'reward': 0.13428552448749542, 'reward_std': 0.21110008656978607, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1675732284784317, 'sampling/sampling_logp_difference/max': 1.9700037240982056, 'sampling/importance_sampling_ratio/min': 0.13945633172988892, 'sampling/importance_sampling_ratio/mean': 1.0281715393066406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2026347368955612, 'clip_ratio/low_mean': 0.0811789813451469, 'clip_ratio/low_min': 0.0811789813451469, 'clip_ratio/high_mean': 0.020270270761102438, 'clip_ratio/high_max': 0.020270270761102438, 'clip_ratio/region_mean': 0.10144925210624933, 'reward_total_mean': 0.13428552448749542, 'reward_meter_mean': 0.228774756193161, 'reward_meter_std': 0.2565818428993225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9844053983688354, 'reward_repeat_penalty_std': 0.017846377566456795, 'reward_repeat_floor_mean': 0.9976608157157898, 'reward_repeat_floor_std': 0.0026769612450152636, 'reward_near_duplicate_penalty_mean': 0.9844053983688354, 'reward_near_duplicate_penalty_std': 0.017846377566456795, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5087499618530273, 'reward_judge_quality_std': 0.2811424434185028, 'reward_total_composite_mean': 0.13428552448749542, 'reward_total_composite_std': 0.21110010147094727, 'epoch': 0.07}

 41%|████▏     | 1368/3300 [5:13:54<6:13:02, 11.59s/it]INFO 04-12 21:12:20 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1369/3300 [5:14:03<5:51:22, 10.92s/it]2026-04-12 21:12:30,050 | INFO | train_grpo_train | metrics_logged mode=train step=1369

                                                       
{'loss': 0.0474, 'grad_norm': 10.18528938293457, 'learning_rate': 5.854545454545455e-06, 'num_tokens': 2423826.0, 'completions/mean_length': 60.625, 'completions/min_length': 54.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.5438340902328491, 'rewards/meter/std': 0.38540804386138916, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.8463025093078613, 'rewards/repeat_penalty/std': 0.10108055174350739, 'rewards/repeat_floor/mean': 0.9810028076171875, 'rewards/repeat_floor/std': 0.011964036151766777, 'rewards/near_duplicate_penalty/mean': 0.9267818927764893, 'rewards/near_duplicate_penalty/std': 0.03889380395412445, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9109505414962769, 'rewards/distinct_2/std': 0.07523106783628464, 'rewards/total_composite/mean': 0.23085661232471466, 'rewards/total_composite/std': 0.1908019483089447, 'reward': 0.23085661232471466, 'reward_std': 0.1908019483089447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1349177211523056, 'sampling/sampling_logp_difference/max': 1.3860890865325928, 'sampling/importance_sampling_ratio/min': 0.2656486928462982, 'sampling/importance_sampling_ratio/mean': 1.022133708000183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9961529895663261, 'clip_ratio/low_mean': 0.0491155069321394, 'clip_ratio/low_min': 0.0491155069321394, 'clip_ratio/high_mean': 0.07743468973785639, 'clip_ratio/high_max': 0.07743468973785639, 'clip_ratio/region_mean': 0.12655019666999578, 'reward_total_mean': 0.23085661232471466, 'reward_meter_mean': 0.5438340902328491, 'reward_meter_std': 0.38540804386138916, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8463025093078613, 'reward_repeat_penalty_std': 0.10108055174350739, 'reward_repeat_floor_mean': 0.9810028076171875, 'reward_repeat_floor_std': 0.011964036151766777, 'reward_near_duplicate_penalty_mean': 0.9267818927764893, 'reward_near_duplicate_penalty_std': 0.03889380395412445, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9109505414962769, 'reward_distinct_2_std': 0.07523106783628464, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.23085661232471466, 'reward_total_composite_std': 0.1908019483089447, 'epoch': 0.07}

 41%|████▏     | 1369/3300 [5:14:03<5:51:22, 10.92s/it]INFO 04-12 21:12:30 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1370/3300 [5:14:13<5:34:58, 10.41s/it]2026-04-12 21:12:39,296 | INFO | train_grpo_train | metrics_logged mode=train step=1370

                                                       
{'loss': 0.0216, 'grad_norm': 10.985562324523926, 'learning_rate': 5.851515151515152e-06, 'num_tokens': 2425781.0, 'completions/mean_length': 68.375, 'completions/min_length': 42.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.7853902578353882, 'rewards/meter/std': 0.345671147108078, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9756944179534912, 'rewards/lexical_plausibility/std': 0.058338064700365067, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.7724920511245728, 'rewards/repeat_penalty/std': 0.17106185853481293, 'rewards/repeat_floor/mean': 0.9715359807014465, 'rewards/repeat_floor/std': 0.023085223510861397, 'rewards/near_duplicate_penalty/mean': 0.9013967514038086, 'rewards/near_duplicate_penalty/std': 0.08656418323516846, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8480718731880188, 'rewards/distinct_2/std': 0.11974718421697617, 'rewards/total_composite/mean': 0.20746755599975586, 'rewards/total_composite/std': 0.1367490440607071, 'reward': 0.20746755599975586, 'reward_std': 0.1367490440607071, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14865341782569885, 'sampling/sampling_logp_difference/max': 1.4660074710845947, 'sampling/importance_sampling_ratio/min': 0.23084531724452972, 'sampling/importance_sampling_ratio/mean': 1.0330126285552979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.332340568304062, 'clip_ratio/low_mean': 0.0659127002581954, 'clip_ratio/low_min': 0.0659127002581954, 'clip_ratio/high_mean': 0.0766943134367466, 'clip_ratio/high_max': 0.0766943134367466, 'clip_ratio/region_mean': 0.142607013694942, 'reward_total_mean': 0.20746755599975586, 'reward_meter_mean': 0.7853902578353882, 'reward_meter_std': 0.345671147108078, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9756944179534912, 'reward_lexical_plausibility_std': 0.058338064700365067, 'reward_repeat_penalty_mean': 0.7724920511245728, 'reward_repeat_penalty_std': 0.17106185853481293, 'reward_repeat_floor_mean': 0.9715359807014465, 'reward_repeat_floor_std': 0.023085223510861397, 'reward_near_duplicate_penalty_mean': 0.9013967514038086, 'reward_near_duplicate_penalty_std': 0.08656418323516846, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8480718731880188, 'reward_distinct_2_std': 0.11974718421697617, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.20746755599975586, 'reward_total_composite_std': 0.1367490440607071, 'epoch': 0.07}

 42%|████▏     | 1370/3300 [5:14:13<5:34:58, 10.41s/it]INFO 04-12 21:12:39 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1371/3300 [5:14:22<5:24:08, 10.08s/it]2026-04-12 21:12:48,607 | INFO | train_grpo_train | metrics_logged mode=train step=1371

                                                       
{'loss': 0.0238, 'grad_norm': 12.143315315246582, 'learning_rate': 5.8484848484848485e-06, 'num_tokens': 2427545.0, 'completions/mean_length': 64.5, 'completions/min_length': 62.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.22518661618232727, 'rewards/meter/std': 0.30453017354011536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9323565363883972, 'rewards/repeat_penalty/std': 0.05915520712733269, 'rewards/repeat_floor/mean': 0.9920791387557983, 'rewards/repeat_floor/std': 0.00636344263330102, 'rewards/near_duplicate_penalty/mean': 0.9719240665435791, 'rewards/near_duplicate_penalty/std': 0.016151413321495056, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9587832689285278, 'rewards/distinct_2/std': 0.0489928312599659, 'rewards/total_composite/mean': 0.10005968064069748, 'rewards/total_composite/std': 0.13596297800540924, 'reward': 0.10005968064069748, 'reward_std': 0.13596297800540924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13936345279216766, 'sampling/sampling_logp_difference/max': 2.1678781509399414, 'sampling/importance_sampling_ratio/min': 0.11442014575004578, 'sampling/importance_sampling_ratio/mean': 1.0268821716308594, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0624381750822067, 'clip_ratio/low_mean': 0.08691003918647766, 'clip_ratio/low_min': 0.08691003918647766, 'clip_ratio/high_mean': 0.028846153989434242, 'clip_ratio/high_max': 0.028846153989434242, 'clip_ratio/region_mean': 0.1157561931759119, 'reward_total_mean': 0.10005968064069748, 'reward_meter_mean': 0.22518661618232727, 'reward_meter_std': 0.30453017354011536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9323565363883972, 'reward_repeat_penalty_std': 0.05915520712733269, 'reward_repeat_floor_mean': 0.9920791387557983, 'reward_repeat_floor_std': 0.00636344263330102, 'reward_near_duplicate_penalty_mean': 0.9719240665435791, 'reward_near_duplicate_penalty_std': 0.016151413321495056, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9587832689285278, 'reward_distinct_2_std': 0.0489928312599659, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.10005968064069748, 'reward_total_composite_std': 0.13596297800540924, 'epoch': 0.07}

 42%|████▏     | 1371/3300 [5:14:22<5:24:08, 10.08s/it]INFO 04-12 21:12:48 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1372/3300 [5:14:31<5:17:42,  9.89s/it]2026-04-12 21:12:58,043 | INFO | train_grpo_train | metrics_logged mode=train step=1372

                                                       
{'loss': 0.0828, 'grad_norm': 8.046740531921387, 'learning_rate': 5.845454545454547e-06, 'num_tokens': 2429674.0, 'completions/mean_length': 81.125, 'completions/min_length': 68.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.18508780002593994, 'rewards/meter/std': 0.32934829592704773, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.18795421719551086, 'rewards/repeat_penalty/mean': 0.9501427412033081, 'rewards/repeat_penalty/std': 0.03616981953382492, 'rewards/repeat_floor/mean': 0.9939243197441101, 'rewards/repeat_floor/std': 0.003354514017701149, 'rewards/near_duplicate_penalty/mean': 0.9742391109466553, 'rewards/near_duplicate_penalty/std': 0.012213360518217087, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9754273891448975, 'rewards/distinct_2/std': 0.03964797407388687, 'rewards/total_composite/mean': 0.1263751983642578, 'rewards/total_composite/std': 0.2864478528499603, 'reward': 0.1263751983642578, 'reward_std': 0.2864478528499603, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16321052610874176, 'sampling/sampling_logp_difference/max': 1.7143049240112305, 'sampling/importance_sampling_ratio/min': 0.1800888627767563, 'sampling/importance_sampling_ratio/mean': 1.0199953317642212, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1901763826608658, 'clip_ratio/low_mean': 0.12560637015849352, 'clip_ratio/low_min': 0.12560637015849352, 'clip_ratio/high_mean': 0.011029412038624287, 'clip_ratio/high_max': 0.011029412038624287, 'clip_ratio/region_mean': 0.1366357821971178, 'reward_total_mean': 0.1263751983642578, 'reward_meter_mean': 0.18508780002593994, 'reward_meter_std': 0.32934829592704773, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9501427412033081, 'reward_repeat_penalty_std': 0.03616981953382492, 'reward_repeat_floor_mean': 0.9939243197441101, 'reward_repeat_floor_std': 0.003354514017701149, 'reward_near_duplicate_penalty_mean': 0.9742391109466553, 'reward_near_duplicate_penalty_std': 0.012213360518217087, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9754273891448975, 'reward_distinct_2_std': 0.03964797407388687, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.18795421719551086, 'reward_total_composite_mean': 0.1263751983642578, 'reward_total_composite_std': 0.2864478528499603, 'epoch': 0.07}

 42%|████▏     | 1372/3300 [5:14:31<5:17:42,  9.89s/it]INFO 04-12 21:12:58 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1373/3300 [5:14:40<5:07:11,  9.56s/it]2026-04-12 21:13:06,934 | INFO | train_grpo_train | metrics_logged mode=train step=1373

                                                       
{'loss': 0.0445, 'grad_norm': 12.134416580200195, 'learning_rate': 5.842424242424243e-06, 'num_tokens': 2431360.0, 'completions/mean_length': 38.75, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.10883817076683044, 'rewards/meter/std': 0.13237561285495758, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.28630343079566956, 'rewards/repeat_penalty/mean': 0.9554584622383118, 'rewards/repeat_penalty/std': 0.06852570176124573, 'rewards/repeat_floor/mean': 0.9942131638526917, 'rewards/repeat_floor/std': 0.007829461246728897, 'rewards/near_duplicate_penalty/mean': 0.9724100828170776, 'rewards/near_duplicate_penalty/std': 0.024116821587085724, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9816849827766418, 'rewards/distinct_2/std': 0.05180269479751587, 'rewards/total_composite/mean': 0.06725022196769714, 'rewards/total_composite/std': 0.09920899569988251, 'reward': 0.06725022196769714, 'reward_std': 0.0992090031504631, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13775679469108582, 'sampling/sampling_logp_difference/max': 1.2311906814575195, 'sampling/importance_sampling_ratio/min': 0.2919447720050812, 'sampling/importance_sampling_ratio/mean': 1.0210661888122559, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.348797246813774, 'clip_ratio/low_mean': 0.11056662816554308, 'clip_ratio/low_min': 0.11056662816554308, 'clip_ratio/high_mean': 0.031759907491505146, 'clip_ratio/high_max': 0.031759907491505146, 'clip_ratio/region_mean': 0.14232653565704823, 'reward_total_mean': 0.06725022196769714, 'reward_meter_mean': 0.10883817076683044, 'reward_meter_std': 0.13237561285495758, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9554584622383118, 'reward_repeat_penalty_std': 0.06852570176124573, 'reward_repeat_floor_mean': 0.9942131638526917, 'reward_repeat_floor_std': 0.007829461246728897, 'reward_near_duplicate_penalty_mean': 0.9724100828170776, 'reward_near_duplicate_penalty_std': 0.024116821587085724, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9816849827766418, 'reward_distinct_2_std': 0.05180269479751587, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.28630343079566956, 'reward_total_composite_mean': 0.06725022196769714, 'reward_total_composite_std': 0.09920899569988251, 'epoch': 0.07}

 42%|████▏     | 1373/3300 [5:14:40<5:07:11,  9.56s/it]INFO 04-12 21:13:07 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1374/3300 [5:14:49<4:57:09,  9.26s/it]2026-04-12 21:13:15,385 | INFO | train_grpo_train | metrics_logged mode=train step=1374

                                                       
{'loss': 0.0664, 'grad_norm': 14.10013484954834, 'learning_rate': 5.83939393939394e-06, 'num_tokens': 2432937.0, 'completions/mean_length': 34.125, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9319015145301819, 'rewards/meter/std': 0.13963213562965393, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9271719455718994, 'rewards/repeat_penalty/std': 0.12586134672164917, 'rewards/repeat_floor/mean': 0.991751492023468, 'rewards/repeat_floor/std': 0.012157819233834743, 'rewards/near_duplicate_penalty/mean': 0.9642972946166992, 'rewards/near_duplicate_penalty/std': 0.04767512530088425, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.37901413440704346, 'rewards/total_composite/std': 0.06541632860898972, 'reward': 0.37901413440704346, 'reward_std': 0.06541632860898972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1121741458773613, 'sampling/sampling_logp_difference/max': 2.3811964988708496, 'sampling/importance_sampling_ratio/min': 0.09243990480899811, 'sampling/importance_sampling_ratio/mean': 1.0020999908447266, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6387466639280319, 'clip_ratio/low_mean': 0.03424250823445618, 'clip_ratio/low_min': 0.03424250823445618, 'clip_ratio/high_mean': 0.04973118286579847, 'clip_ratio/high_max': 0.04973118286579847, 'clip_ratio/region_mean': 0.08397369110025465, 'reward_total_mean': 0.37901413440704346, 'reward_meter_mean': 0.9319015145301819, 'reward_meter_std': 0.13963213562965393, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9271719455718994, 'reward_repeat_penalty_std': 0.12586134672164917, 'reward_repeat_floor_mean': 0.991751492023468, 'reward_repeat_floor_std': 0.012157819233834743, 'reward_near_duplicate_penalty_mean': 0.9642972946166992, 'reward_near_duplicate_penalty_std': 0.04767512530088425, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.37901413440704346, 'reward_total_composite_std': 0.06541632860898972, 'epoch': 0.07}

 42%|████▏     | 1374/3300 [5:14:49<4:57:09,  9.26s/it]INFO 04-12 21:13:15 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1375/3300 [5:15:03<5:43:32, 10.71s/it]2026-04-12 21:13:29,475 | INFO | train_grpo_train | metrics_logged mode=train step=1375

                                                       
{'loss': -0.0394, 'grad_norm': 2.623197317123413, 'learning_rate': 5.836363636363637e-06, 'num_tokens': 2434423.0, 'completions/mean_length': 154.75, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 35.66666793823242, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.44823598861694336, 'rewards/meter/std': 0.38567957282066345, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8820353746414185, 'rewards/lexical_plausibility/std': 0.21875672042369843, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.19537052512168884, 'rewards/repeat_penalty/mean': 0.9160169959068298, 'rewards/repeat_penalty/std': 0.09169948101043701, 'rewards/repeat_floor/mean': 0.9906015396118164, 'rewards/repeat_floor/std': 0.0074348971247673035, 'rewards/near_duplicate_penalty/mean': 0.9565833210945129, 'rewards/near_duplicate_penalty/std': 0.0336182564496994, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9887669086456299, 'rewards/distinct_2/std': 0.028349952772259712, 'rewards/total_composite/mean': 0.12254895269870758, 'rewards/total_composite/std': 0.15505245327949524, 'reward': 0.12254895269870758, 'reward_std': 0.15505243837833405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13303308188915253, 'sampling/sampling_logp_difference/max': 0.940190315246582, 'sampling/importance_sampling_ratio/min': 0.3905535042285919, 'sampling/importance_sampling_ratio/mean': 1.0239512920379639, 'sampling/importance_sampling_ratio/max': 1.8268417119979858, 'entropy': 0.7729804664850235, 'clip_ratio/low_mean': 0.06234984891489148, 'clip_ratio/low_min': 0.06234984891489148, 'clip_ratio/high_mean': 0.03040540497750044, 'clip_ratio/high_max': 0.03040540497750044, 'clip_ratio/region_mean': 0.09275525389239192, 'reward_total_mean': 0.12254895269870758, 'reward_meter_mean': 0.44823598861694336, 'reward_meter_std': 0.38567957282066345, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8820353746414185, 'reward_lexical_plausibility_std': 0.21875672042369843, 'reward_repeat_penalty_mean': 0.9160169959068298, 'reward_repeat_penalty_std': 0.09169948101043701, 'reward_repeat_floor_mean': 0.9906015396118164, 'reward_repeat_floor_std': 0.0074348971247673035, 'reward_near_duplicate_penalty_mean': 0.9565833210945129, 'reward_near_duplicate_penalty_std': 0.0336182564496994, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9887669086456299, 'reward_distinct_2_std': 0.028349952772259712, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.19537052512168884, 'reward_total_composite_mean': 0.12254895269870758, 'reward_total_composite_std': 0.15505245327949524, 'epoch': 0.07}

 42%|████▏     | 1375/3300 [5:15:03<5:43:32, 10.71s/it]INFO 04-12 21:13:29 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1376/3300 [5:15:17<6:16:15, 11.73s/it]2026-04-12 21:13:43,613 | INFO | train_grpo_train | metrics_logged mode=train step=1376

                                                       
{'loss': -0.0935, 'grad_norm': 6.4571051597595215, 'learning_rate': 5.833333333333334e-06, 'num_tokens': 2436490.0, 'completions/mean_length': 139.375, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.14286041259766, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.39170604944229126, 'rewards/meter/std': 0.266248881816864, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.01767767407000065, 'rewards/lexical_plausibility/mean': 0.9198315143585205, 'rewards/lexical_plausibility/std': 0.15002666413784027, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.8627588748931885, 'rewards/repeat_penalty/std': 0.09677524119615555, 'rewards/repeat_floor/mean': 0.9852653741836548, 'rewards/repeat_floor/std': 0.009832357987761497, 'rewards/near_duplicate_penalty/mean': 0.9551376104354858, 'rewards/near_duplicate_penalty/std': 0.03070751577615738, 'rewards/opening_diversity/mean': 0.9624999761581421, 'rewards/opening_diversity/std': 0.05175492912530899, 'rewards/distinct_2/mean': 0.9360523223876953, 'rewards/distinct_2/std': 0.051112402230501175, 'rewards/total_composite/mean': 0.09740854799747467, 'rewards/total_composite/std': 0.07821636646986008, 'reward': 0.09740854799747467, 'reward_std': 0.07821635901927948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16103622317314148, 'sampling/sampling_logp_difference/max': 3.0257763862609863, 'sampling/importance_sampling_ratio/min': 0.04852013662457466, 'sampling/importance_sampling_ratio/mean': 1.0246303081512451, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9950046166777611, 'clip_ratio/low_mean': 0.058809525333344936, 'clip_ratio/low_min': 0.058809525333344936, 'clip_ratio/high_mean': 0.06535284221172333, 'clip_ratio/high_max': 0.06535284221172333, 'clip_ratio/region_mean': 0.12416236754506826, 'reward_total_mean': 0.09740854799747467, 'reward_meter_mean': 0.39170604944229126, 'reward_meter_std': 0.266248881816864, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.01767767407000065, 'reward_lexical_plausibility_mean': 0.9198315143585205, 'reward_lexical_plausibility_std': 0.15002666413784027, 'reward_repeat_penalty_mean': 0.8627588748931885, 'reward_repeat_penalty_std': 0.09677524119615555, 'reward_repeat_floor_mean': 0.9852653741836548, 'reward_repeat_floor_std': 0.009832357987761497, 'reward_near_duplicate_penalty_mean': 0.9551376104354858, 'reward_near_duplicate_penalty_std': 0.03070751577615738, 'reward_opening_diversity_mean': 0.9624999761581421, 'reward_opening_diversity_std': 0.05175492912530899, 'reward_distinct_2_mean': 0.9360523223876953, 'reward_distinct_2_std': 0.051112402230501175, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.09740854799747467, 'reward_total_composite_std': 0.07821636646986008, 'epoch': 0.07}

 42%|████▏     | 1376/3300 [5:15:17<6:16:15, 11.73s/it]INFO 04-12 21:13:43 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1377/3300 [5:15:31<6:41:57, 12.54s/it]2026-04-12 21:13:58,031 | INFO | train_grpo_train | metrics_logged mode=train step=1377

                                                       
{'loss': -0.0428, 'grad_norm': 3.8730804920196533, 'learning_rate': 5.83030303030303e-06, 'num_tokens': 2438551.0, 'completions/mean_length': 190.625, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 83.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.2339208424091339, 'rewards/meter/std': 0.3111248314380646, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.04008918255567551, 'rewards/lexical_plausibility/mean': 0.8983363509178162, 'rewards/lexical_plausibility/std': 0.1886088103055954, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.9519809484481812, 'rewards/repeat_penalty/std': 0.05712372064590454, 'rewards/repeat_floor/mean': 0.9948340654373169, 'rewards/repeat_floor/std': 0.005274706520140171, 'rewards/near_duplicate_penalty/mean': 0.9842805862426758, 'rewards/near_duplicate_penalty/std': 0.012290792539715767, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9740083813667297, 'rewards/distinct_2/std': 0.03490973263978958, 'rewards/total_composite/mean': 0.07561470568180084, 'rewards/total_composite/std': 0.1330641657114029, 'reward': 0.07561470568180084, 'reward_std': 0.1330641657114029, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14303164184093475, 'sampling/sampling_logp_difference/max': 1.6218986511230469, 'sampling/importance_sampling_ratio/min': 0.1975233256816864, 'sampling/importance_sampling_ratio/mean': 1.0300137996673584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9534309580922127, 'clip_ratio/low_mean': 0.07406182214617729, 'clip_ratio/low_min': 0.07406182214617729, 'clip_ratio/high_mean': 0.010416666977107525, 'clip_ratio/high_max': 0.010416666977107525, 'clip_ratio/region_mean': 0.08447848912328482, 'reward_total_mean': 0.07561470568180084, 'reward_meter_mean': 0.2339208424091339, 'reward_meter_std': 0.3111248314380646, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.04008918255567551, 'reward_lexical_plausibility_mean': 0.8983363509178162, 'reward_lexical_plausibility_std': 0.1886088103055954, 'reward_repeat_penalty_mean': 0.9519809484481812, 'reward_repeat_penalty_std': 0.05712372064590454, 'reward_repeat_floor_mean': 0.9948340654373169, 'reward_repeat_floor_std': 0.005274706520140171, 'reward_near_duplicate_penalty_mean': 0.9842805862426758, 'reward_near_duplicate_penalty_std': 0.012290792539715767, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9740083813667297, 'reward_distinct_2_std': 0.03490973263978958, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.07561470568180084, 'reward_total_composite_std': 0.1330641657114029, 'epoch': 0.07}

 42%|████▏     | 1377/3300 [5:15:31<6:41:57, 12.54s/it]INFO 04-12 21:13:58 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1378/3300 [5:15:40<6:01:27, 11.28s/it]2026-04-12 21:14:06,392 | INFO | train_grpo_train | metrics_logged mode=train step=1378

                                                       
{'loss': 0.0106, 'grad_norm': 12.528951644897461, 'learning_rate': 5.8272727272727285e-06, 'num_tokens': 2440160.0, 'completions/mean_length': 35.125, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.11679422110319138, 'rewards/meter/std': 0.10038521885871887, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.3586059510707855, 'rewards/repeat_penalty/mean': 0.9802839756011963, 'rewards/repeat_penalty/std': 0.028267215937376022, 'rewards/repeat_floor/mean': 0.9970425963401794, 'rewards/repeat_floor/std': 0.004240076523274183, 'rewards/near_duplicate_penalty/mean': 0.9802839756011963, 'rewards/near_duplicate_penalty/std': 0.028267215937376022, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.07801657170057297, 'rewards/total_composite/std': 0.10824630409479141, 'reward': 0.07801657170057297, 'reward_std': 0.10824630409479141, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1653551459312439, 'sampling/sampling_logp_difference/max': 1.6351332664489746, 'sampling/importance_sampling_ratio/min': 0.1949263960123062, 'sampling/importance_sampling_ratio/mean': 1.0139355659484863, 'sampling/importance_sampling_ratio/max': 1.9271697998046875, 'entropy': 1.1497474610805511, 'clip_ratio/low_mean': 0.06785545218735933, 'clip_ratio/low_min': 0.06785545218735933, 'clip_ratio/high_mean': 0.03943745046854019, 'clip_ratio/high_max': 0.03943745046854019, 'clip_ratio/region_mean': 0.10729290265589952, 'reward_total_mean': 0.07801657170057297, 'reward_meter_mean': 0.11679422110319138, 'reward_meter_std': 0.10038521885871887, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9802839756011963, 'reward_repeat_penalty_std': 0.028267215937376022, 'reward_repeat_floor_mean': 0.9970425963401794, 'reward_repeat_floor_std': 0.004240076523274183, 'reward_near_duplicate_penalty_mean': 0.9802839756011963, 'reward_near_duplicate_penalty_std': 0.028267215937376022, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.3586059510707855, 'reward_total_composite_mean': 0.07801657170057297, 'reward_total_composite_std': 0.10824630409479141, 'epoch': 0.07}

 42%|████▏     | 1378/3300 [5:15:40<6:01:27, 11.28s/it]INFO 04-12 21:14:06 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1379/3300 [5:15:54<6:30:28, 12.20s/it]2026-04-12 21:14:20,715 | INFO | train_grpo_train | metrics_logged mode=train step=1379

                                                       
{'loss': -0.0275, 'grad_norm': 6.189452648162842, 'learning_rate': 5.824242424242425e-06, 'num_tokens': 2441908.0, 'completions/mean_length': 116.5, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.000003814697266, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4092772603034973, 'rewards/meter/std': 0.3240372836589813, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9319329261779785, 'rewards/lexical_plausibility/std': 0.1623733937740326, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.1927248239517212, 'rewards/repeat_penalty/mean': 0.9625256061553955, 'rewards/repeat_penalty/std': 0.04014236479997635, 'rewards/repeat_floor/mean': 0.9951406121253967, 'rewards/repeat_floor/std': 0.0049741859547793865, 'rewards/near_duplicate_penalty/mean': 0.9759984612464905, 'rewards/near_duplicate_penalty/std': 0.024664591997861862, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9860093593597412, 'rewards/distinct_2/std': 0.025911342352628708, 'rewards/total_composite/mean': 0.19492483139038086, 'rewards/total_composite/std': 0.23800775408744812, 'reward': 0.19492483139038086, 'reward_std': 0.23800775408744812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15827898681163788, 'sampling/sampling_logp_difference/max': 1.7638483047485352, 'sampling/importance_sampling_ratio/min': 0.1713840514421463, 'sampling/importance_sampling_ratio/mean': 1.0424138307571411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9292081221938133, 'clip_ratio/low_mean': 0.059855757281184196, 'clip_ratio/low_min': 0.059855757281184196, 'clip_ratio/high_mean': 0.05416002217680216, 'clip_ratio/high_max': 0.05416002217680216, 'clip_ratio/region_mean': 0.11401577945798635, 'reward_total_mean': 0.19492483139038086, 'reward_meter_mean': 0.4092772603034973, 'reward_meter_std': 0.3240372836589813, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9319329261779785, 'reward_lexical_plausibility_std': 0.1623733937740326, 'reward_repeat_penalty_mean': 0.9625256061553955, 'reward_repeat_penalty_std': 0.04014236479997635, 'reward_repeat_floor_mean': 0.9951406121253967, 'reward_repeat_floor_std': 0.0049741859547793865, 'reward_near_duplicate_penalty_mean': 0.9759984612464905, 'reward_near_duplicate_penalty_std': 0.024664591997861862, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9860093593597412, 'reward_distinct_2_std': 0.025911342352628708, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.1927248239517212, 'reward_total_composite_mean': 0.19492483139038086, 'reward_total_composite_std': 0.23800775408744812, 'epoch': 0.07}

 42%|████▏     | 1379/3300 [5:15:54<6:30:28, 12.20s/it]INFO 04-12 21:14:20 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1380/3300 [5:16:02<5:52:30, 11.02s/it]2026-04-12 21:14:28,981 | INFO | train_grpo_train | metrics_logged mode=train step=1380

                                                       
{'loss': -0.0042, 'grad_norm': 11.943592071533203, 'learning_rate': 5.821212121212122e-06, 'num_tokens': 2443402.0, 'completions/mean_length': 33.75, 'completions/min_length': 31.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5656245350837708, 'rewards/meter/std': 0.3941541910171509, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8062499761581421, 'rewards/judge_quality/std': 0.22012577950954437, 'rewards/repeat_penalty/mean': 0.981610894203186, 'rewards/repeat_penalty/std': 0.018419211730360985, 'rewards/repeat_floor/mean': 0.9972416162490845, 'rewards/repeat_floor/std': 0.0027628839015960693, 'rewards/near_duplicate_penalty/mean': 0.981610894203186, 'rewards/near_duplicate_penalty/std': 0.018419211730360985, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4673754572868347, 'rewards/total_composite/std': 0.3525567948818207, 'reward': 0.4673754572868347, 'reward_std': 0.35255682468414307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10308034718036652, 'sampling/sampling_logp_difference/max': 1.1686080694198608, 'sampling/importance_sampling_ratio/min': 0.3700951337814331, 'sampling/importance_sampling_ratio/mean': 1.032747745513916, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6752573549747467, 'clip_ratio/low_mean': 0.08257925044745207, 'clip_ratio/low_min': 0.08257925044745207, 'clip_ratio/high_mean': 0.05479426775127649, 'clip_ratio/high_max': 0.05479426775127649, 'clip_ratio/region_mean': 0.13737351819872856, 'reward_total_mean': 0.4673754572868347, 'reward_meter_mean': 0.5656245350837708, 'reward_meter_std': 0.3941541910171509, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.981610894203186, 'reward_repeat_penalty_std': 0.018419211730360985, 'reward_repeat_floor_mean': 0.9972416162490845, 'reward_repeat_floor_std': 0.0027628839015960693, 'reward_near_duplicate_penalty_mean': 0.981610894203186, 'reward_near_duplicate_penalty_std': 0.018419211730360985, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8062499761581421, 'reward_judge_quality_std': 0.22012577950954437, 'reward_total_composite_mean': 0.4673754572868347, 'reward_total_composite_std': 0.3525567948818207, 'epoch': 0.07}

 42%|████▏     | 1380/3300 [5:16:02<5:52:30, 11.02s/it]INFO 04-12 21:14:29 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1381/3300 [5:16:17<6:25:37, 12.06s/it]2026-04-12 21:14:43,453 | INFO | train_grpo_train | metrics_logged mode=train step=1381

                                                       
{'loss': -0.0656, 'grad_norm': 8.257170677185059, 'learning_rate': 5.8181818181818185e-06, 'num_tokens': 2445372.0, 'completions/mean_length': 133.25, 'completions/min_length': 70.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 79.14286041259766, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.5085207223892212, 'rewards/meter/std': 0.3166560232639313, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.026516498997807503, 'rewards/lexical_plausibility/mean': 0.9388819932937622, 'rewards/lexical_plausibility/std': 0.1728677898645401, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9647260904312134, 'rewards/repeat_penalty/std': 0.040419209748506546, 'rewards/repeat_floor/mean': 0.9958482384681702, 'rewards/repeat_floor/std': 0.004399863071739674, 'rewards/near_duplicate_penalty/mean': 0.9845747351646423, 'rewards/near_duplicate_penalty/std': 0.01241633202880621, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9795782566070557, 'rewards/distinct_2/std': 0.03128037601709366, 'rewards/total_composite/mean': 0.14327138662338257, 'rewards/total_composite/std': 0.09918083250522614, 'reward': 0.14327138662338257, 'reward_std': 0.09918083250522614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1695936769247055, 'sampling/sampling_logp_difference/max': 1.5722389221191406, 'sampling/importance_sampling_ratio/min': 0.20757991075515747, 'sampling/importance_sampling_ratio/mean': 1.0253328084945679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2447847500443459, 'clip_ratio/low_mean': 0.03497899230569601, 'clip_ratio/low_min': 0.03497899230569601, 'clip_ratio/high_mean': 0.10276469774544239, 'clip_ratio/high_max': 0.10276469774544239, 'clip_ratio/region_mean': 0.1377436900511384, 'reward_total_mean': 0.14327138662338257, 'reward_meter_mean': 0.5085207223892212, 'reward_meter_std': 0.3166560232639313, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.026516498997807503, 'reward_lexical_plausibility_mean': 0.9388819932937622, 'reward_lexical_plausibility_std': 0.1728677898645401, 'reward_repeat_penalty_mean': 0.9647260904312134, 'reward_repeat_penalty_std': 0.040419209748506546, 'reward_repeat_floor_mean': 0.9958482384681702, 'reward_repeat_floor_std': 0.004399863071739674, 'reward_near_duplicate_penalty_mean': 0.9845747351646423, 'reward_near_duplicate_penalty_std': 0.01241633202880621, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9795782566070557, 'reward_distinct_2_std': 0.03128037601709366, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.14327138662338257, 'reward_total_composite_std': 0.09918083250522614, 'epoch': 0.07}

 42%|████▏     | 1381/3300 [5:16:17<6:25:37, 12.06s/it]INFO 04-12 21:14:43 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1382/3300 [5:16:27<6:05:49, 11.44s/it]2026-04-12 21:14:53,466 | INFO | train_grpo_train | metrics_logged mode=train step=1382

                                                       
{'loss': 0.0384, 'grad_norm': 10.212300300598145, 'learning_rate': 5.815151515151516e-06, 'num_tokens': 2447828.0, 'completions/mean_length': 89.0, 'completions/min_length': 82.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.43377628922462463, 'rewards/meter/std': 0.3504347503185272, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.27740830183029175, 'rewards/repeat_penalty/mean': 0.9600228071212769, 'rewards/repeat_penalty/std': 0.031417373567819595, 'rewards/repeat_floor/mean': 0.995720624923706, 'rewards/repeat_floor/std': 0.0031103522051125765, 'rewards/near_duplicate_penalty/mean': 0.98921799659729, 'rewards/near_duplicate_penalty/std': 0.007008589804172516, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9704216718673706, 'rewards/distinct_2/std': 0.02851642481982708, 'rewards/total_composite/mean': 0.23056016862392426, 'rewards/total_composite/std': 0.25362980365753174, 'reward': 0.23056016862392426, 'reward_std': 0.25362980365753174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13449743390083313, 'sampling/sampling_logp_difference/max': 1.7509427070617676, 'sampling/importance_sampling_ratio/min': 0.17361019551753998, 'sampling/importance_sampling_ratio/mean': 1.0187311172485352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0623348504304886, 'clip_ratio/low_mean': 0.06945235095918179, 'clip_ratio/low_min': 0.06945235095918179, 'clip_ratio/high_mean': 0.04316510818898678, 'clip_ratio/high_max': 0.04316510818898678, 'clip_ratio/region_mean': 0.11261745914816856, 'reward_total_mean': 0.23056016862392426, 'reward_meter_mean': 0.43377628922462463, 'reward_meter_std': 0.3504347503185272, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9600228071212769, 'reward_repeat_penalty_std': 0.031417373567819595, 'reward_repeat_floor_mean': 0.995720624923706, 'reward_repeat_floor_std': 0.0031103522051125765, 'reward_near_duplicate_penalty_mean': 0.98921799659729, 'reward_near_duplicate_penalty_std': 0.007008589804172516, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9704216718673706, 'reward_distinct_2_std': 0.02851642481982708, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.27740830183029175, 'reward_total_composite_mean': 0.23056016862392426, 'reward_total_composite_std': 0.25362980365753174, 'epoch': 0.07}

 42%|████▏     | 1382/3300 [5:16:27<6:05:49, 11.44s/it]INFO 04-12 21:14:53 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1383/3300 [5:16:36<5:45:57, 10.83s/it]2026-04-12 21:15:02,853 | INFO | train_grpo_train | metrics_logged mode=train step=1383

                                                       
{'loss': 0.0386, 'grad_norm': 14.913884162902832, 'learning_rate': 5.812121212121212e-06, 'num_tokens': 2449484.0, 'completions/mean_length': 37.0, 'completions/min_length': 35.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.43368643522262573, 'rewards/meter/std': 0.3672539293766022, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.890625, 'rewards/lexical_plausibility/std': 0.20005889236927032, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.08345229178667068, 'rewards/repeat_penalty/mean': 0.8921650052070618, 'rewards/repeat_penalty/std': 0.11480287462472916, 'rewards/repeat_floor/mean': 0.9849240779876709, 'rewards/repeat_floor/std': 0.01506594568490982, 'rewards/near_duplicate_penalty/mean': 0.9197890758514404, 'rewards/near_duplicate_penalty/std': 0.0587652288377285, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9661751389503479, 'rewards/distinct_2/std': 0.07171893864870071, 'rewards/total_composite/mean': 0.2104009985923767, 'rewards/total_composite/std': 0.18752139806747437, 'reward': 0.2104009985923767, 'reward_std': 0.18752138316631317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13715417683124542, 'sampling/sampling_logp_difference/max': 2.148918390274048, 'sampling/importance_sampling_ratio/min': 0.11661022156476974, 'sampling/importance_sampling_ratio/mean': 1.0305767059326172, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0850043445825577, 'clip_ratio/low_mean': 0.04096700111404061, 'clip_ratio/low_min': 0.04096700111404061, 'clip_ratio/high_mean': 0.04990326915867627, 'clip_ratio/high_max': 0.04990326915867627, 'clip_ratio/region_mean': 0.09087027027271688, 'reward_total_mean': 0.2104009985923767, 'reward_meter_mean': 0.43368643522262573, 'reward_meter_std': 0.3672539293766022, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.890625, 'reward_lexical_plausibility_std': 0.20005889236927032, 'reward_repeat_penalty_mean': 0.8921650052070618, 'reward_repeat_penalty_std': 0.11480287462472916, 'reward_repeat_floor_mean': 0.9849240779876709, 'reward_repeat_floor_std': 0.01506594568490982, 'reward_near_duplicate_penalty_mean': 0.9197890758514404, 'reward_near_duplicate_penalty_std': 0.0587652288377285, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9661751389503479, 'reward_distinct_2_std': 0.07171893864870071, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.2104009985923767, 'reward_total_composite_std': 0.18752139806747437, 'epoch': 0.07}

 42%|████▏     | 1383/3300 [5:16:36<5:45:57, 10.83s/it]INFO 04-12 21:15:03 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1384/3300 [5:16:45<5:27:26, 10.25s/it]2026-04-12 21:15:11,770 | INFO | train_grpo_train | metrics_logged mode=train step=1384

                                                       
{'loss': 0.0516, 'grad_norm': 15.159344673156738, 'learning_rate': 5.8090909090909095e-06, 'num_tokens': 2450963.0, 'completions/mean_length': 33.875, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.21777233481407166, 'rewards/meter/std': 0.2769502103328705, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9761722087860107, 'rewards/repeat_penalty/std': 0.023667991161346436, 'rewards/repeat_floor/mean': 0.9964258670806885, 'rewards/repeat_floor/std': 0.003550207242369652, 'rewards/near_duplicate_penalty/mean': 0.9761722087860107, 'rewards/near_duplicate_penalty/std': 0.023667991161346436, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08764076232910156, 'rewards/total_composite/std': 0.12652982771396637, 'reward': 0.08764076232910156, 'reward_std': 0.12652981281280518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1754465103149414, 'sampling/sampling_logp_difference/max': 1.8747997283935547, 'sampling/importance_sampling_ratio/min': 0.1533856838941574, 'sampling/importance_sampling_ratio/mean': 1.0349808931350708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.621047243475914, 'clip_ratio/low_mean': 0.09394368529319763, 'clip_ratio/low_min': 0.09394368529319763, 'clip_ratio/high_mean': 0.034155597910284996, 'clip_ratio/high_max': 0.034155597910284996, 'clip_ratio/region_mean': 0.12809928320348263, 'reward_total_mean': 0.08764076232910156, 'reward_meter_mean': 0.21777233481407166, 'reward_meter_std': 0.2769502103328705, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9761722087860107, 'reward_repeat_penalty_std': 0.023667991161346436, 'reward_repeat_floor_mean': 0.9964258670806885, 'reward_repeat_floor_std': 0.003550207242369652, 'reward_near_duplicate_penalty_mean': 0.9761722087860107, 'reward_near_duplicate_penalty_std': 0.023667991161346436, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.08764076232910156, 'reward_total_composite_std': 0.12652982771396637, 'epoch': 0.07}

 42%|████▏     | 1384/3300 [5:16:45<5:27:26, 10.25s/it]INFO 04-12 21:15:12 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1385/3300 [5:16:55<5:21:32, 10.07s/it]2026-04-12 21:15:21,424 | INFO | train_grpo_train | metrics_logged mode=train step=1385

                                                       
{'loss': -0.0093, 'grad_norm': 16.93290901184082, 'learning_rate': 5.806060606060606e-06, 'num_tokens': 2452194.0, 'completions/mean_length': 20.875, 'completions/min_length': 19.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.6003322005271912, 'rewards/meter/std': 0.42994600534439087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.7139315605163574, 'rewards/repeat_penalty/std': 0.06942594796419144, 'rewards/repeat_floor/mean': 0.9786067008972168, 'rewards/repeat_floor/std': 0.013073219917714596, 'rewards/near_duplicate_penalty/mean': 0.9662534594535828, 'rewards/near_duplicate_penalty/std': 0.08387837558984756, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.3169105648994446, 'rewards/total_composite/std': 0.2148708552122116, 'reward': 0.3169105648994446, 'reward_std': 0.2148708552122116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15424107015132904, 'sampling/sampling_logp_difference/max': 1.8576135635375977, 'sampling/importance_sampling_ratio/min': 0.15604457259178162, 'sampling/importance_sampling_ratio/mean': 1.0220803022384644, 'sampling/importance_sampling_ratio/max': 1.6014626026153564, 'entropy': 1.1259381845593452, 'clip_ratio/low_mean': 0.05131579004228115, 'clip_ratio/low_min': 0.05131579004228115, 'clip_ratio/high_mean': 0.09645593632012606, 'clip_ratio/high_max': 0.09645593632012606, 'clip_ratio/region_mean': 0.1477717263624072, 'reward_total_mean': 0.3169105648994446, 'reward_meter_mean': 0.6003322005271912, 'reward_meter_std': 0.42994600534439087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.7139315605163574, 'reward_repeat_penalty_std': 0.06942594796419144, 'reward_repeat_floor_mean': 0.9786067008972168, 'reward_repeat_floor_std': 0.013073219917714596, 'reward_near_duplicate_penalty_mean': 0.9662534594535828, 'reward_near_duplicate_penalty_std': 0.08387837558984756, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.3169105648994446, 'reward_total_composite_std': 0.2148708552122116, 'epoch': 0.07}

 42%|████▏     | 1385/3300 [5:16:55<5:21:32, 10.07s/it]INFO 04-12 21:15:21 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1386/3300 [5:17:04<5:16:23,  9.92s/it]2026-04-12 21:15:30,980 | INFO | train_grpo_train | metrics_logged mode=train step=1386

                                                       
{'loss': 0.0852, 'grad_norm': 14.418100357055664, 'learning_rate': 5.803030303030304e-06, 'num_tokens': 2453766.0, 'completions/mean_length': 35.5, 'completions/min_length': 25.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6023249626159668, 'rewards/meter/std': 0.38541167974472046, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9407806396484375, 'rewards/repeat_penalty/std': 0.07284440845251083, 'rewards/repeat_floor/mean': 0.9925022125244141, 'rewards/repeat_floor/std': 0.008648661896586418, 'rewards/near_duplicate_penalty/mean': 0.9665827751159668, 'rewards/near_duplicate_penalty/std': 0.03403427451848984, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9723865985870361, 'rewards/distinct_2/std': 0.051303714513778687, 'rewards/total_composite/mean': 0.237629234790802, 'rewards/total_composite/std': 0.15501956641674042, 'reward': 0.237629234790802, 'reward_std': 0.15501955151557922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1717086285352707, 'sampling/sampling_logp_difference/max': 1.6316661834716797, 'sampling/importance_sampling_ratio/min': 0.1956033855676651, 'sampling/importance_sampling_ratio/mean': 1.0156081914901733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2342327982187271, 'clip_ratio/low_mean': 0.06301843374967575, 'clip_ratio/low_min': 0.06301843374967575, 'clip_ratio/high_mean': 0.12470238283276558, 'clip_ratio/high_max': 0.12470238283276558, 'clip_ratio/region_mean': 0.18772081658244133, 'reward_total_mean': 0.237629234790802, 'reward_meter_mean': 0.6023249626159668, 'reward_meter_std': 0.38541167974472046, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9407806396484375, 'reward_repeat_penalty_std': 0.07284440845251083, 'reward_repeat_floor_mean': 0.9925022125244141, 'reward_repeat_floor_std': 0.008648661896586418, 'reward_near_duplicate_penalty_mean': 0.9665827751159668, 'reward_near_duplicate_penalty_std': 0.03403427451848984, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9723865985870361, 'reward_distinct_2_std': 0.051303714513778687, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.237629234790802, 'reward_total_composite_std': 0.15501956641674042, 'epoch': 0.07}

 42%|████▏     | 1386/3300 [5:17:04<5:16:23,  9.92s/it]INFO 04-12 21:15:31 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1387/3300 [5:17:13<5:09:43,  9.71s/it]2026-04-12 21:15:40,207 | INFO | train_grpo_train | metrics_logged mode=train step=1387

                                                       
{'loss': 0.0322, 'grad_norm': 22.800981521606445, 'learning_rate': 5.8e-06, 'num_tokens': 2455203.0, 'completions/mean_length': 15.625, 'completions/min_length': 13.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.625, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.5446288585662842, 'rewards/meter/std': 0.4736005961894989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8100000023841858, 'rewards/judge_quality/std': 0.22258225083351135, 'rewards/repeat_penalty/mean': 0.7348276972770691, 'rewards/repeat_penalty/std': 0.031104182824492455, 'rewards/repeat_floor/mean': 0.9819655418395996, 'rewards/repeat_floor/std': 0.006220838986337185, 'rewards/near_duplicate_penalty/mean': 0.9797701835632324, 'rewards/near_duplicate_penalty/std': 0.04147225245833397, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4695475697517395, 'rewards/total_composite/std': 0.44437405467033386, 'reward': 0.4695475697517395, 'reward_std': 0.4443740248680115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15417081117630005, 'sampling/sampling_logp_difference/max': 1.1479153633117676, 'sampling/importance_sampling_ratio/min': 0.3172975480556488, 'sampling/importance_sampling_ratio/mean': 1.007819652557373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2510360851883888, 'clip_ratio/low_mean': 0.07854920905083418, 'clip_ratio/low_min': 0.07854920905083418, 'clip_ratio/high_mean': 0.09047619253396988, 'clip_ratio/high_max': 0.09047619253396988, 'clip_ratio/region_mean': 0.16902540158480406, 'reward_total_mean': 0.4695475697517395, 'reward_meter_mean': 0.5446288585662842, 'reward_meter_std': 0.4736005961894989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7348276972770691, 'reward_repeat_penalty_std': 0.031104182824492455, 'reward_repeat_floor_mean': 0.9819655418395996, 'reward_repeat_floor_std': 0.006220838986337185, 'reward_near_duplicate_penalty_mean': 0.9797701835632324, 'reward_near_duplicate_penalty_std': 0.04147225245833397, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8100000023841858, 'reward_judge_quality_std': 0.22258225083351135, 'reward_total_composite_mean': 0.4695475697517395, 'reward_total_composite_std': 0.44437405467033386, 'epoch': 0.07}

 42%|████▏     | 1387/3300 [5:17:14<5:09:43,  9.71s/it]INFO 04-12 21:15:40 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1388/3300 [5:17:22<4:59:57,  9.41s/it]2026-04-12 21:15:48,915 | INFO | train_grpo_train | metrics_logged mode=train step=1388

                                                       
{'loss': 0.0466, 'grad_norm': 10.965044021606445, 'learning_rate': 5.796969696969698e-06, 'num_tokens': 2456872.0, 'completions/mean_length': 38.625, 'completions/min_length': 36.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.23914378881454468, 'rewards/meter/std': 0.37572500109672546, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5924999713897705, 'rewards/judge_quality/std': 0.21359175443649292, 'rewards/repeat_penalty/mean': 0.9583553075790405, 'rewards/repeat_penalty/std': 0.04497262090444565, 'rewards/repeat_floor/mean': 0.9943298697471619, 'rewards/repeat_floor/std': 0.005261747166514397, 'rewards/near_duplicate_penalty/mean': 0.9689866304397583, 'rewards/near_duplicate_penalty/std': 0.020228583365678787, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9886877536773682, 'rewards/distinct_2/std': 0.031995780766010284, 'rewards/total_composite/mean': 0.16297760605812073, 'rewards/total_composite/std': 0.2972034215927124, 'reward': 0.16297760605812073, 'reward_std': 0.2972034215927124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11396858841180801, 'sampling/sampling_logp_difference/max': 1.8433475494384766, 'sampling/importance_sampling_ratio/min': 0.1582866758108139, 'sampling/importance_sampling_ratio/mean': 1.0197510719299316, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6444540061056614, 'clip_ratio/low_mean': 0.057240667287260294, 'clip_ratio/low_min': 0.057240667287260294, 'clip_ratio/high_mean': 0.020285087637603283, 'clip_ratio/high_max': 0.020285087637603283, 'clip_ratio/region_mean': 0.07752575492486358, 'reward_total_mean': 0.16297760605812073, 'reward_meter_mean': 0.23914378881454468, 'reward_meter_std': 0.37572500109672546, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.9583553075790405, 'reward_repeat_penalty_std': 0.04497262090444565, 'reward_repeat_floor_mean': 0.9943298697471619, 'reward_repeat_floor_std': 0.005261747166514397, 'reward_near_duplicate_penalty_mean': 0.9689866304397583, 'reward_near_duplicate_penalty_std': 0.020228583365678787, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9886877536773682, 'reward_distinct_2_std': 0.031995780766010284, 'reward_judge_quality_mean': 0.5924999713897705, 'reward_judge_quality_std': 0.21359175443649292, 'reward_total_composite_mean': 0.16297760605812073, 'reward_total_composite_std': 0.2972034215927124, 'epoch': 0.07}

 42%|████▏     | 1388/3300 [5:17:22<4:59:57,  9.41s/it]INFO 04-12 21:15:49 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1389/3300 [5:17:30<4:45:32,  8.97s/it]2026-04-12 21:15:56,837 | INFO | train_grpo_train | metrics_logged mode=train step=1389

                                                       
{'loss': -0.0334, 'grad_norm': 21.580530166625977, 'learning_rate': 5.793939393939394e-06, 'num_tokens': 2458306.0, 'completions/mean_length': 19.25, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.4622415602207184, 'rewards/meter/std': 0.3959689736366272, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9791666269302368, 'rewards/lexical_plausibility/std': 0.0589255727827549, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.734375, 'rewards/repeat_penalty/std': 0.04419417306780815, 'rewards/repeat_floor/mean': 0.9818750023841858, 'rewards/repeat_floor/std': 0.00883884821087122, 'rewards/near_duplicate_penalty/mean': 0.9791666269302368, 'rewards/near_duplicate_penalty/std': 0.0589255727827549, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.22148087620735168, 'rewards/total_composite/std': 0.17652671039104462, 'reward': 0.22148087620735168, 'reward_std': 0.17652671039104462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14687377214431763, 'sampling/sampling_logp_difference/max': 0.9140610694885254, 'sampling/importance_sampling_ratio/min': 0.40089288353919983, 'sampling/importance_sampling_ratio/mean': 1.0245178937911987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2652703747153282, 'clip_ratio/low_mean': 0.05480628600344062, 'clip_ratio/low_min': 0.05480628600344062, 'clip_ratio/high_mean': 0.10457963310182095, 'clip_ratio/high_max': 0.10457963310182095, 'clip_ratio/region_mean': 0.15938591910526156, 'reward_total_mean': 0.22148087620735168, 'reward_meter_mean': 0.4622415602207184, 'reward_meter_std': 0.3959689736366272, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9791666269302368, 'reward_lexical_plausibility_std': 0.0589255727827549, 'reward_repeat_penalty_mean': 0.734375, 'reward_repeat_penalty_std': 0.04419417306780815, 'reward_repeat_floor_mean': 0.9818750023841858, 'reward_repeat_floor_std': 0.00883884821087122, 'reward_near_duplicate_penalty_mean': 0.9791666269302368, 'reward_near_duplicate_penalty_std': 0.0589255727827549, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.22148087620735168, 'reward_total_composite_std': 0.17652671039104462, 'epoch': 0.07}

 42%|████▏     | 1389/3300 [5:17:30<4:45:32,  8.97s/it]INFO 04-12 21:15:57 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1390/3300 [5:17:45<5:44:41, 10.83s/it]2026-04-12 21:16:12,013 | INFO | train_grpo_train | metrics_logged mode=train step=1390

                                                       
{'loss': -0.0699, 'grad_norm': 2.766141176223755, 'learning_rate': 5.790909090909091e-06, 'num_tokens': 2460013.0, 'completions/mean_length': 162.375, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.833335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5453200936317444, 'rewards/meter/std': 0.3683542311191559, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8880355358123779, 'rewards/lexical_plausibility/std': 0.20765499770641327, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.34450533986091614, 'rewards/repeat_penalty/mean': 0.9565514326095581, 'rewards/repeat_penalty/std': 0.08521240949630737, 'rewards/repeat_floor/mean': 0.9962952136993408, 'rewards/repeat_floor/std': 0.005242377519607544, 'rewards/near_duplicate_penalty/mean': 0.9878014326095581, 'rewards/near_duplicate_penalty/std': 0.017887946218252182, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.29501891136169434, 'rewards/total_composite/std': 0.2677532136440277, 'reward': 0.29501891136169434, 'reward_std': 0.2677531838417053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13857948780059814, 'sampling/sampling_logp_difference/max': 1.079117774963379, 'sampling/importance_sampling_ratio/min': 0.36188220977783203, 'sampling/importance_sampling_ratio/mean': 1.023341417312622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6298702880740166, 'clip_ratio/low_mean': 0.039922481402754784, 'clip_ratio/low_min': 0.039922481402754784, 'clip_ratio/high_mean': 0.06304641393944621, 'clip_ratio/high_max': 0.06304641393944621, 'clip_ratio/region_mean': 0.102968895342201, 'reward_total_mean': 0.29501891136169434, 'reward_meter_mean': 0.5453200936317444, 'reward_meter_std': 0.3683542311191559, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8880355358123779, 'reward_lexical_plausibility_std': 0.20765499770641327, 'reward_repeat_penalty_mean': 0.9565514326095581, 'reward_repeat_penalty_std': 0.08521240949630737, 'reward_repeat_floor_mean': 0.9962952136993408, 'reward_repeat_floor_std': 0.005242377519607544, 'reward_near_duplicate_penalty_mean': 0.9878014326095581, 'reward_near_duplicate_penalty_std': 0.017887946218252182, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.34450533986091614, 'reward_total_composite_mean': 0.29501891136169434, 'reward_total_composite_std': 0.2677532136440277, 'epoch': 0.07}

 42%|████▏     | 1390/3300 [5:17:45<5:44:41, 10.83s/it]INFO 04-12 21:16:12 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1391/3300 [5:17:54<5:27:51, 10.30s/it]2026-04-12 21:16:21,102 | INFO | train_grpo_train | metrics_logged mode=train step=1391

                                                       
{'loss': -0.0077, 'grad_norm': 14.41737174987793, 'learning_rate': 5.787878787878788e-06, 'num_tokens': 2461631.0, 'completions/mean_length': 35.25, 'completions/min_length': 31.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7549124360084534, 'rewards/meter/std': 0.36974912881851196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.9899245500564575, 'rewards/repeat_penalty/std': 0.015155394561588764, 'rewards/repeat_floor/mean': 0.9984886646270752, 'rewards/repeat_floor/std': 0.0022733111400157213, 'rewards/near_duplicate_penalty/mean': 0.9899245500564575, 'rewards/near_duplicate_penalty/std': 0.015155394561588764, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.39334648847579956, 'rewards/total_composite/std': 0.27005499601364136, 'reward': 0.39334648847579956, 'reward_std': 0.27005499601364136, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14865972101688385, 'sampling/sampling_logp_difference/max': 1.7956857681274414, 'sampling/importance_sampling_ratio/min': 0.16601356863975525, 'sampling/importance_sampling_ratio/mean': 1.0352250337600708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.225643515586853, 'clip_ratio/low_mean': 0.05947580561041832, 'clip_ratio/low_min': 0.05947580561041832, 'clip_ratio/high_mean': 0.09717901097610593, 'clip_ratio/high_max': 0.09717901097610593, 'clip_ratio/region_mean': 0.15665481658652425, 'reward_total_mean': 0.39334648847579956, 'reward_meter_mean': 0.7549124360084534, 'reward_meter_std': 0.36974912881851196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9899245500564575, 'reward_repeat_penalty_std': 0.015155394561588764, 'reward_repeat_floor_mean': 0.9984886646270752, 'reward_repeat_floor_std': 0.0022733111400157213, 'reward_near_duplicate_penalty_mean': 0.9899245500564575, 'reward_near_duplicate_penalty_std': 0.015155394561588764, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.39334648847579956, 'reward_total_composite_std': 0.27005499601364136, 'epoch': 0.07}

 42%|████▏     | 1391/3300 [5:17:54<5:27:51, 10.30s/it]INFO 04-12 21:16:21 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1392/3300 [5:18:08<6:03:26, 11.43s/it]2026-04-12 21:16:35,152 | INFO | train_grpo_train | metrics_logged mode=train step=1392

                                                       
{'loss': -0.1069, 'grad_norm': 4.598461151123047, 'learning_rate': 5.784848484848486e-06, 'num_tokens': 2463695.0, 'completions/mean_length': 141.0, 'completions/min_length': 75.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 88.00000762939453, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.745202898979187, 'rewards/meter/std': 0.3547601103782654, 'rewards/count_adherence/mean': 0.5833333730697632, 'rewards/count_adherence/std': 0.17817416787147522, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.875, 'rewards/count_floor/std': 0.05345223844051361, 'rewards/lexical_plausibility/mean': 0.935979962348938, 'rewards/lexical_plausibility/std': 0.18107600510120392, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9186392426490784, 'rewards/repeat_penalty/std': 0.07243224233388901, 'rewards/repeat_floor/mean': 0.9918870329856873, 'rewards/repeat_floor/std': 0.004103685263544321, 'rewards/near_duplicate_penalty/mean': 0.972297191619873, 'rewards/near_duplicate_penalty/std': 0.02004527859389782, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9768607020378113, 'rewards/distinct_2/std': 0.021140994504094124, 'rewards/total_composite/mean': 0.2701265215873718, 'rewards/total_composite/std': 0.15338243544101715, 'reward': 0.2701265215873718, 'reward_std': 0.15338243544101715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15899266302585602, 'sampling/sampling_logp_difference/max': 2.2129554748535156, 'sampling/importance_sampling_ratio/min': 0.10937691479921341, 'sampling/importance_sampling_ratio/mean': 1.0513575077056885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3165737092494965, 'clip_ratio/low_mean': 0.02946512121707201, 'clip_ratio/low_min': 0.02946512121707201, 'clip_ratio/high_mean': 0.07904779957607388, 'clip_ratio/high_max': 0.07904779957607388, 'clip_ratio/region_mean': 0.1085129207931459, 'reward_total_mean': 0.2701265215873718, 'reward_meter_mean': 0.745202898979187, 'reward_meter_std': 0.3547601103782654, 'reward_count_adherence_mean': 0.5833333730697632, 'reward_count_adherence_std': 0.17817416787147522, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.875, 'reward_count_floor_std': 0.05345223844051361, 'reward_lexical_plausibility_mean': 0.935979962348938, 'reward_lexical_plausibility_std': 0.18107600510120392, 'reward_repeat_penalty_mean': 0.9186392426490784, 'reward_repeat_penalty_std': 0.07243224233388901, 'reward_repeat_floor_mean': 0.9918870329856873, 'reward_repeat_floor_std': 0.004103685263544321, 'reward_near_duplicate_penalty_mean': 0.972297191619873, 'reward_near_duplicate_penalty_std': 0.02004527859389782, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9768607020378113, 'reward_distinct_2_std': 0.021140994504094124, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.2701265215873718, 'reward_total_composite_std': 0.15338243544101715, 'epoch': 0.07}

 42%|████▏     | 1392/3300 [5:18:09<6:03:26, 11.43s/it]INFO 04-12 21:16:35 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1393/3300 [5:18:23<6:31:37, 12.32s/it]2026-04-12 21:16:49,554 | INFO | train_grpo_train | metrics_logged mode=train step=1393

                                                       
{'loss': -0.0428, 'grad_norm': 4.447368144989014, 'learning_rate': 5.781818181818181e-06, 'num_tokens': 2465196.0, 'completions/mean_length': 100.625, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.85714340209961, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.4454975724220276, 'rewards/meter/std': 0.3531752824783325, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9515846967697144, 'rewards/lexical_plausibility/std': 0.1369391828775406, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.2856946587562561, 'rewards/repeat_penalty/mean': 0.8021814823150635, 'rewards/repeat_penalty/std': 0.1420968621969223, 'rewards/repeat_floor/mean': 0.980101466178894, 'rewards/repeat_floor/std': 0.014370224438607693, 'rewards/near_duplicate_penalty/mean': 0.953133761882782, 'rewards/near_duplicate_penalty/std': 0.04253828153014183, 'rewards/opening_diversity/mean': 0.925000011920929, 'rewards/opening_diversity/std': 0.11338934302330017, 'rewards/distinct_2/mean': 0.9070154428482056, 'rewards/distinct_2/std': 0.08476367592811584, 'rewards/total_composite/mean': 0.2772789001464844, 'rewards/total_composite/std': 0.29282382130622864, 'reward': 0.2772789001464844, 'reward_std': 0.29282379150390625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14055874943733215, 'sampling/sampling_logp_difference/max': 1.5892038345336914, 'sampling/importance_sampling_ratio/min': 0.204088032245636, 'sampling/importance_sampling_ratio/mean': 1.0233770608901978, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9646820798516273, 'clip_ratio/low_mean': 0.03883565682917833, 'clip_ratio/low_min': 0.03883565682917833, 'clip_ratio/high_mean': 0.06705451477319002, 'clip_ratio/high_max': 0.06705451477319002, 'clip_ratio/region_mean': 0.10589017160236835, 'reward_total_mean': 0.2772789001464844, 'reward_meter_mean': 0.4454975724220276, 'reward_meter_std': 0.3531752824783325, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9515846967697144, 'reward_lexical_plausibility_std': 0.1369391828775406, 'reward_repeat_penalty_mean': 0.8021814823150635, 'reward_repeat_penalty_std': 0.1420968621969223, 'reward_repeat_floor_mean': 0.980101466178894, 'reward_repeat_floor_std': 0.014370224438607693, 'reward_near_duplicate_penalty_mean': 0.953133761882782, 'reward_near_duplicate_penalty_std': 0.04253828153014183, 'reward_opening_diversity_mean': 0.925000011920929, 'reward_opening_diversity_std': 0.11338934302330017, 'reward_distinct_2_mean': 0.9070154428482056, 'reward_distinct_2_std': 0.08476367592811584, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.2856946587562561, 'reward_total_composite_mean': 0.2772789001464844, 'reward_total_composite_std': 0.29282382130622864, 'epoch': 0.07}

 42%|████▏     | 1393/3300 [5:18:23<6:31:37, 12.32s/it]INFO 04-12 21:16:49 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1394/3300 [5:18:31<5:56:51, 11.23s/it]2026-04-12 21:16:58,245 | INFO | train_grpo_train | metrics_logged mode=train step=1394

                                                       
{'loss': 0.0387, 'grad_norm': 12.977887153625488, 'learning_rate': 5.7787878787878795e-06, 'num_tokens': 2466842.0, 'completions/mean_length': 35.75, 'completions/min_length': 33.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.31210851669311523, 'rewards/meter/std': 0.3221641480922699, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.7786849737167358, 'rewards/repeat_penalty/std': 0.17316633462905884, 'rewards/repeat_floor/mean': 0.980579137802124, 'rewards/repeat_floor/std': 0.014859617687761784, 'rewards/near_duplicate_penalty/mean': 0.9640665054321289, 'rewards/near_duplicate_penalty/std': 0.03026820346713066, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9482681155204773, 'rewards/distinct_2/std': 0.04287854582071304, 'rewards/total_composite/mean': 0.103264719247818, 'rewards/total_composite/std': 0.1123870387673378, 'reward': 0.103264719247818, 'reward_std': 0.1123870462179184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18829724192619324, 'sampling/sampling_logp_difference/max': 1.236257553100586, 'sampling/importance_sampling_ratio/min': 0.290469229221344, 'sampling/importance_sampling_ratio/mean': 0.9899457097053528, 'sampling/importance_sampling_ratio/max': 1.819714903831482, 'entropy': 1.378494381904602, 'clip_ratio/low_mean': 0.10288808681070805, 'clip_ratio/low_min': 0.10288808681070805, 'clip_ratio/high_mean': 0.04393939487636089, 'clip_ratio/high_max': 0.04393939487636089, 'clip_ratio/region_mean': 0.14682748168706894, 'reward_total_mean': 0.103264719247818, 'reward_meter_mean': 0.31210851669311523, 'reward_meter_std': 0.3221641480922699, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.7786849737167358, 'reward_repeat_penalty_std': 0.17316633462905884, 'reward_repeat_floor_mean': 0.980579137802124, 'reward_repeat_floor_std': 0.014859617687761784, 'reward_near_duplicate_penalty_mean': 0.9640665054321289, 'reward_near_duplicate_penalty_std': 0.03026820346713066, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9482681155204773, 'reward_distinct_2_std': 0.04287854582071304, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.103264719247818, 'reward_total_composite_std': 0.1123870387673378, 'epoch': 0.07}

 42%|████▏     | 1394/3300 [5:18:32<5:56:51, 11.23s/it]INFO 04-12 21:16:58 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1395/3300 [5:18:41<5:38:02, 10.65s/it]2026-04-12 21:17:07,530 | INFO | train_grpo_train | metrics_logged mode=train step=1395

                                                       
{'loss': 0.0203, 'grad_norm': 11.854384422302246, 'learning_rate': 5.775757575757577e-06, 'num_tokens': 2468615.0, 'completions/mean_length': 55.625, 'completions/min_length': 52.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7061129808425903, 'rewards/meter/std': 0.22467495501041412, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.934695303440094, 'rewards/repeat_penalty/std': 0.06023397669196129, 'rewards/repeat_floor/mean': 0.9917354583740234, 'rewards/repeat_floor/std': 0.007443113252520561, 'rewards/near_duplicate_penalty/mean': 0.9630705118179321, 'rewards/near_duplicate_penalty/std': 0.03057786077260971, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9697216749191284, 'rewards/distinct_2/std': 0.03538288548588753, 'rewards/total_composite/mean': 0.26837730407714844, 'rewards/total_composite/std': 0.08743525296449661, 'reward': 0.26837730407714844, 'reward_std': 0.08743525296449661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16467240452766418, 'sampling/sampling_logp_difference/max': 1.2084300518035889, 'sampling/importance_sampling_ratio/min': 0.29909101128578186, 'sampling/importance_sampling_ratio/mean': 1.03805410861969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4293698072433472, 'clip_ratio/low_mean': 0.06443057395517826, 'clip_ratio/low_min': 0.06443057395517826, 'clip_ratio/high_mean': 0.07486285362392664, 'clip_ratio/high_max': 0.07486285362392664, 'clip_ratio/region_mean': 0.1392934275791049, 'reward_total_mean': 0.26837730407714844, 'reward_meter_mean': 0.7061129808425903, 'reward_meter_std': 0.22467495501041412, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.934695303440094, 'reward_repeat_penalty_std': 0.06023397669196129, 'reward_repeat_floor_mean': 0.9917354583740234, 'reward_repeat_floor_std': 0.007443113252520561, 'reward_near_duplicate_penalty_mean': 0.9630705118179321, 'reward_near_duplicate_penalty_std': 0.03057786077260971, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9697216749191284, 'reward_distinct_2_std': 0.03538288548588753, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.26837730407714844, 'reward_total_composite_std': 0.08743525296449661, 'epoch': 0.07}

 42%|████▏     | 1395/3300 [5:18:41<5:38:02, 10.65s/it]INFO 04-12 21:17:07 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1396/3300 [5:18:50<5:26:45, 10.30s/it]2026-04-12 21:17:17,007 | INFO | train_grpo_train | metrics_logged mode=train step=1396

                                                       
{'loss': 0.0021, 'grad_norm': 12.468035697937012, 'learning_rate': 5.772727272727273e-06, 'num_tokens': 2470739.0, 'completions/mean_length': 88.5, 'completions/min_length': 79.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.5, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.8280707001686096, 'rewards/meter/std': 0.2732466459274292, 'rewards/count_adherence/mean': 0.5535714626312256, 'rewards/count_adherence/std': 0.05050763860344887, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8660714626312256, 'rewards/count_floor/std': 0.015152277424931526, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9188646078109741, 'rewards/repeat_penalty/std': 0.063174769282341, 'rewards/repeat_floor/mean': 0.9897975921630859, 'rewards/repeat_floor/std': 0.007479171734303236, 'rewards/near_duplicate_penalty/mean': 0.9554382562637329, 'rewards/near_duplicate_penalty/std': 0.027047229930758476, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9609095454216003, 'rewards/distinct_2/std': 0.043837759643793106, 'rewards/total_composite/mean': 0.26857444643974304, 'rewards/total_composite/std': 0.09987671673297882, 'reward': 0.26857444643974304, 'reward_std': 0.09987672418355942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18102264404296875, 'sampling/sampling_logp_difference/max': 2.163116931915283, 'sampling/importance_sampling_ratio/min': 0.11496621370315552, 'sampling/importance_sampling_ratio/mean': 1.0297086238861084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5483673512935638, 'clip_ratio/low_mean': 0.03540109843015671, 'clip_ratio/low_min': 0.03540109843015671, 'clip_ratio/high_mean': 0.126251928973943, 'clip_ratio/high_max': 0.126251928973943, 'clip_ratio/region_mean': 0.1616530274040997, 'reward_total_mean': 0.26857444643974304, 'reward_meter_mean': 0.8280707001686096, 'reward_meter_std': 0.2732466459274292, 'reward_count_adherence_mean': 0.5535714626312256, 'reward_count_adherence_std': 0.05050763860344887, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8660714626312256, 'reward_count_floor_std': 0.015152277424931526, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9188646078109741, 'reward_repeat_penalty_std': 0.063174769282341, 'reward_repeat_floor_mean': 0.9897975921630859, 'reward_repeat_floor_std': 0.007479171734303236, 'reward_near_duplicate_penalty_mean': 0.9554382562637329, 'reward_near_duplicate_penalty_std': 0.027047229930758476, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9609095454216003, 'reward_distinct_2_std': 0.043837759643793106, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.26857444643974304, 'reward_total_composite_std': 0.09987671673297882, 'epoch': 0.07}

 42%|████▏     | 1396/3300 [5:18:50<5:26:45, 10.30s/it]INFO 04-12 21:17:17 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1397/3300 [5:19:05<6:12:28, 11.74s/it]2026-04-12 21:17:32,140 | INFO | train_grpo_train | metrics_logged mode=train step=1397

                                                       
{'loss': 0.0148, 'grad_norm': 5.432043552398682, 'learning_rate': 5.76969696969697e-06, 'num_tokens': 2473010.0, 'completions/mean_length': 148.875, 'completions/min_length': 84.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 97.00000762939453, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.43902188539505005, 'rewards/meter/std': 0.35895174741744995, 'rewards/count_adherence/mean': 0.6964285969734192, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9089285731315613, 'rewards/count_floor/std': 0.015152297914028168, 'rewards/lexical_plausibility/mean': 0.9462568759918213, 'rewards/lexical_plausibility/std': 0.1520085334777832, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9578856229782104, 'rewards/repeat_penalty/std': 0.03811657428741455, 'rewards/repeat_floor/mean': 0.9951677918434143, 'rewards/repeat_floor/std': 0.004027257207781076, 'rewards/near_duplicate_penalty/mean': 0.9835860729217529, 'rewards/near_duplicate_penalty/std': 0.009844898246228695, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9736655950546265, 'rewards/distinct_2/std': 0.031053410843014717, 'rewards/total_composite/mean': 0.13947707414627075, 'rewards/total_composite/std': 0.12120145559310913, 'reward': 0.13947707414627075, 'reward_std': 0.12120144069194794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1529514491558075, 'sampling/sampling_logp_difference/max': 2.056504249572754, 'sampling/importance_sampling_ratio/min': 0.12790028750896454, 'sampling/importance_sampling_ratio/mean': 1.0159367322921753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.944608248770237, 'clip_ratio/low_mean': 0.06300278846174479, 'clip_ratio/low_min': 0.06300278846174479, 'clip_ratio/high_mean': 0.07215203810483217, 'clip_ratio/high_max': 0.07215203810483217, 'clip_ratio/region_mean': 0.13515482656657696, 'reward_total_mean': 0.13947707414627075, 'reward_meter_mean': 0.43902188539505005, 'reward_meter_std': 0.35895174741744995, 'reward_count_adherence_mean': 0.6964285969734192, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9089285731315613, 'reward_count_floor_std': 0.015152297914028168, 'reward_lexical_plausibility_mean': 0.9462568759918213, 'reward_lexical_plausibility_std': 0.1520085334777832, 'reward_repeat_penalty_mean': 0.9578856229782104, 'reward_repeat_penalty_std': 0.03811657428741455, 'reward_repeat_floor_mean': 0.9951677918434143, 'reward_repeat_floor_std': 0.004027257207781076, 'reward_near_duplicate_penalty_mean': 0.9835860729217529, 'reward_near_duplicate_penalty_std': 0.009844898246228695, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9736655950546265, 'reward_distinct_2_std': 0.031053410843014717, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.13947707414627075, 'reward_total_composite_std': 0.12120145559310913, 'epoch': 0.07}

 42%|████▏     | 1397/3300 [5:19:05<6:12:28, 11.74s/it]INFO 04-12 21:17:32 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1398/3300 [5:19:14<5:44:55, 10.88s/it]2026-04-12 21:17:40,993 | INFO | train_grpo_train | metrics_logged mode=train step=1398

                                                       
{'loss': -0.0021, 'grad_norm': 13.82620620727539, 'learning_rate': 5.766666666666667e-06, 'num_tokens': 2474659.0, 'completions/mean_length': 46.125, 'completions/min_length': 37.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6501312851905823, 'rewards/meter/std': 0.3600732386112213, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9114729166030884, 'rewards/repeat_penalty/std': 0.09514358639717102, 'rewards/repeat_floor/mean': 0.989188551902771, 'rewards/repeat_floor/std': 0.010733425617218018, 'rewards/near_duplicate_penalty/mean': 0.9580814838409424, 'rewards/near_duplicate_penalty/std': 0.029647180810570717, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9497377872467041, 'rewards/distinct_2/std': 0.07673981040716171, 'rewards/total_composite/mean': 0.28176149725914, 'rewards/total_composite/std': 0.15719543397426605, 'reward': 0.28176149725914, 'reward_std': 0.15719541907310486, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11914115399122238, 'sampling/sampling_logp_difference/max': 1.2555313110351562, 'sampling/importance_sampling_ratio/min': 0.2849244475364685, 'sampling/importance_sampling_ratio/mean': 1.006868600845337, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7540392652153969, 'clip_ratio/low_mean': 0.06602221913635731, 'clip_ratio/low_min': 0.06602221913635731, 'clip_ratio/high_mean': 0.09253221191465855, 'clip_ratio/high_max': 0.09253221191465855, 'clip_ratio/region_mean': 0.15855443105101585, 'reward_total_mean': 0.28176149725914, 'reward_meter_mean': 0.6501312851905823, 'reward_meter_std': 0.3600732386112213, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9114729166030884, 'reward_repeat_penalty_std': 0.09514358639717102, 'reward_repeat_floor_mean': 0.989188551902771, 'reward_repeat_floor_std': 0.010733425617218018, 'reward_near_duplicate_penalty_mean': 0.9580814838409424, 'reward_near_duplicate_penalty_std': 0.029647180810570717, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9497377872467041, 'reward_distinct_2_std': 0.07673981040716171, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.28176149725914, 'reward_total_composite_std': 0.15719543397426605, 'epoch': 0.07}

 42%|████▏     | 1398/3300 [5:19:14<5:44:55, 10.88s/it]INFO 04-12 21:17:41 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1399/3300 [5:19:24<5:36:14, 10.61s/it]2026-04-12 21:17:50,986 | INFO | train_grpo_train | metrics_logged mode=train step=1399

                                                       
{'loss': 0.0243, 'grad_norm': 16.468103408813477, 'learning_rate': 5.763636363636365e-06, 'num_tokens': 2476289.0, 'completions/mean_length': 41.75, 'completions/min_length': 38.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.10651238262653351, 'rewards/meter/std': 0.14343857765197754, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9478118419647217, 'rewards/lexical_plausibility/std': 0.12429578602313995, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9843199253082275, 'rewards/repeat_penalty/std': 0.020986884832382202, 'rewards/repeat_floor/mean': 0.9976480007171631, 'rewards/repeat_floor/std': 0.0031480323523283005, 'rewards/near_duplicate_penalty/mean': 0.9843199253082275, 'rewards/near_duplicate_penalty/std': 0.020986884832382202, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.03929522633552551, 'rewards/total_composite/std': 0.04974899813532829, 'reward': 0.03929522633552551, 'reward_std': 0.049748994410037994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1659487932920456, 'sampling/sampling_logp_difference/max': 1.5837574005126953, 'sampling/importance_sampling_ratio/min': 0.2052026242017746, 'sampling/importance_sampling_ratio/mean': 1.0307378768920898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.280714675784111, 'clip_ratio/low_mean': 0.06626155646517873, 'clip_ratio/low_min': 0.06626155646517873, 'clip_ratio/high_mean': 0.034325397573411465, 'clip_ratio/high_max': 0.034325397573411465, 'clip_ratio/region_mean': 0.10058695403859019, 'reward_total_mean': 0.03929522633552551, 'reward_meter_mean': 0.10651238262653351, 'reward_meter_std': 0.14343857765197754, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9478118419647217, 'reward_lexical_plausibility_std': 0.12429578602313995, 'reward_repeat_penalty_mean': 0.9843199253082275, 'reward_repeat_penalty_std': 0.020986884832382202, 'reward_repeat_floor_mean': 0.9976480007171631, 'reward_repeat_floor_std': 0.0031480323523283005, 'reward_near_duplicate_penalty_mean': 0.9843199253082275, 'reward_near_duplicate_penalty_std': 0.020986884832382202, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.03929522633552551, 'reward_total_composite_std': 0.04974899813532829, 'epoch': 0.07}

 42%|████▏     | 1399/3300 [5:19:24<5:36:14, 10.61s/it]INFO 04-12 21:17:51 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1400/3300 [5:19:38<6:09:38, 11.67s/it]2026-04-12 21:18:05,178 | INFO | train_grpo_train | metrics_logged mode=train step=1400

                                                       
{'loss': -0.0703, 'grad_norm': 4.668569564819336, 'learning_rate': 5.760606060606061e-06, 'num_tokens': 2478639.0, 'completions/mean_length': 154.75, 'completions/min_length': 89.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 103.71428680419922, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.29255253076553345, 'rewards/meter/std': 0.22243665158748627, 'rewards/count_adherence/mean': 0.6428571939468384, 'rewards/count_adherence/std': 0.07636035233736038, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8928571343421936, 'rewards/count_floor/std': 0.022908121347427368, 'rewards/lexical_plausibility/mean': 0.953578770160675, 'rewards/lexical_plausibility/std': 0.13129906356334686, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078858494758606, 'rewards/repeat_penalty/mean': 0.9560514688491821, 'rewards/repeat_penalty/std': 0.07047851383686066, 'rewards/repeat_floor/mean': 0.9954732656478882, 'rewards/repeat_floor/std': 0.0063373553566634655, 'rewards/near_duplicate_penalty/mean': 0.9873337149620056, 'rewards/near_duplicate_penalty/std': 0.010340740904211998, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9791463613510132, 'rewards/distinct_2/std': 0.03381364792585373, 'rewards/total_composite/mean': 0.09594075381755829, 'rewards/total_composite/std': 0.08763573318719864, 'reward': 0.09594075381755829, 'reward_std': 0.08763572573661804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11450890451669693, 'sampling/sampling_logp_difference/max': 1.700124740600586, 'sampling/importance_sampling_ratio/min': 0.18266074359416962, 'sampling/importance_sampling_ratio/mean': 1.0187299251556396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6725115776062012, 'clip_ratio/low_mean': 0.05052894912660122, 'clip_ratio/low_min': 0.05052894912660122, 'clip_ratio/high_mean': 0.04498100187629461, 'clip_ratio/high_max': 0.04498100187629461, 'clip_ratio/region_mean': 0.09550995100289583, 'reward_total_mean': 0.09594075381755829, 'reward_meter_mean': 0.29255253076553345, 'reward_meter_std': 0.22243665158748627, 'reward_count_adherence_mean': 0.6428571939468384, 'reward_count_adherence_std': 0.07636035233736038, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8928571343421936, 'reward_count_floor_std': 0.022908121347427368, 'reward_lexical_plausibility_mean': 0.953578770160675, 'reward_lexical_plausibility_std': 0.13129906356334686, 'reward_repeat_penalty_mean': 0.9560514688491821, 'reward_repeat_penalty_std': 0.07047851383686066, 'reward_repeat_floor_mean': 0.9954732656478882, 'reward_repeat_floor_std': 0.0063373553566634655, 'reward_near_duplicate_penalty_mean': 0.9873337149620056, 'reward_near_duplicate_penalty_std': 0.010340740904211998, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9791463613510132, 'reward_distinct_2_std': 0.03381364792585373, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078858494758606, 'reward_total_composite_mean': 0.09594075381755829, 'reward_total_composite_std': 0.08763573318719864, 'epoch': 0.07}

 42%|████▏     | 1400/3300 [5:19:39<6:09:38, 11.67s/it]INFO 04-12 21:18:05 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:09<00:51,  4.68s/it]

 23%|██▎       | 3/13 [00:20<01:14,  7.45s/it]

 31%|███       | 4/13 [00:32<01:21,  9.06s/it]

 38%|███▊      | 5/13 [00:44<01:21, 10.21s/it]

 46%|████▌     | 6/13 [00:56<01:15, 10.72s/it]

 54%|█████▍    | 7/13 [01:08<01:06, 11.14s/it]

 62%|██████▏   | 8/13 [01:20<00:56, 11.32s/it]

 69%|██████▉   | 9/13 [01:31<00:45, 11.35s/it]

 77%|███████▋  | 10/13 [01:43<00:34, 11.36s/it]

 85%|████████▍ | 11/13 [01:55<00:23, 11.59s/it]

 92%|█████████▏| 12/13 [02:06<00:11, 11.36s/it]

100%|██████████| 13/13 [02:17<00:00, 11.40s/it]2026-04-12 21:20:35,237 | INFO | train_grpo_train | metrics_logged mode=eval step=1400

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 149.9257, 'eval_samples_per_second': 0.694, 'eval_steps_per_second': 0.087, 'eval_num_tokens': 2478639.0, 'eval_completions/mean_length': 186.91346153846155, 'eval_completions/min_length': 37.15384615384615, 'eval_completions/max_length': 486.46153846153845, 'eval_completions/clipped_ratio': 0.2403846153846154, 'eval_completions/mean_terminated_length': 85.32271341177133, 'eval_completions/min_terminated_length': 37.15384615384615, 'eval_completions/max_terminated_length': 158.3846153846154, 'eval_rewards/meter/mean': 0.4560626917160474, 'eval_rewards/meter/std': 0.3552854909346654, 'eval_rewards/count_adherence/mean': 0.6974545808938833, 'eval_rewards/count_adherence/std': 0.20306277733582717, 'eval_rewards/arabic_clean/mean': 0.7403846153846154, 'eval_rewards/arabic_clean/std': 0.43055543074241054, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 1.0, 'eval_rewards/arabic_floor/std': 0.0, 'eval_rewards/count_floor/mean': 0.9092363806871268, 'eval_rewards/count_floor/std': 0.06091883649619726, 'eval_rewards/lexical_plausibility/mean': 0.9040030791209295, 'eval_rewards/lexical_plausibility/std': 0.1509883015488203, 'eval_rewards/judge_quality/mean': 0.3560576851551349, 'eval_rewards/judge_quality/std': 0.1958783446596219, 'eval_rewards/repeat_penalty/mean': 0.9114883129413311, 'eval_rewards/repeat_penalty/std': 0.11260175246458787, 'eval_rewards/repeat_floor/mean': 0.9902560802606436, 'eval_rewards/repeat_floor/std': 0.011564714404252859, 'eval_rewards/near_duplicate_penalty/mean': 0.9695396377490118, 'eval_rewards/near_duplicate_penalty/std': 0.031319281014685445, 'eval_rewards/opening_diversity/mean': 0.9863149432035593, 'eval_rewards/opening_diversity/std': 0.030451504656901725, 'eval_rewards/distinct_2/mean': 0.9516249253199651, 'eval_rewards/distinct_2/std': 0.07838768153809585, 'eval_rewards/total_composite/mean': 0.15268873881835204, 'eval_rewards/total_composite/std': 0.15279024638808691, 'eval_reward': 0.15268873881835204, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08628819624965008, 'eval_sampling/sampling_logp_difference/max': 0.9887015269352839, 'eval_sampling/importance_sampling_ratio/min': 0.3846321117419463, 'eval_sampling/importance_sampling_ratio/mean': 1.0270672394679143, 'eval_sampling/importance_sampling_ratio/max': 1.5628197284845204, 'eval_entropy': 1.0947969326606164, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.15268873881835204, 'eval_reward_meter_mean': 0.4560626917160474, 'eval_reward_meter_std': 0.3552854909346654, 'eval_reward_count_adherence_mean': 0.6974545808938833, 'eval_reward_count_adherence_std': 0.20306277733582717, 'eval_reward_arabic_clean_mean': 0.7403846153846154, 'eval_reward_arabic_clean_std': 0.43055543074241054, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 1.0, 'eval_reward_arabic_floor_std': 0.0, 'eval_reward_count_floor_mean': 0.9092363806871268, 'eval_reward_count_floor_std': 0.06091883649619726, 'eval_reward_lexical_plausibility_mean': 0.9040030791209295, 'eval_reward_lexical_plausibility_std': 0.1509883015488203, 'eval_reward_repeat_penalty_mean': 0.9114883129413311, 'eval_reward_repeat_penalty_std': 0.11260175246458787, 'eval_reward_repeat_floor_mean': 0.9902560802606436, 'eval_reward_repeat_floor_std': 0.011564714404252859, 'eval_reward_near_duplicate_penalty_mean': 0.9695396377490118, 'eval_reward_near_duplicate_penalty_std': 0.031319281014685445, 'eval_reward_opening_diversity_mean': 0.9863149432035593, 'eval_reward_opening_diversity_std': 0.030451504656901725, 'eval_reward_distinct_2_mean': 0.9516249253199651, 'eval_reward_distinct_2_std': 0.07838768153809585, 'eval_reward_judge_quality_mean': 0.3560576851551349, 'eval_reward_judge_quality_std': 0.1958783446596219, 'eval_reward_total_composite_mean': 0.15268873881835204, 'eval_reward_total_composite_std': 0.15279024638808691, 'epoch': 0.07}

 42%|████▏     | 1400/3300 [5:22:09<6:09:38, 11.67s/it]

100%|██████████| 13/13 [02:17<00:00, 11.40s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 42%|████▏     | 1401/3300 [5:22:28<31:11:47, 59.14s/it]2026-04-12 21:20:55,049 | INFO | train_grpo_train | metrics_logged mode=train step=1401

                                                        
{'loss': -0.0081, 'grad_norm': 3.8046207427978516, 'learning_rate': 5.7575757575757586e-06, 'num_tokens': 2479960.0, 'completions/mean_length': 81.125, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.571430206298828, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.28679224848747253, 'rewards/meter/std': 0.369350790977478, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9336142539978027, 'rewards/lexical_plausibility/std': 0.16397839784622192, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.26040831208229065, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.12404048442840576, 'rewards/total_composite/std': 0.16422851383686066, 'reward': 0.12404048442840576, 'reward_std': 0.16422851383686066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1284990906715393, 'sampling/sampling_logp_difference/max': 1.0953989028930664, 'sampling/importance_sampling_ratio/min': 0.33440616726875305, 'sampling/importance_sampling_ratio/mean': 1.031387209892273, 'sampling/importance_sampling_ratio/max': 1.5019066333770752, 'entropy': 1.1772644519805908, 'clip_ratio/low_mean': 0.06630117073655128, 'clip_ratio/low_min': 0.06630117073655128, 'clip_ratio/high_mean': 0.03345588315278292, 'clip_ratio/high_max': 0.03345588315278292, 'clip_ratio/region_mean': 0.0997570538893342, 'reward_total_mean': 0.12404048442840576, 'reward_meter_mean': 0.28679224848747253, 'reward_meter_std': 0.369350790977478, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9336142539978027, 'reward_lexical_plausibility_std': 0.16397839784622192, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.26040831208229065, 'reward_total_composite_mean': 0.12404048442840576, 'reward_total_composite_std': 0.16422851383686066, 'epoch': 0.07}

 42%|████▏     | 1401/3300 [5:22:28<31:11:47, 59.14s/it]INFO 04-12 21:20:55 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1402/3300 [5:22:42<23:59:21, 45.50s/it]2026-04-12 21:21:08,734 | INFO | train_grpo_train | metrics_logged mode=train step=1402

                                                        
{'loss': -0.0321, 'grad_norm': 3.853543758392334, 'learning_rate': 5.754545454545455e-06, 'num_tokens': 2481265.0, 'completions/mean_length': 82.125, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.71428680419922, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.6381834745407104, 'rewards/meter/std': 0.48102855682373047, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.13887302577495575, 'rewards/lexical_plausibility/mean': 0.9355192184448242, 'rewards/lexical_plausibility/std': 0.18237926065921783, 'rewards/judge_quality/mean': 0.4962499737739563, 'rewards/judge_quality/std': 0.38799622654914856, 'rewards/repeat_penalty/mean': 0.741767406463623, 'rewards/repeat_penalty/std': 0.19237299263477325, 'rewards/repeat_floor/mean': 0.977429986000061, 'rewards/repeat_floor/std': 0.022154675796628, 'rewards/near_duplicate_penalty/mean': 0.9600359797477722, 'rewards/near_duplicate_penalty/std': 0.0788830816745758, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9408284425735474, 'rewards/distinct_2/std': 0.12651419639587402, 'rewards/total_composite/mean': 0.3597506880760193, 'rewards/total_composite/std': 0.36664658784866333, 'reward': 0.3597506880760193, 'reward_std': 0.36664658784866333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14676345884799957, 'sampling/sampling_logp_difference/max': 0.9957623481750488, 'sampling/importance_sampling_ratio/min': 0.3694417178630829, 'sampling/importance_sampling_ratio/mean': 1.0442439317703247, 'sampling/importance_sampling_ratio/max': 1.798816442489624, 'entropy': 1.2173472940921783, 'clip_ratio/low_mean': 0.03612209111452103, 'clip_ratio/low_min': 0.03612209111452103, 'clip_ratio/high_mean': 0.04114906908944249, 'clip_ratio/high_max': 0.04114906908944249, 'clip_ratio/region_mean': 0.07727116020396352, 'reward_total_mean': 0.3597506880760193, 'reward_meter_mean': 0.6381834745407104, 'reward_meter_std': 0.48102855682373047, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.13887302577495575, 'reward_lexical_plausibility_mean': 0.9355192184448242, 'reward_lexical_plausibility_std': 0.18237926065921783, 'reward_repeat_penalty_mean': 0.741767406463623, 'reward_repeat_penalty_std': 0.19237299263477325, 'reward_repeat_floor_mean': 0.977429986000061, 'reward_repeat_floor_std': 0.022154675796628, 'reward_near_duplicate_penalty_mean': 0.9600359797477722, 'reward_near_duplicate_penalty_std': 0.0788830816745758, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9408284425735474, 'reward_distinct_2_std': 0.12651419639587402, 'reward_judge_quality_mean': 0.4962499737739563, 'reward_judge_quality_std': 0.38799622654914856, 'reward_total_composite_mean': 0.3597506880760193, 'reward_total_composite_std': 0.36664658784866333, 'epoch': 0.07}

 42%|████▏     | 1402/3300 [5:22:42<23:59:21, 45.50s/it]INFO 04-12 21:21:08 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1403/3300 [5:22:56<18:59:29, 36.04s/it]2026-04-12 21:21:22,674 | INFO | train_grpo_train | metrics_logged mode=train step=1403

                                                        
{'loss': -0.0245, 'grad_norm': 2.5551223754882812, 'learning_rate': 5.751515151515152e-06, 'num_tokens': 2482645.0, 'completions/mean_length': 157.5, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 39.333335876464844, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.48084402084350586, 'rewards/meter/std': 0.3628159761428833, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8688361048698425, 'rewards/lexical_plausibility/std': 0.24340996146202087, 'rewards/judge_quality/mean': 0.3787499964237213, 'rewards/judge_quality/std': 0.26840469241142273, 'rewards/repeat_penalty/mean': 0.9991028308868408, 'rewards/repeat_penalty/std': 0.0025374721735715866, 'rewards/repeat_floor/mean': 0.9998654127120972, 'rewards/repeat_floor/std': 0.000380628218408674, 'rewards/near_duplicate_penalty/mean': 0.9991028308868408, 'rewards/near_duplicate_penalty/std': 0.0025374721735715866, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14752501249313354, 'rewards/total_composite/std': 0.17671120166778564, 'reward': 0.14752501249313354, 'reward_std': 0.17671118676662445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14429503679275513, 'sampling/sampling_logp_difference/max': 1.2014061212539673, 'sampling/importance_sampling_ratio/min': 0.30077099800109863, 'sampling/importance_sampling_ratio/mean': 1.0254474878311157, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9729900285601616, 'clip_ratio/low_mean': 0.037819070275872946, 'clip_ratio/low_min': 0.037819070275872946, 'clip_ratio/high_mean': 0.03910875227302313, 'clip_ratio/high_max': 0.03910875227302313, 'clip_ratio/region_mean': 0.07692782254889607, 'reward_total_mean': 0.14752501249313354, 'reward_meter_mean': 0.48084402084350586, 'reward_meter_std': 0.3628159761428833, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8688361048698425, 'reward_lexical_plausibility_std': 0.24340996146202087, 'reward_repeat_penalty_mean': 0.9991028308868408, 'reward_repeat_penalty_std': 0.0025374721735715866, 'reward_repeat_floor_mean': 0.9998654127120972, 'reward_repeat_floor_std': 0.000380628218408674, 'reward_near_duplicate_penalty_mean': 0.9991028308868408, 'reward_near_duplicate_penalty_std': 0.0025374721735715866, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3787499964237213, 'reward_judge_quality_std': 0.26840469241142273, 'reward_total_composite_mean': 0.14752501249313354, 'reward_total_composite_std': 0.17671120166778564, 'epoch': 0.07}

 43%|████▎     | 1403/3300 [5:22:56<18:59:29, 36.04s/it]INFO 04-12 21:21:22 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1404/3300 [5:23:10<15:29:34, 29.42s/it]2026-04-12 21:21:36,660 | INFO | train_grpo_train | metrics_logged mode=train step=1404

                                                        
{'loss': -0.0233, 'grad_norm': 0.686887264251709, 'learning_rate': 5.748484848484849e-06, 'num_tokens': 2483937.0, 'completions/mean_length': 203.5, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 18.399999618530273, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.5003769397735596, 'rewards/meter/std': 0.3927370309829712, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7925765514373779, 'rewards/lexical_plausibility/std': 0.2922649383544922, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.29587340354919434, 'rewards/repeat_penalty/mean': 0.73777174949646, 'rewards/repeat_penalty/std': 0.03458673879504204, 'rewards/repeat_floor/mean': 0.9825543761253357, 'rewards/repeat_floor/std': 0.006917351391166449, 'rewards/near_duplicate_penalty/mean': 0.9836956262588501, 'rewards/near_duplicate_penalty/std': 0.04611567035317421, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24678762257099152, 'rewards/total_composite/std': 0.2963058650493622, 'reward': 0.24678762257099152, 'reward_std': 0.2963058650493622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1610114872455597, 'sampling/sampling_logp_difference/max': 1.126420021057129, 'sampling/importance_sampling_ratio/min': 0.32419177889823914, 'sampling/importance_sampling_ratio/mean': 1.0037342309951782, 'sampling/importance_sampling_ratio/max': 1.7762905359268188, 'entropy': 0.9222633987665176, 'clip_ratio/low_mean': 0.05625000223517418, 'clip_ratio/low_min': 0.05625000223517418, 'clip_ratio/high_mean': 0.07968750037252903, 'clip_ratio/high_max': 0.07968750037252903, 'clip_ratio/region_mean': 0.1359375026077032, 'reward_total_mean': 0.24678762257099152, 'reward_meter_mean': 0.5003769397735596, 'reward_meter_std': 0.3927370309829712, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.7925765514373779, 'reward_lexical_plausibility_std': 0.2922649383544922, 'reward_repeat_penalty_mean': 0.73777174949646, 'reward_repeat_penalty_std': 0.03458673879504204, 'reward_repeat_floor_mean': 0.9825543761253357, 'reward_repeat_floor_std': 0.006917351391166449, 'reward_near_duplicate_penalty_mean': 0.9836956262588501, 'reward_near_duplicate_penalty_std': 0.04611567035317421, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.29587340354919434, 'reward_total_composite_mean': 0.24678762257099152, 'reward_total_composite_std': 0.2963058650493622, 'epoch': 0.07}

 43%|████▎     | 1404/3300 [5:23:10<15:29:34, 29.42s/it]INFO 04-12 21:21:36 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1405/3300 [5:23:24<13:08:21, 24.96s/it]2026-04-12 21:21:51,215 | INFO | train_grpo_train | metrics_logged mode=train step=1405

                                                        
{'loss': -0.0445, 'grad_norm': 4.785051345825195, 'learning_rate': 5.745454545454546e-06, 'num_tokens': 2485551.0, 'completions/mean_length': 98.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.71428680419922, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5539117455482483, 'rewards/meter/std': 0.42445749044418335, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9246437549591064, 'rewards/lexical_plausibility/std': 0.17393621802330017, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.1597989797592163, 'rewards/repeat_penalty/mean': 0.8558858633041382, 'rewards/repeat_penalty/std': 0.15046103298664093, 'rewards/repeat_floor/mean': 0.9876335859298706, 'rewards/repeat_floor/std': 0.012660117819905281, 'rewards/near_duplicate_penalty/mean': 0.9776385426521301, 'rewards/near_duplicate_penalty/std': 0.029019419103860855, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9623643755912781, 'rewards/distinct_2/std': 0.04035434126853943, 'rewards/total_composite/mean': 0.18734672665596008, 'rewards/total_composite/std': 0.17887605726718903, 'reward': 0.18734672665596008, 'reward_std': 0.17887605726718903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18387193977832794, 'sampling/sampling_logp_difference/max': 1.6910743713378906, 'sampling/importance_sampling_ratio/min': 0.18432138860225677, 'sampling/importance_sampling_ratio/mean': 1.0273374319076538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2359858304262161, 'clip_ratio/low_mean': 0.06231571780517697, 'clip_ratio/low_min': 0.06231571780517697, 'clip_ratio/high_mean': 0.06808288767933846, 'clip_ratio/high_max': 0.06808288767933846, 'clip_ratio/region_mean': 0.13039860548451543, 'reward_total_mean': 0.18734672665596008, 'reward_meter_mean': 0.5539117455482483, 'reward_meter_std': 0.42445749044418335, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9246437549591064, 'reward_lexical_plausibility_std': 0.17393621802330017, 'reward_repeat_penalty_mean': 0.8558858633041382, 'reward_repeat_penalty_std': 0.15046103298664093, 'reward_repeat_floor_mean': 0.9876335859298706, 'reward_repeat_floor_std': 0.012660117819905281, 'reward_near_duplicate_penalty_mean': 0.9776385426521301, 'reward_near_duplicate_penalty_std': 0.029019419103860855, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9623643755912781, 'reward_distinct_2_std': 0.04035434126853943, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.1597989797592163, 'reward_total_composite_mean': 0.18734672665596008, 'reward_total_composite_std': 0.17887605726718903, 'epoch': 0.07}

 43%|████▎     | 1405/3300 [5:23:25<13:08:21, 24.96s/it]INFO 04-12 21:21:51 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1406/3300 [5:23:34<10:45:06, 20.44s/it]2026-04-12 21:22:01,173 | INFO | train_grpo_train | metrics_logged mode=train step=1406

                                                        
{'loss': 0.0912, 'grad_norm': 10.590453147888184, 'learning_rate': 5.742424242424242e-06, 'num_tokens': 2487653.0, 'completions/mean_length': 80.75, 'completions/min_length': 70.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.75, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.12493743002414703, 'rewards/meter/std': 0.10775718837976456, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9625000357627869, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4999999701976776, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9810932874679565, 'rewards/repeat_penalty/std': 0.016915034502744675, 'rewards/repeat_floor/mean': 0.9974523782730103, 'rewards/repeat_floor/std': 0.0017508569872006774, 'rewards/near_duplicate_penalty/mean': 0.9860525131225586, 'rewards/near_duplicate_penalty/std': 0.0044753397814929485, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9949392676353455, 'rewards/distinct_2/std': 0.014313912950456142, 'rewards/total_composite/mean': 0.05946630612015724, 'rewards/total_composite/std': 0.0506991371512413, 'reward': 0.05946630612015724, 'reward_std': 0.050699133425951004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1436416506767273, 'sampling/sampling_logp_difference/max': 2.2766549587249756, 'sampling/importance_sampling_ratio/min': 0.10262692719697952, 'sampling/importance_sampling_ratio/mean': 1.0181658267974854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7794519662857056, 'clip_ratio/low_mean': 0.07820118963718414, 'clip_ratio/low_min': 0.07820118963718414, 'clip_ratio/high_mean': 0.05505612678825855, 'clip_ratio/high_max': 0.05505612678825855, 'clip_ratio/region_mean': 0.1332573164254427, 'reward_total_mean': 0.05946630612015724, 'reward_meter_mean': 0.12493743002414703, 'reward_meter_std': 0.10775718837976456, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9625000357627869, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9810932874679565, 'reward_repeat_penalty_std': 0.016915034502744675, 'reward_repeat_floor_mean': 0.9974523782730103, 'reward_repeat_floor_std': 0.0017508569872006774, 'reward_near_duplicate_penalty_mean': 0.9860525131225586, 'reward_near_duplicate_penalty_std': 0.0044753397814929485, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9949392676353455, 'reward_distinct_2_std': 0.014313912950456142, 'reward_judge_quality_mean': 0.4999999701976776, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.05946630612015724, 'reward_total_composite_std': 0.0506991371512413, 'epoch': 0.07}

 43%|████▎     | 1406/3300 [5:23:35<10:45:06, 20.44s/it]INFO 04-12 21:22:01 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1407/3300 [5:23:49<9:50:28, 18.72s/it] 2026-04-12 21:22:15,783 | INFO | train_grpo_train | metrics_logged mode=train step=1407

                                                       
{'loss': -0.0098, 'grad_norm': 5.847170352935791, 'learning_rate': 5.73939393939394e-06, 'num_tokens': 2489642.0, 'completions/mean_length': 134.625, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.71428680419922, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.2960163950920105, 'rewards/meter/std': 0.3309500813484192, 'rewards/count_adherence/mean': 0.7708333730697632, 'rewards/count_adherence/std': 0.0862581729888916, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9312499761581421, 'rewards/count_floor/std': 0.025877466425299644, 'rewards/lexical_plausibility/mean': 0.9628743529319763, 'rewards/lexical_plausibility/std': 0.10500723123550415, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.9943349361419678, 'rewards/repeat_penalty/std': 0.0050043282099068165, 'rewards/repeat_floor/mean': 0.999150276184082, 'rewards/repeat_floor/std': 0.0007506543770432472, 'rewards/near_duplicate_penalty/mean': 0.9943349361419678, 'rewards/near_duplicate_penalty/std': 0.0050043282099068165, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.08322226256132126, 'rewards/total_composite/std': 0.1092400848865509, 'reward': 0.08322226256132126, 'reward_std': 0.1092400774359703, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14656026661396027, 'sampling/sampling_logp_difference/max': 1.395003318786621, 'sampling/importance_sampling_ratio/min': 0.24783220887184143, 'sampling/importance_sampling_ratio/mean': 1.022430658340454, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0365147441625595, 'clip_ratio/low_mean': 0.04992985911667347, 'clip_ratio/low_min': 0.04992985911667347, 'clip_ratio/high_mean': 0.08346572518348694, 'clip_ratio/high_max': 0.08346572518348694, 'clip_ratio/region_mean': 0.1333955843001604, 'reward_total_mean': 0.08322226256132126, 'reward_meter_mean': 0.2960163950920105, 'reward_meter_std': 0.3309500813484192, 'reward_count_adherence_mean': 0.7708333730697632, 'reward_count_adherence_std': 0.0862581729888916, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9312499761581421, 'reward_count_floor_std': 0.025877466425299644, 'reward_lexical_plausibility_mean': 0.9628743529319763, 'reward_lexical_plausibility_std': 0.10500723123550415, 'reward_repeat_penalty_mean': 0.9943349361419678, 'reward_repeat_penalty_std': 0.0050043282099068165, 'reward_repeat_floor_mean': 0.999150276184082, 'reward_repeat_floor_std': 0.0007506543770432472, 'reward_near_duplicate_penalty_mean': 0.9943349361419678, 'reward_near_duplicate_penalty_std': 0.0050043282099068165, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.08322226256132126, 'reward_total_composite_std': 0.1092400848865509, 'epoch': 0.07}

 43%|████▎     | 1407/3300 [5:23:49<9:50:28, 18.72s/it]INFO 04-12 21:22:16 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1408/3300 [5:24:00<8:37:08, 16.40s/it]2026-04-12 21:22:26,775 | INFO | train_grpo_train | metrics_logged mode=train step=1408

                                                       
{'loss': -0.0124, 'grad_norm': 12.23988151550293, 'learning_rate': 5.736363636363637e-06, 'num_tokens': 2491438.0, 'completions/mean_length': 51.5, 'completions/min_length': 45.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7885805368423462, 'rewards/meter/std': 0.2926645278930664, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9124999642372131, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4662500023841858, 'rewards/judge_quality/std': 0.17435084283351898, 'rewards/repeat_penalty/mean': 0.7986570596694946, 'rewards/repeat_penalty/std': 0.12854698300361633, 'rewards/repeat_floor/mean': 0.977705180644989, 'rewards/repeat_floor/std': 0.012192374095320702, 'rewards/near_duplicate_penalty/mean': 0.9311705827713013, 'rewards/near_duplicate_penalty/std': 0.022871362045407295, 'rewards/opening_diversity/mean': 0.9437500238418579, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.904495120048523, 'rewards/distinct_2/std': 0.06510571390390396, 'rewards/total_composite/mean': 0.29079192876815796, 'rewards/total_composite/std': 0.08249565213918686, 'reward': 0.29079192876815796, 'reward_std': 0.08249565213918686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17599311470985413, 'sampling/sampling_logp_difference/max': 2.4814858436584473, 'sampling/importance_sampling_ratio/min': 0.0836188867688179, 'sampling/importance_sampling_ratio/mean': 1.0232452154159546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4229249060153961, 'clip_ratio/low_mean': 0.02803983259946108, 'clip_ratio/low_min': 0.02803983259946108, 'clip_ratio/high_mean': 0.11394911538809538, 'clip_ratio/high_max': 0.11394911538809538, 'clip_ratio/region_mean': 0.14198894798755646, 'reward_total_mean': 0.29079192876815796, 'reward_meter_mean': 0.7885805368423462, 'reward_meter_std': 0.2926645278930664, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9124999642372131, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7986570596694946, 'reward_repeat_penalty_std': 0.12854698300361633, 'reward_repeat_floor_mean': 0.977705180644989, 'reward_repeat_floor_std': 0.012192374095320702, 'reward_near_duplicate_penalty_mean': 0.9311705827713013, 'reward_near_duplicate_penalty_std': 0.022871362045407295, 'reward_opening_diversity_mean': 0.9437500238418579, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.904495120048523, 'reward_distinct_2_std': 0.06510571390390396, 'reward_judge_quality_mean': 0.4662500023841858, 'reward_judge_quality_std': 0.17435084283351898, 'reward_total_composite_mean': 0.29079192876815796, 'reward_total_composite_std': 0.08249565213918686, 'epoch': 0.07}

 43%|████▎     | 1408/3300 [5:24:00<8:37:08, 16.40s/it]INFO 04-12 21:22:27 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1409/3300 [5:24:14<8:13:57, 15.67s/it]2026-04-12 21:22:40,781 | INFO | train_grpo_train | metrics_logged mode=train step=1409

                                                       
{'loss': -0.0446, 'grad_norm': 2.180145263671875, 'learning_rate': 5.733333333333334e-06, 'num_tokens': 2492975.0, 'completions/mean_length': 229.125, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 59.400001525878906, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.17504672706127167, 'rewards/meter/std': 0.3301781415939331, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 0.8438267111778259, 'rewards/lexical_plausibility/std': 0.22010590136051178, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.8769052028656006, 'rewards/repeat_penalty/std': 0.14940084517002106, 'rewards/repeat_floor/mean': 0.9878734946250916, 'rewards/repeat_floor/std': 0.014392049983143806, 'rewards/near_duplicate_penalty/mean': 0.9670385122299194, 'rewards/near_duplicate_penalty/std': 0.034258898347616196, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.96186363697052, 'rewards/distinct_2/std': 0.05792667344212532, 'rewards/total_composite/mean': 0.05738497152924538, 'rewards/total_composite/std': 0.11863113194704056, 'reward': 0.05738497152924538, 'reward_std': 0.11863112449645996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1607528030872345, 'sampling/sampling_logp_difference/max': 1.4583063125610352, 'sampling/importance_sampling_ratio/min': 0.2326299548149109, 'sampling/importance_sampling_ratio/mean': 1.0338927507400513, 'sampling/importance_sampling_ratio/max': 1.8721318244934082, 'entropy': 1.2141825556755066, 'clip_ratio/low_mean': 0.04575598519295454, 'clip_ratio/low_min': 0.04575598519295454, 'clip_ratio/high_mean': 0.051229506731033325, 'clip_ratio/high_max': 0.051229506731033325, 'clip_ratio/region_mean': 0.09698549192398787, 'reward_total_mean': 0.05738497152924538, 'reward_meter_mean': 0.17504672706127167, 'reward_meter_std': 0.3301781415939331, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 0.8438267111778259, 'reward_lexical_plausibility_std': 0.22010590136051178, 'reward_repeat_penalty_mean': 0.8769052028656006, 'reward_repeat_penalty_std': 0.14940084517002106, 'reward_repeat_floor_mean': 0.9878734946250916, 'reward_repeat_floor_std': 0.014392049983143806, 'reward_near_duplicate_penalty_mean': 0.9670385122299194, 'reward_near_duplicate_penalty_std': 0.034258898347616196, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.96186363697052, 'reward_distinct_2_std': 0.05792667344212532, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.05738497152924538, 'reward_total_composite_std': 0.11863113194704056, 'epoch': 0.07}

 43%|████▎     | 1409/3300 [5:24:14<8:13:57, 15.67s/it]INFO 04-12 21:22:41 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1410/3300 [5:24:29<8:07:31, 15.48s/it]2026-04-12 21:22:55,782 | INFO | train_grpo_train | metrics_logged mode=train step=1410

                                                       
{'loss': -0.1162, 'grad_norm': 4.727728366851807, 'learning_rate': 5.7303030303030305e-06, 'num_tokens': 2495150.0, 'completions/mean_length': 136.875, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.28572082519531, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.7368236184120178, 'rewards/meter/std': 0.29515108466148376, 'rewards/count_adherence/mean': 0.6750000715255737, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9024999737739563, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 0.9466106295585632, 'rewards/lexical_plausibility/std': 0.13613402843475342, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9404205083847046, 'rewards/repeat_penalty/std': 0.06177308410406113, 'rewards/repeat_floor/mean': 0.9935632348060608, 'rewards/repeat_floor/std': 0.006866373121738434, 'rewards/near_duplicate_penalty/mean': 0.9842514991760254, 'rewards/near_duplicate_penalty/std': 0.019076792523264885, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9547277092933655, 'rewards/distinct_2/std': 0.04511205852031708, 'rewards/total_composite/mean': 0.21358276903629303, 'rewards/total_composite/std': 0.11908823996782303, 'reward': 0.21358276903629303, 'reward_std': 0.11908823251724243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15063251554965973, 'sampling/sampling_logp_difference/max': 2.350043296813965, 'sampling/importance_sampling_ratio/min': 0.09536503255367279, 'sampling/importance_sampling_ratio/mean': 1.018662452697754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9848711788654327, 'clip_ratio/low_mean': 0.033502329140901566, 'clip_ratio/low_min': 0.033502329140901566, 'clip_ratio/high_mean': 0.10390935186296701, 'clip_ratio/high_max': 0.10390935186296701, 'clip_ratio/region_mean': 0.13741168100386858, 'reward_total_mean': 0.21358276903629303, 'reward_meter_mean': 0.7368236184120178, 'reward_meter_std': 0.29515108466148376, 'reward_count_adherence_mean': 0.6750000715255737, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9024999737739563, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 0.9466106295585632, 'reward_lexical_plausibility_std': 0.13613402843475342, 'reward_repeat_penalty_mean': 0.9404205083847046, 'reward_repeat_penalty_std': 0.06177308410406113, 'reward_repeat_floor_mean': 0.9935632348060608, 'reward_repeat_floor_std': 0.006866373121738434, 'reward_near_duplicate_penalty_mean': 0.9842514991760254, 'reward_near_duplicate_penalty_std': 0.019076792523264885, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9547277092933655, 'reward_distinct_2_std': 0.04511205852031708, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.21358276903629303, 'reward_total_composite_std': 0.11908823996782303, 'epoch': 0.07}

 43%|████▎     | 1410/3300 [5:24:29<8:07:31, 15.48s/it]INFO 04-12 21:22:56 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1411/3300 [5:24:46<8:19:47, 15.87s/it]2026-04-12 21:23:12,580 | INFO | train_grpo_train | metrics_logged mode=train step=1411

                                                       
{'loss': -0.0877, 'grad_norm': 2.2400755882263184, 'learning_rate': 5.727272727272728e-06, 'num_tokens': 2496904.0, 'completions/mean_length': 232.25, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 64.4000015258789, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.49991118907928467, 'rewards/meter/std': 0.357992023229599, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.1602174937725067, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9343750476837158, 'rewards/count_floor/std': 0.04806523770093918, 'rewards/lexical_plausibility/mean': 0.8410376310348511, 'rewards/lexical_plausibility/std': 0.21619945764541626, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.16903086006641388, 'rewards/repeat_penalty/mean': 0.9938429594039917, 'rewards/repeat_penalty/std': 0.006350129842758179, 'rewards/repeat_floor/mean': 0.9990764260292053, 'rewards/repeat_floor/std': 0.00095252605387941, 'rewards/near_duplicate_penalty/mean': 0.9938429594039917, 'rewards/near_duplicate_penalty/std': 0.006350129842758179, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.14490261673927307, 'rewards/total_composite/std': 0.1520993560552597, 'reward': 0.14490261673927307, 'reward_std': 0.1520993709564209, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1719357967376709, 'sampling/sampling_logp_difference/max': 1.1508255004882812, 'sampling/importance_sampling_ratio/min': 0.3163754940032959, 'sampling/importance_sampling_ratio/mean': 1.0270270109176636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1170873790979385, 'clip_ratio/low_mean': 0.009765625, 'clip_ratio/low_min': 0.009765625, 'clip_ratio/high_mean': 0.08834995981305838, 'clip_ratio/high_max': 0.08834995981305838, 'clip_ratio/region_mean': 0.09811558481305838, 'reward_total_mean': 0.14490261673927307, 'reward_meter_mean': 0.49991118907928467, 'reward_meter_std': 0.357992023229599, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.1602174937725067, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9343750476837158, 'reward_count_floor_std': 0.04806523770093918, 'reward_lexical_plausibility_mean': 0.8410376310348511, 'reward_lexical_plausibility_std': 0.21619945764541626, 'reward_repeat_penalty_mean': 0.9938429594039917, 'reward_repeat_penalty_std': 0.006350129842758179, 'reward_repeat_floor_mean': 0.9990764260292053, 'reward_repeat_floor_std': 0.00095252605387941, 'reward_near_duplicate_penalty_mean': 0.9938429594039917, 'reward_near_duplicate_penalty_std': 0.006350129842758179, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.16903086006641388, 'reward_total_composite_mean': 0.14490261673927307, 'reward_total_composite_std': 0.1520993560552597, 'epoch': 0.07}

 43%|████▎     | 1411/3300 [5:24:46<8:19:47, 15.87s/it]INFO 04-12 21:23:12 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1412/3300 [5:24:55<7:17:20, 13.90s/it]2026-04-12 21:23:21,931 | INFO | train_grpo_train | metrics_logged mode=train step=1412

                                                       
{'loss': 0.0286, 'grad_norm': 12.705042839050293, 'learning_rate': 5.724242424242424e-06, 'num_tokens': 2498389.0, 'completions/mean_length': 33.625, 'completions/min_length': 30.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7463754415512085, 'rewards/meter/std': 0.30197617411613464, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4624999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9858032464981079, 'rewards/repeat_penalty/std': 0.014091964811086655, 'rewards/repeat_floor/mean': 0.9978704452514648, 'rewards/repeat_floor/std': 0.0021138025913387537, 'rewards/near_duplicate_penalty/mean': 0.9858032464981079, 'rewards/near_duplicate_penalty/std': 0.014091964811086655, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3601909279823303, 'rewards/total_composite/std': 0.18580298125743866, 'reward': 0.3601909279823303, 'reward_std': 0.18580298125743866, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15604948997497559, 'sampling/sampling_logp_difference/max': 1.7259682416915894, 'sampling/importance_sampling_ratio/min': 0.17800061404705048, 'sampling/importance_sampling_ratio/mean': 1.0356197357177734, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4667446464300156, 'clip_ratio/low_mean': 0.06213459558784962, 'clip_ratio/low_min': 0.06213459558784962, 'clip_ratio/high_mean': 0.07902002241462469, 'clip_ratio/high_max': 0.07902002241462469, 'clip_ratio/region_mean': 0.1411546180024743, 'reward_total_mean': 0.3601909279823303, 'reward_meter_mean': 0.7463754415512085, 'reward_meter_std': 0.30197617411613464, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9858032464981079, 'reward_repeat_penalty_std': 0.014091964811086655, 'reward_repeat_floor_mean': 0.9978704452514648, 'reward_repeat_floor_std': 0.0021138025913387537, 'reward_near_duplicate_penalty_mean': 0.9858032464981079, 'reward_near_duplicate_penalty_std': 0.014091964811086655, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4624999761581421, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.3601909279823303, 'reward_total_composite_std': 0.18580298125743866, 'epoch': 0.07}

 43%|████▎     | 1412/3300 [5:24:55<7:17:20, 13.90s/it]INFO 04-12 21:23:22 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1413/3300 [5:25:04<6:31:22, 12.44s/it]2026-04-12 21:23:30,985 | INFO | train_grpo_train | metrics_logged mode=train step=1413

                                                       
{'loss': 0.0188, 'grad_norm': 16.99075698852539, 'learning_rate': 5.721212121212122e-06, 'num_tokens': 2499900.0, 'completions/mean_length': 19.875, 'completions/min_length': 16.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6092807054519653, 'rewards/meter/std': 0.48635751008987427, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.3123185336589813, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4129544496536255, 'rewards/total_composite/std': 0.42000505328178406, 'reward': 0.4129544496536255, 'reward_std': 0.42000505328178406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1676495373249054, 'sampling/sampling_logp_difference/max': 1.0294160842895508, 'sampling/importance_sampling_ratio/min': 0.3572154939174652, 'sampling/importance_sampling_ratio/mean': 1.055390477180481, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.585373617708683, 'clip_ratio/low_mean': 0.07655888330191374, 'clip_ratio/low_min': 0.07655888330191374, 'clip_ratio/high_mean': 0.08583960123360157, 'clip_ratio/high_max': 0.08583960123360157, 'clip_ratio/region_mean': 0.1623984845355153, 'reward_total_mean': 0.4129544496536255, 'reward_meter_mean': 0.6092807054519653, 'reward_meter_std': 0.48635751008987427, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.3123185336589813, 'reward_total_composite_mean': 0.4129544496536255, 'reward_total_composite_std': 0.42000505328178406, 'epoch': 0.07}

 43%|████▎     | 1413/3300 [5:25:04<6:31:22, 12.44s/it]INFO 04-12 21:23:31 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1414/3300 [5:25:13<5:58:34, 11.41s/it]2026-04-12 21:23:39,896 | INFO | train_grpo_train | metrics_logged mode=train step=1414

                                                       
{'loss': -0.0018, 'grad_norm': 11.382184982299805, 'learning_rate': 5.718181818181819e-06, 'num_tokens': 2501765.0, 'completions/mean_length': 60.125, 'completions/min_length': 39.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.761445164680481, 'rewards/meter/std': 0.3242763578891754, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.9319169521331787, 'rewards/repeat_penalty/std': 0.06158699095249176, 'rewards/repeat_floor/mean': 0.9922173619270325, 'rewards/repeat_floor/std': 0.00667576864361763, 'rewards/near_duplicate_penalty/mean': 0.9745862483978271, 'rewards/near_duplicate_penalty/std': 0.023200392723083496, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9558823108673096, 'rewards/distinct_2/std': 0.052539508789777756, 'rewards/total_composite/mean': 0.3547138571739197, 'rewards/total_composite/std': 0.26118040084838867, 'reward': 0.3547138571739197, 'reward_std': 0.26118040084838867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15082022547721863, 'sampling/sampling_logp_difference/max': 1.1497068405151367, 'sampling/importance_sampling_ratio/min': 0.3167296051979065, 'sampling/importance_sampling_ratio/mean': 1.0293949842453003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.437969148159027, 'clip_ratio/low_mean': 0.06293540028855205, 'clip_ratio/low_min': 0.06293540028855205, 'clip_ratio/high_mean': 0.04487109184265137, 'clip_ratio/high_max': 0.04487109184265137, 'clip_ratio/region_mean': 0.10780649213120341, 'reward_total_mean': 0.3547138571739197, 'reward_meter_mean': 0.761445164680481, 'reward_meter_std': 0.3242763578891754, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9319169521331787, 'reward_repeat_penalty_std': 0.06158699095249176, 'reward_repeat_floor_mean': 0.9922173619270325, 'reward_repeat_floor_std': 0.00667576864361763, 'reward_near_duplicate_penalty_mean': 0.9745862483978271, 'reward_near_duplicate_penalty_std': 0.023200392723083496, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9558823108673096, 'reward_distinct_2_std': 0.052539508789777756, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.3547138571739197, 'reward_total_composite_std': 0.26118040084838867, 'epoch': 0.07}

 43%|████▎     | 1414/3300 [5:25:13<5:58:34, 11.41s/it]INFO 04-12 21:23:40 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1415/3300 [5:25:23<5:44:57, 10.98s/it]2026-04-12 21:23:49,883 | INFO | train_grpo_train | metrics_logged mode=train step=1415

                                                       
{'loss': 0.0526, 'grad_norm': 10.996042251586914, 'learning_rate': 5.715151515151516e-06, 'num_tokens': 2503788.0, 'completions/mean_length': 75.875, 'completions/min_length': 67.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.865246593952179, 'rewards/meter/std': 0.24865683913230896, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9665993452072144, 'rewards/repeat_penalty/std': 0.03361032158136368, 'rewards/repeat_floor/mean': 0.9959484338760376, 'rewards/repeat_floor/std': 0.003226237604394555, 'rewards/near_duplicate_penalty/mean': 0.9829816818237305, 'rewards/near_duplicate_penalty/std': 0.008810799568891525, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9833458662033081, 'rewards/distinct_2/std': 0.03337503597140312, 'rewards/total_composite/mean': 0.3502022922039032, 'rewards/total_composite/std': 0.09896644204854965, 'reward': 0.3502022922039032, 'reward_std': 0.09896643459796906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13796325027942657, 'sampling/sampling_logp_difference/max': 1.9771487712860107, 'sampling/importance_sampling_ratio/min': 0.13846345245838165, 'sampling/importance_sampling_ratio/mean': 1.027772307395935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1937409937381744, 'clip_ratio/low_mean': 0.012345679104328156, 'clip_ratio/low_min': 0.012345679104328156, 'clip_ratio/high_mean': 0.10648495145142078, 'clip_ratio/high_max': 0.10648495145142078, 'clip_ratio/region_mean': 0.11883063055574894, 'reward_total_mean': 0.3502022922039032, 'reward_meter_mean': 0.865246593952179, 'reward_meter_std': 0.24865683913230896, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9665993452072144, 'reward_repeat_penalty_std': 0.03361032158136368, 'reward_repeat_floor_mean': 0.9959484338760376, 'reward_repeat_floor_std': 0.003226237604394555, 'reward_near_duplicate_penalty_mean': 0.9829816818237305, 'reward_near_duplicate_penalty_std': 0.008810799568891525, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9833458662033081, 'reward_distinct_2_std': 0.03337503597140312, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3502022922039032, 'reward_total_composite_std': 0.09896644204854965, 'epoch': 0.07}

 43%|████▎     | 1415/3300 [5:25:23<5:44:57, 10.98s/it]INFO 04-12 21:23:50 [block_pool.py:378] Successfully reset prefix cache
WARNING 04-12 21:23:50 [detokenizer.py:246] Encountered invalid prefix detokenization error for request 14239, resetting decode stream.

 43%|████▎     | 1416/3300 [5:25:45<7:22:47, 14.10s/it]2026-04-12 21:24:11,266 | INFO | train_grpo_train | metrics_logged mode=train step=1416

                                                       
{'loss': -0.0741, 'grad_norm': 5.223179340362549, 'learning_rate': 5.712121212121212e-06, 'num_tokens': 2505517.0, 'completions/mean_length': 109.125, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.57143020629883, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.44829630851745605, 'rewards/meter/std': 0.4435102045536041, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0534522607922554, 'rewards/lexical_plausibility/mean': 0.9490580558776855, 'rewards/lexical_plausibility/std': 0.14408555626869202, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.258370578289032, 'rewards/repeat_penalty/mean': 0.9909513592720032, 'rewards/repeat_penalty/std': 0.016185717657208443, 'rewards/repeat_floor/mean': 0.9986426830291748, 'rewards/repeat_floor/std': 0.0024278489872813225, 'rewards/near_duplicate_penalty/mean': 0.9909513592720032, 'rewards/near_duplicate_penalty/std': 0.016185717657208443, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.1802024096250534, 'rewards/total_composite/std': 0.1855441778898239, 'reward': 0.1802024096250534, 'reward_std': 0.18554414808750153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16368629038333893, 'sampling/sampling_logp_difference/max': 3.311824083328247, 'sampling/importance_sampling_ratio/min': 0.036449626088142395, 'sampling/importance_sampling_ratio/mean': 1.0357638597488403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.105742834508419, 'clip_ratio/low_mean': 0.07330342661589384, 'clip_ratio/low_min': 0.07330342661589384, 'clip_ratio/high_mean': 0.05628856550902128, 'clip_ratio/high_max': 0.05628856550902128, 'clip_ratio/region_mean': 0.12959199212491512, 'reward_total_mean': 0.1802024096250534, 'reward_meter_mean': 0.44829630851745605, 'reward_meter_std': 0.4435102045536041, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0534522607922554, 'reward_lexical_plausibility_mean': 0.9490580558776855, 'reward_lexical_plausibility_std': 0.14408555626869202, 'reward_repeat_penalty_mean': 0.9909513592720032, 'reward_repeat_penalty_std': 0.016185717657208443, 'reward_repeat_floor_mean': 0.9986426830291748, 'reward_repeat_floor_std': 0.0024278489872813225, 'reward_near_duplicate_penalty_mean': 0.9909513592720032, 'reward_near_duplicate_penalty_std': 0.016185717657208443, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.258370578289032, 'reward_total_composite_mean': 0.1802024096250534, 'reward_total_composite_std': 0.1855441778898239, 'epoch': 0.07}

 43%|████▎     | 1416/3300 [5:25:45<7:22:47, 14.10s/it]INFO 04-12 21:24:11 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1417/3300 [5:25:58<7:21:00, 14.05s/it]2026-04-12 21:24:25,206 | INFO | train_grpo_train | metrics_logged mode=train step=1417

                                                       
{'loss': -0.0832, 'grad_norm': 2.090524435043335, 'learning_rate': 5.7090909090909096e-06, 'num_tokens': 2507170.0, 'completions/mean_length': 226.625, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 55.400001525878906, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7362270355224609, 'rewards/meter/std': 0.3507269322872162, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 0.8188418745994568, 'rewards/lexical_plausibility/std': 0.25061509013175964, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1927248239517212, 'rewards/repeat_penalty/mean': 0.917361855506897, 'rewards/repeat_penalty/std': 0.07393814623355865, 'rewards/repeat_floor/mean': 0.9913511276245117, 'rewards/repeat_floor/std': 0.0043833497911691666, 'rewards/near_duplicate_penalty/mean': 0.9649782776832581, 'rewards/near_duplicate_penalty/std': 0.025960108265280724, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9831043481826782, 'rewards/distinct_2/std': 0.029155395925045013, 'rewards/total_composite/mean': 0.20332878828048706, 'rewards/total_composite/std': 0.19292476773262024, 'reward': 0.20332878828048706, 'reward_std': 0.19292476773262024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20183491706848145, 'sampling/sampling_logp_difference/max': 1.5915579795837402, 'sampling/importance_sampling_ratio/min': 0.20360814034938812, 'sampling/importance_sampling_ratio/mean': 1.0448663234710693, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0840475112199783, 'clip_ratio/low_mean': 0.0438209343701601, 'clip_ratio/low_min': 0.0438209343701601, 'clip_ratio/high_mean': 0.06816853396594524, 'clip_ratio/high_max': 0.06816853396594524, 'clip_ratio/region_mean': 0.11198946833610535, 'reward_total_mean': 0.20332878828048706, 'reward_meter_mean': 0.7362270355224609, 'reward_meter_std': 0.3507269322872162, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 0.8188418745994568, 'reward_lexical_plausibility_std': 0.25061509013175964, 'reward_repeat_penalty_mean': 0.917361855506897, 'reward_repeat_penalty_std': 0.07393814623355865, 'reward_repeat_floor_mean': 0.9913511276245117, 'reward_repeat_floor_std': 0.0043833497911691666, 'reward_near_duplicate_penalty_mean': 0.9649782776832581, 'reward_near_duplicate_penalty_std': 0.025960108265280724, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9831043481826782, 'reward_distinct_2_std': 0.029155395925045013, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.1927248239517212, 'reward_total_composite_mean': 0.20332878828048706, 'reward_total_composite_std': 0.19292476773262024, 'epoch': 0.07}

 43%|████▎     | 1417/3300 [5:25:59<7:21:00, 14.05s/it]INFO 04-12 21:24:25 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1418/3300 [5:26:12<7:19:22, 14.01s/it]2026-04-12 21:24:39,108 | INFO | train_grpo_train | metrics_logged mode=train step=1418

                                                       
{'loss': -0.1125, 'grad_norm': 2.6924631595611572, 'learning_rate': 5.706060606060606e-06, 'num_tokens': 2509161.0, 'completions/mean_length': 179.875, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 69.16667175292969, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.7423150539398193, 'rewards/meter/std': 0.3726646900177002, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.10690449923276901, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.8799999952316284, 'rewards/count_floor/std': 0.03207135200500488, 'rewards/lexical_plausibility/mean': 0.8673089742660522, 'rewards/lexical_plausibility/std': 0.2240522801876068, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/repeat_penalty/mean': 0.9715592861175537, 'rewards/repeat_penalty/std': 0.03581685945391655, 'rewards/repeat_floor/mean': 0.9962940216064453, 'rewards/repeat_floor/std': 0.003976048436015844, 'rewards/near_duplicate_penalty/mean': 0.981366753578186, 'rewards/near_duplicate_penalty/std': 0.015348234213888645, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.20716926455497742, 'rewards/total_composite/std': 0.1545560508966446, 'reward': 0.20716926455497742, 'reward_std': 0.1545560508966446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1818479299545288, 'sampling/sampling_logp_difference/max': 1.5515947341918945, 'sampling/importance_sampling_ratio/min': 0.21190977096557617, 'sampling/importance_sampling_ratio/mean': 1.0296623706817627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1521000415086746, 'clip_ratio/low_mean': 0.023076923564076424, 'clip_ratio/low_min': 0.023076923564076424, 'clip_ratio/high_mean': 0.11405929550528526, 'clip_ratio/high_max': 0.11405929550528526, 'clip_ratio/region_mean': 0.1371362190693617, 'reward_total_mean': 0.20716926455497742, 'reward_meter_mean': 0.7423150539398193, 'reward_meter_std': 0.3726646900177002, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.10690449923276901, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.8799999952316284, 'reward_count_floor_std': 0.03207135200500488, 'reward_lexical_plausibility_mean': 0.8673089742660522, 'reward_lexical_plausibility_std': 0.2240522801876068, 'reward_repeat_penalty_mean': 0.9715592861175537, 'reward_repeat_penalty_std': 0.03581685945391655, 'reward_repeat_floor_mean': 0.9962940216064453, 'reward_repeat_floor_std': 0.003976048436015844, 'reward_near_duplicate_penalty_mean': 0.981366753578186, 'reward_near_duplicate_penalty_std': 0.015348234213888645, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.20716926455497742, 'reward_total_composite_std': 0.1545560508966446, 'epoch': 0.07}

 43%|████▎     | 1418/3300 [5:26:12<7:19:22, 14.01s/it]INFO 04-12 21:24:39 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1419/3300 [5:26:27<7:23:03, 14.13s/it]2026-04-12 21:24:53,528 | INFO | train_grpo_train | metrics_logged mode=train step=1419

                                                       
{'loss': -0.0712, 'grad_norm': 3.0327870845794678, 'learning_rate': 5.703030303030303e-06, 'num_tokens': 2511066.0, 'completions/mean_length': 172.125, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 58.833335876464844, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.46629881858825684, 'rewards/meter/std': 0.39893838763237, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.8942897319793701, 'rewards/lexical_plausibility/std': 0.19582656025886536, 'rewards/judge_quality/mean': 0.2499999850988388, 'rewards/judge_quality/std': 0.16903084516525269, 'rewards/repeat_penalty/mean': 0.9377177953720093, 'rewards/repeat_penalty/std': 0.054013270884752274, 'rewards/repeat_floor/mean': 0.991885781288147, 'rewards/repeat_floor/std': 0.006311929319053888, 'rewards/near_duplicate_penalty/mean': 0.9604518413543701, 'rewards/near_duplicate_penalty/std': 0.022788435220718384, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9757559299468994, 'rewards/distinct_2/std': 0.03709888085722923, 'rewards/total_composite/mean': 0.15347814559936523, 'rewards/total_composite/std': 0.16966484487056732, 'reward': 0.15347814559936523, 'reward_std': 0.16966484487056732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14197510480880737, 'sampling/sampling_logp_difference/max': 1.073725700378418, 'sampling/importance_sampling_ratio/min': 0.3417329490184784, 'sampling/importance_sampling_ratio/mean': 1.0360581874847412, 'sampling/importance_sampling_ratio/max': 1.882956624031067, 'entropy': 1.0979614108800888, 'clip_ratio/low_mean': 0.04020364489406347, 'clip_ratio/low_min': 0.04020364489406347, 'clip_ratio/high_mean': 0.040330946911126375, 'clip_ratio/high_max': 0.040330946911126375, 'clip_ratio/region_mean': 0.08053459180518985, 'reward_total_mean': 0.15347814559936523, 'reward_meter_mean': 0.46629881858825684, 'reward_meter_std': 0.39893838763237, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.8942897319793701, 'reward_lexical_plausibility_std': 0.19582656025886536, 'reward_repeat_penalty_mean': 0.9377177953720093, 'reward_repeat_penalty_std': 0.054013270884752274, 'reward_repeat_floor_mean': 0.991885781288147, 'reward_repeat_floor_std': 0.006311929319053888, 'reward_near_duplicate_penalty_mean': 0.9604518413543701, 'reward_near_duplicate_penalty_std': 0.022788435220718384, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9757559299468994, 'reward_distinct_2_std': 0.03709888085722923, 'reward_judge_quality_mean': 0.2499999850988388, 'reward_judge_quality_std': 0.16903084516525269, 'reward_total_composite_mean': 0.15347814559936523, 'reward_total_composite_std': 0.16966484487056732, 'epoch': 0.07}

 43%|████▎     | 1419/3300 [5:26:27<7:23:03, 14.13s/it]INFO 04-12 21:24:53 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1420/3300 [5:26:42<7:33:27, 14.47s/it]2026-04-12 21:25:08,813 | INFO | train_grpo_train | metrics_logged mode=train step=1420

                                                       
{'loss': -0.0982, 'grad_norm': 2.394282102584839, 'learning_rate': 5.7e-06, 'num_tokens': 2512947.0, 'completions/mean_length': 247.125, 'completions/min_length': 79.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 88.20000457763672, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.25709956884384155, 'rewards/meter/std': 0.21952390670776367, 'rewards/count_adherence/mean': 0.675000011920929, 'rewards/count_adherence/std': 0.23754701018333435, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9025000333786011, 'rewards/count_floor/std': 0.0712641030550003, 'rewards/lexical_plausibility/mean': 0.8210036754608154, 'rewards/lexical_plausibility/std': 0.2591615319252014, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.2070196568965912, 'rewards/repeat_penalty/mean': 0.923556923866272, 'rewards/repeat_penalty/std': 0.08092190325260162, 'rewards/repeat_floor/mean': 0.9922753572463989, 'rewards/repeat_floor/std': 0.005748991388827562, 'rewards/near_duplicate_penalty/mean': 0.9716846942901611, 'rewards/near_duplicate_penalty/std': 0.02343595214188099, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9821963906288147, 'rewards/distinct_2/std': 0.023891696706414223, 'rewards/total_composite/mean': 0.08479015529155731, 'rewards/total_composite/std': 0.09267206490039825, 'reward': 0.08479015529155731, 'reward_std': 0.09267206490039825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16568785905838013, 'sampling/sampling_logp_difference/max': 1.5872840881347656, 'sampling/importance_sampling_ratio/min': 0.20448020100593567, 'sampling/importance_sampling_ratio/mean': 1.011500358581543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8012524545192719, 'clip_ratio/low_mean': 0.036545682698488235, 'clip_ratio/low_min': 0.036545682698488235, 'clip_ratio/high_mean': 0.05723891407251358, 'clip_ratio/high_max': 0.05723891407251358, 'clip_ratio/region_mean': 0.09378459677100182, 'reward_total_mean': 0.08479015529155731, 'reward_meter_mean': 0.25709956884384155, 'reward_meter_std': 0.21952390670776367, 'reward_count_adherence_mean': 0.675000011920929, 'reward_count_adherence_std': 0.23754701018333435, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9025000333786011, 'reward_count_floor_std': 0.0712641030550003, 'reward_lexical_plausibility_mean': 0.8210036754608154, 'reward_lexical_plausibility_std': 0.2591615319252014, 'reward_repeat_penalty_mean': 0.923556923866272, 'reward_repeat_penalty_std': 0.08092190325260162, 'reward_repeat_floor_mean': 0.9922753572463989, 'reward_repeat_floor_std': 0.005748991388827562, 'reward_near_duplicate_penalty_mean': 0.9716846942901611, 'reward_near_duplicate_penalty_std': 0.02343595214188099, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9821963906288147, 'reward_distinct_2_std': 0.023891696706414223, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.2070196568965912, 'reward_total_composite_mean': 0.08479015529155731, 'reward_total_composite_std': 0.09267206490039825, 'epoch': 0.07}

 43%|████▎     | 1420/3300 [5:26:42<7:33:27, 14.47s/it]INFO 04-12 21:25:09 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1421/3300 [5:26:56<7:28:45, 14.33s/it]2026-04-12 21:25:22,816 | INFO | train_grpo_train | metrics_logged mode=train step=1421

                                                       
{'loss': -0.1351, 'grad_norm': 2.1773228645324707, 'learning_rate': 5.696969696969698e-06, 'num_tokens': 2515080.0, 'completions/mean_length': 256.625, 'completions/min_length': 88.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 103.4000015258789, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7861090302467346, 'rewards/meter/std': 0.3016795516014099, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8499999642372131, 'rewards/count_floor/std': 0.040089186280965805, 'rewards/lexical_plausibility/mean': 0.8296400308609009, 'rewards/lexical_plausibility/std': 0.24067197740077972, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/repeat_penalty/mean': 0.880561888217926, 'rewards/repeat_penalty/std': 0.10514307022094727, 'rewards/repeat_floor/mean': 0.9877530932426453, 'rewards/repeat_floor/std': 0.009224310517311096, 'rewards/near_duplicate_penalty/mean': 0.962739109992981, 'rewards/near_duplicate_penalty/std': 0.026324665173888206, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.9551913738250732, 'rewards/distinct_2/std': 0.03991148620843887, 'rewards/total_composite/mean': 0.18429060280323029, 'rewards/total_composite/std': 0.13176029920578003, 'reward': 0.18429060280323029, 'reward_std': 0.13176031410694122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14812849462032318, 'sampling/sampling_logp_difference/max': 1.5598716735839844, 'sampling/importance_sampling_ratio/min': 0.21016304194927216, 'sampling/importance_sampling_ratio/mean': 1.0318808555603027, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8566962331533432, 'clip_ratio/low_mean': 0.02350427396595478, 'clip_ratio/low_min': 0.02350427396595478, 'clip_ratio/high_mean': 0.0722195915877819, 'clip_ratio/high_max': 0.0722195915877819, 'clip_ratio/region_mean': 0.09572386555373669, 'reward_total_mean': 0.18429060280323029, 'reward_meter_mean': 0.7861090302467346, 'reward_meter_std': 0.3016795516014099, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.13363061845302582, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8499999642372131, 'reward_count_floor_std': 0.040089186280965805, 'reward_lexical_plausibility_mean': 0.8296400308609009, 'reward_lexical_plausibility_std': 0.24067197740077972, 'reward_repeat_penalty_mean': 0.880561888217926, 'reward_repeat_penalty_std': 0.10514307022094727, 'reward_repeat_floor_mean': 0.9877530932426453, 'reward_repeat_floor_std': 0.009224310517311096, 'reward_near_duplicate_penalty_mean': 0.962739109992981, 'reward_near_duplicate_penalty_std': 0.026324665173888206, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.9551913738250732, 'reward_distinct_2_std': 0.03991148620843887, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.18429060280323029, 'reward_total_composite_std': 0.13176029920578003, 'epoch': 0.07}

 43%|████▎     | 1421/3300 [5:26:56<7:28:45, 14.33s/it]INFO 04-12 21:25:23 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1422/3300 [5:27:10<7:26:26, 14.26s/it]2026-04-12 21:25:37,026 | INFO | train_grpo_train | metrics_logged mode=train step=1422

                                                       
{'loss': -0.0525, 'grad_norm': 4.93320369720459, 'learning_rate': 5.693939393939394e-06, 'num_tokens': 2517334.0, 'completions/mean_length': 144.75, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 92.28572082519531, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.5939919352531433, 'rewards/meter/std': 0.31222984194755554, 'rewards/count_adherence/mean': 0.6071429252624512, 'rewards/count_adherence/std': 0.10101525485515594, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8821429014205933, 'rewards/count_floor/std': 0.03030458092689514, 'rewards/lexical_plausibility/mean': 0.9413983821868896, 'rewards/lexical_plausibility/std': 0.16575030982494354, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8164722919464111, 'rewards/repeat_penalty/std': 0.13678307831287384, 'rewards/repeat_floor/mean': 0.9817279577255249, 'rewards/repeat_floor/std': 0.011806267313659191, 'rewards/near_duplicate_penalty/mean': 0.958777666091919, 'rewards/near_duplicate_penalty/std': 0.015340616926550865, 'rewards/opening_diversity/mean': 0.9255208373069763, 'rewards/opening_diversity/std': 0.11379190534353256, 'rewards/distinct_2/mean': 0.9153342247009277, 'rewards/distinct_2/std': 0.04667168855667114, 'rewards/total_composite/mean': 0.19468994438648224, 'rewards/total_composite/std': 0.13996662199497223, 'reward': 0.19468994438648224, 'reward_std': 0.13996662199497223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13757051527500153, 'sampling/sampling_logp_difference/max': 2.128413677215576, 'sampling/importance_sampling_ratio/min': 0.11902596056461334, 'sampling/importance_sampling_ratio/mean': 1.0129565000534058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9266053065657616, 'clip_ratio/low_mean': 0.056586208287626505, 'clip_ratio/low_min': 0.056586208287626505, 'clip_ratio/high_mean': 0.05667359568178654, 'clip_ratio/high_max': 0.05667359568178654, 'clip_ratio/region_mean': 0.11325980396941304, 'reward_total_mean': 0.19468994438648224, 'reward_meter_mean': 0.5939919352531433, 'reward_meter_std': 0.31222984194755554, 'reward_count_adherence_mean': 0.6071429252624512, 'reward_count_adherence_std': 0.10101525485515594, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8821429014205933, 'reward_count_floor_std': 0.03030458092689514, 'reward_lexical_plausibility_mean': 0.9413983821868896, 'reward_lexical_plausibility_std': 0.16575030982494354, 'reward_repeat_penalty_mean': 0.8164722919464111, 'reward_repeat_penalty_std': 0.13678307831287384, 'reward_repeat_floor_mean': 0.9817279577255249, 'reward_repeat_floor_std': 0.011806267313659191, 'reward_near_duplicate_penalty_mean': 0.958777666091919, 'reward_near_duplicate_penalty_std': 0.015340616926550865, 'reward_opening_diversity_mean': 0.9255208373069763, 'reward_opening_diversity_std': 0.11379190534353256, 'reward_distinct_2_mean': 0.9153342247009277, 'reward_distinct_2_std': 0.04667168855667114, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.19468994438648224, 'reward_total_composite_std': 0.13996662199497223, 'epoch': 0.07}

 43%|████▎     | 1422/3300 [5:27:10<7:26:26, 14.26s/it]INFO 04-12 21:25:37 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1423/3300 [5:27:24<7:21:51, 14.12s/it]2026-04-12 21:25:50,712 | INFO | train_grpo_train | metrics_logged mode=train step=1423

                                                       
{'loss': -0.0403, 'grad_norm': 4.625883102416992, 'learning_rate': 5.690909090909091e-06, 'num_tokens': 2518794.0, 'completions/mean_length': 95.5, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.32400161027908325, 'rewards/meter/std': 0.3609822988510132, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9517061710357666, 'rewards/lexical_plausibility/std': 0.13659556210041046, 'rewards/judge_quality/mean': 0.5512499809265137, 'rewards/judge_quality/std': 0.3296508193016052, 'rewards/repeat_penalty/mean': 0.975520133972168, 'rewards/repeat_penalty/std': 0.019892379641532898, 'rewards/repeat_floor/mean': 0.9963279962539673, 'rewards/repeat_floor/std': 0.00298385345377028, 'rewards/near_duplicate_penalty/mean': 0.975520133972168, 'rewards/near_duplicate_penalty/std': 0.019892379641532898, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.15501461923122406, 'rewards/total_composite/std': 0.1385594606399536, 'reward': 0.15501461923122406, 'reward_std': 0.1385594606399536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1565273255109787, 'sampling/sampling_logp_difference/max': 3.7504048347473145, 'sampling/importance_sampling_ratio/min': 0.023508228361606598, 'sampling/importance_sampling_ratio/mean': 1.002364158630371, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6725737601518631, 'clip_ratio/low_mean': 0.05694728344678879, 'clip_ratio/low_min': 0.05694728344678879, 'clip_ratio/high_mean': 0.08240068983286619, 'clip_ratio/high_max': 0.08240068983286619, 'clip_ratio/region_mean': 0.13934797327965498, 'reward_total_mean': 0.15501461923122406, 'reward_meter_mean': 0.32400161027908325, 'reward_meter_std': 0.3609822988510132, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9517061710357666, 'reward_lexical_plausibility_std': 0.13659556210041046, 'reward_repeat_penalty_mean': 0.975520133972168, 'reward_repeat_penalty_std': 0.019892379641532898, 'reward_repeat_floor_mean': 0.9963279962539673, 'reward_repeat_floor_std': 0.00298385345377028, 'reward_near_duplicate_penalty_mean': 0.975520133972168, 'reward_near_duplicate_penalty_std': 0.019892379641532898, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5512499809265137, 'reward_judge_quality_std': 0.3296508193016052, 'reward_total_composite_mean': 0.15501461923122406, 'reward_total_composite_std': 0.1385594606399536, 'epoch': 0.07}

 43%|████▎     | 1423/3300 [5:27:24<7:21:51, 14.12s/it]INFO 04-12 21:25:50 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1424/3300 [5:27:34<6:41:35, 12.84s/it]2026-04-12 21:26:00,566 | INFO | train_grpo_train | metrics_logged mode=train step=1424

                                                       
{'loss': 0.0166, 'grad_norm': 11.637589454650879, 'learning_rate': 5.687878787878789e-06, 'num_tokens': 2520349.0, 'completions/mean_length': 36.375, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8646896481513977, 'rewards/meter/std': 0.21324127912521362, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.5698262453079224, 'rewards/repeat_penalty/std': 0.1878826916217804, 'rewards/repeat_floor/mean': 0.9449223279953003, 'rewards/repeat_floor/std': 0.02574588544666767, 'rewards/near_duplicate_penalty/mean': 0.805144190788269, 'rewards/near_duplicate_penalty/std': 0.10607056319713593, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.7961185574531555, 'rewards/distinct_2/std': 0.10696239769458771, 'rewards/total_composite/mean': 0.2628685235977173, 'rewards/total_composite/std': 0.0961027517914772, 'reward': 0.2628685235977173, 'reward_std': 0.0961027517914772, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13617858290672302, 'sampling/sampling_logp_difference/max': 1.6272382736206055, 'sampling/importance_sampling_ratio/min': 0.1964714229106903, 'sampling/importance_sampling_ratio/mean': 1.0462281703948975, 'sampling/importance_sampling_ratio/max': 1.905685305595398, 'entropy': 1.3303654342889786, 'clip_ratio/low_mean': 0.055872919503599405, 'clip_ratio/low_min': 0.055872919503599405, 'clip_ratio/high_mean': 0.0699146525003016, 'clip_ratio/high_max': 0.0699146525003016, 'clip_ratio/region_mean': 0.125787572003901, 'reward_total_mean': 0.2628685235977173, 'reward_meter_mean': 0.8646896481513977, 'reward_meter_std': 0.21324127912521362, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.5698262453079224, 'reward_repeat_penalty_std': 0.1878826916217804, 'reward_repeat_floor_mean': 0.9449223279953003, 'reward_repeat_floor_std': 0.02574588544666767, 'reward_near_duplicate_penalty_mean': 0.805144190788269, 'reward_near_duplicate_penalty_std': 0.10607056319713593, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.7961185574531555, 'reward_distinct_2_std': 0.10696239769458771, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.2628685235977173, 'reward_total_composite_std': 0.0961027517914772, 'epoch': 0.07}

 43%|████▎     | 1424/3300 [5:27:34<6:41:35, 12.84s/it]INFO 04-12 21:26:00 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1425/3300 [5:27:44<6:12:55, 11.93s/it]2026-04-12 21:26:10,380 | INFO | train_grpo_train | metrics_logged mode=train step=1425

                                                       
{'loss': 0.0416, 'grad_norm': 14.353424072265625, 'learning_rate': 5.684848484848485e-06, 'num_tokens': 2521967.0, 'completions/mean_length': 42.25, 'completions/min_length': 37.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7335023283958435, 'rewards/meter/std': 0.3352108597755432, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.9804375171661377, 'rewards/repeat_penalty/std': 0.0216483473777771, 'rewards/repeat_floor/mean': 0.9970656037330627, 'rewards/repeat_floor/std': 0.0032472480088472366, 'rewards/near_duplicate_penalty/mean': 0.9804375171661377, 'rewards/near_duplicate_penalty/std': 0.0216483473777771, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3722691535949707, 'rewards/total_composite/std': 0.18204954266548157, 'reward': 0.3722691535949707, 'reward_std': 0.18204952776432037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1473720371723175, 'sampling/sampling_logp_difference/max': 1.531935691833496, 'sampling/importance_sampling_ratio/min': 0.21611693501472473, 'sampling/importance_sampling_ratio/mean': 1.0309715270996094, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3216603696346283, 'clip_ratio/low_mean': 0.04184154234826565, 'clip_ratio/low_min': 0.04184154234826565, 'clip_ratio/high_mean': 0.08992148283869028, 'clip_ratio/high_max': 0.08992148283869028, 'clip_ratio/region_mean': 0.13176302518695593, 'reward_total_mean': 0.3722691535949707, 'reward_meter_mean': 0.7335023283958435, 'reward_meter_std': 0.3352108597755432, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9804375171661377, 'reward_repeat_penalty_std': 0.0216483473777771, 'reward_repeat_floor_mean': 0.9970656037330627, 'reward_repeat_floor_std': 0.0032472480088472366, 'reward_near_duplicate_penalty_mean': 0.9804375171661377, 'reward_near_duplicate_penalty_std': 0.0216483473777771, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.3722691535949707, 'reward_total_composite_std': 0.18204954266548157, 'epoch': 0.07}

 43%|████▎     | 1425/3300 [5:27:44<6:12:55, 11.93s/it]INFO 04-12 21:26:10 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1426/3300 [5:27:57<6:27:23, 12.40s/it]2026-04-12 21:26:23,879 | INFO | train_grpo_train | metrics_logged mode=train step=1426

                                                       
{'loss': -0.0966, 'grad_norm': 2.5535073280334473, 'learning_rate': 5.681818181818183e-06, 'num_tokens': 2523679.0, 'completions/mean_length': 233.0, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 65.5999984741211, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7778260111808777, 'rewards/meter/std': 0.3299368917942047, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.053032997995615005, 'rewards/lexical_plausibility/mean': 0.8410569429397583, 'rewards/lexical_plausibility/std': 0.2204350382089615, 'rewards/judge_quality/mean': 0.32874998450279236, 'rewards/judge_quality/std': 0.284677118062973, 'rewards/repeat_penalty/mean': 0.7332978248596191, 'rewards/repeat_penalty/std': 0.28560400009155273, 'rewards/repeat_floor/mean': 0.9716359972953796, 'rewards/repeat_floor/std': 0.03159656748175621, 'rewards/near_duplicate_penalty/mean': 0.9330688714981079, 'rewards/near_duplicate_penalty/std': 0.06271221488714218, 'rewards/opening_diversity/mean': 0.8696428537368774, 'rewards/opening_diversity/std': 0.15559300780296326, 'rewards/distinct_2/mean': 0.88330078125, 'rewards/distinct_2/std': 0.18401280045509338, 'rewards/total_composite/mean': 0.2714325487613678, 'rewards/total_composite/std': 0.26420652866363525, 'reward': 0.2714325487613678, 'reward_std': 0.26420652866363525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13892415165901184, 'sampling/sampling_logp_difference/max': 1.608747124671936, 'sampling/importance_sampling_ratio/min': 0.20013821125030518, 'sampling/importance_sampling_ratio/mean': 1.0274556875228882, 'sampling/importance_sampling_ratio/max': 1.9897373914718628, 'entropy': 0.7331947535276413, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.059100549668073654, 'clip_ratio/high_max': 0.059100549668073654, 'clip_ratio/region_mean': 0.059100549668073654, 'reward_total_mean': 0.2714325487613678, 'reward_meter_mean': 0.7778260111808777, 'reward_meter_std': 0.3299368917942047, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.053032997995615005, 'reward_lexical_plausibility_mean': 0.8410569429397583, 'reward_lexical_plausibility_std': 0.2204350382089615, 'reward_repeat_penalty_mean': 0.7332978248596191, 'reward_repeat_penalty_std': 0.28560400009155273, 'reward_repeat_floor_mean': 0.9716359972953796, 'reward_repeat_floor_std': 0.03159656748175621, 'reward_near_duplicate_penalty_mean': 0.9330688714981079, 'reward_near_duplicate_penalty_std': 0.06271221488714218, 'reward_opening_diversity_mean': 0.8696428537368774, 'reward_opening_diversity_std': 0.15559300780296326, 'reward_distinct_2_mean': 0.88330078125, 'reward_distinct_2_std': 0.18401280045509338, 'reward_judge_quality_mean': 0.32874998450279236, 'reward_judge_quality_std': 0.284677118062973, 'reward_total_composite_mean': 0.2714325487613678, 'reward_total_composite_std': 0.26420652866363525, 'epoch': 0.08}

 43%|████▎     | 1426/3300 [5:27:57<6:27:23, 12.40s/it]INFO 04-12 21:26:24 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1427/3300 [5:28:11<6:45:48, 13.00s/it]2026-04-12 21:26:38,291 | INFO | train_grpo_train | metrics_logged mode=train step=1427

                                                       
{'loss': -0.0987, 'grad_norm': 3.4083316326141357, 'learning_rate': 5.67878787878788e-06, 'num_tokens': 2525468.0, 'completions/mean_length': 113.625, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.71428680419922, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6425893902778625, 'rewards/meter/std': 0.3112960755825043, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9448069334030151, 'rewards/lexical_plausibility/std': 0.15610963106155396, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8449782133102417, 'rewards/repeat_penalty/std': 0.08276111632585526, 'rewards/repeat_floor/mean': 0.9810770750045776, 'rewards/repeat_floor/std': 0.009634961374104023, 'rewards/near_duplicate_penalty/mean': 0.9288694858551025, 'rewards/near_duplicate_penalty/std': 0.02703668735921383, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9082965850830078, 'rewards/distinct_2/std': 0.06722333282232285, 'rewards/total_composite/mean': 0.24714139103889465, 'rewards/total_composite/std': 0.13522407412528992, 'reward': 0.24714139103889465, 'reward_std': 0.13522405922412872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15562821924686432, 'sampling/sampling_logp_difference/max': 2.1100234985351562, 'sampling/importance_sampling_ratio/min': 0.12123511731624603, 'sampling/importance_sampling_ratio/mean': 1.0316572189331055, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0921587944030762, 'clip_ratio/low_mean': 0.021780303679406643, 'clip_ratio/low_min': 0.021780303679406643, 'clip_ratio/high_mean': 0.09094184171408415, 'clip_ratio/high_max': 0.09094184171408415, 'clip_ratio/region_mean': 0.11272214539349079, 'reward_total_mean': 0.24714139103889465, 'reward_meter_mean': 0.6425893902778625, 'reward_meter_std': 0.3112960755825043, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9448069334030151, 'reward_lexical_plausibility_std': 0.15610963106155396, 'reward_repeat_penalty_mean': 0.8449782133102417, 'reward_repeat_penalty_std': 0.08276111632585526, 'reward_repeat_floor_mean': 0.9810770750045776, 'reward_repeat_floor_std': 0.009634961374104023, 'reward_near_duplicate_penalty_mean': 0.9288694858551025, 'reward_near_duplicate_penalty_std': 0.02703668735921383, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9082965850830078, 'reward_distinct_2_std': 0.06722333282232285, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.24714139103889465, 'reward_total_composite_std': 0.13522407412528992, 'epoch': 0.08}

 43%|████▎     | 1427/3300 [5:28:12<6:45:48, 13.00s/it]INFO 04-12 21:26:38 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1428/3300 [5:28:21<6:12:29, 11.94s/it]2026-04-12 21:26:47,735 | INFO | train_grpo_train | metrics_logged mode=train step=1428

                                                       
{'loss': 0.1098, 'grad_norm': 11.471367835998535, 'learning_rate': 5.675757575757577e-06, 'num_tokens': 2527285.0, 'completions/mean_length': 51.125, 'completions/min_length': 42.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6151893734931946, 'rewards/meter/std': 0.3938708007335663, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6212500333786011, 'rewards/judge_quality/std': 0.19134393334388733, 'rewards/repeat_penalty/mean': 0.9832082390785217, 'rewards/repeat_penalty/std': 0.02674015797674656, 'rewards/repeat_floor/mean': 0.9977906942367554, 'rewards/repeat_floor/std': 0.0032035009935498238, 'rewards/near_duplicate_penalty/mean': 0.9886648654937744, 'rewards/near_duplicate_penalty/std': 0.013352897018194199, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9943438768386841, 'rewards/distinct_2/std': 0.015997890383005142, 'rewards/total_composite/mean': 0.4150158166885376, 'rewards/total_composite/std': 0.3356136381626129, 'reward': 0.4150158166885376, 'reward_std': 0.3356136381626129, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12866315245628357, 'sampling/sampling_logp_difference/max': 1.5372400283813477, 'sampling/importance_sampling_ratio/min': 0.21497361361980438, 'sampling/importance_sampling_ratio/mean': 1.0516523122787476, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1902798116207123, 'clip_ratio/low_mean': 0.09253580030053854, 'clip_ratio/low_min': 0.09253580030053854, 'clip_ratio/high_mean': 0.029715219978243113, 'clip_ratio/high_max': 0.029715219978243113, 'clip_ratio/region_mean': 0.12225102027878165, 'reward_total_mean': 0.4150158166885376, 'reward_meter_mean': 0.6151893734931946, 'reward_meter_std': 0.3938708007335663, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9832082390785217, 'reward_repeat_penalty_std': 0.02674015797674656, 'reward_repeat_floor_mean': 0.9977906942367554, 'reward_repeat_floor_std': 0.0032035009935498238, 'reward_near_duplicate_penalty_mean': 0.9886648654937744, 'reward_near_duplicate_penalty_std': 0.013352897018194199, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9943438768386841, 'reward_distinct_2_std': 0.015997890383005142, 'reward_judge_quality_mean': 0.6212500333786011, 'reward_judge_quality_std': 0.19134393334388733, 'reward_total_composite_mean': 0.4150158166885376, 'reward_total_composite_std': 0.3356136381626129, 'epoch': 0.08}

 43%|████▎     | 1428/3300 [5:28:21<6:12:29, 11.94s/it]INFO 04-12 21:26:47 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1429/3300 [5:28:35<6:29:07, 12.48s/it]2026-04-12 21:27:01,473 | INFO | train_grpo_train | metrics_logged mode=train step=1429

                                                       
{'loss': -0.0657, 'grad_norm': 4.530540943145752, 'learning_rate': 5.672727272727273e-06, 'num_tokens': 2529140.0, 'completions/mean_length': 119.875, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.857147216796875, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.39749693870544434, 'rewards/meter/std': 0.3820936977863312, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.0707106739282608, 'rewards/lexical_plausibility/mean': 0.9280478358268738, 'rewards/lexical_plausibility/std': 0.20351143181324005, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.34280356764793396, 'rewards/repeat_penalty/mean': 0.7069588899612427, 'rewards/repeat_penalty/std': 0.2796173691749573, 'rewards/repeat_floor/mean': 0.9735864400863647, 'rewards/repeat_floor/std': 0.027258770540356636, 'rewards/near_duplicate_penalty/mean': 0.9582749605178833, 'rewards/near_duplicate_penalty/std': 0.051941584795713425, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.28149792551994324, 'rewards/distinct_2/mean': 0.9218907356262207, 'rewards/distinct_2/std': 0.07085563987493515, 'rewards/total_composite/mean': 0.3610021471977234, 'rewards/total_composite/std': 0.35330790281295776, 'reward': 0.3610021471977234, 'reward_std': 0.3533078730106354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13061189651489258, 'sampling/sampling_logp_difference/max': 3.2633328437805176, 'sampling/importance_sampling_ratio/min': 0.038260672241449356, 'sampling/importance_sampling_ratio/mean': 1.01331627368927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5737249925732613, 'clip_ratio/low_mean': 0.06658971542492509, 'clip_ratio/low_min': 0.06658971542492509, 'clip_ratio/high_mean': 0.04152960516512394, 'clip_ratio/high_max': 0.04152960516512394, 'clip_ratio/region_mean': 0.10811932059004903, 'reward_total_mean': 0.3610021471977234, 'reward_meter_mean': 0.39749693870544434, 'reward_meter_std': 0.3820936977863312, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.0707106739282608, 'reward_lexical_plausibility_mean': 0.9280478358268738, 'reward_lexical_plausibility_std': 0.20351143181324005, 'reward_repeat_penalty_mean': 0.7069588899612427, 'reward_repeat_penalty_std': 0.2796173691749573, 'reward_repeat_floor_mean': 0.9735864400863647, 'reward_repeat_floor_std': 0.027258770540356636, 'reward_near_duplicate_penalty_mean': 0.9582749605178833, 'reward_near_duplicate_penalty_std': 0.051941584795713425, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.28149792551994324, 'reward_distinct_2_mean': 0.9218907356262207, 'reward_distinct_2_std': 0.07085563987493515, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.34280356764793396, 'reward_total_composite_mean': 0.3610021471977234, 'reward_total_composite_std': 0.35330790281295776, 'epoch': 0.08}

 43%|████▎     | 1429/3300 [5:28:35<6:29:07, 12.48s/it]INFO 04-12 21:27:01 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1430/3300 [5:28:49<6:48:09, 13.10s/it]2026-04-12 21:27:16,008 | INFO | train_grpo_train | metrics_logged mode=train step=1430

                                                       
{'loss': -0.0997, 'grad_norm': 2.382554531097412, 'learning_rate': 5.6696969696969705e-06, 'num_tokens': 2530843.0, 'completions/mean_length': 287.875, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7674127817153931, 'rewards/meter/std': 0.37033113837242126, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.08017838001251221, 'rewards/lexical_plausibility/mean': 0.7275892496109009, 'rewards/lexical_plausibility/std': 0.2960568368434906, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.19809359312057495, 'rewards/repeat_penalty/mean': 0.961026132106781, 'rewards/repeat_penalty/std': 0.03477640822529793, 'rewards/repeat_floor/mean': 0.9950242638587952, 'rewards/repeat_floor/std': 0.004474334418773651, 'rewards/near_duplicate_penalty/mean': 0.9762563705444336, 'rewards/near_duplicate_penalty/std': 0.02430010214447975, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9842867255210876, 'rewards/distinct_2/std': 0.020308425650000572, 'rewards/total_composite/mean': 0.18792983889579773, 'rewards/total_composite/std': 0.18820473551750183, 'reward': 0.18792983889579773, 'reward_std': 0.18820473551750183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18625052273273468, 'sampling/sampling_logp_difference/max': 4.266085624694824, 'sampling/importance_sampling_ratio/min': 0.01403662096709013, 'sampling/importance_sampling_ratio/mean': 1.0408921241760254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9037230908870697, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06337859574705362, 'clip_ratio/high_max': 0.06337859574705362, 'clip_ratio/region_mean': 0.06337859574705362, 'reward_total_mean': 0.18792983889579773, 'reward_meter_mean': 0.7674127817153931, 'reward_meter_std': 0.37033113837242126, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.26726123690605164, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.08017838001251221, 'reward_lexical_plausibility_mean': 0.7275892496109009, 'reward_lexical_plausibility_std': 0.2960568368434906, 'reward_repeat_penalty_mean': 0.961026132106781, 'reward_repeat_penalty_std': 0.03477640822529793, 'reward_repeat_floor_mean': 0.9950242638587952, 'reward_repeat_floor_std': 0.004474334418773651, 'reward_near_duplicate_penalty_mean': 0.9762563705444336, 'reward_near_duplicate_penalty_std': 0.02430010214447975, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9842867255210876, 'reward_distinct_2_std': 0.020308425650000572, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.19809359312057495, 'reward_total_composite_mean': 0.18792983889579773, 'reward_total_composite_std': 0.18820473551750183, 'epoch': 0.08}

 43%|████▎     | 1430/3300 [5:28:49<6:48:09, 13.10s/it]INFO 04-12 21:27:16 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1431/3300 [5:29:00<6:26:37, 12.41s/it]2026-04-12 21:27:26,849 | INFO | train_grpo_train | metrics_logged mode=train step=1431

                                                       
{'loss': 0.025, 'grad_norm': 13.66073226928711, 'learning_rate': 5.666666666666667e-06, 'num_tokens': 2532534.0, 'completions/mean_length': 34.375, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.772925615310669, 'rewards/meter/std': 0.4008168876171112, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6387499570846558, 'rewards/judge_quality/std': 0.2639501094818115, 'rewards/repeat_penalty/mean': 0.9684585928916931, 'rewards/repeat_penalty/std': 0.03413161262869835, 'rewards/repeat_floor/mean': 0.9956202507019043, 'rewards/repeat_floor/std': 0.0043912623077631, 'rewards/near_duplicate_penalty/mean': 0.9747804403305054, 'rewards/near_duplicate_penalty/std': 0.023137658834457397, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9933686852455139, 'rewards/distinct_2/std': 0.01875614933669567, 'rewards/total_composite/mean': 0.4710928201675415, 'rewards/total_composite/std': 0.3410051465034485, 'reward': 0.4710928201675415, 'reward_std': 0.3410051465034485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14480949938297272, 'sampling/sampling_logp_difference/max': 1.6082849502563477, 'sampling/importance_sampling_ratio/min': 0.20023071765899658, 'sampling/importance_sampling_ratio/mean': 0.9996879696846008, 'sampling/importance_sampling_ratio/max': 1.7699857950210571, 'entropy': 1.0403863191604614, 'clip_ratio/low_mean': 0.08572594448924065, 'clip_ratio/low_min': 0.08572594448924065, 'clip_ratio/high_mean': 0.059783825650811195, 'clip_ratio/high_max': 0.059783825650811195, 'clip_ratio/region_mean': 0.14550977014005184, 'reward_total_mean': 0.4710928201675415, 'reward_meter_mean': 0.772925615310669, 'reward_meter_std': 0.4008168876171112, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9684585928916931, 'reward_repeat_penalty_std': 0.03413161262869835, 'reward_repeat_floor_mean': 0.9956202507019043, 'reward_repeat_floor_std': 0.0043912623077631, 'reward_near_duplicate_penalty_mean': 0.9747804403305054, 'reward_near_duplicate_penalty_std': 0.023137658834457397, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9933686852455139, 'reward_distinct_2_std': 0.01875614933669567, 'reward_judge_quality_mean': 0.6387499570846558, 'reward_judge_quality_std': 0.2639501094818115, 'reward_total_composite_mean': 0.4710928201675415, 'reward_total_composite_std': 0.3410051465034485, 'epoch': 0.08}

 43%|████▎     | 1431/3300 [5:29:00<6:26:37, 12.41s/it]INFO 04-12 21:27:27 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1432/3300 [5:29:14<6:41:18, 12.89s/it]2026-04-12 21:27:40,821 | INFO | train_grpo_train | metrics_logged mode=train step=1432

                                                       
{'loss': -0.0843, 'grad_norm': 3.8338301181793213, 'learning_rate': 5.663636363636364e-06, 'num_tokens': 2534201.0, 'completions/mean_length': 116.375, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.857147216796875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9751989841461182, 'rewards/meter/std': 0.027916021645069122, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.9547233581542969, 'rewards/lexical_plausibility/std': 0.12806160748004913, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.14880475401878357, 'rewards/repeat_penalty/mean': 0.893029510974884, 'rewards/repeat_penalty/std': 0.08625410497188568, 'rewards/repeat_floor/mean': 0.9868086576461792, 'rewards/repeat_floor/std': 0.009888304397463799, 'rewards/near_duplicate_penalty/mean': 0.9464174509048462, 'rewards/near_duplicate_penalty/std': 0.03807366266846657, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9427334070205688, 'rewards/distinct_2/std': 0.07041586190462112, 'rewards/total_composite/mean': 0.2611513137817383, 'rewards/total_composite/std': 0.14535030722618103, 'reward': 0.2611513137817383, 'reward_std': 0.14535030722618103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14724820852279663, 'sampling/sampling_logp_difference/max': 1.667689323425293, 'sampling/importance_sampling_ratio/min': 0.18868255615234375, 'sampling/importance_sampling_ratio/mean': 1.0426467657089233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1246240213513374, 'clip_ratio/low_mean': 0.044358803890645504, 'clip_ratio/low_min': 0.044358803890645504, 'clip_ratio/high_mean': 0.06831552786752582, 'clip_ratio/high_max': 0.06831552786752582, 'clip_ratio/region_mean': 0.11267433175817132, 'reward_total_mean': 0.2611513137817383, 'reward_meter_mean': 0.9751989841461182, 'reward_meter_std': 0.027916021645069122, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.9547233581542969, 'reward_lexical_plausibility_std': 0.12806160748004913, 'reward_repeat_penalty_mean': 0.893029510974884, 'reward_repeat_penalty_std': 0.08625410497188568, 'reward_repeat_floor_mean': 0.9868086576461792, 'reward_repeat_floor_std': 0.009888304397463799, 'reward_near_duplicate_penalty_mean': 0.9464174509048462, 'reward_near_duplicate_penalty_std': 0.03807366266846657, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9427334070205688, 'reward_distinct_2_std': 0.07041586190462112, 'reward_judge_quality_mean': 0.2750000059604645, 'reward_judge_quality_std': 0.14880475401878357, 'reward_total_composite_mean': 0.2611513137817383, 'reward_total_composite_std': 0.14535030722618103, 'epoch': 0.08}

 43%|████▎     | 1432/3300 [5:29:14<6:41:18, 12.89s/it]INFO 04-12 21:27:41 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1433/3300 [5:29:26<6:36:20, 12.74s/it]2026-04-12 21:27:53,211 | INFO | train_grpo_train | metrics_logged mode=train step=1433

                                                       
{'loss': 0.0155, 'grad_norm': 10.815335273742676, 'learning_rate': 5.6606060606060606e-06, 'num_tokens': 2535821.0, 'completions/mean_length': 47.5, 'completions/min_length': 44.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8035632371902466, 'rewards/meter/std': 0.35357198119163513, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.925196647644043, 'rewards/repeat_penalty/std': 0.09327761828899384, 'rewards/repeat_floor/mean': 0.9913779497146606, 'rewards/repeat_floor/std': 0.011129924096167088, 'rewards/near_duplicate_penalty/mean': 0.97325599193573, 'rewards/near_duplicate_penalty/std': 0.03835231438279152, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9487727284431458, 'rewards/distinct_2/std': 0.06601634621620178, 'rewards/total_composite/mean': 0.559435248374939, 'rewards/total_composite/std': 0.31340456008911133, 'reward': 0.559435248374939, 'reward_std': 0.3134045898914337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12842796742916107, 'sampling/sampling_logp_difference/max': 1.2949838638305664, 'sampling/importance_sampling_ratio/min': 0.2739022970199585, 'sampling/importance_sampling_ratio/mean': 1.0260000228881836, 'sampling/importance_sampling_ratio/max': 1.9964464902877808, 'entropy': 0.9089308679103851, 'clip_ratio/low_mean': 0.09015991259366274, 'clip_ratio/low_min': 0.09015991259366274, 'clip_ratio/high_mean': 0.029072696343064308, 'clip_ratio/high_max': 0.029072696343064308, 'clip_ratio/region_mean': 0.11923260893672705, 'reward_total_mean': 0.559435248374939, 'reward_meter_mean': 0.8035632371902466, 'reward_meter_std': 0.35357198119163513, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.925196647644043, 'reward_repeat_penalty_std': 0.09327761828899384, 'reward_repeat_floor_mean': 0.9913779497146606, 'reward_repeat_floor_std': 0.011129924096167088, 'reward_near_duplicate_penalty_mean': 0.97325599193573, 'reward_near_duplicate_penalty_std': 0.03835231438279152, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9487727284431458, 'reward_distinct_2_std': 0.06601634621620178, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.559435248374939, 'reward_total_composite_std': 0.31340456008911133, 'epoch': 0.08}

 43%|████▎     | 1433/3300 [5:29:27<6:36:20, 12.74s/it]INFO 04-12 21:27:53 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1434/3300 [5:29:35<5:59:19, 11.55s/it]2026-04-12 21:28:02,011 | INFO | train_grpo_train | metrics_logged mode=train step=1434

                                                       
{'loss': 0.0235, 'grad_norm': 12.809391975402832, 'learning_rate': 5.657575757575759e-06, 'num_tokens': 2537327.0, 'completions/mean_length': 42.25, 'completions/min_length': 40.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6077009439468384, 'rewards/meter/std': 0.4388333261013031, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5674999952316284, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/repeat_penalty/mean': 0.9670984745025635, 'rewards/repeat_penalty/std': 0.05196331813931465, 'rewards/repeat_floor/mean': 0.9955167174339294, 'rewards/repeat_floor/std': 0.006778073497116566, 'rewards/near_duplicate_penalty/mean': 0.9756060242652893, 'rewards/near_duplicate_penalty/std': 0.03486240282654762, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.3359838128089905, 'rewards/total_composite/std': 0.2881816327571869, 'reward': 0.3359838128089905, 'reward_std': 0.2881816327571869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12209069728851318, 'sampling/sampling_logp_difference/max': 1.4627957344055176, 'sampling/importance_sampling_ratio/min': 0.23158791661262512, 'sampling/importance_sampling_ratio/mean': 1.0027520656585693, 'sampling/importance_sampling_ratio/max': 1.8436205387115479, 'entropy': 0.7831537052989006, 'clip_ratio/low_mean': 0.047203767113387585, 'clip_ratio/low_min': 0.047203767113387585, 'clip_ratio/high_mean': 0.05719283316284418, 'clip_ratio/high_max': 0.05719283316284418, 'clip_ratio/region_mean': 0.10439660027623177, 'reward_total_mean': 0.3359838128089905, 'reward_meter_mean': 0.6077009439468384, 'reward_meter_std': 0.4388333261013031, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9670984745025635, 'reward_repeat_penalty_std': 0.05196331813931465, 'reward_repeat_floor_mean': 0.9955167174339294, 'reward_repeat_floor_std': 0.006778073497116566, 'reward_near_duplicate_penalty_mean': 0.9756060242652893, 'reward_near_duplicate_penalty_std': 0.03486240282654762, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.5674999952316284, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.3359838128089905, 'reward_total_composite_std': 0.2881816327571869, 'epoch': 0.08}

 43%|████▎     | 1434/3300 [5:29:35<5:59:19, 11.55s/it]INFO 04-12 21:28:02 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1435/3300 [5:29:45<5:41:01, 10.97s/it]2026-04-12 21:28:11,684 | INFO | train_grpo_train | metrics_logged mode=train step=1435

                                                       
{'loss': -0.0376, 'grad_norm': 11.316720962524414, 'learning_rate': 5.654545454545455e-06, 'num_tokens': 2539044.0, 'completions/mean_length': 42.625, 'completions/min_length': 34.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7937230467796326, 'rewards/meter/std': 0.2995172142982483, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/repeat_penalty/mean': 0.9687439799308777, 'rewards/repeat_penalty/std': 0.02794233150780201, 'rewards/repeat_floor/mean': 0.9953116178512573, 'rewards/repeat_floor/std': 0.004191352054476738, 'rewards/near_duplicate_penalty/mean': 0.9687439799308777, 'rewards/near_duplicate_penalty/std': 0.02794233150780201, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.33091306686401367, 'rewards/total_composite/std': 0.12386972457170486, 'reward': 0.33091306686401367, 'reward_std': 0.12386971712112427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15386852622032166, 'sampling/sampling_logp_difference/max': 1.954709768295288, 'sampling/importance_sampling_ratio/min': 0.14160557091236115, 'sampling/importance_sampling_ratio/mean': 1.0349780321121216, 'sampling/importance_sampling_ratio/max': 1.8183578252792358, 'entropy': 1.1979315429925919, 'clip_ratio/low_mean': 0.020582706667482853, 'clip_ratio/low_min': 0.020582706667482853, 'clip_ratio/high_mean': 0.11941523663699627, 'clip_ratio/high_max': 0.11941523663699627, 'clip_ratio/region_mean': 0.13999794330447912, 'reward_total_mean': 0.33091306686401367, 'reward_meter_mean': 0.7937230467796326, 'reward_meter_std': 0.2995172142982483, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9687439799308777, 'reward_repeat_penalty_std': 0.02794233150780201, 'reward_repeat_floor_mean': 0.9953116178512573, 'reward_repeat_floor_std': 0.004191352054476738, 'reward_near_duplicate_penalty_mean': 0.9687439799308777, 'reward_near_duplicate_penalty_std': 0.02794233150780201, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.33091306686401367, 'reward_total_composite_std': 0.12386972457170486, 'epoch': 0.08}

 43%|████▎     | 1435/3300 [5:29:45<5:41:01, 10.97s/it]INFO 04-12 21:28:11 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1436/3300 [5:30:00<6:16:28, 12.12s/it]2026-04-12 21:28:26,414 | INFO | train_grpo_train | metrics_logged mode=train step=1436

                                                       
{'loss': -0.1499, 'grad_norm': 2.8543624877929688, 'learning_rate': 5.651515151515152e-06, 'num_tokens': 2540968.0, 'completions/mean_length': 253.5, 'completions/min_length': 88.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 98.4000015258789, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.8084597587585449, 'rewards/meter/std': 0.3178073465824127, 'rewards/count_adherence/mean': 0.5357142686843872, 'rewards/count_adherence/std': 0.12662933766841888, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.860714316368103, 'rewards/count_floor/std': 0.03798879683017731, 'rewards/lexical_plausibility/mean': 0.8361128568649292, 'rewards/lexical_plausibility/std': 0.2306813895702362, 'rewards/judge_quality/mean': 0.20624999701976776, 'rewards/judge_quality/std': 0.1590990275144577, 'rewards/repeat_penalty/mean': 0.9289405345916748, 'rewards/repeat_penalty/std': 0.053262174129486084, 'rewards/repeat_floor/mean': 0.9920302629470825, 'rewards/repeat_floor/std': 0.005751530174165964, 'rewards/near_duplicate_penalty/mean': 0.9761865139007568, 'rewards/near_duplicate_penalty/std': 0.015039334073662758, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9511362314224243, 'rewards/distinct_2/std': 0.04154196381568909, 'rewards/total_composite/mean': 0.16314910352230072, 'rewards/total_composite/std': 0.13803774118423462, 'reward': 0.16314910352230072, 'reward_std': 0.13803774118423462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20392070710659027, 'sampling/sampling_logp_difference/max': 2.021611213684082, 'sampling/importance_sampling_ratio/min': 0.132441908121109, 'sampling/importance_sampling_ratio/mean': 1.0531235933303833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1845258176326752, 'clip_ratio/low_mean': 0.01315789483487606, 'clip_ratio/low_min': 0.01315789483487606, 'clip_ratio/high_mean': 0.07894021738320589, 'clip_ratio/high_max': 0.07894021738320589, 'clip_ratio/region_mean': 0.09209811221808195, 'reward_total_mean': 0.16314910352230072, 'reward_meter_mean': 0.8084597587585449, 'reward_meter_std': 0.3178073465824127, 'reward_count_adherence_mean': 0.5357142686843872, 'reward_count_adherence_std': 0.12662933766841888, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.860714316368103, 'reward_count_floor_std': 0.03798879683017731, 'reward_lexical_plausibility_mean': 0.8361128568649292, 'reward_lexical_plausibility_std': 0.2306813895702362, 'reward_repeat_penalty_mean': 0.9289405345916748, 'reward_repeat_penalty_std': 0.053262174129486084, 'reward_repeat_floor_mean': 0.9920302629470825, 'reward_repeat_floor_std': 0.005751530174165964, 'reward_near_duplicate_penalty_mean': 0.9761865139007568, 'reward_near_duplicate_penalty_std': 0.015039334073662758, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9511362314224243, 'reward_distinct_2_std': 0.04154196381568909, 'reward_judge_quality_mean': 0.20624999701976776, 'reward_judge_quality_std': 0.1590990275144577, 'reward_total_composite_mean': 0.16314910352230072, 'reward_total_composite_std': 0.13803774118423462, 'epoch': 0.08}

 44%|████▎     | 1436/3300 [5:30:00<6:16:28, 12.12s/it]INFO 04-12 21:28:26 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1437/3300 [5:30:09<5:48:36, 11.23s/it]2026-04-12 21:28:35,554 | INFO | train_grpo_train | metrics_logged mode=train step=1437

                                                       
{'loss': 0.0239, 'grad_norm': 11.36780071258545, 'learning_rate': 5.648484848484849e-06, 'num_tokens': 2542758.0, 'completions/mean_length': 58.75, 'completions/min_length': 51.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.852945864200592, 'rewards/meter/std': 0.28729549050331116, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.10690448433160782, 'rewards/repeat_penalty/mean': 0.9559943675994873, 'rewards/repeat_penalty/std': 0.07664373517036438, 'rewards/repeat_floor/mean': 0.9953839778900146, 'rewards/repeat_floor/std': 0.007098164409399033, 'rewards/near_duplicate_penalty/mean': 0.9863051176071167, 'rewards/near_duplicate_penalty/std': 0.014090240933001041, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9798692464828491, 'rewards/distinct_2/std': 0.0397031269967556, 'rewards/total_composite/mean': 0.3271406292915344, 'rewards/total_composite/std': 0.14686590433120728, 'reward': 0.3271406292915344, 'reward_std': 0.14686588943004608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14686284959316254, 'sampling/sampling_logp_difference/max': 1.56541109085083, 'sampling/importance_sampling_ratio/min': 0.2090020626783371, 'sampling/importance_sampling_ratio/mean': 1.029215693473816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0047723203897476, 'clip_ratio/low_mean': 0.029373468831181526, 'clip_ratio/low_min': 0.029373468831181526, 'clip_ratio/high_mean': 0.11431877501308918, 'clip_ratio/high_max': 0.11431877501308918, 'clip_ratio/region_mean': 0.1436922438442707, 'reward_total_mean': 0.3271406292915344, 'reward_meter_mean': 0.852945864200592, 'reward_meter_std': 0.28729549050331116, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9559943675994873, 'reward_repeat_penalty_std': 0.07664373517036438, 'reward_repeat_floor_mean': 0.9953839778900146, 'reward_repeat_floor_std': 0.007098164409399033, 'reward_near_duplicate_penalty_mean': 0.9863051176071167, 'reward_near_duplicate_penalty_std': 0.014090240933001041, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9798692464828491, 'reward_distinct_2_std': 0.0397031269967556, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.10690448433160782, 'reward_total_composite_mean': 0.3271406292915344, 'reward_total_composite_std': 0.14686590433120728, 'epoch': 0.08}

 44%|████▎     | 1437/3300 [5:30:09<5:48:36, 11.23s/it]INFO 04-12 21:28:35 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1438/3300 [5:30:23<6:20:01, 12.25s/it]2026-04-12 21:28:50,179 | INFO | train_grpo_train | metrics_logged mode=train step=1438

                                                       
{'loss': -0.1688, 'grad_norm': 2.5985569953918457, 'learning_rate': 5.645454545454546e-06, 'num_tokens': 2545097.0, 'completions/mean_length': 267.375, 'completions/min_length': 108.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 120.5999984741211, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.8278605341911316, 'rewards/meter/std': 0.2633938491344452, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.06681530922651291, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.020044591277837753, 'rewards/lexical_plausibility/mean': 0.8392350673675537, 'rewards/lexical_plausibility/std': 0.22488006949424744, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9627857208251953, 'rewards/repeat_penalty/std': 0.035533227026462555, 'rewards/repeat_floor/mean': 0.9952646493911743, 'rewards/repeat_floor/std': 0.0038585616275668144, 'rewards/near_duplicate_penalty/mean': 0.9776012897491455, 'rewards/near_duplicate_penalty/std': 0.013215497136116028, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9847158193588257, 'rewards/distinct_2/std': 0.029535382986068726, 'rewards/total_composite/mean': 0.18558116257190704, 'rewards/total_composite/std': 0.12809965014457703, 'reward': 0.18558116257190704, 'reward_std': 0.12809965014457703, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16897639632225037, 'sampling/sampling_logp_difference/max': 3.12882661819458, 'sampling/importance_sampling_ratio/min': 0.043769124895334244, 'sampling/importance_sampling_ratio/mean': 1.0234577655792236, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8054457157850266, 'clip_ratio/low_mean': 0.02187499962747097, 'clip_ratio/low_min': 0.02187499962747097, 'clip_ratio/high_mean': 0.07969456352293491, 'clip_ratio/high_max': 0.07969456352293491, 'clip_ratio/region_mean': 0.10156956315040588, 'reward_total_mean': 0.18558116257190704, 'reward_meter_mean': 0.8278605341911316, 'reward_meter_std': 0.2633938491344452, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.06681530922651291, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.020044591277837753, 'reward_lexical_plausibility_mean': 0.8392350673675537, 'reward_lexical_plausibility_std': 0.22488006949424744, 'reward_repeat_penalty_mean': 0.9627857208251953, 'reward_repeat_penalty_std': 0.035533227026462555, 'reward_repeat_floor_mean': 0.9952646493911743, 'reward_repeat_floor_std': 0.0038585616275668144, 'reward_near_duplicate_penalty_mean': 0.9776012897491455, 'reward_near_duplicate_penalty_std': 0.013215497136116028, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9847158193588257, 'reward_distinct_2_std': 0.029535382986068726, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.18558116257190704, 'reward_total_composite_std': 0.12809965014457703, 'epoch': 0.08}

 44%|████▎     | 1438/3300 [5:30:24<6:20:01, 12.25s/it]INFO 04-12 21:28:50 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1439/3300 [5:30:32<5:44:43, 11.11s/it]2026-04-12 21:28:58,653 | INFO | train_grpo_train | metrics_logged mode=train step=1439

                                                       
{'loss': 0.0401, 'grad_norm': 15.01860237121582, 'learning_rate': 5.642424242424242e-06, 'num_tokens': 2546841.0, 'completions/mean_length': 38.0, 'completions/min_length': 35.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9474092721939087, 'rewards/meter/std': 0.10505027323961258, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6212500333786011, 'rewards/judge_quality/std': 0.19134393334388733, 'rewards/repeat_penalty/mean': 0.8722543716430664, 'rewards/repeat_penalty/std': 0.16646598279476166, 'rewards/repeat_floor/mean': 0.9894043803215027, 'rewards/repeat_floor/std': 0.013198469765484333, 'rewards/near_duplicate_penalty/mean': 0.9841701984405518, 'rewards/near_duplicate_penalty/std': 0.015411601401865482, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9711538553237915, 'rewards/distinct_2/std': 0.039811473339796066, 'rewards/total_composite/mean': 0.5893003344535828, 'rewards/total_composite/std': 0.20974105596542358, 'reward': 0.5893003344535828, 'reward_std': 0.20974107086658478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1472310572862625, 'sampling/sampling_logp_difference/max': 1.385605812072754, 'sampling/importance_sampling_ratio/min': 0.2501721978187561, 'sampling/importance_sampling_ratio/mean': 1.0171046257019043, 'sampling/importance_sampling_ratio/max': 1.8697993755340576, 'entropy': 1.0474810674786568, 'clip_ratio/low_mean': 0.05385982058942318, 'clip_ratio/low_min': 0.05385982058942318, 'clip_ratio/high_mean': 0.08811751566827297, 'clip_ratio/high_max': 0.08811751566827297, 'clip_ratio/region_mean': 0.14197733625769615, 'reward_total_mean': 0.5893003344535828, 'reward_meter_mean': 0.9474092721939087, 'reward_meter_std': 0.10505027323961258, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8722543716430664, 'reward_repeat_penalty_std': 0.16646598279476166, 'reward_repeat_floor_mean': 0.9894043803215027, 'reward_repeat_floor_std': 0.013198469765484333, 'reward_near_duplicate_penalty_mean': 0.9841701984405518, 'reward_near_duplicate_penalty_std': 0.015411601401865482, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9711538553237915, 'reward_distinct_2_std': 0.039811473339796066, 'reward_judge_quality_mean': 0.6212500333786011, 'reward_judge_quality_std': 0.19134393334388733, 'reward_total_composite_mean': 0.5893003344535828, 'reward_total_composite_std': 0.20974105596542358, 'epoch': 0.08}

 44%|████▎     | 1439/3300 [5:30:32<5:44:43, 11.11s/it]INFO 04-12 21:28:58 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1440/3300 [5:30:42<5:33:45, 10.77s/it]2026-04-12 21:29:08,611 | INFO | train_grpo_train | metrics_logged mode=train step=1440

                                                       
{'loss': 0.069, 'grad_norm': 12.151432991027832, 'learning_rate': 5.6393939393939405e-06, 'num_tokens': 2548666.0, 'completions/mean_length': 62.125, 'completions/min_length': 58.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.6526681184768677, 'rewards/meter/std': 0.29940924048423767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9337679147720337, 'rewards/repeat_penalty/std': 0.046768590807914734, 'rewards/repeat_floor/mean': 0.992020845413208, 'rewards/repeat_floor/std': 0.005467830691486597, 'rewards/near_duplicate_penalty/mean': 0.9685639142990112, 'rewards/near_duplicate_penalty/std': 0.02111426740884781, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9637367129325867, 'rewards/distinct_2/std': 0.03472144901752472, 'rewards/total_composite/mean': 0.22972485423088074, 'rewards/total_composite/std': 0.09843901544809341, 'reward': 0.22972485423088074, 'reward_std': 0.09843901544809341, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15524837374687195, 'sampling/sampling_logp_difference/max': 1.846414566040039, 'sampling/importance_sampling_ratio/min': 0.15780194103717804, 'sampling/importance_sampling_ratio/mean': 1.0249295234680176, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8301663920283318, 'clip_ratio/low_mean': 0.0634230524301529, 'clip_ratio/low_min': 0.0634230524301529, 'clip_ratio/high_mean': 0.07499332632869482, 'clip_ratio/high_max': 0.07499332632869482, 'clip_ratio/region_mean': 0.1384163787588477, 'reward_total_mean': 0.22972485423088074, 'reward_meter_mean': 0.6526681184768677, 'reward_meter_std': 0.29940924048423767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9337679147720337, 'reward_repeat_penalty_std': 0.046768590807914734, 'reward_repeat_floor_mean': 0.992020845413208, 'reward_repeat_floor_std': 0.005467830691486597, 'reward_near_duplicate_penalty_mean': 0.9685639142990112, 'reward_near_duplicate_penalty_std': 0.02111426740884781, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9637367129325867, 'reward_distinct_2_std': 0.03472144901752472, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.22972485423088074, 'reward_total_composite_std': 0.09843901544809341, 'epoch': 0.08}

 44%|████▎     | 1440/3300 [5:30:42<5:33:45, 10.77s/it]INFO 04-12 21:29:08 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1441/3300 [5:30:55<5:59:58, 11.62s/it]2026-04-12 21:29:22,215 | INFO | train_grpo_train | metrics_logged mode=train step=1441

                                                       
{'loss': -0.0184, 'grad_norm': 5.862575531005859, 'learning_rate': 5.636363636363636e-06, 'num_tokens': 2550185.0, 'completions/mean_length': 90.875, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 30.71428680419922, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.4679519534111023, 'rewards/meter/std': 0.40002721548080444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9398874044418335, 'rewards/lexical_plausibility/std': 0.170024111866951, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.16690458357334137, 'rewards/repeat_penalty/mean': 0.9132769107818604, 'rewards/repeat_penalty/std': 0.1240047886967659, 'rewards/repeat_floor/mean': 0.9895557761192322, 'rewards/repeat_floor/std': 0.01305301021784544, 'rewards/near_duplicate_penalty/mean': 0.9486858248710632, 'rewards/near_duplicate_penalty/std': 0.054175253957509995, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9903097748756409, 'rewards/distinct_2/std': 0.025729352608323097, 'rewards/total_composite/mean': 0.20433664321899414, 'rewards/total_composite/std': 0.22805427014827728, 'reward': 0.20433664321899414, 'reward_std': 0.22805428504943848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16726423799991608, 'sampling/sampling_logp_difference/max': 1.3706603050231934, 'sampling/importance_sampling_ratio/min': 0.2539392113685608, 'sampling/importance_sampling_ratio/mean': 1.0363671779632568, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.016862913966179, 'clip_ratio/low_mean': 0.042008758056908846, 'clip_ratio/low_min': 0.042008758056908846, 'clip_ratio/high_mean': 0.06507230084389448, 'clip_ratio/high_max': 0.06507230084389448, 'clip_ratio/region_mean': 0.10708105890080333, 'reward_total_mean': 0.20433664321899414, 'reward_meter_mean': 0.4679519534111023, 'reward_meter_std': 0.40002721548080444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9398874044418335, 'reward_lexical_plausibility_std': 0.170024111866951, 'reward_repeat_penalty_mean': 0.9132769107818604, 'reward_repeat_penalty_std': 0.1240047886967659, 'reward_repeat_floor_mean': 0.9895557761192322, 'reward_repeat_floor_std': 0.01305301021784544, 'reward_near_duplicate_penalty_mean': 0.9486858248710632, 'reward_near_duplicate_penalty_std': 0.054175253957509995, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9903097748756409, 'reward_distinct_2_std': 0.025729352608323097, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.16690458357334137, 'reward_total_composite_mean': 0.20433664321899414, 'reward_total_composite_std': 0.22805427014827728, 'epoch': 0.08}

 44%|████▎     | 1441/3300 [5:30:56<5:59:58, 11.62s/it]INFO 04-12 21:29:22 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1442/3300 [5:31:11<6:34:14, 12.73s/it]2026-04-12 21:29:37,549 | INFO | train_grpo_train | metrics_logged mode=train step=1442

                                                       
{'loss': -0.0929, 'grad_norm': 1.5667822360992432, 'learning_rate': 5.633333333333334e-06, 'num_tokens': 2551771.0, 'completions/mean_length': 221.25, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 46.79999923706055, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8781077861785889, 'rewards/meter/std': 0.23861871659755707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.827507495880127, 'rewards/lexical_plausibility/std': 0.2392042726278305, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.19955307245254517, 'rewards/repeat_penalty/mean': 0.9672141075134277, 'rewards/repeat_penalty/std': 0.049560315907001495, 'rewards/repeat_floor/mean': 0.9958791732788086, 'rewards/repeat_floor/std': 0.006192892324179411, 'rewards/near_duplicate_penalty/mean': 0.9816303253173828, 'rewards/near_duplicate_penalty/std': 0.028096236288547516, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9848290085792542, 'rewards/distinct_2/std': 0.026992013677954674, 'rewards/total_composite/mean': 0.2753852903842926, 'rewards/total_composite/std': 0.20176035165786743, 'reward': 0.2753852903842926, 'reward_std': 0.20176035165786743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16209512948989868, 'sampling/sampling_logp_difference/max': 1.6074860095977783, 'sampling/importance_sampling_ratio/min': 0.2003907561302185, 'sampling/importance_sampling_ratio/mean': 1.0195982456207275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7363856211304665, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08630292490124702, 'clip_ratio/high_max': 0.08630292490124702, 'clip_ratio/region_mean': 0.08630292490124702, 'reward_total_mean': 0.2753852903842926, 'reward_meter_mean': 0.8781077861785889, 'reward_meter_std': 0.23861871659755707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.827507495880127, 'reward_lexical_plausibility_std': 0.2392042726278305, 'reward_repeat_penalty_mean': 0.9672141075134277, 'reward_repeat_penalty_std': 0.049560315907001495, 'reward_repeat_floor_mean': 0.9958791732788086, 'reward_repeat_floor_std': 0.006192892324179411, 'reward_near_duplicate_penalty_mean': 0.9816303253173828, 'reward_near_duplicate_penalty_std': 0.028096236288547516, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9848290085792542, 'reward_distinct_2_std': 0.026992013677954674, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.19955307245254517, 'reward_total_composite_mean': 0.2753852903842926, 'reward_total_composite_std': 0.20176035165786743, 'epoch': 0.08}

 44%|████▎     | 1442/3300 [5:31:11<6:34:14, 12.73s/it]INFO 04-12 21:29:37 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1443/3300 [5:31:20<6:04:05, 11.76s/it]2026-04-12 21:29:47,050 | INFO | train_grpo_train | metrics_logged mode=train step=1443

                                                       
{'loss': 0.0226, 'grad_norm': 18.40782928466797, 'learning_rate': 5.630303030303031e-06, 'num_tokens': 2553212.0, 'completions/mean_length': 31.125, 'completions/min_length': 19.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6279726028442383, 'rewards/meter/std': 0.43741652369499207, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.16035674512386322, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.8512497544288635, 'rewards/repeat_penalty/std': 0.14278069138526917, 'rewards/repeat_floor/mean': 0.9887473583221436, 'rewards/repeat_floor/std': 0.010075317695736885, 'rewards/near_duplicate_penalty/mean': 0.9810548424720764, 'rewards/near_duplicate_penalty/std': 0.01773335039615631, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.15707695484161377, 'rewards/total_composite/std': 0.14221549034118652, 'reward': 0.15707695484161377, 'reward_std': 0.14221547544002533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14728401601314545, 'sampling/sampling_logp_difference/max': 1.8133232593536377, 'sampling/importance_sampling_ratio/min': 0.16311118006706238, 'sampling/importance_sampling_ratio/mean': 1.0340102910995483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.921648733317852, 'clip_ratio/low_mean': 0.07453275099396706, 'clip_ratio/low_min': 0.07453275099396706, 'clip_ratio/high_mean': 0.056966845877468586, 'clip_ratio/high_max': 0.056966845877468586, 'clip_ratio/region_mean': 0.13149959687143564, 'reward_total_mean': 0.15707695484161377, 'reward_meter_mean': 0.6279726028442383, 'reward_meter_std': 0.43741652369499207, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.16035674512386322, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8512497544288635, 'reward_repeat_penalty_std': 0.14278069138526917, 'reward_repeat_floor_mean': 0.9887473583221436, 'reward_repeat_floor_std': 0.010075317695736885, 'reward_near_duplicate_penalty_mean': 0.9810548424720764, 'reward_near_duplicate_penalty_std': 0.01773335039615631, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.15707695484161377, 'reward_total_composite_std': 0.14221549034118652, 'epoch': 0.08}

 44%|████▎     | 1443/3300 [5:31:20<6:04:05, 11.76s/it]INFO 04-12 21:29:47 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1444/3300 [5:31:34<6:23:14, 12.39s/it]2026-04-12 21:30:00,914 | INFO | train_grpo_train | metrics_logged mode=train step=1444

                                                       
{'loss': -0.0213, 'grad_norm': 5.70648717880249, 'learning_rate': 5.627272727272728e-06, 'num_tokens': 2554822.0, 'completions/mean_length': 93.25, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.42857360839844, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.12548930943012238, 'rewards/meter/std': 0.16755937039852142, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9460666179656982, 'rewards/lexical_plausibility/std': 0.1525467336177826, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.35087648034095764, 'rewards/repeat_penalty/mean': 0.7169134616851807, 'rewards/repeat_penalty/std': 0.2304912954568863, 'rewards/repeat_floor/mean': 0.9687427282333374, 'rewards/repeat_floor/std': 0.02760075405240059, 'rewards/near_duplicate_penalty/mean': 0.9079367518424988, 'rewards/near_duplicate_penalty/std': 0.07836468517780304, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9103026390075684, 'rewards/distinct_2/std': 0.11174733191728592, 'rewards/total_composite/mean': 0.103962741792202, 'rewards/total_composite/std': 0.15143881738185883, 'reward': 0.103962741792202, 'reward_std': 0.15143880248069763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10993009060621262, 'sampling/sampling_logp_difference/max': 0.8074464797973633, 'sampling/importance_sampling_ratio/min': 0.4459955096244812, 'sampling/importance_sampling_ratio/mean': 1.0364493131637573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.621724434196949, 'clip_ratio/low_mean': 0.06739372527226806, 'clip_ratio/low_min': 0.06739372527226806, 'clip_ratio/high_mean': 0.01481729093939066, 'clip_ratio/high_max': 0.01481729093939066, 'clip_ratio/region_mean': 0.08221101621165872, 'reward_total_mean': 0.103962741792202, 'reward_meter_mean': 0.12548930943012238, 'reward_meter_std': 0.16755937039852142, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9460666179656982, 'reward_lexical_plausibility_std': 0.1525467336177826, 'reward_repeat_penalty_mean': 0.7169134616851807, 'reward_repeat_penalty_std': 0.2304912954568863, 'reward_repeat_floor_mean': 0.9687427282333374, 'reward_repeat_floor_std': 0.02760075405240059, 'reward_near_duplicate_penalty_mean': 0.9079367518424988, 'reward_near_duplicate_penalty_std': 0.07836468517780304, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9103026390075684, 'reward_distinct_2_std': 0.11174733191728592, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.35087648034095764, 'reward_total_composite_mean': 0.103962741792202, 'reward_total_composite_std': 0.15143881738185883, 'epoch': 0.08}

 44%|████▍     | 1444/3300 [5:31:34<6:23:14, 12.39s/it]INFO 04-12 21:30:01 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1445/3300 [5:31:48<6:38:34, 12.89s/it]2026-04-12 21:30:14,955 | INFO | train_grpo_train | metrics_logged mode=train step=1445

                                                       
{'loss': -0.0874, 'grad_norm': 2.906337261199951, 'learning_rate': 5.624242424242424e-06, 'num_tokens': 2556547.0, 'completions/mean_length': 98.625, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.57143020629883, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8600119352340698, 'rewards/meter/std': 0.34526491165161133, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9370453357696533, 'rewards/lexical_plausibility/std': 0.17806260287761688, 'rewards/judge_quality/mean': 0.42500001192092896, 'rewards/judge_quality/std': 0.19086270034313202, 'rewards/repeat_penalty/mean': 0.9552587270736694, 'rewards/repeat_penalty/std': 0.08494306355714798, 'rewards/repeat_floor/mean': 0.9961013197898865, 'rewards/repeat_floor/std': 0.005262878257781267, 'rewards/near_duplicate_penalty/mean': 0.9865087270736694, 'rewards/near_duplicate_penalty/std': 0.01914379559457302, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4105956554412842, 'rewards/total_composite/std': 0.2022566795349121, 'reward': 0.4105956554412842, 'reward_std': 0.2022566944360733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15023012459278107, 'sampling/sampling_logp_difference/max': 1.6220568418502808, 'sampling/importance_sampling_ratio/min': 0.1974920630455017, 'sampling/importance_sampling_ratio/mean': 1.0156630277633667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9702880829572678, 'clip_ratio/low_mean': 0.01923076994717121, 'clip_ratio/low_min': 0.01923076994717121, 'clip_ratio/high_mean': 0.12259170645847917, 'clip_ratio/high_max': 0.12259170645847917, 'clip_ratio/region_mean': 0.14182247640565038, 'reward_total_mean': 0.4105956554412842, 'reward_meter_mean': 0.8600119352340698, 'reward_meter_std': 0.34526491165161133, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9370453357696533, 'reward_lexical_plausibility_std': 0.17806260287761688, 'reward_repeat_penalty_mean': 0.9552587270736694, 'reward_repeat_penalty_std': 0.08494306355714798, 'reward_repeat_floor_mean': 0.9961013197898865, 'reward_repeat_floor_std': 0.005262878257781267, 'reward_near_duplicate_penalty_mean': 0.9865087270736694, 'reward_near_duplicate_penalty_std': 0.01914379559457302, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42500001192092896, 'reward_judge_quality_std': 0.19086270034313202, 'reward_total_composite_mean': 0.4105956554412842, 'reward_total_composite_std': 0.2022566795349121, 'epoch': 0.08}

 44%|████▍     | 1445/3300 [5:31:48<6:38:34, 12.89s/it]INFO 04-12 21:30:15 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1446/3300 [5:31:57<6:04:59, 11.81s/it]2026-04-12 21:30:24,254 | INFO | train_grpo_train | metrics_logged mode=train step=1446

                                                       
{'loss': 0.0723, 'grad_norm': 10.862899780273438, 'learning_rate': 5.6212121212121215e-06, 'num_tokens': 2558358.0, 'completions/mean_length': 68.375, 'completions/min_length': 62.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.6186676025390625, 'rewards/meter/std': 0.28509220480918884, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.14078859984874725, 'rewards/repeat_penalty/mean': 0.9670566916465759, 'rewards/repeat_penalty/std': 0.05191989988088608, 'rewards/repeat_floor/mean': 0.9961279630661011, 'rewards/repeat_floor/std': 0.005744150374084711, 'rewards/near_duplicate_penalty/mean': 0.9859901666641235, 'rewards/near_duplicate_penalty/std': 0.016596216708421707, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9803271293640137, 'rewards/distinct_2/std': 0.039262957870960236, 'rewards/total_composite/mean': 0.29716455936431885, 'rewards/total_composite/std': 0.19716593623161316, 'reward': 0.29716455936431885, 'reward_std': 0.19716592133045197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10598771274089813, 'sampling/sampling_logp_difference/max': 1.25407075881958, 'sampling/importance_sampling_ratio/min': 0.2853408753871918, 'sampling/importance_sampling_ratio/mean': 1.0259636640548706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8285898119211197, 'clip_ratio/low_mean': 0.06666329968720675, 'clip_ratio/low_min': 0.06666329968720675, 'clip_ratio/high_mean': 0.033678301610052586, 'clip_ratio/high_max': 0.033678301610052586, 'clip_ratio/region_mean': 0.10034160129725933, 'reward_total_mean': 0.29716455936431885, 'reward_meter_mean': 0.6186676025390625, 'reward_meter_std': 0.28509220480918884, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9670566916465759, 'reward_repeat_penalty_std': 0.05191989988088608, 'reward_repeat_floor_mean': 0.9961279630661011, 'reward_repeat_floor_std': 0.005744150374084711, 'reward_near_duplicate_penalty_mean': 0.9859901666641235, 'reward_near_duplicate_penalty_std': 0.016596216708421707, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9803271293640137, 'reward_distinct_2_std': 0.039262957870960236, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.14078859984874725, 'reward_total_composite_mean': 0.29716455936431885, 'reward_total_composite_std': 0.19716593623161316, 'epoch': 0.08}

 44%|████▍     | 1446/3300 [5:31:58<6:04:59, 11.81s/it]INFO 04-12 21:30:24 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1447/3300 [5:32:07<5:43:40, 11.13s/it]2026-04-12 21:30:33,860 | INFO | train_grpo_train | metrics_logged mode=train step=1447

                                                       
{'loss': 0.2793, 'grad_norm': 12.900527954101562, 'learning_rate': 5.618181818181818e-06, 'num_tokens': 2559791.0, 'completions/mean_length': 33.125, 'completions/min_length': 19.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9007810354232788, 'rewards/meter/std': 0.13912810385227203, 'rewards/count_adherence/mean': 0.375, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8125, 'rewards/count_floor/std': 0.15526476502418518, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.26868730783462524, 'rewards/repeat_penalty/mean': 0.8558162450790405, 'rewards/repeat_penalty/std': 0.14621999859809875, 'rewards/repeat_floor/mean': 0.9869076013565063, 'rewards/repeat_floor/std': 0.01457811426371336, 'rewards/near_duplicate_penalty/mean': 0.9607065916061401, 'rewards/near_duplicate_penalty/std': 0.06303125619888306, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9825174808502197, 'rewards/distinct_2/std': 0.049448031932115555, 'rewards/total_composite/mean': 0.42034912109375, 'rewards/total_composite/std': 0.30853331089019775, 'reward': 0.42034912109375, 'reward_std': 0.30853328108787537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14884044229984283, 'sampling/sampling_logp_difference/max': 2.026609420776367, 'sampling/importance_sampling_ratio/min': 0.13178159296512604, 'sampling/importance_sampling_ratio/mean': 1.0054724216461182, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8013309985399246, 'clip_ratio/low_mean': 0.07981733372434974, 'clip_ratio/low_min': 0.07981733372434974, 'clip_ratio/high_mean': 0.044407895766198635, 'clip_ratio/high_max': 0.044407895766198635, 'clip_ratio/region_mean': 0.12422522949054837, 'reward_total_mean': 0.42034912109375, 'reward_meter_mean': 0.9007810354232788, 'reward_meter_std': 0.13912810385227203, 'reward_count_adherence_mean': 0.375, 'reward_count_adherence_std': 0.5175492167472839, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8125, 'reward_count_floor_std': 0.15526476502418518, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8558162450790405, 'reward_repeat_penalty_std': 0.14621999859809875, 'reward_repeat_floor_mean': 0.9869076013565063, 'reward_repeat_floor_std': 0.01457811426371336, 'reward_near_duplicate_penalty_mean': 0.9607065916061401, 'reward_near_duplicate_penalty_std': 0.06303125619888306, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9825174808502197, 'reward_distinct_2_std': 0.049448031932115555, 'reward_judge_quality_mean': 0.5174999833106995, 'reward_judge_quality_std': 0.26868730783462524, 'reward_total_composite_mean': 0.42034912109375, 'reward_total_composite_std': 0.30853331089019775, 'epoch': 0.08}

 44%|████▍     | 1447/3300 [5:32:07<5:43:40, 11.13s/it]INFO 04-12 21:30:34 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1448/3300 [5:32:21<6:10:24, 12.00s/it]2026-04-12 21:30:47,828 | INFO | train_grpo_train | metrics_logged mode=train step=1448

                                                       
{'loss': -0.0362, 'grad_norm': 6.789714336395264, 'learning_rate': 5.615151515151516e-06, 'num_tokens': 2561280.0, 'completions/mean_length': 95.125, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.57143020629883, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.3045957684516907, 'rewards/meter/std': 0.21977011859416962, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9153863191604614, 'rewards/lexical_plausibility/std': 0.16365289688110352, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.2585087716579437, 'rewards/repeat_penalty/mean': 0.8437164425849915, 'rewards/repeat_penalty/std': 0.18094368278980255, 'rewards/repeat_floor/mean': 0.984000563621521, 'rewards/repeat_floor/std': 0.0171316210180521, 'rewards/near_duplicate_penalty/mean': 0.9471008777618408, 'rewards/near_duplicate_penalty/std': 0.04772946611046791, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9728937745094299, 'rewards/distinct_2/std': 0.04889722540974617, 'rewards/total_composite/mean': 0.12607242166996002, 'rewards/total_composite/std': 0.09509876370429993, 'reward': 0.12607242166996002, 'reward_std': 0.09509876370429993, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13479502499103546, 'sampling/sampling_logp_difference/max': 1.485246181488037, 'sampling/importance_sampling_ratio/min': 0.22644659876823425, 'sampling/importance_sampling_ratio/mean': 1.0138142108917236, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.735356479883194, 'clip_ratio/low_mean': 0.023489835672080517, 'clip_ratio/low_min': 0.023489835672080517, 'clip_ratio/high_mean': 0.043710211757570505, 'clip_ratio/high_max': 0.043710211757570505, 'clip_ratio/region_mean': 0.06720004742965102, 'reward_total_mean': 0.12607242166996002, 'reward_meter_mean': 0.3045957684516907, 'reward_meter_std': 0.21977011859416962, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9153863191604614, 'reward_lexical_plausibility_std': 0.16365289688110352, 'reward_repeat_penalty_mean': 0.8437164425849915, 'reward_repeat_penalty_std': 0.18094368278980255, 'reward_repeat_floor_mean': 0.984000563621521, 'reward_repeat_floor_std': 0.0171316210180521, 'reward_near_duplicate_penalty_mean': 0.9471008777618408, 'reward_near_duplicate_penalty_std': 0.04772946611046791, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9728937745094299, 'reward_distinct_2_std': 0.04889722540974617, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.2585087716579437, 'reward_total_composite_mean': 0.12607242166996002, 'reward_total_composite_std': 0.09509876370429993, 'epoch': 0.08}

 44%|████▍     | 1448/3300 [5:32:21<6:10:24, 12.00s/it]INFO 04-12 21:30:48 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1449/3300 [5:32:35<6:28:04, 12.58s/it]2026-04-12 21:31:01,755 | INFO | train_grpo_train | metrics_logged mode=train step=1449

                                                       
{'loss': -0.05, 'grad_norm': 6.012845039367676, 'learning_rate': 5.612121212121212e-06, 'num_tokens': 2562859.0, 'completions/mean_length': 94.375, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.71428680419922, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.5517988204956055, 'rewards/meter/std': 0.3980918228626251, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9247732162475586, 'rewards/lexical_plausibility/std': 0.17357861995697021, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.1552647352218628, 'rewards/repeat_penalty/mean': 0.9935404062271118, 'rewards/repeat_penalty/std': 0.0078064752742648125, 'rewards/repeat_floor/mean': 0.9990310668945312, 'rewards/repeat_floor/std': 0.0011709689861163497, 'rewards/near_duplicate_penalty/mean': 0.9935404062271118, 'rewards/near_duplicate_penalty/std': 0.0078064752742648125, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.18930363655090332, 'rewards/total_composite/std': 0.15464147925376892, 'reward': 0.18930363655090332, 'reward_std': 0.15464146435260773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15794429183006287, 'sampling/sampling_logp_difference/max': 1.5865812301635742, 'sampling/importance_sampling_ratio/min': 0.2046239674091339, 'sampling/importance_sampling_ratio/mean': 1.0130412578582764, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9310573041439056, 'clip_ratio/low_mean': 0.07002380397170782, 'clip_ratio/low_min': 0.07002380397170782, 'clip_ratio/high_mean': 0.03570431470870972, 'clip_ratio/high_max': 0.03570431470870972, 'clip_ratio/region_mean': 0.10572811868041754, 'reward_total_mean': 0.18930363655090332, 'reward_meter_mean': 0.5517988204956055, 'reward_meter_std': 0.3980918228626251, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9247732162475586, 'reward_lexical_plausibility_std': 0.17357861995697021, 'reward_repeat_penalty_mean': 0.9935404062271118, 'reward_repeat_penalty_std': 0.0078064752742648125, 'reward_repeat_floor_mean': 0.9990310668945312, 'reward_repeat_floor_std': 0.0011709689861163497, 'reward_near_duplicate_penalty_mean': 0.9935404062271118, 'reward_near_duplicate_penalty_std': 0.0078064752742648125, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.1552647352218628, 'reward_total_composite_mean': 0.18930363655090332, 'reward_total_composite_std': 0.15464147925376892, 'epoch': 0.08}

 44%|████▍     | 1449/3300 [5:32:35<6:28:04, 12.58s/it]INFO 04-12 21:31:02 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1450/3300 [5:32:44<5:59:14, 11.65s/it]2026-04-12 21:31:11,239 | INFO | train_grpo_train | metrics_logged mode=train step=1450

                                                       
{'loss': 0.1182, 'grad_norm': 13.827315330505371, 'learning_rate': 5.60909090909091e-06, 'num_tokens': 2564522.0, 'completions/mean_length': 50.875, 'completions/min_length': 43.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9243301749229431, 'rewards/meter/std': 0.06853026896715164, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9375, 'rewards/count_floor/std': 0.05175492912530899, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.30933108925819397, 'rewards/repeat_penalty/mean': 0.9664291143417358, 'rewards/repeat_penalty/std': 0.06967398524284363, 'rewards/repeat_floor/mean': 0.9959158897399902, 'rewards/repeat_floor/std': 0.007901218719780445, 'rewards/near_duplicate_penalty/mean': 0.9839391112327576, 'rewards/near_duplicate_penalty/std': 0.026036009192466736, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.981389582157135, 'rewards/distinct_2/std': 0.052638232707977295, 'rewards/total_composite/mean': 0.5376390814781189, 'rewards/total_composite/std': 0.24836792051792145, 'reward': 0.5376390814781189, 'reward_std': 0.24836792051792145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11453325301408768, 'sampling/sampling_logp_difference/max': 1.2618434429168701, 'sampling/importance_sampling_ratio/min': 0.28313159942626953, 'sampling/importance_sampling_ratio/mean': 1.0313094854354858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8031835108995438, 'clip_ratio/low_mean': 0.07638262957334518, 'clip_ratio/low_min': 0.07638262957334518, 'clip_ratio/high_mean': 0.0561265810392797, 'clip_ratio/high_max': 0.0561265810392797, 'clip_ratio/region_mean': 0.13250921061262488, 'reward_total_mean': 0.5376390814781189, 'reward_meter_mean': 0.9243301749229431, 'reward_meter_std': 0.06853026896715164, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9375, 'reward_count_floor_std': 0.05175492912530899, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9664291143417358, 'reward_repeat_penalty_std': 0.06967398524284363, 'reward_repeat_floor_mean': 0.9959158897399902, 'reward_repeat_floor_std': 0.007901218719780445, 'reward_near_duplicate_penalty_mean': 0.9839391112327576, 'reward_near_duplicate_penalty_std': 0.026036009192466736, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.981389582157135, 'reward_distinct_2_std': 0.052638232707977295, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.30933108925819397, 'reward_total_composite_mean': 0.5376390814781189, 'reward_total_composite_std': 0.24836792051792145, 'epoch': 0.08}

 44%|████▍     | 1450/3300 [5:32:45<5:59:14, 11.65s/it]INFO 04-12 21:31:11 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:42,  3.89s/it]

 23%|██▎       | 3/13 [00:13<00:46,  4.61s/it]

 31%|███       | 4/13 [00:21<00:52,  5.79s/it]

 38%|███▊      | 5/13 [00:29<00:53,  6.70s/it]

 46%|████▌     | 6/13 [00:37<00:49,  7.01s/it]

 54%|█████▍    | 7/13 [00:45<00:43,  7.29s/it]

 62%|██████▏   | 8/13 [00:56<00:43,  8.60s/it]

 69%|██████▉   | 9/13 [01:07<00:37,  9.33s/it]

 77%|███████▋  | 10/13 [01:19<00:30, 10.08s/it]

 85%|████████▍ | 11/13 [01:26<00:18,  9.26s/it]

 92%|█████████▏| 12/13 [01:33<00:08,  8.67s/it]

100%|██████████| 13/13 [01:45<00:00,  9.39s/it]2026-04-12 21:33:08,248 | INFO | train_grpo_train | metrics_logged mode=eval step=1450

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 116.8578, 'eval_samples_per_second': 0.89, 'eval_steps_per_second': 0.111, 'eval_num_tokens': 2564522.0, 'eval_completions/mean_length': 123.38461538461539, 'eval_completions/min_length': 36.07692307692308, 'eval_completions/max_length': 314.84615384615387, 'eval_completions/clipped_ratio': 0.057692307692307696, 'eval_completions/mean_terminated_length': 99.989012498122, 'eval_completions/min_terminated_length': 36.07692307692308, 'eval_completions/max_terminated_length': 193.53846153846155, 'eval_rewards/meter/mean': 0.4948628590657161, 'eval_rewards/meter/std': 0.3793173890847426, 'eval_rewards/count_adherence/mean': 0.7469225571705744, 'eval_rewards/count_adherence/std': 0.1673047777551871, 'eval_rewards/arabic_clean/mean': 0.9038461538461539, 'eval_rewards/arabic_clean/std': 0.21561105434711164, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/arabic_floor/mean': 0.9985576913906977, 'eval_rewards/arabic_floor/std': 0.004079461670838869, 'eval_rewards/count_floor/mean': 0.9240767772381122, 'eval_rewards/count_floor/std': 0.050191439115084134, 'eval_rewards/lexical_plausibility/mean': 0.974945774445167, 'eval_rewards/lexical_plausibility/std': 0.06377692921803547, 'eval_rewards/judge_quality/mean': 0.42788461309212905, 'eval_rewards/judge_quality/std': 0.1789561833899755, 'eval_rewards/repeat_penalty/mean': 0.8445180150178763, 'eval_rewards/repeat_penalty/std': 0.17082588661175507, 'eval_rewards/repeat_floor/mean': 0.9832773712965158, 'eval_rewards/repeat_floor/std': 0.018256844236300543, 'eval_rewards/near_duplicate_penalty/mean': 0.9598250480798575, 'eval_rewards/near_duplicate_penalty/std': 0.03725613632167761, 'eval_rewards/opening_diversity/mean': 0.9624566665062537, 'eval_rewards/opening_diversity/std': 0.08163573478276913, 'eval_rewards/distinct_2/mean': 0.9061803038303669, 'eval_rewards/distinct_2/std': 0.1165417289504638, 'eval_rewards/total_composite/mean': 0.20722378847690728, 'eval_rewards/total_composite/std': 0.21472871933992094, 'eval_reward': 0.20722378847690728, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07099351286888123, 'eval_sampling/sampling_logp_difference/max': 1.083869824042687, 'eval_sampling/importance_sampling_ratio/min': 0.3429002177256804, 'eval_sampling/importance_sampling_ratio/mean': 1.0191162091035109, 'eval_sampling/importance_sampling_ratio/max': 1.4892533467366145, 'eval_entropy': 0.8099038279973544, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.20722378847690728, 'eval_reward_meter_mean': 0.4948628590657161, 'eval_reward_meter_std': 0.3793173890847426, 'eval_reward_count_adherence_mean': 0.7469225571705744, 'eval_reward_count_adherence_std': 0.1673047777551871, 'eval_reward_arabic_clean_mean': 0.9038461538461539, 'eval_reward_arabic_clean_std': 0.21561105434711164, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_arabic_floor_mean': 0.9985576913906977, 'eval_reward_arabic_floor_std': 0.004079461670838869, 'eval_reward_count_floor_mean': 0.9240767772381122, 'eval_reward_count_floor_std': 0.050191439115084134, 'eval_reward_lexical_plausibility_mean': 0.974945774445167, 'eval_reward_lexical_plausibility_std': 0.06377692921803547, 'eval_reward_repeat_penalty_mean': 0.8445180150178763, 'eval_reward_repeat_penalty_std': 0.17082588661175507, 'eval_reward_repeat_floor_mean': 0.9832773712965158, 'eval_reward_repeat_floor_std': 0.018256844236300543, 'eval_reward_near_duplicate_penalty_mean': 0.9598250480798575, 'eval_reward_near_duplicate_penalty_std': 0.03725613632167761, 'eval_reward_opening_diversity_mean': 0.9624566665062537, 'eval_reward_opening_diversity_std': 0.08163573478276913, 'eval_reward_distinct_2_mean': 0.9061803038303669, 'eval_reward_distinct_2_std': 0.1165417289504638, 'eval_reward_judge_quality_mean': 0.42788461309212905, 'eval_reward_judge_quality_std': 0.1789561833899755, 'eval_reward_total_composite_mean': 0.20722378847690728, 'eval_reward_total_composite_std': 0.21472871933992094, 'epoch': 0.08}

 44%|████▍     | 1450/3300 [5:34:42<5:59:14, 11.65s/it]

100%|██████████| 13/13 [01:45<00:00,  9.39s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 44%|████▍     | 1451/3300 [5:34:53<24:01:45, 46.79s/it]2026-04-12 21:33:20,010 | INFO | train_grpo_train | metrics_logged mode=train step=1451

                                                        
{'loss': -0.0245, 'grad_norm': 15.212316513061523, 'learning_rate': 5.606060606060606e-06, 'num_tokens': 2565929.0, 'completions/mean_length': 18.875, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.6307111978530884, 'rewards/meter/std': 0.5018659234046936, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.852787971496582, 'rewards/lexical_plausibility/std': 0.17906473577022552, 'rewards/judge_quality/mean': 0.8462499976158142, 'rewards/judge_quality/std': 0.28152328729629517, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.48879387974739075, 'rewards/total_composite/std': 0.46672311425209045, 'reward': 0.48879387974739075, 'reward_std': 0.46672311425209045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10528997331857681, 'sampling/sampling_logp_difference/max': 0.9088325500488281, 'sampling/importance_sampling_ratio/min': 0.40660393238067627, 'sampling/importance_sampling_ratio/mean': 1.0172096490859985, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5657436288893223, 'clip_ratio/low_mean': 0.0486111119389534, 'clip_ratio/low_min': 0.0486111119389534, 'clip_ratio/high_mean': 0.042888471856713295, 'clip_ratio/high_max': 0.042888471856713295, 'clip_ratio/region_mean': 0.0914995837956667, 'reward_total_mean': 0.48879387974739075, 'reward_meter_mean': 0.6307111978530884, 'reward_meter_std': 0.5018659234046936, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.852787971496582, 'reward_lexical_plausibility_std': 0.17906473577022552, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.8462499976158142, 'reward_judge_quality_std': 0.28152328729629517, 'reward_total_composite_mean': 0.48879387974739075, 'reward_total_composite_std': 0.46672311425209045, 'epoch': 0.08}

 44%|████▍     | 1451/3300 [5:34:53<24:01:45, 46.79s/it]INFO 04-12 21:33:20 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1452/3300 [5:35:03<18:16:17, 35.59s/it]2026-04-12 21:33:29,496 | INFO | train_grpo_train | metrics_logged mode=train step=1452

                                                        
{'loss': 0.0865, 'grad_norm': 15.49891471862793, 'learning_rate': 5.603030303030303e-06, 'num_tokens': 2567597.0, 'completions/mean_length': 42.5, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7769848108291626, 'rewards/meter/std': 0.3613482415676117, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9822304248809814, 'rewards/lexical_plausibility/std': 0.0502600334584713, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.98424232006073, 'rewards/repeat_penalty/std': 0.03573727235198021, 'rewards/repeat_floor/mean': 0.9981409311294556, 'rewards/repeat_floor/std': 0.003948426805436611, 'rewards/near_duplicate_penalty/mean': 0.9931008815765381, 'rewards/near_duplicate_penalty/std': 0.011313963681459427, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9908424615859985, 'rewards/distinct_2/std': 0.025901345536112785, 'rewards/total_composite/mean': 0.27035024762153625, 'rewards/total_composite/std': 0.15232446789741516, 'reward': 0.27035024762153625, 'reward_std': 0.15232446789741516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1547611951828003, 'sampling/sampling_logp_difference/max': 1.599407434463501, 'sampling/importance_sampling_ratio/min': 0.2020161896944046, 'sampling/importance_sampling_ratio/mean': 1.0268700122833252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8347392901778221, 'clip_ratio/low_mean': 0.05114811006933451, 'clip_ratio/low_min': 0.05114811006933451, 'clip_ratio/high_mean': 0.0972203891724348, 'clip_ratio/high_max': 0.0972203891724348, 'clip_ratio/region_mean': 0.1483684992417693, 'reward_total_mean': 0.27035024762153625, 'reward_meter_mean': 0.7769848108291626, 'reward_meter_std': 0.3613482415676117, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9822304248809814, 'reward_lexical_plausibility_std': 0.0502600334584713, 'reward_repeat_penalty_mean': 0.98424232006073, 'reward_repeat_penalty_std': 0.03573727235198021, 'reward_repeat_floor_mean': 0.9981409311294556, 'reward_repeat_floor_std': 0.003948426805436611, 'reward_near_duplicate_penalty_mean': 0.9931008815765381, 'reward_near_duplicate_penalty_std': 0.011313963681459427, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9908424615859985, 'reward_distinct_2_std': 0.025901345536112785, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.27035024762153625, 'reward_total_composite_std': 0.15232446789741516, 'epoch': 0.08}

 44%|████▍     | 1452/3300 [5:35:03<18:16:17, 35.59s/it]INFO 04-12 21:33:29 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1453/3300 [5:35:12<14:15:53, 27.80s/it]2026-04-12 21:33:39,130 | INFO | train_grpo_train | metrics_logged mode=train step=1453

                                                        
{'loss': 0.0524, 'grad_norm': 12.834370613098145, 'learning_rate': 5.600000000000001e-06, 'num_tokens': 2569359.0, 'completions/mean_length': 53.25, 'completions/min_length': 45.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6973351240158081, 'rewards/meter/std': 0.4103483259677887, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9479973316192627, 'rewards/repeat_penalty/std': 0.06794706732034683, 'rewards/repeat_floor/mean': 0.9937500953674316, 'rewards/repeat_floor/std': 0.007863028906285763, 'rewards/near_duplicate_penalty/mean': 0.976303219795227, 'rewards/near_duplicate_penalty/std': 0.026468992233276367, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9700506925582886, 'rewards/distinct_2/std': 0.04615097492933273, 'rewards/total_composite/mean': 0.2968471944332123, 'rewards/total_composite/std': 0.17945891618728638, 'reward': 0.2968471944332123, 'reward_std': 0.17945891618728638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14051197469234467, 'sampling/sampling_logp_difference/max': 1.5314531326293945, 'sampling/importance_sampling_ratio/min': 0.21622124314308167, 'sampling/importance_sampling_ratio/mean': 1.0060518980026245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9625870659947395, 'clip_ratio/low_mean': 0.04444444552063942, 'clip_ratio/low_min': 0.04444444552063942, 'clip_ratio/high_mean': 0.07614768436178565, 'clip_ratio/high_max': 0.07614768436178565, 'clip_ratio/region_mean': 0.12059212988242507, 'reward_total_mean': 0.2968471944332123, 'reward_meter_mean': 0.6973351240158081, 'reward_meter_std': 0.4103483259677887, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9479973316192627, 'reward_repeat_penalty_std': 0.06794706732034683, 'reward_repeat_floor_mean': 0.9937500953674316, 'reward_repeat_floor_std': 0.007863028906285763, 'reward_near_duplicate_penalty_mean': 0.976303219795227, 'reward_near_duplicate_penalty_std': 0.026468992233276367, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9700506925582886, 'reward_distinct_2_std': 0.04615097492933273, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.2968471944332123, 'reward_total_composite_std': 0.17945891618728638, 'epoch': 0.08}

 44%|████▍     | 1453/3300 [5:35:12<14:15:53, 27.80s/it]INFO 04-12 21:33:39 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1454/3300 [5:35:22<11:30:49, 22.45s/it]2026-04-12 21:33:49,079 | INFO | train_grpo_train | metrics_logged mode=train step=1454

                                                        
{'loss': 0.0128, 'grad_norm': 9.163919448852539, 'learning_rate': 5.596969696969697e-06, 'num_tokens': 2571715.0, 'completions/mean_length': 87.5, 'completions/min_length': 77.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.5, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.844443678855896, 'rewards/meter/std': 0.23527929186820984, 'rewards/count_adherence/mean': 0.5892857313156128, 'rewards/count_adherence/std': 0.05050762742757797, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8767856955528259, 'rewards/count_floor/std': 0.015152297914028168, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9178019165992737, 'rewards/repeat_penalty/std': 0.025992587208747864, 'rewards/repeat_floor/mean': 0.9889581203460693, 'rewards/repeat_floor/std': 0.0027362657710909843, 'rewards/near_duplicate_penalty/mean': 0.9412369728088379, 'rewards/near_duplicate_penalty/std': 0.014868483878672123, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9752515554428101, 'rewards/distinct_2/std': 0.029101964086294174, 'rewards/total_composite/mean': 0.32958167791366577, 'rewards/total_composite/std': 0.09200555831193924, 'reward': 0.32958167791366577, 'reward_std': 0.09200555086135864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1267203837633133, 'sampling/sampling_logp_difference/max': 4.170647144317627, 'sampling/importance_sampling_ratio/min': 0.015442264266312122, 'sampling/importance_sampling_ratio/mean': 0.9946044683456421, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5722476579248905, 'clip_ratio/low_mean': 0.02560240961611271, 'clip_ratio/low_min': 0.02560240961611271, 'clip_ratio/high_mean': 0.09039441123604774, 'clip_ratio/high_max': 0.09039441123604774, 'clip_ratio/region_mean': 0.11599682085216045, 'reward_total_mean': 0.32958167791366577, 'reward_meter_mean': 0.844443678855896, 'reward_meter_std': 0.23527929186820984, 'reward_count_adherence_mean': 0.5892857313156128, 'reward_count_adherence_std': 0.05050762742757797, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8767856955528259, 'reward_count_floor_std': 0.015152297914028168, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9178019165992737, 'reward_repeat_penalty_std': 0.025992587208747864, 'reward_repeat_floor_mean': 0.9889581203460693, 'reward_repeat_floor_std': 0.0027362657710909843, 'reward_near_duplicate_penalty_mean': 0.9412369728088379, 'reward_near_duplicate_penalty_std': 0.014868483878672123, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9752515554428101, 'reward_distinct_2_std': 0.029101964086294174, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.32958167791366577, 'reward_total_composite_std': 0.09200555831193924, 'epoch': 0.08}

 44%|████▍     | 1454/3300 [5:35:22<11:30:49, 22.45s/it]INFO 04-12 21:33:49 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1455/3300 [5:35:36<10:10:38, 19.86s/it]2026-04-12 21:34:02,886 | INFO | train_grpo_train | metrics_logged mode=train step=1455

                                                        
{'loss': -0.0348, 'grad_norm': 4.8910017013549805, 'learning_rate': 5.593939393939395e-06, 'num_tokens': 2573015.0, 'completions/mean_length': 81.5, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.7246298789978027, 'rewards/meter/std': 0.4203140139579773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.7343491315841675, 'rewards/lexical_plausibility/std': 0.3033807873725891, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.33161941170692444, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.544529139995575, 'rewards/total_composite/std': 0.38631463050842285, 'reward': 0.544529139995575, 'reward_std': 0.38631460070610046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18220755457878113, 'sampling/sampling_logp_difference/max': 1.6146444082260132, 'sampling/importance_sampling_ratio/min': 0.19896140694618225, 'sampling/importance_sampling_ratio/mean': 1.0120280981063843, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8820024505257607, 'clip_ratio/low_mean': 0.03759398451074958, 'clip_ratio/low_min': 0.03759398451074958, 'clip_ratio/high_mean': 0.05915404297411442, 'clip_ratio/high_max': 0.05915404297411442, 'clip_ratio/region_mean': 0.096748027484864, 'reward_total_mean': 0.544529139995575, 'reward_meter_mean': 0.7246298789978027, 'reward_meter_std': 0.4203140139579773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.7343491315841675, 'reward_lexical_plausibility_std': 0.3033807873725891, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.33161941170692444, 'reward_total_composite_mean': 0.544529139995575, 'reward_total_composite_std': 0.38631463050842285, 'epoch': 0.08}

 44%|████▍     | 1455/3300 [5:35:36<10:10:38, 19.86s/it]INFO 04-12 21:34:03 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1456/3300 [5:35:46<8:34:32, 16.74s/it] 2026-04-12 21:34:12,356 | INFO | train_grpo_train | metrics_logged mode=train step=1456

                                                       
{'loss': -0.0021, 'grad_norm': 10.771652221679688, 'learning_rate': 5.5909090909090915e-06, 'num_tokens': 2574660.0, 'completions/mean_length': 59.625, 'completions/min_length': 51.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7960054874420166, 'rewards/meter/std': 0.33846190571784973, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.7404118180274963, 'rewards/repeat_penalty/std': 0.21907556056976318, 'rewards/repeat_floor/mean': 0.973548412322998, 'rewards/repeat_floor/std': 0.021930187940597534, 'rewards/near_duplicate_penalty/mean': 0.9327205419540405, 'rewards/near_duplicate_penalty/std': 0.045159511268138885, 'rewards/opening_diversity/mean': 0.9071428775787354, 'rewards/opening_diversity/std': 0.173793226480484, 'rewards/distinct_2/mean': 0.8801308274269104, 'rewards/distinct_2/std': 0.08492591232061386, 'rewards/total_composite/mean': 0.2411336600780487, 'rewards/total_composite/std': 0.14135272800922394, 'reward': 0.2411336600780487, 'reward_std': 0.14135274291038513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12735137343406677, 'sampling/sampling_logp_difference/max': 1.7067031860351562, 'sampling/importance_sampling_ratio/min': 0.1814630627632141, 'sampling/importance_sampling_ratio/mean': 1.014919638633728, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7837313786149025, 'clip_ratio/low_mean': 0.02834524679929018, 'clip_ratio/low_min': 0.02834524679929018, 'clip_ratio/high_mean': 0.06697310274466872, 'clip_ratio/high_max': 0.06697310274466872, 'clip_ratio/region_mean': 0.0953183495439589, 'reward_total_mean': 0.2411336600780487, 'reward_meter_mean': 0.7960054874420166, 'reward_meter_std': 0.33846190571784973, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7404118180274963, 'reward_repeat_penalty_std': 0.21907556056976318, 'reward_repeat_floor_mean': 0.973548412322998, 'reward_repeat_floor_std': 0.021930187940597534, 'reward_near_duplicate_penalty_mean': 0.9327205419540405, 'reward_near_duplicate_penalty_std': 0.045159511268138885, 'reward_opening_diversity_mean': 0.9071428775787354, 'reward_opening_diversity_std': 0.173793226480484, 'reward_distinct_2_mean': 0.8801308274269104, 'reward_distinct_2_std': 0.08492591232061386, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.2411336600780487, 'reward_total_composite_std': 0.14135272800922394, 'epoch': 0.08}

 44%|████▍     | 1456/3300 [5:35:46<8:34:32, 16.74s/it]INFO 04-12 21:34:12 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1457/3300 [5:35:56<7:33:42, 14.77s/it]2026-04-12 21:34:22,538 | INFO | train_grpo_train | metrics_logged mode=train step=1457

                                                       
{'loss': -0.0128, 'grad_norm': 10.029252052307129, 'learning_rate': 5.587878787878789e-06, 'num_tokens': 2577283.0, 'completions/mean_length': 103.875, 'completions/min_length': 80.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.875, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.5449814796447754, 'rewards/meter/std': 0.34511974453926086, 'rewards/count_adherence/mean': 0.7291666865348816, 'rewards/count_adherence/std': 0.0862581729888916, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9187499284744263, 'rewards/count_floor/std': 0.025877466425299644, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.20152543485164642, 'rewards/repeat_penalty/mean': 0.992195188999176, 'rewards/repeat_penalty/std': 0.004715257324278355, 'rewards/repeat_floor/mean': 0.9988293051719666, 'rewards/repeat_floor/std': 0.0007072848384268582, 'rewards/near_duplicate_penalty/mean': 0.992195188999176, 'rewards/near_duplicate_penalty/std': 0.004715257324278355, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20172077417373657, 'rewards/total_composite/std': 0.12398242205381393, 'reward': 0.20172077417373657, 'reward_std': 0.12398242205381393, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14147800207138062, 'sampling/sampling_logp_difference/max': 4.002121448516846, 'sampling/importance_sampling_ratio/min': 0.018276823684573174, 'sampling/importance_sampling_ratio/mean': 1.0025047063827515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7168181985616684, 'clip_ratio/low_mean': 0.026059217751026154, 'clip_ratio/low_min': 0.026059217751026154, 'clip_ratio/high_mean': 0.06202992983162403, 'clip_ratio/high_max': 0.06202992983162403, 'clip_ratio/region_mean': 0.08808914758265018, 'reward_total_mean': 0.20172077417373657, 'reward_meter_mean': 0.5449814796447754, 'reward_meter_std': 0.34511974453926086, 'reward_count_adherence_mean': 0.7291666865348816, 'reward_count_adherence_std': 0.0862581729888916, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9187499284744263, 'reward_count_floor_std': 0.025877466425299644, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.992195188999176, 'reward_repeat_penalty_std': 0.004715257324278355, 'reward_repeat_floor_mean': 0.9988293051719666, 'reward_repeat_floor_std': 0.0007072848384268582, 'reward_near_duplicate_penalty_mean': 0.992195188999176, 'reward_near_duplicate_penalty_std': 0.004715257324278355, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.20152543485164642, 'reward_total_composite_mean': 0.20172077417373657, 'reward_total_composite_std': 0.12398242205381393, 'epoch': 0.08}

 44%|████▍     | 1457/3300 [5:35:56<7:33:42, 14.77s/it]INFO 04-12 21:34:22 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1458/3300 [5:36:10<7:30:13, 14.67s/it]2026-04-12 21:34:36,944 | INFO | train_grpo_train | metrics_logged mode=train step=1458

                                                       
{'loss': -0.0324, 'grad_norm': 4.993852138519287, 'learning_rate': 5.584848484848485e-06, 'num_tokens': 2579045.0, 'completions/mean_length': 114.25, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.42857360839844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.24118903279304504, 'rewards/meter/std': 0.21816526353359222, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.07440238445997238, 'rewards/lexical_plausibility/mean': 0.9310699105262756, 'rewards/lexical_plausibility/std': 0.1949637234210968, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.24041630327701569, 'rewards/repeat_penalty/mean': 0.9280463457107544, 'rewards/repeat_penalty/std': 0.07874472439289093, 'rewards/repeat_floor/mean': 0.9930820465087891, 'rewards/repeat_floor/std': 0.004956471733748913, 'rewards/near_duplicate_penalty/mean': 0.977810800075531, 'rewards/near_duplicate_penalty/std': 0.019915655255317688, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.980948805809021, 'rewards/distinct_2/std': 0.026327626779675484, 'rewards/total_composite/mean': 0.14034301042556763, 'rewards/total_composite/std': 0.20952698588371277, 'reward': 0.14034301042556763, 'reward_std': 0.20952698588371277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1216459646821022, 'sampling/sampling_logp_difference/max': 1.603621482849121, 'sampling/importance_sampling_ratio/min': 0.20116667449474335, 'sampling/importance_sampling_ratio/mean': 1.011099934577942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6306528896093369, 'clip_ratio/low_mean': 0.0635438822209835, 'clip_ratio/low_min': 0.0635438822209835, 'clip_ratio/high_mean': 0.013392857275903225, 'clip_ratio/high_max': 0.013392857275903225, 'clip_ratio/region_mean': 0.07693673949688673, 'reward_total_mean': 0.14034301042556763, 'reward_meter_mean': 0.24118903279304504, 'reward_meter_std': 0.21816526353359222, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.07440238445997238, 'reward_lexical_plausibility_mean': 0.9310699105262756, 'reward_lexical_plausibility_std': 0.1949637234210968, 'reward_repeat_penalty_mean': 0.9280463457107544, 'reward_repeat_penalty_std': 0.07874472439289093, 'reward_repeat_floor_mean': 0.9930820465087891, 'reward_repeat_floor_std': 0.004956471733748913, 'reward_near_duplicate_penalty_mean': 0.977810800075531, 'reward_near_duplicate_penalty_std': 0.019915655255317688, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.980948805809021, 'reward_distinct_2_std': 0.026327626779675484, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.24041630327701569, 'reward_total_composite_mean': 0.14034301042556763, 'reward_total_composite_std': 0.20952698588371277, 'epoch': 0.08}

 44%|████▍     | 1458/3300 [5:36:10<7:30:13, 14.67s/it]INFO 04-12 21:34:37 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1459/3300 [5:36:21<6:51:06, 13.40s/it]2026-04-12 21:34:47,392 | INFO | train_grpo_train | metrics_logged mode=train step=1459

                                                       
{'loss': 0.0482, 'grad_norm': 11.061907768249512, 'learning_rate': 5.5818181818181824e-06, 'num_tokens': 2580937.0, 'completions/mean_length': 65.5, 'completions/min_length': 62.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.797038197517395, 'rewards/meter/std': 0.2192562371492386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6212500333786011, 'rewards/judge_quality/std': 0.26626721024513245, 'rewards/repeat_penalty/mean': 0.9703969359397888, 'rewards/repeat_penalty/std': 0.034107450395822525, 'rewards/repeat_floor/mean': 0.9966459274291992, 'rewards/repeat_floor/std': 0.003730526426807046, 'rewards/near_duplicate_penalty/mean': 0.9890539050102234, 'rewards/near_duplicate_penalty/std': 0.012349693104624748, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9809759855270386, 'rewards/distinct_2/std': 0.026260556653141975, 'rewards/total_composite/mean': 0.5113263726234436, 'rewards/total_composite/std': 0.29251784086227417, 'reward': 0.5113263726234436, 'reward_std': 0.29251787066459656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1326427310705185, 'sampling/sampling_logp_difference/max': 2.559915542602539, 'sampling/importance_sampling_ratio/min': 0.07731126993894577, 'sampling/importance_sampling_ratio/mean': 1.0033767223358154, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6349594257771969, 'clip_ratio/low_mean': 0.07067729264963418, 'clip_ratio/low_min': 0.07067729264963418, 'clip_ratio/high_mean': 0.04168027266860008, 'clip_ratio/high_max': 0.04168027266860008, 'clip_ratio/region_mean': 0.11235756531823426, 'reward_total_mean': 0.5113263726234436, 'reward_meter_mean': 0.797038197517395, 'reward_meter_std': 0.2192562371492386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9703969359397888, 'reward_repeat_penalty_std': 0.034107450395822525, 'reward_repeat_floor_mean': 0.9966459274291992, 'reward_repeat_floor_std': 0.003730526426807046, 'reward_near_duplicate_penalty_mean': 0.9890539050102234, 'reward_near_duplicate_penalty_std': 0.012349693104624748, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9809759855270386, 'reward_distinct_2_std': 0.026260556653141975, 'reward_judge_quality_mean': 0.6212500333786011, 'reward_judge_quality_std': 0.26626721024513245, 'reward_total_composite_mean': 0.5113263726234436, 'reward_total_composite_std': 0.29251784086227417, 'epoch': 0.08}

 44%|████▍     | 1459/3300 [5:36:21<6:51:06, 13.40s/it]INFO 04-12 21:34:47 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1460/3300 [5:36:28<5:59:53, 11.74s/it]2026-04-12 21:34:55,251 | INFO | train_grpo_train | metrics_logged mode=train step=1460

                                                       
{'loss': 0.0541, 'grad_norm': 18.399173736572266, 'learning_rate': 5.578787878787879e-06, 'num_tokens': 2582536.0, 'completions/mean_length': 34.875, 'completions/min_length': 31.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.26310494542121887, 'rewards/meter/std': 0.23746247589588165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9997294545173645, 'rewards/repeat_penalty/std': 0.0007652603671886027, 'rewards/repeat_floor/mean': 0.999959409236908, 'rewards/repeat_floor/std': 0.00011478694068500772, 'rewards/near_duplicate_penalty/mean': 0.9997294545173645, 'rewards/near_duplicate_penalty/std': 0.0007652603671886027, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.24205397069454193, 'rewards/total_composite/std': 0.2184678614139557, 'reward': 0.24205397069454193, 'reward_std': 0.21846787631511688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1296447068452835, 'sampling/sampling_logp_difference/max': 1.6458253860473633, 'sampling/importance_sampling_ratio/min': 0.19285331666469574, 'sampling/importance_sampling_ratio/mean': 0.9987260103225708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6240876615047455, 'clip_ratio/low_mean': 0.060707886004820466, 'clip_ratio/low_min': 0.060707886004820466, 'clip_ratio/high_mean': 0.06086601410061121, 'clip_ratio/high_max': 0.06086601410061121, 'clip_ratio/region_mean': 0.12157390010543168, 'reward_total_mean': 0.24205397069454193, 'reward_meter_mean': 0.26310494542121887, 'reward_meter_std': 0.23746247589588165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9997294545173645, 'reward_repeat_penalty_std': 0.0007652603671886027, 'reward_repeat_floor_mean': 0.999959409236908, 'reward_repeat_floor_std': 0.00011478694068500772, 'reward_near_duplicate_penalty_mean': 0.9997294545173645, 'reward_near_duplicate_penalty_std': 0.0007652603671886027, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.24205397069454193, 'reward_total_composite_std': 0.2184678614139557, 'epoch': 0.08}

 44%|████▍     | 1460/3300 [5:36:29<5:59:53, 11.74s/it]INFO 04-12 21:34:55 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1461/3300 [5:36:37<5:32:00, 10.83s/it]2026-04-12 21:35:03,975 | INFO | train_grpo_train | metrics_logged mode=train step=1461

                                                       
{'loss': 0.0321, 'grad_norm': 16.66950798034668, 'learning_rate': 5.575757575757577e-06, 'num_tokens': 2584043.0, 'completions/mean_length': 39.375, 'completions/min_length': 36.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6944570541381836, 'rewards/meter/std': 0.38515704870224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.2796298861503601, 'rewards/repeat_penalty/mean': 0.9830665588378906, 'rewards/repeat_penalty/std': 0.028471464291214943, 'rewards/repeat_floor/mean': 0.9980672597885132, 'rewards/repeat_floor/std': 0.0028028227388858795, 'rewards/near_duplicate_penalty/mean': 0.9931880235671997, 'rewards/near_duplicate_penalty/std': 0.012192907743155956, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.3138124346733093, 'rewards/total_composite/std': 0.27168527245521545, 'reward': 0.3138124346733093, 'reward_std': 0.27168530225753784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14927981793880463, 'sampling/sampling_logp_difference/max': 1.4698505401611328, 'sampling/importance_sampling_ratio/min': 0.22995984554290771, 'sampling/importance_sampling_ratio/mean': 0.9951277375221252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8751384690403938, 'clip_ratio/low_mean': 0.08426490612328053, 'clip_ratio/low_min': 0.08426490612328053, 'clip_ratio/high_mean': 0.058296571020036936, 'clip_ratio/high_max': 0.058296571020036936, 'clip_ratio/region_mean': 0.14256147714331746, 'reward_total_mean': 0.3138124346733093, 'reward_meter_mean': 0.6944570541381836, 'reward_meter_std': 0.38515704870224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9830665588378906, 'reward_repeat_penalty_std': 0.028471464291214943, 'reward_repeat_floor_mean': 0.9980672597885132, 'reward_repeat_floor_std': 0.0028028227388858795, 'reward_near_duplicate_penalty_mean': 0.9931880235671997, 'reward_near_duplicate_penalty_std': 0.012192907743155956, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.2796298861503601, 'reward_total_composite_mean': 0.3138124346733093, 'reward_total_composite_std': 0.27168527245521545, 'epoch': 0.08}

 44%|████▍     | 1461/3300 [5:36:37<5:32:00, 10.83s/it]INFO 04-12 21:35:04 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1462/3300 [5:36:46<5:15:30, 10.30s/it]2026-04-12 21:35:13,043 | INFO | train_grpo_train | metrics_logged mode=train step=1462

                                                       
{'loss': 0.0702, 'grad_norm': 16.789833068847656, 'learning_rate': 5.572727272727273e-06, 'num_tokens': 2585502.0, 'completions/mean_length': 34.375, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7005495429039001, 'rewards/meter/std': 0.37462905049324036, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6725000143051147, 'rewards/judge_quality/std': 0.2666056752204895, 'rewards/repeat_penalty/mean': 0.9376727342605591, 'rewards/repeat_penalty/std': 0.06537135690450668, 'rewards/repeat_floor/mean': 0.9915980100631714, 'rewards/repeat_floor/std': 0.008354512974619865, 'rewards/near_duplicate_penalty/mean': 0.9558917284011841, 'rewards/near_duplicate_penalty/std': 0.03984785079956055, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9801586866378784, 'rewards/distinct_2/std': 0.036883533000946045, 'rewards/total_composite/mean': 0.48158371448516846, 'rewards/total_composite/std': 0.3427993059158325, 'reward': 0.48158371448516846, 'reward_std': 0.3427993059158325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1685110330581665, 'sampling/sampling_logp_difference/max': 1.707540512084961, 'sampling/importance_sampling_ratio/min': 0.18131117522716522, 'sampling/importance_sampling_ratio/mean': 1.0167186260223389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1643597856163979, 'clip_ratio/low_mean': 0.09971459209918976, 'clip_ratio/low_min': 0.09971459209918976, 'clip_ratio/high_mean': 0.04715909156948328, 'clip_ratio/high_max': 0.04715909156948328, 'clip_ratio/region_mean': 0.14687368366867304, 'reward_total_mean': 0.48158371448516846, 'reward_meter_mean': 0.7005495429039001, 'reward_meter_std': 0.37462905049324036, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9376727342605591, 'reward_repeat_penalty_std': 0.06537135690450668, 'reward_repeat_floor_mean': 0.9915980100631714, 'reward_repeat_floor_std': 0.008354512974619865, 'reward_near_duplicate_penalty_mean': 0.9558917284011841, 'reward_near_duplicate_penalty_std': 0.03984785079956055, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9801586866378784, 'reward_distinct_2_std': 0.036883533000946045, 'reward_judge_quality_mean': 0.6725000143051147, 'reward_judge_quality_std': 0.2666056752204895, 'reward_total_composite_mean': 0.48158371448516846, 'reward_total_composite_std': 0.3427993059158325, 'epoch': 0.08}

 44%|████▍     | 1462/3300 [5:36:46<5:15:30, 10.30s/it]INFO 04-12 21:35:13 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1463/3300 [5:36:55<5:01:54,  9.86s/it]2026-04-12 21:35:21,870 | INFO | train_grpo_train | metrics_logged mode=train step=1463

                                                       
{'loss': -0.0049, 'grad_norm': 17.285612106323242, 'learning_rate': 5.569696969696971e-06, 'num_tokens': 2587018.0, 'completions/mean_length': 31.5, 'completions/min_length': 29.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.41260722279548645, 'rewards/meter/std': 0.40489640831947327, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5037499666213989, 'rewards/judge_quality/std': 0.2209678441286087, 'rewards/repeat_penalty/mean': 0.9876722693443298, 'rewards/repeat_penalty/std': 0.01332098338752985, 'rewards/repeat_floor/mean': 0.9981508255004883, 'rewards/repeat_floor/std': 0.0019981428049504757, 'rewards/near_duplicate_penalty/mean': 0.9876722693443298, 'rewards/near_duplicate_penalty/std': 0.01332098338752985, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.19124242663383484, 'rewards/total_composite/std': 0.24342593550682068, 'reward': 0.19124242663383484, 'reward_std': 0.24342593550682068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11594981700181961, 'sampling/sampling_logp_difference/max': 1.809114933013916, 'sampling/importance_sampling_ratio/min': 0.16379904747009277, 'sampling/importance_sampling_ratio/mean': 0.9999021887779236, 'sampling/importance_sampling_ratio/max': 1.8358376026153564, 'entropy': 0.7771306931972504, 'clip_ratio/low_mean': 0.09983001556247473, 'clip_ratio/low_min': 0.09983001556247473, 'clip_ratio/high_mean': 0.03042140230536461, 'clip_ratio/high_max': 0.03042140230536461, 'clip_ratio/region_mean': 0.13025141786783934, 'reward_total_mean': 0.19124242663383484, 'reward_meter_mean': 0.41260722279548645, 'reward_meter_std': 0.40489640831947327, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9876722693443298, 'reward_repeat_penalty_std': 0.01332098338752985, 'reward_repeat_floor_mean': 0.9981508255004883, 'reward_repeat_floor_std': 0.0019981428049504757, 'reward_near_duplicate_penalty_mean': 0.9876722693443298, 'reward_near_duplicate_penalty_std': 0.01332098338752985, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5037499666213989, 'reward_judge_quality_std': 0.2209678441286087, 'reward_total_composite_mean': 0.19124242663383484, 'reward_total_composite_std': 0.24342593550682068, 'epoch': 0.08}

 44%|████▍     | 1463/3300 [5:36:55<5:01:54,  9.86s/it]INFO 04-12 21:35:22 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1464/3300 [5:37:10<5:45:07, 11.28s/it]2026-04-12 21:35:36,463 | INFO | train_grpo_train | metrics_logged mode=train step=1464

                                                       
{'loss': -0.053, 'grad_norm': 6.3476481437683105, 'learning_rate': 5.566666666666667e-06, 'num_tokens': 2588812.0, 'completions/mean_length': 121.25, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.42857360839844, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.42968156933784485, 'rewards/meter/std': 0.3897289037704468, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9489843845367432, 'rewards/lexical_plausibility/std': 0.14429405331611633, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, 'rewards/repeat_penalty/mean': 0.8424714207649231, 'rewards/repeat_penalty/std': 0.1445269137620926, 'rewards/repeat_floor/mean': 0.9826183319091797, 'rewards/repeat_floor/std': 0.014140552841126919, 'rewards/near_duplicate_penalty/mean': 0.9461427330970764, 'rewards/near_duplicate_penalty/std': 0.030269332230091095, 'rewards/opening_diversity/mean': 0.9508928656578064, 'rewards/opening_diversity/std': 0.09533105045557022, 'rewards/distinct_2/mean': 0.9293700456619263, 'rewards/distinct_2/std': 0.06268090754747391, 'rewards/total_composite/mean': 0.1584310233592987, 'rewards/total_composite/std': 0.17256559431552887, 'reward': 0.1584310233592987, 'reward_std': 0.17256557941436768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15980832278728485, 'sampling/sampling_logp_difference/max': 1.949110984802246, 'sampling/importance_sampling_ratio/min': 0.1424006074666977, 'sampling/importance_sampling_ratio/mean': 1.0366225242614746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8533193469047546, 'clip_ratio/low_mean': 0.06362336687743664, 'clip_ratio/low_min': 0.06362336687743664, 'clip_ratio/high_mean': 0.05051523260772228, 'clip_ratio/high_max': 0.05051523260772228, 'clip_ratio/region_mean': 0.11413859948515892, 'reward_total_mean': 0.1584310233592987, 'reward_meter_mean': 0.42968156933784485, 'reward_meter_std': 0.3897289037704468, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9489843845367432, 'reward_lexical_plausibility_std': 0.14429405331611633, 'reward_repeat_penalty_mean': 0.8424714207649231, 'reward_repeat_penalty_std': 0.1445269137620926, 'reward_repeat_floor_mean': 0.9826183319091797, 'reward_repeat_floor_std': 0.014140552841126919, 'reward_near_duplicate_penalty_mean': 0.9461427330970764, 'reward_near_duplicate_penalty_std': 0.030269332230091095, 'reward_opening_diversity_mean': 0.9508928656578064, 'reward_opening_diversity_std': 0.09533105045557022, 'reward_distinct_2_mean': 0.9293700456619263, 'reward_distinct_2_std': 0.06268090754747391, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.1584310233592987, 'reward_total_composite_std': 0.17256559431552887, 'epoch': 0.08}

 44%|████▍     | 1464/3300 [5:37:10<5:45:07, 11.28s/it]INFO 04-12 21:35:36 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1465/3300 [5:37:19<5:25:04, 10.63s/it]2026-04-12 21:35:45,563 | INFO | train_grpo_train | metrics_logged mode=train step=1465

                                                       
{'loss': 0.0986, 'grad_norm': 23.177236557006836, 'learning_rate': 5.563636363636364e-06, 'num_tokens': 2590313.0, 'completions/mean_length': 38.625, 'completions/min_length': 32.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.827404260635376, 'rewards/meter/std': 0.28178417682647705, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.4749999940395355, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9689664244651794, 'rewards/repeat_penalty/std': 0.04249228909611702, 'rewards/repeat_floor/mean': 0.9963927865028381, 'rewards/repeat_floor/std': 0.004672999028116465, 'rewards/near_duplicate_penalty/mean': 0.987041711807251, 'rewards/near_duplicate_penalty/std': 0.017708996310830116, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9815173149108887, 'rewards/distinct_2/std': 0.03442956507205963, 'rewards/total_composite/mean': 0.3947297930717468, 'rewards/total_composite/std': 0.17811062932014465, 'reward': 0.3947297930717468, 'reward_std': 0.17811061441898346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14125335216522217, 'sampling/sampling_logp_difference/max': 1.6986949443817139, 'sampling/importance_sampling_ratio/min': 0.21040359139442444, 'sampling/importance_sampling_ratio/mean': 1.0032693147659302, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9236818850040436, 'clip_ratio/low_mean': 0.03358563641086221, 'clip_ratio/low_min': 0.03358563641086221, 'clip_ratio/high_mean': 0.07814845256507397, 'clip_ratio/high_max': 0.07814845256507397, 'clip_ratio/region_mean': 0.11173408897593617, 'reward_total_mean': 0.3947297930717468, 'reward_meter_mean': 0.827404260635376, 'reward_meter_std': 0.28178417682647705, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9689664244651794, 'reward_repeat_penalty_std': 0.04249228909611702, 'reward_repeat_floor_mean': 0.9963927865028381, 'reward_repeat_floor_std': 0.004672999028116465, 'reward_near_duplicate_penalty_mean': 0.987041711807251, 'reward_near_duplicate_penalty_std': 0.017708996310830116, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9815173149108887, 'reward_distinct_2_std': 0.03442956507205963, 'reward_judge_quality_mean': 0.4749999940395355, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.3947297930717468, 'reward_total_composite_std': 0.17811062932014465, 'epoch': 0.08}

 44%|████▍     | 1465/3300 [5:37:19<5:25:04, 10.63s/it]INFO 04-12 21:35:45 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1466/3300 [5:37:33<6:01:33, 11.83s/it]2026-04-12 21:36:00,191 | INFO | train_grpo_train | metrics_logged mode=train step=1466

                                                       
{'loss': -0.079, 'grad_norm': 4.824151992797852, 'learning_rate': 5.560606060606061e-06, 'num_tokens': 2591807.0, 'completions/mean_length': 94.75, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.142860412597656, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7681292295455933, 'rewards/meter/std': 0.39944857358932495, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9221006631851196, 'rewards/lexical_plausibility/std': 0.22033251821994781, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.18516400456428528, 'rewards/repeat_penalty/mean': 0.8969316482543945, 'rewards/repeat_penalty/std': 0.12550103664398193, 'rewards/repeat_floor/mean': 0.9893254041671753, 'rewards/repeat_floor/std': 0.013125483877956867, 'rewards/near_duplicate_penalty/mean': 0.9673749804496765, 'rewards/near_duplicate_penalty/std': 0.043096814304590225, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9566013813018799, 'rewards/distinct_2/std': 0.08165136724710464, 'rewards/total_composite/mean': 0.38969165086746216, 'rewards/total_composite/std': 0.22181619703769684, 'reward': 0.38969165086746216, 'reward_std': 0.22181621193885803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14818120002746582, 'sampling/sampling_logp_difference/max': 1.7204985618591309, 'sampling/importance_sampling_ratio/min': 0.1789768934249878, 'sampling/importance_sampling_ratio/mean': 1.0239019393920898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.78403040766716, 'clip_ratio/low_mean': 0.011363636702299118, 'clip_ratio/low_min': 0.011363636702299118, 'clip_ratio/high_mean': 0.09900515992194414, 'clip_ratio/high_max': 0.09900515992194414, 'clip_ratio/region_mean': 0.11036879662424326, 'reward_total_mean': 0.38969165086746216, 'reward_meter_mean': 0.7681292295455933, 'reward_meter_std': 0.39944857358932495, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9221006631851196, 'reward_lexical_plausibility_std': 0.22033251821994781, 'reward_repeat_penalty_mean': 0.8969316482543945, 'reward_repeat_penalty_std': 0.12550103664398193, 'reward_repeat_floor_mean': 0.9893254041671753, 'reward_repeat_floor_std': 0.013125483877956867, 'reward_near_duplicate_penalty_mean': 0.9673749804496765, 'reward_near_duplicate_penalty_std': 0.043096814304590225, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9566013813018799, 'reward_distinct_2_std': 0.08165136724710464, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.18516400456428528, 'reward_total_composite_mean': 0.38969165086746216, 'reward_total_composite_std': 0.22181619703769684, 'epoch': 0.08}

 44%|████▍     | 1466/3300 [5:37:34<6:01:33, 11.83s/it]INFO 04-12 21:36:00 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1467/3300 [5:37:42<5:35:09, 10.97s/it]2026-04-12 21:36:09,156 | INFO | train_grpo_train | metrics_logged mode=train step=1467

                                                       
{'loss': 0.026, 'grad_norm': 10.648513793945312, 'learning_rate': 5.557575757575758e-06, 'num_tokens': 2593293.0, 'completions/mean_length': 40.75, 'completions/min_length': 38.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7763723134994507, 'rewards/meter/std': 0.3710571825504303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9292891621589661, 'rewards/repeat_penalty/std': 0.11720864474773407, 'rewards/repeat_floor/mean': 0.9919396042823792, 'rewards/repeat_floor/std': 0.01083588507026434, 'rewards/near_duplicate_penalty/mean': 0.9632017016410828, 'rewards/near_duplicate_penalty/std': 0.03623322397470474, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9926035404205322, 'rewards/distinct_2/std': 0.020920326933264732, 'rewards/total_composite/mean': 0.27405762672424316, 'rewards/total_composite/std': 0.12334898114204407, 'reward': 0.27405762672424316, 'reward_std': 0.12334898114204407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15289442241191864, 'sampling/sampling_logp_difference/max': 1.4374723434448242, 'sampling/importance_sampling_ratio/min': 0.23752740025520325, 'sampling/importance_sampling_ratio/mean': 1.0133510828018188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0599054545164108, 'clip_ratio/low_mean': 0.044391026720404625, 'clip_ratio/low_min': 0.044391026720404625, 'clip_ratio/high_mean': 0.12404589354991913, 'clip_ratio/high_max': 0.12404589354991913, 'clip_ratio/region_mean': 0.16843692027032375, 'reward_total_mean': 0.27405762672424316, 'reward_meter_mean': 0.7763723134994507, 'reward_meter_std': 0.3710571825504303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9292891621589661, 'reward_repeat_penalty_std': 0.11720864474773407, 'reward_repeat_floor_mean': 0.9919396042823792, 'reward_repeat_floor_std': 0.01083588507026434, 'reward_near_duplicate_penalty_mean': 0.9632017016410828, 'reward_near_duplicate_penalty_std': 0.03623322397470474, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9926035404205322, 'reward_distinct_2_std': 0.020920326933264732, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.27405762672424316, 'reward_total_composite_std': 0.12334898114204407, 'epoch': 0.08}

 44%|████▍     | 1467/3300 [5:37:43<5:35:09, 10.97s/it]INFO 04-12 21:36:09 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1468/3300 [5:37:52<5:23:18, 10.59s/it]2026-04-12 21:36:18,865 | INFO | train_grpo_train | metrics_logged mode=train step=1468

                                                       
{'loss': 0.0396, 'grad_norm': 14.380463600158691, 'learning_rate': 5.554545454545454e-06, 'num_tokens': 2595749.0, 'completions/mean_length': 93.0, 'completions/min_length': 84.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.8517998456954956, 'rewards/meter/std': 0.23740869760513306, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.887499988079071, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4662500023841858, 'rewards/judge_quality/std': 0.1743508279323578, 'rewards/repeat_penalty/mean': 0.9787132143974304, 'rewards/repeat_penalty/std': 0.03088396228849888, 'rewards/repeat_floor/mean': 0.9975749850273132, 'rewards/repeat_floor/std': 0.003145806025713682, 'rewards/near_duplicate_penalty/mean': 0.9918221831321716, 'rewards/near_duplicate_penalty/std': 0.007289330009371042, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9866847991943359, 'rewards/distinct_2/std': 0.026523606851696968, 'rewards/total_composite/mean': 0.3590654730796814, 'rewards/total_composite/std': 0.18661890923976898, 'reward': 0.3590654730796814, 'reward_std': 0.18661890923976898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13440656661987305, 'sampling/sampling_logp_difference/max': 1.8946889638900757, 'sampling/importance_sampling_ratio/min': 0.15036509931087494, 'sampling/importance_sampling_ratio/mean': 1.013350009918213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7228589132428169, 'clip_ratio/low_mean': 0.06182474084198475, 'clip_ratio/low_min': 0.06182474084198475, 'clip_ratio/high_mean': 0.05634537059813738, 'clip_ratio/high_max': 0.05634537059813738, 'clip_ratio/region_mean': 0.11817011144012213, 'reward_total_mean': 0.3590654730796814, 'reward_meter_mean': 0.8517998456954956, 'reward_meter_std': 0.23740869760513306, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.887499988079071, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9787132143974304, 'reward_repeat_penalty_std': 0.03088396228849888, 'reward_repeat_floor_mean': 0.9975749850273132, 'reward_repeat_floor_std': 0.003145806025713682, 'reward_near_duplicate_penalty_mean': 0.9918221831321716, 'reward_near_duplicate_penalty_std': 0.007289330009371042, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9866847991943359, 'reward_distinct_2_std': 0.026523606851696968, 'reward_judge_quality_mean': 0.4662500023841858, 'reward_judge_quality_std': 0.1743508279323578, 'reward_total_composite_mean': 0.3590654730796814, 'reward_total_composite_std': 0.18661890923976898, 'epoch': 0.08}

 44%|████▍     | 1468/3300 [5:37:52<5:23:18, 10.59s/it]INFO 04-12 21:36:19 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1469/3300 [5:38:02<5:13:41, 10.28s/it]2026-04-12 21:36:28,424 | INFO | train_grpo_train | metrics_logged mode=train step=1469

                                                       
{'loss': 0.054, 'grad_norm': 12.183830261230469, 'learning_rate': 5.5515151515151524e-06, 'num_tokens': 2597553.0, 'completions/mean_length': 53.5, 'completions/min_length': 46.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6609489917755127, 'rewards/meter/std': 0.36367860436439514, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.9624999761581421, 'rewards/arabic_floor/std': 0.06943649053573608, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093073666095734, 'rewards/repeat_penalty/mean': 0.9753797054290771, 'rewards/repeat_penalty/std': 0.014399535022675991, 'rewards/repeat_floor/mean': 0.9963070154190063, 'rewards/repeat_floor/std': 0.0021599384490400553, 'rewards/near_duplicate_penalty/mean': 0.9753797054290771, 'rewards/near_duplicate_penalty/std': 0.014399535022675991, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2325470745563507, 'rewards/total_composite/std': 0.17948530614376068, 'reward': 0.2325470745563507, 'reward_std': 0.17948530614376068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16152890026569366, 'sampling/sampling_logp_difference/max': 1.8105525970458984, 'sampling/importance_sampling_ratio/min': 0.16356372833251953, 'sampling/importance_sampling_ratio/mean': 1.010656476020813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9364163056015968, 'clip_ratio/low_mean': 0.07524390611797571, 'clip_ratio/low_min': 0.07524390611797571, 'clip_ratio/high_mean': 0.07295853178948164, 'clip_ratio/high_max': 0.07295853178948164, 'clip_ratio/region_mean': 0.14820243790745735, 'reward_total_mean': 0.2325470745563507, 'reward_meter_mean': 0.6609489917755127, 'reward_meter_std': 0.36367860436439514, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.9624999761581421, 'reward_arabic_floor_std': 0.06943649053573608, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9753797054290771, 'reward_repeat_penalty_std': 0.014399535022675991, 'reward_repeat_floor_mean': 0.9963070154190063, 'reward_repeat_floor_std': 0.0021599384490400553, 'reward_near_duplicate_penalty_mean': 0.9753797054290771, 'reward_near_duplicate_penalty_std': 0.014399535022675991, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.13093073666095734, 'reward_total_composite_mean': 0.2325470745563507, 'reward_total_composite_std': 0.17948530614376068, 'epoch': 0.08}

 45%|████▍     | 1469/3300 [5:38:02<5:13:41, 10.28s/it]INFO 04-12 21:36:28 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1470/3300 [5:38:11<5:01:17,  9.88s/it]2026-04-12 21:36:37,353 | INFO | train_grpo_train | metrics_logged mode=train step=1470

                                                       
{'loss': 0.0716, 'grad_norm': 14.52413272857666, 'learning_rate': 5.548484848484849e-06, 'num_tokens': 2599172.0, 'completions/mean_length': 37.375, 'completions/min_length': 31.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6708568930625916, 'rewards/meter/std': 0.37073376774787903, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/repeat_penalty/mean': 0.9427042007446289, 'rewards/repeat_penalty/std': 0.11969783157110214, 'rewards/repeat_floor/mean': 0.9924949407577515, 'rewards/repeat_floor/std': 0.014933905564248562, 'rewards/near_duplicate_penalty/mean': 0.9674487709999084, 'rewards/near_duplicate_penalty/std': 0.052374377846717834, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9708625078201294, 'rewards/distinct_2/std': 0.08241338282823563, 'rewards/total_composite/mean': 0.4312814176082611, 'rewards/total_composite/std': 0.3087232708930969, 'reward': 0.4312814176082611, 'reward_std': 0.3087233006954193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14285679161548615, 'sampling/sampling_logp_difference/max': 1.5482988357543945, 'sampling/importance_sampling_ratio/min': 0.21260933578014374, 'sampling/importance_sampling_ratio/mean': 1.0080362558364868, 'sampling/importance_sampling_ratio/max': 1.7865575551986694, 'entropy': 0.9456106573343277, 'clip_ratio/low_mean': 0.09007683023810387, 'clip_ratio/low_min': 0.09007683023810387, 'clip_ratio/high_mean': 0.0358193286228925, 'clip_ratio/high_max': 0.0358193286228925, 'clip_ratio/region_mean': 0.12589615886099637, 'reward_total_mean': 0.4312814176082611, 'reward_meter_mean': 0.6708568930625916, 'reward_meter_std': 0.37073376774787903, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9427042007446289, 'reward_repeat_penalty_std': 0.11969783157110214, 'reward_repeat_floor_mean': 0.9924949407577515, 'reward_repeat_floor_std': 0.014933905564248562, 'reward_near_duplicate_penalty_mean': 0.9674487709999084, 'reward_near_duplicate_penalty_std': 0.052374377846717834, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9708625078201294, 'reward_distinct_2_std': 0.08241338282823563, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.4312814176082611, 'reward_total_composite_std': 0.3087232708930969, 'epoch': 0.08}

 45%|████▍     | 1470/3300 [5:38:11<5:01:17,  9.88s/it]INFO 04-12 21:36:37 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1471/3300 [5:38:20<4:53:43,  9.64s/it]2026-04-12 21:36:46,436 | INFO | train_grpo_train | metrics_logged mode=train step=1471

                                                       
{'loss': 0.0502, 'grad_norm': 13.896950721740723, 'learning_rate': 5.545454545454546e-06, 'num_tokens': 2600713.0, 'completions/mean_length': 33.625, 'completions/min_length': 28.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.903566300868988, 'rewards/meter/std': 0.12202103435993195, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6637500524520874, 'rewards/judge_quality/std': 0.2351861149072647, 'rewards/repeat_penalty/mean': 0.980202853679657, 'rewards/repeat_penalty/std': 0.025147607550024986, 'rewards/repeat_floor/mean': 0.9970304369926453, 'rewards/repeat_floor/std': 0.003772138385102153, 'rewards/near_duplicate_penalty/mean': 0.980202853679657, 'rewards/near_duplicate_penalty/std': 0.025147607550024986, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.604533314704895, 'rewards/total_composite/std': 0.24003581702709198, 'reward': 0.604533314704895, 'reward_std': 0.24003581702709198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11794211715459824, 'sampling/sampling_logp_difference/max': 1.3657530546188354, 'sampling/importance_sampling_ratio/min': 0.2551884353160858, 'sampling/importance_sampling_ratio/mean': 1.035339593887329, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8341148272156715, 'clip_ratio/low_mean': 0.04921622667461634, 'clip_ratio/low_min': 0.04921622667461634, 'clip_ratio/high_mean': 0.059879032894968987, 'clip_ratio/high_max': 0.059879032894968987, 'clip_ratio/region_mean': 0.10909525956958532, 'reward_total_mean': 0.604533314704895, 'reward_meter_mean': 0.903566300868988, 'reward_meter_std': 0.12202103435993195, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.980202853679657, 'reward_repeat_penalty_std': 0.025147607550024986, 'reward_repeat_floor_mean': 0.9970304369926453, 'reward_repeat_floor_std': 0.003772138385102153, 'reward_near_duplicate_penalty_mean': 0.980202853679657, 'reward_near_duplicate_penalty_std': 0.025147607550024986, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6637500524520874, 'reward_judge_quality_std': 0.2351861149072647, 'reward_total_composite_mean': 0.604533314704895, 'reward_total_composite_std': 0.24003581702709198, 'epoch': 0.08}

 45%|████▍     | 1471/3300 [5:38:20<4:53:43,  9.64s/it]INFO 04-12 21:36:46 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1472/3300 [5:38:34<5:34:08, 10.97s/it]2026-04-12 21:37:00,605 | INFO | train_grpo_train | metrics_logged mode=train step=1472

                                                       
{'loss': -0.0901, 'grad_norm': 1.7235727310180664, 'learning_rate': 5.5424242424242425e-06, 'num_tokens': 2602192.0, 'completions/mean_length': 219.875, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 44.60000228881836, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.5801786780357361, 'rewards/meter/std': 0.3490472435951233, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8448355197906494, 'rewards/lexical_plausibility/std': 0.22178836166858673, 'rewards/judge_quality/mean': 0.2874999940395355, 'rewards/judge_quality/std': 0.19955307245254517, 'rewards/repeat_penalty/mean': 0.8767648339271545, 'rewards/repeat_penalty/std': 0.09850464761257172, 'rewards/repeat_floor/mean': 0.9871397614479065, 'rewards/repeat_floor/std': 0.009073498658835888, 'rewards/near_duplicate_penalty/mean': 0.9392648339271545, 'rewards/near_duplicate_penalty/std': 0.07061699777841568, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.21015127003192902, 'rewards/total_composite/std': 0.17606578767299652, 'reward': 0.21015127003192902, 'reward_std': 0.17606577277183533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1457851082086563, 'sampling/sampling_logp_difference/max': 1.569849967956543, 'sampling/importance_sampling_ratio/min': 0.2080763876438141, 'sampling/importance_sampling_ratio/mean': 1.0451232194900513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7417897805571556, 'clip_ratio/low_mean': 0.027027027681469917, 'clip_ratio/low_min': 0.027027027681469917, 'clip_ratio/high_mean': 0.07155514415353537, 'clip_ratio/high_max': 0.07155514415353537, 'clip_ratio/region_mean': 0.09858217183500528, 'reward_total_mean': 0.21015127003192902, 'reward_meter_mean': 0.5801786780357361, 'reward_meter_std': 0.3490472435951233, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8448355197906494, 'reward_lexical_plausibility_std': 0.22178836166858673, 'reward_repeat_penalty_mean': 0.8767648339271545, 'reward_repeat_penalty_std': 0.09850464761257172, 'reward_repeat_floor_mean': 0.9871397614479065, 'reward_repeat_floor_std': 0.009073498658835888, 'reward_near_duplicate_penalty_mean': 0.9392648339271545, 'reward_near_duplicate_penalty_std': 0.07061699777841568, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.2874999940395355, 'reward_judge_quality_std': 0.19955307245254517, 'reward_total_composite_mean': 0.21015127003192902, 'reward_total_composite_std': 0.17606578767299652, 'epoch': 0.08}

 45%|████▍     | 1472/3300 [5:38:34<5:34:08, 10.97s/it]INFO 04-12 21:37:00 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1473/3300 [5:38:47<5:56:18, 11.70s/it]2026-04-12 21:37:13,926 | INFO | train_grpo_train | metrics_logged mode=train step=1473

                                                       
{'loss': -0.0381, 'grad_norm': 4.544206142425537, 'learning_rate': 5.53939393939394e-06, 'num_tokens': 2603787.0, 'completions/mean_length': 106.375, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.42857360839844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.3125974237918854, 'rewards/meter/std': 0.3656372129917145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.942984938621521, 'rewards/lexical_plausibility/std': 0.16126291453838348, 'rewards/judge_quality/mean': 0.7599999904632568, 'rewards/judge_quality/std': 0.3326946496963501, 'rewards/repeat_penalty/mean': 0.9846710562705994, 'rewards/repeat_penalty/std': 0.025053514167666435, 'rewards/repeat_floor/mean': 0.9982251524925232, 'rewards/repeat_floor/std': 0.0025504808872938156, 'rewards/near_duplicate_penalty/mean': 0.9934123158454895, 'rewards/near_duplicate_penalty/std': 0.012166498228907585, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.2731981575489044, 'rewards/total_composite/std': 0.34586209058761597, 'reward': 0.2731981575489044, 'reward_std': 0.34586209058761597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13730488717556, 'sampling/sampling_logp_difference/max': 1.5517511367797852, 'sampling/importance_sampling_ratio/min': 0.21187661588191986, 'sampling/importance_sampling_ratio/mean': 1.007899284362793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8874154984951019, 'clip_ratio/low_mean': 0.06453917641192675, 'clip_ratio/low_min': 0.06453917641192675, 'clip_ratio/high_mean': 0.03988925367593765, 'clip_ratio/high_max': 0.03988925367593765, 'clip_ratio/region_mean': 0.1044284300878644, 'reward_total_mean': 0.2731981575489044, 'reward_meter_mean': 0.3125974237918854, 'reward_meter_std': 0.3656372129917145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.942984938621521, 'reward_lexical_plausibility_std': 0.16126291453838348, 'reward_repeat_penalty_mean': 0.9846710562705994, 'reward_repeat_penalty_std': 0.025053514167666435, 'reward_repeat_floor_mean': 0.9982251524925232, 'reward_repeat_floor_std': 0.0025504808872938156, 'reward_near_duplicate_penalty_mean': 0.9934123158454895, 'reward_near_duplicate_penalty_std': 0.012166498228907585, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.7599999904632568, 'reward_judge_quality_std': 0.3326946496963501, 'reward_total_composite_mean': 0.2731981575489044, 'reward_total_composite_std': 0.34586209058761597, 'epoch': 0.08}

 45%|████▍     | 1473/3300 [5:38:47<5:56:18, 11.70s/it]INFO 04-12 21:37:14 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1474/3300 [5:38:56<5:26:05, 10.72s/it]2026-04-12 21:37:22,336 | INFO | train_grpo_train | metrics_logged mode=train step=1474

                                                       
{'loss': 0.0519, 'grad_norm': 16.62087059020996, 'learning_rate': 5.536363636363636e-06, 'num_tokens': 2605207.0, 'completions/mean_length': 29.5, 'completions/min_length': 20.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.29854297637939453, 'rewards/meter/std': 0.39993488788604736, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8875000476837158, 'rewards/count_floor/std': 0.15526476502418518, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5549999475479126, 'rewards/judge_quality/std': 0.22790977358818054, 'rewards/repeat_penalty/mean': 0.9033284783363342, 'rewards/repeat_penalty/std': 0.12703187763690948, 'rewards/repeat_floor/mean': 0.9939367771148682, 'rewards/repeat_floor/std': 0.007424952462315559, 'rewards/near_duplicate_penalty/mean': 0.9970784783363342, 'rewards/near_duplicate_penalty/std': 0.0046976786106824875, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.11308514326810837, 'rewards/total_composite/std': 0.15752579271793365, 'reward': 0.11308514326810837, 'reward_std': 0.15752580761909485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16352230310440063, 'sampling/sampling_logp_difference/max': 1.9650976657867432, 'sampling/importance_sampling_ratio/min': 0.14014220237731934, 'sampling/importance_sampling_ratio/mean': 1.01524019241333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8454306125640869, 'clip_ratio/low_mean': 0.11931326892226934, 'clip_ratio/low_min': 0.11931326892226934, 'clip_ratio/high_mean': 0.031746032647788525, 'clip_ratio/high_max': 0.031746032647788525, 'clip_ratio/region_mean': 0.15105930157005787, 'reward_total_mean': 0.11308514326810837, 'reward_meter_mean': 0.29854297637939453, 'reward_meter_std': 0.39993488788604736, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8875000476837158, 'reward_count_floor_std': 0.15526476502418518, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9033284783363342, 'reward_repeat_penalty_std': 0.12703187763690948, 'reward_repeat_floor_mean': 0.9939367771148682, 'reward_repeat_floor_std': 0.007424952462315559, 'reward_near_duplicate_penalty_mean': 0.9970784783363342, 'reward_near_duplicate_penalty_std': 0.0046976786106824875, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5549999475479126, 'reward_judge_quality_std': 0.22790977358818054, 'reward_total_composite_mean': 0.11308514326810837, 'reward_total_composite_std': 0.15752579271793365, 'epoch': 0.08}

 45%|████▍     | 1474/3300 [5:38:56<5:26:05, 10.72s/it]INFO 04-12 21:37:22 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1475/3300 [5:39:06<5:26:20, 10.73s/it]2026-04-12 21:37:33,097 | INFO | train_grpo_train | metrics_logged mode=train step=1475

                                                       
{'loss': 0.0404, 'grad_norm': 9.530497550964355, 'learning_rate': 5.533333333333334e-06, 'num_tokens': 2608300.0, 'completions/mean_length': 133.625, 'completions/min_length': 122.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.625, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.9528020620346069, 'rewards/meter/std': 0.08151319622993469, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.0578637570142746, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8968750238418579, 'rewards/count_floor/std': 0.017359137535095215, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9439680576324463, 'rewards/repeat_penalty/std': 0.04404127225279808, 'rewards/repeat_floor/mean': 0.9939197301864624, 'rewards/repeat_floor/std': 0.004599173087626696, 'rewards/near_duplicate_penalty/mean': 0.9840735197067261, 'rewards/near_duplicate_penalty/std': 0.0099300816655159, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9589853286743164, 'rewards/distinct_2/std': 0.03691066429018974, 'rewards/total_composite/mean': 0.33041584491729736, 'rewards/total_composite/std': 0.061515677720308304, 'reward': 0.33041584491729736, 'reward_std': 0.0615156814455986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14891010522842407, 'sampling/sampling_logp_difference/max': 2.926743984222412, 'sampling/importance_sampling_ratio/min': 0.0535711832344532, 'sampling/importance_sampling_ratio/mean': 1.0132691860198975, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9275376498699188, 'clip_ratio/low_mean': 0.0870408657938242, 'clip_ratio/low_min': 0.0870408657938242, 'clip_ratio/high_mean': 0.04037346690893173, 'clip_ratio/high_max': 0.04037346690893173, 'clip_ratio/region_mean': 0.12741433270275593, 'reward_total_mean': 0.33041584491729736, 'reward_meter_mean': 0.9528020620346069, 'reward_meter_std': 0.08151319622993469, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.0578637570142746, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8968750238418579, 'reward_count_floor_std': 0.017359137535095215, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9439680576324463, 'reward_repeat_penalty_std': 0.04404127225279808, 'reward_repeat_floor_mean': 0.9939197301864624, 'reward_repeat_floor_std': 0.004599173087626696, 'reward_near_duplicate_penalty_mean': 0.9840735197067261, 'reward_near_duplicate_penalty_std': 0.0099300816655159, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9589853286743164, 'reward_distinct_2_std': 0.03691066429018974, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.33041584491729736, 'reward_total_composite_std': 0.061515677720308304, 'epoch': 0.08}

 45%|████▍     | 1475/3300 [5:39:06<5:26:20, 10.73s/it]INFO 04-12 21:37:33 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1476/3300 [5:39:20<5:55:20, 11.69s/it]2026-04-12 21:37:47,024 | INFO | train_grpo_train | metrics_logged mode=train step=1476

                                                       
{'loss': -0.0059, 'grad_norm': 2.84600567817688, 'learning_rate': 5.530303030303031e-06, 'num_tokens': 2609522.0, 'completions/mean_length': 144.75, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 22.33333396911621, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7057888507843018, 'rewards/meter/std': 0.3767654597759247, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.8561461567878723, 'rewards/lexical_plausibility/std': 0.2663852572441101, 'rewards/judge_quality/mean': 0.32124999165534973, 'rewards/judge_quality/std': 0.2923519015312195, 'rewards/repeat_penalty/mean': 0.8104128241539001, 'rewards/repeat_penalty/std': 0.16286340355873108, 'rewards/repeat_floor/mean': 0.9842416048049927, 'rewards/repeat_floor/std': 0.01870684139430523, 'rewards/near_duplicate_penalty/mean': 0.9574443101882935, 'rewards/near_duplicate_penalty/std': 0.10259455442428589, 'rewards/opening_diversity/mean': 0.84375, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2286793738603592, 'rewards/total_composite/std': 0.29948046803474426, 'reward': 0.2286793738603592, 'reward_std': 0.29948046803474426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15166312456130981, 'sampling/sampling_logp_difference/max': 1.2029943466186523, 'sampling/importance_sampling_ratio/min': 0.3002936840057373, 'sampling/importance_sampling_ratio/mean': 1.0375717878341675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0161238610744476, 'clip_ratio/low_mean': 0.06118742423132062, 'clip_ratio/low_min': 0.06118742423132062, 'clip_ratio/high_mean': 0.014705882407724857, 'clip_ratio/high_max': 0.014705882407724857, 'clip_ratio/region_mean': 0.07589330663904548, 'reward_total_mean': 0.2286793738603592, 'reward_meter_mean': 0.7057888507843018, 'reward_meter_std': 0.3767654597759247, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.8561461567878723, 'reward_lexical_plausibility_std': 0.2663852572441101, 'reward_repeat_penalty_mean': 0.8104128241539001, 'reward_repeat_penalty_std': 0.16286340355873108, 'reward_repeat_floor_mean': 0.9842416048049927, 'reward_repeat_floor_std': 0.01870684139430523, 'reward_near_duplicate_penalty_mean': 0.9574443101882935, 'reward_near_duplicate_penalty_std': 0.10259455442428589, 'reward_opening_diversity_mean': 0.84375, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.32124999165534973, 'reward_judge_quality_std': 0.2923519015312195, 'reward_total_composite_mean': 0.2286793738603592, 'reward_total_composite_std': 0.29948046803474426, 'epoch': 0.08}

 45%|████▍     | 1476/3300 [5:39:20<5:55:20, 11.69s/it]INFO 04-12 21:37:47 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1477/3300 [5:39:29<5:24:39, 10.69s/it]2026-04-12 21:37:55,377 | INFO | train_grpo_train | metrics_logged mode=train step=1477

                                                       
{'loss': 0.0347, 'grad_norm': 23.032712936401367, 'learning_rate': 5.527272727272728e-06, 'num_tokens': 2610904.0, 'completions/mean_length': 20.75, 'completions/min_length': 17.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.37311288714408875, 'rewards/meter/std': 0.47168493270874023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.953125, 'rewards/lexical_plausibility/std': 0.13258251547813416, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.2796298861503601, 'rewards/repeat_penalty/mean': 0.7299373149871826, 'rewards/repeat_penalty/std': 0.04513396695256233, 'rewards/repeat_floor/mean': 0.9822695255279541, 'rewards/repeat_floor/std': 0.005590358749032021, 'rewards/near_duplicate_penalty/mean': 0.994617223739624, 'rewards/near_duplicate_penalty/std': 0.015224790200591087, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.1464983969926834, 'rewards/total_composite/std': 0.17938017845153809, 'reward': 0.1464983969926834, 'reward_std': 0.17938017845153809, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15937772393226624, 'sampling/sampling_logp_difference/max': 1.2131202220916748, 'sampling/importance_sampling_ratio/min': 0.2972683012485504, 'sampling/importance_sampling_ratio/mean': 1.016886591911316, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8530717939138412, 'clip_ratio/low_mean': 0.09069953998550773, 'clip_ratio/low_min': 0.09069953998550773, 'clip_ratio/high_mean': 0.06439394131302834, 'clip_ratio/high_max': 0.06439394131302834, 'clip_ratio/region_mean': 0.15509348129853606, 'reward_total_mean': 0.1464983969926834, 'reward_meter_mean': 0.37311288714408875, 'reward_meter_std': 0.47168493270874023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.953125, 'reward_lexical_plausibility_std': 0.13258251547813416, 'reward_repeat_penalty_mean': 0.7299373149871826, 'reward_repeat_penalty_std': 0.04513396695256233, 'reward_repeat_floor_mean': 0.9822695255279541, 'reward_repeat_floor_std': 0.005590358749032021, 'reward_near_duplicate_penalty_mean': 0.994617223739624, 'reward_near_duplicate_penalty_std': 0.015224790200591087, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.2796298861503601, 'reward_total_composite_mean': 0.1464983969926834, 'reward_total_composite_std': 0.17938017845153809, 'epoch': 0.08}

 45%|████▍     | 1477/3300 [5:39:29<5:24:39, 10.69s/it]INFO 04-12 21:37:55 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1478/3300 [5:39:43<5:59:53, 11.85s/it]2026-04-12 21:38:09,926 | INFO | train_grpo_train | metrics_logged mode=train step=1478

                                                       
{'loss': -0.158, 'grad_norm': 4.232932090759277, 'learning_rate': 5.524242424242424e-06, 'num_tokens': 2613641.0, 'completions/mean_length': 197.125, 'completions/min_length': 137.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 152.1428680419922, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.7354157567024231, 'rewards/meter/std': 0.16244816780090332, 'rewards/count_adherence/mean': 0.7625000476837158, 'rewards/count_adherence/std': 0.05175492912530899, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9287500381469727, 'rewards/count_floor/std': 0.015526460483670235, 'rewards/lexical_plausibility/mean': 0.9670959711074829, 'rewards/lexical_plausibility/std': 0.0930667370557785, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.6656760573387146, 'rewards/repeat_penalty/std': 0.16788245737552643, 'rewards/repeat_floor/mean': 0.9651086330413818, 'rewards/repeat_floor/std': 0.017652025446295738, 'rewards/near_duplicate_penalty/mean': 0.9291567802429199, 'rewards/near_duplicate_penalty/std': 0.02458607591688633, 'rewards/opening_diversity/mean': 0.8861607313156128, 'rewards/opening_diversity/std': 0.0864999070763588, 'rewards/distinct_2/mean': 0.8062828183174133, 'rewards/distinct_2/std': 0.11577828228473663, 'rewards/total_composite/mean': 0.19909149408340454, 'rewards/total_composite/std': 0.04298942908644676, 'reward': 0.19909149408340454, 'reward_std': 0.04298942908644676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1250392496585846, 'sampling/sampling_logp_difference/max': 2.6524553298950195, 'sampling/importance_sampling_ratio/min': 0.07047795504331589, 'sampling/importance_sampling_ratio/mean': 1.0133577585220337, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6019513010978699, 'clip_ratio/low_mean': 0.04865333251655102, 'clip_ratio/low_min': 0.04865333251655102, 'clip_ratio/high_mean': 0.06369003001600504, 'clip_ratio/high_max': 0.06369003001600504, 'clip_ratio/region_mean': 0.11234336253255606, 'reward_total_mean': 0.19909149408340454, 'reward_meter_mean': 0.7354157567024231, 'reward_meter_std': 0.16244816780090332, 'reward_count_adherence_mean': 0.7625000476837158, 'reward_count_adherence_std': 0.05175492912530899, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9287500381469727, 'reward_count_floor_std': 0.015526460483670235, 'reward_lexical_plausibility_mean': 0.9670959711074829, 'reward_lexical_plausibility_std': 0.0930667370557785, 'reward_repeat_penalty_mean': 0.6656760573387146, 'reward_repeat_penalty_std': 0.16788245737552643, 'reward_repeat_floor_mean': 0.9651086330413818, 'reward_repeat_floor_std': 0.017652025446295738, 'reward_near_duplicate_penalty_mean': 0.9291567802429199, 'reward_near_duplicate_penalty_std': 0.02458607591688633, 'reward_opening_diversity_mean': 0.8861607313156128, 'reward_opening_diversity_std': 0.0864999070763588, 'reward_distinct_2_mean': 0.8062828183174133, 'reward_distinct_2_std': 0.11577828228473663, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.19909149408340454, 'reward_total_composite_std': 0.04298942908644676, 'epoch': 0.08}

 45%|████▍     | 1478/3300 [5:39:43<5:59:53, 11.85s/it]INFO 04-12 21:38:10 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1479/3300 [5:39:53<5:38:37, 11.16s/it]2026-04-12 21:38:19,465 | INFO | train_grpo_train | metrics_logged mode=train step=1479

                                                       
{'loss': 0.0326, 'grad_norm': 12.372693061828613, 'learning_rate': 5.521212121212122e-06, 'num_tokens': 2615541.0, 'completions/mean_length': 65.5, 'completions/min_length': 51.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.6939051747322083, 'rewards/meter/std': 0.30842527747154236, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4375, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8855779767036438, 'rewards/repeat_penalty/std': 0.0997515395283699, 'rewards/repeat_floor/mean': 0.987216591835022, 'rewards/repeat_floor/std': 0.010126245208084583, 'rewards/near_duplicate_penalty/mean': 0.9627809524536133, 'rewards/near_duplicate_penalty/std': 0.020330019295215607, 'rewards/opening_diversity/mean': 0.9921875, 'rewards/opening_diversity/std': 0.022097086533904076, 'rewards/distinct_2/mean': 0.9252022504806519, 'rewards/distinct_2/std': 0.08017771691083908, 'rewards/total_composite/mean': 0.30763912200927734, 'rewards/total_composite/std': 0.18136948347091675, 'reward': 0.30763912200927734, 'reward_std': 0.18136949837207794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15677587687969208, 'sampling/sampling_logp_difference/max': 2.489776134490967, 'sampling/importance_sampling_ratio/min': 0.08292853087186813, 'sampling/importance_sampling_ratio/mean': 1.008183240890503, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8283580169081688, 'clip_ratio/low_mean': 0.07551028952002525, 'clip_ratio/low_min': 0.07551028952002525, 'clip_ratio/high_mean': 0.0631263256072998, 'clip_ratio/high_max': 0.0631263256072998, 'clip_ratio/region_mean': 0.13863661512732506, 'reward_total_mean': 0.30763912200927734, 'reward_meter_mean': 0.6939051747322083, 'reward_meter_std': 0.30842527747154236, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8855779767036438, 'reward_repeat_penalty_std': 0.0997515395283699, 'reward_repeat_floor_mean': 0.987216591835022, 'reward_repeat_floor_std': 0.010126245208084583, 'reward_near_duplicate_penalty_mean': 0.9627809524536133, 'reward_near_duplicate_penalty_std': 0.020330019295215607, 'reward_opening_diversity_mean': 0.9921875, 'reward_opening_diversity_std': 0.022097086533904076, 'reward_distinct_2_mean': 0.9252022504806519, 'reward_distinct_2_std': 0.08017771691083908, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.30763912200927734, 'reward_total_composite_std': 0.18136948347091675, 'epoch': 0.08}

 45%|████▍     | 1479/3300 [5:39:53<5:38:37, 11.16s/it]INFO 04-12 21:38:19 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1480/3300 [5:40:02<5:17:59, 10.48s/it]2026-04-12 21:38:28,369 | INFO | train_grpo_train | metrics_logged mode=train step=1480

                                                       
{'loss': 0.0687, 'grad_norm': 13.060455322265625, 'learning_rate': 5.518181818181818e-06, 'num_tokens': 2617105.0, 'completions/mean_length': 37.5, 'completions/min_length': 31.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9345614910125732, 'rewards/meter/std': 0.07272852957248688, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5462499856948853, 'rewards/judge_quality/std': 0.18392060697078705, 'rewards/repeat_penalty/mean': 0.85737144947052, 'rewards/repeat_penalty/std': 0.08898493647575378, 'rewards/repeat_floor/mean': 0.9819572567939758, 'rewards/repeat_floor/std': 0.010493194684386253, 'rewards/near_duplicate_penalty/mean': 0.923769474029541, 'rewards/near_duplicate_penalty/std': 0.03479602187871933, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9265761971473694, 'rewards/distinct_2/std': 0.06919492781162262, 'rewards/total_composite/mean': 0.4943440854549408, 'rewards/total_composite/std': 0.1385333091020584, 'reward': 0.4943440854549408, 'reward_std': 0.1385333389043808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12273003160953522, 'sampling/sampling_logp_difference/max': 1.4212586879730225, 'sampling/importance_sampling_ratio/min': 0.24140998721122742, 'sampling/importance_sampling_ratio/mean': 1.0026049613952637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6941500902175903, 'clip_ratio/low_mean': 0.07864060625433922, 'clip_ratio/low_min': 0.07864060625433922, 'clip_ratio/high_mean': 0.01785714295692742, 'clip_ratio/high_max': 0.01785714295692742, 'clip_ratio/region_mean': 0.09649774921126664, 'reward_total_mean': 0.4943440854549408, 'reward_meter_mean': 0.9345614910125732, 'reward_meter_std': 0.07272852957248688, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.85737144947052, 'reward_repeat_penalty_std': 0.08898493647575378, 'reward_repeat_floor_mean': 0.9819572567939758, 'reward_repeat_floor_std': 0.010493194684386253, 'reward_near_duplicate_penalty_mean': 0.923769474029541, 'reward_near_duplicate_penalty_std': 0.03479602187871933, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9265761971473694, 'reward_distinct_2_std': 0.06919492781162262, 'reward_judge_quality_mean': 0.5462499856948853, 'reward_judge_quality_std': 0.18392060697078705, 'reward_total_composite_mean': 0.4943440854549408, 'reward_total_composite_std': 0.1385333091020584, 'epoch': 0.08}

 45%|████▍     | 1480/3300 [5:40:02<5:17:59, 10.48s/it]INFO 04-12 21:38:28 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1481/3300 [5:40:10<4:55:53,  9.76s/it]2026-04-12 21:38:36,446 | INFO | train_grpo_train | metrics_logged mode=train step=1481

                                                       
{'loss': 0.0926, 'grad_norm': 20.302675247192383, 'learning_rate': 5.515151515151515e-06, 'num_tokens': 2618461.0, 'completions/mean_length': 30.5, 'completions/min_length': 27.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8041801452636719, 'rewards/meter/std': 0.24927420914173126, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.976553738117218, 'rewards/repeat_penalty/std': 0.03615066409111023, 'rewards/repeat_floor/mean': 0.9969298243522644, 'rewards/repeat_floor/std': 0.004829831421375275, 'rewards/near_duplicate_penalty/mean': 0.9841476082801819, 'rewards/near_duplicate_penalty/std': 0.029976798221468925, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923076629638672, 'rewards/distinct_2/std': 0.021757129579782486, 'rewards/total_composite/mean': 0.38019245862960815, 'rewards/total_composite/std': 0.13928818702697754, 'reward': 0.38019245862960815, 'reward_std': 0.13928818702697754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18236510455608368, 'sampling/sampling_logp_difference/max': 1.8239991664886475, 'sampling/importance_sampling_ratio/min': 0.16137908399105072, 'sampling/importance_sampling_ratio/mean': 0.9616653323173523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8683034181594849, 'clip_ratio/low_mean': 0.053973354399204254, 'clip_ratio/low_min': 0.053973354399204254, 'clip_ratio/high_mean': 0.13541666604578495, 'clip_ratio/high_max': 0.13541666604578495, 'clip_ratio/region_mean': 0.1893900204449892, 'reward_total_mean': 0.38019245862960815, 'reward_meter_mean': 0.8041801452636719, 'reward_meter_std': 0.24927420914173126, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.976553738117218, 'reward_repeat_penalty_std': 0.03615066409111023, 'reward_repeat_floor_mean': 0.9969298243522644, 'reward_repeat_floor_std': 0.004829831421375275, 'reward_near_duplicate_penalty_mean': 0.9841476082801819, 'reward_near_duplicate_penalty_std': 0.029976798221468925, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923076629638672, 'reward_distinct_2_std': 0.021757129579782486, 'reward_judge_quality_mean': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.38019245862960815, 'reward_total_composite_std': 0.13928818702697754, 'epoch': 0.08}

 45%|████▍     | 1481/3300 [5:40:10<4:55:53,  9.76s/it]INFO 04-12 21:38:36 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1482/3300 [5:40:23<5:28:09, 10.83s/it]2026-04-12 21:38:49,771 | INFO | train_grpo_train | metrics_logged mode=train step=1482

                                                       
{'loss': -0.0613, 'grad_norm': 12.659420013427734, 'learning_rate': 5.512121212121213e-06, 'num_tokens': 2619834.0, 'completions/mean_length': 98.625, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.57143020629883, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6670266389846802, 'rewards/meter/std': 0.3609144687652588, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9494751691818237, 'rewards/lexical_plausibility/std': 0.14290589094161987, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.8963013291358948, 'rewards/repeat_penalty/std': 0.1631539762020111, 'rewards/repeat_floor/mean': 0.9881108403205872, 'rewards/repeat_floor/std': 0.01716705597937107, 'rewards/near_duplicate_penalty/mean': 0.9567710757255554, 'rewards/near_duplicate_penalty/std': 0.03095054067671299, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.9691130518913269, 'rewards/distinct_2/std': 0.08568475395441055, 'rewards/total_composite/mean': 0.24383217096328735, 'rewards/total_composite/std': 0.1715419441461563, 'reward': 0.24383217096328735, 'reward_std': 0.1715419441461563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1116553246974945, 'sampling/sampling_logp_difference/max': 1.348264217376709, 'sampling/importance_sampling_ratio/min': 0.25969064235687256, 'sampling/importance_sampling_ratio/mean': 1.0265294313430786, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7966594099998474, 'clip_ratio/low_mean': 0.036683558486402035, 'clip_ratio/low_min': 0.036683558486402035, 'clip_ratio/high_mean': 0.06691124010831118, 'clip_ratio/high_max': 0.06691124010831118, 'clip_ratio/region_mean': 0.10359479859471321, 'reward_total_mean': 0.24383217096328735, 'reward_meter_mean': 0.6670266389846802, 'reward_meter_std': 0.3609144687652588, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9494751691818237, 'reward_lexical_plausibility_std': 0.14290589094161987, 'reward_repeat_penalty_mean': 0.8963013291358948, 'reward_repeat_penalty_std': 0.1631539762020111, 'reward_repeat_floor_mean': 0.9881108403205872, 'reward_repeat_floor_std': 0.01716705597937107, 'reward_near_duplicate_penalty_mean': 0.9567710757255554, 'reward_near_duplicate_penalty_std': 0.03095054067671299, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.9691130518913269, 'reward_distinct_2_std': 0.08568475395441055, 'reward_judge_quality_mean': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.24383217096328735, 'reward_total_composite_std': 0.1715419441461563, 'epoch': 0.08}

 45%|████▍     | 1482/3300 [5:40:23<5:28:09, 10.83s/it]INFO 04-12 21:38:50 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1483/3300 [5:40:31<5:04:23, 10.05s/it]2026-04-12 21:38:58,012 | INFO | train_grpo_train | metrics_logged mode=train step=1483

                                                       
{'loss': 0.0302, 'grad_norm': 15.642985343933105, 'learning_rate': 5.50909090909091e-06, 'num_tokens': 2621541.0, 'completions/mean_length': 53.375, 'completions/min_length': 50.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.4391917586326599, 'rewards/meter/std': 0.31837061047554016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/repeat_penalty/mean': 0.8100943565368652, 'rewards/repeat_penalty/std': 0.1832367479801178, 'rewards/repeat_floor/mean': 0.981307864189148, 'rewards/repeat_floor/std': 0.01715531386435032, 'rewards/near_duplicate_penalty/mean': 0.9617388248443604, 'rewards/near_duplicate_penalty/std': 0.028309646993875504, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.918578565120697, 'rewards/distinct_2/std': 0.0732157975435257, 'rewards/total_composite/mean': 0.3797881305217743, 'rewards/total_composite/std': 0.3076392710208893, 'reward': 0.3797881305217743, 'reward_std': 0.3076392412185669, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12488719820976257, 'sampling/sampling_logp_difference/max': 1.6914520263671875, 'sampling/importance_sampling_ratio/min': 0.1842517852783203, 'sampling/importance_sampling_ratio/mean': 1.0219050645828247, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6671174019575119, 'clip_ratio/low_mean': 0.06702635623514652, 'clip_ratio/low_min': 0.06702635623514652, 'clip_ratio/high_mean': 0.06447712425142527, 'clip_ratio/high_max': 0.06447712425142527, 'clip_ratio/region_mean': 0.1315034804865718, 'reward_total_mean': 0.3797881305217743, 'reward_meter_mean': 0.4391917586326599, 'reward_meter_std': 0.31837061047554016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8100943565368652, 'reward_repeat_penalty_std': 0.1832367479801178, 'reward_repeat_floor_mean': 0.981307864189148, 'reward_repeat_floor_std': 0.01715531386435032, 'reward_near_duplicate_penalty_mean': 0.9617388248443604, 'reward_near_duplicate_penalty_std': 0.028309646993875504, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.918578565120697, 'reward_distinct_2_std': 0.0732157975435257, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.3797881305217743, 'reward_total_composite_std': 0.3076392710208893, 'epoch': 0.08}

 45%|████▍     | 1483/3300 [5:40:31<5:04:23, 10.05s/it]INFO 04-12 21:38:58 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1484/3300 [5:40:42<5:07:37, 10.16s/it]2026-04-12 21:39:08,442 | INFO | train_grpo_train | metrics_logged mode=train step=1484

                                                       
{'loss': 0.0481, 'grad_norm': 9.162497520446777, 'learning_rate': 5.506060606060607e-06, 'num_tokens': 2624119.0, 'completions/mean_length': 119.25, 'completions/min_length': 103.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.25, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.8776699900627136, 'rewards/meter/std': 0.16596528887748718, 'rewards/count_adherence/mean': 0.671875, 'rewards/count_adherence/std': 0.06469365209341049, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.901562511920929, 'rewards/count_floor/std': 0.019408106803894043, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9527665376663208, 'rewards/repeat_penalty/std': 0.040672667324543, 'rewards/repeat_floor/mean': 0.9944452047348022, 'rewards/repeat_floor/std': 0.004806766752153635, 'rewards/near_duplicate_penalty/mean': 0.9795374870300293, 'rewards/near_duplicate_penalty/std': 0.019689787179231644, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9723844528198242, 'rewards/distinct_2/std': 0.026663072407245636, 'rewards/total_composite/mean': 0.31343966722488403, 'rewards/total_composite/std': 0.06773490458726883, 'reward': 0.31343966722488403, 'reward_std': 0.06773491203784943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1395310014486313, 'sampling/sampling_logp_difference/max': 2.542202949523926, 'sampling/importance_sampling_ratio/min': 0.07869285345077515, 'sampling/importance_sampling_ratio/mean': 1.0206793546676636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8869970813393593, 'clip_ratio/low_mean': 0.07193025480955839, 'clip_ratio/low_min': 0.07193025480955839, 'clip_ratio/high_mean': 0.06542260386049747, 'clip_ratio/high_max': 0.06542260386049747, 'clip_ratio/region_mean': 0.13735285867005587, 'reward_total_mean': 0.31343966722488403, 'reward_meter_mean': 0.8776699900627136, 'reward_meter_std': 0.16596528887748718, 'reward_count_adherence_mean': 0.671875, 'reward_count_adherence_std': 0.06469365209341049, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.901562511920929, 'reward_count_floor_std': 0.019408106803894043, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9527665376663208, 'reward_repeat_penalty_std': 0.040672667324543, 'reward_repeat_floor_mean': 0.9944452047348022, 'reward_repeat_floor_std': 0.004806766752153635, 'reward_near_duplicate_penalty_mean': 0.9795374870300293, 'reward_near_duplicate_penalty_std': 0.019689787179231644, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9723844528198242, 'reward_distinct_2_std': 0.026663072407245636, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.05345224589109421, 'reward_total_composite_mean': 0.31343966722488403, 'reward_total_composite_std': 0.06773490458726883, 'epoch': 0.08}

 45%|████▍     | 1484/3300 [5:40:42<5:07:37, 10.16s/it]INFO 04-12 21:39:08 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1485/3300 [5:40:55<5:37:20, 11.15s/it]2026-04-12 21:39:21,899 | INFO | train_grpo_train | metrics_logged mode=train step=1485

                                                       
{'loss': -0.0462, 'grad_norm': 3.93593692779541, 'learning_rate': 5.5030303030303034e-06, 'num_tokens': 2625619.0, 'completions/mean_length': 164.5, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 48.66666793823242, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5685986280441284, 'rewards/meter/std': 0.3968857228755951, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8860464692115784, 'rewards/lexical_plausibility/std': 0.22089797258377075, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.155264750123024, 'rewards/repeat_penalty/mean': 0.8654284477233887, 'rewards/repeat_penalty/std': 0.13551020622253418, 'rewards/repeat_floor/mean': 0.9869046211242676, 'rewards/repeat_floor/std': 0.011987159959971905, 'rewards/near_duplicate_penalty/mean': 0.9661963582038879, 'rewards/near_duplicate_penalty/std': 0.027471110224723816, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9525013566017151, 'rewards/distinct_2/std': 0.053281817585229874, 'rewards/total_composite/mean': 0.12828464806079865, 'rewards/total_composite/std': 0.15320289134979248, 'reward': 0.12828464806079865, 'reward_std': 0.15320289134979248, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14334425330162048, 'sampling/sampling_logp_difference/max': 1.959779977798462, 'sampling/importance_sampling_ratio/min': 0.14088942110538483, 'sampling/importance_sampling_ratio/mean': 1.039678692817688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.930069625377655, 'clip_ratio/low_mean': 0.054975331760942936, 'clip_ratio/low_min': 0.054975331760942936, 'clip_ratio/high_mean': 0.033018868416547775, 'clip_ratio/high_max': 0.033018868416547775, 'clip_ratio/region_mean': 0.08799420017749071, 'reward_total_mean': 0.12828464806079865, 'reward_meter_mean': 0.5685986280441284, 'reward_meter_std': 0.3968857228755951, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8860464692115784, 'reward_lexical_plausibility_std': 0.22089797258377075, 'reward_repeat_penalty_mean': 0.8654284477233887, 'reward_repeat_penalty_std': 0.13551020622253418, 'reward_repeat_floor_mean': 0.9869046211242676, 'reward_repeat_floor_std': 0.011987159959971905, 'reward_near_duplicate_penalty_mean': 0.9661963582038879, 'reward_near_duplicate_penalty_std': 0.027471110224723816, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9525013566017151, 'reward_distinct_2_std': 0.053281817585229874, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.155264750123024, 'reward_total_composite_mean': 0.12828464806079865, 'reward_total_composite_std': 0.15320289134979248, 'epoch': 0.08}

 45%|████▌     | 1485/3300 [5:40:55<5:37:20, 11.15s/it]INFO 04-12 21:39:22 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1486/3300 [5:41:08<5:56:21, 11.79s/it]2026-04-12 21:39:35,166 | INFO | train_grpo_train | metrics_logged mode=train step=1486

                                                       
{'loss': -0.0118, 'grad_norm': 1.9544090032577515, 'learning_rate': 5.500000000000001e-06, 'num_tokens': 2627037.0, 'completions/mean_length': 148.25, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6240063309669495, 'rewards/meter/std': 0.43916282057762146, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8500000238418579, 'rewards/count_floor/std': 0.16035674512386322, 'rewards/lexical_plausibility/mean': 0.873565673828125, 'rewards/lexical_plausibility/std': 0.23518669605255127, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/repeat_penalty/mean': 0.7838403582572937, 'rewards/repeat_penalty/std': 0.1432766318321228, 'rewards/repeat_floor/mean': 0.9840992093086243, 'rewards/repeat_floor/std': 0.01313428208231926, 'rewards/near_duplicate_penalty/mean': 0.9821920990943909, 'rewards/near_duplicate_penalty/std': 0.038857847452163696, 'rewards/opening_diversity/mean': 0.8125, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.978003978729248, 'rewards/distinct_2/std': 0.04075079411268234, 'rewards/total_composite/mean': 0.21810519695281982, 'rewards/total_composite/std': 0.1997922956943512, 'reward': 0.21810519695281982, 'reward_std': 0.1997922956943512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15405714511871338, 'sampling/sampling_logp_difference/max': 0.9355897903442383, 'sampling/importance_sampling_ratio/min': 0.39235439896583557, 'sampling/importance_sampling_ratio/mean': 1.0509032011032104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2151992172002792, 'clip_ratio/low_mean': 0.021042471285909414, 'clip_ratio/low_min': 0.021042471285909414, 'clip_ratio/high_mean': 0.07353148516267538, 'clip_ratio/high_max': 0.07353148516267538, 'clip_ratio/region_mean': 0.0945739564485848, 'reward_total_mean': 0.21810519695281982, 'reward_meter_mean': 0.6240063309669495, 'reward_meter_std': 0.43916282057762146, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8500000238418579, 'reward_count_floor_std': 0.16035674512386322, 'reward_lexical_plausibility_mean': 0.873565673828125, 'reward_lexical_plausibility_std': 0.23518669605255127, 'reward_repeat_penalty_mean': 0.7838403582572937, 'reward_repeat_penalty_std': 0.1432766318321228, 'reward_repeat_floor_mean': 0.9840992093086243, 'reward_repeat_floor_std': 0.01313428208231926, 'reward_near_duplicate_penalty_mean': 0.9821920990943909, 'reward_near_duplicate_penalty_std': 0.038857847452163696, 'reward_opening_diversity_mean': 0.8125, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.978003978729248, 'reward_distinct_2_std': 0.04075079411268234, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.21810519695281982, 'reward_total_composite_std': 0.1997922956943512, 'epoch': 0.08}

 45%|████▌     | 1486/3300 [5:41:09<5:56:21, 11.79s/it]INFO 04-12 21:39:35 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1487/3300 [5:41:22<6:15:51, 12.44s/it]2026-04-12 21:39:49,126 | INFO | train_grpo_train | metrics_logged mode=train step=1487

                                                       
{'loss': -0.1709, 'grad_norm': 3.3017332553863525, 'learning_rate': 5.496969696969697e-06, 'num_tokens': 2629177.0, 'completions/mean_length': 145.5, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 93.14286041259766, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.8524502515792847, 'rewards/meter/std': 0.1989641934633255, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.949999988079071, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9373282194137573, 'rewards/lexical_plausibility/std': 0.1772625595331192, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.12817397713661194, 'rewards/repeat_penalty/mean': 0.7223489284515381, 'rewards/repeat_penalty/std': 0.1824721097946167, 'rewards/repeat_floor/mean': 0.9715101718902588, 'rewards/repeat_floor/std': 0.017117882147431374, 'rewards/near_duplicate_penalty/mean': 0.9337801933288574, 'rewards/near_duplicate_penalty/std': 0.02307550422847271, 'rewards/opening_diversity/mean': 0.9323863983154297, 'rewards/opening_diversity/std': 0.08456561714410782, 'rewards/distinct_2/mean': 0.8388886451721191, 'rewards/distinct_2/std': 0.11587636917829514, 'rewards/total_composite/mean': 0.2683011591434479, 'rewards/total_composite/std': 0.1135934367775917, 'reward': 0.2683011591434479, 'reward_std': 0.1135934367775917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1571616530418396, 'sampling/sampling_logp_difference/max': 3.443772792816162, 'sampling/importance_sampling_ratio/min': 0.03194394335150719, 'sampling/importance_sampling_ratio/mean': 1.0295419692993164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7630656883120537, 'clip_ratio/low_mean': 0.021341463550925255, 'clip_ratio/low_min': 0.021341463550925255, 'clip_ratio/high_mean': 0.09213289897888899, 'clip_ratio/high_max': 0.09213289897888899, 'clip_ratio/region_mean': 0.11347436252981424, 'reward_total_mean': 0.2683011591434479, 'reward_meter_mean': 0.8524502515792847, 'reward_meter_std': 0.1989641934633255, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.949999988079071, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9373282194137573, 'reward_lexical_plausibility_std': 0.1772625595331192, 'reward_repeat_penalty_mean': 0.7223489284515381, 'reward_repeat_penalty_std': 0.1824721097946167, 'reward_repeat_floor_mean': 0.9715101718902588, 'reward_repeat_floor_std': 0.017117882147431374, 'reward_near_duplicate_penalty_mean': 0.9337801933288574, 'reward_near_duplicate_penalty_std': 0.02307550422847271, 'reward_opening_diversity_mean': 0.9323863983154297, 'reward_opening_diversity_std': 0.08456561714410782, 'reward_distinct_2_mean': 0.8388886451721191, 'reward_distinct_2_std': 0.11587636917829514, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.12817397713661194, 'reward_total_composite_mean': 0.2683011591434479, 'reward_total_composite_std': 0.1135934367775917, 'epoch': 0.08}

 45%|████▌     | 1487/3300 [5:41:22<6:15:51, 12.44s/it]INFO 04-12 21:39:49 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1488/3300 [5:41:31<5:45:49, 11.45s/it]2026-04-12 21:39:58,256 | INFO | train_grpo_train | metrics_logged mode=train step=1488

                                                       
{'loss': 0.0485, 'grad_norm': 11.664468765258789, 'learning_rate': 5.493939393939395e-06, 'num_tokens': 2631049.0, 'completions/mean_length': 70.0, 'completions/min_length': 64.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.5656070113182068, 'rewards/meter/std': 0.38506126403808594, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41249996423721313, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.8139246702194214, 'rewards/repeat_penalty/std': 0.09003651887178421, 'rewards/repeat_floor/mean': 0.9799196124076843, 'rewards/repeat_floor/std': 0.010779653675854206, 'rewards/near_duplicate_penalty/mean': 0.9567646384239197, 'rewards/near_duplicate_penalty/std': 0.031647518277168274, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8489433526992798, 'rewards/distinct_2/std': 0.07170760631561279, 'rewards/total_composite/mean': 0.21787984669208527, 'rewards/total_composite/std': 0.15066687762737274, 'reward': 0.21787984669208527, 'reward_std': 0.15066687762737274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16216185688972473, 'sampling/sampling_logp_difference/max': 1.274740219116211, 'sampling/importance_sampling_ratio/min': 0.27950355410575867, 'sampling/importance_sampling_ratio/mean': 1.0116816759109497, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.190341092646122, 'clip_ratio/low_mean': 0.05594387277960777, 'clip_ratio/low_min': 0.05594387277960777, 'clip_ratio/high_mean': 0.12318585999310017, 'clip_ratio/high_max': 0.12318585999310017, 'clip_ratio/region_mean': 0.17912973277270794, 'reward_total_mean': 0.21787984669208527, 'reward_meter_mean': 0.5656070113182068, 'reward_meter_std': 0.38506126403808594, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8139246702194214, 'reward_repeat_penalty_std': 0.09003651887178421, 'reward_repeat_floor_mean': 0.9799196124076843, 'reward_repeat_floor_std': 0.010779653675854206, 'reward_near_duplicate_penalty_mean': 0.9567646384239197, 'reward_near_duplicate_penalty_std': 0.031647518277168274, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8489433526992798, 'reward_distinct_2_std': 0.07170760631561279, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.21787984669208527, 'reward_total_composite_std': 0.15066687762737274, 'epoch': 0.08}

 45%|████▌     | 1488/3300 [5:41:32<5:45:49, 11.45s/it]INFO 04-12 21:39:58 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1489/3300 [5:41:47<6:22:47, 12.68s/it]2026-04-12 21:40:13,905 | INFO | train_grpo_train | metrics_logged mode=train step=1489

                                                       
{'loss': -0.0086, 'grad_norm': 13.06687068939209, 'learning_rate': 5.490909090909091e-06, 'num_tokens': 2632709.0, 'completions/mean_length': 42.5, 'completions/min_length': 37.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.3592393100261688, 'rewards/meter/std': 0.37753790616989136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.731249988079071, 'rewards/judge_quality/std': 0.2623213827610016, 'rewards/repeat_penalty/mean': 0.929060697555542, 'rewards/repeat_penalty/std': 0.05514172464609146, 'rewards/repeat_floor/mean': 0.9909193515777588, 'rewards/repeat_floor/std': 0.0063325450755655766, 'rewards/near_duplicate_penalty/mean': 0.9576809406280518, 'rewards/near_duplicate_penalty/std': 0.02281358651816845, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9696356058120728, 'rewards/distinct_2/std': 0.04190681502223015, 'rewards/total_composite/mean': 0.26232224702835083, 'rewards/total_composite/std': 0.30432623624801636, 'reward': 0.26232224702835083, 'reward_std': 0.30432623624801636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12602795660495758, 'sampling/sampling_logp_difference/max': 1.3015408515930176, 'sampling/importance_sampling_ratio/min': 0.2721121907234192, 'sampling/importance_sampling_ratio/mean': 1.0090632438659668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6958280131220818, 'clip_ratio/low_mean': 0.0829486520960927, 'clip_ratio/low_min': 0.0829486520960927, 'clip_ratio/high_mean': 0.03374537592753768, 'clip_ratio/high_max': 0.03374537592753768, 'clip_ratio/region_mean': 0.11669402802363038, 'reward_total_mean': 0.26232224702835083, 'reward_meter_mean': 0.3592393100261688, 'reward_meter_std': 0.37753790616989136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.929060697555542, 'reward_repeat_penalty_std': 0.05514172464609146, 'reward_repeat_floor_mean': 0.9909193515777588, 'reward_repeat_floor_std': 0.0063325450755655766, 'reward_near_duplicate_penalty_mean': 0.9576809406280518, 'reward_near_duplicate_penalty_std': 0.02281358651816845, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9696356058120728, 'reward_distinct_2_std': 0.04190681502223015, 'reward_judge_quality_mean': 0.731249988079071, 'reward_judge_quality_std': 0.2623213827610016, 'reward_total_composite_mean': 0.26232224702835083, 'reward_total_composite_std': 0.30432623624801636, 'epoch': 0.08}

 45%|████▌     | 1489/3300 [5:41:47<6:22:47, 12.68s/it]INFO 04-12 21:40:14 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1490/3300 [5:42:01<6:36:34, 13.15s/it]2026-04-12 21:40:28,046 | INFO | train_grpo_train | metrics_logged mode=train step=1490

                                                       
{'loss': -0.0428, 'grad_norm': 1.673332929611206, 'learning_rate': 5.487878787878789e-06, 'num_tokens': 2634104.0, 'completions/mean_length': 143.375, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 20.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.7184266448020935, 'rewards/meter/std': 0.4443516433238983, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.8509981632232666, 'rewards/lexical_plausibility/std': 0.2576138377189636, 'rewards/judge_quality/mean': 0.6000000238418579, 'rewards/judge_quality/std': 0.44352486729621887, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.5575048923492432, 'rewards/total_composite/std': 0.42457643151283264, 'reward': 0.5575048923492432, 'reward_std': 0.42457643151283264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10896066576242447, 'sampling/sampling_logp_difference/max': 0.9587488174438477, 'sampling/importance_sampling_ratio/min': 0.3833722770214081, 'sampling/importance_sampling_ratio/mean': 1.0135163068771362, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5921568423509598, 'clip_ratio/low_mean': 0.02380952425301075, 'clip_ratio/low_min': 0.02380952425301075, 'clip_ratio/high_mean': 0.030654762405902147, 'clip_ratio/high_max': 0.030654762405902147, 'clip_ratio/region_mean': 0.0544642866589129, 'reward_total_mean': 0.5575048923492432, 'reward_meter_mean': 0.7184266448020935, 'reward_meter_std': 0.4443516433238983, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.8509981632232666, 'reward_lexical_plausibility_std': 0.2576138377189636, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.6000000238418579, 'reward_judge_quality_std': 0.44352486729621887, 'reward_total_composite_mean': 0.5575048923492432, 'reward_total_composite_std': 0.42457643151283264, 'epoch': 0.08}

 45%|████▌     | 1490/3300 [5:42:01<6:36:34, 13.15s/it]INFO 04-12 21:40:28 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1491/3300 [5:42:11<6:06:36, 12.16s/it]2026-04-12 21:40:37,904 | INFO | train_grpo_train | metrics_logged mode=train step=1491

                                                       
{'loss': 0.0606, 'grad_norm': 8.14633560180664, 'learning_rate': 5.484848484848485e-06, 'num_tokens': 2636447.0, 'completions/mean_length': 89.875, 'completions/min_length': 75.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.875, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.707213282585144, 'rewards/meter/std': 0.2729416787624359, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9550000429153442, 'rewards/count_floor/std': 0.02777460590004921, 'rewards/lexical_plausibility/mean': 0.9965277910232544, 'rewards/lexical_plausibility/std': 0.009820932522416115, 'rewards/judge_quality/mean': 0.5537499785423279, 'rewards/judge_quality/std': 0.18469570577144623, 'rewards/repeat_penalty/mean': 0.7755419611930847, 'rewards/repeat_penalty/std': 0.13996149599552155, 'rewards/repeat_floor/mean': 0.9788610339164734, 'rewards/repeat_floor/std': 0.013142640702426434, 'rewards/near_duplicate_penalty/mean': 0.9672986268997192, 'rewards/near_duplicate_penalty/std': 0.02037934772670269, 'rewards/opening_diversity/mean': 0.8932291269302368, 'rewards/opening_diversity/std': 0.10232751816511154, 'rewards/distinct_2/mean': 0.8908056616783142, 'rewards/distinct_2/std': 0.05121969059109688, 'rewards/total_composite/mean': 0.34950780868530273, 'rewards/total_composite/std': 0.20279620587825775, 'reward': 0.34950780868530273, 'reward_std': 0.20279620587825775, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1235978901386261, 'sampling/sampling_logp_difference/max': 1.5745749473571777, 'sampling/importance_sampling_ratio/min': 0.20709556341171265, 'sampling/importance_sampling_ratio/mean': 0.9896231889724731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6195723824203014, 'clip_ratio/low_mean': 0.09080421086400747, 'clip_ratio/low_min': 0.09080421086400747, 'clip_ratio/high_mean': 0.033072915859520435, 'clip_ratio/high_max': 0.033072915859520435, 'clip_ratio/region_mean': 0.12387712672352791, 'reward_total_mean': 0.34950780868530273, 'reward_meter_mean': 0.707213282585144, 'reward_meter_std': 0.2729416787624359, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9550000429153442, 'reward_count_floor_std': 0.02777460590004921, 'reward_lexical_plausibility_mean': 0.9965277910232544, 'reward_lexical_plausibility_std': 0.009820932522416115, 'reward_repeat_penalty_mean': 0.7755419611930847, 'reward_repeat_penalty_std': 0.13996149599552155, 'reward_repeat_floor_mean': 0.9788610339164734, 'reward_repeat_floor_std': 0.013142640702426434, 'reward_near_duplicate_penalty_mean': 0.9672986268997192, 'reward_near_duplicate_penalty_std': 0.02037934772670269, 'reward_opening_diversity_mean': 0.8932291269302368, 'reward_opening_diversity_std': 0.10232751816511154, 'reward_distinct_2_mean': 0.8908056616783142, 'reward_distinct_2_std': 0.05121969059109688, 'reward_judge_quality_mean': 0.5537499785423279, 'reward_judge_quality_std': 0.18469570577144623, 'reward_total_composite_mean': 0.34950780868530273, 'reward_total_composite_std': 0.20279620587825775, 'epoch': 0.08}

 45%|████▌     | 1491/3300 [5:42:11<6:06:36, 12.16s/it]INFO 04-12 21:40:38 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1492/3300 [5:42:21<5:43:20, 11.39s/it]2026-04-12 21:40:47,518 | INFO | train_grpo_train | metrics_logged mode=train step=1492

                                                       
{'loss': -0.0241, 'grad_norm': 12.305021286010742, 'learning_rate': 5.4818181818181825e-06, 'num_tokens': 2638512.0, 'completions/mean_length': 82.125, 'completions/min_length': 73.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.8114791512489319, 'rewards/meter/std': 0.27180203795433044, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9399999976158142, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.7008771300315857, 'rewards/repeat_penalty/std': 0.07706035673618317, 'rewards/repeat_floor/mean': 0.9699922800064087, 'rewards/repeat_floor/std': 0.007471021264791489, 'rewards/near_duplicate_penalty/mean': 0.9324653148651123, 'rewards/near_duplicate_penalty/std': 0.02045949548482895, 'rewards/opening_diversity/mean': 0.9765625, 'rewards/opening_diversity/std': 0.032346826046705246, 'rewards/distinct_2/mean': 0.7947638034820557, 'rewards/distinct_2/std': 0.07021769881248474, 'rewards/total_composite/mean': 0.26756078004837036, 'rewards/total_composite/std': 0.10327073931694031, 'reward': 0.26756078004837036, 'reward_std': 0.10327073931694031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1463538408279419, 'sampling/sampling_logp_difference/max': 2.0251224040985107, 'sampling/importance_sampling_ratio/min': 0.13197769224643707, 'sampling/importance_sampling_ratio/mean': 1.0007835626602173, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7512584701180458, 'clip_ratio/low_mean': 0.024697049520909786, 'clip_ratio/low_min': 0.024697049520909786, 'clip_ratio/high_mean': 0.09512386005371809, 'clip_ratio/high_max': 0.09512386005371809, 'clip_ratio/region_mean': 0.11982090957462788, 'reward_total_mean': 0.26756078004837036, 'reward_meter_mean': 0.8114791512489319, 'reward_meter_std': 0.27180203795433044, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9399999976158142, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7008771300315857, 'reward_repeat_penalty_std': 0.07706035673618317, 'reward_repeat_floor_mean': 0.9699922800064087, 'reward_repeat_floor_std': 0.007471021264791489, 'reward_near_duplicate_penalty_mean': 0.9324653148651123, 'reward_near_duplicate_penalty_std': 0.02045949548482895, 'reward_opening_diversity_mean': 0.9765625, 'reward_opening_diversity_std': 0.032346826046705246, 'reward_distinct_2_mean': 0.7947638034820557, 'reward_distinct_2_std': 0.07021769881248474, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.26756078004837036, 'reward_total_composite_std': 0.10327073931694031, 'epoch': 0.08}

 45%|████▌     | 1492/3300 [5:42:21<5:43:20, 11.39s/it]INFO 04-12 21:40:47 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1493/3300 [5:42:34<6:03:49, 12.08s/it]2026-04-12 21:41:01,204 | INFO | train_grpo_train | metrics_logged mode=train step=1493

                                                       
{'loss': 0.001, 'grad_norm': 6.434122562408447, 'learning_rate': 5.478787878787879e-06, 'num_tokens': 2639919.0, 'completions/mean_length': 82.875, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.571430206298828, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.6664613485336304, 'rewards/meter/std': 0.3683927059173584, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9344984292984009, 'rewards/lexical_plausibility/std': 0.18526646494865417, 'rewards/judge_quality/mean': 0.668749988079071, 'rewards/judge_quality/std': 0.3589046001434326, 'rewards/repeat_penalty/mean': 0.78125, 'rewards/repeat_penalty/std': 0.0883883461356163, 'rewards/repeat_floor/mean': 0.9868749976158142, 'rewards/repeat_floor/std': 0.00530329579487443, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.42003118991851807, 'rewards/total_composite/std': 0.3424651026725769, 'reward': 0.42003118991851807, 'reward_std': 0.3424651026725769, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15595580637454987, 'sampling/sampling_logp_difference/max': 2.0770986080169678, 'sampling/importance_sampling_ratio/min': 0.12529321014881134, 'sampling/importance_sampling_ratio/mean': 1.0601035356521606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7379826530814171, 'clip_ratio/low_mean': 0.10134804295375943, 'clip_ratio/low_min': 0.10134804295375943, 'clip_ratio/high_mean': 0.05871212296187878, 'clip_ratio/high_max': 0.05871212296187878, 'clip_ratio/region_mean': 0.1600601659156382, 'reward_total_mean': 0.42003118991851807, 'reward_meter_mean': 0.6664613485336304, 'reward_meter_std': 0.3683927059173584, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9344984292984009, 'reward_lexical_plausibility_std': 0.18526646494865417, 'reward_repeat_penalty_mean': 0.78125, 'reward_repeat_penalty_std': 0.0883883461356163, 'reward_repeat_floor_mean': 0.9868749976158142, 'reward_repeat_floor_std': 0.00530329579487443, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.668749988079071, 'reward_judge_quality_std': 0.3589046001434326, 'reward_total_composite_mean': 0.42003118991851807, 'reward_total_composite_std': 0.3424651026725769, 'epoch': 0.08}

 45%|████▌     | 1493/3300 [5:42:35<6:03:49, 12.08s/it]INFO 04-12 21:41:01 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1494/3300 [5:42:48<6:16:44, 12.52s/it]2026-04-12 21:41:14,726 | INFO | train_grpo_train | metrics_logged mode=train step=1494

                                                       
{'loss': -0.0158, 'grad_norm': 4.789248466491699, 'learning_rate': 5.475757575757576e-06, 'num_tokens': 2641538.0, 'completions/mean_length': 100.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.57143020629883, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5468735098838806, 'rewards/meter/std': 0.4477143883705139, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.9311972260475159, 'rewards/lexical_plausibility/std': 0.19460363686084747, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.27523690462112427, 'rewards/repeat_penalty/mean': 0.9528023600578308, 'rewards/repeat_penalty/std': 0.08412477374076843, 'rewards/repeat_floor/mean': 0.995732843875885, 'rewards/repeat_floor/std': 0.005191707983613014, 'rewards/near_duplicate_penalty/mean': 0.9840523600578308, 'rewards/near_duplicate_penalty/std': 0.02008967101573944, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3218480050563812, 'rewards/total_composite/std': 0.31258994340896606, 'reward': 0.3218480050563812, 'reward_std': 0.31258994340896606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13717229664325714, 'sampling/sampling_logp_difference/max': 1.1928362846374512, 'sampling/importance_sampling_ratio/min': 0.30335962772369385, 'sampling/importance_sampling_ratio/mean': 1.0361531972885132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8859149366617203, 'clip_ratio/low_mean': 0.040857032872736454, 'clip_ratio/low_min': 0.040857032872736454, 'clip_ratio/high_mean': 0.055024946574121714, 'clip_ratio/high_max': 0.055024946574121714, 'clip_ratio/region_mean': 0.09588197944685817, 'reward_total_mean': 0.3218480050563812, 'reward_meter_mean': 0.5468735098838806, 'reward_meter_std': 0.4477143883705139, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.9311972260475159, 'reward_lexical_plausibility_std': 0.19460363686084747, 'reward_repeat_penalty_mean': 0.9528023600578308, 'reward_repeat_penalty_std': 0.08412477374076843, 'reward_repeat_floor_mean': 0.995732843875885, 'reward_repeat_floor_std': 0.005191707983613014, 'reward_near_duplicate_penalty_mean': 0.9840523600578308, 'reward_near_duplicate_penalty_std': 0.02008967101573944, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.27523690462112427, 'reward_total_composite_mean': 0.3218480050563812, 'reward_total_composite_std': 0.31258994340896606, 'epoch': 0.08}

 45%|████▌     | 1494/3300 [5:42:48<6:16:44, 12.52s/it]INFO 04-12 21:41:14 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1495/3300 [5:42:58<5:52:36, 11.72s/it]2026-04-12 21:41:24,699 | INFO | train_grpo_train | metrics_logged mode=train step=1495

                                                       
{'loss': -0.03, 'grad_norm': 10.359344482421875, 'learning_rate': 5.472727272727273e-06, 'num_tokens': 2643390.0, 'completions/mean_length': 60.5, 'completions/min_length': 54.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7368300557136536, 'rewards/meter/std': 0.31711530685424805, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.48749998211860657, 'rewards/judge_quality/std': 0.1060660257935524, 'rewards/repeat_penalty/mean': 0.8162379860877991, 'rewards/repeat_penalty/std': 0.18616178631782532, 'rewards/repeat_floor/mean': 0.9810099601745605, 'rewards/repeat_floor/std': 0.018245793879032135, 'rewards/near_duplicate_penalty/mean': 0.9489593505859375, 'rewards/near_duplicate_penalty/std': 0.03879024088382721, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9365675449371338, 'rewards/distinct_2/std': 0.06015472486615181, 'rewards/total_composite/mean': 0.3489229083061218, 'rewards/total_composite/std': 0.14641989767551422, 'reward': 0.3489229083061218, 'reward_std': 0.14641989767551422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1227312907576561, 'sampling/sampling_logp_difference/max': 1.6666741371154785, 'sampling/importance_sampling_ratio/min': 0.18887419998645782, 'sampling/importance_sampling_ratio/mean': 1.0128438472747803, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9465381503105164, 'clip_ratio/low_mean': 0.036210318095982075, 'clip_ratio/low_min': 0.036210318095982075, 'clip_ratio/high_mean': 0.10332394111901522, 'clip_ratio/high_max': 0.10332394111901522, 'clip_ratio/region_mean': 0.1395342592149973, 'reward_total_mean': 0.3489229083061218, 'reward_meter_mean': 0.7368300557136536, 'reward_meter_std': 0.31711530685424805, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8162379860877991, 'reward_repeat_penalty_std': 0.18616178631782532, 'reward_repeat_floor_mean': 0.9810099601745605, 'reward_repeat_floor_std': 0.018245793879032135, 'reward_near_duplicate_penalty_mean': 0.9489593505859375, 'reward_near_duplicate_penalty_std': 0.03879024088382721, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9365675449371338, 'reward_distinct_2_std': 0.06015472486615181, 'reward_judge_quality_mean': 0.48749998211860657, 'reward_judge_quality_std': 0.1060660257935524, 'reward_total_composite_mean': 0.3489229083061218, 'reward_total_composite_std': 0.14641989767551422, 'epoch': 0.08}

 45%|████▌     | 1495/3300 [5:42:58<5:52:36, 11.72s/it]INFO 04-12 21:41:25 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1496/3300 [5:43:07<5:33:15, 11.08s/it]2026-04-12 21:41:34,197 | INFO | train_grpo_train | metrics_logged mode=train step=1496

                                                       
{'loss': 0.0842, 'grad_norm': 14.571981430053711, 'learning_rate': 5.469696969696971e-06, 'num_tokens': 2645079.0, 'completions/mean_length': 41.125, 'completions/min_length': 36.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7244601845741272, 'rewards/meter/std': 0.35150283575057983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.550000011920929, 'rewards/judge_quality/std': 0.21902382373809814, 'rewards/repeat_penalty/mean': 0.965584397315979, 'rewards/repeat_penalty/std': 0.031008455902338028, 'rewards/repeat_floor/mean': 0.9948376417160034, 'rewards/repeat_floor/std': 0.004651259630918503, 'rewards/near_duplicate_penalty/mean': 0.965584397315979, 'rewards/near_duplicate_penalty/std': 0.031008455902338028, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.41951608657836914, 'rewards/total_composite/std': 0.30211377143859863, 'reward': 0.41951608657836914, 'reward_std': 0.30211377143859863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1708325445652008, 'sampling/sampling_logp_difference/max': 2.2356536388397217, 'sampling/importance_sampling_ratio/min': 0.1069222167134285, 'sampling/importance_sampling_ratio/mean': 1.0017690658569336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.046118751168251, 'clip_ratio/low_mean': 0.06589591037482023, 'clip_ratio/low_min': 0.06589591037482023, 'clip_ratio/high_mean': 0.053077406249940395, 'clip_ratio/high_max': 0.053077406249940395, 'clip_ratio/region_mean': 0.11897331662476063, 'reward_total_mean': 0.41951608657836914, 'reward_meter_mean': 0.7244601845741272, 'reward_meter_std': 0.35150283575057983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.965584397315979, 'reward_repeat_penalty_std': 0.031008455902338028, 'reward_repeat_floor_mean': 0.9948376417160034, 'reward_repeat_floor_std': 0.004651259630918503, 'reward_near_duplicate_penalty_mean': 0.965584397315979, 'reward_near_duplicate_penalty_std': 0.031008455902338028, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.550000011920929, 'reward_judge_quality_std': 0.21902382373809814, 'reward_total_composite_mean': 0.41951608657836914, 'reward_total_composite_std': 0.30211377143859863, 'epoch': 0.08}

 45%|████▌     | 1496/3300 [5:43:08<5:33:15, 11.08s/it]INFO 04-12 21:41:34 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1497/3300 [5:43:21<5:53:51, 11.78s/it]2026-04-12 21:41:47,614 | INFO | train_grpo_train | metrics_logged mode=train step=1497

                                                       
{'loss': -0.0631, 'grad_norm': 2.9713034629821777, 'learning_rate': 5.466666666666667e-06, 'num_tokens': 2646603.0, 'completions/mean_length': 155.5, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 36.66666793823242, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.003530373563989997, 'rewards/meter/std': 0.003565644845366478, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.8572268486022949, 'rewards/lexical_plausibility/std': 0.24691884219646454, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.38873422145843506, 'rewards/repeat_penalty/mean': 0.9400744438171387, 'rewards/repeat_penalty/std': 0.10531332343816757, 'rewards/repeat_floor/mean': 0.9924274682998657, 'rewards/repeat_floor/std': 0.013014971278607845, 'rewards/near_duplicate_penalty/mean': 0.9691842794418335, 'rewards/near_duplicate_penalty/std': 0.046577438712120056, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9672203063964844, 'rewards/distinct_2/std': 0.06894759088754654, 'rewards/total_composite/mean': 0.0021841018460690975, 'rewards/total_composite/std': 0.0018069708021357656, 'reward': 0.0021841018460690975, 'reward_std': 0.0018069706857204437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17462275922298431, 'sampling/sampling_logp_difference/max': 3.2488083839416504, 'sampling/importance_sampling_ratio/min': 0.03882044181227684, 'sampling/importance_sampling_ratio/mean': 1.0238977670669556, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9586880058050156, 'clip_ratio/low_mean': 0.05899064056575298, 'clip_ratio/low_min': 0.05899064056575298, 'clip_ratio/high_mean': 0.09824727475643158, 'clip_ratio/high_max': 0.09824727475643158, 'clip_ratio/region_mean': 0.15723791532218456, 'reward_total_mean': 0.0021841018460690975, 'reward_meter_mean': 0.003530373563989997, 'reward_meter_std': 0.003565644845366478, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.8572268486022949, 'reward_lexical_plausibility_std': 0.24691884219646454, 'reward_repeat_penalty_mean': 0.9400744438171387, 'reward_repeat_penalty_std': 0.10531332343816757, 'reward_repeat_floor_mean': 0.9924274682998657, 'reward_repeat_floor_std': 0.013014971278607845, 'reward_near_duplicate_penalty_mean': 0.9691842794418335, 'reward_near_duplicate_penalty_std': 0.046577438712120056, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9672203063964844, 'reward_distinct_2_std': 0.06894759088754654, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.38873422145843506, 'reward_total_composite_mean': 0.0021841018460690975, 'reward_total_composite_std': 0.0018069708021357656, 'epoch': 0.08}

 45%|████▌     | 1497/3300 [5:43:21<5:53:51, 11.78s/it]INFO 04-12 21:41:47 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1498/3300 [5:43:30<5:28:46, 10.95s/it]2026-04-12 21:41:56,597 | INFO | train_grpo_train | metrics_logged mode=train step=1498

                                                       
{'loss': 0.0322, 'grad_norm': 12.15373420715332, 'learning_rate': 5.463636363636364e-06, 'num_tokens': 2648308.0, 'completions/mean_length': 56.125, 'completions/min_length': 53.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8618424534797668, 'rewards/meter/std': 0.21899884939193726, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.44999998807907104, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9447248578071594, 'rewards/repeat_penalty/std': 0.06420300900936127, 'rewards/repeat_floor/mean': 0.9930490255355835, 'rewards/repeat_floor/std': 0.007730742450803518, 'rewards/near_duplicate_penalty/mean': 0.9694256782531738, 'rewards/near_duplicate_penalty/std': 0.031231814995408058, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9737240076065063, 'rewards/distinct_2/std': 0.04157579690217972, 'rewards/total_composite/mean': 0.38480257987976074, 'rewards/total_composite/std': 0.09682688117027283, 'reward': 0.38480257987976074, 'reward_std': 0.09682688862085342, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10703235119581223, 'sampling/sampling_logp_difference/max': 1.5997743606567383, 'sampling/importance_sampling_ratio/min': 0.2019420713186264, 'sampling/importance_sampling_ratio/mean': 1.0039345026016235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6510071456432343, 'clip_ratio/low_mean': 0.021558132953941822, 'clip_ratio/low_min': 0.021558132953941822, 'clip_ratio/high_mean': 0.06633547181263566, 'clip_ratio/high_max': 0.06633547181263566, 'clip_ratio/region_mean': 0.08789360476657748, 'reward_total_mean': 0.38480257987976074, 'reward_meter_mean': 0.8618424534797668, 'reward_meter_std': 0.21899884939193726, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9447248578071594, 'reward_repeat_penalty_std': 0.06420300900936127, 'reward_repeat_floor_mean': 0.9930490255355835, 'reward_repeat_floor_std': 0.007730742450803518, 'reward_near_duplicate_penalty_mean': 0.9694256782531738, 'reward_near_duplicate_penalty_std': 0.031231814995408058, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9737240076065063, 'reward_distinct_2_std': 0.04157579690217972, 'reward_judge_quality_mean': 0.44999998807907104, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.38480257987976074, 'reward_total_composite_std': 0.09682688117027283, 'epoch': 0.08}

 45%|████▌     | 1498/3300 [5:43:30<5:28:46, 10.95s/it]INFO 04-12 21:41:56 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1499/3300 [5:43:43<5:51:45, 11.72s/it]2026-04-12 21:42:10,127 | INFO | train_grpo_train | metrics_logged mode=train step=1499

                                                       
{'loss': -0.0718, 'grad_norm': 5.17856502532959, 'learning_rate': 5.460606060606061e-06, 'num_tokens': 2649814.0, 'completions/mean_length': 97.25, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7419392466545105, 'rewards/meter/std': 0.35429903864860535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9380473494529724, 'rewards/lexical_plausibility/std': 0.17522861063480377, 'rewards/judge_quality/mean': 0.4712499976158142, 'rewards/judge_quality/std': 0.26384180784225464, 'rewards/repeat_penalty/mean': 0.9611736536026001, 'rewards/repeat_penalty/std': 0.06628385186195374, 'rewards/repeat_floor/mean': 0.9945216178894043, 'rewards/repeat_floor/std': 0.009005018509924412, 'rewards/near_duplicate_penalty/mean': 0.968722403049469, 'rewards/near_duplicate_penalty/std': 0.04608163982629776, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9912587404251099, 'rewards/distinct_2/std': 0.024724015966057777, 'rewards/total_composite/mean': 0.3988385498523712, 'rewards/total_composite/std': 0.2777070701122284, 'reward': 0.3988385498523712, 'reward_std': 0.2777070701122284, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14017628133296967, 'sampling/sampling_logp_difference/max': 1.3838778734207153, 'sampling/importance_sampling_ratio/min': 0.2506048381328583, 'sampling/importance_sampling_ratio/mean': 1.0108575820922852, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8696282655000687, 'clip_ratio/low_mean': 0.04787644650787115, 'clip_ratio/low_min': 0.04787644650787115, 'clip_ratio/high_mean': 0.06470645405352116, 'clip_ratio/high_max': 0.06470645405352116, 'clip_ratio/region_mean': 0.11258290056139231, 'reward_total_mean': 0.3988385498523712, 'reward_meter_mean': 0.7419392466545105, 'reward_meter_std': 0.35429903864860535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9380473494529724, 'reward_lexical_plausibility_std': 0.17522861063480377, 'reward_repeat_penalty_mean': 0.9611736536026001, 'reward_repeat_penalty_std': 0.06628385186195374, 'reward_repeat_floor_mean': 0.9945216178894043, 'reward_repeat_floor_std': 0.009005018509924412, 'reward_near_duplicate_penalty_mean': 0.968722403049469, 'reward_near_duplicate_penalty_std': 0.04608163982629776, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9912587404251099, 'reward_distinct_2_std': 0.024724015966057777, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.26384180784225464, 'reward_total_composite_mean': 0.3988385498523712, 'reward_total_composite_std': 0.2777070701122284, 'epoch': 0.08}

 45%|████▌     | 1499/3300 [5:43:43<5:51:45, 11.72s/it]INFO 04-12 21:42:10 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1500/3300 [5:43:53<5:28:54, 10.96s/it]2026-04-12 21:42:19,327 | INFO | train_grpo_train | metrics_logged mode=train step=1500

                                                       
{'loss': 0.0934, 'grad_norm': 11.850004196166992, 'learning_rate': 5.457575757575758e-06, 'num_tokens': 2651717.0, 'completions/mean_length': 65.875, 'completions/min_length': 62.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.828018844127655, 'rewards/meter/std': 0.33673205971717834, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4374999701976776, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.7612658739089966, 'rewards/repeat_penalty/std': 0.1380801945924759, 'rewards/repeat_floor/mean': 0.9732581377029419, 'rewards/repeat_floor/std': 0.013619935140013695, 'rewards/near_duplicate_penalty/mean': 0.9182708263397217, 'rewards/near_duplicate_penalty/std': 0.036135610193014145, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.8807505369186401, 'rewards/distinct_2/std': 0.06856482475996017, 'rewards/total_composite/mean': 0.32490718364715576, 'rewards/total_composite/std': 0.15903796255588531, 'reward': 0.32490718364715576, 'reward_std': 0.15903796255588531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1432938128709793, 'sampling/sampling_logp_difference/max': 1.5977137088775635, 'sampling/importance_sampling_ratio/min': 0.2155953198671341, 'sampling/importance_sampling_ratio/mean': 1.0119035243988037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.857727088034153, 'clip_ratio/low_mean': 0.08351230248808861, 'clip_ratio/low_min': 0.08351230248808861, 'clip_ratio/high_mean': 0.04989119339734316, 'clip_ratio/high_max': 0.04989119339734316, 'clip_ratio/region_mean': 0.13340349588543177, 'reward_total_mean': 0.32490718364715576, 'reward_meter_mean': 0.828018844127655, 'reward_meter_std': 0.33673205971717834, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7612658739089966, 'reward_repeat_penalty_std': 0.1380801945924759, 'reward_repeat_floor_mean': 0.9732581377029419, 'reward_repeat_floor_std': 0.013619935140013695, 'reward_near_duplicate_penalty_mean': 0.9182708263397217, 'reward_near_duplicate_penalty_std': 0.036135610193014145, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.8807505369186401, 'reward_distinct_2_std': 0.06856482475996017, 'reward_judge_quality_mean': 0.4374999701976776, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.32490718364715576, 'reward_total_composite_std': 0.15903796255588531, 'epoch': 0.08}

 45%|████▌     | 1500/3300 [5:43:53<5:28:54, 10.96s/it]INFO 04-12 21:42:19 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:07<00:43,  3.97s/it]

 23%|██▎       | 3/13 [00:18<01:08,  6.87s/it]

 31%|███       | 4/13 [00:26<01:03,  7.03s/it]

 38%|███▊      | 5/13 [00:34<00:58,  7.35s/it]

 46%|████▌     | 6/13 [00:50<01:12, 10.34s/it]

 54%|█████▍    | 7/13 [01:01<01:03, 10.57s/it]

 62%|██████▏   | 8/13 [01:13<00:54, 10.94s/it]

 69%|██████▉   | 9/13 [01:24<00:44, 11.06s/it]

 77%|███████▋  | 10/13 [01:31<00:29,  9.80s/it]

 85%|████████▍ | 11/13 [01:38<00:17,  8.82s/it]

 92%|█████████▏| 12/13 [01:45<00:08,  8.47s/it]

100%|██████████| 13/13 [01:56<00:00,  9.27s/it]2026-04-12 21:44:28,341 | INFO | train_grpo_train | metrics_logged mode=eval step=1500

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 128.8856, 'eval_samples_per_second': 0.807, 'eval_steps_per_second': 0.101, 'eval_num_tokens': 2651717.0, 'eval_completions/mean_length': 138.44230769230768, 'eval_completions/min_length': 34.46153846153846, 'eval_completions/max_length': 363.38461538461536, 'eval_completions/clipped_ratio': 0.09615384615384616, 'eval_completions/mean_terminated_length': 98.25925181462215, 'eval_completions/min_terminated_length': 34.46153846153846, 'eval_completions/max_terminated_length': 186.53846153846155, 'eval_rewards/meter/mean': 0.5205199214128348, 'eval_rewards/meter/std': 0.3734464542223857, 'eval_rewards/count_adherence/mean': 0.7607057553071243, 'eval_rewards/count_adherence/std': 0.19632262048813012, 'eval_rewards/arabic_clean/mean': 0.8653846153846154, 'eval_rewards/arabic_clean/std': 0.320187671826436, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9971153827813956, 'eval_rewards/arabic_floor/std': 0.008158923341677738, 'eval_rewards/count_floor/mean': 0.9282117256751428, 'eval_rewards/count_floor/std': 0.05889678803774027, 'eval_rewards/lexical_plausibility/mean': 0.952255340722891, 'eval_rewards/lexical_plausibility/std': 0.10383106080385354, 'eval_rewards/judge_quality/mean': 0.40288461630160993, 'eval_rewards/judge_quality/std': 0.191096341667267, 'eval_rewards/repeat_penalty/mean': 0.8246570871426508, 'eval_rewards/repeat_penalty/std': 0.18076111978063217, 'eval_rewards/repeat_floor/mean': 0.9809717169174781, 'eval_rewards/repeat_floor/std': 0.019479196697760087, 'eval_rewards/near_duplicate_penalty/mean': 0.952797912634336, 'eval_rewards/near_duplicate_penalty/std': 0.042314441301501714, 'eval_rewards/opening_diversity/mean': 0.9644750494223374, 'eval_rewards/opening_diversity/std': 0.08255957468197896, 'eval_rewards/distinct_2/mean': 0.8909281308834369, 'eval_rewards/distinct_2/std': 0.13125740163601363, 'eval_rewards/total_composite/mean': 0.18016218164792427, 'eval_rewards/total_composite/std': 0.16037252316108117, 'eval_reward': 0.18016218164792427, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07613115356518672, 'eval_sampling/sampling_logp_difference/max': 1.0884063977461595, 'eval_sampling/importance_sampling_ratio/min': 0.3465814143419266, 'eval_sampling/importance_sampling_ratio/mean': 1.0177436058337872, 'eval_sampling/importance_sampling_ratio/max': 1.5422767400741577, 'eval_entropy': 0.830715876359206, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.18016218164792427, 'eval_reward_meter_mean': 0.5205199214128348, 'eval_reward_meter_std': 0.3734464542223857, 'eval_reward_count_adherence_mean': 0.7607057553071243, 'eval_reward_count_adherence_std': 0.19632262048813012, 'eval_reward_arabic_clean_mean': 0.8653846153846154, 'eval_reward_arabic_clean_std': 0.320187671826436, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9971153827813956, 'eval_reward_arabic_floor_std': 0.008158923341677738, 'eval_reward_count_floor_mean': 0.9282117256751428, 'eval_reward_count_floor_std': 0.05889678803774027, 'eval_reward_lexical_plausibility_mean': 0.952255340722891, 'eval_reward_lexical_plausibility_std': 0.10383106080385354, 'eval_reward_repeat_penalty_mean': 0.8246570871426508, 'eval_reward_repeat_penalty_std': 0.18076111978063217, 'eval_reward_repeat_floor_mean': 0.9809717169174781, 'eval_reward_repeat_floor_std': 0.019479196697760087, 'eval_reward_near_duplicate_penalty_mean': 0.952797912634336, 'eval_reward_near_duplicate_penalty_std': 0.042314441301501714, 'eval_reward_opening_diversity_mean': 0.9644750494223374, 'eval_reward_opening_diversity_std': 0.08255957468197896, 'eval_reward_distinct_2_mean': 0.8909281308834369, 'eval_reward_distinct_2_std': 0.13125740163601363, 'eval_reward_judge_quality_mean': 0.40288461630160993, 'eval_reward_judge_quality_std': 0.191096341667267, 'eval_reward_total_composite_mean': 0.18016218164792427, 'eval_reward_total_composite_std': 0.16037252316108117, 'epoch': 0.08}

 45%|████▌     | 1500/3300 [5:46:02<5:28:54, 10.96s/it]

100%|██████████| 13/13 [01:57<00:00,  9.27s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 45%|████▌     | 1501/3300 [5:46:19<25:51:27, 51.74s/it]2026-04-12 21:44:46,232 | INFO | train_grpo_train | metrics_logged mode=train step=1501

                                                        
{'loss': -0.0327, 'grad_norm': 1.1844478845596313, 'learning_rate': 5.4545454545454545e-06, 'num_tokens': 2653108.0, 'completions/mean_length': 334.875, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 39.66666793823242, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.4636211395263672, 'rewards/meter/std': 0.41752588748931885, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.722219705581665, 'rewards/lexical_plausibility/std': 0.23276673257350922, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.3881365656852722, 'rewards/repeat_penalty/mean': 0.9706431031227112, 'rewards/repeat_penalty/std': 0.06275924295186996, 'rewards/repeat_floor/mean': 0.9960339069366455, 'rewards/repeat_floor/std': 0.008184741251170635, 'rewards/near_duplicate_penalty/mean': 0.9796320796012878, 'rewards/near_duplicate_penalty/std': 0.03757540136575699, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9898785352706909, 'rewards/distinct_2/std': 0.028627805411815643, 'rewards/total_composite/mean': 0.15541711449623108, 'rewards/total_composite/std': 0.2921048700809479, 'reward': 0.15541711449623108, 'reward_std': 0.2921048700809479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11035260558128357, 'sampling/sampling_logp_difference/max': 1.0038478374481201, 'sampling/importance_sampling_ratio/min': 0.4140963554382324, 'sampling/importance_sampling_ratio/mean': 1.0581837892532349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31593991816043854, 'clip_ratio/low_mean': 0.00937500037252903, 'clip_ratio/low_min': 0.00937500037252903, 'clip_ratio/high_mean': 0.02227564202621579, 'clip_ratio/high_max': 0.02227564202621579, 'clip_ratio/region_mean': 0.03165064239874482, 'reward_total_mean': 0.15541711449623108, 'reward_meter_mean': 0.4636211395263672, 'reward_meter_std': 0.41752588748931885, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.722219705581665, 'reward_lexical_plausibility_std': 0.23276673257350922, 'reward_repeat_penalty_mean': 0.9706431031227112, 'reward_repeat_penalty_std': 0.06275924295186996, 'reward_repeat_floor_mean': 0.9960339069366455, 'reward_repeat_floor_std': 0.008184741251170635, 'reward_near_duplicate_penalty_mean': 0.9796320796012878, 'reward_near_duplicate_penalty_std': 0.03757540136575699, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9898785352706909, 'reward_distinct_2_std': 0.028627805411815643, 'reward_judge_quality_mean': 0.3125, 'reward_judge_quality_std': 0.3881365656852722, 'reward_total_composite_mean': 0.15541711449623108, 'reward_total_composite_std': 0.2921048700809479, 'epoch': 0.08}

 45%|████▌     | 1501/3300 [5:46:20<25:51:27, 51.74s/it]INFO 04-12 21:44:46 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1502/3300 [5:46:28<19:19:48, 38.70s/it]2026-04-12 21:44:54,503 | INFO | train_grpo_train | metrics_logged mode=train step=1502

                                                        
{'loss': 0.0201, 'grad_norm': 19.518024444580078, 'learning_rate': 5.451515151515152e-06, 'num_tokens': 2654656.0, 'completions/mean_length': 22.5, 'completions/min_length': 20.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.7743388414382935, 'rewards/meter/std': 0.3545885384082794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.7211898565292358, 'rewards/repeat_penalty/std': 0.05438433960080147, 'rewards/repeat_floor/mean': 0.979237973690033, 'rewards/repeat_floor/std': 0.010876880958676338, 'rewards/near_duplicate_penalty/mean': 0.9615864753723145, 'rewards/near_duplicate_penalty/std': 0.0725124403834343, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4979093074798584, 'rewards/total_composite/std': 0.33453577756881714, 'reward': 0.4979093074798584, 'reward_std': 0.33453577756881714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13807715475559235, 'sampling/sampling_logp_difference/max': 1.2705497741699219, 'sampling/importance_sampling_ratio/min': 0.28067725896835327, 'sampling/importance_sampling_ratio/mean': 1.0295438766479492, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3630553856492043, 'clip_ratio/low_mean': 0.0863395887427032, 'clip_ratio/low_min': 0.0863395887427032, 'clip_ratio/high_mean': 0.05358391720801592, 'clip_ratio/high_max': 0.05358391720801592, 'clip_ratio/region_mean': 0.13992350595071912, 'reward_total_mean': 0.4979093074798584, 'reward_meter_mean': 0.7743388414382935, 'reward_meter_std': 0.3545885384082794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7211898565292358, 'reward_repeat_penalty_std': 0.05438433960080147, 'reward_repeat_floor_mean': 0.979237973690033, 'reward_repeat_floor_std': 0.010876880958676338, 'reward_near_duplicate_penalty_mean': 0.9615864753723145, 'reward_near_duplicate_penalty_std': 0.0725124403834343, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.3091896176338196, 'reward_total_composite_mean': 0.4979093074798584, 'reward_total_composite_std': 0.33453577756881714, 'epoch': 0.08}

 46%|████▌     | 1502/3300 [5:46:28<19:19:48, 38.70s/it]INFO 04-12 21:44:54 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1503/3300 [5:46:44<16:01:11, 32.09s/it]2026-04-12 21:45:11,172 | INFO | train_grpo_train | metrics_logged mode=train step=1503

                                                        
{'loss': -0.0226, 'grad_norm': 5.816728115081787, 'learning_rate': 5.448484848484848e-06, 'num_tokens': 2656318.0, 'completions/mean_length': 105.75, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.71428680419922, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.4968388080596924, 'rewards/meter/std': 0.4825288653373718, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9437500238418579, 'rewards/count_floor/std': 0.0776323676109314, 'rewards/lexical_plausibility/mean': 0.935044527053833, 'rewards/lexical_plausibility/std': 0.18372172117233276, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.2841905951499939, 'rewards/repeat_penalty/mean': 0.8561388254165649, 'rewards/repeat_penalty/std': 0.19052931666374207, 'rewards/repeat_floor/mean': 0.9834905862808228, 'rewards/repeat_floor/std': 0.022917786613106728, 'rewards/near_duplicate_penalty/mean': 0.9394563436508179, 'rewards/near_duplicate_penalty/std': 0.07918263226747513, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9591346383094788, 'rewards/distinct_2/std': 0.08378557860851288, 'rewards/total_composite/mean': 0.3224206566810608, 'rewards/total_composite/std': 0.37920036911964417, 'reward': 0.3224206566810608, 'reward_std': 0.37920036911964417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14845776557922363, 'sampling/sampling_logp_difference/max': 3.0565102100372314, 'sampling/importance_sampling_ratio/min': 0.04705160856246948, 'sampling/importance_sampling_ratio/mean': 1.0048623085021973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5755541920661926, 'clip_ratio/low_mean': 0.039290034445002675, 'clip_ratio/low_min': 0.039290034445002675, 'clip_ratio/high_mean': 0.04572284361347556, 'clip_ratio/high_max': 0.04572284361347556, 'clip_ratio/region_mean': 0.08501287805847824, 'reward_total_mean': 0.3224206566810608, 'reward_meter_mean': 0.4968388080596924, 'reward_meter_std': 0.4825288653373718, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9437500238418579, 'reward_count_floor_std': 0.0776323676109314, 'reward_lexical_plausibility_mean': 0.935044527053833, 'reward_lexical_plausibility_std': 0.18372172117233276, 'reward_repeat_penalty_mean': 0.8561388254165649, 'reward_repeat_penalty_std': 0.19052931666374207, 'reward_repeat_floor_mean': 0.9834905862808228, 'reward_repeat_floor_std': 0.022917786613106728, 'reward_near_duplicate_penalty_mean': 0.9394563436508179, 'reward_near_duplicate_penalty_std': 0.07918263226747513, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9591346383094788, 'reward_distinct_2_std': 0.08378557860851288, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.2841905951499939, 'reward_total_composite_mean': 0.3224206566810608, 'reward_total_composite_std': 0.37920036911964417, 'epoch': 0.08}

 46%|████▌     | 1503/3300 [5:46:45<16:01:11, 32.09s/it]INFO 04-12 21:45:11 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1504/3300 [5:46:53<12:34:08, 25.19s/it]2026-04-12 21:45:20,268 | INFO | train_grpo_train | metrics_logged mode=train step=1504

                                                        
{'loss': 0.0467, 'grad_norm': 9.911951065063477, 'learning_rate': 5.445454545454546e-06, 'num_tokens': 2658307.0, 'completions/mean_length': 77.625, 'completions/min_length': 73.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.8497236371040344, 'rewards/meter/std': 0.20450463891029358, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9174999594688416, 'rewards/count_floor/std': 0.031052952632308006, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36249998211860657, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.8421530723571777, 'rewards/repeat_penalty/std': 0.1271183043718338, 'rewards/repeat_floor/mean': 0.9819009304046631, 'rewards/repeat_floor/std': 0.014206822961568832, 'rewards/near_duplicate_penalty/mean': 0.9481204748153687, 'rewards/near_duplicate_penalty/std': 0.03183474764227867, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8853652477264404, 'rewards/distinct_2/std': 0.10835153609514236, 'rewards/total_composite/mean': 0.27107545733451843, 'rewards/total_composite/std': 0.09206779301166534, 'reward': 0.27107545733451843, 'reward_std': 0.09206779301166534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13171453773975372, 'sampling/sampling_logp_difference/max': 4.11968994140625, 'sampling/importance_sampling_ratio/min': 0.016249552369117737, 'sampling/importance_sampling_ratio/mean': 1.0106518268585205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9071650430560112, 'clip_ratio/low_mean': 0.04211115650832653, 'clip_ratio/low_min': 0.04211115650832653, 'clip_ratio/high_mean': 0.07172557269223034, 'clip_ratio/high_max': 0.07172557269223034, 'clip_ratio/region_mean': 0.11383672920055687, 'reward_total_mean': 0.27107545733451843, 'reward_meter_mean': 0.8497236371040344, 'reward_meter_std': 0.20450463891029358, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.1035098284482956, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9174999594688416, 'reward_count_floor_std': 0.031052952632308006, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8421530723571777, 'reward_repeat_penalty_std': 0.1271183043718338, 'reward_repeat_floor_mean': 0.9819009304046631, 'reward_repeat_floor_std': 0.014206822961568832, 'reward_near_duplicate_penalty_mean': 0.9481204748153687, 'reward_near_duplicate_penalty_std': 0.03183474764227867, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8853652477264404, 'reward_distinct_2_std': 0.10835153609514236, 'reward_judge_quality_mean': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.27107545733451843, 'reward_total_composite_std': 0.09206779301166534, 'epoch': 0.08}

 46%|████▌     | 1504/3300 [5:46:54<12:34:08, 25.19s/it]INFO 04-12 21:45:20 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1505/3300 [5:47:03<10:09:44, 20.38s/it]2026-04-12 21:45:29,449 | INFO | train_grpo_train | metrics_logged mode=train step=1505

                                                        
{'loss': -0.0028, 'grad_norm': 15.74963092803955, 'learning_rate': 5.442424242424243e-06, 'num_tokens': 2659760.0, 'completions/mean_length': 32.625, 'completions/min_length': 26.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8095187544822693, 'rewards/meter/std': 0.2534782290458679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.192831352353096, 'rewards/repeat_penalty/mean': 0.9941810965538025, 'rewards/repeat_penalty/std': 0.009612584486603737, 'rewards/repeat_floor/mean': 0.9991271495819092, 'rewards/repeat_floor/std': 0.0014418804785236716, 'rewards/near_duplicate_penalty/mean': 0.9941810965538025, 'rewards/near_duplicate_penalty/std': 0.009612584486603737, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.3758554756641388, 'rewards/total_composite/std': 0.22956426441669464, 'reward': 0.3758554756641388, 'reward_std': 0.22956423461437225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18263065814971924, 'sampling/sampling_logp_difference/max': 1.4986176490783691, 'sampling/importance_sampling_ratio/min': 0.2234388291835785, 'sampling/importance_sampling_ratio/mean': 1.0401262044906616, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.265030175447464, 'clip_ratio/low_mean': 0.1291364375501871, 'clip_ratio/low_min': 0.1291364375501871, 'clip_ratio/high_mean': 0.049573054537177086, 'clip_ratio/high_max': 0.049573054537177086, 'clip_ratio/region_mean': 0.1787094920873642, 'reward_total_mean': 0.3758554756641388, 'reward_meter_mean': 0.8095187544822693, 'reward_meter_std': 0.2534782290458679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9941810965538025, 'reward_repeat_penalty_std': 0.009612584486603737, 'reward_repeat_floor_mean': 0.9991271495819092, 'reward_repeat_floor_std': 0.0014418804785236716, 'reward_near_duplicate_penalty_mean': 0.9941810965538025, 'reward_near_duplicate_penalty_std': 0.009612584486603737, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.192831352353096, 'reward_total_composite_mean': 0.3758554756641388, 'reward_total_composite_std': 0.22956426441669464, 'epoch': 0.08}

 46%|████▌     | 1505/3300 [5:47:03<10:09:44, 20.38s/it]INFO 04-12 21:45:29 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1506/3300 [5:47:12<8:28:30, 17.01s/it] 2026-04-12 21:45:38,553 | INFO | train_grpo_train | metrics_logged mode=train step=1506

                                                       
{'loss': 0.0277, 'grad_norm': 13.640207290649414, 'learning_rate': 5.43939393939394e-06, 'num_tokens': 2661249.0, 'completions/mean_length': 44.125, 'completions/min_length': 33.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7546921968460083, 'rewards/meter/std': 0.30970534682273865, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.36250001192092896, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.8982670307159424, 'rewards/repeat_penalty/std': 0.09006664901971817, 'rewards/repeat_floor/mean': 0.987313985824585, 'rewards/repeat_floor/std': 0.011077425442636013, 'rewards/near_duplicate_penalty/mean': 0.9483059644699097, 'rewards/near_duplicate_penalty/std': 0.04114063084125519, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9452006220817566, 'rewards/distinct_2/std': 0.05571816489100456, 'rewards/total_composite/mean': 0.24329471588134766, 'rewards/total_composite/std': 0.12118619680404663, 'reward': 0.24329471588134766, 'reward_std': 0.12118618935346603, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12625190615653992, 'sampling/sampling_logp_difference/max': 2.643170118331909, 'sampling/importance_sampling_ratio/min': 0.07113540917634964, 'sampling/importance_sampling_ratio/mean': 0.9974692463874817, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6669752933084965, 'clip_ratio/low_mean': 0.052506458945572376, 'clip_ratio/low_min': 0.052506458945572376, 'clip_ratio/high_mean': 0.05770864710211754, 'clip_ratio/high_max': 0.05770864710211754, 'clip_ratio/region_mean': 0.11021510604768991, 'reward_total_mean': 0.24329471588134766, 'reward_meter_mean': 0.7546921968460083, 'reward_meter_std': 0.30970534682273865, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8982670307159424, 'reward_repeat_penalty_std': 0.09006664901971817, 'reward_repeat_floor_mean': 0.987313985824585, 'reward_repeat_floor_std': 0.011077425442636013, 'reward_near_duplicate_penalty_mean': 0.9483059644699097, 'reward_near_duplicate_penalty_std': 0.04114063084125519, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9452006220817566, 'reward_distinct_2_std': 0.05571816489100456, 'reward_judge_quality_mean': 0.36250001192092896, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.24329471588134766, 'reward_total_composite_std': 0.12118619680404663, 'epoch': 0.08}

 46%|████▌     | 1506/3300 [5:47:12<8:28:30, 17.01s/it]INFO 04-12 21:45:38 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1507/3300 [5:47:20<7:11:25, 14.44s/it]2026-04-12 21:45:46,993 | INFO | train_grpo_train | metrics_logged mode=train step=1507

                                                       
{'loss': -0.0711, 'grad_norm': 15.191036224365234, 'learning_rate': 5.436363636363636e-06, 'num_tokens': 2662557.0, 'completions/mean_length': 19.5, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9796422719955444, 'rewards/meter/std': 0.020406179130077362, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.75, 'rewards/repeat_penalty/std': 0.0, 'rewards/repeat_floor/mean': 0.9850000143051147, 'rewards/repeat_floor/std': 0.0, 'rewards/near_duplicate_penalty/mean': 1.0, 'rewards/near_duplicate_penalty/std': 0.0, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.20541267096996307, 'rewards/total_composite/std': 0.11572809517383575, 'reward': 0.20541267096996307, 'reward_std': 0.11572809517383575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.047722648829221725, 'sampling/sampling_logp_difference/max': 0.8962516784667969, 'sampling/importance_sampling_ratio/min': 0.40809646248817444, 'sampling/importance_sampling_ratio/mean': 1.0133461952209473, 'sampling/importance_sampling_ratio/max': 1.3490544557571411, 'entropy': 0.39085325971245766, 'clip_ratio/low_mean': 0.03484477149322629, 'clip_ratio/low_min': 0.03484477149322629, 'clip_ratio/high_mean': 0.02766798436641693, 'clip_ratio/high_max': 0.02766798436641693, 'clip_ratio/region_mean': 0.06251275585964322, 'reward_total_mean': 0.20541267096996307, 'reward_meter_mean': 0.9796422719955444, 'reward_meter_std': 0.020406179130077362, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.75, 'reward_repeat_penalty_std': 0.0, 'reward_repeat_floor_mean': 0.9850000143051147, 'reward_repeat_floor_std': 0.0, 'reward_near_duplicate_penalty_mean': 1.0, 'reward_near_duplicate_penalty_std': 0.0, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.20541267096996307, 'reward_total_composite_std': 0.11572809517383575, 'epoch': 0.08}

 46%|████▌     | 1507/3300 [5:47:20<7:11:25, 14.44s/it]INFO 04-12 21:45:47 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1508/3300 [5:47:35<7:15:12, 14.57s/it]2026-04-12 21:46:01,879 | INFO | train_grpo_train | metrics_logged mode=train step=1508

                                                       
{'loss': -0.1422, 'grad_norm': 3.514066457748413, 'learning_rate': 5.4333333333333335e-06, 'num_tokens': 2664914.0, 'completions/mean_length': 209.625, 'completions/min_length': 101.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 108.83333587646484, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.4659782648086548, 'rewards/meter/std': 0.30951109528541565, 'rewards/count_adherence/mean': 0.4861111342906952, 'rewards/count_adherence/std': 0.20520134270191193, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.8458333611488342, 'rewards/count_floor/std': 0.06156039610505104, 'rewards/lexical_plausibility/mean': 0.8928514719009399, 'rewards/lexical_plausibility/std': 0.19861266016960144, 'rewards/judge_quality/mean': 0.24374999105930328, 'rewards/judge_quality/std': 0.17816023528575897, 'rewards/repeat_penalty/mean': 0.8520583510398865, 'rewards/repeat_penalty/std': 0.12227675318717957, 'rewards/repeat_floor/mean': 0.983232319355011, 'rewards/repeat_floor/std': 0.013877196237444878, 'rewards/near_duplicate_penalty/mean': 0.9538599252700806, 'rewards/near_duplicate_penalty/std': 0.033677831292152405, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8905925154685974, 'rewards/distinct_2/std': 0.10430935770273209, 'rewards/total_composite/mean': 0.11469920724630356, 'rewards/total_composite/std': 0.1024836003780365, 'reward': 0.11469920724630356, 'reward_std': 0.1024836003780365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13628514111042023, 'sampling/sampling_logp_difference/max': 1.3663530349731445, 'sampling/importance_sampling_ratio/min': 0.2550353705883026, 'sampling/importance_sampling_ratio/mean': 1.014146327972412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.697959378361702, 'clip_ratio/low_mean': 0.029232016764581203, 'clip_ratio/low_min': 0.029232016764581203, 'clip_ratio/high_mean': 0.06191588006913662, 'clip_ratio/high_max': 0.06191588006913662, 'clip_ratio/region_mean': 0.09114789683371782, 'reward_total_mean': 0.11469920724630356, 'reward_meter_mean': 0.4659782648086548, 'reward_meter_std': 0.30951109528541565, 'reward_count_adherence_mean': 0.4861111342906952, 'reward_count_adherence_std': 0.20520134270191193, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.8458333611488342, 'reward_count_floor_std': 0.06156039610505104, 'reward_lexical_plausibility_mean': 0.8928514719009399, 'reward_lexical_plausibility_std': 0.19861266016960144, 'reward_repeat_penalty_mean': 0.8520583510398865, 'reward_repeat_penalty_std': 0.12227675318717957, 'reward_repeat_floor_mean': 0.983232319355011, 'reward_repeat_floor_std': 0.013877196237444878, 'reward_near_duplicate_penalty_mean': 0.9538599252700806, 'reward_near_duplicate_penalty_std': 0.033677831292152405, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8905925154685974, 'reward_distinct_2_std': 0.10430935770273209, 'reward_judge_quality_mean': 0.24374999105930328, 'reward_judge_quality_std': 0.17816023528575897, 'reward_total_composite_mean': 0.11469920724630356, 'reward_total_composite_std': 0.1024836003780365, 'epoch': 0.08}

 46%|████▌     | 1508/3300 [5:47:35<7:15:12, 14.57s/it]INFO 04-12 21:46:02 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1509/3300 [5:47:44<6:23:29, 12.85s/it]2026-04-12 21:46:10,707 | INFO | train_grpo_train | metrics_logged mode=train step=1509

                                                       
{'loss': 0.0063, 'grad_norm': 13.724625587463379, 'learning_rate': 5.430303030303032e-06, 'num_tokens': 2666489.0, 'completions/mean_length': 35.875, 'completions/min_length': 33.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8653483390808105, 'rewards/meter/std': 0.24270948767662048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.47499996423721313, 'rewards/judge_quality/std': 0.1164964810013771, 'rewards/repeat_penalty/mean': 0.9800074696540833, 'rewards/repeat_penalty/std': 0.032433781772851944, 'rewards/repeat_floor/mean': 0.9970011115074158, 'rewards/repeat_floor/std': 0.0048650591634213924, 'rewards/near_duplicate_penalty/mean': 0.9800074696540833, 'rewards/near_duplicate_penalty/std': 0.032433781772851944, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.4095819890499115, 'rewards/total_composite/std': 0.15033851563930511, 'reward': 0.4095819890499115, 'reward_std': 0.15033851563930511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16545505821704865, 'sampling/sampling_logp_difference/max': 1.5893678665161133, 'sampling/importance_sampling_ratio/min': 0.2040545642375946, 'sampling/importance_sampling_ratio/mean': 1.0241625308990479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9418702125549316, 'clip_ratio/low_mean': 0.042293233796954155, 'clip_ratio/low_min': 0.042293233796954155, 'clip_ratio/high_mean': 0.07090657064691186, 'clip_ratio/high_max': 0.07090657064691186, 'clip_ratio/region_mean': 0.11319980444386601, 'reward_total_mean': 0.4095819890499115, 'reward_meter_mean': 0.8653483390808105, 'reward_meter_std': 0.24270948767662048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9800074696540833, 'reward_repeat_penalty_std': 0.032433781772851944, 'reward_repeat_floor_mean': 0.9970011115074158, 'reward_repeat_floor_std': 0.0048650591634213924, 'reward_near_duplicate_penalty_mean': 0.9800074696540833, 'reward_near_duplicate_penalty_std': 0.032433781772851944, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.47499996423721313, 'reward_judge_quality_std': 0.1164964810013771, 'reward_total_composite_mean': 0.4095819890499115, 'reward_total_composite_std': 0.15033851563930511, 'epoch': 0.08}

 46%|████▌     | 1509/3300 [5:47:44<6:23:29, 12.85s/it]INFO 04-12 21:46:10 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1510/3300 [5:47:56<6:20:01, 12.74s/it]2026-04-12 21:46:23,237 | INFO | train_grpo_train | metrics_logged mode=train step=1510

                                                       
{'loss': -0.0542, 'grad_norm': 3.9751882553100586, 'learning_rate': 5.427272727272728e-06, 'num_tokens': 2668030.0, 'completions/mean_length': 97.625, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.42857360839844, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.533435046672821, 'rewards/meter/std': 0.4034217596054077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9351696372032166, 'rewards/lexical_plausibility/std': 0.18336790800094604, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.20469054579734802, 'rewards/repeat_penalty/mean': 0.8853815793991089, 'rewards/repeat_penalty/std': 0.07688191533088684, 'rewards/repeat_floor/mean': 0.9850950241088867, 'rewards/repeat_floor/std': 0.010502473451197147, 'rewards/near_duplicate_penalty/mean': 0.9283722639083862, 'rewards/near_duplicate_penalty/std': 0.06372316926717758, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.953769326210022, 'rewards/distinct_2/std': 0.05108838528394699, 'rewards/total_composite/mean': 0.21596336364746094, 'rewards/total_composite/std': 0.18015703558921814, 'reward': 0.21596336364746094, 'reward_std': 0.18015703558921814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11728202551603317, 'sampling/sampling_logp_difference/max': 0.9407529830932617, 'sampling/importance_sampling_ratio/min': 0.39033380150794983, 'sampling/importance_sampling_ratio/mean': 1.0317085981369019, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7383372858166695, 'clip_ratio/low_mean': 0.05126689188182354, 'clip_ratio/low_min': 0.05126689188182354, 'clip_ratio/high_mean': 0.044203305849805474, 'clip_ratio/high_max': 0.044203305849805474, 'clip_ratio/region_mean': 0.09547019773162901, 'reward_total_mean': 0.21596336364746094, 'reward_meter_mean': 0.533435046672821, 'reward_meter_std': 0.4034217596054077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9351696372032166, 'reward_lexical_plausibility_std': 0.18336790800094604, 'reward_repeat_penalty_mean': 0.8853815793991089, 'reward_repeat_penalty_std': 0.07688191533088684, 'reward_repeat_floor_mean': 0.9850950241088867, 'reward_repeat_floor_std': 0.010502473451197147, 'reward_near_duplicate_penalty_mean': 0.9283722639083862, 'reward_near_duplicate_penalty_std': 0.06372316926717758, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.953769326210022, 'reward_distinct_2_std': 0.05108838528394699, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.20469054579734802, 'reward_total_composite_mean': 0.21596336364746094, 'reward_total_composite_std': 0.18015703558921814, 'epoch': 0.08}

 46%|████▌     | 1510/3300 [5:47:57<6:20:01, 12.74s/it]INFO 04-12 21:46:23 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1511/3300 [5:48:08<6:13:33, 12.53s/it]2026-04-12 21:46:35,254 | INFO | train_grpo_train | metrics_logged mode=train step=1511

                                                       
{'loss': -0.1036, 'grad_norm': 6.0721116065979, 'learning_rate': 5.424242424242425e-06, 'num_tokens': 2670306.0, 'completions/mean_length': 156.5, 'completions/min_length': 95.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 105.71428680419922, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.5891748666763306, 'rewards/meter/std': 0.3977961540222168, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.2878147065639496, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.90625, 'rewards/count_floor/std': 0.086344413459301, 'rewards/lexical_plausibility/mean': 0.9368640780448914, 'rewards/lexical_plausibility/std': 0.17857539653778076, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/repeat_penalty/mean': 0.9631444215774536, 'rewards/repeat_penalty/std': 0.034029312431812286, 'rewards/repeat_floor/mean': 0.9954419136047363, 'rewards/repeat_floor/std': 0.004139741417020559, 'rewards/near_duplicate_penalty/mean': 0.9800870418548584, 'rewards/near_duplicate_penalty/std': 0.019236844033002853, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9825426340103149, 'rewards/distinct_2/std': 0.021234121173620224, 'rewards/total_composite/mean': 0.2614912986755371, 'rewards/total_composite/std': 0.20411521196365356, 'reward': 0.2614912986755371, 'reward_std': 0.20411522686481476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12111324071884155, 'sampling/sampling_logp_difference/max': 2.742553949356079, 'sampling/importance_sampling_ratio/min': 0.0644056424498558, 'sampling/importance_sampling_ratio/mean': 1.0143250226974487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.752310611307621, 'clip_ratio/low_mean': 0.03820488788187504, 'clip_ratio/low_min': 0.03820488788187504, 'clip_ratio/high_mean': 0.05722587741911411, 'clip_ratio/high_max': 0.05722587741911411, 'clip_ratio/region_mean': 0.09543076530098915, 'reward_total_mean': 0.2614912986755371, 'reward_meter_mean': 0.5891748666763306, 'reward_meter_std': 0.3977961540222168, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.2878147065639496, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.90625, 'reward_count_floor_std': 0.086344413459301, 'reward_lexical_plausibility_mean': 0.9368640780448914, 'reward_lexical_plausibility_std': 0.17857539653778076, 'reward_repeat_penalty_mean': 0.9631444215774536, 'reward_repeat_penalty_std': 0.034029312431812286, 'reward_repeat_floor_mean': 0.9954419136047363, 'reward_repeat_floor_std': 0.004139741417020559, 'reward_near_duplicate_penalty_mean': 0.9800870418548584, 'reward_near_duplicate_penalty_std': 0.019236844033002853, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9825426340103149, 'reward_distinct_2_std': 0.021234121173620224, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.2614912986755371, 'reward_total_composite_std': 0.20411521196365356, 'epoch': 0.08}

 46%|████▌     | 1511/3300 [5:48:09<6:13:33, 12.53s/it]INFO 04-12 21:46:35 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1512/3300 [5:48:20<6:05:07, 12.25s/it]2026-04-12 21:46:47,030 | INFO | train_grpo_train | metrics_logged mode=train step=1512

                                                       
{'loss': -0.0397, 'grad_norm': 5.631893634796143, 'learning_rate': 5.421212121212122e-06, 'num_tokens': 2671739.0, 'completions/mean_length': 91.125, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.000001907348633, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.2916238307952881, 'rewards/meter/std': 0.23950965702533722, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.950294554233551, 'rewards/lexical_plausibility/std': 0.1405882090330124, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.28965190052986145, 'rewards/repeat_penalty/mean': 0.9739413261413574, 'rewards/repeat_penalty/std': 0.025588728487491608, 'rewards/repeat_floor/mean': 0.9961215257644653, 'rewards/repeat_floor/std': 0.003854460082948208, 'rewards/near_duplicate_penalty/mean': 0.9744541645050049, 'rewards/near_duplicate_penalty/std': 0.025886302813887596, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9994816780090332, 'rewards/distinct_2/std': 0.0014661203604191542, 'rewards/total_composite/mean': 0.13043023645877838, 'rewards/total_composite/std': 0.10295972228050232, 'reward': 0.13043023645877838, 'reward_std': 0.10295972228050232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16739794611930847, 'sampling/sampling_logp_difference/max': 2.154484748840332, 'sampling/importance_sampling_ratio/min': 0.1159629225730896, 'sampling/importance_sampling_ratio/mean': 0.9954019784927368, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7971791326999664, 'clip_ratio/low_mean': 0.02935222629457712, 'clip_ratio/low_min': 0.02935222629457712, 'clip_ratio/high_mean': 0.08494540583342314, 'clip_ratio/high_max': 0.08494540583342314, 'clip_ratio/region_mean': 0.11429763212800026, 'reward_total_mean': 0.13043023645877838, 'reward_meter_mean': 0.2916238307952881, 'reward_meter_std': 0.23950965702533722, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.950294554233551, 'reward_lexical_plausibility_std': 0.1405882090330124, 'reward_repeat_penalty_mean': 0.9739413261413574, 'reward_repeat_penalty_std': 0.025588728487491608, 'reward_repeat_floor_mean': 0.9961215257644653, 'reward_repeat_floor_std': 0.003854460082948208, 'reward_near_duplicate_penalty_mean': 0.9744541645050049, 'reward_near_duplicate_penalty_std': 0.025886302813887596, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9994816780090332, 'reward_distinct_2_std': 0.0014661203604191542, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.28965190052986145, 'reward_total_composite_mean': 0.13043023645877838, 'reward_total_composite_std': 0.10295972228050232, 'epoch': 0.08}

 46%|████▌     | 1512/3300 [5:48:20<6:05:07, 12.25s/it]INFO 04-12 21:46:47 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1513/3300 [5:48:28<5:27:15, 10.99s/it]2026-04-12 21:46:54,930 | INFO | train_grpo_train | metrics_logged mode=train step=1513

                                                       
{'loss': 0.0059, 'grad_norm': 11.002195358276367, 'learning_rate': 5.418181818181819e-06, 'num_tokens': 2673848.0, 'completions/mean_length': 94.625, 'completions/min_length': 86.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.625, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.822895884513855, 'rewards/meter/std': 0.21343597769737244, 'rewards/count_adherence/mean': 0.7708333134651184, 'rewards/count_adherence/std': 0.0862581729888916, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9312499761581421, 'rewards/count_floor/std': 0.025877464562654495, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6074999570846558, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/repeat_penalty/mean': 0.8126115202903748, 'rewards/repeat_penalty/std': 0.13901011645793915, 'rewards/repeat_floor/mean': 0.9795634746551514, 'rewards/repeat_floor/std': 0.01371503435075283, 'rewards/near_duplicate_penalty/mean': 0.9438788294792175, 'rewards/near_duplicate_penalty/std': 0.030512496829032898, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.866461992263794, 'rewards/distinct_2/std': 0.1032836064696312, 'rewards/total_composite/mean': 0.4423291087150574, 'rewards/total_composite/std': 0.1346423327922821, 'reward': 0.4423291087150574, 'reward_std': 0.1346423327922821, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13650833070278168, 'sampling/sampling_logp_difference/max': 2.0763885974884033, 'sampling/importance_sampling_ratio/min': 0.125382199883461, 'sampling/importance_sampling_ratio/mean': 1.0190232992172241, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7314606606960297, 'clip_ratio/low_mean': 0.07198336906731129, 'clip_ratio/low_min': 0.07198336906731129, 'clip_ratio/high_mean': 0.048810635693371296, 'clip_ratio/high_max': 0.048810635693371296, 'clip_ratio/region_mean': 0.12079400476068258, 'reward_total_mean': 0.4423291087150574, 'reward_meter_mean': 0.822895884513855, 'reward_meter_std': 0.21343597769737244, 'reward_count_adherence_mean': 0.7708333134651184, 'reward_count_adherence_std': 0.0862581729888916, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9312499761581421, 'reward_count_floor_std': 0.025877464562654495, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8126115202903748, 'reward_repeat_penalty_std': 0.13901011645793915, 'reward_repeat_floor_mean': 0.9795634746551514, 'reward_repeat_floor_std': 0.01371503435075283, 'reward_near_duplicate_penalty_mean': 0.9438788294792175, 'reward_near_duplicate_penalty_std': 0.030512496829032898, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.866461992263794, 'reward_distinct_2_std': 0.1032836064696312, 'reward_judge_quality_mean': 0.6074999570846558, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.4423291087150574, 'reward_total_composite_std': 0.1346423327922821, 'epoch': 0.08}

 46%|████▌     | 1513/3300 [5:48:28<5:27:15, 10.99s/it]INFO 04-12 21:46:55 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1514/3300 [5:48:40<5:33:28, 11.20s/it]2026-04-12 21:47:06,584 | INFO | train_grpo_train | metrics_logged mode=train step=1514

                                                       
{'loss': -0.0423, 'grad_norm': 4.147073745727539, 'learning_rate': 5.415151515151515e-06, 'num_tokens': 2675335.0, 'completions/mean_length': 93.875, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.142860412597656, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.47963839769363403, 'rewards/meter/std': 0.36328837275505066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9479241967201233, 'rewards/lexical_plausibility/std': 0.1472925841808319, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.1345893144607544, 'rewards/repeat_penalty/mean': 0.7378515005111694, 'rewards/repeat_penalty/std': 0.2528228461742401, 'rewards/repeat_floor/mean': 0.9738674163818359, 'rewards/repeat_floor/std': 0.023671019822359085, 'rewards/near_duplicate_penalty/mean': 0.9177777767181396, 'rewards/near_duplicate_penalty/std': 0.06464187055826187, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.9300079345703125, 'rewards/distinct_2/std': 0.09348642826080322, 'rewards/total_composite/mean': 0.15357424318790436, 'rewards/total_composite/std': 0.1161903589963913, 'reward': 0.15357424318790436, 'reward_std': 0.1161903589963913, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15072131156921387, 'sampling/sampling_logp_difference/max': 1.581021785736084, 'sampling/importance_sampling_ratio/min': 0.2057647407054901, 'sampling/importance_sampling_ratio/mean': 1.0171388387680054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8746337741613388, 'clip_ratio/low_mean': 0.036672474816441536, 'clip_ratio/low_min': 0.036672474816441536, 'clip_ratio/high_mean': 0.0700973947532475, 'clip_ratio/high_max': 0.0700973947532475, 'clip_ratio/region_mean': 0.10676986956968904, 'reward_total_mean': 0.15357424318790436, 'reward_meter_mean': 0.47963839769363403, 'reward_meter_std': 0.36328837275505066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9479241967201233, 'reward_lexical_plausibility_std': 0.1472925841808319, 'reward_repeat_penalty_mean': 0.7378515005111694, 'reward_repeat_penalty_std': 0.2528228461742401, 'reward_repeat_floor_mean': 0.9738674163818359, 'reward_repeat_floor_std': 0.023671019822359085, 'reward_near_duplicate_penalty_mean': 0.9177777767181396, 'reward_near_duplicate_penalty_std': 0.06464187055826187, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.9300079345703125, 'reward_distinct_2_std': 0.09348642826080322, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.1345893144607544, 'reward_total_composite_mean': 0.15357424318790436, 'reward_total_composite_std': 0.1161903589963913, 'epoch': 0.08}

 46%|████▌     | 1514/3300 [5:48:40<5:33:28, 11.20s/it]INFO 04-12 21:47:06 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1515/3300 [5:48:46<4:49:27,  9.73s/it]2026-04-12 21:47:12,924 | INFO | train_grpo_train | metrics_logged mode=train step=1515

                                                       
{'loss': -0.0147, 'grad_norm': 18.05291175842285, 'learning_rate': 5.412121212121213e-06, 'num_tokens': 2676658.0, 'completions/mean_length': 16.375, 'completions/min_length': 15.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.375, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.6703294515609741, 'rewards/meter/std': 0.4293660819530487, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/repeat_penalty/mean': 0.7236356735229492, 'rewards/repeat_penalty/std': 0.04696202650666237, 'rewards/repeat_floor/mean': 0.9797271490097046, 'rewards/repeat_floor/std': 0.009392405860126019, 'rewards/near_duplicate_penalty/mean': 0.9648475646972656, 'rewards/near_duplicate_penalty/std': 0.06261603534221649, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.28660327196121216, 'rewards/total_composite/std': 0.186248779296875, 'reward': 0.28660327196121216, 'reward_std': 0.1862487643957138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1236686110496521, 'sampling/sampling_logp_difference/max': 1.156153678894043, 'sampling/importance_sampling_ratio/min': 0.31469428539276123, 'sampling/importance_sampling_ratio/mean': 1.0033175945281982, 'sampling/importance_sampling_ratio/max': 1.6952159404754639, 'entropy': 0.7346522063016891, 'clip_ratio/low_mean': 0.0474571087397635, 'clip_ratio/low_min': 0.0474571087397635, 'clip_ratio/high_mean': 0.060661765281111, 'clip_ratio/high_max': 0.060661765281111, 'clip_ratio/region_mean': 0.1081188740208745, 'reward_total_mean': 0.28660327196121216, 'reward_meter_mean': 0.6703294515609741, 'reward_meter_std': 0.4293660819530487, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7236356735229492, 'reward_repeat_penalty_std': 0.04696202650666237, 'reward_repeat_floor_mean': 0.9797271490097046, 'reward_repeat_floor_std': 0.009392405860126019, 'reward_near_duplicate_penalty_mean': 0.9648475646972656, 'reward_near_duplicate_penalty_std': 0.06261603534221649, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.28660327196121216, 'reward_total_composite_std': 0.186248779296875, 'epoch': 0.08}

 46%|████▌     | 1515/3300 [5:48:46<4:49:27,  9.73s/it]INFO 04-12 21:47:13 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1516/3300 [5:48:54<4:36:11,  9.29s/it]2026-04-12 21:47:21,136 | INFO | train_grpo_train | metrics_logged mode=train step=1516

                                                       
{'loss': -0.0229, 'grad_norm': 15.29035758972168, 'learning_rate': 5.409090909090909e-06, 'num_tokens': 2678562.0, 'completions/mean_length': 62.0, 'completions/min_length': 50.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9803504347801208, 'rewards/meter/std': 0.00986937154084444, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955310225486755, 'rewards/repeat_penalty/mean': 0.7859803438186646, 'rewards/repeat_penalty/std': 0.11170168220996857, 'rewards/repeat_floor/mean': 0.9761965274810791, 'rewards/repeat_floor/std': 0.012812917120754719, 'rewards/near_duplicate_penalty/mean': 0.938115119934082, 'rewards/near_duplicate_penalty/std': 0.03744295984506607, 'rewards/opening_diversity/mean': 0.9750000238418579, 'rewards/opening_diversity/std': 0.04629101976752281, 'rewards/distinct_2/mean': 0.8553248643875122, 'rewards/distinct_2/std': 0.0700458213686943, 'rewards/total_composite/mean': 0.5239804983139038, 'rewards/total_composite/std': 0.20022590458393097, 'reward': 0.5239804983139038, 'reward_std': 0.20022590458393097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1125568151473999, 'sampling/sampling_logp_difference/max': 1.176687240600586, 'sampling/importance_sampling_ratio/min': 0.30829837918281555, 'sampling/importance_sampling_ratio/mean': 1.0040172338485718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8430889993906021, 'clip_ratio/low_mean': 0.06538659706711769, 'clip_ratio/low_min': 0.06538659706711769, 'clip_ratio/high_mean': 0.0486111119389534, 'clip_ratio/high_max': 0.0486111119389534, 'clip_ratio/region_mean': 0.11399770900607109, 'reward_total_mean': 0.5239804983139038, 'reward_meter_mean': 0.9803504347801208, 'reward_meter_std': 0.00986937154084444, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7859803438186646, 'reward_repeat_penalty_std': 0.11170168220996857, 'reward_repeat_floor_mean': 0.9761965274810791, 'reward_repeat_floor_std': 0.012812917120754719, 'reward_near_duplicate_penalty_mean': 0.938115119934082, 'reward_near_duplicate_penalty_std': 0.03744295984506607, 'reward_opening_diversity_mean': 0.9750000238418579, 'reward_opening_diversity_std': 0.04629101976752281, 'reward_distinct_2_mean': 0.8553248643875122, 'reward_distinct_2_std': 0.0700458213686943, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955310225486755, 'reward_total_composite_mean': 0.5239804983139038, 'reward_total_composite_std': 0.20022590458393097, 'epoch': 0.08}

 46%|████▌     | 1516/3300 [5:48:54<4:36:11,  9.29s/it]INFO 04-12 21:47:21 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1517/3300 [5:49:07<5:03:38, 10.22s/it]2026-04-12 21:47:33,525 | INFO | train_grpo_train | metrics_logged mode=train step=1517

                                                       
{'loss': -0.0504, 'grad_norm': 4.940978050231934, 'learning_rate': 5.406060606060607e-06, 'num_tokens': 2679924.0, 'completions/mean_length': 91.25, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.142858505249023, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7435610294342041, 'rewards/meter/std': 0.3742123246192932, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9426130056381226, 'rewards/lexical_plausibility/std': 0.16231493651866913, 'rewards/judge_quality/mean': 0.3812500238418579, 'rewards/judge_quality/std': 0.13452960550785065, 'rewards/repeat_penalty/mean': 0.8528552651405334, 'rewards/repeat_penalty/std': 0.1847434788942337, 'rewards/repeat_floor/mean': 0.9854755401611328, 'rewards/repeat_floor/std': 0.018479827791452408, 'rewards/near_duplicate_penalty/mean': 0.9577878713607788, 'rewards/near_duplicate_penalty/std': 0.05102771893143654, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.971470832824707, 'rewards/distinct_2/std': 0.05287402868270874, 'rewards/total_composite/mean': 0.30584797263145447, 'rewards/total_composite/std': 0.1784461885690689, 'reward': 0.30584797263145447, 'reward_std': 0.1784461885690689, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12514343857765198, 'sampling/sampling_logp_difference/max': 1.3416500091552734, 'sampling/importance_sampling_ratio/min': 0.2614139914512634, 'sampling/importance_sampling_ratio/mean': 1.0078914165496826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5223672538995743, 'clip_ratio/low_mean': 0.014285714365541935, 'clip_ratio/low_min': 0.014285714365541935, 'clip_ratio/high_mean': 0.09308868274092674, 'clip_ratio/high_max': 0.09308868274092674, 'clip_ratio/region_mean': 0.10737439710646868, 'reward_total_mean': 0.30584797263145447, 'reward_meter_mean': 0.7435610294342041, 'reward_meter_std': 0.3742123246192932, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9426130056381226, 'reward_lexical_plausibility_std': 0.16231493651866913, 'reward_repeat_penalty_mean': 0.8528552651405334, 'reward_repeat_penalty_std': 0.1847434788942337, 'reward_repeat_floor_mean': 0.9854755401611328, 'reward_repeat_floor_std': 0.018479827791452408, 'reward_near_duplicate_penalty_mean': 0.9577878713607788, 'reward_near_duplicate_penalty_std': 0.05102771893143654, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.971470832824707, 'reward_distinct_2_std': 0.05287402868270874, 'reward_judge_quality_mean': 0.3812500238418579, 'reward_judge_quality_std': 0.13452960550785065, 'reward_total_composite_mean': 0.30584797263145447, 'reward_total_composite_std': 0.1784461885690689, 'epoch': 0.08}

 46%|████▌     | 1517/3300 [5:49:07<5:03:38, 10.22s/it]INFO 04-12 21:47:33 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1518/3300 [5:49:19<5:17:59, 10.71s/it]2026-04-12 21:47:45,357 | INFO | train_grpo_train | metrics_logged mode=train step=1518

                                                       
{'loss': -0.0533, 'grad_norm': 3.6835150718688965, 'learning_rate': 5.4030303030303036e-06, 'num_tokens': 2681565.0, 'completions/mean_length': 164.125, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 48.16666793823242, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.30135756731033325, 'rewards/meter/std': 0.32846879959106445, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.8866062760353088, 'rewards/lexical_plausibility/std': 0.21001556515693665, 'rewards/judge_quality/mean': 0.5, 'rewards/judge_quality/std': 0.33445051312446594, 'rewards/repeat_penalty/mean': 0.9626946449279785, 'rewards/repeat_penalty/std': 0.037383463233709335, 'rewards/repeat_floor/mean': 0.9948590397834778, 'rewards/repeat_floor/std': 0.0048016090877354145, 'rewards/near_duplicate_penalty/mean': 0.9707438349723816, 'rewards/near_duplicate_penalty/std': 0.027035821229219437, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9916387796401978, 'rewards/distinct_2/std': 0.0236490610986948, 'rewards/total_composite/mean': 0.20959311723709106, 'rewards/total_composite/std': 0.2918854057788849, 'reward': 0.20959311723709106, 'reward_std': 0.2918853759765625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13603657484054565, 'sampling/sampling_logp_difference/max': 1.2135319709777832, 'sampling/importance_sampling_ratio/min': 0.29714590311050415, 'sampling/importance_sampling_ratio/mean': 1.031583309173584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7734980285167694, 'clip_ratio/low_mean': 0.04430373292416334, 'clip_ratio/low_min': 0.04430373292416334, 'clip_ratio/high_mean': 0.0339311808347702, 'clip_ratio/high_max': 0.0339311808347702, 'clip_ratio/region_mean': 0.07823491375893354, 'reward_total_mean': 0.20959311723709106, 'reward_meter_mean': 0.30135756731033325, 'reward_meter_std': 0.32846879959106445, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.8866062760353088, 'reward_lexical_plausibility_std': 0.21001556515693665, 'reward_repeat_penalty_mean': 0.9626946449279785, 'reward_repeat_penalty_std': 0.037383463233709335, 'reward_repeat_floor_mean': 0.9948590397834778, 'reward_repeat_floor_std': 0.0048016090877354145, 'reward_near_duplicate_penalty_mean': 0.9707438349723816, 'reward_near_duplicate_penalty_std': 0.027035821229219437, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9916387796401978, 'reward_distinct_2_std': 0.0236490610986948, 'reward_judge_quality_mean': 0.5, 'reward_judge_quality_std': 0.33445051312446594, 'reward_total_composite_mean': 0.20959311723709106, 'reward_total_composite_std': 0.2918854057788849, 'epoch': 0.08}

 46%|████▌     | 1518/3300 [5:49:19<5:17:59, 10.71s/it]INFO 04-12 21:47:45 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1519/3300 [5:49:30<5:23:59, 10.91s/it]2026-04-12 21:47:56,754 | INFO | train_grpo_train | metrics_logged mode=train step=1519

                                                       
{'loss': -0.0457, 'grad_norm': 4.085677146911621, 'learning_rate': 5.400000000000001e-06, 'num_tokens': 2683274.0, 'completions/mean_length': 103.625, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.28571701049805, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7756665945053101, 'rewards/meter/std': 0.3952201008796692, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9567331075668335, 'rewards/lexical_plausibility/std': 0.12237721681594849, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.18845234811306, 'rewards/repeat_penalty/mean': 0.9726606607437134, 'rewards/repeat_penalty/std': 0.05877441540360451, 'rewards/repeat_floor/mean': 0.9966561794281006, 'rewards/repeat_floor/std': 0.006725291721522808, 'rewards/near_duplicate_penalty/mean': 0.9865835905075073, 'rewards/near_duplicate_penalty/std': 0.021426426246762276, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.3402079939842224, 'rewards/total_composite/std': 0.21770277619361877, 'reward': 0.3402079939842224, 'reward_std': 0.21770276129245758, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15535330772399902, 'sampling/sampling_logp_difference/max': 1.7656736373901367, 'sampling/importance_sampling_ratio/min': 0.17107151448726654, 'sampling/importance_sampling_ratio/mean': 0.9973017573356628, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9880257174372673, 'clip_ratio/low_mean': 0.028384079225361347, 'clip_ratio/low_min': 0.028384079225361347, 'clip_ratio/high_mean': 0.12003011628985405, 'clip_ratio/high_max': 0.12003011628985405, 'clip_ratio/region_mean': 0.1484141955152154, 'reward_total_mean': 0.3402079939842224, 'reward_meter_mean': 0.7756665945053101, 'reward_meter_std': 0.3952201008796692, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9567331075668335, 'reward_lexical_plausibility_std': 0.12237721681594849, 'reward_repeat_penalty_mean': 0.9726606607437134, 'reward_repeat_penalty_std': 0.05877441540360451, 'reward_repeat_floor_mean': 0.9966561794281006, 'reward_repeat_floor_std': 0.006725291721522808, 'reward_near_duplicate_penalty_mean': 0.9865835905075073, 'reward_near_duplicate_penalty_std': 0.021426426246762276, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.18845234811306, 'reward_total_composite_mean': 0.3402079939842224, 'reward_total_composite_std': 0.21770277619361877, 'epoch': 0.08}

 46%|████▌     | 1519/3300 [5:49:30<5:23:59, 10.91s/it]INFO 04-12 21:47:57 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1520/3300 [5:49:36<4:43:27,  9.55s/it]2026-04-12 21:48:03,145 | INFO | train_grpo_train | metrics_logged mode=train step=1520

                                                       
{'loss': 0.0235, 'grad_norm': 12.409919738769531, 'learning_rate': 5.396969696969697e-06, 'num_tokens': 2684852.0, 'completions/mean_length': 39.25, 'completions/min_length': 37.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7924405336380005, 'rewards/meter/std': 0.34022170305252075, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.7737500667572021, 'rewards/judge_quality/std': 0.22032040357589722, 'rewards/repeat_penalty/mean': 0.9224278330802917, 'rewards/repeat_penalty/std': 0.05909670516848564, 'rewards/repeat_floor/mean': 0.9906152486801147, 'rewards/repeat_floor/std': 0.006676937453448772, 'rewards/near_duplicate_penalty/mean': 0.9625922441482544, 'rewards/near_duplicate_penalty/std': 0.02779064141213894, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9580711126327515, 'rewards/distinct_2/std': 0.05110624432563782, 'rewards/total_composite/mean': 0.5837881565093994, 'rewards/total_composite/std': 0.30689722299575806, 'reward': 0.5837881565093994, 'reward_std': 0.30689719319343567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0965077131986618, 'sampling/sampling_logp_difference/max': 1.8102085590362549, 'sampling/importance_sampling_ratio/min': 0.1636200100183487, 'sampling/importance_sampling_ratio/mean': 1.0129787921905518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5542140938341618, 'clip_ratio/low_mean': 0.037042682990431786, 'clip_ratio/low_min': 0.037042682990431786, 'clip_ratio/high_mean': 0.04445488750934601, 'clip_ratio/high_max': 0.04445488750934601, 'clip_ratio/region_mean': 0.0814975704997778, 'reward_total_mean': 0.5837881565093994, 'reward_meter_mean': 0.7924405336380005, 'reward_meter_std': 0.34022170305252075, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9224278330802917, 'reward_repeat_penalty_std': 0.05909670516848564, 'reward_repeat_floor_mean': 0.9906152486801147, 'reward_repeat_floor_std': 0.006676937453448772, 'reward_near_duplicate_penalty_mean': 0.9625922441482544, 'reward_near_duplicate_penalty_std': 0.02779064141213894, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9580711126327515, 'reward_distinct_2_std': 0.05110624432563782, 'reward_judge_quality_mean': 0.7737500667572021, 'reward_judge_quality_std': 0.22032040357589722, 'reward_total_composite_mean': 0.5837881565093994, 'reward_total_composite_std': 0.30689722299575806, 'epoch': 0.08}

 46%|████▌     | 1520/3300 [5:49:37<4:43:27,  9.55s/it]INFO 04-12 21:48:03 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1521/3300 [5:49:43<4:16:18,  8.64s/it]2026-04-12 21:48:09,713 | INFO | train_grpo_train | metrics_logged mode=train step=1521

                                                       
{'loss': 0.0872, 'grad_norm': 14.308231353759766, 'learning_rate': 5.3939393939393945e-06, 'num_tokens': 2686391.0, 'completions/mean_length': 40.375, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7918498516082764, 'rewards/meter/std': 0.26750698685646057, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/repeat_penalty/mean': 0.8959201574325562, 'rewards/repeat_penalty/std': 0.09392093122005463, 'rewards/repeat_floor/mean': 0.9866591691970825, 'rewards/repeat_floor/std': 0.011625532060861588, 'rewards/near_duplicate_penalty/mean': 0.9387994408607483, 'rewards/near_duplicate_penalty/std': 0.05887793004512787, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9537689685821533, 'rewards/distinct_2/std': 0.07117754966020584, 'rewards/total_composite/mean': 0.49097132682800293, 'rewards/total_composite/std': 0.28832054138183594, 'reward': 0.49097132682800293, 'reward_std': 0.28832054138183594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14265866577625275, 'sampling/sampling_logp_difference/max': 1.7881091833114624, 'sampling/importance_sampling_ratio/min': 0.16727615892887115, 'sampling/importance_sampling_ratio/mean': 1.0254676342010498, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9759261161088943, 'clip_ratio/low_mean': 0.06911526061594486, 'clip_ratio/low_min': 0.06911526061594486, 'clip_ratio/high_mean': 0.07159130461513996, 'clip_ratio/high_max': 0.07159130461513996, 'clip_ratio/region_mean': 0.14070656523108482, 'reward_total_mean': 0.49097132682800293, 'reward_meter_mean': 0.7918498516082764, 'reward_meter_std': 0.26750698685646057, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8959201574325562, 'reward_repeat_penalty_std': 0.09392093122005463, 'reward_repeat_floor_mean': 0.9866591691970825, 'reward_repeat_floor_std': 0.011625532060861588, 'reward_near_duplicate_penalty_mean': 0.9387994408607483, 'reward_near_duplicate_penalty_std': 0.05887793004512787, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9537689685821533, 'reward_distinct_2_std': 0.07117754966020584, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.49097132682800293, 'reward_total_composite_std': 0.28832054138183594, 'epoch': 0.08}

 46%|████▌     | 1521/3300 [5:49:43<4:16:18,  8.64s/it]INFO 04-12 21:48:10 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1522/3300 [5:49:50<4:01:06,  8.14s/it]2026-04-12 21:48:16,645 | INFO | train_grpo_train | metrics_logged mode=train step=1522

                                                       
{'loss': -0.0108, 'grad_norm': 10.884950637817383, 'learning_rate': 5.390909090909091e-06, 'num_tokens': 2688332.0, 'completions/mean_length': 72.625, 'completions/min_length': 58.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.946377694606781, 'rewards/meter/std': 0.04489784687757492, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.1524970978498459, 'rewards/repeat_penalty/mean': 0.6585756540298462, 'rewards/repeat_penalty/std': 0.1976325511932373, 'rewards/repeat_floor/mean': 0.9586641192436218, 'rewards/repeat_floor/std': 0.026731295511126518, 'rewards/near_duplicate_penalty/mean': 0.8806854486465454, 'rewards/near_duplicate_penalty/std': 0.08739843219518661, 'rewards/opening_diversity/mean': 0.9140625, 'rewards/opening_diversity/std': 0.10531362891197205, 'rewards/distinct_2/mean': 0.7968617677688599, 'rewards/distinct_2/std': 0.10690400749444962, 'rewards/total_composite/mean': 0.47932302951812744, 'rewards/total_composite/std': 0.14711469411849976, 'reward': 0.47932302951812744, 'reward_std': 0.14711467921733856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13120333850383759, 'sampling/sampling_logp_difference/max': 2.2665791511535645, 'sampling/importance_sampling_ratio/min': 0.10366620123386383, 'sampling/importance_sampling_ratio/mean': 1.0189564228057861, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7270428612828255, 'clip_ratio/low_mean': 0.06438386719673872, 'clip_ratio/low_min': 0.06438386719673872, 'clip_ratio/high_mean': 0.04815789498388767, 'clip_ratio/high_max': 0.04815789498388767, 'clip_ratio/region_mean': 0.11254176218062639, 'reward_total_mean': 0.47932302951812744, 'reward_meter_mean': 0.946377694606781, 'reward_meter_std': 0.04489784687757492, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6585756540298462, 'reward_repeat_penalty_std': 0.1976325511932373, 'reward_repeat_floor_mean': 0.9586641192436218, 'reward_repeat_floor_std': 0.026731295511126518, 'reward_near_duplicate_penalty_mean': 0.8806854486465454, 'reward_near_duplicate_penalty_std': 0.08739843219518661, 'reward_opening_diversity_mean': 0.9140625, 'reward_opening_diversity_std': 0.10531362891197205, 'reward_distinct_2_mean': 0.7968617677688599, 'reward_distinct_2_std': 0.10690400749444962, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.1524970978498459, 'reward_total_composite_mean': 0.47932302951812744, 'reward_total_composite_std': 0.14711469411849976, 'epoch': 0.08}

 46%|████▌     | 1522/3300 [5:49:50<4:01:06,  8.14s/it]INFO 04-12 21:48:17 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1523/3300 [5:50:02<4:32:30,  9.20s/it]2026-04-12 21:48:28,298 | INFO | train_grpo_train | metrics_logged mode=train step=1523

                                                       
{'loss': -0.0684, 'grad_norm': 3.554558753967285, 'learning_rate': 5.387878787878789e-06, 'num_tokens': 2690015.0, 'completions/mean_length': 169.375, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 55.16666793823242, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.465198814868927, 'rewards/meter/std': 0.39417970180511475, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.925000011920929, 'rewards/count_floor/std': 0.1060660183429718, 'rewards/lexical_plausibility/mean': 0.8780081272125244, 'rewards/lexical_plausibility/std': 0.22701644897460938, 'rewards/judge_quality/mean': 0.32124999165534973, 'rewards/judge_quality/std': 0.18333710730075836, 'rewards/repeat_penalty/mean': 0.9469808340072632, 'rewards/repeat_penalty/std': 0.060167331248521805, 'rewards/repeat_floor/mean': 0.9938617944717407, 'rewards/repeat_floor/std': 0.0064409468322992325, 'rewards/near_duplicate_penalty/mean': 0.9790488481521606, 'rewards/near_duplicate_penalty/std': 0.016430774703621864, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9667167663574219, 'rewards/distinct_2/std': 0.04897624999284744, 'rewards/total_composite/mean': 0.17819607257843018, 'rewards/total_composite/std': 0.17153175175189972, 'reward': 0.17819607257843018, 'reward_std': 0.17153175175189972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1527625173330307, 'sampling/sampling_logp_difference/max': 2.240691661834717, 'sampling/importance_sampling_ratio/min': 0.10638489574193954, 'sampling/importance_sampling_ratio/mean': 1.0010536909103394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6078944504261017, 'clip_ratio/low_mean': 0.02384559903293848, 'clip_ratio/low_min': 0.02384559903293848, 'clip_ratio/high_mean': 0.07753779739141464, 'clip_ratio/high_max': 0.07753779739141464, 'clip_ratio/region_mean': 0.10138339642435312, 'reward_total_mean': 0.17819607257843018, 'reward_meter_mean': 0.465198814868927, 'reward_meter_std': 0.39417970180511475, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.925000011920929, 'reward_count_floor_std': 0.1060660183429718, 'reward_lexical_plausibility_mean': 0.8780081272125244, 'reward_lexical_plausibility_std': 0.22701644897460938, 'reward_repeat_penalty_mean': 0.9469808340072632, 'reward_repeat_penalty_std': 0.060167331248521805, 'reward_repeat_floor_mean': 0.9938617944717407, 'reward_repeat_floor_std': 0.0064409468322992325, 'reward_near_duplicate_penalty_mean': 0.9790488481521606, 'reward_near_duplicate_penalty_std': 0.016430774703621864, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9667167663574219, 'reward_distinct_2_std': 0.04897624999284744, 'reward_judge_quality_mean': 0.32124999165534973, 'reward_judge_quality_std': 0.18333710730075836, 'reward_total_composite_mean': 0.17819607257843018, 'reward_total_composite_std': 0.17153175175189972, 'epoch': 0.08}

 46%|████▌     | 1523/3300 [5:50:02<4:32:30,  9.20s/it]INFO 04-12 21:48:28 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1524/3300 [5:50:08<4:10:23,  8.46s/it]2026-04-12 21:48:35,014 | INFO | train_grpo_train | metrics_logged mode=train step=1524

                                                       
{'loss': 0.0089, 'grad_norm': 10.494470596313477, 'learning_rate': 5.384848484848485e-06, 'num_tokens': 2691956.0, 'completions/mean_length': 75.625, 'completions/min_length': 65.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.6106592416763306, 'rewards/meter/std': 0.34382274746894836, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.03471824526786804, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.1566559225320816, 'rewards/repeat_penalty/mean': 0.833951473236084, 'rewards/repeat_penalty/std': 0.10136827826499939, 'rewards/repeat_floor/mean': 0.9835786819458008, 'rewards/repeat_floor/std': 0.009157607331871986, 'rewards/near_duplicate_penalty/mean': 0.9619607329368591, 'rewards/near_duplicate_penalty/std': 0.01770230010151863, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8809397220611572, 'rewards/distinct_2/std': 0.075276218354702, 'rewards/total_composite/mean': 0.3298325538635254, 'rewards/total_composite/std': 0.22314615547657013, 'reward': 0.3298325538635254, 'reward_std': 0.22314615547657013, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14376334846019745, 'sampling/sampling_logp_difference/max': 1.522964358329773, 'sampling/importance_sampling_ratio/min': 0.21806451678276062, 'sampling/importance_sampling_ratio/mean': 1.0001769065856934, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6671876981854439, 'clip_ratio/low_mean': 0.06491773435845971, 'clip_ratio/low_min': 0.06491773435845971, 'clip_ratio/high_mean': 0.04397420305758715, 'clip_ratio/high_max': 0.04397420305758715, 'clip_ratio/region_mean': 0.10889193741604686, 'reward_total_mean': 0.3298325538635254, 'reward_meter_mean': 0.6106592416763306, 'reward_meter_std': 0.34382274746894836, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.03471824526786804, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.833951473236084, 'reward_repeat_penalty_std': 0.10136827826499939, 'reward_repeat_floor_mean': 0.9835786819458008, 'reward_repeat_floor_std': 0.009157607331871986, 'reward_near_duplicate_penalty_mean': 0.9619607329368591, 'reward_near_duplicate_penalty_std': 0.01770230010151863, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8809397220611572, 'reward_distinct_2_std': 0.075276218354702, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.1566559225320816, 'reward_total_composite_mean': 0.3298325538635254, 'reward_total_composite_std': 0.22314615547657013, 'epoch': 0.08}

 46%|████▌     | 1524/3300 [5:50:08<4:10:23,  8.46s/it]INFO 04-12 21:48:35 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1525/3300 [5:50:14<3:50:19,  7.79s/it]2026-04-12 21:48:41,235 | INFO | train_grpo_train | metrics_logged mode=train step=1525

                                                       
{'loss': 0.0859, 'grad_norm': 12.723620414733887, 'learning_rate': 5.381818181818183e-06, 'num_tokens': 2693578.0, 'completions/mean_length': 39.75, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9614500999450684, 'rewards/meter/std': 0.049186501652002335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5149999856948853, 'rewards/judge_quality/std': 0.1803172081708908, 'rewards/repeat_penalty/mean': 0.9003968238830566, 'rewards/repeat_penalty/std': 0.14561787247657776, 'rewards/repeat_floor/mean': 0.9882755279541016, 'rewards/repeat_floor/std': 0.015324287116527557, 'rewards/near_duplicate_penalty/mean': 0.9550296068191528, 'rewards/near_duplicate_penalty/std': 0.04247147589921951, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9655120372772217, 'rewards/distinct_2/std': 0.057574495673179626, 'rewards/total_composite/mean': 0.49208956956863403, 'rewards/total_composite/std': 0.1803610622882843, 'reward': 0.49208956956863403, 'reward_std': 0.1803610622882843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1361437290906906, 'sampling/sampling_logp_difference/max': 2.4802467823028564, 'sampling/importance_sampling_ratio/min': 0.0837225615978241, 'sampling/importance_sampling_ratio/mean': 0.9915043711662292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5932756625115871, 'clip_ratio/low_mean': 0.07083026552572846, 'clip_ratio/low_min': 0.07083026552572846, 'clip_ratio/high_mean': 0.039403016678988934, 'clip_ratio/high_max': 0.039403016678988934, 'clip_ratio/region_mean': 0.1102332822047174, 'reward_total_mean': 0.49208956956863403, 'reward_meter_mean': 0.9614500999450684, 'reward_meter_std': 0.049186501652002335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9003968238830566, 'reward_repeat_penalty_std': 0.14561787247657776, 'reward_repeat_floor_mean': 0.9882755279541016, 'reward_repeat_floor_std': 0.015324287116527557, 'reward_near_duplicate_penalty_mean': 0.9550296068191528, 'reward_near_duplicate_penalty_std': 0.04247147589921951, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9655120372772217, 'reward_distinct_2_std': 0.057574495673179626, 'reward_judge_quality_mean': 0.5149999856948853, 'reward_judge_quality_std': 0.1803172081708908, 'reward_total_composite_mean': 0.49208956956863403, 'reward_total_composite_std': 0.1803610622882843, 'epoch': 0.08}

 46%|████▌     | 1525/3300 [5:50:15<3:50:19,  7.79s/it]INFO 04-12 21:48:41 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1526/3300 [5:50:26<4:24:59,  8.96s/it]2026-04-12 21:48:52,959 | INFO | train_grpo_train | metrics_logged mode=train step=1526

                                                       
{'loss': -0.1279, 'grad_norm': 3.381234884262085, 'learning_rate': 5.378787878787879e-06, 'num_tokens': 2695563.0, 'completions/mean_length': 124.125, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 68.71428680419922, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8804305791854858, 'rewards/meter/std': 0.2439807653427124, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 0.9360970258712769, 'rewards/lexical_plausibility/std': 0.180744931101799, 'rewards/judge_quality/mean': 0.4150000214576721, 'rewards/judge_quality/std': 0.18031719326972961, 'rewards/repeat_penalty/mean': 0.8676546812057495, 'rewards/repeat_penalty/std': 0.1129741445183754, 'rewards/repeat_floor/mean': 0.9842111468315125, 'rewards/repeat_floor/std': 0.013527601025998592, 'rewards/near_duplicate_penalty/mean': 0.9460961818695068, 'rewards/near_duplicate_penalty/std': 0.045496616512537, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9144079685211182, 'rewards/distinct_2/std': 0.0847252905368805, 'rewards/total_composite/mean': 0.36114463210105896, 'rewards/total_composite/std': 0.19667382538318634, 'reward': 0.36114463210105896, 'reward_std': 0.19667382538318634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13826672732830048, 'sampling/sampling_logp_difference/max': 4.088789939880371, 'sampling/importance_sampling_ratio/min': 0.01675950177013874, 'sampling/importance_sampling_ratio/mean': 1.0007166862487793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6977327093482018, 'clip_ratio/low_mean': 0.02083333395421505, 'clip_ratio/low_min': 0.02083333395421505, 'clip_ratio/high_mean': 0.0925633879378438, 'clip_ratio/high_max': 0.0925633879378438, 'clip_ratio/region_mean': 0.11339672189205885, 'reward_total_mean': 0.36114463210105896, 'reward_meter_mean': 0.8804305791854858, 'reward_meter_std': 0.2439807653427124, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 0.9360970258712769, 'reward_lexical_plausibility_std': 0.180744931101799, 'reward_repeat_penalty_mean': 0.8676546812057495, 'reward_repeat_penalty_std': 0.1129741445183754, 'reward_repeat_floor_mean': 0.9842111468315125, 'reward_repeat_floor_std': 0.013527601025998592, 'reward_near_duplicate_penalty_mean': 0.9460961818695068, 'reward_near_duplicate_penalty_std': 0.045496616512537, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9144079685211182, 'reward_distinct_2_std': 0.0847252905368805, 'reward_judge_quality_mean': 0.4150000214576721, 'reward_judge_quality_std': 0.18031719326972961, 'reward_total_composite_mean': 0.36114463210105896, 'reward_total_composite_std': 0.19667382538318634, 'epoch': 0.08}

 46%|████▌     | 1526/3300 [5:50:26<4:24:59,  8.96s/it]INFO 04-12 21:48:53 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1527/3300 [5:50:38<4:47:08,  9.72s/it]2026-04-12 21:49:04,428 | INFO | train_grpo_train | metrics_logged mode=train step=1527

                                                       
{'loss': -0.0648, 'grad_norm': 2.375276565551758, 'learning_rate': 5.375757575757576e-06, 'num_tokens': 2696998.0, 'completions/mean_length': 159.375, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 41.833335876464844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6439677476882935, 'rewards/meter/std': 0.4702589809894562, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.75, 'rewards/arabic_clean/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.852933943271637, 'rewards/lexical_plausibility/std': 0.27269622683525085, 'rewards/judge_quality/mean': 0.3725000023841858, 'rewards/judge_quality/std': 0.22282280027866364, 'rewards/repeat_penalty/mean': 0.8634754419326782, 'rewards/repeat_penalty/std': 0.12386002391576767, 'rewards/repeat_floor/mean': 0.9863673448562622, 'rewards/repeat_floor/std': 0.013182880356907845, 'rewards/near_duplicate_penalty/mean': 0.9531238675117493, 'rewards/near_duplicate_penalty/std': 0.06239024177193642, 'rewards/opening_diversity/mean': 0.9375, 'rewards/opening_diversity/std': 0.1157275140285492, 'rewards/distinct_2/mean': 0.9683189392089844, 'rewards/distinct_2/std': 0.06502888351678848, 'rewards/total_composite/mean': 0.29991188645362854, 'rewards/total_composite/std': 0.2601035535335541, 'reward': 0.29991188645362854, 'reward_std': 0.2601035535335541, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1712958961725235, 'sampling/sampling_logp_difference/max': 1.1802854537963867, 'sampling/importance_sampling_ratio/min': 0.30719104409217834, 'sampling/importance_sampling_ratio/mean': 1.0464965105056763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0569485574960709, 'clip_ratio/low_mean': 0.01595744676887989, 'clip_ratio/low_min': 0.01595744676887989, 'clip_ratio/high_mean': 0.07877714512869716, 'clip_ratio/high_max': 0.07877714512869716, 'clip_ratio/region_mean': 0.09473459189757705, 'reward_total_mean': 0.29991188645362854, 'reward_meter_mean': 0.6439677476882935, 'reward_meter_std': 0.4702589809894562, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.75, 'reward_arabic_clean_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.852933943271637, 'reward_lexical_plausibility_std': 0.27269622683525085, 'reward_repeat_penalty_mean': 0.8634754419326782, 'reward_repeat_penalty_std': 0.12386002391576767, 'reward_repeat_floor_mean': 0.9863673448562622, 'reward_repeat_floor_std': 0.013182880356907845, 'reward_near_duplicate_penalty_mean': 0.9531238675117493, 'reward_near_duplicate_penalty_std': 0.06239024177193642, 'reward_opening_diversity_mean': 0.9375, 'reward_opening_diversity_std': 0.1157275140285492, 'reward_distinct_2_mean': 0.9683189392089844, 'reward_distinct_2_std': 0.06502888351678848, 'reward_judge_quality_mean': 0.3725000023841858, 'reward_judge_quality_std': 0.22282280027866364, 'reward_total_composite_mean': 0.29991188645362854, 'reward_total_composite_std': 0.2601035535335541, 'epoch': 0.08}

 46%|████▋     | 1527/3300 [5:50:38<4:47:08,  9.72s/it]INFO 04-12 21:49:04 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1528/3300 [5:50:49<5:01:39, 10.21s/it]2026-04-12 21:49:15,811 | INFO | train_grpo_train | metrics_logged mode=train step=1528

                                                       
{'loss': -0.042, 'grad_norm': 5.34360408782959, 'learning_rate': 5.372727272727273e-06, 'num_tokens': 2698262.0, 'completions/mean_length': 82.0, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.571430206298828, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.619267463684082, 'rewards/meter/std': 0.40426722168922424, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.1060660257935524, 'rewards/lexical_plausibility/mean': 0.9295322895050049, 'rewards/lexical_plausibility/std': 0.1754276156425476, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/repeat_penalty/mean': 0.7707452774047852, 'rewards/repeat_penalty/std': 0.09718868881464005, 'rewards/repeat_floor/mean': 0.9847740530967712, 'rewards/repeat_floor/std': 0.008511229418218136, 'rewards/near_duplicate_penalty/mean': 0.9859936237335205, 'rewards/near_duplicate_penalty/std': 0.03961591050028801, 'rewards/opening_diversity/mean': 0.78125, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 1.0, 'rewards/distinct_2/std': 0.0, 'rewards/total_composite/mean': 0.2586647868156433, 'rewards/total_composite/std': 0.1693427413702011, 'reward': 0.2586647868156433, 'reward_std': 0.1693427562713623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14291003346443176, 'sampling/sampling_logp_difference/max': 1.4642868041992188, 'sampling/importance_sampling_ratio/min': 0.23124286532402039, 'sampling/importance_sampling_ratio/mean': 1.034210443496704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7955408729612827, 'clip_ratio/low_mean': 0.023989899549633265, 'clip_ratio/low_min': 0.023989899549633265, 'clip_ratio/high_mean': 0.06154306326061487, 'clip_ratio/high_max': 0.06154306326061487, 'clip_ratio/region_mean': 0.08553296281024814, 'reward_total_mean': 0.2586647868156433, 'reward_meter_mean': 0.619267463684082, 'reward_meter_std': 0.40426722168922424, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.1060660257935524, 'reward_lexical_plausibility_mean': 0.9295322895050049, 'reward_lexical_plausibility_std': 0.1754276156425476, 'reward_repeat_penalty_mean': 0.7707452774047852, 'reward_repeat_penalty_std': 0.09718868881464005, 'reward_repeat_floor_mean': 0.9847740530967712, 'reward_repeat_floor_std': 0.008511229418218136, 'reward_near_duplicate_penalty_mean': 0.9859936237335205, 'reward_near_duplicate_penalty_std': 0.03961591050028801, 'reward_opening_diversity_mean': 0.78125, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 1.0, 'reward_distinct_2_std': 0.0, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.2586647868156433, 'reward_total_composite_std': 0.1693427413702011, 'epoch': 0.08}

 46%|████▋     | 1528/3300 [5:50:49<5:01:39, 10.21s/it]INFO 04-12 21:49:16 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1529/3300 [5:50:55<4:25:53,  9.01s/it]2026-04-12 21:49:21,993 | INFO | train_grpo_train | metrics_logged mode=train step=1529

                                                       
{'loss': 0.023, 'grad_norm': 18.271020889282227, 'learning_rate': 5.36969696969697e-06, 'num_tokens': 2699938.0, 'completions/mean_length': 37.5, 'completions/min_length': 34.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8695324063301086, 'rewards/meter/std': 0.26024389266967773, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.35999998450279236, 'rewards/judge_quality/std': 0.09165150672197342, 'rewards/repeat_penalty/mean': 0.7619063854217529, 'rewards/repeat_penalty/std': 0.24325452744960785, 'rewards/repeat_floor/mean': 0.9734892845153809, 'rewards/repeat_floor/std': 0.027102071791887283, 'rewards/near_duplicate_penalty/mean': 0.9235715866088867, 'rewards/near_duplicate_penalty/std': 0.07062521576881409, 'rewards/opening_diversity/mean': 0.875, 'rewards/opening_diversity/std': 0.13363061845302582, 'rewards/distinct_2/mean': 0.916150689125061, 'rewards/distinct_2/std': 0.10905730724334717, 'rewards/total_composite/mean': 0.29942119121551514, 'rewards/total_composite/std': 0.11908740550279617, 'reward': 0.29942119121551514, 'reward_std': 0.11908739805221558, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12328280508518219, 'sampling/sampling_logp_difference/max': 1.3762907981872559, 'sampling/importance_sampling_ratio/min': 0.25251346826553345, 'sampling/importance_sampling_ratio/mean': 1.0183333158493042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6989334225654602, 'clip_ratio/low_mean': 0.060727136209607124, 'clip_ratio/low_min': 0.060727136209607124, 'clip_ratio/high_mean': 0.06254699267446995, 'clip_ratio/high_max': 0.06254699267446995, 'clip_ratio/region_mean': 0.12327412888407707, 'reward_total_mean': 0.29942119121551514, 'reward_meter_mean': 0.8695324063301086, 'reward_meter_std': 0.26024389266967773, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7619063854217529, 'reward_repeat_penalty_std': 0.24325452744960785, 'reward_repeat_floor_mean': 0.9734892845153809, 'reward_repeat_floor_std': 0.027102071791887283, 'reward_near_duplicate_penalty_mean': 0.9235715866088867, 'reward_near_duplicate_penalty_std': 0.07062521576881409, 'reward_opening_diversity_mean': 0.875, 'reward_opening_diversity_std': 0.13363061845302582, 'reward_distinct_2_mean': 0.916150689125061, 'reward_distinct_2_std': 0.10905730724334717, 'reward_judge_quality_mean': 0.35999998450279236, 'reward_judge_quality_std': 0.09165150672197342, 'reward_total_composite_mean': 0.29942119121551514, 'reward_total_composite_std': 0.11908740550279617, 'epoch': 0.08}

 46%|████▋     | 1529/3300 [5:50:55<4:25:53,  9.01s/it]INFO 04-12 21:49:22 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1530/3300 [5:51:01<3:58:01,  8.07s/it]2026-04-12 21:49:27,867 | INFO | train_grpo_train | metrics_logged mode=train step=1530

                                                       
{'loss': -0.0312, 'grad_norm': 15.39715576171875, 'learning_rate': 5.366666666666666e-06, 'num_tokens': 2701348.0, 'completions/mean_length': 22.25, 'completions/min_length': 20.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6570097804069519, 'rewards/meter/std': 0.36284562945365906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.844009280204773, 'rewards/lexical_plausibility/std': 0.16080482304096222, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.738130509853363, 'rewards/repeat_penalty/std': 0.031143037602305412, 'rewards/repeat_floor/mean': 0.9835137128829956, 'rewards/repeat_floor/std': 0.003728361800312996, 'rewards/near_duplicate_penalty/mean': 0.9989669322967529, 'rewards/near_duplicate_penalty/std': 0.002921935636550188, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9852070808410645, 'rewards/distinct_2/std': 0.041840631514787674, 'rewards/total_composite/mean': 0.2716251015663147, 'rewards/total_composite/std': 0.15022878348827362, 'reward': 0.2716251015663147, 'reward_std': 0.15022878348827362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10107531398534775, 'sampling/sampling_logp_difference/max': 0.9791464805603027, 'sampling/importance_sampling_ratio/min': 0.37563157081604004, 'sampling/importance_sampling_ratio/mean': 1.0228551626205444, 'sampling/importance_sampling_ratio/max': 1.581586241722107, 'entropy': 0.9554580301046371, 'clip_ratio/low_mean': 0.018452381249517202, 'clip_ratio/low_min': 0.018452381249517202, 'clip_ratio/high_mean': 0.09584980458021164, 'clip_ratio/high_max': 0.09584980458021164, 'clip_ratio/region_mean': 0.11430218582972884, 'reward_total_mean': 0.2716251015663147, 'reward_meter_mean': 0.6570097804069519, 'reward_meter_std': 0.36284562945365906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.844009280204773, 'reward_lexical_plausibility_std': 0.16080482304096222, 'reward_repeat_penalty_mean': 0.738130509853363, 'reward_repeat_penalty_std': 0.031143037602305412, 'reward_repeat_floor_mean': 0.9835137128829956, 'reward_repeat_floor_std': 0.003728361800312996, 'reward_near_duplicate_penalty_mean': 0.9989669322967529, 'reward_near_duplicate_penalty_std': 0.002921935636550188, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9852070808410645, 'reward_distinct_2_std': 0.041840631514787674, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.2716251015663147, 'reward_total_composite_std': 0.15022878348827362, 'epoch': 0.08}

 46%|████▋     | 1530/3300 [5:51:01<3:58:01,  8.07s/it]INFO 04-12 21:49:28 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1531/3300 [5:51:07<3:41:37,  7.52s/it]2026-04-12 21:49:34,094 | INFO | train_grpo_train | metrics_logged mode=train step=1531

                                                       
{'loss': 0.1039, 'grad_norm': 24.86589813232422, 'learning_rate': 5.3636363636363645e-06, 'num_tokens': 2702624.0, 'completions/mean_length': 19.5, 'completions/min_length': 14.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.5760490894317627, 'rewards/meter/std': 0.3285415768623352, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/repeat_penalty/mean': 0.7218390107154846, 'rewards/repeat_penalty/std': 0.054339636117219925, 'rewards/repeat_floor/mean': 0.980582594871521, 'rewards/repeat_floor/std': 0.00819238182157278, 'rewards/near_duplicate_penalty/mean': 0.9833712577819824, 'rewards/near_duplicate_penalty/std': 0.03924896568059921, 'rewards/opening_diversity/mean': 0.75, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9786324501037598, 'rewards/distinct_2/std': 0.06043647602200508, 'rewards/total_composite/mean': 0.24635758996009827, 'rewards/total_composite/std': 0.1442408263683319, 'reward': 0.24635758996009827, 'reward_std': 0.1442408263683319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1799771934747696, 'sampling/sampling_logp_difference/max': 2.439213514328003, 'sampling/importance_sampling_ratio/min': 0.087229423224926, 'sampling/importance_sampling_ratio/mean': 0.9935725331306458, 'sampling/importance_sampling_ratio/max': 1.798694372177124, 'entropy': 0.9039500951766968, 'clip_ratio/low_mean': 0.07151016779243946, 'clip_ratio/low_min': 0.07151016779243946, 'clip_ratio/high_mean': 0.0652725575491786, 'clip_ratio/high_max': 0.0652725575491786, 'clip_ratio/region_mean': 0.13678272534161806, 'reward_total_mean': 0.24635758996009827, 'reward_meter_mean': 0.5760490894317627, 'reward_meter_std': 0.3285415768623352, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7218390107154846, 'reward_repeat_penalty_std': 0.054339636117219925, 'reward_repeat_floor_mean': 0.980582594871521, 'reward_repeat_floor_std': 0.00819238182157278, 'reward_near_duplicate_penalty_mean': 0.9833712577819824, 'reward_near_duplicate_penalty_std': 0.03924896568059921, 'reward_opening_diversity_mean': 0.75, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9786324501037598, 'reward_distinct_2_std': 0.06043647602200508, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.24635758996009827, 'reward_total_composite_std': 0.1442408263683319, 'epoch': 0.08}

 46%|████▋     | 1531/3300 [5:51:07<3:41:37,  7.52s/it]INFO 04-12 21:49:34 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1532/3300 [5:51:19<4:20:18,  8.83s/it]2026-04-12 21:49:46,010 | INFO | train_grpo_train | metrics_logged mode=train step=1532

                                                       
{'loss': -0.1104, 'grad_norm': 3.404844284057617, 'learning_rate': 5.360606060606061e-06, 'num_tokens': 2704641.0, 'completions/mean_length': 193.125, 'completions/min_length': 80.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 86.83333587646484, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.513882577419281, 'rewards/meter/std': 0.366549015045166, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9474999904632568, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.8779207468032837, 'rewards/lexical_plausibility/std': 0.2269989401102066, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.20817232131958008, 'rewards/repeat_penalty/mean': 0.8992990255355835, 'rewards/repeat_penalty/std': 0.060248106718063354, 'rewards/repeat_floor/mean': 0.9890507459640503, 'rewards/repeat_floor/std': 0.005993724800646305, 'rewards/near_duplicate_penalty/mean': 0.9720340371131897, 'rewards/near_duplicate_penalty/std': 0.01015514973551035, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.924951434135437, 'rewards/distinct_2/std': 0.05770958960056305, 'rewards/total_composite/mean': 0.1910392791032791, 'rewards/total_composite/std': 0.16687238216400146, 'reward': 0.1910392791032791, 'reward_std': 0.16687238216400146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1429230123758316, 'sampling/sampling_logp_difference/max': 2.1384706497192383, 'sampling/importance_sampling_ratio/min': 0.1178349107503891, 'sampling/importance_sampling_ratio/mean': 1.00698721408844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7193223834037781, 'clip_ratio/low_mean': 0.03324353415518999, 'clip_ratio/low_min': 0.03324353415518999, 'clip_ratio/high_mean': 0.07775518763810396, 'clip_ratio/high_max': 0.07775518763810396, 'clip_ratio/region_mean': 0.11099872179329395, 'reward_total_mean': 0.1910392791032791, 'reward_meter_mean': 0.513882577419281, 'reward_meter_std': 0.366549015045166, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9474999904632568, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.8779207468032837, 'reward_lexical_plausibility_std': 0.2269989401102066, 'reward_repeat_penalty_mean': 0.8992990255355835, 'reward_repeat_penalty_std': 0.060248106718063354, 'reward_repeat_floor_mean': 0.9890507459640503, 'reward_repeat_floor_std': 0.005993724800646305, 'reward_near_duplicate_penalty_mean': 0.9720340371131897, 'reward_near_duplicate_penalty_std': 0.01015514973551035, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.924951434135437, 'reward_distinct_2_std': 0.05770958960056305, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.20817232131958008, 'reward_total_composite_mean': 0.1910392791032791, 'reward_total_composite_std': 0.16687238216400146, 'epoch': 0.08}

 46%|████▋     | 1532/3300 [5:51:19<4:20:18,  8.83s/it]INFO 04-12 21:49:46 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1533/3300 [5:51:31<4:43:08,  9.61s/it]2026-04-12 21:49:57,450 | INFO | train_grpo_train | metrics_logged mode=train step=1533

                                                       
{'loss': -0.1195, 'grad_norm': 3.960251808166504, 'learning_rate': 5.357575757575758e-06, 'num_tokens': 2706610.0, 'completions/mean_length': 127.125, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 72.14286041259766, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.8311418294906616, 'rewards/meter/std': 0.22106477618217468, 'rewards/count_adherence/mean': 0.7750000357627869, 'rewards/count_adherence/std': 0.12817399203777313, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.03845219686627388, 'rewards/lexical_plausibility/mean': 0.9463132619857788, 'rewards/lexical_plausibility/std': 0.15184898674488068, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/repeat_penalty/mean': 0.8302541971206665, 'rewards/repeat_penalty/std': 0.15448428690433502, 'rewards/repeat_floor/mean': 0.9817454218864441, 'rewards/repeat_floor/std': 0.013800125569105148, 'rewards/near_duplicate_penalty/mean': 0.9470402598381042, 'rewards/near_duplicate_penalty/std': 0.023172305896878242, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8854376077651978, 'rewards/distinct_2/std': 0.12076510488986969, 'rewards/total_composite/mean': 0.2855883240699768, 'rewards/total_composite/std': 0.12879391014575958, 'reward': 0.2855883240699768, 'reward_std': 0.12879391014575958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14837683737277985, 'sampling/sampling_logp_difference/max': 2.100227117538452, 'sampling/importance_sampling_ratio/min': 0.12242861092090607, 'sampling/importance_sampling_ratio/mean': 1.0206729173660278, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8774540796875954, 'clip_ratio/low_mean': 0.035305808298289776, 'clip_ratio/low_min': 0.035305808298289776, 'clip_ratio/high_mean': 0.0816015973687172, 'clip_ratio/high_max': 0.0816015973687172, 'clip_ratio/region_mean': 0.11690740566700697, 'reward_total_mean': 0.2855883240699768, 'reward_meter_mean': 0.8311418294906616, 'reward_meter_std': 0.22106477618217468, 'reward_count_adherence_mean': 0.7750000357627869, 'reward_count_adherence_std': 0.12817399203777313, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.03845219686627388, 'reward_lexical_plausibility_mean': 0.9463132619857788, 'reward_lexical_plausibility_std': 0.15184898674488068, 'reward_repeat_penalty_mean': 0.8302541971206665, 'reward_repeat_penalty_std': 0.15448428690433502, 'reward_repeat_floor_mean': 0.9817454218864441, 'reward_repeat_floor_std': 0.013800125569105148, 'reward_near_duplicate_penalty_mean': 0.9470402598381042, 'reward_near_duplicate_penalty_std': 0.023172305896878242, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8854376077651978, 'reward_distinct_2_std': 0.12076510488986969, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.2855883240699768, 'reward_total_composite_std': 0.12879391014575958, 'epoch': 0.08}

 46%|████▋     | 1533/3300 [5:51:31<4:43:08,  9.61s/it]INFO 04-12 21:49:57 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1534/3300 [5:51:42<4:59:03, 10.16s/it]2026-04-12 21:50:08,872 | INFO | train_grpo_train | metrics_logged mode=train step=1534

                                                       
{'loss': -0.0644, 'grad_norm': 5.147063255310059, 'learning_rate': 5.3545454545454546e-06, 'num_tokens': 2708651.0, 'completions/mean_length': 138.125, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 84.71428680419922, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.45128393173217773, 'rewards/meter/std': 0.32359185814857483, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9285992383956909, 'rewards/lexical_plausibility/std': 0.2019517868757248, 'rewards/judge_quality/mean': 0.44875001907348633, 'rewards/judge_quality/std': 0.2105392962694168, 'rewards/repeat_penalty/mean': 0.8362984657287598, 'rewards/repeat_penalty/std': 0.21853581070899963, 'rewards/repeat_floor/mean': 0.9808658361434937, 'rewards/repeat_floor/std': 0.027302125468850136, 'rewards/near_duplicate_penalty/mean': 0.9547817707061768, 'rewards/near_duplicate_penalty/std': 0.06124158203601837, 'rewards/opening_diversity/mean': 0.9583333730697632, 'rewards/opening_diversity/std': 0.117851123213768, 'rewards/distinct_2/mean': 0.8905392289161682, 'rewards/distinct_2/std': 0.1314607858657837, 'rewards/total_composite/mean': 0.212234765291214, 'rewards/total_composite/std': 0.15200981497764587, 'reward': 0.212234765291214, 'reward_std': 0.15200981497764587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13178750872612, 'sampling/sampling_logp_difference/max': 2.0805277824401855, 'sampling/importance_sampling_ratio/min': 0.12486429512500763, 'sampling/importance_sampling_ratio/mean': 1.015284776687622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6982308626174927, 'clip_ratio/low_mean': 0.03898688219487667, 'clip_ratio/low_min': 0.03898688219487667, 'clip_ratio/high_mean': 0.06643756851553917, 'clip_ratio/high_max': 0.06643756851553917, 'clip_ratio/region_mean': 0.10542445071041584, 'reward_total_mean': 0.212234765291214, 'reward_meter_mean': 0.45128393173217773, 'reward_meter_std': 0.32359185814857483, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9285992383956909, 'reward_lexical_plausibility_std': 0.2019517868757248, 'reward_repeat_penalty_mean': 0.8362984657287598, 'reward_repeat_penalty_std': 0.21853581070899963, 'reward_repeat_floor_mean': 0.9808658361434937, 'reward_repeat_floor_std': 0.027302125468850136, 'reward_near_duplicate_penalty_mean': 0.9547817707061768, 'reward_near_duplicate_penalty_std': 0.06124158203601837, 'reward_opening_diversity_mean': 0.9583333730697632, 'reward_opening_diversity_std': 0.117851123213768, 'reward_distinct_2_mean': 0.8905392289161682, 'reward_distinct_2_std': 0.1314607858657837, 'reward_judge_quality_mean': 0.44875001907348633, 'reward_judge_quality_std': 0.2105392962694168, 'reward_total_composite_mean': 0.212234765291214, 'reward_total_composite_std': 0.15200981497764587, 'epoch': 0.08}

 46%|████▋     | 1534/3300 [5:51:42<4:59:03, 10.16s/it]INFO 04-12 21:50:09 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1535/3300 [5:51:54<5:16:25, 10.76s/it]2026-04-12 21:50:21,024 | INFO | train_grpo_train | metrics_logged mode=train step=1535

                                                       
{'loss': -0.0511, 'grad_norm': 1.641182780265808, 'learning_rate': 5.351515151515152e-06, 'num_tokens': 2710091.0, 'completions/mean_length': 278.0, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.45078086853027344, 'rewards/meter/std': 0.37090060114860535, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.5, 'rewards/arabic_clean/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.774194061756134, 'rewards/lexical_plausibility/std': 0.2528378665447235, 'rewards/judge_quality/mean': 0.2212499976158142, 'rewards/judge_quality/std': 0.19357076287269592, 'rewards/repeat_penalty/mean': 0.9106108546257019, 'rewards/repeat_penalty/std': 0.18619275093078613, 'rewards/repeat_floor/mean': 0.9908612966537476, 'rewards/repeat_floor/std': 0.019462091848254204, 'rewards/near_duplicate_penalty/mean': 0.9902702569961548, 'rewards/near_duplicate_penalty/std': 0.022085433825850487, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9230077862739563, 'rewards/distinct_2/std': 0.15756137669086456, 'rewards/total_composite/mean': 0.118416428565979, 'rewards/total_composite/std': 0.14938929677009583, 'reward': 0.118416428565979, 'reward_std': 0.14938926696777344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14951294660568237, 'sampling/sampling_logp_difference/max': 1.6516304016113281, 'sampling/importance_sampling_ratio/min': 0.19173705577850342, 'sampling/importance_sampling_ratio/mean': 1.0335068702697754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6646767035126686, 'clip_ratio/low_mean': 0.011111111380159855, 'clip_ratio/low_min': 0.011111111380159855, 'clip_ratio/high_mean': 0.05758567899465561, 'clip_ratio/high_max': 0.05758567899465561, 'clip_ratio/region_mean': 0.06869679037481546, 'reward_total_mean': 0.118416428565979, 'reward_meter_mean': 0.45078086853027344, 'reward_meter_std': 0.37090060114860535, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.5, 'reward_arabic_clean_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.774194061756134, 'reward_lexical_plausibility_std': 0.2528378665447235, 'reward_repeat_penalty_mean': 0.9106108546257019, 'reward_repeat_penalty_std': 0.18619275093078613, 'reward_repeat_floor_mean': 0.9908612966537476, 'reward_repeat_floor_std': 0.019462091848254204, 'reward_near_duplicate_penalty_mean': 0.9902702569961548, 'reward_near_duplicate_penalty_std': 0.022085433825850487, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9230077862739563, 'reward_distinct_2_std': 0.15756137669086456, 'reward_judge_quality_mean': 0.2212499976158142, 'reward_judge_quality_std': 0.19357076287269592, 'reward_total_composite_mean': 0.118416428565979, 'reward_total_composite_std': 0.14938929677009583, 'epoch': 0.08}

 47%|████▋     | 1535/3300 [5:51:54<5:16:25, 10.76s/it]INFO 04-12 21:50:21 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1536/3300 [5:52:01<4:42:46,  9.62s/it]2026-04-12 21:50:27,980 | INFO | train_grpo_train | metrics_logged mode=train step=1536

                                                       
{'loss': 0.0556, 'grad_norm': 17.375411987304688, 'learning_rate': 5.348484848484848e-06, 'num_tokens': 2711644.0, 'completions/mean_length': 38.125, 'completions/min_length': 35.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.20822304487228394, 'rewards/meter/std': 0.19521218538284302, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9866071343421936, 'rewards/lexical_plausibility/std': 0.03788072615861893, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/repeat_penalty/mean': 0.9726599454879761, 'rewards/repeat_penalty/std': 0.02959694154560566, 'rewards/repeat_floor/mean': 0.9969107508659363, 'rewards/repeat_floor/std': 0.0026783111970871687, 'rewards/near_duplicate_penalty/mean': 0.9896665811538696, 'rewards/near_duplicate_penalty/std': 0.008280138485133648, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9828975200653076, 'rewards/distinct_2/std': 0.031678035855293274, 'rewards/total_composite/mean': 0.10866338759660721, 'rewards/total_composite/std': 0.11096411943435669, 'reward': 0.10866338759660721, 'reward_std': 0.11096411943435669, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20482344925403595, 'sampling/sampling_logp_difference/max': 3.809886932373047, 'sampling/importance_sampling_ratio/min': 0.022150682285428047, 'sampling/importance_sampling_ratio/mean': 1.000345230102539, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0398316383361816, 'clip_ratio/low_mean': 0.044513950822874904, 'clip_ratio/low_min': 0.044513950822874904, 'clip_ratio/high_mean': 0.05062741320580244, 'clip_ratio/high_max': 0.05062741320580244, 'clip_ratio/region_mean': 0.09514136402867734, 'reward_total_mean': 0.10866338759660721, 'reward_meter_mean': 0.20822304487228394, 'reward_meter_std': 0.19521218538284302, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9866071343421936, 'reward_lexical_plausibility_std': 0.03788072615861893, 'reward_repeat_penalty_mean': 0.9726599454879761, 'reward_repeat_penalty_std': 0.02959694154560566, 'reward_repeat_floor_mean': 0.9969107508659363, 'reward_repeat_floor_std': 0.0026783111970871687, 'reward_near_duplicate_penalty_mean': 0.9896665811538696, 'reward_near_duplicate_penalty_std': 0.008280138485133648, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9828975200653076, 'reward_distinct_2_std': 0.031678035855293274, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.10866338759660721, 'reward_total_composite_std': 0.11096411943435669, 'epoch': 0.08}

 47%|████▋     | 1536/3300 [5:52:01<4:42:46,  9.62s/it]INFO 04-12 21:50:28 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1537/3300 [5:52:14<5:06:37, 10.44s/it]2026-04-12 21:50:40,319 | INFO | train_grpo_train | metrics_logged mode=train step=1537

                                                       
{'loss': -0.0954, 'grad_norm': 4.557192802429199, 'learning_rate': 5.3454545454545455e-06, 'num_tokens': 2713585.0, 'completions/mean_length': 135.625, 'completions/min_length': 70.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.85714721679688, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.5788543224334717, 'rewards/meter/std': 0.25308576226234436, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9325000047683716, 'rewards/count_floor/std': 0.02121320553123951, 'rewards/lexical_plausibility/mean': 0.9413644075393677, 'rewards/lexical_plausibility/std': 0.1658465713262558, 'rewards/judge_quality/mean': 0.4150000214576721, 'rewards/judge_quality/std': 0.18031719326972961, 'rewards/repeat_penalty/mean': 0.856420636177063, 'rewards/repeat_penalty/std': 0.0703580304980278, 'rewards/repeat_floor/mean': 0.9833903312683105, 'rewards/repeat_floor/std': 0.008245175704360008, 'rewards/near_duplicate_penalty/mean': 0.946567714214325, 'rewards/near_duplicate_penalty/std': 0.03625441715121269, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9045015573501587, 'rewards/distinct_2/std': 0.061601005494594574, 'rewards/total_composite/mean': 0.20196250081062317, 'rewards/total_composite/std': 0.10801875591278076, 'reward': 0.20196250081062317, 'reward_std': 0.10801874846220016, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15305712819099426, 'sampling/sampling_logp_difference/max': 1.261347770690918, 'sampling/importance_sampling_ratio/min': 0.28327199816703796, 'sampling/importance_sampling_ratio/mean': 1.0075916051864624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0446966513991356, 'clip_ratio/low_mean': 0.03248993679881096, 'clip_ratio/low_min': 0.03248993679881096, 'clip_ratio/high_mean': 0.10087534971535206, 'clip_ratio/high_max': 0.10087534971535206, 'clip_ratio/region_mean': 0.13336528651416302, 'reward_total_mean': 0.20196250081062317, 'reward_meter_mean': 0.5788543224334717, 'reward_meter_std': 0.25308576226234436, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9325000047683716, 'reward_count_floor_std': 0.02121320553123951, 'reward_lexical_plausibility_mean': 0.9413644075393677, 'reward_lexical_plausibility_std': 0.1658465713262558, 'reward_repeat_penalty_mean': 0.856420636177063, 'reward_repeat_penalty_std': 0.0703580304980278, 'reward_repeat_floor_mean': 0.9833903312683105, 'reward_repeat_floor_std': 0.008245175704360008, 'reward_near_duplicate_penalty_mean': 0.946567714214325, 'reward_near_duplicate_penalty_std': 0.03625441715121269, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9045015573501587, 'reward_distinct_2_std': 0.061601005494594574, 'reward_judge_quality_mean': 0.4150000214576721, 'reward_judge_quality_std': 0.18031719326972961, 'reward_total_composite_mean': 0.20196250081062317, 'reward_total_composite_std': 0.10801875591278076, 'epoch': 0.08}

 47%|████▋     | 1537/3300 [5:52:14<5:06:37, 10.44s/it]INFO 04-12 21:50:40 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1538/3300 [5:52:26<5:21:51, 10.96s/it]2026-04-12 21:50:52,507 | INFO | train_grpo_train | metrics_logged mode=train step=1538

                                                       
{'loss': -0.1273, 'grad_norm': 2.0342001914978027, 'learning_rate': 5.342424242424244e-06, 'num_tokens': 2715336.0, 'completions/mean_length': 233.875, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.6564201712608337, 'rewards/meter/std': 0.28901103138923645, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/arabic_clean/mean': 0.625, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9750000238418579, 'rewards/count_floor/std': 0.04629101976752281, 'rewards/lexical_plausibility/mean': 0.8234190940856934, 'rewards/lexical_plausibility/std': 0.24398906528949738, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.31057265400886536, 'rewards/repeat_penalty/mean': 0.9463145732879639, 'rewards/repeat_penalty/std': 0.03588086739182472, 'rewards/repeat_floor/mean': 0.9932747483253479, 'rewards/repeat_floor/std': 0.004338892176747322, 'rewards/near_duplicate_penalty/mean': 0.969734787940979, 'rewards/near_duplicate_penalty/std': 0.02084873802959919, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9757170081138611, 'rewards/distinct_2/std': 0.02509191259741783, 'rewards/total_composite/mean': 0.29934364557266235, 'rewards/total_composite/std': 0.240284726023674, 'reward': 0.29934364557266235, 'reward_std': 0.24028471112251282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15088796615600586, 'sampling/sampling_logp_difference/max': 1.6127660274505615, 'sampling/importance_sampling_ratio/min': 0.1993354707956314, 'sampling/importance_sampling_ratio/mean': 1.037274956703186, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.719944953918457, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07552995346486568, 'clip_ratio/high_max': 0.07552995346486568, 'clip_ratio/region_mean': 0.07552995346486568, 'reward_total_mean': 0.29934364557266235, 'reward_meter_mean': 0.6564201712608337, 'reward_meter_std': 0.28901103138923645, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_arabic_clean_mean': 0.625, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9750000238418579, 'reward_count_floor_std': 0.04629101976752281, 'reward_lexical_plausibility_mean': 0.8234190940856934, 'reward_lexical_plausibility_std': 0.24398906528949738, 'reward_repeat_penalty_mean': 0.9463145732879639, 'reward_repeat_penalty_std': 0.03588086739182472, 'reward_repeat_floor_mean': 0.9932747483253479, 'reward_repeat_floor_std': 0.004338892176747322, 'reward_near_duplicate_penalty_mean': 0.969734787940979, 'reward_near_duplicate_penalty_std': 0.02084873802959919, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9757170081138611, 'reward_distinct_2_std': 0.02509191259741783, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.31057265400886536, 'reward_total_composite_mean': 0.29934364557266235, 'reward_total_composite_std': 0.240284726023674, 'epoch': 0.08}

 47%|████▋     | 1538/3300 [5:52:26<5:21:51, 10.96s/it]INFO 04-12 21:50:52 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1539/3300 [5:52:33<4:47:51,  9.81s/it]2026-04-12 21:50:59,634 | INFO | train_grpo_train | metrics_logged mode=train step=1539

                                                       
{'loss': 0.0604, 'grad_norm': 10.385834693908691, 'learning_rate': 5.33939393939394e-06, 'num_tokens': 2717305.0, 'completions/mean_length': 65.125, 'completions/min_length': 58.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8148790597915649, 'rewards/meter/std': 0.23742079734802246, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/repeat_penalty/mean': 0.6886733770370483, 'rewards/repeat_penalty/std': 0.15798814594745636, 'rewards/repeat_floor/mean': 0.9628767371177673, 'rewards/repeat_floor/std': 0.01890617609024048, 'rewards/near_duplicate_penalty/mean': 0.8828556537628174, 'rewards/near_duplicate_penalty/std': 0.052996646612882614, 'rewards/opening_diversity/mean': 0.956250011920929, 'rewards/opening_diversity/std': 0.09038607776165009, 'rewards/distinct_2/mean': 0.8119268417358398, 'rewards/distinct_2/std': 0.13330933451652527, 'rewards/total_composite/mean': 0.27073222398757935, 'rewards/total_composite/std': 0.10145965963602066, 'reward': 0.27073222398757935, 'reward_std': 0.10145965963602066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13211476802825928, 'sampling/sampling_logp_difference/max': 1.193953275680542, 'sampling/importance_sampling_ratio/min': 0.3030209541320801, 'sampling/importance_sampling_ratio/mean': 1.01863694190979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9416583776473999, 'clip_ratio/low_mean': 0.07649038266390562, 'clip_ratio/low_min': 0.07649038266390562, 'clip_ratio/high_mean': 0.07816583476960659, 'clip_ratio/high_max': 0.07816583476960659, 'clip_ratio/region_mean': 0.1546562174335122, 'reward_total_mean': 0.27073222398757935, 'reward_meter_mean': 0.8148790597915649, 'reward_meter_std': 0.23742079734802246, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.6886733770370483, 'reward_repeat_penalty_std': 0.15798814594745636, 'reward_repeat_floor_mean': 0.9628767371177673, 'reward_repeat_floor_std': 0.01890617609024048, 'reward_near_duplicate_penalty_mean': 0.8828556537628174, 'reward_near_duplicate_penalty_std': 0.052996646612882614, 'reward_opening_diversity_mean': 0.956250011920929, 'reward_opening_diversity_std': 0.09038607776165009, 'reward_distinct_2_mean': 0.8119268417358398, 'reward_distinct_2_std': 0.13330933451652527, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.27073222398757935, 'reward_total_composite_std': 0.10145965963602066, 'epoch': 0.08}

 47%|████▋     | 1539/3300 [5:52:33<4:47:51,  9.81s/it]INFO 04-12 21:50:59 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1540/3300 [5:52:45<5:10:27, 10.58s/it]2026-04-12 21:51:12,017 | INFO | train_grpo_train | metrics_logged mode=train step=1540

                                                       
{'loss': -0.0652, 'grad_norm': 3.296373128890991, 'learning_rate': 5.336363636363637e-06, 'num_tokens': 2718945.0, 'completions/mean_length': 100.0, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.142860412597656, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8488795757293701, 'rewards/meter/std': 0.24881407618522644, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9446155428886414, 'rewards/lexical_plausibility/std': 0.15665094554424286, 'rewards/judge_quality/mean': 0.6237499713897705, 'rewards/judge_quality/std': 0.33907175064086914, 'rewards/repeat_penalty/mean': 0.6673555374145508, 'rewards/repeat_penalty/std': 0.18606480956077576, 'rewards/repeat_floor/mean': 0.953986644744873, 'rewards/repeat_floor/std': 0.026274489238858223, 'rewards/near_duplicate_penalty/mean': 0.8011376857757568, 'rewards/near_duplicate_penalty/std': 0.09802062064409256, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8201777338981628, 'rewards/distinct_2/std': 0.1391981989145279, 'rewards/total_composite/mean': 0.5084881782531738, 'rewards/total_composite/std': 0.3138238191604614, 'reward': 0.5084881782531738, 'reward_std': 0.31382378935813904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1106218621134758, 'sampling/sampling_logp_difference/max': 1.3291430473327637, 'sampling/importance_sampling_ratio/min': 0.26470398902893066, 'sampling/importance_sampling_ratio/mean': 1.0372222661972046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6519262492656708, 'clip_ratio/low_mean': 0.03848958807066083, 'clip_ratio/low_min': 0.03848958807066083, 'clip_ratio/high_mean': 0.03529640194028616, 'clip_ratio/high_max': 0.03529640194028616, 'clip_ratio/region_mean': 0.07378599001094699, 'reward_total_mean': 0.5084881782531738, 'reward_meter_mean': 0.8488795757293701, 'reward_meter_std': 0.24881407618522644, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9446155428886414, 'reward_lexical_plausibility_std': 0.15665094554424286, 'reward_repeat_penalty_mean': 0.6673555374145508, 'reward_repeat_penalty_std': 0.18606480956077576, 'reward_repeat_floor_mean': 0.953986644744873, 'reward_repeat_floor_std': 0.026274489238858223, 'reward_near_duplicate_penalty_mean': 0.8011376857757568, 'reward_near_duplicate_penalty_std': 0.09802062064409256, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8201777338981628, 'reward_distinct_2_std': 0.1391981989145279, 'reward_judge_quality_mean': 0.6237499713897705, 'reward_judge_quality_std': 0.33907175064086914, 'reward_total_composite_mean': 0.5084881782531738, 'reward_total_composite_std': 0.3138238191604614, 'epoch': 0.08}

 47%|████▋     | 1540/3300 [5:52:45<5:10:27, 10.58s/it]INFO 04-12 21:51:12 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1541/3300 [5:52:56<5:08:22, 10.52s/it]2026-04-12 21:51:22,392 | INFO | train_grpo_train | metrics_logged mode=train step=1541

                                                       
{'loss': 0.016, 'grad_norm': 11.169768333435059, 'learning_rate': 5.333333333333334e-06, 'num_tokens': 2720765.0, 'completions/mean_length': 58.5, 'completions/min_length': 50.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.5, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.29664599895477295, 'rewards/meter/std': 0.31480297446250916, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.987500011920929, 'rewards/count_floor/std': 0.0353553481400013, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906257808208466, 'rewards/repeat_penalty/mean': 0.7345658540725708, 'rewards/repeat_penalty/std': 0.15851762890815735, 'rewards/repeat_floor/mean': 0.968478262424469, 'rewards/repeat_floor/std': 0.018966805189847946, 'rewards/near_duplicate_penalty/mean': 0.9067052602767944, 'rewards/near_duplicate_penalty/std': 0.043794550001621246, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8052495718002319, 'rewards/distinct_2/std': 0.14887332916259766, 'rewards/total_composite/mean': 0.12190087139606476, 'rewards/total_composite/std': 0.12960219383239746, 'reward': 0.12190087139606476, 'reward_std': 0.12960217893123627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14092068374156952, 'sampling/sampling_logp_difference/max': 1.8606610298156738, 'sampling/importance_sampling_ratio/min': 0.1830008178949356, 'sampling/importance_sampling_ratio/mean': 1.0127097368240356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7807648330926895, 'clip_ratio/low_mean': 0.07900669239461422, 'clip_ratio/low_min': 0.07900669239461422, 'clip_ratio/high_mean': 0.03483376791700721, 'clip_ratio/high_max': 0.03483376791700721, 'clip_ratio/region_mean': 0.11384046031162143, 'reward_total_mean': 0.12190087139606476, 'reward_meter_mean': 0.29664599895477295, 'reward_meter_std': 0.31480297446250916, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.987500011920929, 'reward_count_floor_std': 0.0353553481400013, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7345658540725708, 'reward_repeat_penalty_std': 0.15851762890815735, 'reward_repeat_floor_mean': 0.968478262424469, 'reward_repeat_floor_std': 0.018966805189847946, 'reward_near_duplicate_penalty_mean': 0.9067052602767944, 'reward_near_duplicate_penalty_std': 0.043794550001621246, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8052495718002319, 'reward_distinct_2_std': 0.14887332916259766, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906257808208466, 'reward_total_composite_mean': 0.12190087139606476, 'reward_total_composite_std': 0.12960219383239746, 'epoch': 0.08}

 47%|████▋     | 1541/3300 [5:52:56<5:08:22, 10.52s/it]INFO 04-12 21:51:22 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1542/3300 [5:53:07<5:19:18, 10.90s/it]2026-04-12 21:51:34,169 | INFO | train_grpo_train | metrics_logged mode=train step=1542

                                                       
{'loss': -0.0495, 'grad_norm': 1.1691104173660278, 'learning_rate': 5.330303030303031e-06, 'num_tokens': 2722148.0, 'completions/mean_length': 333.875, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.36853447556495667, 'rewards/meter/std': 0.5003435611724854, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/arabic_clean/mean': 0.375, 'rewards/arabic_clean/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 0.981249988079071, 'rewards/arabic_floor/std': 0.053033001720905304, 'rewards/count_floor/mean': 0.9624999761581421, 'rewards/count_floor/std': 0.06943649053573608, 'rewards/lexical_plausibility/mean': 0.6625404357910156, 'rewards/lexical_plausibility/std': 0.28458768129348755, 'rewards/judge_quality/mean': 0.1887499988079071, 'rewards/judge_quality/std': 0.1914931833744049, 'rewards/repeat_penalty/mean': 0.8250202536582947, 'rewards/repeat_penalty/std': 0.24050584435462952, 'rewards/repeat_floor/mean': 0.9841594099998474, 'rewards/repeat_floor/std': 0.0224430151283741, 'rewards/near_duplicate_penalty/mean': 0.9556620121002197, 'rewards/near_duplicate_penalty/std': 0.06206832081079483, 'rewards/opening_diversity/mean': 0.90625, 'rewards/opening_diversity/std': 0.12938730418682098, 'rewards/distinct_2/mean': 0.9603902101516724, 'rewards/distinct_2/std': 0.10834155976772308, 'rewards/total_composite/mean': 0.14881381392478943, 'rewards/total_composite/std': 0.2050238698720932, 'reward': 0.14881381392478943, 'reward_std': 0.2050238698720932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13622038066387177, 'sampling/sampling_logp_difference/max': 1.5734450817108154, 'sampling/importance_sampling_ratio/min': 0.2073296755552292, 'sampling/importance_sampling_ratio/mean': 0.995415449142456, 'sampling/importance_sampling_ratio/max': 1.5963773727416992, 'entropy': 0.3542150482535362, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.05432233866304159, 'clip_ratio/high_max': 0.05432233866304159, 'clip_ratio/region_mean': 0.05432233866304159, 'reward_total_mean': 0.14881381392478943, 'reward_meter_mean': 0.36853447556495667, 'reward_meter_std': 0.5003435611724854, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_arabic_clean_mean': 0.375, 'reward_arabic_clean_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 0.981249988079071, 'reward_arabic_floor_std': 0.053033001720905304, 'reward_count_floor_mean': 0.9624999761581421, 'reward_count_floor_std': 0.06943649053573608, 'reward_lexical_plausibility_mean': 0.6625404357910156, 'reward_lexical_plausibility_std': 0.28458768129348755, 'reward_repeat_penalty_mean': 0.8250202536582947, 'reward_repeat_penalty_std': 0.24050584435462952, 'reward_repeat_floor_mean': 0.9841594099998474, 'reward_repeat_floor_std': 0.0224430151283741, 'reward_near_duplicate_penalty_mean': 0.9556620121002197, 'reward_near_duplicate_penalty_std': 0.06206832081079483, 'reward_opening_diversity_mean': 0.90625, 'reward_opening_diversity_std': 0.12938730418682098, 'reward_distinct_2_mean': 0.9603902101516724, 'reward_distinct_2_std': 0.10834155976772308, 'reward_judge_quality_mean': 0.1887499988079071, 'reward_judge_quality_std': 0.1914931833744049, 'reward_total_composite_mean': 0.14881381392478943, 'reward_total_composite_std': 0.2050238698720932, 'epoch': 0.08}

 47%|████▋     | 1542/3300 [5:53:08<5:19:18, 10.90s/it]INFO 04-12 21:51:34 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1543/3300 [5:53:14<4:39:01,  9.53s/it]2026-04-12 21:51:40,507 | INFO | train_grpo_train | metrics_logged mode=train step=1543

                                                       
{'loss': 0.0231, 'grad_norm': 19.730863571166992, 'learning_rate': 5.327272727272727e-06, 'num_tokens': 2723802.0, 'completions/mean_length': 27.75, 'completions/min_length': 26.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5575295090675354, 'rewards/meter/std': 0.3906041383743286, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/repeat_penalty/mean': 0.9326024055480957, 'rewards/repeat_penalty/std': 0.08621253073215485, 'rewards/repeat_floor/mean': 0.990406334400177, 'rewards/repeat_floor/std': 0.011759807355701923, 'rewards/near_duplicate_penalty/mean': 0.9442841410636902, 'rewards/near_duplicate_penalty/std': 0.06192629039287567, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9862637519836426, 'rewards/distinct_2/std': 0.0388520210981369, 'rewards/total_composite/mean': 0.23389223217964172, 'rewards/total_composite/std': 0.16279156506061554, 'reward': 0.23389223217964172, 'reward_std': 0.16279156506061554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13585558533668518, 'sampling/sampling_logp_difference/max': 1.5361089706420898, 'sampling/importance_sampling_ratio/min': 0.21521688997745514, 'sampling/importance_sampling_ratio/mean': 1.0094985961914062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7222808226943016, 'clip_ratio/low_mean': 0.0888594202697277, 'clip_ratio/low_min': 0.0888594202697277, 'clip_ratio/high_mean': 0.06620598584413528, 'clip_ratio/high_max': 0.06620598584413528, 'clip_ratio/region_mean': 0.155065406113863, 'reward_total_mean': 0.23389223217964172, 'reward_meter_mean': 0.5575295090675354, 'reward_meter_std': 0.3906041383743286, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9326024055480957, 'reward_repeat_penalty_std': 0.08621253073215485, 'reward_repeat_floor_mean': 0.990406334400177, 'reward_repeat_floor_std': 0.011759807355701923, 'reward_near_duplicate_penalty_mean': 0.9442841410636902, 'reward_near_duplicate_penalty_std': 0.06192629039287567, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9862637519836426, 'reward_distinct_2_std': 0.0388520210981369, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.23389223217964172, 'reward_total_composite_std': 0.16279156506061554, 'epoch': 0.08}

 47%|████▋     | 1543/3300 [5:53:14<4:39:01,  9.53s/it]INFO 04-12 21:51:40 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1544/3300 [5:53:27<5:08:14, 10.53s/it]2026-04-12 21:51:53,381 | INFO | train_grpo_train | metrics_logged mode=train step=1544

                                                       
{'loss': -0.0805, 'grad_norm': 5.6606221199035645, 'learning_rate': 5.324242424242425e-06, 'num_tokens': 2726132.0, 'completions/mean_length': 153.25, 'completions/min_length': 90.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 102.00000762939453, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.5970669984817505, 'rewards/meter/std': 0.40424844622612, 'rewards/count_adherence/mean': 0.7291666269302368, 'rewards/count_adherence/std': 0.1767766773700714, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.918749988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9492233991622925, 'rewards/lexical_plausibility/std': 0.14361795783042908, 'rewards/judge_quality/mean': 0.44874998927116394, 'rewards/judge_quality/std': 0.2105392962694168, 'rewards/repeat_penalty/mean': 0.8727701902389526, 'rewards/repeat_penalty/std': 0.10709965974092484, 'rewards/repeat_floor/mean': 0.9863566160202026, 'rewards/repeat_floor/std': 0.010722039267420769, 'rewards/near_duplicate_penalty/mean': 0.9649642705917358, 'rewards/near_duplicate_penalty/std': 0.021786026656627655, 'rewards/opening_diversity/mean': 0.9791666269302368, 'rewards/opening_diversity/std': 0.0589255727827549, 'rewards/distinct_2/mean': 0.9206880927085876, 'rewards/distinct_2/std': 0.060860540717840195, 'rewards/total_composite/mean': 0.20749783515930176, 'rewards/total_composite/std': 0.16850578784942627, 'reward': 0.20749783515930176, 'reward_std': 0.16850577294826508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12931886315345764, 'sampling/sampling_logp_difference/max': 2.968108654022217, 'sampling/importance_sampling_ratio/min': 0.05140043795108795, 'sampling/importance_sampling_ratio/mean': 1.0104999542236328, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6217343434691429, 'clip_ratio/low_mean': 0.05353942792862654, 'clip_ratio/low_min': 0.05353942792862654, 'clip_ratio/high_mean': 0.06009184010326862, 'clip_ratio/high_max': 0.06009184010326862, 'clip_ratio/region_mean': 0.11363126803189516, 'reward_total_mean': 0.20749783515930176, 'reward_meter_mean': 0.5970669984817505, 'reward_meter_std': 0.40424844622612, 'reward_count_adherence_mean': 0.7291666269302368, 'reward_count_adherence_std': 0.1767766773700714, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.918749988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9492233991622925, 'reward_lexical_plausibility_std': 0.14361795783042908, 'reward_repeat_penalty_mean': 0.8727701902389526, 'reward_repeat_penalty_std': 0.10709965974092484, 'reward_repeat_floor_mean': 0.9863566160202026, 'reward_repeat_floor_std': 0.010722039267420769, 'reward_near_duplicate_penalty_mean': 0.9649642705917358, 'reward_near_duplicate_penalty_std': 0.021786026656627655, 'reward_opening_diversity_mean': 0.9791666269302368, 'reward_opening_diversity_std': 0.0589255727827549, 'reward_distinct_2_mean': 0.9206880927085876, 'reward_distinct_2_std': 0.060860540717840195, 'reward_judge_quality_mean': 0.44874998927116394, 'reward_judge_quality_std': 0.2105392962694168, 'reward_total_composite_mean': 0.20749783515930176, 'reward_total_composite_std': 0.16850578784942627, 'epoch': 0.08}

 47%|████▋     | 1544/3300 [5:53:27<5:08:14, 10.53s/it]INFO 04-12 21:51:53 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1545/3300 [5:53:38<5:12:37, 10.69s/it]2026-04-12 21:52:04,437 | INFO | train_grpo_train | metrics_logged mode=train step=1545

                                                       
{'loss': 0.0558, 'grad_norm': 13.716522216796875, 'learning_rate': 5.321212121212122e-06, 'num_tokens': 2727969.0, 'completions/mean_length': 65.625, 'completions/min_length': 54.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.39185115694999695, 'rewards/meter/std': 0.25724393129348755, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9906250238418579, 'rewards/count_floor/std': 0.026516500860452652, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9120625853538513, 'rewards/repeat_penalty/std': 0.07507210969924927, 'rewards/repeat_floor/mean': 0.9899458885192871, 'rewards/repeat_floor/std': 0.008587034419178963, 'rewards/near_duplicate_penalty/mean': 0.968421459197998, 'rewards/near_duplicate_penalty/std': 0.031119329854846, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9409183859825134, 'rewards/distinct_2/std': 0.056111954152584076, 'rewards/total_composite/mean': 0.1604272872209549, 'rewards/total_composite/std': 0.10421840846538544, 'reward': 0.1604272872209549, 'reward_std': 0.10421840846538544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14631056785583496, 'sampling/sampling_logp_difference/max': 3.738355875015259, 'sampling/importance_sampling_ratio/min': 0.023793190717697144, 'sampling/importance_sampling_ratio/mean': 0.9791350364685059, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6832848936319351, 'clip_ratio/low_mean': 0.04576245602220297, 'clip_ratio/low_min': 0.04576245602220297, 'clip_ratio/high_mean': 0.08005691785365343, 'clip_ratio/high_max': 0.08005691785365343, 'clip_ratio/region_mean': 0.1258193738758564, 'reward_total_mean': 0.1604272872209549, 'reward_meter_mean': 0.39185115694999695, 'reward_meter_std': 0.25724393129348755, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9906250238418579, 'reward_count_floor_std': 0.026516500860452652, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9120625853538513, 'reward_repeat_penalty_std': 0.07507210969924927, 'reward_repeat_floor_mean': 0.9899458885192871, 'reward_repeat_floor_std': 0.008587034419178963, 'reward_near_duplicate_penalty_mean': 0.968421459197998, 'reward_near_duplicate_penalty_std': 0.031119329854846, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9409183859825134, 'reward_distinct_2_std': 0.056111954152584076, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.1604272872209549, 'reward_total_composite_std': 0.10421840846538544, 'epoch': 0.08}

 47%|████▋     | 1545/3300 [5:53:38<5:12:37, 10.69s/it]INFO 04-12 21:52:04 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1546/3300 [5:53:49<5:22:21, 11.03s/it]2026-04-12 21:52:16,251 | INFO | train_grpo_train | metrics_logged mode=train step=1546

                                                       
{'loss': -0.1122, 'grad_norm': 4.126603603363037, 'learning_rate': 5.318181818181819e-06, 'num_tokens': 2730157.0, 'completions/mean_length': 145.5, 'completions/min_length': 89.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 93.14286041259766, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.8042241334915161, 'rewards/meter/std': 0.30065563321113586, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 0.9537556767463684, 'rewards/lexical_plausibility/std': 0.13079874217510223, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/repeat_penalty/mean': 0.7879989147186279, 'rewards/repeat_penalty/std': 0.12193939834833145, 'rewards/repeat_floor/mean': 0.9760335087776184, 'rewards/repeat_floor/std': 0.014990242198109627, 'rewards/near_duplicate_penalty/mean': 0.9388830661773682, 'rewards/near_duplicate_penalty/std': 0.04608868807554245, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8355673551559448, 'rewards/distinct_2/std': 0.09226034581661224, 'rewards/total_composite/mean': 0.2881458103656769, 'rewards/total_composite/std': 0.15814697742462158, 'reward': 0.2881458103656769, 'reward_std': 0.15814697742462158, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.139773428440094, 'sampling/sampling_logp_difference/max': 2.6370255947113037, 'sampling/importance_sampling_ratio/min': 0.07157384604215622, 'sampling/importance_sampling_ratio/mean': 1.0097938776016235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6622270792722702, 'clip_ratio/low_mean': 0.03209064435213804, 'clip_ratio/low_min': 0.03209064435213804, 'clip_ratio/high_mean': 0.08805008511990309, 'clip_ratio/high_max': 0.08805008511990309, 'clip_ratio/region_mean': 0.12014072947204113, 'reward_total_mean': 0.2881458103656769, 'reward_meter_mean': 0.8042241334915161, 'reward_meter_std': 0.30065563321113586, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 0.9537556767463684, 'reward_lexical_plausibility_std': 0.13079874217510223, 'reward_repeat_penalty_mean': 0.7879989147186279, 'reward_repeat_penalty_std': 0.12193939834833145, 'reward_repeat_floor_mean': 0.9760335087776184, 'reward_repeat_floor_std': 0.014990242198109627, 'reward_near_duplicate_penalty_mean': 0.9388830661773682, 'reward_near_duplicate_penalty_std': 0.04608868807554245, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8355673551559448, 'reward_distinct_2_std': 0.09226034581661224, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.2881458103656769, 'reward_total_composite_std': 0.15814697742462158, 'epoch': 0.08}

 47%|████▋     | 1546/3300 [5:53:50<5:22:21, 11.03s/it]INFO 04-12 21:52:16 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1547/3300 [5:53:56<4:41:21,  9.63s/it]2026-04-12 21:52:22,622 | INFO | train_grpo_train | metrics_logged mode=train step=1547

                                                       
{'loss': 0.0312, 'grad_norm': 16.875259399414062, 'learning_rate': 5.3151515151515155e-06, 'num_tokens': 2731716.0, 'completions/mean_length': 42.875, 'completions/min_length': 37.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.90336012840271, 'rewards/meter/std': 0.2108646035194397, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/repeat_penalty/mean': 0.8804962038993835, 'rewards/repeat_penalty/std': 0.12388050556182861, 'rewards/repeat_floor/mean': 0.986414909362793, 'rewards/repeat_floor/std': 0.01430308073759079, 'rewards/near_duplicate_penalty/mean': 0.9592976570129395, 'rewards/near_duplicate_penalty/std': 0.05108080804347992, 'rewards/opening_diversity/mean': 0.987500011920929, 'rewards/opening_diversity/std': 0.0353553481400013, 'rewards/distinct_2/mean': 0.9252249002456665, 'rewards/distinct_2/std': 0.07454216480255127, 'rewards/total_composite/mean': 0.37330663204193115, 'rewards/total_composite/std': 0.08980835974216461, 'reward': 0.37330663204193115, 'reward_std': 0.08980835974216461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15155208110809326, 'sampling/sampling_logp_difference/max': 3.999438762664795, 'sampling/importance_sampling_ratio/min': 0.018325921148061752, 'sampling/importance_sampling_ratio/mean': 1.0104871988296509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7958294898271561, 'clip_ratio/low_mean': 0.02589285746216774, 'clip_ratio/low_min': 0.02589285746216774, 'clip_ratio/high_mean': 0.10588439274579287, 'clip_ratio/high_max': 0.10588439274579287, 'clip_ratio/region_mean': 0.1317772502079606, 'reward_total_mean': 0.37330663204193115, 'reward_meter_mean': 0.90336012840271, 'reward_meter_std': 0.2108646035194397, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8804962038993835, 'reward_repeat_penalty_std': 0.12388050556182861, 'reward_repeat_floor_mean': 0.986414909362793, 'reward_repeat_floor_std': 0.01430308073759079, 'reward_near_duplicate_penalty_mean': 0.9592976570129395, 'reward_near_duplicate_penalty_std': 0.05108080804347992, 'reward_opening_diversity_mean': 0.987500011920929, 'reward_opening_diversity_std': 0.0353553481400013, 'reward_distinct_2_mean': 0.9252249002456665, 'reward_distinct_2_std': 0.07454216480255127, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.37330663204193115, 'reward_total_composite_std': 0.08980835974216461, 'epoch': 0.08}

 47%|████▋     | 1547/3300 [5:53:56<4:41:21,  9.63s/it]INFO 04-12 21:52:22 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1548/3300 [5:54:02<4:12:18,  8.64s/it]2026-04-12 21:52:28,954 | INFO | train_grpo_train | metrics_logged mode=train step=1548

                                                       
{'loss': 0.0933, 'grad_norm': 15.812986373901367, 'learning_rate': 5.312121212121213e-06, 'num_tokens': 2733321.0, 'completions/mean_length': 34.625, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7568606734275818, 'rewards/meter/std': 0.20686504244804382, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.981249988079071, 'rewards/count_floor/std': 0.053033001720905304, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/repeat_penalty/mean': 0.9419490098953247, 'rewards/repeat_penalty/std': 0.06766816973686218, 'rewards/repeat_floor/mean': 0.9926952719688416, 'rewards/repeat_floor/std': 0.007710225880146027, 'rewards/near_duplicate_penalty/mean': 0.9677453637123108, 'rewards/near_duplicate_penalty/std': 0.025681620463728905, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9725937843322754, 'rewards/distinct_2/std': 0.0523630827665329, 'rewards/total_composite/mean': 0.3215841054916382, 'rewards/total_composite/std': 0.10116592794656754, 'reward': 0.3215841054916382, 'reward_std': 0.10116592049598694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09823382645845413, 'sampling/sampling_logp_difference/max': 1.6047916412353516, 'sampling/importance_sampling_ratio/min': 0.20093141496181488, 'sampling/importance_sampling_ratio/mean': 1.0040779113769531, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3879966586828232, 'clip_ratio/low_mean': 0.018481182865798473, 'clip_ratio/low_min': 0.018481182865798473, 'clip_ratio/high_mean': 0.05277685262262821, 'clip_ratio/high_max': 0.05277685262262821, 'clip_ratio/region_mean': 0.07125803548842669, 'reward_total_mean': 0.3215841054916382, 'reward_meter_mean': 0.7568606734275818, 'reward_meter_std': 0.20686504244804382, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.981249988079071, 'reward_count_floor_std': 0.053033001720905304, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9419490098953247, 'reward_repeat_penalty_std': 0.06766816973686218, 'reward_repeat_floor_mean': 0.9926952719688416, 'reward_repeat_floor_std': 0.007710225880146027, 'reward_near_duplicate_penalty_mean': 0.9677453637123108, 'reward_near_duplicate_penalty_std': 0.025681620463728905, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9725937843322754, 'reward_distinct_2_std': 0.0523630827665329, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.3215841054916382, 'reward_total_composite_std': 0.10116592794656754, 'epoch': 0.08}

 47%|████▋     | 1548/3300 [5:54:02<4:12:18,  8.64s/it]INFO 04-12 21:52:29 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1549/3300 [5:54:08<3:51:01,  7.92s/it]2026-04-12 21:52:35,178 | INFO | train_grpo_train | metrics_logged mode=train step=1549

                                                       
{'loss': 0.0296, 'grad_norm': 12.052055358886719, 'learning_rate': 5.309090909090909e-06, 'num_tokens': 2734856.0, 'completions/mean_length': 44.875, 'completions/min_length': 41.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.47071734070777893, 'rewards/meter/std': 0.3999786078929901, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/repeat_penalty/mean': 0.7076892852783203, 'rewards/repeat_penalty/std': 0.16064846515655518, 'rewards/repeat_floor/mean': 0.9636876583099365, 'rewards/repeat_floor/std': 0.01906200684607029, 'rewards/near_duplicate_penalty/mean': 0.8726427555084229, 'rewards/near_duplicate_penalty/std': 0.05706513300538063, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.8296248912811279, 'rewards/distinct_2/std': 0.10697593539953232, 'rewards/total_composite/mean': 0.18172188103199005, 'rewards/total_composite/std': 0.16507014632225037, 'reward': 0.18172188103199005, 'reward_std': 0.16507014632225037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13479001820087433, 'sampling/sampling_logp_difference/max': 3.075570583343506, 'sampling/importance_sampling_ratio/min': 0.04616328328847885, 'sampling/importance_sampling_ratio/mean': 1.0149714946746826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8421830460429192, 'clip_ratio/low_mean': 0.09485042933374643, 'clip_ratio/low_min': 0.09485042933374643, 'clip_ratio/high_mean': 0.04511863738298416, 'clip_ratio/high_max': 0.04511863738298416, 'clip_ratio/region_mean': 0.1399690667167306, 'reward_total_mean': 0.18172188103199005, 'reward_meter_mean': 0.47071734070777893, 'reward_meter_std': 0.3999786078929901, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.7076892852783203, 'reward_repeat_penalty_std': 0.16064846515655518, 'reward_repeat_floor_mean': 0.9636876583099365, 'reward_repeat_floor_std': 0.01906200684607029, 'reward_near_duplicate_penalty_mean': 0.8726427555084229, 'reward_near_duplicate_penalty_std': 0.05706513300538063, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.8296248912811279, 'reward_distinct_2_std': 0.10697593539953232, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.18172188103199005, 'reward_total_composite_std': 0.16507014632225037, 'epoch': 0.08}

 47%|████▋     | 1549/3300 [5:54:09<3:51:01,  7.92s/it]INFO 04-12 21:52:35 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1550/3300 [5:54:16<3:47:48,  7.81s/it]2026-04-12 21:52:42,741 | INFO | train_grpo_train | metrics_logged mode=train step=1550

                                                       
{'loss': -0.0245, 'grad_norm': 12.013070106506348, 'learning_rate': 5.306060606060606e-06, 'num_tokens': 2736447.0, 'completions/mean_length': 40.875, 'completions/min_length': 35.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.1458166390657425, 'rewards/meter/std': 0.15230119228363037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/repeat_penalty/mean': 0.8864622116088867, 'rewards/repeat_penalty/std': 0.13986869156360626, 'rewards/repeat_floor/mean': 0.9862712025642395, 'rewards/repeat_floor/std': 0.01546245813369751, 'rewards/near_duplicate_penalty/mean': 0.9411143660545349, 'rewards/near_duplicate_penalty/std': 0.0577935054898262, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9664342403411865, 'rewards/distinct_2/std': 0.048219744116067886, 'rewards/total_composite/mean': 0.06861890107393265, 'rewards/total_composite/std': 0.06434255838394165, 'reward': 0.06861890107393265, 'reward_std': 0.06434255093336105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11489810049533844, 'sampling/sampling_logp_difference/max': 1.5114288330078125, 'sampling/importance_sampling_ratio/min': 0.22059455513954163, 'sampling/importance_sampling_ratio/mean': 1.0160824060440063, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8503337800502777, 'clip_ratio/low_mean': 0.07352757221087813, 'clip_ratio/low_min': 0.07352757221087813, 'clip_ratio/high_mean': 0.025568182580173016, 'clip_ratio/high_max': 0.025568182580173016, 'clip_ratio/region_mean': 0.09909575479105115, 'reward_total_mean': 0.06861890107393265, 'reward_meter_mean': 0.1458166390657425, 'reward_meter_std': 0.15230119228363037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8864622116088867, 'reward_repeat_penalty_std': 0.13986869156360626, 'reward_repeat_floor_mean': 0.9862712025642395, 'reward_repeat_floor_std': 0.01546245813369751, 'reward_near_duplicate_penalty_mean': 0.9411143660545349, 'reward_near_duplicate_penalty_std': 0.0577935054898262, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9664342403411865, 'reward_distinct_2_std': 0.048219744116067886, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.06861890107393265, 'reward_total_composite_std': 0.06434255838394165, 'epoch': 0.08}

 47%|████▋     | 1550/3300 [5:54:16<3:47:48,  7.81s/it]INFO 04-12 21:52:42 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/13 [00:00<?, ?it/s]

 15%|█▌        | 2/13 [00:08<00:48,  4.45s/it]

 23%|██▎       | 3/13 [00:12<00:41,  4.17s/it]

 31%|███       | 4/13 [00:22<00:55,  6.12s/it]

 38%|███▊      | 5/13 [00:31<00:57,  7.18s/it]

 46%|████▌     | 6/13 [00:37<00:47,  6.72s/it]

 54%|█████▍    | 7/13 [00:46<00:44,  7.50s/it]

 62%|██████▏   | 8/13 [00:54<00:39,  7.85s/it]

 69%|██████▉   | 9/13 [01:07<00:37,  9.47s/it]

 77%|███████▋  | 10/13 [01:16<00:27,  9.32s/it]

 85%|████████▍ | 11/13 [01:26<00:18,  9.39s/it]

 92%|█████████▏| 12/13 [01:35<00:09,  9.23s/it]

100%|██████████| 13/13 [01:40<00:00,  7.96s/it]2026-04-12 21:54:29,472 | INFO | train_grpo_train | metrics_logged mode=eval step=1550

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 106.5921, 'eval_samples_per_second': 0.976, 'eval_steps_per_second': 0.122, 'eval_num_tokens': 2736447.0, 'eval_completions/mean_length': 148.26923076923077, 'eval_completions/min_length': 37.69230769230769, 'eval_completions/max_length': 353.2307692307692, 'eval_completions/clipped_ratio': 0.09615384615384616, 'eval_completions/mean_terminated_length': 109.64001993032602, 'eval_completions/min_terminated_length': 37.69230769230769, 'eval_completions/max_terminated_length': 209.6153846153846, 'eval_rewards/meter/mean': 0.5370853772530189, 'eval_rewards/meter/std': 0.3735564396931575, 'eval_rewards/count_adherence/mean': 0.7950394474543058, 'eval_rewards/count_adherence/std': 0.1830053117412787, 'eval_rewards/arabic_clean/mean': 0.8653846153846154, 'eval_rewards/arabic_clean/std': 0.2755125302534837, 'eval_rewards/hard_gate/mean': 0.9903846153846154, 'eval_rewards/hard_gate/std': 0.027196414195574246, 'eval_rewards/arabic_floor/mean': 0.9956730741720933, 'eval_rewards/arabic_floor/std': 0.012238385012516608, 'eval_rewards/count_floor/mean': 0.9385118301098163, 'eval_rewards/count_floor/std': 0.05490159529906053, 'eval_rewards/lexical_plausibility/mean': 0.9523813541118915, 'eval_rewards/lexical_plausibility/std': 0.09585086657450749, 'eval_rewards/judge_quality/mean': 0.43240384413645816, 'eval_rewards/judge_quality/std': 0.16614713770552322, 'eval_rewards/repeat_penalty/mean': 0.7907334153468792, 'eval_rewards/repeat_penalty/std': 0.19797226843925622, 'eval_rewards/repeat_floor/mean': 0.9776892845447247, 'eval_rewards/repeat_floor/std': 0.021466942933889534, 'eval_rewards/near_duplicate_penalty/mean': 0.9485286657626812, 'eval_rewards/near_duplicate_penalty/std': 0.04545352521997232, 'eval_rewards/opening_diversity/mean': 0.9847964552732614, 'eval_rewards/opening_diversity/std': 0.040863610517520174, 'eval_rewards/distinct_2/mean': 0.8480243361913241, 'eval_rewards/distinct_2/std': 0.17115985401547873, 'eval_rewards/total_composite/mean': 0.2187344225553366, 'eval_rewards/total_composite/std': 0.1744088610777488, 'eval_reward': 0.2187344225553366, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06427777701845536, 'eval_sampling/sampling_logp_difference/max': 1.0461443937741792, 'eval_sampling/importance_sampling_ratio/min': 0.355395835179549, 'eval_sampling/importance_sampling_ratio/mean': 1.0153421805455134, 'eval_sampling/importance_sampling_ratio/max': 1.4949395656585693, 'eval_entropy': 0.6981543944432185, 'eval_clip_ratio/low_mean': 0.0, 'eval_clip_ratio/low_min': 0.0, 'eval_clip_ratio/high_mean': 0.0, 'eval_clip_ratio/high_max': 0.0, 'eval_clip_ratio/region_mean': 0.0, 'eval_reward_total_mean': 0.2187344225553366, 'eval_reward_meter_mean': 0.5370853772530189, 'eval_reward_meter_std': 0.3735564396931575, 'eval_reward_count_adherence_mean': 0.7950394474543058, 'eval_reward_count_adherence_std': 0.1830053117412787, 'eval_reward_arabic_clean_mean': 0.8653846153846154, 'eval_reward_arabic_clean_std': 0.2755125302534837, 'eval_reward_hard_gate_mean': 0.9903846153846154, 'eval_reward_hard_gate_std': 0.027196414195574246, 'eval_reward_arabic_floor_mean': 0.9956730741720933, 'eval_reward_arabic_floor_std': 0.012238385012516608, 'eval_reward_count_floor_mean': 0.9385118301098163, 'eval_reward_count_floor_std': 0.05490159529906053, 'eval_reward_lexical_plausibility_mean': 0.9523813541118915, 'eval_reward_lexical_plausibility_std': 0.09585086657450749, 'eval_reward_repeat_penalty_mean': 0.7907334153468792, 'eval_reward_repeat_penalty_std': 0.19797226843925622, 'eval_reward_repeat_floor_mean': 0.9776892845447247, 'eval_reward_repeat_floor_std': 0.021466942933889534, 'eval_reward_near_duplicate_penalty_mean': 0.9485286657626812, 'eval_reward_near_duplicate_penalty_std': 0.04545352521997232, 'eval_reward_opening_diversity_mean': 0.9847964552732614, 'eval_reward_opening_diversity_std': 0.040863610517520174, 'eval_reward_distinct_2_mean': 0.8480243361913241, 'eval_reward_distinct_2_std': 0.17115985401547873, 'eval_reward_judge_quality_mean': 0.43240384413645816, 'eval_reward_judge_quality_std': 0.16614713770552322, 'eval_reward_total_composite_mean': 0.2187344225553366, 'eval_reward_total_composite_std': 0.1744088610777488, 'epoch': 0.08}

 47%|████▋     | 1550/3300 [5:56:03<3:47:48,  7.81s/it]

100%|██████████| 13/13 [01:40<00:00,  7.96s/it]

                                               /root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/torch/utils/checkpoint.py:85: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
  warnings.warn(

 47%|████▋     | 1551/3300 [5:56:17<20:19:09, 41.82s/it]2026-04-12 21:54:43,931 | INFO | train_grpo_train | metrics_logged mode=train step=1551

                                                        
{'loss': 0.024, 'grad_norm': 11.444884300231934, 'learning_rate': 5.303030303030303e-06, 'num_tokens': 2738063.0, 'completions/mean_length': 43.0, 'completions/min_length': 39.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8957365155220032, 'rewards/meter/std': 0.20929378271102905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/repeat_penalty/mean': 0.8869351148605347, 'rewards/repeat_penalty/std': 0.14403527975082397, 'rewards/repeat_floor/mean': 0.9862526059150696, 'rewards/repeat_floor/std': 0.016092751175165176, 'rewards/near_duplicate_penalty/mean': 0.943255603313446, 'rewards/near_duplicate_penalty/std': 0.05065969377756119, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9626585245132446, 'rewards/distinct_2/std': 0.045198049396276474, 'rewards/total_composite/mean': 0.44593286514282227, 'rewards/total_composite/std': 0.18317784368991852, 'reward': 0.44593286514282227, 'reward_std': 0.18317782878875732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13498082756996155, 'sampling/sampling_logp_difference/max': 1.7982535362243652, 'sampling/importance_sampling_ratio/min': 0.16558782756328583, 'sampling/importance_sampling_ratio/mean': 0.9999086260795593, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.872445285320282, 'clip_ratio/low_mean': 0.09172903839498758, 'clip_ratio/low_min': 0.09172903839498758, 'clip_ratio/high_mean': 0.043604651466012, 'clip_ratio/high_max': 0.043604651466012, 'clip_ratio/region_mean': 0.13533368986099958, 'reward_total_mean': 0.44593286514282227, 'reward_meter_mean': 0.8957365155220032, 'reward_meter_std': 0.20929378271102905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8869351148605347, 'reward_repeat_penalty_std': 0.14403527975082397, 'reward_repeat_floor_mean': 0.9862526059150696, 'reward_repeat_floor_std': 0.016092751175165176, 'reward_near_duplicate_penalty_mean': 0.943255603313446, 'reward_near_duplicate_penalty_std': 0.05065969377756119, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9626585245132446, 'reward_distinct_2_std': 0.045198049396276474, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.44593286514282227, 'reward_total_composite_std': 0.18317784368991852, 'epoch': 0.08}

 47%|████▋     | 1551/3300 [5:56:17<20:19:09, 41.82s/it]INFO 04-12 21:54:44 [block_pool.py:378] Successfully reset prefix cache