INFO 04-12 15:23:21 [__init__.py:216] Automatically detected platform cuda.
2026-04-12 15:23:22,570 | INFO | train_grpo_train | mode=train

Fetching 128 files:   0%|          | 0/128 [00:00<?, ?it/s]
Fetching 128 files: 100%|██████████| 128/128 [00:00<00:00, 12564.85it/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,  1.35it/s]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:01<00:00,  1.37it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.44it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.42it/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-12 15:25:00 [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 15:25:00 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-12 15:25:00 [model.py:1510] Using max model len 1536
INFO 04-12 15:25:01 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-12 15:25:01 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-12 15:25:01 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-12 15:25:01 [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 15:25:02 [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 15:25:02 [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 15:25:02 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-12 15:25:03 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-12 15:25:03 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-12 15:25:03 [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<00:00,  2.21it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.67it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.51it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.58it/s]

INFO 04-12 15:25:05 [default_loader.py:267] Loading weights took 2.09 seconds
INFO 04-12 15:25:06 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.438951 seconds
INFO 04-12 15:25:11 [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 15:25:11 [backends.py:559] Dynamo bytecode transform time: 5.27 s
INFO 04-12 15:25:15 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.684 s
INFO 04-12 15:25:16 [monitor.py:34] torch.compile takes 5.27 s in total
INFO 04-12 15:25:17 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-12 15:25:17 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-12 15:25:17 [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, 17.50it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  80%|████████  | 4/5 [00:00<00:00, 18.15it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 16.57it/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, 19.95it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 20.62it/s]
INFO 04-12 15:25:19 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-12 15:25:19 [core.py:210] init engine (profile, create kv cache, warmup model) took 13.09 seconds
INFO 04-12 15:25:19 [llm.py:306] Supported_tasks: ('generate',)
2026-04-12 15:25:19,508 | INFO | train_grpo_train | dataset_id=Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1 train_size=24897 eval_size=104
2026-04-12 15:25:19,509 | 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 15:25:19,509 | 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 15:25:19,509 | 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 15:25:19,509 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo-friend-v1
2026-04-12 15:25:19,509 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_drop_trio/cap_3000/selected_manifest.csv
2026-04-12 15:25:19,509 | INFO | train_grpo_train | hard_diagnostic_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_drop_trio/hard_diagnostic_cap_256/selected_manifest.csv
2026-04-12 15:25:19,509 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-12 15:25:19,509 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-12T15:25:00Z", "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 15:25:19,509 | 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 15:25:20 [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 [00:06<6:04:41,  6.63s/it]2026-04-12 15:25:26,556 | INFO | train_grpo_train | metrics_logged mode=train step=1

                                                  
{'loss': 0.1191, 'grad_norm': 10.253804206848145, 'learning_rate': 1e-05, 'num_tokens': 2463.0, 'completions/mean_length': 121.875, 'completions/min_length': 81.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.875, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.42592284083366394, 'rewards/meter/std': 0.33591756224632263, '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': 0.38749998807907104, 'rewards/judge_quality/std': 0.0517549142241478, 'rewards/repeat_penalty/mean': 0.9937220811843872, 'rewards/repeat_penalty/std': 0.004517034627497196, 'rewards/repeat_floor/mean': 0.9990583658218384, 'rewards/repeat_floor/std': 0.0006775603396818042, 'rewards/near_duplicate_penalty/mean': 0.9937220811843872, 'rewards/near_duplicate_penalty/std': 0.004517034627497196, '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': 0.15973103046417236, 'rewards/total_composite/std': 0.12557268142700195, 'reward': 0.15973103046417236, 'reward_std': 0.12557268142700195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2232884168624878, 'sampling/sampling_logp_difference/max': 2.2234439849853516, 'sampling/importance_sampling_ratio/min': 0.10823570936918259, 'sampling/importance_sampling_ratio/mean': 1.0212712287902832, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.127237632870674, 'clip_ratio/low_mean': 0.1246343944221735, 'clip_ratio/low_min': 0.1246343944221735, 'clip_ratio/high_mean': 0.08433323167264462, 'clip_ratio/high_max': 0.08433323167264462, 'clip_ratio/region_mean': 0.20896762609481812, 'reward_total_mean': 0.15973103046417236, 'reward_meter_mean': 0.42592284083366394, 'reward_meter_std': 0.33591756224632263, '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': 0.9937220811843872, 'reward_repeat_penalty_std': 0.004517034627497196, 'reward_repeat_floor_mean': 0.9990583658218384, 'reward_repeat_floor_std': 0.0006775603396818042, 'reward_near_duplicate_penalty_mean': 0.9937220811843872, 'reward_near_duplicate_penalty_std': 0.004517034627497196, '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': 0.15973103046417236, 'reward_total_composite_std': 0.12557268142700195, 'epoch': 0.0}

  0%|          | 1/3300 [00:06<6:04:41,  6.63s/it]INFO 04-12 15:25:26 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 2/3300 [00:16<7:50:25,  8.56s/it]2026-04-12 15:25:36,455 | INFO | train_grpo_train | metrics_logged mode=train step=2

                                                  
{'loss': 0.0996, 'grad_norm': 14.527668952941895, 'learning_rate': 9.996969696969698e-06, 'num_tokens': 4959.0, 'completions/mean_length': 111.0, 'completions/min_length': 96.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.718535304069519, 'rewards/meter/std': 0.17709191143512726, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, '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': 0.9838680028915405, 'rewards/repeat_penalty/std': 0.006402861792594194, 'rewards/repeat_floor/mean': 0.9975801706314087, 'rewards/repeat_floor/std': 0.0009604145889170468, 'rewards/near_duplicate_penalty/mean': 0.9838680028915405, 'rewards/near_duplicate_penalty/std': 0.006402861792594194, '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': 0.5599968433380127, 'rewards/total_composite/std': 0.22703492641448975, 'reward': 0.5599968433380127, 'reward_std': 0.22703492641448975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1736430674791336, 'sampling/sampling_logp_difference/max': 3.8660693168640137, 'sampling/importance_sampling_ratio/min': 0.020940518006682396, 'sampling/importance_sampling_ratio/mean': 0.9950367212295532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6591976061463356, 'clip_ratio/low_mean': 0.06499460339546204, 'clip_ratio/low_min': 0.06499460339546204, 'clip_ratio/high_mean': 0.0667342699598521, 'clip_ratio/high_max': 0.0667342699598521, 'clip_ratio/region_mean': 0.13172887335531414, 'reward_total_mean': 0.5599968433380127, 'reward_meter_mean': 0.718535304069519, 'reward_meter_std': 0.17709191143512726, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9838680028915405, 'reward_repeat_penalty_std': 0.006402861792594194, 'reward_repeat_floor_mean': 0.9975801706314087, 'reward_repeat_floor_std': 0.0009604145889170468, 'reward_near_duplicate_penalty_mean': 0.9838680028915405, 'reward_near_duplicate_penalty_std': 0.006402861792594194, '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': 0.21767853200435638, 'reward_total_composite_mean': 0.5599968433380127, 'reward_total_composite_std': 0.22703492641448975, 'epoch': 0.0}

  0%|          | 2/3300 [00:16<7:50:25,  8.56s/it]INFO 04-12 15:25:36 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 3/3300 [00:24<7:41:48,  8.40s/it]2026-04-12 15:25:44,675 | INFO | train_grpo_train | metrics_logged mode=train step=3

                                                  
{'loss': 0.1111, 'grad_norm': 16.40032196044922, 'learning_rate': 9.993939393939395e-06, 'num_tokens': 6509.0, 'completions/mean_length': 38.75, 'completions/min_length': 30.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.35528749227523804, 'rewards/meter/std': 0.3985433578491211, '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': 0.9887152910232544, 'rewards/lexical_plausibility/std': 0.022863244637846947, 'rewards/judge_quality/mean': 0.6262500286102295, 'rewards/judge_quality/std': 0.3325631320476532, 'rewards/repeat_penalty/mean': 0.9862549304962158, 'rewards/repeat_penalty/std': 0.020482772961258888, 'rewards/repeat_floor/mean': 0.9979382753372192, 'rewards/repeat_floor/std': 0.0030724238604307175, 'rewards/near_duplicate_penalty/mean': 0.9862549304962158, 'rewards/near_duplicate_penalty/std': 0.020482772961258888, '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': 0.28549739718437195, 'rewards/total_composite/std': 0.39841151237487793, 'reward': 0.28549739718437195, 'reward_std': 0.39841148257255554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2212861031293869, 'sampling/sampling_logp_difference/max': 1.8380317687988281, 'sampling/importance_sampling_ratio/min': 0.15913031995296478, 'sampling/importance_sampling_ratio/mean': 1.007591962814331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8705281987786293, 'clip_ratio/low_mean': 0.15355645678937435, 'clip_ratio/low_min': 0.15355645678937435, 'clip_ratio/high_mean': 0.0538194440305233, 'clip_ratio/high_max': 0.0538194440305233, 'clip_ratio/region_mean': 0.20737590081989765, 'reward_total_mean': 0.28549739718437195, 'reward_meter_mean': 0.35528749227523804, 'reward_meter_std': 0.3985433578491211, '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': 0.9887152910232544, 'reward_lexical_plausibility_std': 0.022863244637846947, 'reward_repeat_penalty_mean': 0.9862549304962158, 'reward_repeat_penalty_std': 0.020482772961258888, 'reward_repeat_floor_mean': 0.9979382753372192, 'reward_repeat_floor_std': 0.0030724238604307175, 'reward_near_duplicate_penalty_mean': 0.9862549304962158, 'reward_near_duplicate_penalty_std': 0.020482772961258888, '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': 0.6262500286102295, 'reward_judge_quality_std': 0.3325631320476532, 'reward_total_composite_mean': 0.28549739718437195, 'reward_total_composite_std': 0.39841151237487793, 'epoch': 0.0}

  0%|          | 3/3300 [00:24<7:41:48,  8.40s/it]INFO 04-12 15:25:44 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 4/3300 [00:36<8:54:42,  9.73s/it]2026-04-12 15:25:56,461 | INFO | train_grpo_train | metrics_logged mode=train step=4

                                                  
{'loss': -0.1386, 'grad_norm': 8.324551582336426, 'learning_rate': 9.990909090909093e-06, 'num_tokens': 9508.0, 'completions/mean_length': 159.875, 'completions/min_length': 96.0, 'completions/max_length': 229.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.875, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 229.0, 'rewards/meter/mean': 0.5181947946548462, 'rewards/meter/std': 0.3963140845298767, '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': 0.9840686321258545, 'rewards/lexical_plausibility/std': 0.03130572661757469, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/repeat_penalty/mean': 0.9750292897224426, 'rewards/repeat_penalty/std': 0.055296771228313446, 'rewards/repeat_floor/mean': 0.9972373247146606, 'rewards/repeat_floor/std': 0.0055242059752345085, 'rewards/near_duplicate_penalty/mean': 0.9920720458030701, 'rewards/near_duplicate_penalty/std': 0.007942475378513336, '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': 0.17476558685302734, 'rewards/total_composite/std': 0.14435797929763794, 'reward': 0.17476558685302734, 'reward_std': 0.14435797929763794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22558651864528656, 'sampling/sampling_logp_difference/max': 2.412989616394043, 'sampling/importance_sampling_ratio/min': 0.08954718708992004, 'sampling/importance_sampling_ratio/mean': 1.0321437120437622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5734728425741196, 'clip_ratio/low_mean': 0.12947455421090126, 'clip_ratio/low_min': 0.12947455421090126, 'clip_ratio/high_mean': 0.07497359253466129, 'clip_ratio/high_max': 0.07497359253466129, 'clip_ratio/region_mean': 0.20444814674556255, 'reward_total_mean': 0.17476558685302734, 'reward_meter_mean': 0.5181947946548462, 'reward_meter_std': 0.3963140845298767, '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': 0.9840686321258545, 'reward_lexical_plausibility_std': 0.03130572661757469, 'reward_repeat_penalty_mean': 0.9750292897224426, 'reward_repeat_penalty_std': 0.055296771228313446, 'reward_repeat_floor_mean': 0.9972373247146606, 'reward_repeat_floor_std': 0.0055242059752345085, 'reward_near_duplicate_penalty_mean': 0.9920720458030701, 'reward_near_duplicate_penalty_std': 0.007942475378513336, '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': 0.32499998807907104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.17476558685302734, 'reward_total_composite_std': 0.14435797929763794, 'epoch': 0.0}

  0%|          | 4/3300 [00:36<8:54:42,  9.73s/it]INFO 04-12 15:25:56 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 5/3300 [00:45<8:37:34,  9.42s/it]2026-04-12 15:26:05,324 | INFO | train_grpo_train | metrics_logged mode=train step=5

                                                  
{'loss': -0.0003, 'grad_norm': 17.628894805908203, 'learning_rate': 9.987878787878788e-06, 'num_tokens': 11109.0, 'completions/mean_length': 46.125, 'completions/min_length': 37.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5038740634918213, 'rewards/meter/std': 0.48447510600090027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.5087500214576721, 'rewards/judge_quality/std': 0.21820290386676788, 'rewards/repeat_penalty/mean': 0.992081880569458, 'rewards/repeat_penalty/std': 0.020084861665964127, 'rewards/repeat_floor/mean': 0.9988123178482056, 'rewards/repeat_floor/std': 0.003012728178873658, 'rewards/near_duplicate_penalty/mean': 0.992081880569458, 'rewards/near_duplicate_penalty/std': 0.020084861665964127, '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': 0.25947105884552, 'rewards/total_composite/std': 0.29764559864997864, 'reward': 0.25947105884552, 'reward_std': 0.29764559864997864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2340587079524994, 'sampling/sampling_logp_difference/max': 1.5304131507873535, 'sampling/importance_sampling_ratio/min': 0.21644623577594757, 'sampling/importance_sampling_ratio/mean': 1.0601122379302979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2103342562913895, 'clip_ratio/low_mean': 0.12461328227072954, 'clip_ratio/low_min': 0.12461328227072954, 'clip_ratio/high_mean': 0.058544520288705826, 'clip_ratio/high_max': 0.058544520288705826, 'clip_ratio/region_mean': 0.18315780255943537, 'reward_total_mean': 0.25947105884552, 'reward_meter_mean': 0.5038740634918213, 'reward_meter_std': 0.48447510600090027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.992081880569458, 'reward_repeat_penalty_std': 0.020084861665964127, 'reward_repeat_floor_mean': 0.9988123178482056, 'reward_repeat_floor_std': 0.003012728178873658, 'reward_near_duplicate_penalty_mean': 0.992081880569458, 'reward_near_duplicate_penalty_std': 0.020084861665964127, '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': 0.5087500214576721, 'reward_judge_quality_std': 0.21820290386676788, 'reward_total_composite_mean': 0.25947105884552, 'reward_total_composite_std': 0.29764559864997864, 'epoch': 0.0}

  0%|          | 5/3300 [00:45<8:37:34,  9.42s/it]INFO 04-12 15:26:05 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 6/3300 [00:54<8:33:55,  9.36s/it]2026-04-12 15:26:14,563 | INFO | train_grpo_train | metrics_logged mode=train step=6

                                                  
{'loss': -0.0588, 'grad_norm': 18.911636352539062, 'learning_rate': 9.984848484848485e-06, 'num_tokens': 12709.0, 'completions/mean_length': 42.0, 'completions/min_length': 33.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.550910234451294, 'rewards/meter/std': 0.4658724069595337, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.2745385766029358, 'rewards/repeat_penalty/mean': 0.9660366773605347, 'rewards/repeat_penalty/std': 0.049621619284152985, 'rewards/repeat_floor/mean': 0.9952383637428284, 'rewards/repeat_floor/std': 0.006644129753112793, 'rewards/near_duplicate_penalty/mean': 0.9725293517112732, 'rewards/near_duplicate_penalty/std': 0.034934960305690765, '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': 0.29271385073661804, 'rewards/total_composite/std': 0.29651957750320435, 'reward': 0.29271385073661804, 'reward_std': 0.29651951789855957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24054525792598724, 'sampling/sampling_logp_difference/max': 1.5981216430664062, 'sampling/importance_sampling_ratio/min': 0.2022761106491089, 'sampling/importance_sampling_ratio/mean': 1.0655957460403442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4212131202220917, 'clip_ratio/low_mean': 0.08407547883689404, 'clip_ratio/low_min': 0.08407547883689404, 'clip_ratio/high_mean': 0.09894526936113834, 'clip_ratio/high_max': 0.09894526936113834, 'clip_ratio/region_mean': 0.18302074819803238, 'reward_total_mean': 0.29271385073661804, 'reward_meter_mean': 0.550910234451294, 'reward_meter_std': 0.4658724069595337, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9660366773605347, 'reward_repeat_penalty_std': 0.049621619284152985, 'reward_repeat_floor_mean': 0.9952383637428284, 'reward_repeat_floor_std': 0.006644129753112793, 'reward_near_duplicate_penalty_mean': 0.9725293517112732, 'reward_near_duplicate_penalty_std': 0.034934960305690765, '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': 0.5049999952316284, 'reward_judge_quality_std': 0.2745385766029358, 'reward_total_composite_mean': 0.29271385073661804, 'reward_total_composite_std': 0.29651957750320435, 'epoch': 0.0}

  0%|          | 6/3300 [00:54<8:33:55,  9.36s/it]INFO 04-12 15:26:14 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 7/3300 [01:03<8:31:09,  9.31s/it]2026-04-12 15:26:23,776 | INFO | train_grpo_train | metrics_logged mode=train step=7

                                                  
{'loss': 0.0227, 'grad_norm': 20.267526626586914, 'learning_rate': 9.981818181818183e-06, 'num_tokens': 14297.0, 'completions/mean_length': 39.5, 'completions/min_length': 35.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6312246322631836, 'rewards/meter/std': 0.4418559670448303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.12464234977960587, 'rewards/repeat_penalty/mean': 0.990004301071167, 'rewards/repeat_penalty/std': 0.01144260074943304, 'rewards/repeat_floor/mean': 0.998500645160675, 'rewards/repeat_floor/std': 0.001716392464004457, 'rewards/near_duplicate_penalty/mean': 0.990004301071167, 'rewards/near_duplicate_penalty/std': 0.01144260074943304, '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': 0.26341506838798523, 'rewards/total_composite/std': 0.17551356554031372, 'reward': 0.26341506838798523, 'reward_std': 0.17551355063915253, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2223813384771347, 'sampling/sampling_logp_difference/max': 1.7176733016967773, 'sampling/importance_sampling_ratio/min': 0.17948326468467712, 'sampling/importance_sampling_ratio/mean': 1.0373649597167969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9773674011230469, 'clip_ratio/low_mean': 0.07656641863286495, 'clip_ratio/low_min': 0.07656641863286495, 'clip_ratio/high_mean': 0.10327381175011396, 'clip_ratio/high_max': 0.10327381175011396, 'clip_ratio/region_mean': 0.17984023038297892, 'reward_total_mean': 0.26341506838798523, 'reward_meter_mean': 0.6312246322631836, 'reward_meter_std': 0.4418559670448303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.990004301071167, 'reward_repeat_penalty_std': 0.01144260074943304, 'reward_repeat_floor_mean': 0.998500645160675, 'reward_repeat_floor_std': 0.001716392464004457, 'reward_near_duplicate_penalty_mean': 0.990004301071167, 'reward_near_duplicate_penalty_std': 0.01144260074943304, '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': 0.12464234977960587, 'reward_total_composite_mean': 0.26341506838798523, 'reward_total_composite_std': 0.17551356554031372, 'epoch': 0.0}

  0%|          | 7/3300 [01:03<8:31:09,  9.31s/it]INFO 04-12 15:26:24 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 8/3300 [01:17<9:44:01, 10.64s/it]2026-04-12 15:26:37,271 | INFO | train_grpo_train | metrics_logged mode=train step=8

                                                  
{'loss': -0.0994, 'grad_norm': 5.759873867034912, 'learning_rate': 9.97878787878788e-06, 'num_tokens': 18531.0, 'completions/mean_length': 334.25, 'completions/min_length': 242.0, 'completions/max_length': 388.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 334.25, 'completions/min_terminated_length': 242.0, 'completions/max_terminated_length': 388.0, 'rewards/meter/mean': 0.7083072662353516, 'rewards/meter/std': 0.20774205029010773, 'rewards/count_adherence/mean': 0.9090909361839294, 'rewards/count_adherence/std': 0.048592954874038696, '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': 0.9727272987365723, 'rewards/count_floor/std': 0.014577888883650303, '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': 0.9861735701560974, 'rewards/repeat_penalty/std': 0.01844516023993492, 'rewards/repeat_floor/mean': 0.9983733296394348, 'rewards/repeat_floor/std': 0.0017793928273022175, 'rewards/near_duplicate_penalty/mean': 0.9937269687652588, 'rewards/near_duplicate_penalty/std': 0.003104675328359008, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9923807382583618, 'rewards/distinct_2/std': 0.01718667894601822, 'rewards/total_composite/mean': 0.22860600054264069, 'rewards/total_composite/std': 0.08173606544733047, 'reward': 0.22860600054264069, 'reward_std': 0.08173606544733047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20303091406822205, 'sampling/sampling_logp_difference/max': 2.9986186027526855, 'sampling/importance_sampling_ratio/min': 0.04985589161515236, 'sampling/importance_sampling_ratio/mean': 1.024811029434204, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.436163306236267, 'clip_ratio/low_mean': 0.07238915748894215, 'clip_ratio/low_min': 0.07238915748894215, 'clip_ratio/high_mean': 0.13320296816527843, 'clip_ratio/high_max': 0.13320296816527843, 'clip_ratio/region_mean': 0.20559212565422058, 'reward_total_mean': 0.22860600054264069, 'reward_meter_mean': 0.7083072662353516, 'reward_meter_std': 0.20774205029010773, 'reward_count_adherence_mean': 0.9090909361839294, 'reward_count_adherence_std': 0.048592954874038696, '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': 0.9727272987365723, 'reward_count_floor_std': 0.014577888883650303, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9861735701560974, 'reward_repeat_penalty_std': 0.01844516023993492, 'reward_repeat_floor_mean': 0.9983733296394348, 'reward_repeat_floor_std': 0.0017793928273022175, 'reward_near_duplicate_penalty_mean': 0.9937269687652588, 'reward_near_duplicate_penalty_std': 0.003104675328359008, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9923807382583618, 'reward_distinct_2_std': 0.01718667894601822, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.08345229178667068, 'reward_total_composite_mean': 0.22860600054264069, 'reward_total_composite_std': 0.08173606544733047, 'epoch': 0.0}

  0%|          | 8/3300 [01:17<9:44:01, 10.64s/it]INFO 04-12 15:26:37 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 9/3300 [01:28<9:47:33, 10.71s/it]2026-04-12 15:26:48,133 | INFO | train_grpo_train | metrics_logged mode=train step=9

                                                  
{'loss': 0.0052, 'grad_norm': 10.515203475952148, 'learning_rate': 9.975757575757577e-06, 'num_tokens': 21015.0, 'completions/mean_length': 127.5, 'completions/min_length': 110.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.5, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.656243085861206, 'rewards/meter/std': 0.28146716952323914, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, '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': 0.5637500286102295, 'rewards/judge_quality/std': 0.3137759864330292, 'rewards/repeat_penalty/mean': 0.9839040637016296, 'rewards/repeat_penalty/std': 0.014884334988892078, 'rewards/repeat_floor/mean': 0.997585654258728, 'rewards/repeat_floor/std': 0.002232639119029045, 'rewards/near_duplicate_penalty/mean': 0.9839040637016296, 'rewards/near_duplicate_penalty/std': 0.014884334988892078, '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': 0.364164799451828, 'rewards/total_composite/std': 0.25800395011901855, 'reward': 0.364164799451828, 'reward_std': 0.25800392031669617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23039698600769043, 'sampling/sampling_logp_difference/max': 2.888472080230713, 'sampling/importance_sampling_ratio/min': 0.055661194026470184, 'sampling/importance_sampling_ratio/mean': 0.9615637063980103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7885613832622766, 'clip_ratio/low_mean': 0.09423136431723833, 'clip_ratio/low_min': 0.09423136431723833, 'clip_ratio/high_mean': 0.08434331836178899, 'clip_ratio/high_max': 0.08434331836178899, 'clip_ratio/region_mean': 0.17857468267902732, 'reward_total_mean': 0.364164799451828, 'reward_meter_mean': 0.656243085861206, 'reward_meter_std': 0.28146716952323914, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, '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': 0.9839040637016296, 'reward_repeat_penalty_std': 0.014884334988892078, 'reward_repeat_floor_mean': 0.997585654258728, 'reward_repeat_floor_std': 0.002232639119029045, 'reward_near_duplicate_penalty_mean': 0.9839040637016296, 'reward_near_duplicate_penalty_std': 0.014884334988892078, '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': 0.3137759864330292, 'reward_total_composite_mean': 0.364164799451828, 'reward_total_composite_std': 0.25800395011901855, 'epoch': 0.0}

  0%|          | 9/3300 [01:28<9:47:33, 10.71s/it]INFO 04-12 15:26:48 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 10/3300 [01:38<9:39:01, 10.56s/it]2026-04-12 15:26:58,369 | INFO | train_grpo_train | metrics_logged mode=train step=10

                                                   
{'loss': 0.0707, 'grad_norm': 13.917400360107422, 'learning_rate': 9.972727272727274e-06, 'num_tokens': 23094.0, 'completions/mean_length': 85.875, 'completions/min_length': 64.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.7418884038925171, 'rewards/meter/std': 0.28886109590530396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.0353553369641304, 'rewards/repeat_penalty/mean': 0.9732223749160767, 'rewards/repeat_penalty/std': 0.02280677855014801, 'rewards/repeat_floor/mean': 0.9965280294418335, 'rewards/repeat_floor/std': 0.0028661310207098722, 'rewards/near_duplicate_penalty/mean': 0.9824955463409424, 'rewards/near_duplicate_penalty/std': 0.017405029386281967, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9905967712402344, 'rewards/distinct_2/std': 0.017439009621739388, 'rewards/total_composite/mean': 0.32619839906692505, 'rewards/total_composite/std': 0.136269211769104, 'reward': 0.32619839906692505, 'reward_std': 0.1362692266702652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.230515256524086, 'sampling/sampling_logp_difference/max': 1.4770076274871826, 'sampling/importance_sampling_ratio/min': 0.2283198982477188, 'sampling/importance_sampling_ratio/mean': 1.0057719945907593, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.104541003704071, 'clip_ratio/low_mean': 0.07275313138961792, 'clip_ratio/low_min': 0.07275313138961792, 'clip_ratio/high_mean': 0.14823600091040134, 'clip_ratio/high_max': 0.14823600091040134, 'clip_ratio/region_mean': 0.22098913230001926, 'reward_total_mean': 0.32619839906692505, 'reward_meter_mean': 0.7418884038925171, 'reward_meter_std': 0.28886109590530396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9732223749160767, 'reward_repeat_penalty_std': 0.02280677855014801, 'reward_repeat_floor_mean': 0.9965280294418335, 'reward_repeat_floor_std': 0.0028661310207098722, 'reward_near_duplicate_penalty_mean': 0.9824955463409424, 'reward_near_duplicate_penalty_std': 0.017405029386281967, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9905967712402344, 'reward_distinct_2_std': 0.017439009621739388, 'reward_judge_quality_mean': 0.4375, 'reward_judge_quality_std': 0.0353553369641304, 'reward_total_composite_mean': 0.32619839906692505, 'reward_total_composite_std': 0.136269211769104, 'epoch': 0.0}

  0%|          | 10/3300 [01:38<9:39:01, 10.56s/it]INFO 04-12 15:26:58 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 11/3300 [01:47<9:05:54,  9.96s/it]2026-04-12 15:27:06,948 | INFO | train_grpo_train | metrics_logged mode=train step=11

                                                   
{'loss': 0.026, 'grad_norm': 14.798701286315918, 'learning_rate': 9.96969696969697e-06, 'num_tokens': 24829.0, 'completions/mean_length': 50.875, 'completions/min_length': 28.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.3353729248046875, 'rewards/meter/std': 0.3401470482349396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.26249998807907104, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9405688643455505, 'rewards/repeat_penalty/std': 0.1435241848230362, 'rewards/repeat_floor/mean': 0.993354320526123, 'rewards/repeat_floor/std': 0.015140877105295658, 'rewards/near_duplicate_penalty/mean': 0.9754068851470947, 'rewards/near_duplicate_penalty/std': 0.04600444808602333, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9879807829856873, 'rewards/distinct_2/std': 0.03399552404880524, 'rewards/total_composite/mean': 0.07768288999795914, 'rewards/total_composite/std': 0.0734281837940216, 'reward': 0.07768288999795914, 'reward_std': 0.07342817634344101, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18992814421653748, 'sampling/sampling_logp_difference/max': 1.4641127586364746, 'sampling/importance_sampling_ratio/min': 0.23128311336040497, 'sampling/importance_sampling_ratio/mean': 1.0147560834884644, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.485662467777729, 'clip_ratio/low_mean': 0.10995643772184849, 'clip_ratio/low_min': 0.10995643772184849, 'clip_ratio/high_mean': 0.04779450502246618, 'clip_ratio/high_max': 0.04779450502246618, 'clip_ratio/region_mean': 0.15775094274431467, 'reward_total_mean': 0.07768288999795914, 'reward_meter_mean': 0.3353729248046875, 'reward_meter_std': 0.3401470482349396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9405688643455505, 'reward_repeat_penalty_std': 0.1435241848230362, 'reward_repeat_floor_mean': 0.993354320526123, 'reward_repeat_floor_std': 0.015140877105295658, 'reward_near_duplicate_penalty_mean': 0.9754068851470947, 'reward_near_duplicate_penalty_std': 0.04600444808602333, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9879807829856873, 'reward_distinct_2_std': 0.03399552404880524, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.07768288999795914, 'reward_total_composite_std': 0.0734281837940216, 'epoch': 0.0}

  0%|          | 11/3300 [01:47<9:05:54,  9.96s/it]INFO 04-12 15:27:07 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 12/3300 [02:01<10:17:14, 11.26s/it]2026-04-12 15:27:21,196 | INFO | train_grpo_train | metrics_logged mode=train step=12

                                                    
{'loss': 0.1291, 'grad_norm': 7.750916481018066, 'learning_rate': 9.966666666666667e-06, 'num_tokens': 29320.0, 'completions/mean_length': 336.375, 'completions/min_length': 224.0, 'completions/max_length': 459.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 336.375, 'completions/min_terminated_length': 224.0, 'completions/max_terminated_length': 459.0, 'rewards/meter/mean': 0.6111171245574951, 'rewards/meter/std': 0.25081607699394226, 'rewards/count_adherence/mean': 0.9519230723381042, 'rewards/count_adherence/std': 0.05723259598016739, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.985576868057251, 'rewards/count_floor/std': 0.017169784754514694, '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': 0.9864811897277832, 'rewards/repeat_penalty/std': 0.017781730741262436, 'rewards/repeat_floor/mean': 0.998455286026001, 'rewards/repeat_floor/std': 0.0017400516662746668, 'rewards/near_duplicate_penalty/mean': 0.9946328997612, 'rewards/near_duplicate_penalty/std': 0.003203645348548889, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9917812347412109, 'rewards/distinct_2/std': 0.01605413667857647, 'rewards/total_composite/mean': 0.12012744694948196, 'rewards/total_composite/std': 0.08017784357070923, 'reward': 0.12012744694948196, 'reward_std': 0.08017783612012863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22707612812519073, 'sampling/sampling_logp_difference/max': 2.110173225402832, 'sampling/importance_sampling_ratio/min': 0.12121696770191193, 'sampling/importance_sampling_ratio/mean': 1.0317015647888184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.876276522874832, 'clip_ratio/low_mean': 0.13203903660178185, 'clip_ratio/low_min': 0.13203903660178185, 'clip_ratio/high_mean': 0.06098760478198528, 'clip_ratio/high_max': 0.06098760478198528, 'clip_ratio/region_mean': 0.19302664138376713, 'reward_total_mean': 0.12012744694948196, 'reward_meter_mean': 0.6111171245574951, 'reward_meter_std': 0.25081607699394226, 'reward_count_adherence_mean': 0.9519230723381042, 'reward_count_adherence_std': 0.05723259598016739, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.985576868057251, 'reward_count_floor_std': 0.017169784754514694, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9864811897277832, 'reward_repeat_penalty_std': 0.017781730741262436, 'reward_repeat_floor_mean': 0.998455286026001, 'reward_repeat_floor_std': 0.0017400516662746668, 'reward_near_duplicate_penalty_mean': 0.9946328997612, 'reward_near_duplicate_penalty_std': 0.003203645348548889, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9917812347412109, 'reward_distinct_2_std': 0.01605413667857647, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.12012744694948196, 'reward_total_composite_std': 0.08017784357070923, 'epoch': 0.0}

  0%|          | 12/3300 [02:01<10:17:14, 11.26s/it]INFO 04-12 15:27:21 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 13/3300 [02:10<9:49:32, 10.76s/it] 2026-04-12 15:27:30,804 | INFO | train_grpo_train | metrics_logged mode=train step=13

                                                   
{'loss': 0.1698, 'grad_norm': 16.573932647705078, 'learning_rate': 9.963636363636364e-06, 'num_tokens': 30974.0, 'completions/mean_length': 46.75, 'completions/min_length': 30.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.5358141660690308, 'rewards/meter/std': 0.40260711312294006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8793129920959473, 'rewards/lexical_plausibility/std': 0.16006377339363098, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9961528182029724, 'rewards/repeat_penalty/std': 0.005420883186161518, 'rewards/repeat_floor/mean': 0.9994229078292847, 'rewards/repeat_floor/std': 0.0008131422218866646, 'rewards/near_duplicate_penalty/mean': 0.9961528182029724, 'rewards/near_duplicate_penalty/std': 0.005420883186161518, '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': 0.17765599489212036, 'rewards/total_composite/std': 0.14073699712753296, 'reward': 0.17765599489212036, 'reward_std': 0.14073699712753296, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23024636507034302, 'sampling/sampling_logp_difference/max': 1.6995067596435547, 'sampling/importance_sampling_ratio/min': 0.1827736496925354, 'sampling/importance_sampling_ratio/mean': 1.0083376169204712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.80851811170578, 'clip_ratio/low_mean': 0.11393210291862488, 'clip_ratio/low_min': 0.11393210291862488, 'clip_ratio/high_mean': 0.07991453260183334, 'clip_ratio/high_max': 0.07991453260183334, 'clip_ratio/region_mean': 0.19384663552045822, 'reward_total_mean': 0.17765599489212036, 'reward_meter_mean': 0.5358141660690308, 'reward_meter_std': 0.40260711312294006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8793129920959473, 'reward_lexical_plausibility_std': 0.16006377339363098, 'reward_repeat_penalty_mean': 0.9961528182029724, 'reward_repeat_penalty_std': 0.005420883186161518, 'reward_repeat_floor_mean': 0.9994229078292847, 'reward_repeat_floor_std': 0.0008131422218866646, 'reward_near_duplicate_penalty_mean': 0.9961528182029724, 'reward_near_duplicate_penalty_std': 0.005420883186161518, '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': 0.17765599489212036, 'reward_total_composite_std': 0.14073699712753296, 'epoch': 0.0}

  0%|          | 13/3300 [02:10<9:49:32, 10.76s/it]INFO 04-12 15:27:31 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 14/3300 [02:20<9:30:18, 10.41s/it]2026-04-12 15:27:40,412 | INFO | train_grpo_train | metrics_logged mode=train step=14

                                                   
{'loss': -0.0211, 'grad_norm': 16.62190818786621, 'learning_rate': 9.960606060606062e-06, 'num_tokens': 32583.0, 'completions/mean_length': 42.125, 'completions/min_length': 29.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.5649515390396118, 'rewards/meter/std': 0.45746996998786926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.5787500143051147, 'rewards/judge_quality/std': 0.22433632612228394, 'rewards/repeat_penalty/mean': 0.9921665191650391, 'rewards/repeat_penalty/std': 0.01081833429634571, 'rewards/repeat_floor/mean': 0.998824954032898, 'rewards/repeat_floor/std': 0.0016227505402639508, 'rewards/near_duplicate_penalty/mean': 0.9921665191650391, 'rewards/near_duplicate_penalty/std': 0.01081833429634571, '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': 0.3522951602935791, 'rewards/total_composite/std': 0.317210853099823, 'reward': 0.3522951602935791, 'reward_std': 0.317210853099823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19130516052246094, 'sampling/sampling_logp_difference/max': 1.6429328918457031, 'sampling/importance_sampling_ratio/min': 0.1934119462966919, 'sampling/importance_sampling_ratio/mean': 1.0410844087600708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6727047264575958, 'clip_ratio/low_mean': 0.10082429647445679, 'clip_ratio/low_min': 0.10082429647445679, 'clip_ratio/high_mean': 0.11190067790448666, 'clip_ratio/high_max': 0.11190067790448666, 'clip_ratio/region_mean': 0.21272497437894344, 'reward_total_mean': 0.3522951602935791, 'reward_meter_mean': 0.5649515390396118, 'reward_meter_std': 0.45746996998786926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9921665191650391, 'reward_repeat_penalty_std': 0.01081833429634571, 'reward_repeat_floor_mean': 0.998824954032898, 'reward_repeat_floor_std': 0.0016227505402639508, 'reward_near_duplicate_penalty_mean': 0.9921665191650391, 'reward_near_duplicate_penalty_std': 0.01081833429634571, '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': 0.5787500143051147, 'reward_judge_quality_std': 0.22433632612228394, 'reward_total_composite_mean': 0.3522951602935791, 'reward_total_composite_std': 0.317210853099823, 'epoch': 0.0}

  0%|          | 14/3300 [02:20<9:30:18, 10.41s/it]INFO 04-12 15:27:40 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 15/3300 [02:30<9:20:19, 10.23s/it]2026-04-12 15:27:50,233 | INFO | train_grpo_train | metrics_logged mode=train step=15

                                                   
{'loss': -0.123, 'grad_norm': 15.516450881958008, 'learning_rate': 9.957575757575757e-06, 'num_tokens': 34370.0, 'completions/mean_length': 53.375, 'completions/min_length': 10.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.375, 'completions/min_terminated_length': 10.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.30706876516342163, 'rewards/meter/std': 0.4386492073535919, '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': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.3091896176338196, 'rewards/repeat_penalty/mean': 0.9929812550544739, 'rewards/repeat_penalty/std': 0.019852008670568466, 'rewards/repeat_floor/mean': 0.9989472031593323, 'rewards/repeat_floor/std': 0.002977801486849785, 'rewards/near_duplicate_penalty/mean': 0.9929812550544739, 'rewards/near_duplicate_penalty/std': 0.019852008670568466, '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': 0.22470685839653015, 'rewards/total_composite/std': 0.3384774923324585, 'reward': 0.22470685839653015, 'reward_std': 0.3384774923324585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2112843096256256, 'sampling/sampling_logp_difference/max': 2.0283632278442383, 'sampling/importance_sampling_ratio/min': 0.13155066967010498, 'sampling/importance_sampling_ratio/mean': 1.0471327304840088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.098313555121422, 'clip_ratio/low_mean': 0.1331961303949356, 'clip_ratio/low_min': 0.1331961303949356, 'clip_ratio/high_mean': 0.06151947192847729, 'clip_ratio/high_max': 0.06151947192847729, 'clip_ratio/region_mean': 0.1947156023234129, 'reward_total_mean': 0.22470685839653015, 'reward_meter_mean': 0.30706876516342163, 'reward_meter_std': 0.4386492073535919, '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': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.9929812550544739, 'reward_repeat_penalty_std': 0.019852008670568466, 'reward_repeat_floor_mean': 0.9989472031593323, 'reward_repeat_floor_std': 0.002977801486849785, 'reward_near_duplicate_penalty_mean': 0.9929812550544739, 'reward_near_duplicate_penalty_std': 0.019852008670568466, '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': 0.22470685839653015, 'reward_total_composite_std': 0.3384774923324585, 'epoch': 0.0}

  0%|          | 15/3300 [02:30<9:20:19, 10.23s/it]INFO 04-12 15:27:50 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 16/3300 [02:46<10:56:50, 12.00s/it]2026-04-12 15:28:06,334 | INFO | train_grpo_train | metrics_logged mode=train step=16

                                                    
{'loss': -0.1233, 'grad_norm': 4.472887992858887, 'learning_rate': 9.954545454545456e-06, 'num_tokens': 38592.0, 'completions/mean_length': 380.75, 'completions/min_length': 285.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 362.0000305175781, 'completions/min_terminated_length': 285.0, 'completions/max_terminated_length': 489.0, 'rewards/meter/mean': 0.646077036857605, 'rewards/meter/std': 0.22905045747756958, 'rewards/count_adherence/mean': 0.8958333730697632, 'rewards/count_adherence/std': 0.058925580233335495, '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': 0.96875, 'rewards/count_floor/std': 0.017677679657936096, 'rewards/lexical_plausibility/mean': 0.9700763821601868, 'rewards/lexical_plausibility/std': 0.08350753784179688, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9905528426170349, 'rewards/repeat_penalty/std': 0.011145573109388351, 'rewards/repeat_floor/mean': 0.9989702701568604, 'rewards/repeat_floor/std': 0.001023361925035715, 'rewards/near_duplicate_penalty/mean': 0.9970415830612183, 'rewards/near_duplicate_penalty/std': 0.0005189756629988551, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9934895634651184, 'rewards/distinct_2/std': 0.01092198584228754, 'rewards/total_composite/mean': 0.20287978649139404, 'rewards/total_composite/std': 0.10549522191286087, 'reward': 0.20287978649139404, 'reward_std': 0.10549522191286087, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22138601541519165, 'sampling/sampling_logp_difference/max': 1.7725324630737305, 'sampling/importance_sampling_ratio/min': 0.16990217566490173, 'sampling/importance_sampling_ratio/mean': 1.0467000007629395, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.758598208427429, 'clip_ratio/low_mean': 0.04998158477246761, 'clip_ratio/low_min': 0.04998158477246761, 'clip_ratio/high_mean': 0.13737528584897518, 'clip_ratio/high_max': 0.13737528584897518, 'clip_ratio/region_mean': 0.1873568706214428, 'reward_total_mean': 0.20287978649139404, 'reward_meter_mean': 0.646077036857605, 'reward_meter_std': 0.22905045747756958, 'reward_count_adherence_mean': 0.8958333730697632, 'reward_count_adherence_std': 0.058925580233335495, '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': 0.96875, 'reward_count_floor_std': 0.017677679657936096, 'reward_lexical_plausibility_mean': 0.9700763821601868, 'reward_lexical_plausibility_std': 0.08350753784179688, 'reward_repeat_penalty_mean': 0.9905528426170349, 'reward_repeat_penalty_std': 0.011145573109388351, 'reward_repeat_floor_mean': 0.9989702701568604, 'reward_repeat_floor_std': 0.001023361925035715, 'reward_near_duplicate_penalty_mean': 0.9970415830612183, 'reward_near_duplicate_penalty_std': 0.0005189756629988551, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9934895634651184, 'reward_distinct_2_std': 0.01092198584228754, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.20287978649139404, 'reward_total_composite_std': 0.10549522191286087, 'epoch': 0.0}

  0%|          | 16/3300 [02:46<10:56:50, 12.00s/it]INFO 04-12 15:28:06 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 17/3300 [02:55<10:14:12, 11.23s/it]2026-04-12 15:28:15,756 | INFO | train_grpo_train | metrics_logged mode=train step=17

                                                    
{'loss': 0.1407, 'grad_norm': 14.15239143371582, 'learning_rate': 9.951515151515152e-06, 'num_tokens': 40533.0, 'completions/mean_length': 65.625, 'completions/min_length': 47.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.625, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.4984070062637329, 'rewards/meter/std': 0.42103222012519836, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9718775153160095, 'rewards/lexical_plausibility/std': 0.05715625733137131, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9909399747848511, 'rewards/repeat_penalty/std': 0.017375122755765915, 'rewards/repeat_floor/mean': 0.9986410140991211, 'rewards/repeat_floor/std': 0.0026062640827149153, 'rewards/near_duplicate_penalty/mean': 0.9909399747848511, 'rewards/near_duplicate_penalty/std': 0.017375122755765915, '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': 0.1791306436061859, 'rewards/total_composite/std': 0.17350420355796814, 'reward': 0.1791306436061859, 'reward_std': 0.17350421845912933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20403529703617096, 'sampling/sampling_logp_difference/max': 1.0976190567016602, 'sampling/importance_sampling_ratio/min': 0.33366456627845764, 'sampling/importance_sampling_ratio/mean': 1.026328444480896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2219748347997665, 'clip_ratio/low_mean': 0.1221171673387289, 'clip_ratio/low_min': 0.1221171673387289, 'clip_ratio/high_mean': 0.07865174673497677, 'clip_ratio/high_max': 0.07865174673497677, 'clip_ratio/region_mean': 0.20076891407370567, 'reward_total_mean': 0.1791306436061859, 'reward_meter_mean': 0.4984070062637329, 'reward_meter_std': 0.42103222012519836, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9718775153160095, 'reward_lexical_plausibility_std': 0.05715625733137131, 'reward_repeat_penalty_mean': 0.9909399747848511, 'reward_repeat_penalty_std': 0.017375122755765915, 'reward_repeat_floor_mean': 0.9986410140991211, 'reward_repeat_floor_std': 0.0026062640827149153, 'reward_near_duplicate_penalty_mean': 0.9909399747848511, 'reward_near_duplicate_penalty_std': 0.017375122755765915, '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': 0.1791306436061859, 'reward_total_composite_std': 0.17350420355796814, 'epoch': 0.0}

  1%|          | 17/3300 [02:55<10:14:12, 11.23s/it]INFO 04-12 15:28:16 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 18/3300 [03:06<10:10:36, 11.16s/it]2026-04-12 15:28:26,774 | INFO | train_grpo_train | metrics_logged mode=train step=18

                                                    
{'loss': 0.0279, 'grad_norm': 8.760952949523926, 'learning_rate': 9.948484848484849e-06, 'num_tokens': 43125.0, 'completions/mean_length': 125.0, 'completions/min_length': 82.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.0, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.3913823366165161, 'rewards/meter/std': 0.29120635986328125, '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': 0.9837819933891296, 'rewards/lexical_plausibility/std': 0.040401432663202286, 'rewards/judge_quality/mean': 0.3837500214576721, 'rewards/judge_quality/std': 0.2579555809497833, 'rewards/repeat_penalty/mean': 0.9826048612594604, 'rewards/repeat_penalty/std': 0.018295876681804657, 'rewards/repeat_floor/mean': 0.9979451894760132, 'rewards/repeat_floor/std': 0.001798599842004478, 'rewards/near_duplicate_penalty/mean': 0.9919602870941162, 'rewards/near_duplicate_penalty/std': 0.005748011637479067, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9905686378479004, 'rewards/distinct_2/std': 0.01752571202814579, 'rewards/total_composite/mean': 0.14996148645877838, 'rewards/total_composite/std': 0.1391526311635971, 'reward': 0.14996148645877838, 'reward_std': 0.1391526311635971, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21279513835906982, 'sampling/sampling_logp_difference/max': 1.3045446872711182, 'sampling/importance_sampling_ratio/min': 0.29111722111701965, 'sampling/importance_sampling_ratio/mean': 1.0226211547851562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.263792857527733, 'clip_ratio/low_mean': 0.12045164220035076, 'clip_ratio/low_min': 0.12045164220035076, 'clip_ratio/high_mean': 0.06518849171698093, 'clip_ratio/high_max': 0.06518849171698093, 'clip_ratio/region_mean': 0.1856401339173317, 'reward_total_mean': 0.14996148645877838, 'reward_meter_mean': 0.3913823366165161, 'reward_meter_std': 0.29120635986328125, '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': 0.9837819933891296, 'reward_lexical_plausibility_std': 0.040401432663202286, 'reward_repeat_penalty_mean': 0.9826048612594604, 'reward_repeat_penalty_std': 0.018295876681804657, 'reward_repeat_floor_mean': 0.9979451894760132, 'reward_repeat_floor_std': 0.001798599842004478, 'reward_near_duplicate_penalty_mean': 0.9919602870941162, 'reward_near_duplicate_penalty_std': 0.005748011637479067, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9905686378479004, 'reward_distinct_2_std': 0.01752571202814579, 'reward_judge_quality_mean': 0.3837500214576721, 'reward_judge_quality_std': 0.2579555809497833, 'reward_total_composite_mean': 0.14996148645877838, 'reward_total_composite_std': 0.1391526311635971, 'epoch': 0.0}

  1%|          | 18/3300 [03:06<10:10:36, 11.16s/it]INFO 04-12 15:28:27 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 19/3300 [03:22<11:21:03, 12.45s/it]2026-04-12 15:28:42,237 | INFO | train_grpo_train | metrics_logged mode=train step=19

                                                    
{'loss': -0.0062, 'grad_norm': 4.123045444488525, 'learning_rate': 9.945454545454546e-06, 'num_tokens': 46986.0, 'completions/mean_length': 463.625, 'completions/min_length': 351.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 434.6000061035156, 'completions/min_terminated_length': 351.0, 'completions/max_terminated_length': 509.0, 'rewards/meter/mean': 0.7735102772712708, 'rewards/meter/std': 0.18511484563350677, 'rewards/count_adherence/mean': 0.953125, 'rewards/count_adherence/std': 0.05540033057332039, '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': 0.9859374761581421, 'rewards/count_floor/std': 0.016620110720396042, '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': 0.9906105995178223, 'rewards/repeat_penalty/std': 0.006933511700481176, 'rewards/repeat_floor/mean': 0.9989017248153687, 'rewards/repeat_floor/std': 0.0006824919255450368, 'rewards/near_duplicate_penalty/mean': 0.9958175420761108, 'rewards/near_duplicate_penalty/std': 0.001465248642489314, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9947676658630371, 'rewards/distinct_2/std': 0.006208095233887434, 'rewards/total_composite/mean': 0.13303399085998535, 'rewards/total_composite/std': 0.05913260579109192, 'reward': 0.13303399085998535, 'reward_std': 0.05913260951638222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22717344760894775, 'sampling/sampling_logp_difference/max': 1.536581039428711, 'sampling/importance_sampling_ratio/min': 0.21511532366275787, 'sampling/importance_sampling_ratio/mean': 1.0486820936203003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.131060630083084, 'clip_ratio/low_mean': 0.08844146691262722, 'clip_ratio/low_min': 0.08844146691262722, 'clip_ratio/high_mean': 0.028282828629016876, 'clip_ratio/high_max': 0.028282828629016876, 'clip_ratio/region_mean': 0.1167242955416441, 'reward_total_mean': 0.13303399085998535, 'reward_meter_mean': 0.7735102772712708, 'reward_meter_std': 0.18511484563350677, 'reward_count_adherence_mean': 0.953125, 'reward_count_adherence_std': 0.05540033057332039, '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': 0.9859374761581421, 'reward_count_floor_std': 0.016620110720396042, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9906105995178223, 'reward_repeat_penalty_std': 0.006933511700481176, 'reward_repeat_floor_mean': 0.9989017248153687, 'reward_repeat_floor_std': 0.0006824919255450368, 'reward_near_duplicate_penalty_mean': 0.9958175420761108, 'reward_near_duplicate_penalty_std': 0.001465248642489314, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9947676658630371, 'reward_distinct_2_std': 0.006208095233887434, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.13303399085998535, 'reward_total_composite_std': 0.05913260579109192, 'epoch': 0.0}

  1%|          | 19/3300 [03:22<11:21:03, 12.45s/it]INFO 04-12 15:28:42 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 20/3300 [03:33<10:52:30, 11.94s/it]2026-04-12 15:28:52,967 | INFO | train_grpo_train | metrics_logged mode=train step=20

                                                    
{'loss': 0.0119, 'grad_norm': 13.167397499084473, 'learning_rate': 9.942424242424244e-06, 'num_tokens': 48850.0, 'completions/mean_length': 79.0, 'completions/min_length': 52.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7391309142112732, 'rewards/meter/std': 0.3295347988605499, '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': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9912289381027222, 'rewards/repeat_penalty/std': 0.017576977610588074, 'rewards/repeat_floor/mean': 0.9990683794021606, 'rewards/repeat_floor/std': 0.001569906948134303, 'rewards/near_duplicate_penalty/mean': 0.9976356029510498, 'rewards/near_duplicate_penalty/std': 0.002671419410035014, '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': 0.25470268726348877, 'rewards/total_composite/std': 0.13675996661186218, 'reward': 0.25470268726348877, 'reward_std': 0.13675996661186218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1914476603269577, 'sampling/sampling_logp_difference/max': 1.754624366760254, 'sampling/importance_sampling_ratio/min': 0.1729722023010254, 'sampling/importance_sampling_ratio/mean': 1.0130658149719238, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.942652314901352, 'clip_ratio/low_mean': 0.0659028347581625, 'clip_ratio/low_min': 0.0659028347581625, 'clip_ratio/high_mean': 0.12095401994884014, 'clip_ratio/high_max': 0.12095401994884014, 'clip_ratio/region_mean': 0.18685685470700264, 'reward_total_mean': 0.25470268726348877, 'reward_meter_mean': 0.7391309142112732, 'reward_meter_std': 0.3295347988605499, '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': 0.9912289381027222, 'reward_repeat_penalty_std': 0.017576977610588074, 'reward_repeat_floor_mean': 0.9990683794021606, 'reward_repeat_floor_std': 0.001569906948134303, 'reward_near_duplicate_penalty_mean': 0.9976356029510498, 'reward_near_duplicate_penalty_std': 0.002671419410035014, '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': 0.36249998211860657, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.25470268726348877, 'reward_total_composite_std': 0.13675996661186218, 'epoch': 0.0}

  1%|          | 20/3300 [03:33<10:52:30, 11.94s/it]INFO 04-12 15:28:53 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 21/3300 [03:43<10:31:33, 11.56s/it]2026-04-12 15:29:03,637 | INFO | train_grpo_train | metrics_logged mode=train step=21

                                                    
{'loss': 0.1395, 'grad_norm': 18.839717864990234, 'learning_rate': 9.939393939393939e-06, 'num_tokens': 50629.0, 'completions/mean_length': 54.375, 'completions/min_length': 42.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.16253086924552917, 'rewards/meter/std': 0.32640475034713745, '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': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.4962500035762787, 'rewards/judge_quality/std': 0.17476005852222443, 'rewards/repeat_penalty/mean': 0.9408199787139893, 'rewards/repeat_penalty/std': 0.10722612589597702, 'rewards/repeat_floor/mean': 0.9939693212509155, 'rewards/repeat_floor/std': 0.008080698549747467, 'rewards/near_duplicate_penalty/mean': 0.9694105386734009, 'rewards/near_duplicate_penalty/std': 0.027310004457831383, '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': 0.12064145505428314, 'rewards/total_composite/std': 0.28272712230682373, 'reward': 0.12064145505428314, 'reward_std': 0.28272712230682373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.31002137064933777, 'sampling/sampling_logp_difference/max': 1.9811716079711914, 'sampling/importance_sampling_ratio/min': 0.13790756464004517, 'sampling/importance_sampling_ratio/mean': 1.0304527282714844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9839466661214828, 'clip_ratio/low_mean': 0.21813260950148106, 'clip_ratio/low_min': 0.21813260950148106, 'clip_ratio/high_mean': 0.02083333395421505, 'clip_ratio/high_max': 0.02083333395421505, 'clip_ratio/region_mean': 0.2389659434556961, 'reward_total_mean': 0.12064145505428314, 'reward_meter_mean': 0.16253086924552917, 'reward_meter_std': 0.32640475034713745, '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': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9408199787139893, 'reward_repeat_penalty_std': 0.10722612589597702, 'reward_repeat_floor_mean': 0.9939693212509155, 'reward_repeat_floor_std': 0.008080698549747467, 'reward_near_duplicate_penalty_mean': 0.9694105386734009, 'reward_near_duplicate_penalty_std': 0.027310004457831383, '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': 0.4962500035762787, 'reward_judge_quality_std': 0.17476005852222443, 'reward_total_composite_mean': 0.12064145505428314, 'reward_total_composite_std': 0.28272712230682373, 'epoch': 0.0}

  1%|          | 21/3300 [03:43<10:31:33, 11.56s/it]INFO 04-12 15:29:03 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 22/3300 [03:53<10:06:58, 11.11s/it]2026-04-12 15:29:13,704 | INFO | train_grpo_train | metrics_logged mode=train step=22

                                                    
{'loss': 0.1318, 'grad_norm': 18.062366485595703, 'learning_rate': 9.936363636363638e-06, 'num_tokens': 52412.0, 'completions/mean_length': 39.875, 'completions/min_length': 32.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7118128538131714, 'rewards/meter/std': 0.30599501729011536, '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': 0.6012499928474426, 'rewards/judge_quality/std': 0.2671777307987213, 'rewards/repeat_penalty/mean': 0.9786217212677002, 'rewards/repeat_penalty/std': 0.03314605727791786, 'rewards/repeat_floor/mean': 0.9973337650299072, 'rewards/repeat_floor/std': 0.0035276152193546295, 'rewards/near_duplicate_penalty/mean': 0.9879943132400513, 'rewards/near_duplicate_penalty/std': 0.010588407516479492, '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': 0.4631122946739197, 'rewards/total_composite/std': 0.3377132713794708, 'reward': 0.4631122946739197, 'reward_std': 0.3377132713794708, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18866823613643646, 'sampling/sampling_logp_difference/max': 1.7111029624938965, 'sampling/importance_sampling_ratio/min': 0.18066641688346863, 'sampling/importance_sampling_ratio/mean': 1.0093622207641602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3397293128073215, 'clip_ratio/low_mean': 0.09336858335882425, 'clip_ratio/low_min': 0.09336858335882425, 'clip_ratio/high_mean': 0.02302631549537182, 'clip_ratio/high_max': 0.02302631549537182, 'clip_ratio/region_mean': 0.11639489885419607, 'reward_total_mean': 0.4631122946739197, 'reward_meter_mean': 0.7118128538131714, 'reward_meter_std': 0.30599501729011536, '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': 0.9786217212677002, 'reward_repeat_penalty_std': 0.03314605727791786, 'reward_repeat_floor_mean': 0.9973337650299072, 'reward_repeat_floor_std': 0.0035276152193546295, 'reward_near_duplicate_penalty_mean': 0.9879943132400513, 'reward_near_duplicate_penalty_std': 0.010588407516479492, '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': 0.6012499928474426, 'reward_judge_quality_std': 0.2671777307987213, 'reward_total_composite_mean': 0.4631122946739197, 'reward_total_composite_std': 0.3377132713794708, 'epoch': 0.0}

  1%|          | 22/3300 [03:53<10:06:58, 11.11s/it]INFO 04-12 15:29:13 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 23/3300 [04:10<11:32:09, 12.67s/it]2026-04-12 15:29:30,023 | INFO | train_grpo_train | metrics_logged mode=train step=23

                                                    
{'loss': 0.1579, 'grad_norm': 1.7903239727020264, 'learning_rate': 9.933333333333334e-06, 'num_tokens': 55561.0, 'completions/mean_length': 509.625, 'completions/min_length': 500.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.625, 'completions/mean_terminated_length': 505.66668701171875, 'completions/min_terminated_length': 500.0, 'completions/max_terminated_length': 511.0, 'rewards/meter/mean': 0.7652940154075623, 'rewards/meter/std': 0.17092706263065338, 'rewards/count_adherence/mean': 0.9910714626312256, 'rewards/count_adherence/std': 0.025253823027014732, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9973214268684387, 'rewards/count_floor/std': 0.007576148957014084, 'rewards/lexical_plausibility/mean': 0.9985464811325073, 'rewards/lexical_plausibility/std': 0.004111087881028652, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/repeat_penalty/mean': 0.9815077185630798, 'rewards/repeat_penalty/std': 0.03388342633843422, 'rewards/repeat_floor/mean': 0.9980646371841431, 'rewards/repeat_floor/std': 0.003204014850780368, 'rewards/near_duplicate_penalty/mean': 0.9956598281860352, 'rewards/near_duplicate_penalty/std': 0.002639841055497527, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9857295751571655, 'rewards/distinct_2/std': 0.03214246407151222, 'rewards/total_composite/mean': 0.19023799896240234, 'rewards/total_composite/std': 0.08524578809738159, 'reward': 0.19023799896240234, 'reward_std': 0.08524578809738159, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17710649967193604, 'sampling/sampling_logp_difference/max': 1.5074372291564941, 'sampling/importance_sampling_ratio/min': 0.22147683799266815, 'sampling/importance_sampling_ratio/mean': 1.0306750535964966, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.85990309715271, 'clip_ratio/low_mean': 0.04566798359155655, 'clip_ratio/low_min': 0.04566798359155655, 'clip_ratio/high_mean': 0.020792564377188683, 'clip_ratio/high_max': 0.020792564377188683, 'clip_ratio/region_mean': 0.06646054796874523, 'reward_total_mean': 0.19023799896240234, 'reward_meter_mean': 0.7652940154075623, 'reward_meter_std': 0.17092706263065338, 'reward_count_adherence_mean': 0.9910714626312256, 'reward_count_adherence_std': 0.025253823027014732, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9973214268684387, 'reward_count_floor_std': 0.007576148957014084, 'reward_lexical_plausibility_mean': 0.9985464811325073, 'reward_lexical_plausibility_std': 0.004111087881028652, 'reward_repeat_penalty_mean': 0.9815077185630798, 'reward_repeat_penalty_std': 0.03388342633843422, 'reward_repeat_floor_mean': 0.9980646371841431, 'reward_repeat_floor_std': 0.003204014850780368, 'reward_near_duplicate_penalty_mean': 0.9956598281860352, 'reward_near_duplicate_penalty_std': 0.002639841055497527, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9857295751571655, 'reward_distinct_2_std': 0.03214246407151222, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.19023799896240234, 'reward_total_composite_std': 0.08524578809738159, 'epoch': 0.0}

  1%|          | 23/3300 [04:10<11:32:09, 12.67s/it]INFO 04-12 15:29:30 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 24/3300 [04:22<11:23:04, 12.51s/it]2026-04-12 15:29:42,156 | INFO | train_grpo_train | metrics_logged mode=train step=24

                                                    
{'loss': 0.171, 'grad_norm': 15.77658462524414, 'learning_rate': 9.930303030303031e-06, 'num_tokens': 57564.0, 'completions/mean_length': 83.375, 'completions/min_length': 61.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.375, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.7255039215087891, 'rewards/meter/std': 0.34764692187309265, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.9726814031600952, 'rewards/repeat_penalty/std': 0.029803479090332985, 'rewards/repeat_floor/mean': 0.996475338935852, 'rewards/repeat_floor/std': 0.003498756093904376, 'rewards/near_duplicate_penalty/mean': 0.9827389121055603, 'rewards/near_duplicate_penalty/std': 0.013923194259405136, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9896050095558167, 'rewards/distinct_2/std': 0.019247809424996376, 'rewards/total_composite/mean': 0.293908953666687, 'rewards/total_composite/std': 0.14653658866882324, 'reward': 0.293908953666687, 'reward_std': 0.14653658866882324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2087927609682083, 'sampling/sampling_logp_difference/max': 1.8262858390808105, 'sampling/importance_sampling_ratio/min': 0.1610104739665985, 'sampling/importance_sampling_ratio/mean': 1.0376325845718384, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.086565062403679, 'clip_ratio/low_mean': 0.07746557146310806, 'clip_ratio/low_min': 0.07746557146310806, 'clip_ratio/high_mean': 0.12272364646196365, 'clip_ratio/high_max': 0.12272364646196365, 'clip_ratio/region_mean': 0.20018921792507172, 'reward_total_mean': 0.293908953666687, 'reward_meter_mean': 0.7255039215087891, 'reward_meter_std': 0.34764692187309265, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9726814031600952, 'reward_repeat_penalty_std': 0.029803479090332985, 'reward_repeat_floor_mean': 0.996475338935852, 'reward_repeat_floor_std': 0.003498756093904376, 'reward_near_duplicate_penalty_mean': 0.9827389121055603, 'reward_near_duplicate_penalty_std': 0.013923194259405136, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9896050095558167, 'reward_distinct_2_std': 0.019247809424996376, 'reward_judge_quality_mean': 0.41249996423721313, 'reward_judge_quality_std': 0.0517549142241478, 'reward_total_composite_mean': 0.293908953666687, 'reward_total_composite_std': 0.14653658866882324, 'epoch': 0.0}

  1%|          | 24/3300 [04:22<11:23:04, 12.51s/it]INFO 04-12 15:29:42 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 25/3300 [04:36<11:56:02, 13.12s/it]2026-04-12 15:29:56,689 | INFO | train_grpo_train | metrics_logged mode=train step=25

                                                    
{'loss': 0.0352, 'grad_norm': 14.53252124786377, 'learning_rate': 9.927272727272728e-06, 'num_tokens': 59742.0, 'completions/mean_length': 87.25, 'completions/min_length': 76.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.25, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.6735032796859741, 'rewards/meter/std': 0.3561065196990967, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.958823561668396, 'rewards/lexical_plausibility/std': 0.11646464467048645, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.05345224589109421, 'rewards/repeat_penalty/mean': 0.9920245409011841, 'rewards/repeat_penalty/std': 0.00808724295347929, 'rewards/repeat_floor/mean': 0.9988037347793579, 'rewards/repeat_floor/std': 0.0012130799004808068, 'rewards/near_duplicate_penalty/mean': 0.9920245409011841, 'rewards/near_duplicate_penalty/std': 0.00808724295347929, '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': 0.26689326763153076, 'rewards/total_composite/std': 0.1506560742855072, 'reward': 0.26689326763153076, 'reward_std': 0.1506560742855072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1992534101009369, 'sampling/sampling_logp_difference/max': 1.8375164270401, 'sampling/importance_sampling_ratio/min': 0.15921235084533691, 'sampling/importance_sampling_ratio/mean': 1.0229802131652832, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6130416989326477, 'clip_ratio/low_mean': 0.09780654683709145, 'clip_ratio/low_min': 0.09780654683709145, 'clip_ratio/high_mean': 0.10350676998496056, 'clip_ratio/high_max': 0.10350676998496056, 'clip_ratio/region_mean': 0.201313316822052, 'reward_total_mean': 0.26689326763153076, 'reward_meter_mean': 0.6735032796859741, 'reward_meter_std': 0.3561065196990967, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.958823561668396, 'reward_lexical_plausibility_std': 0.11646464467048645, 'reward_repeat_penalty_mean': 0.9920245409011841, 'reward_repeat_penalty_std': 0.00808724295347929, 'reward_repeat_floor_mean': 0.9988037347793579, 'reward_repeat_floor_std': 0.0012130799004808068, 'reward_near_duplicate_penalty_mean': 0.9920245409011841, 'reward_near_duplicate_penalty_std': 0.00808724295347929, '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': 0.26689326763153076, 'reward_total_composite_std': 0.1506560742855072, 'epoch': 0.0}

  1%|          | 25/3300 [04:36<11:56:02, 13.12s/it]INFO 04-12 15:29:56 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 26/3300 [04:47<11:11:41, 12.31s/it]2026-04-12 15:30:07,112 | INFO | train_grpo_train | metrics_logged mode=train step=26

                                                    
{'loss': 0.055, 'grad_norm': 12.390909194946289, 'learning_rate': 9.924242424242425e-06, 'num_tokens': 61561.0, 'completions/mean_length': 68.375, 'completions/min_length': 44.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.375, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.3102201819419861, 'rewards/meter/std': 0.34473279118537903, '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': 0.981586754322052, 'rewards/lexical_plausibility/std': 0.05208052322268486, 'rewards/judge_quality/mean': 0.5174999833106995, 'rewards/judge_quality/std': 0.2522329092025757, 'rewards/repeat_penalty/mean': 0.9818671941757202, 'rewards/repeat_penalty/std': 0.015078977681696415, 'rewards/repeat_floor/mean': 0.9972801208496094, 'rewards/repeat_floor/std': 0.002261844929307699, 'rewards/near_duplicate_penalty/mean': 0.9818671941757202, 'rewards/near_duplicate_penalty/std': 0.015078977681696415, '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': 0.16526462137699127, 'rewards/total_composite/std': 0.21345311403274536, 'reward': 0.16526462137699127, 'reward_std': 0.21345311403274536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20714733004570007, 'sampling/sampling_logp_difference/max': 1.583723545074463, 'sampling/importance_sampling_ratio/min': 0.20520955324172974, 'sampling/importance_sampling_ratio/mean': 1.0312914848327637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8172488510608673, 'clip_ratio/low_mean': 0.1448319312185049, 'clip_ratio/low_min': 0.1448319312185049, 'clip_ratio/high_mean': 0.04946471005678177, 'clip_ratio/high_max': 0.04946471005678177, 'clip_ratio/region_mean': 0.19429664127528667, 'reward_total_mean': 0.16526462137699127, 'reward_meter_mean': 0.3102201819419861, 'reward_meter_std': 0.34473279118537903, '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': 0.981586754322052, 'reward_lexical_plausibility_std': 0.05208052322268486, 'reward_repeat_penalty_mean': 0.9818671941757202, 'reward_repeat_penalty_std': 0.015078977681696415, 'reward_repeat_floor_mean': 0.9972801208496094, 'reward_repeat_floor_std': 0.002261844929307699, 'reward_near_duplicate_penalty_mean': 0.9818671941757202, 'reward_near_duplicate_penalty_std': 0.015078977681696415, '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': 0.16526462137699127, 'reward_total_composite_std': 0.21345311403274536, 'epoch': 0.0}

  1%|          | 26/3300 [04:47<11:11:41, 12.31s/it]INFO 04-12 15:30:07 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 27/3300 [05:03<12:10:16, 13.39s/it]2026-04-12 15:30:23,014 | INFO | train_grpo_train | metrics_logged mode=train step=27

                                                    
{'loss': 0.1186, 'grad_norm': 17.233257293701172, 'learning_rate': 9.921212121212121e-06, 'num_tokens': 63455.0, 'completions/mean_length': 61.75, 'completions/min_length': 48.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.6537008285522461, 'rewards/meter/std': 0.4205484986305237, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9908088445663452, 'rewards/lexical_plausibility/std': 0.025996576994657516, 'rewards/judge_quality/mean': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9815423488616943, 'rewards/repeat_penalty/std': 0.02558419480919838, 'rewards/repeat_floor/mean': 0.9976505041122437, 'rewards/repeat_floor/std': 0.0026978568639606237, 'rewards/near_duplicate_penalty/mean': 0.9887744188308716, 'rewards/near_duplicate_penalty/std': 0.007379098795354366, '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': 0.23420360684394836, 'rewards/total_composite/std': 0.16933497786521912, 'reward': 0.23420360684394836, 'reward_std': 0.16933497786521912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23756131529808044, 'sampling/sampling_logp_difference/max': 2.162611961364746, 'sampling/importance_sampling_ratio/min': 0.11502428352832794, 'sampling/importance_sampling_ratio/mean': 1.0507766008377075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0019322484731674, 'clip_ratio/low_mean': 0.09170935861766338, 'clip_ratio/low_min': 0.09170935861766338, 'clip_ratio/high_mean': 0.15436112694442272, 'clip_ratio/high_max': 0.15436112694442272, 'clip_ratio/region_mean': 0.2460704855620861, 'reward_total_mean': 0.23420360684394836, 'reward_meter_mean': 0.6537008285522461, 'reward_meter_std': 0.4205484986305237, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9908088445663452, 'reward_lexical_plausibility_std': 0.025996576994657516, 'reward_repeat_penalty_mean': 0.9815423488616943, 'reward_repeat_penalty_std': 0.02558419480919838, 'reward_repeat_floor_mean': 0.9976505041122437, 'reward_repeat_floor_std': 0.0026978568639606237, 'reward_near_duplicate_penalty_mean': 0.9887744188308716, 'reward_near_duplicate_penalty_std': 0.007379098795354366, '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': 0.3125, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.23420360684394836, 'reward_total_composite_std': 0.16933497786521912, 'epoch': 0.0}

  1%|          | 27/3300 [05:03<12:10:16, 13.39s/it]INFO 04-12 15:30:23 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 28/3300 [05:13<11:19:26, 12.46s/it]2026-04-12 15:30:33,308 | INFO | train_grpo_train | metrics_logged mode=train step=28

                                                    
{'loss': 0.0357, 'grad_norm': 11.514175415039062, 'learning_rate': 9.918181818181818e-06, 'num_tokens': 65139.0, 'completions/mean_length': 63.5, 'completions/min_length': 55.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.962704062461853, 'rewards/meter/std': 0.06590832024812698, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.998999297618866, 'rewards/repeat_penalty/std': 0.002014234196394682, 'rewards/repeat_floor/mean': 0.9998499155044556, 'rewards/repeat_floor/std': 0.00030213952413760126, 'rewards/near_duplicate_penalty/mean': 0.998999297618866, 'rewards/near_duplicate_penalty/std': 0.002014234196394682, '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': 0.31323373317718506, 'rewards/total_composite/std': 0.11740633845329285, 'reward': 0.31323373317718506, 'reward_std': 0.11740633100271225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2057531177997589, 'sampling/sampling_logp_difference/max': 1.7660636901855469, 'sampling/importance_sampling_ratio/min': 0.17100480198860168, 'sampling/importance_sampling_ratio/mean': 1.002128005027771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8552557826042175, 'clip_ratio/low_mean': 0.05429360922425985, 'clip_ratio/low_min': 0.05429360922425985, 'clip_ratio/high_mean': 0.13596890307962894, 'clip_ratio/high_max': 0.13596890307962894, 'clip_ratio/region_mean': 0.1902625123038888, 'reward_total_mean': 0.31323373317718506, 'reward_meter_mean': 0.962704062461853, 'reward_meter_std': 0.06590832024812698, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.998999297618866, 'reward_repeat_penalty_std': 0.002014234196394682, 'reward_repeat_floor_mean': 0.9998499155044556, 'reward_repeat_floor_std': 0.00030213952413760126, 'reward_near_duplicate_penalty_mean': 0.998999297618866, 'reward_near_duplicate_penalty_std': 0.002014234196394682, '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': 0.31323373317718506, 'reward_total_composite_std': 0.11740633845329285, 'epoch': 0.0}

  1%|          | 28/3300 [05:13<11:19:26, 12.46s/it]INFO 04-12 15:30:33 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 29/3300 [05:26<11:29:02, 12.64s/it]2026-04-12 15:30:46,367 | INFO | train_grpo_train | metrics_logged mode=train step=29

                                                    
{'loss': 0.1162, 'grad_norm': 9.114456176757812, 'learning_rate': 9.915151515151515e-06, 'num_tokens': 68295.0, 'completions/mean_length': 157.5, 'completions/min_length': 117.0, 'completions/max_length': 212.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.5, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.13019584119319916, 'rewards/meter/std': 0.09870223701000214, 'rewards/count_adherence/mean': 0.984375, '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': 0.995312511920929, 'rewards/count_floor/std': 0.013258260674774647, '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': 0.9913126230239868, 'rewards/repeat_penalty/std': 0.007308544125407934, 'rewards/repeat_floor/mean': 0.99884432554245, 'rewards/repeat_floor/std': 0.0006962743354961276, 'rewards/near_duplicate_penalty/mean': 0.9937939643859863, 'rewards/near_duplicate_penalty/std': 0.0016443743370473385, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9975025057792664, 'rewards/distinct_2/std': 0.0070640011690557, 'rewards/total_composite/mean': 0.0525735467672348, 'rewards/total_composite/std': 0.04172529652714729, 'reward': 0.0525735467672348, 'reward_std': 0.04172529652714729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23808272182941437, 'sampling/sampling_logp_difference/max': 1.764352798461914, 'sampling/importance_sampling_ratio/min': 0.17582903802394867, 'sampling/importance_sampling_ratio/mean': 1.0574243068695068, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5588284730911255, 'clip_ratio/low_mean': 0.1294011827558279, 'clip_ratio/low_min': 0.1294011827558279, 'clip_ratio/high_mean': 0.11166917532682419, 'clip_ratio/high_max': 0.11166917532682419, 'clip_ratio/region_mean': 0.2410703580826521, 'reward_total_mean': 0.0525735467672348, 'reward_meter_mean': 0.13019584119319916, 'reward_meter_std': 0.09870223701000214, 'reward_count_adherence_mean': 0.984375, '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': 0.995312511920929, 'reward_count_floor_std': 0.013258260674774647, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9913126230239868, 'reward_repeat_penalty_std': 0.007308544125407934, 'reward_repeat_floor_mean': 0.99884432554245, 'reward_repeat_floor_std': 0.0006962743354961276, 'reward_near_duplicate_penalty_mean': 0.9937939643859863, 'reward_near_duplicate_penalty_std': 0.0016443743370473385, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9975025057792664, 'reward_distinct_2_std': 0.0070640011690557, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.07440237700939178, 'reward_total_composite_mean': 0.0525735467672348, 'reward_total_composite_std': 0.04172529652714729, 'epoch': 0.0}

  1%|          | 29/3300 [05:26<11:29:02, 12.64s/it]INFO 04-12 15:30:46 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 30/3300 [05:37<11:01:29, 12.14s/it]2026-04-12 15:30:57,375 | INFO | train_grpo_train | metrics_logged mode=train step=30

                                                    
{'loss': 0.1002, 'grad_norm': 11.642829895019531, 'learning_rate': 9.912121212121213e-06, 'num_tokens': 70145.0, 'completions/mean_length': 67.25, 'completions/min_length': 53.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.6778810024261475, 'rewards/meter/std': 0.33140578866004944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.9941116571426392, 'rewards/repeat_penalty/std': 0.005025853868573904, 'rewards/repeat_floor/mean': 0.9991167783737183, 'rewards/repeat_floor/std': 0.000753876578528434, 'rewards/near_duplicate_penalty/mean': 0.9941116571426392, 'rewards/near_duplicate_penalty/std': 0.005025853868573904, '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': 0.467239111661911, 'rewards/total_composite/std': 0.3392532467842102, 'reward': 0.467239111661911, 'reward_std': 0.3392532169818878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16195835173130035, 'sampling/sampling_logp_difference/max': 1.5645742416381836, 'sampling/importance_sampling_ratio/min': 0.20917704701423645, 'sampling/importance_sampling_ratio/mean': 1.00449538230896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3717585802078247, 'clip_ratio/low_mean': 0.10288059245795012, 'clip_ratio/low_min': 0.10288059245795012, 'clip_ratio/high_mean': 0.05898871552199125, 'clip_ratio/high_max': 0.05898871552199125, 'clip_ratio/region_mean': 0.16186930797994137, 'reward_total_mean': 0.467239111661911, 'reward_meter_mean': 0.6778810024261475, 'reward_meter_std': 0.33140578866004944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9941116571426392, 'reward_repeat_penalty_std': 0.005025853868573904, 'reward_repeat_floor_mean': 0.9991167783737183, 'reward_repeat_floor_std': 0.000753876578528434, 'reward_near_duplicate_penalty_mean': 0.9941116571426392, 'reward_near_duplicate_penalty_std': 0.005025853868573904, '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': 0.2805097699165344, 'reward_total_composite_mean': 0.467239111661911, 'reward_total_composite_std': 0.3392532467842102, 'epoch': 0.0}

  1%|          | 30/3300 [05:37<11:01:29, 12.14s/it]INFO 04-12 15:30:57 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 31/3300 [05:48<10:46:55, 11.87s/it]2026-04-12 15:31:08,594 | INFO | train_grpo_train | metrics_logged mode=train step=31

                                                    
{'loss': 0.0048, 'grad_norm': 13.63132095336914, 'learning_rate': 9.90909090909091e-06, 'num_tokens': 71800.0, 'completions/mean_length': 56.875, 'completions/min_length': 34.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.4364202320575714, 'rewards/meter/std': 0.3727029860019684, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.9707009792327881, 'rewards/repeat_penalty/std': 0.05309678614139557, 'rewards/repeat_floor/mean': 0.9960089325904846, 'rewards/repeat_floor/std': 0.006874763872474432, 'rewards/near_duplicate_penalty/mean': 0.9786378145217896, 'rewards/near_duplicate_penalty/std': 0.03214501217007637, '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': 0.22145074605941772, 'rewards/total_composite/std': 0.2818836569786072, 'reward': 0.22145074605941772, 'reward_std': 0.28188368678092957, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19524186849594116, 'sampling/sampling_logp_difference/max': 2.38645076751709, 'sampling/importance_sampling_ratio/min': 0.09195548295974731, 'sampling/importance_sampling_ratio/mean': 1.041056513786316, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7630411237478256, 'clip_ratio/low_mean': 0.13699123077094555, 'clip_ratio/low_min': 0.13699123077094555, 'clip_ratio/high_mean': 0.03808922600001097, 'clip_ratio/high_max': 0.03808922600001097, 'clip_ratio/region_mean': 0.17508045677095652, 'reward_total_mean': 0.22145074605941772, 'reward_meter_mean': 0.4364202320575714, 'reward_meter_std': 0.3727029860019684, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9707009792327881, 'reward_repeat_penalty_std': 0.05309678614139557, 'reward_repeat_floor_mean': 0.9960089325904846, 'reward_repeat_floor_std': 0.006874763872474432, 'reward_near_duplicate_penalty_mean': 0.9786378145217896, 'reward_near_duplicate_penalty_std': 0.03214501217007637, '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': 0.4925000071525574, 'reward_judge_quality_std': 0.26385873556137085, 'reward_total_composite_mean': 0.22145074605941772, 'reward_total_composite_std': 0.2818836569786072, 'epoch': 0.0}

  1%|          | 31/3300 [05:48<10:46:55, 11.87s/it]INFO 04-12 15:31:08 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 32/3300 [05:59<10:22:42, 11.43s/it]2026-04-12 15:31:18,996 | INFO | train_grpo_train | metrics_logged mode=train step=32

                                                    
{'loss': 0.1396, 'grad_norm': 12.520078659057617, 'learning_rate': 9.906060606060607e-06, 'num_tokens': 73554.0, 'completions/mean_length': 57.25, 'completions/min_length': 32.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6373588442802429, 'rewards/meter/std': 0.42701539397239685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.99669349193573, 'rewards/repeat_penalty/std': 0.00636273343116045, 'rewards/repeat_floor/mean': 0.9995039701461792, 'rewards/repeat_floor/std': 0.000954415590967983, 'rewards/near_duplicate_penalty/mean': 0.99669349193573, 'rewards/near_duplicate_penalty/std': 0.00636273343116045, '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': 0.1878948211669922, 'rewards/total_composite/std': 0.1571483612060547, 'reward': 0.1878948211669922, 'reward_std': 0.1571483612060547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19454073905944824, 'sampling/sampling_logp_difference/max': 1.4174580574035645, 'sampling/importance_sampling_ratio/min': 0.24232923984527588, 'sampling/importance_sampling_ratio/mean': 1.019760251045227, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9404871463775635, 'clip_ratio/low_mean': 0.1199952345341444, 'clip_ratio/low_min': 0.1199952345341444, 'clip_ratio/high_mean': 0.0848543792963028, 'clip_ratio/high_max': 0.0848543792963028, 'clip_ratio/region_mean': 0.2048496138304472, 'reward_total_mean': 0.1878948211669922, 'reward_meter_mean': 0.6373588442802429, 'reward_meter_std': 0.42701539397239685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.99669349193573, 'reward_repeat_penalty_std': 0.00636273343116045, 'reward_repeat_floor_mean': 0.9995039701461792, 'reward_repeat_floor_std': 0.000954415590967983, 'reward_near_duplicate_penalty_mean': 0.99669349193573, 'reward_near_duplicate_penalty_std': 0.00636273343116045, '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': 0.13887301087379456, 'reward_total_composite_mean': 0.1878948211669922, 'reward_total_composite_std': 0.1571483612060547, 'epoch': 0.0}

  1%|          | 32/3300 [05:59<10:22:42, 11.43s/it]INFO 04-12 15:31:19 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 33/3300 [06:11<10:38:35, 11.73s/it]2026-04-12 15:31:31,415 | INFO | train_grpo_train | metrics_logged mode=train step=33

                                                    
{'loss': 0.0278, 'grad_norm': 8.586396217346191, 'learning_rate': 9.903030303030305e-06, 'num_tokens': 76304.0, 'completions/mean_length': 154.75, 'completions/min_length': 108.0, 'completions/max_length': 204.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.75, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 204.0, 'rewards/meter/mean': 0.36529800295829773, 'rewards/meter/std': 0.22657908499240875, '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': 0.9947916269302368, 'rewards/lexical_plausibility/std': 0.014731398783624172, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9920752644538879, 'rewards/repeat_penalty/std': 0.014160385355353355, 'rewards/repeat_floor/mean': 0.9990055561065674, 'rewards/repeat_floor/std': 0.0015801809495314956, 'rewards/near_duplicate_penalty/mean': 0.9953945875167847, 'rewards/near_duplicate_penalty/std': 0.004976656287908554, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9966261982917786, 'rewards/distinct_2/std': 0.009542594663798809, 'rewards/total_composite/mean': 0.0919654443860054, 'rewards/total_composite/std': 0.06707508116960526, 'reward': 0.0919654443860054, 'reward_std': 0.06707508116960526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23216886818408966, 'sampling/sampling_logp_difference/max': 1.5689716339111328, 'sampling/importance_sampling_ratio/min': 0.2082592397928238, 'sampling/importance_sampling_ratio/mean': 1.0328764915466309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7506909668445587, 'clip_ratio/low_mean': 0.09479101933538914, 'clip_ratio/low_min': 0.09479101933538914, 'clip_ratio/high_mean': 0.11930768936872482, 'clip_ratio/high_max': 0.11930768936872482, 'clip_ratio/region_mean': 0.21409870870411396, 'reward_total_mean': 0.0919654443860054, 'reward_meter_mean': 0.36529800295829773, 'reward_meter_std': 0.22657908499240875, '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': 0.9947916269302368, 'reward_lexical_plausibility_std': 0.014731398783624172, 'reward_repeat_penalty_mean': 0.9920752644538879, 'reward_repeat_penalty_std': 0.014160385355353355, 'reward_repeat_floor_mean': 0.9990055561065674, 'reward_repeat_floor_std': 0.0015801809495314956, 'reward_near_duplicate_penalty_mean': 0.9953945875167847, 'reward_near_duplicate_penalty_std': 0.004976656287908554, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9966261982917786, 'reward_distinct_2_std': 0.009542594663798809, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.0919654443860054, 'reward_total_composite_std': 0.06707508116960526, 'epoch': 0.0}

  1%|          | 33/3300 [06:11<10:38:35, 11.73s/it]INFO 04-12 15:31:31 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 34/3300 [06:22<10:27:22, 11.53s/it]2026-04-12 15:31:42,468 | INFO | train_grpo_train | metrics_logged mode=train step=34

                                                    
{'loss': 0.0938, 'grad_norm': 10.180171012878418, 'learning_rate': 9.9e-06, 'num_tokens': 78449.0, 'completions/mean_length': 103.125, 'completions/min_length': 65.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.411704421043396, 'rewards/meter/std': 0.3591403663158417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.3412500023841858, 'rewards/judge_quality/std': 0.2391316145658493, 'rewards/repeat_penalty/mean': 0.9769642353057861, 'rewards/repeat_penalty/std': 0.01614268682897091, 'rewards/repeat_floor/mean': 0.9971598386764526, 'rewards/repeat_floor/std': 0.001443410525098443, 'rewards/near_duplicate_penalty/mean': 0.9873068332672119, 'rewards/near_duplicate_penalty/std': 0.007589713204652071, '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': 0.09900060296058655, 'rewards/total_composite/std': 0.07295992225408554, 'reward': 0.09900060296058655, 'reward_std': 0.07295991480350494, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22360895574092865, 'sampling/sampling_logp_difference/max': 2.222749710083008, 'sampling/importance_sampling_ratio/min': 0.10831087827682495, 'sampling/importance_sampling_ratio/mean': 1.0236644744873047, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.525913953781128, 'clip_ratio/low_mean': 0.1048739105463028, 'clip_ratio/low_min': 0.1048739105463028, 'clip_ratio/high_mean': 0.11431151069700718, 'clip_ratio/high_max': 0.11431151069700718, 'clip_ratio/region_mean': 0.21918542124330997, 'reward_total_mean': 0.09900060296058655, 'reward_meter_mean': 0.411704421043396, 'reward_meter_std': 0.3591403663158417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9769642353057861, 'reward_repeat_penalty_std': 0.01614268682897091, 'reward_repeat_floor_mean': 0.9971598386764526, 'reward_repeat_floor_std': 0.001443410525098443, 'reward_near_duplicate_penalty_mean': 0.9873068332672119, 'reward_near_duplicate_penalty_std': 0.007589713204652071, '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': 0.3412500023841858, 'reward_judge_quality_std': 0.2391316145658493, 'reward_total_composite_mean': 0.09900060296058655, 'reward_total_composite_std': 0.07295992225408554, 'epoch': 0.0}

  1%|          | 34/3300 [06:22<10:27:22, 11.53s/it]INFO 04-12 15:31:42 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 35/3300 [06:38<11:31:15, 12.70s/it]2026-04-12 15:31:57,918 | INFO | train_grpo_train | metrics_logged mode=train step=35

                                                    
{'loss': 0.0145, 'grad_norm': 5.9968180656433105, 'learning_rate': 9.896969696969699e-06, 'num_tokens': 82036.0, 'completions/mean_length': 238.375, 'completions/min_length': 160.0, 'completions/max_length': 372.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 238.375, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 372.0, 'rewards/meter/mean': 0.38229402899742126, 'rewards/meter/std': 0.2463291436433792, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.05750546231865883, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, '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': 0.32499998807907104, 'rewards/judge_quality/std': 0.08864051848649979, 'rewards/repeat_penalty/mean': 0.9921417236328125, 'rewards/repeat_penalty/std': 0.0073619126342237, 'rewards/repeat_floor/mean': 0.9989433288574219, 'rewards/repeat_floor/std': 0.000841090630274266, 'rewards/near_duplicate_penalty/mean': 0.9941961169242859, 'rewards/near_duplicate_penalty/std': 0.00413910299539566, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9979321956634521, 'rewards/distinct_2/std': 0.0058486973866820335, 'rewards/total_composite/mean': 0.1344568133354187, 'rewards/total_composite/std': 0.11377713084220886, 'reward': 0.1344568133354187, 'reward_std': 0.11377713084220886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2276795357465744, 'sampling/sampling_logp_difference/max': 1.8151664733886719, 'sampling/importance_sampling_ratio/min': 0.1628108024597168, 'sampling/importance_sampling_ratio/mean': 1.056570053100586, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9400649964809418, 'clip_ratio/low_mean': 0.1554820965975523, 'clip_ratio/low_min': 0.1554820965975523, 'clip_ratio/high_mean': 0.0467352420091629, 'clip_ratio/high_max': 0.0467352420091629, 'clip_ratio/region_mean': 0.2022173386067152, 'reward_total_mean': 0.1344568133354187, 'reward_meter_mean': 0.38229402899742126, 'reward_meter_std': 0.2463291436433792, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.05750546231865883, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, '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': 0.9921417236328125, 'reward_repeat_penalty_std': 0.0073619126342237, 'reward_repeat_floor_mean': 0.9989433288574219, 'reward_repeat_floor_std': 0.000841090630274266, 'reward_near_duplicate_penalty_mean': 0.9941961169242859, 'reward_near_duplicate_penalty_std': 0.00413910299539566, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9979321956634521, 'reward_distinct_2_std': 0.0058486973866820335, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.08864051848649979, 'reward_total_composite_mean': 0.1344568133354187, 'reward_total_composite_std': 0.11377713084220886, 'epoch': 0.0}

  1%|          | 35/3300 [06:38<11:31:15, 12.70s/it]INFO 04-12 15:31:58 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 36/3300 [06:54<12:34:01, 13.86s/it]2026-04-12 15:32:14,479 | INFO | train_grpo_train | metrics_logged mode=train step=36

                                                    
{'loss': -0.1239, 'grad_norm': 3.8850080966949463, 'learning_rate': 9.893939393939395e-06, 'num_tokens': 85617.0, 'completions/mean_length': 417.625, 'completions/min_length': 273.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 361.0, 'completions/min_terminated_length': 273.0, 'completions/max_terminated_length': 441.0, 'rewards/meter/mean': 0.3704602122306824, 'rewards/meter/std': 0.2523605525493622, 'rewards/count_adherence/mean': 0.904411792755127, 'rewards/count_adherence/std': 0.05388972535729408, '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': 0.9713234901428223, 'rewards/count_floor/std': 0.016166917979717255, 'rewards/lexical_plausibility/mean': 0.9926470518112183, 'rewards/lexical_plausibility/std': 0.020797256380319595, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.11259915679693222, 'rewards/repeat_penalty/mean': 0.9598931074142456, 'rewards/repeat_penalty/std': 0.07228787988424301, 'rewards/repeat_floor/mean': 0.9958555102348328, 'rewards/repeat_floor/std': 0.006967020686715841, 'rewards/near_duplicate_penalty/mean': 0.990498960018158, 'rewards/near_duplicate_penalty/std': 0.009715107269585133, 'rewards/opening_diversity/mean': 0.9934210777282715, 'rewards/opening_diversity/std': 0.01860806532204151, 'rewards/distinct_2/mean': 0.9741712212562561, 'rewards/distinct_2/std': 0.04915395379066467, 'rewards/total_composite/mean': 0.09012877196073532, 'rewards/total_composite/std': 0.07985422015190125, 'reward': 0.09012877196073532, 'reward_std': 0.07985421270132065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23293013870716095, 'sampling/sampling_logp_difference/max': 1.5956039428710938, 'sampling/importance_sampling_ratio/min': 0.20278601348400116, 'sampling/importance_sampling_ratio/mean': 1.050012469291687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.258246600627899, 'clip_ratio/low_mean': 0.06290536001324654, 'clip_ratio/low_min': 0.06290536001324654, 'clip_ratio/high_mean': 0.053880227729678154, 'clip_ratio/high_max': 0.053880227729678154, 'clip_ratio/region_mean': 0.11678558774292469, 'reward_total_mean': 0.09012877196073532, 'reward_meter_mean': 0.3704602122306824, 'reward_meter_std': 0.2523605525493622, 'reward_count_adherence_mean': 0.904411792755127, 'reward_count_adherence_std': 0.05388972535729408, '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': 0.9713234901428223, 'reward_count_floor_std': 0.016166917979717255, 'reward_lexical_plausibility_mean': 0.9926470518112183, 'reward_lexical_plausibility_std': 0.020797256380319595, 'reward_repeat_penalty_mean': 0.9598931074142456, 'reward_repeat_penalty_std': 0.07228787988424301, 'reward_repeat_floor_mean': 0.9958555102348328, 'reward_repeat_floor_std': 0.006967020686715841, 'reward_near_duplicate_penalty_mean': 0.990498960018158, 'reward_near_duplicate_penalty_std': 0.009715107269585133, 'reward_opening_diversity_mean': 0.9934210777282715, 'reward_opening_diversity_std': 0.01860806532204151, 'reward_distinct_2_mean': 0.9741712212562561, 'reward_distinct_2_std': 0.04915395379066467, 'reward_judge_quality_mean': 0.23750001192092896, 'reward_judge_quality_std': 0.11259915679693222, 'reward_total_composite_mean': 0.09012877196073532, 'reward_total_composite_std': 0.07985422015190125, 'epoch': 0.0}

  1%|          | 36/3300 [06:54<12:34:01, 13.86s/it]INFO 04-12 15:32:14 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 37/3300 [07:05<11:52:58, 13.11s/it]2026-04-12 15:32:25,839 | INFO | train_grpo_train | metrics_logged mode=train step=37

                                                    
{'loss': 0.0721, 'grad_norm': 10.824390411376953, 'learning_rate': 9.890909090909092e-06, 'num_tokens': 87528.0, 'completions/mean_length': 62.875, 'completions/min_length': 43.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8028993606567383, 'rewards/meter/std': 0.3583672344684601, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.3125, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/repeat_penalty/mean': 0.9925150275230408, 'rewards/repeat_penalty/std': 0.01191670261323452, 'rewards/repeat_floor/mean': 0.998877227306366, 'rewards/repeat_floor/std': 0.001787504879757762, 'rewards/near_duplicate_penalty/mean': 0.9925150275230408, 'rewards/near_duplicate_penalty/std': 0.01191670261323452, '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': 0.2672617733478546, 'rewards/total_composite/std': 0.14594782888889313, 'reward': 0.2672617733478546, 'reward_std': 0.14594782888889313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21163925528526306, 'sampling/sampling_logp_difference/max': 1.3112602233886719, 'sampling/importance_sampling_ratio/min': 0.26948022842407227, 'sampling/importance_sampling_ratio/mean': 1.0406631231307983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2507003843784332, 'clip_ratio/low_mean': 0.07544384896755219, 'clip_ratio/low_min': 0.07544384896755219, 'clip_ratio/high_mean': 0.16523032821714878, 'clip_ratio/high_max': 0.16523032821714878, 'clip_ratio/region_mean': 0.24067417718470097, 'reward_total_mean': 0.2672617733478546, 'reward_meter_mean': 0.8028993606567383, 'reward_meter_std': 0.3583672344684601, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9925150275230408, 'reward_repeat_penalty_std': 0.01191670261323452, 'reward_repeat_floor_mean': 0.998877227306366, 'reward_repeat_floor_std': 0.001787504879757762, 'reward_near_duplicate_penalty_mean': 0.9925150275230408, 'reward_near_duplicate_penalty_std': 0.01191670261323452, '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': 0.2672617733478546, 'reward_total_composite_std': 0.14594782888889313, 'epoch': 0.0}

  1%|          | 37/3300 [07:05<11:52:58, 13.11s/it]INFO 04-12 15:32:26 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 38/3300 [07:15<11:02:36, 12.19s/it]2026-04-12 15:32:35,875 | INFO | train_grpo_train | metrics_logged mode=train step=38

                                                    
{'loss': -0.019, 'grad_norm': 14.67089557647705, 'learning_rate': 9.887878787878789e-06, 'num_tokens': 89053.0, 'completions/mean_length': 45.625, 'completions/min_length': 33.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.40140506625175476, 'rewards/meter/std': 0.4438471496105194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9613440036773682, 'rewards/lexical_plausibility/std': 0.09522879123687744, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.19668231904506683, 'rewards/repeat_penalty/mean': 0.9098960161209106, 'rewards/repeat_penalty/std': 0.1974615901708603, 'rewards/repeat_floor/mean': 0.9886674880981445, 'rewards/repeat_floor/std': 0.02346639521420002, 'rewards/near_duplicate_penalty/mean': 0.9612413644790649, 'rewards/near_duplicate_penalty/std': 0.053749240934848785, 'rewards/opening_diversity/mean': 0.96875, 'rewards/opening_diversity/std': 0.0883883461356163, 'rewards/distinct_2/mean': 0.9595141410827637, 'rewards/distinct_2/std': 0.11451122164726257, 'rewards/total_composite/mean': 0.16040226817131042, 'rewards/total_composite/std': 0.1722187101840973, 'reward': 0.16040226817131042, 'reward_std': 0.1722187101840973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2073492407798767, 'sampling/sampling_logp_difference/max': 1.552976131439209, 'sampling/importance_sampling_ratio/min': 0.21161723136901855, 'sampling/importance_sampling_ratio/mean': 1.0489842891693115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7606774419546127, 'clip_ratio/low_mean': 0.13604292180389166, 'clip_ratio/low_min': 0.13604292180389166, 'clip_ratio/high_mean': 0.06482307612895966, 'clip_ratio/high_max': 0.06482307612895966, 'clip_ratio/region_mean': 0.20086599793285131, 'reward_total_mean': 0.16040226817131042, 'reward_meter_mean': 0.40140506625175476, 'reward_meter_std': 0.4438471496105194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9613440036773682, 'reward_lexical_plausibility_std': 0.09522879123687744, 'reward_repeat_penalty_mean': 0.9098960161209106, 'reward_repeat_penalty_std': 0.1974615901708603, 'reward_repeat_floor_mean': 0.9886674880981445, 'reward_repeat_floor_std': 0.02346639521420002, 'reward_near_duplicate_penalty_mean': 0.9612413644790649, 'reward_near_duplicate_penalty_std': 0.053749240934848785, 'reward_opening_diversity_mean': 0.96875, 'reward_opening_diversity_std': 0.0883883461356163, 'reward_distinct_2_mean': 0.9595141410827637, 'reward_distinct_2_std': 0.11451122164726257, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.19668231904506683, 'reward_total_composite_mean': 0.16040226817131042, 'reward_total_composite_std': 0.1722187101840973, 'epoch': 0.0}

  1%|          | 38/3300 [07:15<11:02:36, 12.19s/it]INFO 04-12 15:32:36 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 39/3300 [07:31<11:55:26, 13.16s/it]2026-04-12 15:32:51,316 | INFO | train_grpo_train | metrics_logged mode=train step=39

                                                    
{'loss': 0.1693, 'grad_norm': 3.7285618782043457, 'learning_rate': 9.884848484848486e-06, 'num_tokens': 93483.0, 'completions/mean_length': 436.75, 'completions/min_length': 336.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 411.66668701171875, 'completions/min_terminated_length': 336.0, 'completions/max_terminated_length': 478.0, 'rewards/meter/mean': 0.7699059247970581, 'rewards/meter/std': 0.3402743935585022, 'rewards/count_adherence/mean': 0.9615384340286255, 'rewards/count_adherence/std': 0.041117113083601, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9884614944458008, 'rewards/count_floor/std': 0.012335150502622128, 'rewards/lexical_plausibility/mean': 0.9794768691062927, 'rewards/lexical_plausibility/std': 0.05804818123579025, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.07440237700939178, 'rewards/repeat_penalty/mean': 0.9960432052612305, 'rewards/repeat_penalty/std': 0.0015128182712942362, 'rewards/repeat_floor/mean': 0.9994064569473267, 'rewards/repeat_floor/std': 0.00022692015045322478, 'rewards/near_duplicate_penalty/mean': 0.9960432052612305, 'rewards/near_duplicate_penalty/std': 0.0015128182712942362, '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': 0.1492643654346466, 'rewards/total_composite/std': 0.0966491624712944, 'reward': 0.1492643654346466, 'reward_std': 0.0966491550207138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21291592717170715, 'sampling/sampling_logp_difference/max': 1.6394968032836914, 'sampling/importance_sampling_ratio/min': 0.19407767057418823, 'sampling/importance_sampling_ratio/mean': 1.039568305015564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.262940913438797, 'clip_ratio/low_mean': 0.10346917249262333, 'clip_ratio/low_min': 0.10346917249262333, 'clip_ratio/high_mean': 0.02666666731238365, 'clip_ratio/high_max': 0.02666666731238365, 'clip_ratio/region_mean': 0.13013583980500698, 'reward_total_mean': 0.1492643654346466, 'reward_meter_mean': 0.7699059247970581, 'reward_meter_std': 0.3402743935585022, 'reward_count_adherence_mean': 0.9615384340286255, 'reward_count_adherence_std': 0.041117113083601, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9884614944458008, 'reward_count_floor_std': 0.012335150502622128, 'reward_lexical_plausibility_mean': 0.9794768691062927, 'reward_lexical_plausibility_std': 0.05804818123579025, 'reward_repeat_penalty_mean': 0.9960432052612305, 'reward_repeat_penalty_std': 0.0015128182712942362, 'reward_repeat_floor_mean': 0.9994064569473267, 'reward_repeat_floor_std': 0.00022692015045322478, 'reward_near_duplicate_penalty_mean': 0.9960432052612305, 'reward_near_duplicate_penalty_std': 0.0015128182712942362, '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': 0.07440237700939178, 'reward_total_composite_mean': 0.1492643654346466, 'reward_total_composite_std': 0.0966491624712944, 'epoch': 0.0}

  1%|          | 39/3300 [07:31<11:55:26, 13.16s/it]INFO 04-12 15:32:51 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 40/3300 [07:42<11:17:13, 12.46s/it]2026-04-12 15:33:02,149 | INFO | train_grpo_train | metrics_logged mode=train step=40

                                                    
{'loss': 0.0421, 'grad_norm': 14.545975685119629, 'learning_rate': 9.881818181818182e-06, 'num_tokens': 95288.0, 'completions/mean_length': 52.625, 'completions/min_length': 36.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.6627398729324341, 'rewards/meter/std': 0.4211808443069458, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9653618931770325, 'rewards/lexical_plausibility/std': 0.06416625529527664, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.10350983589887619, 'rewards/repeat_penalty/mean': 0.9966648817062378, 'rewards/repeat_penalty/std': 0.006504828110337257, 'rewards/repeat_floor/mean': 0.9994997382164001, 'rewards/repeat_floor/std': 0.0009757138323038816, 'rewards/near_duplicate_penalty/mean': 0.9966648817062378, 'rewards/near_duplicate_penalty/std': 0.006504828110337257, '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': 0.24780182540416718, 'rewards/total_composite/std': 0.18882529437541962, 'reward': 0.24780182540416718, 'reward_std': 0.1888253092765808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2167918086051941, 'sampling/sampling_logp_difference/max': 1.2831764221191406, 'sampling/importance_sampling_ratio/min': 0.2771555483341217, 'sampling/importance_sampling_ratio/mean': 1.0531648397445679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5625782012939453, 'clip_ratio/low_mean': 0.09673448745161295, 'clip_ratio/low_min': 0.09673448745161295, 'clip_ratio/high_mean': 0.10798456333577633, 'clip_ratio/high_max': 0.10798456333577633, 'clip_ratio/region_mean': 0.20471905078738928, 'reward_total_mean': 0.24780182540416718, 'reward_meter_mean': 0.6627398729324341, 'reward_meter_std': 0.4211808443069458, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9653618931770325, 'reward_lexical_plausibility_std': 0.06416625529527664, 'reward_repeat_penalty_mean': 0.9966648817062378, 'reward_repeat_penalty_std': 0.006504828110337257, 'reward_repeat_floor_mean': 0.9994997382164001, 'reward_repeat_floor_std': 0.0009757138323038816, 'reward_near_duplicate_penalty_mean': 0.9966648817062378, 'reward_near_duplicate_penalty_std': 0.006504828110337257, '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': 0.10350983589887619, 'reward_total_composite_mean': 0.24780182540416718, 'reward_total_composite_std': 0.18882529437541962, 'epoch': 0.0}

  1%|          | 40/3300 [07:42<11:17:13, 12.46s/it]INFO 04-12 15:33:02 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 41/3300 [07:48<9:38:20, 10.65s/it] 2026-04-12 15:33:08,556 | INFO | train_grpo_train | metrics_logged mode=train step=41

                                                   
{'loss': 0.0762, 'grad_norm': 18.142820358276367, 'learning_rate': 9.87878787878788e-06, 'num_tokens': 96765.0, 'completions/mean_length': 28.625, 'completions/min_length': 28.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6262686252593994, 'rewards/meter/std': 0.21932648122310638, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9275000095367432, 'rewards/judge_quality/std': 0.013887288980185986, 'rewards/repeat_penalty/mean': 0.995519757270813, 'rewards/repeat_penalty/std': 0.001612812397070229, 'rewards/repeat_floor/mean': 0.9993279576301575, 'rewards/repeat_floor/std': 0.0002419229131191969, 'rewards/near_duplicate_penalty/mean': 0.995519757270813, 'rewards/near_duplicate_penalty/std': 0.001612812397070229, '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': 0.5809922814369202, 'rewards/total_composite/std': 0.20403531193733215, 'reward': 0.5809922814369202, 'reward_std': 0.20403531193733215, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04791252315044403, 'sampling/sampling_logp_difference/max': 1.212748646736145, 'sampling/importance_sampling_ratio/min': 0.2973787784576416, 'sampling/importance_sampling_ratio/mean': 0.9945252537727356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.19142003450542688, 'clip_ratio/low_mean': 0.022727273404598236, 'clip_ratio/low_min': 0.022727273404598236, 'clip_ratio/high_mean': 0.022321429569274187, 'clip_ratio/high_max': 0.022321429569274187, 'clip_ratio/region_mean': 0.04504870297387242, 'reward_total_mean': 0.5809922814369202, 'reward_meter_mean': 0.6262686252593994, 'reward_meter_std': 0.21932648122310638, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.995519757270813, 'reward_repeat_penalty_std': 0.001612812397070229, 'reward_repeat_floor_mean': 0.9993279576301575, 'reward_repeat_floor_std': 0.0002419229131191969, 'reward_near_duplicate_penalty_mean': 0.995519757270813, 'reward_near_duplicate_penalty_std': 0.001612812397070229, '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': 0.9275000095367432, 'reward_judge_quality_std': 0.013887288980185986, 'reward_total_composite_mean': 0.5809922814369202, 'reward_total_composite_std': 0.20403531193733215, 'epoch': 0.0}

  1%|          | 41/3300 [07:48<9:38:20, 10.65s/it]INFO 04-12 15:33:08 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 42/3300 [08:03<10:47:46, 11.93s/it]2026-04-12 15:33:23,478 | INFO | train_grpo_train | metrics_logged mode=train step=42

                                                    
{'loss': 0.1345, 'grad_norm': 1.6544698476791382, 'learning_rate': 9.875757575757576e-06, 'num_tokens': 99084.0, 'completions/mean_length': 507.875, 'completions/min_length': 479.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.875, 'completions/mean_terminated_length': 479.0, 'completions/min_terminated_length': 479.0, 'completions/max_terminated_length': 479.0, 'rewards/meter/mean': 0.6511483192443848, 'rewards/meter/std': 0.3560481667518616, 'rewards/count_adherence/mean': 0.9500000476837158, 'rewards/count_adherence/std': 0.030860668048262596, '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': 0.009258192032575607, 'rewards/lexical_plausibility/mean': 0.9631125926971436, 'rewards/lexical_plausibility/std': 0.04655386880040169, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9927043914794922, 'rewards/repeat_penalty/std': 0.0060564796440303326, 'rewards/repeat_floor/mean': 0.9990572929382324, 'rewards/repeat_floor/std': 0.0006409144261851907, 'rewards/near_duplicate_penalty/mean': 0.9952572584152222, 'rewards/near_duplicate_penalty/std': 0.001910837716422975, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9974303245544434, 'rewards/distinct_2/std': 0.004765058867633343, 'rewards/total_composite/mean': 0.1181429922580719, 'rewards/total_composite/std': 0.10310676693916321, 'reward': 0.1181429922580719, 'reward_std': 0.10310675948858261, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21991652250289917, 'sampling/sampling_logp_difference/max': 1.5076560974121094, 'sampling/importance_sampling_ratio/min': 0.22142839431762695, 'sampling/importance_sampling_ratio/mean': 1.0317631959915161, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3663713037967682, 'clip_ratio/low_mean': 0.019050104543566704, 'clip_ratio/low_min': 0.019050104543566704, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.019050104543566704, 'reward_total_mean': 0.1181429922580719, 'reward_meter_mean': 0.6511483192443848, 'reward_meter_std': 0.3560481667518616, 'reward_count_adherence_mean': 0.9500000476837158, 'reward_count_adherence_std': 0.030860668048262596, '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': 0.009258192032575607, 'reward_lexical_plausibility_mean': 0.9631125926971436, 'reward_lexical_plausibility_std': 0.04655386880040169, 'reward_repeat_penalty_mean': 0.9927043914794922, 'reward_repeat_penalty_std': 0.0060564796440303326, 'reward_repeat_floor_mean': 0.9990572929382324, 'reward_repeat_floor_std': 0.0006409144261851907, 'reward_near_duplicate_penalty_mean': 0.9952572584152222, 'reward_near_duplicate_penalty_std': 0.001910837716422975, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9974303245544434, 'reward_distinct_2_std': 0.004765058867633343, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660183429718, 'reward_total_composite_mean': 0.1181429922580719, 'reward_total_composite_std': 0.10310676693916321, 'epoch': 0.0}

  1%|▏         | 42/3300 [08:03<10:47:46, 11.93s/it]INFO 04-12 15:33:23 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 43/3300 [08:14<10:33:18, 11.67s/it]2026-04-12 15:33:34,532 | INFO | train_grpo_train | metrics_logged mode=train step=43

                                                    
{'loss': 0.0228, 'grad_norm': 13.056751251220703, 'learning_rate': 9.872727272727274e-06, 'num_tokens': 101057.0, 'completions/mean_length': 76.625, 'completions/min_length': 42.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.625, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.5245823860168457, 'rewards/meter/std': 0.3338170349597931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9497177600860596, 'rewards/lexical_plausibility/std': 0.08389555662870407, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9922226071357727, 'rewards/repeat_penalty/std': 0.005637835245579481, 'rewards/repeat_floor/mean': 0.9988333582878113, 'rewards/repeat_floor/std': 0.0008456670329906046, 'rewards/near_duplicate_penalty/mean': 0.9922226071357727, 'rewards/near_duplicate_penalty/std': 0.005637835245579481, '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': 0.14095836877822876, 'rewards/total_composite/std': 0.10865902900695801, 'reward': 0.14095836877822876, 'reward_std': 0.10865902900695801, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2289411872625351, 'sampling/sampling_logp_difference/max': 1.0979537963867188, 'sampling/importance_sampling_ratio/min': 0.33355289697647095, 'sampling/importance_sampling_ratio/mean': 1.0378375053405762, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9542465806007385, 'clip_ratio/low_mean': 0.10622397344559431, 'clip_ratio/low_min': 0.10622397344559431, 'clip_ratio/high_mean': 0.07670014910399914, 'clip_ratio/high_max': 0.07670014910399914, 'clip_ratio/region_mean': 0.18292412254959345, 'reward_total_mean': 0.14095836877822876, 'reward_meter_mean': 0.5245823860168457, 'reward_meter_std': 0.3338170349597931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9497177600860596, 'reward_lexical_plausibility_std': 0.08389555662870407, 'reward_repeat_penalty_mean': 0.9922226071357727, 'reward_repeat_penalty_std': 0.005637835245579481, 'reward_repeat_floor_mean': 0.9988333582878113, 'reward_repeat_floor_std': 0.0008456670329906046, 'reward_near_duplicate_penalty_mean': 0.9922226071357727, 'reward_near_duplicate_penalty_std': 0.005637835245579481, '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': 0.13887301087379456, 'reward_total_composite_mean': 0.14095836877822876, 'reward_total_composite_std': 0.10865902900695801, 'epoch': 0.0}

  1%|▏         | 43/3300 [08:14<10:33:18, 11.67s/it]INFO 04-12 15:33:34 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 44/3300 [08:25<10:15:07, 11.34s/it]2026-04-12 15:33:45,093 | INFO | train_grpo_train | metrics_logged mode=train step=44

                                                    
{'loss': 0.0465, 'grad_norm': 10.664681434631348, 'learning_rate': 9.869696969696971e-06, 'num_tokens': 103524.0, 'completions/mean_length': 133.375, 'completions/min_length': 125.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.375, 'completions/min_terminated_length': 125.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.6796356439590454, 'rewards/meter/std': 0.3681916892528534, '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': 0.9957386255264282, 'rewards/lexical_plausibility/std': 0.012052967213094234, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.9921809434890747, 'rewards/repeat_penalty/std': 0.00745220435783267, 'rewards/repeat_floor/mean': 0.9988270998001099, 'rewards/repeat_floor/std': 0.0011178392451256514, 'rewards/near_duplicate_penalty/mean': 0.9921809434890747, 'rewards/near_duplicate_penalty/std': 0.00745220435783267, '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': 0.25574326515197754, 'rewards/total_composite/std': 0.160466730594635, 'reward': 0.25574326515197754, 'reward_std': 0.160466730594635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20719045400619507, 'sampling/sampling_logp_difference/max': 2.266843318939209, 'sampling/importance_sampling_ratio/min': 0.10363882035017014, 'sampling/importance_sampling_ratio/mean': 1.025524616241455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9783300161361694, 'clip_ratio/low_mean': 0.0797614585608244, 'clip_ratio/low_min': 0.0797614585608244, 'clip_ratio/high_mean': 0.12393800355494022, 'clip_ratio/high_max': 0.12393800355494022, 'clip_ratio/region_mean': 0.20369946211576462, 'reward_total_mean': 0.25574326515197754, 'reward_meter_mean': 0.6796356439590454, 'reward_meter_std': 0.3681916892528534, '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': 0.9957386255264282, 'reward_lexical_plausibility_std': 0.012052967213094234, 'reward_repeat_penalty_mean': 0.9921809434890747, 'reward_repeat_penalty_std': 0.00745220435783267, 'reward_repeat_floor_mean': 0.9988270998001099, 'reward_repeat_floor_std': 0.0011178392451256514, 'reward_near_duplicate_penalty_mean': 0.9921809434890747, 'reward_near_duplicate_penalty_std': 0.00745220435783267, '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': 0.09258200228214264, 'reward_total_composite_mean': 0.25574326515197754, 'reward_total_composite_std': 0.160466730594635, 'epoch': 0.0}

  1%|▏         | 44/3300 [08:25<10:15:07, 11.34s/it]INFO 04-12 15:33:45 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 45/3300 [08:36<10:06:57, 11.19s/it]2026-04-12 15:33:55,941 | INFO | train_grpo_train | metrics_logged mode=train step=45

                                                    
{'loss': 0.0274, 'grad_norm': 13.253972053527832, 'learning_rate': 9.866666666666668e-06, 'num_tokens': 105392.0, 'completions/mean_length': 60.5, 'completions/min_length': 45.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.5734286308288574, 'rewards/meter/std': 0.4587654173374176, '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': 0.9921802282333374, 'rewards/repeat_penalty/std': 0.00954827107489109, 'rewards/repeat_floor/mean': 0.9988269805908203, 'rewards/repeat_floor/std': 0.0014322433853521943, 'rewards/near_duplicate_penalty/mean': 0.9921802282333374, 'rewards/near_duplicate_penalty/std': 0.00954827107489109, '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': 0.24287740886211395, 'rewards/total_composite/std': 0.19906795024871826, 'reward': 0.24287740886211395, 'reward_std': 0.19906793534755707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2012096345424652, 'sampling/sampling_logp_difference/max': 1.835031509399414, 'sampling/importance_sampling_ratio/min': 0.15960848331451416, 'sampling/importance_sampling_ratio/mean': 1.0404279232025146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1705244183540344, 'clip_ratio/low_mean': 0.08109668269753456, 'clip_ratio/low_min': 0.08109668269753456, 'clip_ratio/high_mean': 0.1209473479539156, 'clip_ratio/high_max': 0.1209473479539156, 'clip_ratio/region_mean': 0.20204403065145016, 'reward_total_mean': 0.24287740886211395, 'reward_meter_mean': 0.5734286308288574, 'reward_meter_std': 0.4587654173374176, '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': 0.9921802282333374, 'reward_repeat_penalty_std': 0.00954827107489109, 'reward_repeat_floor_mean': 0.9988269805908203, 'reward_repeat_floor_std': 0.0014322433853521943, 'reward_near_duplicate_penalty_mean': 0.9921802282333374, 'reward_near_duplicate_penalty_std': 0.00954827107489109, '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': 0.24287740886211395, 'reward_total_composite_std': 0.19906795024871826, 'epoch': 0.0}

  1%|▏         | 45/3300 [08:36<10:06:57, 11.19s/it]INFO 04-12 15:33:56 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 46/3300 [08:45<9:46:39, 10.82s/it] 2026-04-12 15:34:05,889 | INFO | train_grpo_train | metrics_logged mode=train step=46

                                                   
{'loss': 0.2277, 'grad_norm': 18.42459487915039, 'learning_rate': 9.863636363636364e-06, 'num_tokens': 106794.0, 'completions/mean_length': 53.25, 'completions/min_length': 34.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.5852382779121399, 'rewards/meter/std': 0.4186255633831024, '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': 0.7778487801551819, 'rewards/lexical_plausibility/std': 0.20808249711990356, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9214205741882324, 'rewards/repeat_penalty/std': 0.12742295861244202, 'rewards/repeat_floor/mean': 0.9904617667198181, 'rewards/repeat_floor/std': 0.01590562053024769, 'rewards/near_duplicate_penalty/mean': 0.9674508571624756, 'rewards/near_duplicate_penalty/std': 0.05521593987941742, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9482681155204773, 'rewards/distinct_2/std': 0.08636173605918884, 'rewards/total_composite/mean': 0.08721469342708588, 'rewards/total_composite/std': 0.06235937774181366, 'reward': 0.08721469342708588, 'reward_std': 0.06235937401652336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2607600688934326, 'sampling/sampling_logp_difference/max': 2.129840850830078, 'sampling/importance_sampling_ratio/min': 0.11885620653629303, 'sampling/importance_sampling_ratio/mean': 1.0428297519683838, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.413048952817917, 'clip_ratio/low_mean': 0.08067348599433899, 'clip_ratio/low_min': 0.08067348599433899, 'clip_ratio/high_mean': 0.1449983101338148, 'clip_ratio/high_max': 0.1449983101338148, 'clip_ratio/region_mean': 0.2256717961281538, 'reward_total_mean': 0.08721469342708588, 'reward_meter_mean': 0.5852382779121399, 'reward_meter_std': 0.4186255633831024, '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': 0.7778487801551819, 'reward_lexical_plausibility_std': 0.20808249711990356, 'reward_repeat_penalty_mean': 0.9214205741882324, 'reward_repeat_penalty_std': 0.12742295861244202, 'reward_repeat_floor_mean': 0.9904617667198181, 'reward_repeat_floor_std': 0.01590562053024769, 'reward_near_duplicate_penalty_mean': 0.9674508571624756, 'reward_near_duplicate_penalty_std': 0.05521593987941742, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9482681155204773, 'reward_distinct_2_std': 0.08636173605918884, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.08721469342708588, 'reward_total_composite_std': 0.06235937774181366, 'epoch': 0.0}

  1%|▏         | 46/3300 [08:45<9:46:39, 10.82s/it]INFO 04-12 15:34:06 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 47/3300 [08:56<9:34:58, 10.61s/it]2026-04-12 15:34:16,001 | INFO | train_grpo_train | metrics_logged mode=train step=47

                                                   
{'loss': 0.079, 'grad_norm': 24.17237663269043, 'learning_rate': 9.860606060606061e-06, 'num_tokens': 108435.0, 'completions/mean_length': 35.125, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.4072815477848053, 'rewards/meter/std': 0.3229300379753113, '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': 0.9438661336898804, 'rewards/lexical_plausibility/std': 0.06871479749679565, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.2173829823732376, 'rewards/repeat_penalty/mean': 0.990066647529602, 'rewards/repeat_penalty/std': 0.014448706060647964, 'rewards/repeat_floor/mean': 0.9985100030899048, 'rewards/repeat_floor/std': 0.0021673087030649185, 'rewards/near_duplicate_penalty/mean': 0.990066647529602, 'rewards/near_duplicate_penalty/std': 0.014448706060647964, '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': 0.18267777562141418, 'rewards/total_composite/std': 0.14632606506347656, 'reward': 0.18267777562141418, 'reward_std': 0.14632606506347656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23760807514190674, 'sampling/sampling_logp_difference/max': 1.9303264617919922, 'sampling/importance_sampling_ratio/min': 0.14510083198547363, 'sampling/importance_sampling_ratio/mean': 0.996955156326294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5569427162408829, 'clip_ratio/low_mean': 0.12464997917413712, 'clip_ratio/low_min': 0.12464997917413712, 'clip_ratio/high_mean': 0.06870941631495953, 'clip_ratio/high_max': 0.06870941631495953, 'clip_ratio/region_mean': 0.19335939548909664, 'reward_total_mean': 0.18267777562141418, 'reward_meter_mean': 0.4072815477848053, 'reward_meter_std': 0.3229300379753113, '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': 0.9438661336898804, 'reward_lexical_plausibility_std': 0.06871479749679565, 'reward_repeat_penalty_mean': 0.990066647529602, 'reward_repeat_penalty_std': 0.014448706060647964, 'reward_repeat_floor_mean': 0.9985100030899048, 'reward_repeat_floor_std': 0.0021673087030649185, 'reward_near_duplicate_penalty_mean': 0.990066647529602, 'reward_near_duplicate_penalty_std': 0.014448706060647964, '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': 0.18267777562141418, 'reward_total_composite_std': 0.14632606506347656, 'epoch': 0.0}

  1%|▏         | 47/3300 [08:56<9:34:58, 10.61s/it]INFO 04-12 15:34:16 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 48/3300 [09:07<9:54:26, 10.97s/it]2026-04-12 15:34:27,813 | INFO | train_grpo_train | metrics_logged mode=train step=48

                                                   
{'loss': 0.1594, 'grad_norm': 9.763179779052734, 'learning_rate': 9.857575757575758e-06, 'num_tokens': 110717.0, 'completions/mean_length': 108.25, 'completions/min_length': 76.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.25, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.5324523448944092, 'rewards/meter/std': 0.2819362282752991, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.4375, 'rewards/judge_quality/std': 0.172688826918602, 'rewards/repeat_penalty/mean': 0.9970675706863403, 'rewards/repeat_penalty/std': 0.0017854582984000444, 'rewards/repeat_floor/mean': 0.9995601177215576, 'rewards/repeat_floor/std': 0.0002678222372196615, 'rewards/near_duplicate_penalty/mean': 0.9970675706863403, 'rewards/near_duplicate_penalty/std': 0.0017854582984000444, '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': 0.2149481475353241, 'rewards/total_composite/std': 0.09053788334131241, 'reward': 0.2149481475353241, 'reward_std': 0.09053788334131241, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17114581167697906, 'sampling/sampling_logp_difference/max': 1.2221838235855103, 'sampling/importance_sampling_ratio/min': 0.2945861518383026, 'sampling/importance_sampling_ratio/mean': 1.0245773792266846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5544067025184631, 'clip_ratio/low_mean': 0.07282837852835655, 'clip_ratio/low_min': 0.07282837852835655, 'clip_ratio/high_mean': 0.08689650148153305, 'clip_ratio/high_max': 0.08689650148153305, 'clip_ratio/region_mean': 0.1597248800098896, 'reward_total_mean': 0.2149481475353241, 'reward_meter_mean': 0.5324523448944092, 'reward_meter_std': 0.2819362282752991, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9970675706863403, 'reward_repeat_penalty_std': 0.0017854582984000444, 'reward_repeat_floor_mean': 0.9995601177215576, 'reward_repeat_floor_std': 0.0002678222372196615, 'reward_near_duplicate_penalty_mean': 0.9970675706863403, 'reward_near_duplicate_penalty_std': 0.0017854582984000444, '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': 0.4375, 'reward_judge_quality_std': 0.172688826918602, 'reward_total_composite_mean': 0.2149481475353241, 'reward_total_composite_std': 0.09053788334131241, 'epoch': 0.0}

  1%|▏         | 48/3300 [09:07<9:54:26, 10.97s/it]INFO 04-12 15:34:28 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 49/3300 [09:19<9:59:35, 11.07s/it]2026-04-12 15:34:39,112 | INFO | train_grpo_train | metrics_logged mode=train step=49

                                                   
{'loss': 0.1162, 'grad_norm': 15.116177558898926, 'learning_rate': 9.854545454545456e-06, 'num_tokens': 112441.0, 'completions/mean_length': 51.5, 'completions/min_length': 35.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.3499334156513214, 'rewards/meter/std': 0.37027695775032043, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.35374999046325684, 'rewards/judge_quality/std': 0.24224767088890076, 'rewards/repeat_penalty/mean': 0.9889936447143555, 'rewards/repeat_penalty/std': 0.009213130921125412, 'rewards/repeat_floor/mean': 0.9983490705490112, 'rewards/repeat_floor/std': 0.0013819633750244975, 'rewards/near_duplicate_penalty/mean': 0.9889936447143555, 'rewards/near_duplicate_penalty/std': 0.009213130921125412, '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': 0.09926129877567291, 'rewards/total_composite/std': 0.1285303235054016, 'reward': 0.09926129877567291, 'reward_std': 0.12853030860424042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24942898750305176, 'sampling/sampling_logp_difference/max': 2.068516254425049, 'sampling/importance_sampling_ratio/min': 0.12637315690517426, 'sampling/importance_sampling_ratio/mean': 1.0445268154144287, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.220777377486229, 'clip_ratio/low_mean': 0.14375014416873455, 'clip_ratio/low_min': 0.14375014416873455, 'clip_ratio/high_mean': 0.0872842688113451, 'clip_ratio/high_max': 0.0872842688113451, 'clip_ratio/region_mean': 0.23103441298007965, 'reward_total_mean': 0.09926129877567291, 'reward_meter_mean': 0.3499334156513214, 'reward_meter_std': 0.37027695775032043, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9889936447143555, 'reward_repeat_penalty_std': 0.009213130921125412, 'reward_repeat_floor_mean': 0.9983490705490112, 'reward_repeat_floor_std': 0.0013819633750244975, 'reward_near_duplicate_penalty_mean': 0.9889936447143555, 'reward_near_duplicate_penalty_std': 0.009213130921125412, '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': 0.35374999046325684, 'reward_judge_quality_std': 0.24224767088890076, 'reward_total_composite_mean': 0.09926129877567291, 'reward_total_composite_std': 0.1285303235054016, 'epoch': 0.0}

  1%|▏         | 49/3300 [09:19<9:59:35, 11.07s/it]INFO 04-12 15:34:39 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 50/3300 [09:29<9:43:34, 10.77s/it]2026-04-12 15:34:49,201 | INFO | train_grpo_train | metrics_logged mode=train step=50

                                                   
{'loss': 0.0165, 'grad_norm': 12.628697395324707, 'learning_rate': 9.851515151515151e-06, 'num_tokens': 114218.0, 'completions/mean_length': 65.125, 'completions/min_length': 50.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.17972107231616974, 'rewards/meter/std': 0.2734915316104889, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.5837500095367432, 'rewards/judge_quality/std': 0.2557307481765747, 'rewards/repeat_penalty/mean': 0.9935485124588013, 'rewards/repeat_penalty/std': 0.013540077954530716, 'rewards/repeat_floor/mean': 0.9990322589874268, 'rewards/repeat_floor/std': 0.002030999632552266, 'rewards/near_duplicate_penalty/mean': 0.9935485124588013, 'rewards/near_duplicate_penalty/std': 0.013540077954530716, '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': 0.09714251756668091, 'rewards/total_composite/std': 0.12042219936847687, 'reward': 0.09714251756668091, 'reward_std': 0.12042219936847687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2302437573671341, 'sampling/sampling_logp_difference/max': 1.4848899841308594, 'sampling/importance_sampling_ratio/min': 0.22652727365493774, 'sampling/importance_sampling_ratio/mean': 1.0109827518463135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2434477508068085, 'clip_ratio/low_mean': 0.1374474922195077, 'clip_ratio/low_min': 0.1374474922195077, 'clip_ratio/high_mean': 0.07342592626810074, 'clip_ratio/high_max': 0.07342592626810074, 'clip_ratio/region_mean': 0.21087341848760843, 'reward_total_mean': 0.09714251756668091, 'reward_meter_mean': 0.17972107231616974, 'reward_meter_std': 0.2734915316104889, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9935485124588013, 'reward_repeat_penalty_std': 0.013540077954530716, 'reward_repeat_floor_mean': 0.9990322589874268, 'reward_repeat_floor_std': 0.002030999632552266, 'reward_near_duplicate_penalty_mean': 0.9935485124588013, 'reward_near_duplicate_penalty_std': 0.013540077954530716, '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': 0.2557307481765747, 'reward_total_composite_mean': 0.09714251756668091, 'reward_total_composite_std': 0.12042219936847687, 'epoch': 0.0}

  2%|▏         | 50/3300 [09:29<9:43:34, 10.77s/it]INFO 04-12 15:34:49 [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<01:05,  5.94s/it]

 23%|██▎       | 3/13 [00:20<01:11,  7.18s/it]

 31%|███       | 4/13 [00:36<01:34, 10.48s/it]

 38%|███▊      | 5/13 [00:49<01:30, 11.26s/it]

 46%|████▌     | 6/13 [01:03<01:24, 12.08s/it]

 54%|█████▍    | 7/13 [01:14<01:10, 11.70s/it]

 62%|██████▏   | 8/13 [01:28<01:02, 12.58s/it]

 69%|██████▉   | 9/13 [01:39<00:48, 12.16s/it]

 77%|███████▋  | 10/13 [01:50<00:35, 11.79s/it]

 85%|████████▍ | 11/13 [02:04<00:24, 12.42s/it]

 92%|█████████▏| 12/13 [02:16<00:12, 12.34s/it]

100%|██████████| 13/13 [02:30<00:00, 12.66s/it]2026-04-12 15:37:32,865 | INFO | train_grpo_train | metrics_logged mode=eval step=50

                                                   

                                               
{'eval_loss': nan, 'eval_runtime': 163.655, 'eval_samples_per_second': 0.635, 'eval_steps_per_second': 0.079, 'eval_num_tokens': 114218.0, 'eval_completions/mean_length': 175.16346153846155, 'eval_completions/min_length': 38.23076923076923, 'eval_completions/max_length': 416.61538461538464, 'eval_completions/clipped_ratio': 0.04807692307692308, 'eval_completions/mean_terminated_length': 158.04258493276743, 'eval_completions/min_terminated_length': 38.23076923076923, 'eval_completions/max_terminated_length': 333.9230769230769, 'eval_rewards/meter/mean': 0.31429795691600215, 'eval_rewards/meter/std': 0.32833211009319013, 'eval_rewards/count_adherence/mean': 0.9764095040468069, 'eval_rewards/count_adherence/std': 0.04316165212255258, 'eval_rewards/arabic_clean/mean': 0.9230769230769231, 'eval_rewards/arabic_clean/std': 0.1987869510283837, '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': 0.992922842502594, 'eval_rewards/count_floor/std': 0.01294849870296625, 'eval_rewards/lexical_plausibility/mean': 0.9796242438829862, 'eval_rewards/lexical_plausibility/std': 0.045520682783367544, 'eval_rewards/judge_quality/mean': 0.36173075896043044, 'eval_rewards/judge_quality/std': 0.1917116017295764, 'eval_rewards/repeat_penalty/mean': 0.9708329714261569, 'eval_rewards/repeat_penalty/std': 0.0467274531077307, 'eval_rewards/repeat_floor/mean': 0.9970654936937186, 'eval_rewards/repeat_floor/std': 0.004295328734979893, 'eval_rewards/near_duplicate_penalty/mean': 0.989183797286107, 'eval_rewards/near_duplicate_penalty/std': 0.014242813295613114, 'eval_rewards/opening_diversity/mean': 0.9921875, 'eval_rewards/opening_diversity/std': 0.022097086533904076, 'eval_rewards/distinct_2/mean': 0.9906296409093417, 'eval_rewards/distinct_2/std': 0.019931892918136258, 'eval_rewards/total_composite/mean': 0.10798603410904224, 'eval_rewards/total_composite/std': 0.12781854145801985, 'eval_reward': 0.10798603410904224, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15063365606161264, 'eval_sampling/sampling_logp_difference/max': 1.2357948743380034, 'eval_sampling/importance_sampling_ratio/min': 0.29498302936553955, 'eval_sampling/importance_sampling_ratio/mean': 1.0426076375521147, 'eval_sampling/importance_sampling_ratio/max': 1.624151321557852, 'eval_entropy': 2.499734410872826, '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': 0.10798603410904224, 'eval_reward_meter_mean': 0.31429795691600215, 'eval_reward_meter_std': 0.32833211009319013, 'eval_reward_count_adherence_mean': 0.9764095040468069, 'eval_reward_count_adherence_std': 0.04316165212255258, 'eval_reward_arabic_clean_mean': 0.9230769230769231, 'eval_reward_arabic_clean_std': 0.1987869510283837, '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': 0.992922842502594, 'eval_reward_count_floor_std': 0.01294849870296625, 'eval_reward_lexical_plausibility_mean': 0.9796242438829862, 'eval_reward_lexical_plausibility_std': 0.045520682783367544, 'eval_reward_repeat_penalty_mean': 0.9708329714261569, 'eval_reward_repeat_penalty_std': 0.0467274531077307, 'eval_reward_repeat_floor_mean': 0.9970654936937186, 'eval_reward_repeat_floor_std': 0.004295328734979893, 'eval_reward_near_duplicate_penalty_mean': 0.989183797286107, 'eval_reward_near_duplicate_penalty_std': 0.014242813295613114, 'eval_reward_opening_diversity_mean': 0.9921875, 'eval_reward_opening_diversity_std': 0.022097086533904076, 'eval_reward_distinct_2_mean': 0.9906296409093417, 'eval_reward_distinct_2_std': 0.019931892918136258, 'eval_reward_judge_quality_mean': 0.36173075896043044, 'eval_reward_judge_quality_std': 0.1917116017295764, 'eval_reward_total_composite_mean': 0.10798603410904224, 'eval_reward_total_composite_std': 0.12781854145801985, 'epoch': 0.0}

  2%|▏         | 50/3300 [12:12<9:43:34, 10.77s/it]

100%|██████████| 13/13 [02:30<00:00, 12.66s/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 [12:27<55:04:09, 61.02s/it]2026-04-12 15:37:47,465 | INFO | train_grpo_train | metrics_logged mode=train step=51

                                                    
{'loss': 0.1796, 'grad_norm': 10.217948913574219, 'learning_rate': 9.84848484848485e-06, 'num_tokens': 116472.0, 'completions/mean_length': 108.75, 'completions/min_length': 65.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.2087373584508896, 'rewards/meter/std': 0.27834978699684143, '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': 0.9901316165924072, 'rewards/lexical_plausibility/std': 0.027912110090255737, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/repeat_penalty/mean': 0.9795576930046082, 'rewards/repeat_penalty/std': 0.03321705386042595, 'rewards/repeat_floor/mean': 0.997545599937439, 'rewards/repeat_floor/std': 0.003270600223913789, 'rewards/near_duplicate_penalty/mean': 0.9900475740432739, 'rewards/near_duplicate_penalty/std': 0.005164245143532753, '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': 0.07415324449539185, 'rewards/total_composite/std': 0.10081373155117035, 'reward': 0.07415324449539185, 'reward_std': 0.10081371665000916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2235293835401535, 'sampling/sampling_logp_difference/max': 1.7983970642089844, 'sampling/importance_sampling_ratio/min': 0.1995221972465515, 'sampling/importance_sampling_ratio/mean': 1.035662055015564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.48785662651062, 'clip_ratio/low_mean': 0.10533129796385765, 'clip_ratio/low_min': 0.10533129796385765, 'clip_ratio/high_mean': 0.07935485057532787, 'clip_ratio/high_max': 0.07935485057532787, 'clip_ratio/region_mean': 0.18468614853918552, 'reward_total_mean': 0.07415324449539185, 'reward_meter_mean': 0.2087373584508896, 'reward_meter_std': 0.27834978699684143, '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': 0.9901316165924072, 'reward_lexical_plausibility_std': 0.027912110090255737, 'reward_repeat_penalty_mean': 0.9795576930046082, 'reward_repeat_penalty_std': 0.03321705386042595, 'reward_repeat_floor_mean': 0.997545599937439, 'reward_repeat_floor_std': 0.003270600223913789, 'reward_near_duplicate_penalty_mean': 0.9900475740432739, 'reward_near_duplicate_penalty_std': 0.005164245143532753, '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': 0.3375000059604645, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.07415324449539185, 'reward_total_composite_std': 0.10081373155117035, 'epoch': 0.0}

  2%|▏         | 51/3300 [12:27<55:04:09, 61.02s/it]INFO 04-12 15:37:47 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 52/3300 [12:37<41:18:13, 45.78s/it]2026-04-12 15:37:57,683 | INFO | train_grpo_train | metrics_logged mode=train step=52

                                                    
{'loss': -0.0216, 'grad_norm': 24.365684509277344, 'learning_rate': 9.845454545454546e-06, 'num_tokens': 117883.0, 'completions/mean_length': 23.375, 'completions/min_length': 17.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.4259107708930969, 'rewards/meter/std': 0.39908939599990845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, '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': 0.19917435944080353, 'rewards/total_composite/std': 0.17407074570655823, 'reward': 0.19917435944080353, 'reward_std': 0.17407073080539703, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18808428943157196, 'sampling/sampling_logp_difference/max': 1.4671764373779297, 'sampling/importance_sampling_ratio/min': 0.23057562112808228, 'sampling/importance_sampling_ratio/mean': 1.0117603540420532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7564258873462677, 'clip_ratio/low_mean': 0.07030866295099258, 'clip_ratio/low_min': 0.07030866295099258, 'clip_ratio/high_mean': 0.07976588793098927, 'clip_ratio/high_max': 0.07976588793098927, 'clip_ratio/region_mean': 0.15007455088198185, 'reward_total_mean': 0.19917435944080353, 'reward_meter_mean': 0.4259107708930969, 'reward_meter_std': 0.39908939599990845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.75, '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': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.19917435944080353, 'reward_total_composite_std': 0.17407074570655823, 'epoch': 0.0}

  2%|▏         | 52/3300 [12:37<41:18:13, 45.78s/it]INFO 04-12 15:37:57 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 53/3300 [12:49<32:10:59, 35.68s/it]2026-04-12 15:38:09,803 | INFO | train_grpo_train | metrics_logged mode=train step=53

                                                    
{'loss': 0.0065, 'grad_norm': 7.9538373947143555, 'learning_rate': 9.842424242424243e-06, 'num_tokens': 120549.0, 'completions/mean_length': 154.25, 'completions/min_length': 105.0, 'completions/max_length': 196.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.25, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 196.0, 'rewards/meter/mean': 0.5221389532089233, 'rewards/meter/std': 0.36941981315612793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8817651867866516, 'rewards/lexical_plausibility/std': 0.1428799331188202, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.995241641998291, 'rewards/repeat_penalty/std': 0.005499041173607111, 'rewards/repeat_floor/mean': 0.9992862939834595, 'rewards/repeat_floor/std': 0.0008248515077866614, 'rewards/near_duplicate_penalty/mean': 0.995241641998291, 'rewards/near_duplicate_penalty/std': 0.005499041173607111, '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': 0.1543811410665512, 'rewards/total_composite/std': 0.11782321333885193, 'reward': 0.1543811410665512, 'reward_std': 0.11782321333885193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2051299810409546, 'sampling/sampling_logp_difference/max': 1.1630926132202148, 'sampling/importance_sampling_ratio/min': 0.3125181794166565, 'sampling/importance_sampling_ratio/mean': 1.04629647731781, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.848069667816162, 'clip_ratio/low_mean': 0.11896091513335705, 'clip_ratio/low_min': 0.11896091513335705, 'clip_ratio/high_mean': 0.07578452862799168, 'clip_ratio/high_max': 0.07578452862799168, 'clip_ratio/region_mean': 0.19474544376134872, 'reward_total_mean': 0.1543811410665512, 'reward_meter_mean': 0.5221389532089233, 'reward_meter_std': 0.36941981315612793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8817651867866516, 'reward_lexical_plausibility_std': 0.1428799331188202, 'reward_repeat_penalty_mean': 0.995241641998291, 'reward_repeat_penalty_std': 0.005499041173607111, 'reward_repeat_floor_mean': 0.9992862939834595, 'reward_repeat_floor_std': 0.0008248515077866614, 'reward_near_duplicate_penalty_mean': 0.995241641998291, 'reward_near_duplicate_penalty_std': 0.005499041173607111, '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': 0.09258200228214264, 'reward_total_composite_mean': 0.1543811410665512, 'reward_total_composite_std': 0.11782321333885193, 'epoch': 0.0}

  2%|▏         | 53/3300 [12:49<32:10:59, 35.68s/it]INFO 04-12 15:38:10 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 54/3300 [13:00<25:27:16, 28.23s/it]2026-04-12 15:38:20,650 | INFO | train_grpo_train | metrics_logged mode=train step=54

                                                    
{'loss': 0.0778, 'grad_norm': 13.751712799072266, 'learning_rate': 9.83939393939394e-06, 'num_tokens': 122237.0, 'completions/mean_length': 47.0, 'completions/min_length': 35.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.6743103265762329, 'rewards/meter/std': 0.37354472279548645, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9696969985961914, 'rewards/lexical_plausibility/std': 0.08570991456508636, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.990012526512146, 'rewards/repeat_penalty/std': 0.027903828769922256, 'rewards/repeat_floor/mean': 0.9990427494049072, 'rewards/repeat_floor/std': 0.0026558798272162676, 'rewards/near_duplicate_penalty/mean': 0.9991129636764526, 'rewards/near_duplicate_penalty/std': 0.0021831849589943886, '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': 0.2338332235813141, 'rewards/total_composite/std': 0.16309471428394318, 'reward': 0.2338332235813141, 'reward_std': 0.16309471428394318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23054999113082886, 'sampling/sampling_logp_difference/max': 1.2963180541992188, 'sampling/importance_sampling_ratio/min': 0.2735370695590973, 'sampling/importance_sampling_ratio/mean': 1.0494039058685303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.69976544380188, 'clip_ratio/low_mean': 0.11132789682596922, 'clip_ratio/low_min': 0.11132789682596922, 'clip_ratio/high_mean': 0.07033322565257549, 'clip_ratio/high_max': 0.07033322565257549, 'clip_ratio/region_mean': 0.1816611224785447, 'reward_total_mean': 0.2338332235813141, 'reward_meter_mean': 0.6743103265762329, 'reward_meter_std': 0.37354472279548645, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9696969985961914, 'reward_lexical_plausibility_std': 0.08570991456508636, 'reward_repeat_penalty_mean': 0.990012526512146, 'reward_repeat_penalty_std': 0.027903828769922256, 'reward_repeat_floor_mean': 0.9990427494049072, 'reward_repeat_floor_std': 0.0026558798272162676, 'reward_near_duplicate_penalty_mean': 0.9991129636764526, 'reward_near_duplicate_penalty_std': 0.0021831849589943886, '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': 0.375, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.2338332235813141, 'reward_total_composite_std': 0.16309471428394318, 'epoch': 0.0}

  2%|▏         | 54/3300 [13:00<25:27:16, 28.23s/it]INFO 04-12 15:38:20 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 55/3300 [13:16<22:05:18, 24.50s/it]2026-04-12 15:38:36,459 | INFO | train_grpo_train | metrics_logged mode=train step=55

                                                    
{'loss': 0.117, 'grad_norm': 3.5803184509277344, 'learning_rate': 9.836363636363637e-06, 'num_tokens': 126215.0, 'completions/mean_length': 416.25, 'completions/min_length': 259.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 384.3333435058594, 'completions/min_terminated_length': 259.0, 'completions/max_terminated_length': 503.0, 'rewards/meter/mean': 0.3558233380317688, 'rewards/meter/std': 0.33072230219841003, 'rewards/count_adherence/mean': 0.8671875, 'rewards/count_adherence/std': 0.0521576851606369, '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': 0.9601562023162842, 'rewards/count_floor/std': 0.015647288411855698, 'rewards/lexical_plausibility/mean': 0.9793447256088257, 'rewards/lexical_plausibility/std': 0.04138510674238205, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/repeat_penalty/mean': 0.989062488079071, 'rewards/repeat_penalty/std': 0.009795082733035088, 'rewards/repeat_floor/mean': 0.9987688064575195, 'rewards/repeat_floor/std': 0.0009688546415418386, 'rewards/near_duplicate_penalty/mean': 0.9959406852722168, 'rewards/near_duplicate_penalty/std': 0.0019094451563432813, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.993085503578186, 'rewards/distinct_2/std': 0.008636143989861012, 'rewards/total_composite/mean': 0.06600721180438995, 'rewards/total_composite/std': 0.06121479719877243, 'reward': 0.06600721180438995, 'reward_std': 0.06121479719877243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23353834450244904, 'sampling/sampling_logp_difference/max': 1.7532901763916016, 'sampling/importance_sampling_ratio/min': 0.17320314049720764, 'sampling/importance_sampling_ratio/mean': 1.0502129793167114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0999017655849457, 'clip_ratio/low_mean': 0.08828209713101387, 'clip_ratio/low_min': 0.08828209713101387, 'clip_ratio/high_mean': 0.056370774284005165, 'clip_ratio/high_max': 0.056370774284005165, 'clip_ratio/region_mean': 0.14465287141501904, 'reward_total_mean': 0.06600721180438995, 'reward_meter_mean': 0.3558233380317688, 'reward_meter_std': 0.33072230219841003, 'reward_count_adherence_mean': 0.8671875, 'reward_count_adherence_std': 0.0521576851606369, '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': 0.9601562023162842, 'reward_count_floor_std': 0.015647288411855698, 'reward_lexical_plausibility_mean': 0.9793447256088257, 'reward_lexical_plausibility_std': 0.04138510674238205, 'reward_repeat_penalty_mean': 0.989062488079071, 'reward_repeat_penalty_std': 0.009795082733035088, 'reward_repeat_floor_mean': 0.9987688064575195, 'reward_repeat_floor_std': 0.0009688546415418386, 'reward_near_duplicate_penalty_mean': 0.9959406852722168, 'reward_near_duplicate_penalty_std': 0.0019094451563432813, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.993085503578186, 'reward_distinct_2_std': 0.008636143989861012, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.06600721180438995, 'reward_total_composite_std': 0.06121479719877243, 'epoch': 0.0}

  2%|▏         | 55/3300 [13:16<22:05:18, 24.50s/it]INFO 04-12 15:38:36 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 56/3300 [13:25<18:00:35, 19.99s/it]2026-04-12 15:38:45,901 | INFO | train_grpo_train | metrics_logged mode=train step=56

                                                    
{'loss': 0.1691, 'grad_norm': 15.630742073059082, 'learning_rate': 9.833333333333333e-06, 'num_tokens': 127853.0, 'completions/mean_length': 44.75, 'completions/min_length': 34.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8525811433792114, 'rewards/meter/std': 0.29226991534233093, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9888150691986084, 'rewards/lexical_plausibility/std': 0.022782549262046814, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.2796139121055603, 'rewards/repeat_penalty/mean': 0.9708560705184937, 'rewards/repeat_penalty/std': 0.052419934421777725, 'rewards/repeat_floor/mean': 0.996009349822998, 'rewards/repeat_floor/std': 0.006806257180869579, 'rewards/near_duplicate_penalty/mean': 0.9784122705459595, 'rewards/near_duplicate_penalty/std': 0.03164125978946686, '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': 0.3470025658607483, 'rewards/total_composite/std': 0.27790409326553345, 'reward': 0.3470025658607483, 'reward_std': 0.27790409326553345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21946589648723602, 'sampling/sampling_logp_difference/max': 1.6893329620361328, 'sampling/importance_sampling_ratio/min': 0.18464264273643494, 'sampling/importance_sampling_ratio/mean': 1.0393718481063843, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.309250071644783, 'clip_ratio/low_mean': 0.15036188997328281, 'clip_ratio/low_min': 0.15036188997328281, 'clip_ratio/high_mean': 0.049383943900465965, 'clip_ratio/high_max': 0.049383943900465965, 'clip_ratio/region_mean': 0.19974583387374878, 'reward_total_mean': 0.3470025658607483, 'reward_meter_mean': 0.8525811433792114, 'reward_meter_std': 0.29226991534233093, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9888150691986084, 'reward_lexical_plausibility_std': 0.022782549262046814, 'reward_repeat_penalty_mean': 0.9708560705184937, 'reward_repeat_penalty_std': 0.052419934421777725, 'reward_repeat_floor_mean': 0.996009349822998, 'reward_repeat_floor_std': 0.006806257180869579, 'reward_near_duplicate_penalty_mean': 0.9784122705459595, 'reward_near_duplicate_penalty_std': 0.03164125978946686, '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': 0.4087499976158142, 'reward_judge_quality_std': 0.2796139121055603, 'reward_total_composite_mean': 0.3470025658607483, 'reward_total_composite_std': 0.27790409326553345, 'epoch': 0.0}

  2%|▏         | 56/3300 [13:25<18:00:35, 19.99s/it]INFO 04-12 15:38:46 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 57/3300 [13:36<15:28:52, 17.19s/it]2026-04-12 15:38:56,557 | INFO | train_grpo_train | metrics_logged mode=train step=57

                                                    
{'loss': 0.039, 'grad_norm': 10.930910110473633, 'learning_rate': 9.830303030303032e-06, 'num_tokens': 129740.0, 'completions/mean_length': 64.875, 'completions/min_length': 42.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8454275727272034, 'rewards/meter/std': 0.3148094713687897, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9343099594116211, 'rewards/lexical_plausibility/std': 0.09829296171665192, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.268537700176239, 'rewards/repeat_penalty/mean': 0.993340015411377, 'rewards/repeat_penalty/std': 0.01577628031373024, 'rewards/repeat_floor/mean': 0.9990010261535645, 'rewards/repeat_floor/std': 0.0023664399050176144, 'rewards/near_duplicate_penalty/mean': 0.993340015411377, 'rewards/near_duplicate_penalty/std': 0.01577628031373024, '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': 0.3210640847682953, 'rewards/total_composite/std': 0.2879931926727295, 'reward': 0.3210640847682953, 'reward_std': 0.2879932224750519, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20600314438343048, 'sampling/sampling_logp_difference/max': 1.6941933631896973, 'sampling/importance_sampling_ratio/min': 0.3185790479183197, 'sampling/importance_sampling_ratio/mean': 1.0374006032943726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.514877498149872, 'clip_ratio/low_mean': 0.11329697445034981, 'clip_ratio/low_min': 0.11329697445034981, 'clip_ratio/high_mean': 0.12848603911697865, 'clip_ratio/high_max': 0.12848603911697865, 'clip_ratio/region_mean': 0.24178301356732845, 'reward_total_mean': 0.3210640847682953, 'reward_meter_mean': 0.8454275727272034, 'reward_meter_std': 0.3148094713687897, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9343099594116211, 'reward_lexical_plausibility_std': 0.09829296171665192, 'reward_repeat_penalty_mean': 0.993340015411377, 'reward_repeat_penalty_std': 0.01577628031373024, 'reward_repeat_floor_mean': 0.9990010261535645, 'reward_repeat_floor_std': 0.0023664399050176144, 'reward_near_duplicate_penalty_mean': 0.993340015411377, 'reward_near_duplicate_penalty_std': 0.01577628031373024, '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': 0.268537700176239, 'reward_total_composite_mean': 0.3210640847682953, 'reward_total_composite_std': 0.2879931926727295, 'epoch': 0.0}

  2%|▏         | 57/3300 [13:36<15:28:52, 17.19s/it]INFO 04-12 15:38:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 58/3300 [13:46<13:29:56, 14.99s/it]2026-04-12 15:39:06,416 | INFO | train_grpo_train | metrics_logged mode=train step=58

                                                    
{'loss': 0.1269, 'grad_norm': 15.922411918640137, 'learning_rate': 9.827272727272729e-06, 'num_tokens': 131445.0, 'completions/mean_length': 46.125, 'completions/min_length': 33.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.25518494844436646, 'rewards/meter/std': 0.3260885775089264, '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': 0.9750437140464783, 'rewards/repeat_penalty/std': 0.06004312261939049, 'rewards/repeat_floor/mean': 0.9966843128204346, 'rewards/repeat_floor/std': 0.007807428482919931, 'rewards/near_duplicate_penalty/mean': 0.9836643934249878, 'rewards/near_duplicate_penalty/std': 0.03597417101264, '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': 0.1558072566986084, 'rewards/total_composite/std': 0.26507842540740967, 'reward': 0.1558072566986084, 'reward_std': 0.26507842540740967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2382640540599823, 'sampling/sampling_logp_difference/max': 1.806563377380371, 'sampling/importance_sampling_ratio/min': 0.1642175316810608, 'sampling/importance_sampling_ratio/mean': 1.0559062957763672, 'sampling/importance_sampling_ratio/max': 1.9906350374221802, 'entropy': 2.646505504846573, 'clip_ratio/low_mean': 0.13163077924400568, 'clip_ratio/low_min': 0.13163077924400568, 'clip_ratio/high_mean': 0.0778409093618393, 'clip_ratio/high_max': 0.0778409093618393, 'clip_ratio/region_mean': 0.20947168860584497, 'reward_total_mean': 0.1558072566986084, 'reward_meter_mean': 0.25518494844436646, 'reward_meter_std': 0.3260885775089264, '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': 0.9750437140464783, 'reward_repeat_penalty_std': 0.06004312261939049, 'reward_repeat_floor_mean': 0.9966843128204346, 'reward_repeat_floor_std': 0.007807428482919931, 'reward_near_duplicate_penalty_mean': 0.9836643934249878, 'reward_near_duplicate_penalty_std': 0.03597417101264, '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': 0.42124998569488525, 'reward_judge_quality_std': 0.22177450358867645, 'reward_total_composite_mean': 0.1558072566986084, 'reward_total_composite_std': 0.26507842540740967, 'epoch': 0.0}

  2%|▏         | 58/3300 [13:46<13:29:56, 14.99s/it]INFO 04-12 15:39:06 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 59/3300 [13:55<11:45:23, 13.06s/it]2026-04-12 15:39:14,971 | INFO | train_grpo_train | metrics_logged mode=train step=59

                                                    
{'loss': 0.0084, 'grad_norm': 20.86791229248047, 'learning_rate': 9.824242424242425e-06, 'num_tokens': 132945.0, 'completions/mean_length': 32.5, 'completions/min_length': 27.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.5404990911483765, 'rewards/meter/std': 0.3343353271484375, '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': 0.9921875, 'rewards/lexical_plausibility/std': 0.022097086533904076, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.32275325059890747, 'rewards/repeat_penalty/mean': 0.9985888004302979, 'rewards/repeat_penalty/std': 0.00263221119530499, 'rewards/repeat_floor/mean': 0.9997882843017578, 'rewards/repeat_floor/std': 0.0003948377270717174, 'rewards/near_duplicate_penalty/mean': 0.9985888004302979, 'rewards/near_duplicate_penalty/std': 0.00263221119530499, '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': 0.44012579321861267, 'rewards/total_composite/std': 0.344654381275177, 'reward': 0.44012579321861267, 'reward_std': 0.344654381275177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22032026946544647, 'sampling/sampling_logp_difference/max': 2.4319496154785156, 'sampling/importance_sampling_ratio/min': 0.08786536008119583, 'sampling/importance_sampling_ratio/mean': 0.9949856996536255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.05552601441741, 'clip_ratio/low_mean': 0.08308576047420502, 'clip_ratio/low_min': 0.08308576047420502, 'clip_ratio/high_mean': 0.09974937606602907, 'clip_ratio/high_max': 0.09974937606602907, 'clip_ratio/region_mean': 0.1828351365402341, 'reward_total_mean': 0.44012579321861267, 'reward_meter_mean': 0.5404990911483765, 'reward_meter_std': 0.3343353271484375, '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': 0.9921875, 'reward_lexical_plausibility_std': 0.022097086533904076, 'reward_repeat_penalty_mean': 0.9985888004302979, 'reward_repeat_penalty_std': 0.00263221119530499, 'reward_repeat_floor_mean': 0.9997882843017578, 'reward_repeat_floor_std': 0.0003948377270717174, 'reward_near_duplicate_penalty_mean': 0.9985888004302979, 'reward_near_duplicate_penalty_std': 0.00263221119530499, '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': 0.6937500238418579, 'reward_judge_quality_std': 0.32275325059890747, 'reward_total_composite_mean': 0.44012579321861267, 'reward_total_composite_std': 0.344654381275177, 'epoch': 0.0}

  2%|▏         | 59/3300 [13:55<11:45:23, 13.06s/it]INFO 04-12 15:39:15 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 60/3300 [14:06<11:17:31, 12.55s/it]2026-04-12 15:39:26,323 | INFO | train_grpo_train | metrics_logged mode=train step=60

                                                    
{'loss': 0.1221, 'grad_norm': 10.750651359558105, 'learning_rate': 9.821212121212122e-06, 'num_tokens': 135227.0, 'completions/mean_length': 99.25, 'completions/min_length': 70.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.8280071020126343, 'rewards/meter/std': 0.2582924962043762, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.11877349019050598, 'rewards/repeat_penalty/mean': 0.9965113997459412, 'rewards/repeat_penalty/std': 0.0037187119014561176, 'rewards/repeat_floor/mean': 0.9994766712188721, 'rewards/repeat_floor/std': 0.0005578127456828952, 'rewards/near_duplicate_penalty/mean': 0.9965113997459412, 'rewards/near_duplicate_penalty/std': 0.0037187119014561176, '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': 0.222237229347229, 'rewards/total_composite/std': 0.10033319145441055, 'reward': 0.222237229347229, 'reward_std': 0.10033319145441055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21776022017002106, 'sampling/sampling_logp_difference/max': 1.4216468334197998, 'sampling/importance_sampling_ratio/min': 0.2413162887096405, 'sampling/importance_sampling_ratio/mean': 1.0483505725860596, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8715760707855225, 'clip_ratio/low_mean': 0.11581573635339737, 'clip_ratio/low_min': 0.11581573635339737, 'clip_ratio/high_mean': 0.0863803569227457, 'clip_ratio/high_max': 0.0863803569227457, 'clip_ratio/region_mean': 0.20219609327614307, 'reward_total_mean': 0.222237229347229, 'reward_meter_mean': 0.8280071020126343, 'reward_meter_std': 0.2582924962043762, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9965113997459412, 'reward_repeat_penalty_std': 0.0037187119014561176, 'reward_repeat_floor_mean': 0.9994766712188721, 'reward_repeat_floor_std': 0.0005578127456828952, 'reward_near_duplicate_penalty_mean': 0.9965113997459412, 'reward_near_duplicate_penalty_std': 0.0037187119014561176, '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': 0.11877349019050598, 'reward_total_composite_mean': 0.222237229347229, 'reward_total_composite_std': 0.10033319145441055, 'epoch': 0.0}

  2%|▏         | 60/3300 [14:06<11:17:31, 12.55s/it]INFO 04-12 15:39:26 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 61/3300 [14:18<11:08:50, 12.39s/it]2026-04-12 15:39:38,345 | INFO | train_grpo_train | metrics_logged mode=train step=61

                                                    
{'loss': 0.0213, 'grad_norm': 11.486438751220703, 'learning_rate': 9.81818181818182e-06, 'num_tokens': 137117.0, 'completions/mean_length': 71.25, 'completions/min_length': 48.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.39732348918914795, 'rewards/meter/std': 0.3453885316848755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.4712499976158142, 'rewards/judge_quality/std': 0.20951901376247406, 'rewards/repeat_penalty/mean': 0.9746606349945068, 'rewards/repeat_penalty/std': 0.03274624049663544, 'rewards/repeat_floor/mean': 0.9970314502716064, 'rewards/repeat_floor/std': 0.0034739612601697445, 'rewards/near_duplicate_penalty/mean': 0.9889459013938904, 'rewards/near_duplicate_penalty/std': 0.010812339372932911, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9854395389556885, 'rewards/distinct_2/std': 0.027018221095204353, 'rewards/total_composite/mean': 0.22805894911289215, 'rewards/total_composite/std': 0.27717325091362, 'reward': 0.22805894911289215, 'reward_std': 0.2771732211112976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19503915309906006, 'sampling/sampling_logp_difference/max': 1.6394710540771484, 'sampling/importance_sampling_ratio/min': 0.19408267736434937, 'sampling/importance_sampling_ratio/mean': 1.0252107381820679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.917788788676262, 'clip_ratio/low_mean': 0.10975828021764755, 'clip_ratio/low_min': 0.10975828021764755, 'clip_ratio/high_mean': 0.07531298883259296, 'clip_ratio/high_max': 0.07531298883259296, 'clip_ratio/region_mean': 0.18507126905024052, 'reward_total_mean': 0.22805894911289215, 'reward_meter_mean': 0.39732348918914795, 'reward_meter_std': 0.3453885316848755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9746606349945068, 'reward_repeat_penalty_std': 0.03274624049663544, 'reward_repeat_floor_mean': 0.9970314502716064, 'reward_repeat_floor_std': 0.0034739612601697445, 'reward_near_duplicate_penalty_mean': 0.9889459013938904, 'reward_near_duplicate_penalty_std': 0.010812339372932911, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9854395389556885, 'reward_distinct_2_std': 0.027018221095204353, 'reward_judge_quality_mean': 0.4712499976158142, 'reward_judge_quality_std': 0.20951901376247406, 'reward_total_composite_mean': 0.22805894911289215, 'reward_total_composite_std': 0.27717325091362, 'epoch': 0.0}

  2%|▏         | 61/3300 [14:18<11:08:50, 12.39s/it]INFO 04-12 15:39:38 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 62/3300 [14:29<10:47:50, 12.00s/it]2026-04-12 15:39:49,455 | INFO | train_grpo_train | metrics_logged mode=train step=62

                                                    
{'loss': 0.1336, 'grad_norm': 8.1593599319458, 'learning_rate': 9.815151515151516e-06, 'num_tokens': 139449.0, 'completions/mean_length': 126.5, 'completions/min_length': 72.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.5, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.4721960127353668, 'rewards/meter/std': 0.36472707986831665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9060581922531128, 'rewards/lexical_plausibility/std': 0.07539358735084534, 'rewards/judge_quality/mean': 0.23750001192092896, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9983223676681519, 'rewards/repeat_penalty/std': 0.002827560994774103, 'rewards/repeat_floor/mean': 0.9997483491897583, 'rewards/repeat_floor/std': 0.00042413041228428483, 'rewards/near_duplicate_penalty/mean': 0.9983223676681519, 'rewards/near_duplicate_penalty/std': 0.002827560994774103, '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': 0.12167593836784363, 'rewards/total_composite/std': 0.114832803606987, 'reward': 0.12167593836784363, 'reward_std': 0.1148327961564064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21668383479118347, 'sampling/sampling_logp_difference/max': 2.096221923828125, 'sampling/importance_sampling_ratio/min': 0.1229199469089508, 'sampling/importance_sampling_ratio/mean': 1.0373650789260864, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5818874835968018, 'clip_ratio/low_mean': 0.13090897910296917, 'clip_ratio/low_min': 0.13090897910296917, 'clip_ratio/high_mean': 0.07697806879878044, 'clip_ratio/high_max': 0.07697806879878044, 'clip_ratio/region_mean': 0.2078870479017496, 'reward_total_mean': 0.12167593836784363, 'reward_meter_mean': 0.4721960127353668, 'reward_meter_std': 0.36472707986831665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9060581922531128, 'reward_lexical_plausibility_std': 0.07539358735084534, 'reward_repeat_penalty_mean': 0.9983223676681519, 'reward_repeat_penalty_std': 0.002827560994774103, 'reward_repeat_floor_mean': 0.9997483491897583, 'reward_repeat_floor_std': 0.00042413041228428483, 'reward_near_duplicate_penalty_mean': 0.9983223676681519, 'reward_near_duplicate_penalty_std': 0.002827560994774103, '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': 0.09910311549901962, 'reward_total_composite_mean': 0.12167593836784363, 'reward_total_composite_std': 0.114832803606987, 'epoch': 0.0}

  2%|▏         | 62/3300 [14:29<10:47:50, 12.00s/it]INFO 04-12 15:39:49 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 63/3300 [14:39<10:15:42, 11.41s/it]2026-04-12 15:39:59,484 | INFO | train_grpo_train | metrics_logged mode=train step=63

                                                    
{'loss': 0.1348, 'grad_norm': 15.880659103393555, 'learning_rate': 9.812121212121212e-06, 'num_tokens': 141089.0, 'completions/mean_length': 42.0, 'completions/min_length': 30.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.3279654383659363, 'rewards/meter/std': 0.31591081619262695, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.9835828542709351, 'rewards/repeat_penalty/std': 0.02570706605911255, 'rewards/repeat_floor/mean': 0.9975374341011047, 'rewards/repeat_floor/std': 0.0038560586981475353, 'rewards/near_duplicate_penalty/mean': 0.9835828542709351, 'rewards/near_duplicate_penalty/std': 0.02570706605911255, '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': 0.1425551027059555, 'rewards/total_composite/std': 0.18220064043998718, 'reward': 0.1425551027059555, 'reward_std': 0.18220064043998718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22257696092128754, 'sampling/sampling_logp_difference/max': 2.073049545288086, 'sampling/importance_sampling_ratio/min': 0.12580156326293945, 'sampling/importance_sampling_ratio/mean': 1.0265105962753296, 'sampling/importance_sampling_ratio/max': 1.9564517736434937, 'entropy': 2.4436209201812744, 'clip_ratio/low_mean': 0.1373596116900444, 'clip_ratio/low_min': 0.1373596116900444, 'clip_ratio/high_mean': 0.04627193138003349, 'clip_ratio/high_max': 0.04627193138003349, 'clip_ratio/region_mean': 0.1836315430700779, 'reward_total_mean': 0.1425551027059555, 'reward_meter_mean': 0.3279654383659363, 'reward_meter_std': 0.31591081619262695, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9835828542709351, 'reward_repeat_penalty_std': 0.02570706605911255, 'reward_repeat_floor_mean': 0.9975374341011047, 'reward_repeat_floor_std': 0.0038560586981475353, 'reward_near_duplicate_penalty_mean': 0.9835828542709351, 'reward_near_duplicate_penalty_std': 0.02570706605911255, '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': 0.1425551027059555, 'reward_total_composite_std': 0.18220064043998718, 'epoch': 0.0}

  2%|▏         | 63/3300 [14:39<10:15:42, 11.41s/it]INFO 04-12 15:39:59 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 64/3300 [14:50<10:07:58, 11.27s/it]2026-04-12 15:40:10,429 | INFO | train_grpo_train | metrics_logged mode=train step=64

                                                    
{'loss': -0.0646, 'grad_norm': 8.46484661102295, 'learning_rate': 9.809090909090911e-06, 'num_tokens': 143675.0, 'completions/mean_length': 138.25, 'completions/min_length': 101.0, 'completions/max_length': 170.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.25, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 170.0, 'rewards/meter/mean': 0.9050223231315613, 'rewards/meter/std': 0.18551981449127197, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9393068552017212, 'rewards/lexical_plausibility/std': 0.095570407807827, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9850629568099976, 'rewards/repeat_penalty/std': 0.025699971243739128, 'rewards/repeat_floor/mean': 0.9983018636703491, 'rewards/repeat_floor/std': 0.0028515057638287544, 'rewards/near_duplicate_penalty/mean': 0.9943779706954956, 'rewards/near_duplicate_penalty/std': 0.008487415499985218, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.990501880645752, 'rewards/distinct_2/std': 0.0175889004021883, 'rewards/total_composite/mean': 0.13550083339214325, 'rewards/total_composite/std': 0.027694441378116608, 'reward': 0.13550083339214325, 'reward_std': 0.027694443240761757, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19049707055091858, 'sampling/sampling_logp_difference/max': 2.0743465423583984, 'sampling/importance_sampling_ratio/min': 0.12563849985599518, 'sampling/importance_sampling_ratio/mean': 1.0256162881851196, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9644009470939636, 'clip_ratio/low_mean': 0.023514851927757263, 'clip_ratio/low_min': 0.023514851927757263, 'clip_ratio/high_mean': 0.20162308029830456, 'clip_ratio/high_max': 0.20162308029830456, 'clip_ratio/region_mean': 0.22513793222606182, 'reward_total_mean': 0.13550083339214325, 'reward_meter_mean': 0.9050223231315613, 'reward_meter_std': 0.18551981449127197, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9393068552017212, 'reward_lexical_plausibility_std': 0.095570407807827, 'reward_repeat_penalty_mean': 0.9850629568099976, 'reward_repeat_penalty_std': 0.025699971243739128, 'reward_repeat_floor_mean': 0.9983018636703491, 'reward_repeat_floor_std': 0.0028515057638287544, 'reward_near_duplicate_penalty_mean': 0.9943779706954956, 'reward_near_duplicate_penalty_std': 0.008487415499985218, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.990501880645752, 'reward_distinct_2_std': 0.0175889004021883, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.13550083339214325, 'reward_total_composite_std': 0.027694441378116608, 'epoch': 0.0}

  2%|▏         | 64/3300 [14:50<10:07:58, 11.27s/it]INFO 04-12 15:40:10 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 65/3300 [15:01<10:04:21, 11.21s/it]2026-04-12 15:40:21,491 | INFO | train_grpo_train | metrics_logged mode=train step=65

                                                    
{'loss': 0.0429, 'grad_norm': 9.566802024841309, 'learning_rate': 9.806060606060607e-06, 'num_tokens': 145911.0, 'completions/mean_length': 86.5, 'completions/min_length': 59.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.26148083806037903, 'rewards/meter/std': 0.3580758571624756, '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': 0.9787825345993042, 'rewards/lexical_plausibility/std': 0.03981919214129448, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9905944466590881, 'rewards/repeat_penalty/std': 0.006443301681429148, 'rewards/repeat_floor/mean': 0.9985891580581665, 'rewards/repeat_floor/std': 0.0009664942626841366, 'rewards/near_duplicate_penalty/mean': 0.9905944466590881, 'rewards/near_duplicate_penalty/std': 0.006443301681429148, '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': 0.05834570527076721, 'rewards/total_composite/std': 0.08223804831504822, 'reward': 0.05834570527076721, 'reward_std': 0.08223804831504822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21089264750480652, 'sampling/sampling_logp_difference/max': 1.4266748428344727, 'sampling/importance_sampling_ratio/min': 0.2401059865951538, 'sampling/importance_sampling_ratio/mean': 1.0309840440750122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6042764484882355, 'clip_ratio/low_mean': 0.15297048538923264, 'clip_ratio/low_min': 0.15297048538923264, 'clip_ratio/high_mean': 0.04924657475203276, 'clip_ratio/high_max': 0.04924657475203276, 'clip_ratio/region_mean': 0.2022170601412654, 'reward_total_mean': 0.05834570527076721, 'reward_meter_mean': 0.26148083806037903, 'reward_meter_std': 0.3580758571624756, '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': 0.9787825345993042, 'reward_lexical_plausibility_std': 0.03981919214129448, 'reward_repeat_penalty_mean': 0.9905944466590881, 'reward_repeat_penalty_std': 0.006443301681429148, 'reward_repeat_floor_mean': 0.9985891580581665, 'reward_repeat_floor_std': 0.0009664942626841366, 'reward_near_duplicate_penalty_mean': 0.9905944466590881, 'reward_near_duplicate_penalty_std': 0.006443301681429148, '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': 0.05834570527076721, 'reward_total_composite_std': 0.08223804831504822, 'epoch': 0.0}

  2%|▏         | 65/3300 [15:01<10:04:21, 11.21s/it]INFO 04-12 15:40:21 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 66/3300 [15:10<9:20:44, 10.40s/it] 2026-04-12 15:40:30,014 | INFO | train_grpo_train | metrics_logged mode=train step=66

                                                   
{'loss': 0.0282, 'grad_norm': 11.81692886352539, 'learning_rate': 9.803030303030304e-06, 'num_tokens': 147708.0, 'completions/mean_length': 44.625, 'completions/min_length': 44.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9933011531829834, 'rewards/meter/std': 0.011455508880317211, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.949999988079071, 'rewards/judge_quality/std': 0.0, 'rewards/repeat_penalty/mean': 0.9623221158981323, 'rewards/repeat_penalty/std': 0.020256616175174713, 'rewards/repeat_floor/mean': 0.9943482875823975, 'rewards/repeat_floor/std': 0.0030385067220777273, 'rewards/near_duplicate_penalty/mean': 0.9623221158981323, 'rewards/near_duplicate_penalty/std': 0.020256616175174713, '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': 0.9382868409156799, 'rewards/total_composite/std': 0.009567228145897388, 'reward': 0.9382868409156799, 'reward_std': 0.009567230939865112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0628151223063469, 'sampling/sampling_logp_difference/max': 1.1648931503295898, 'sampling/importance_sampling_ratio/min': 0.311955988407135, 'sampling/importance_sampling_ratio/mean': 0.9966855049133301, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24568559974431992, 'clip_ratio/low_mean': 0.013586956076323986, 'clip_ratio/low_min': 0.013586956076323986, 'clip_ratio/high_mean': 0.044823233503848314, 'clip_ratio/high_max': 0.044823233503848314, 'clip_ratio/region_mean': 0.0584101895801723, 'reward_total_mean': 0.9382868409156799, 'reward_meter_mean': 0.9933011531829834, 'reward_meter_std': 0.011455508880317211, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9623221158981323, 'reward_repeat_penalty_std': 0.020256616175174713, 'reward_repeat_floor_mean': 0.9943482875823975, 'reward_repeat_floor_std': 0.0030385067220777273, 'reward_near_duplicate_penalty_mean': 0.9623221158981323, 'reward_near_duplicate_penalty_std': 0.020256616175174713, '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': 0.949999988079071, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.9382868409156799, 'reward_total_composite_std': 0.009567228145897388, 'epoch': 0.0}

  2%|▏         | 66/3300 [15:10<9:20:44, 10.40s/it]INFO 04-12 15:40:30 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 67/3300 [15:20<9:26:30, 10.51s/it]2026-04-12 15:40:40,784 | INFO | train_grpo_train | metrics_logged mode=train step=67

                                                   
{'loss': 0.2793, 'grad_norm': 15.087583541870117, 'learning_rate': 9.800000000000001e-06, 'num_tokens': 149338.0, 'completions/mean_length': 45.75, 'completions/min_length': 28.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.26053667068481445, 'rewards/meter/std': 0.4058288037776947, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.29232752323150635, 'rewards/repeat_penalty/mean': 0.9497696757316589, 'rewards/repeat_penalty/std': 0.030571458861231804, 'rewards/repeat_floor/mean': 0.9924654960632324, 'rewards/repeat_floor/std': 0.004585713148117065, 'rewards/near_duplicate_penalty/mean': 0.9497696757316589, 'rewards/near_duplicate_penalty/std': 0.030571458861231804, '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': 0.11962468922138214, 'rewards/total_composite/std': 0.16466636955738068, 'reward': 0.11962468922138214, 'reward_std': 0.16466636955738068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2025369107723236, 'sampling/sampling_logp_difference/max': 1.726677417755127, 'sampling/importance_sampling_ratio/min': 0.17787444591522217, 'sampling/importance_sampling_ratio/mean': 1.036473035812378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.786854311823845, 'clip_ratio/low_mean': 0.08405422745272517, 'clip_ratio/low_min': 0.08405422745272517, 'clip_ratio/high_mean': 0.09464126825332642, 'clip_ratio/high_max': 0.09464126825332642, 'clip_ratio/region_mean': 0.1786954957060516, 'reward_total_mean': 0.11962468922138214, 'reward_meter_mean': 0.26053667068481445, 'reward_meter_std': 0.4058288037776947, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9497696757316589, 'reward_repeat_penalty_std': 0.030571458861231804, 'reward_repeat_floor_mean': 0.9924654960632324, 'reward_repeat_floor_std': 0.004585713148117065, 'reward_near_duplicate_penalty_mean': 0.9497696757316589, 'reward_near_duplicate_penalty_std': 0.030571458861231804, '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': 0.29232752323150635, 'reward_total_composite_mean': 0.11962468922138214, 'reward_total_composite_std': 0.16466636955738068, 'epoch': 0.0}

  2%|▏         | 67/3300 [15:20<9:26:30, 10.51s/it]INFO 04-12 15:40:41 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 68/3300 [15:30<9:13:29, 10.28s/it]2026-04-12 15:40:50,504 | INFO | train_grpo_train | metrics_logged mode=train step=68

                                                   
{'loss': 0.0546, 'grad_norm': 16.317995071411133, 'learning_rate': 9.796969696969698e-06, 'num_tokens': 150970.0, 'completions/mean_length': 49.0, 'completions/min_length': 32.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.24453140795230865, 'rewards/meter/std': 0.2780782878398895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9454997181892395, 'rewards/lexical_plausibility/std': 0.13229869306087494, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.25670650601387024, 'rewards/repeat_penalty/mean': 0.9586745500564575, 'rewards/repeat_penalty/std': 0.03985991328954697, 'rewards/repeat_floor/mean': 0.9938011765480042, 'rewards/repeat_floor/std': 0.005978988483548164, 'rewards/near_duplicate_penalty/mean': 0.9586745500564575, 'rewards/near_duplicate_penalty/std': 0.03985991328954697, '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': 0.04745898395776749, 'rewards/total_composite/std': 0.03720048442482948, 'reward': 0.04745898395776749, 'reward_std': 0.037200480699539185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2399846315383911, 'sampling/sampling_logp_difference/max': 1.7153024673461914, 'sampling/importance_sampling_ratio/min': 0.28902629017829895, 'sampling/importance_sampling_ratio/mean': 1.050329327583313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2161257565021515, 'clip_ratio/low_mean': 0.10240742564201355, 'clip_ratio/low_min': 0.10240742564201355, 'clip_ratio/high_mean': 0.0988451074808836, 'clip_ratio/high_max': 0.0988451074808836, 'clip_ratio/region_mean': 0.20125253312289715, 'reward_total_mean': 0.04745898395776749, 'reward_meter_mean': 0.24453140795230865, 'reward_meter_std': 0.2780782878398895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9454997181892395, 'reward_lexical_plausibility_std': 0.13229869306087494, 'reward_repeat_penalty_mean': 0.9586745500564575, 'reward_repeat_penalty_std': 0.03985991328954697, 'reward_repeat_floor_mean': 0.9938011765480042, 'reward_repeat_floor_std': 0.005978988483548164, 'reward_near_duplicate_penalty_mean': 0.9586745500564575, 'reward_near_duplicate_penalty_std': 0.03985991328954697, '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': 0.04745898395776749, 'reward_total_composite_std': 0.03720048442482948, 'epoch': 0.0}

  2%|▏         | 68/3300 [15:30<9:13:29, 10.28s/it]INFO 04-12 15:40:50 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 69/3300 [15:37<8:25:27,  9.39s/it]2026-04-12 15:40:57,817 | INFO | train_grpo_train | metrics_logged mode=train step=69

                                                   
{'loss': 0.1148, 'grad_norm': 23.034942626953125, 'learning_rate': 9.793939393939394e-06, 'num_tokens': 152365.0, 'completions/mean_length': 26.375, 'completions/min_length': 23.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8718483448028564, 'rewards/meter/std': 0.15568596124649048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 1.0, 'rewards/lexical_plausibility/std': 0.0, 'rewards/judge_quality/mean': 0.9087499976158142, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/repeat_penalty/mean': 0.982042670249939, 'rewards/repeat_penalty/std': 0.014344698749482632, 'rewards/repeat_floor/mean': 0.9973064064979553, 'rewards/repeat_floor/std': 0.0021517009008675814, 'rewards/near_duplicate_penalty/mean': 0.982042670249939, 'rewards/near_duplicate_penalty/std': 0.014344698749482632, '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': 0.8043801188468933, 'rewards/total_composite/std': 0.19763554632663727, 'reward': 0.8043801188468933, 'reward_std': 0.19763551652431488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06679823249578476, 'sampling/sampling_logp_difference/max': 1.2693195343017578, 'sampling/importance_sampling_ratio/min': 0.2810227870941162, 'sampling/importance_sampling_ratio/mean': 0.994265615940094, 'sampling/importance_sampling_ratio/max': 1.9343113899230957, 'entropy': 0.247740825638175, 'clip_ratio/low_mean': 0.0037878789007663727, 'clip_ratio/low_min': 0.0037878789007663727, 'clip_ratio/high_mean': 0.053493162617087364, 'clip_ratio/high_max': 0.053493162617087364, 'clip_ratio/region_mean': 0.05728104151785374, 'reward_total_mean': 0.8043801188468933, 'reward_meter_mean': 0.8718483448028564, 'reward_meter_std': 0.15568596124649048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.982042670249939, 'reward_repeat_penalty_std': 0.014344698749482632, 'reward_repeat_floor_mean': 0.9973064064979553, 'reward_repeat_floor_std': 0.0021517009008675814, 'reward_near_duplicate_penalty_mean': 0.982042670249939, 'reward_near_duplicate_penalty_std': 0.014344698749482632, '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': 0.9087499976158142, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.8043801188468933, 'reward_total_composite_std': 0.19763554632663727, 'epoch': 0.0}

  2%|▏         | 69/3300 [15:37<8:25:27,  9.39s/it]INFO 04-12 15:40:58 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 70/3300 [15:54<10:26:00, 11.63s/it]2026-04-12 15:41:14,678 | INFO | train_grpo_train | metrics_logged mode=train step=70

                                                    
{'loss': 0.1482, 'grad_norm': 3.9465537071228027, 'learning_rate': 9.790909090909093e-06, 'num_tokens': 156555.0, 'completions/mean_length': 431.75, 'completions/min_length': 291.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 405.0, 'completions/min_terminated_length': 291.0, 'completions/max_terminated_length': 507.0, 'rewards/meter/mean': 0.8073446750640869, 'rewards/meter/std': 0.13133519887924194, 'rewards/count_adherence/mean': 0.8653846383094788, 'rewards/count_adherence/std': 0.05439282953739166, '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': 0.9596153497695923, 'rewards/count_floor/std': 0.01631784252822399, 'rewards/lexical_plausibility/mean': 0.9824336767196655, 'rewards/lexical_plausibility/std': 0.04754986613988876, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.09425459802150726, 'rewards/repeat_penalty/mean': 0.9940862655639648, 'rewards/repeat_penalty/std': 0.005090711172670126, 'rewards/repeat_floor/mean': 0.999253511428833, 'rewards/repeat_floor/std': 0.0005315811140462756, 'rewards/near_duplicate_penalty/mean': 0.9964444041252136, 'rewards/near_duplicate_penalty/std': 0.0019520486239343882, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9976320266723633, 'rewards/distinct_2/std': 0.004385976120829582, 'rewards/total_composite/mean': 0.12153446674346924, 'rewards/total_composite/std': 0.08290655165910721, 'reward': 0.12153446674346924, 'reward_std': 0.08290655165910721, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2274150252342224, 'sampling/sampling_logp_difference/max': 1.5501213073730469, 'sampling/importance_sampling_ratio/min': 0.21222223341464996, 'sampling/importance_sampling_ratio/mean': 1.0611239671707153, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9055804908275604, 'clip_ratio/low_mean': 0.06976692005991936, 'clip_ratio/low_min': 0.06976692005991936, 'clip_ratio/high_mean': 0.04721180908381939, 'clip_ratio/high_max': 0.04721180908381939, 'clip_ratio/region_mean': 0.11697872914373875, 'reward_total_mean': 0.12153446674346924, 'reward_meter_mean': 0.8073446750640869, 'reward_meter_std': 0.13133519887924194, 'reward_count_adherence_mean': 0.8653846383094788, 'reward_count_adherence_std': 0.05439282953739166, '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': 0.9596153497695923, 'reward_count_floor_std': 0.01631784252822399, 'reward_lexical_plausibility_mean': 0.9824336767196655, 'reward_lexical_plausibility_std': 0.04754986613988876, 'reward_repeat_penalty_mean': 0.9940862655639648, 'reward_repeat_penalty_std': 0.005090711172670126, 'reward_repeat_floor_mean': 0.999253511428833, 'reward_repeat_floor_std': 0.0005315811140462756, 'reward_near_duplicate_penalty_mean': 0.9964444041252136, 'reward_near_duplicate_penalty_std': 0.0019520486239343882, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9976320266723633, 'reward_distinct_2_std': 0.004385976120829582, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.09425459802150726, 'reward_total_composite_mean': 0.12153446674346924, 'reward_total_composite_std': 0.08290655165910721, 'epoch': 0.0}

  2%|▏         | 70/3300 [15:54<10:26:00, 11.63s/it]INFO 04-12 15:41:14 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 71/3300 [16:04<9:55:32, 11.07s/it] 2026-04-12 15:41:24,430 | INFO | train_grpo_train | metrics_logged mode=train step=71

                                                   
{'loss': 0.0477, 'grad_norm': 13.31476879119873, 'learning_rate': 9.787878787878788e-06, 'num_tokens': 158053.0, 'completions/mean_length': 28.25, 'completions/min_length': 23.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.748448371887207, 'rewards/meter/std': 0.2588638663291931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.8858761787414551, 'rewards/repeat_penalty/std': 0.09615682810544968, 'rewards/repeat_floor/mean': 0.9891153573989868, 'rewards/repeat_floor/std': 0.008976382203400135, 'rewards/near_duplicate_penalty/mean': 0.9914911985397339, 'rewards/near_duplicate_penalty/std': 0.009008262306451797, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.8932409286499023, 'rewards/distinct_2/std': 0.09393897652626038, 'rewards/total_composite/mean': 0.642677366733551, 'rewards/total_composite/std': 0.2716851532459259, 'reward': 0.642677366733551, 'reward_std': 0.2716851532459259, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11929537355899811, 'sampling/sampling_logp_difference/max': 1.7623071670532227, 'sampling/importance_sampling_ratio/min': 0.17164838314056396, 'sampling/importance_sampling_ratio/mean': 1.0300935506820679, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8854544442147017, 'clip_ratio/low_mean': 0.0507183913141489, 'clip_ratio/low_min': 0.0507183913141489, 'clip_ratio/high_mean': 0.05932382959872484, 'clip_ratio/high_max': 0.05932382959872484, 'clip_ratio/region_mean': 0.11004222091287374, 'reward_total_mean': 0.642677366733551, 'reward_meter_mean': 0.748448371887207, 'reward_meter_std': 0.2588638663291931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.8858761787414551, 'reward_repeat_penalty_std': 0.09615682810544968, 'reward_repeat_floor_mean': 0.9891153573989868, 'reward_repeat_floor_std': 0.008976382203400135, 'reward_near_duplicate_penalty_mean': 0.9914911985397339, 'reward_near_duplicate_penalty_std': 0.009008262306451797, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.8932409286499023, 'reward_distinct_2_std': 0.09393897652626038, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.642677366733551, 'reward_total_composite_std': 0.2716851532459259, 'epoch': 0.0}

  2%|▏         | 71/3300 [16:04<9:55:32, 11.07s/it]INFO 04-12 15:41:24 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 72/3300 [16:16<10:09:22, 11.33s/it]2026-04-12 15:41:36,368 | INFO | train_grpo_train | metrics_logged mode=train step=72

                                                    
{'loss': -0.0528, 'grad_norm': 12.80850887298584, 'learning_rate': 9.784848484848486e-06, 'num_tokens': 159897.0, 'completions/mean_length': 54.5, 'completions/min_length': 42.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6490570306777954, 'rewards/meter/std': 0.3876386284828186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9399245977401733, 'rewards/lexical_plausibility/std': 0.11370430886745453, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9924507737159729, 'rewards/repeat_penalty/std': 0.01201132033020258, 'rewards/repeat_floor/mean': 0.9988676309585571, 'rewards/repeat_floor/std': 0.0018016992835327983, 'rewards/near_duplicate_penalty/mean': 0.9924507737159729, 'rewards/near_duplicate_penalty/std': 0.01201132033020258, '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': 0.20356422662734985, 'rewards/total_composite/std': 0.1556379348039627, 'reward': 0.20356422662734985, 'reward_std': 0.1556379348039627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19354785978794098, 'sampling/sampling_logp_difference/max': 1.5727412700653076, 'sampling/importance_sampling_ratio/min': 0.20747564733028412, 'sampling/importance_sampling_ratio/mean': 1.0351853370666504, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.354126736521721, 'clip_ratio/low_mean': 0.09373548906296492, 'clip_ratio/low_min': 0.09373548906296492, 'clip_ratio/high_mean': 0.06477318424731493, 'clip_ratio/high_max': 0.06477318424731493, 'clip_ratio/region_mean': 0.15850867331027985, 'reward_total_mean': 0.20356422662734985, 'reward_meter_mean': 0.6490570306777954, 'reward_meter_std': 0.3876386284828186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9399245977401733, 'reward_lexical_plausibility_std': 0.11370430886745453, 'reward_repeat_penalty_mean': 0.9924507737159729, 'reward_repeat_penalty_std': 0.01201132033020258, 'reward_repeat_floor_mean': 0.9988676309585571, 'reward_repeat_floor_std': 0.0018016992835327983, 'reward_near_duplicate_penalty_mean': 0.9924507737159729, 'reward_near_duplicate_penalty_std': 0.01201132033020258, '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': 0.20356422662734985, 'reward_total_composite_std': 0.1556379348039627, 'epoch': 0.0}

  2%|▏         | 72/3300 [16:16<10:09:22, 11.33s/it]INFO 04-12 15:41:36 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 73/3300 [16:26<9:52:02, 11.01s/it] 2026-04-12 15:41:46,630 | INFO | train_grpo_train | metrics_logged mode=train step=73

                                                   
{'loss': -0.0327, 'grad_norm': 14.568339347839355, 'learning_rate': 9.781818181818183e-06, 'num_tokens': 161446.0, 'completions/mean_length': 43.625, 'completions/min_length': 30.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.691016435623169, 'rewards/meter/std': 0.39511194825172424, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.13562026619911194, 'rewards/repeat_penalty/mean': 0.9733875393867493, 'rewards/repeat_penalty/std': 0.033725347369909286, 'rewards/repeat_floor/mean': 0.9960081577301025, 'rewards/repeat_floor/std': 0.005058803129941225, 'rewards/near_duplicate_penalty/mean': 0.9733875393867493, 'rewards/near_duplicate_penalty/std': 0.033725347369909286, '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': 0.2363760620355606, 'rewards/total_composite/std': 0.16948167979717255, 'reward': 0.2363760620355606, 'reward_std': 0.16948167979717255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22721387445926666, 'sampling/sampling_logp_difference/max': 1.0508317947387695, 'sampling/importance_sampling_ratio/min': 0.34964677691459656, 'sampling/importance_sampling_ratio/mean': 1.0778717994689941, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.77302822470665, 'clip_ratio/low_mean': 0.09030991233885288, 'clip_ratio/low_min': 0.09030991233885288, 'clip_ratio/high_mean': 0.10500114597380161, 'clip_ratio/high_max': 0.10500114597380161, 'clip_ratio/region_mean': 0.1953110583126545, 'reward_total_mean': 0.2363760620355606, 'reward_meter_mean': 0.691016435623169, 'reward_meter_std': 0.39511194825172424, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9733875393867493, 'reward_repeat_penalty_std': 0.033725347369909286, 'reward_repeat_floor_mean': 0.9960081577301025, 'reward_repeat_floor_std': 0.005058803129941225, 'reward_near_duplicate_penalty_mean': 0.9733875393867493, 'reward_near_duplicate_penalty_std': 0.033725347369909286, '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': 0.13562026619911194, 'reward_total_composite_mean': 0.2363760620355606, 'reward_total_composite_std': 0.16948167979717255, 'epoch': 0.0}

  2%|▏         | 73/3300 [16:26<9:52:02, 11.01s/it]INFO 04-12 15:41:46 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 74/3300 [16:36<9:25:51, 10.52s/it]2026-04-12 15:41:56,026 | INFO | train_grpo_train | metrics_logged mode=train step=74

                                                   
{'loss': 0.1199, 'grad_norm': 14.732952117919922, 'learning_rate': 9.77878787878788e-06, 'num_tokens': 163147.0, 'completions/mean_length': 44.625, 'completions/min_length': 38.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.2882691025733948, 'rewards/meter/std': 0.3729732632637024, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.45375001430511475, 'rewards/judge_quality/std': 0.17919962108135223, 'rewards/repeat_penalty/mean': 0.9798584580421448, 'rewards/repeat_penalty/std': 0.03511206805706024, 'rewards/repeat_floor/mean': 0.9974900484085083, 'rewards/repeat_floor/std': 0.00400077598169446, 'rewards/near_duplicate_penalty/mean': 0.9887615442276001, 'rewards/near_duplicate_penalty/std': 0.01619119755923748, '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': 0.15449298918247223, 'rewards/total_composite/std': 0.23706699907779694, 'reward': 0.15449298918247223, 'reward_std': 0.23706696927547455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21905067563056946, 'sampling/sampling_logp_difference/max': 1.7052898406982422, 'sampling/importance_sampling_ratio/min': 0.18171970546245575, 'sampling/importance_sampling_ratio/mean': 1.0297564268112183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2557131201028824, 'clip_ratio/low_mean': 0.14629038237035275, 'clip_ratio/low_min': 0.14629038237035275, 'clip_ratio/high_mean': 0.04858299531042576, 'clip_ratio/high_max': 0.04858299531042576, 'clip_ratio/region_mean': 0.1948733776807785, 'reward_total_mean': 0.15449298918247223, 'reward_meter_mean': 0.2882691025733948, 'reward_meter_std': 0.3729732632637024, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9798584580421448, 'reward_repeat_penalty_std': 0.03511206805706024, 'reward_repeat_floor_mean': 0.9974900484085083, 'reward_repeat_floor_std': 0.00400077598169446, 'reward_near_duplicate_penalty_mean': 0.9887615442276001, 'reward_near_duplicate_penalty_std': 0.01619119755923748, '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': 0.45375001430511475, 'reward_judge_quality_std': 0.17919962108135223, 'reward_total_composite_mean': 0.15449298918247223, 'reward_total_composite_std': 0.23706699907779694, 'epoch': 0.0}

  2%|▏         | 74/3300 [16:36<9:25:51, 10.52s/it]INFO 04-12 15:41:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 75/3300 [16:52<11:00:14, 12.28s/it]2026-04-12 15:42:12,416 | INFO | train_grpo_train | metrics_logged mode=train step=75

                                                    
{'loss': -0.0741, 'grad_norm': 3.5783498287200928, 'learning_rate': 9.775757575757576e-06, 'num_tokens': 166702.0, 'completions/mean_length': 490.375, 'completions/min_length': 424.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 468.75, 'completions/min_terminated_length': 424.0, 'completions/max_terminated_length': 493.0, 'rewards/meter/mean': 0.47243809700012207, 'rewards/meter/std': 0.24235592782497406, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.09977654367685318, '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': 0.9343750476837158, 'rewards/count_floor/std': 0.02993295155465603, 'rewards/lexical_plausibility/mean': 0.9230168461799622, 'rewards/lexical_plausibility/std': 0.09823575615882874, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/repeat_penalty/mean': 0.9928020238876343, 'rewards/repeat_penalty/std': 0.0057755764573812485, 'rewards/repeat_floor/mean': 0.9991464614868164, 'rewards/repeat_floor/std': 0.0005958321271464229, 'rewards/near_duplicate_penalty/mean': 0.9965962767601013, 'rewards/near_duplicate_penalty/std': 0.0015001635765656829, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.996188759803772, 'rewards/distinct_2/std': 0.004702497273683548, 'rewards/total_composite/mean': 0.05822772905230522, 'rewards/total_composite/std': 0.03356262668967247, 'reward': 0.05822772905230522, 'reward_std': 0.03356262668967247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23616062104701996, 'sampling/sampling_logp_difference/max': 1.720937728881836, 'sampling/importance_sampling_ratio/min': 0.17889831960201263, 'sampling/importance_sampling_ratio/mean': 1.0708262920379639, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.286005437374115, 'clip_ratio/low_mean': 0.03473265469074249, 'clip_ratio/low_min': 0.03473265469074249, 'clip_ratio/high_mean': 0.044642968103289604, 'clip_ratio/high_max': 0.044642968103289604, 'clip_ratio/region_mean': 0.0793756227940321, 'reward_total_mean': 0.05822772905230522, 'reward_meter_mean': 0.47243809700012207, 'reward_meter_std': 0.24235592782497406, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.09977654367685318, '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': 0.9343750476837158, 'reward_count_floor_std': 0.02993295155465603, 'reward_lexical_plausibility_mean': 0.9230168461799622, 'reward_lexical_plausibility_std': 0.09823575615882874, 'reward_repeat_penalty_mean': 0.9928020238876343, 'reward_repeat_penalty_std': 0.0057755764573812485, 'reward_repeat_floor_mean': 0.9991464614868164, 'reward_repeat_floor_std': 0.0005958321271464229, 'reward_near_duplicate_penalty_mean': 0.9965962767601013, 'reward_near_duplicate_penalty_std': 0.0015001635765656829, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.996188759803772, 'reward_distinct_2_std': 0.004702497273683548, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.05822772905230522, 'reward_total_composite_std': 0.03356262668967247, 'epoch': 0.0}

  2%|▏         | 75/3300 [16:52<11:00:14, 12.28s/it]INFO 04-12 15:42:12 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 76/3300 [17:03<10:45:47, 12.02s/it]2026-04-12 15:42:23,813 | INFO | train_grpo_train | metrics_logged mode=train step=76

                                                    
{'loss': -0.0453, 'grad_norm': 10.16315746307373, 'learning_rate': 9.772727272727273e-06, 'num_tokens': 168886.0, 'completions/mean_length': 84.0, 'completions/min_length': 68.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.18300165235996246, 'rewards/meter/std': 0.1538061946630478, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9643131494522095, 'rewards/lexical_plausibility/std': 0.0722564160823822, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.9956752061843872, 'rewards/repeat_penalty/std': 0.004371351096779108, 'rewards/repeat_floor/mean': 0.9993512630462646, 'rewards/repeat_floor/std': 0.0006557052838616073, 'rewards/near_duplicate_penalty/mean': 0.9956752061843872, 'rewards/near_duplicate_penalty/std': 0.004371351096779108, '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': 0.04776401072740555, 'rewards/total_composite/std': 0.05731336027383804, 'reward': 0.04776401072740555, 'reward_std': 0.05731336399912834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23373211920261383, 'sampling/sampling_logp_difference/max': 1.646296501159668, 'sampling/importance_sampling_ratio/min': 0.19276247918605804, 'sampling/importance_sampling_ratio/mean': 1.0429555177688599, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4562574326992035, 'clip_ratio/low_mean': 0.12433180119842291, 'clip_ratio/low_min': 0.12433180119842291, 'clip_ratio/high_mean': 0.05970418266952038, 'clip_ratio/high_max': 0.05970418266952038, 'clip_ratio/region_mean': 0.1840359838679433, 'reward_total_mean': 0.04776401072740555, 'reward_meter_mean': 0.18300165235996246, 'reward_meter_std': 0.1538061946630478, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9643131494522095, 'reward_lexical_plausibility_std': 0.0722564160823822, 'reward_repeat_penalty_mean': 0.9956752061843872, 'reward_repeat_penalty_std': 0.004371351096779108, 'reward_repeat_floor_mean': 0.9993512630462646, 'reward_repeat_floor_std': 0.0006557052838616073, 'reward_near_duplicate_penalty_mean': 0.9956752061843872, 'reward_near_duplicate_penalty_std': 0.004371351096779108, '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': 0.04776401072740555, 'reward_total_composite_std': 0.05731336027383804, 'epoch': 0.0}

  2%|▏         | 76/3300 [17:03<10:45:47, 12.02s/it]INFO 04-12 15:42:24 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 77/3300 [17:15<10:42:55, 11.97s/it]2026-04-12 15:42:35,678 | INFO | train_grpo_train | metrics_logged mode=train step=77

                                                    
{'loss': 0.1764, 'grad_norm': 11.057759284973145, 'learning_rate': 9.76969696969697e-06, 'num_tokens': 170779.0, 'completions/mean_length': 72.625, 'completions/min_length': 45.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.4394851624965668, 'rewards/meter/std': 0.4574509263038635, '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': 0.927527666091919, 'rewards/lexical_plausibility/std': 0.11085552722215652, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/repeat_penalty/mean': 0.995111882686615, 'rewards/repeat_penalty/std': 0.009327364154160023, 'rewards/repeat_floor/mean': 0.9992667436599731, 'rewards/repeat_floor/std': 0.001399110071361065, 'rewards/near_duplicate_penalty/mean': 0.995111882686615, 'rewards/near_duplicate_penalty/std': 0.009327364154160023, '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': 0.16453397274017334, 'rewards/total_composite/std': 0.18075242638587952, 'reward': 0.16453397274017334, 'reward_std': 0.18075241148471832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23954260349273682, 'sampling/sampling_logp_difference/max': 1.5823326110839844, 'sampling/importance_sampling_ratio/min': 0.2054952085018158, 'sampling/importance_sampling_ratio/mean': 1.0527844429016113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.435519188642502, 'clip_ratio/low_mean': 0.10869949869811535, 'clip_ratio/low_min': 0.10869949869811535, 'clip_ratio/high_mean': 0.1150579359382391, 'clip_ratio/high_max': 0.1150579359382391, 'clip_ratio/region_mean': 0.22375743463635445, 'reward_total_mean': 0.16453397274017334, 'reward_meter_mean': 0.4394851624965668, 'reward_meter_std': 0.4574509263038635, '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': 0.927527666091919, 'reward_lexical_plausibility_std': 0.11085552722215652, 'reward_repeat_penalty_mean': 0.995111882686615, 'reward_repeat_penalty_std': 0.009327364154160023, 'reward_repeat_floor_mean': 0.9992667436599731, 'reward_repeat_floor_std': 0.001399110071361065, 'reward_near_duplicate_penalty_mean': 0.995111882686615, 'reward_near_duplicate_penalty_std': 0.009327364154160023, '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': 0.16453397274017334, 'reward_total_composite_std': 0.18075242638587952, 'epoch': 0.0}

  2%|▏         | 77/3300 [17:15<10:42:55, 11.97s/it]INFO 04-12 15:42:35 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 78/3300 [17:28<10:48:33, 12.08s/it]2026-04-12 15:42:47,998 | INFO | train_grpo_train | metrics_logged mode=train step=78

                                                    
{'loss': 0.0939, 'grad_norm': 8.838769912719727, 'learning_rate': 9.766666666666667e-06, 'num_tokens': 173451.0, 'completions/mean_length': 128.0, 'completions/min_length': 94.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.0, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.758094072341919, 'rewards/meter/std': 0.31945016980171204, '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': 0.9223395586013794, 'rewards/lexical_plausibility/std': 0.09644214063882828, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9941157102584839, 'rewards/repeat_penalty/std': 0.005136305466294289, 'rewards/repeat_floor/mean': 0.999117374420166, 'rewards/repeat_floor/std': 0.0007704399758949876, 'rewards/near_duplicate_penalty/mean': 0.9941157102584839, 'rewards/near_duplicate_penalty/std': 0.005136305466294289, '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': 0.18528631329536438, 'rewards/total_composite/std': 0.13080137968063354, 'reward': 0.18528631329536438, 'reward_std': 0.13080137968063354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23385238647460938, 'sampling/sampling_logp_difference/max': 1.537796974182129, 'sampling/importance_sampling_ratio/min': 0.2148539274930954, 'sampling/importance_sampling_ratio/mean': 1.0435125827789307, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3352877497673035, 'clip_ratio/low_mean': 0.1070121843367815, 'clip_ratio/low_min': 0.1070121843367815, 'clip_ratio/high_mean': 0.07725760340690613, 'clip_ratio/high_max': 0.07725760340690613, 'clip_ratio/region_mean': 0.18426978774368763, 'reward_total_mean': 0.18528631329536438, 'reward_meter_mean': 0.758094072341919, 'reward_meter_std': 0.31945016980171204, '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': 0.9223395586013794, 'reward_lexical_plausibility_std': 0.09644214063882828, 'reward_repeat_penalty_mean': 0.9941157102584839, 'reward_repeat_penalty_std': 0.005136305466294289, 'reward_repeat_floor_mean': 0.999117374420166, 'reward_repeat_floor_std': 0.0007704399758949876, 'reward_near_duplicate_penalty_mean': 0.9941157102584839, 'reward_near_duplicate_penalty_std': 0.005136305466294289, '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': 0.22499999403953552, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.18528631329536438, 'reward_total_composite_std': 0.13080137968063354, 'epoch': 0.0}

  2%|▏         | 78/3300 [17:28<10:48:33, 12.08s/it]INFO 04-12 15:42:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 79/3300 [17:40<10:49:02, 12.09s/it]2026-04-12 15:43:00,117 | INFO | train_grpo_train | metrics_logged mode=train step=79

                                                    
{'loss': 0.1918, 'grad_norm': 7.480940341949463, 'learning_rate': 9.763636363636365e-06, 'num_tokens': 176286.0, 'completions/mean_length': 151.375, 'completions/min_length': 113.0, 'completions/max_length': 263.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.375, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 263.0, 'rewards/meter/mean': 0.505182147026062, 'rewards/meter/std': 0.33270955085754395, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/arabic_clean/mean': 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': 0.023145509883761406, 'rewards/lexical_plausibility/mean': 0.8953814506530762, 'rewards/lexical_plausibility/std': 0.1638709306716919, 'rewards/judge_quality/mean': 0.2499999850988388, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9931508302688599, 'rewards/repeat_penalty/std': 0.009013802744448185, 'rewards/repeat_floor/mean': 0.9991541504859924, 'rewards/repeat_floor/std': 0.0008856873610056937, 'rewards/near_duplicate_penalty/mean': 0.9961925745010376, 'rewards/near_duplicate_penalty/std': 0.003039491130039096, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9969475269317627, 'rewards/distinct_2/std': 0.008633782155811787, 'rewards/total_composite/mean': 0.15535502135753632, 'rewards/total_composite/std': 0.13758698105812073, 'reward': 0.15535502135753632, 'reward_std': 0.13758698105812073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23767443001270294, 'sampling/sampling_logp_difference/max': 1.6421737670898438, 'sampling/importance_sampling_ratio/min': 0.19355882704257965, 'sampling/importance_sampling_ratio/mean': 1.0542786121368408, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9226368367671967, 'clip_ratio/low_mean': 0.09021471440792084, 'clip_ratio/low_min': 0.09021471440792084, 'clip_ratio/high_mean': 0.10830162838101387, 'clip_ratio/high_max': 0.10830162838101387, 'clip_ratio/region_mean': 0.1985163427889347, 'reward_total_mean': 0.15535502135753632, 'reward_meter_mean': 0.505182147026062, 'reward_meter_std': 0.33270955085754395, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_arabic_clean_mean': 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': 0.023145509883761406, 'reward_lexical_plausibility_mean': 0.8953814506530762, 'reward_lexical_plausibility_std': 0.1638709306716919, 'reward_repeat_penalty_mean': 0.9931508302688599, 'reward_repeat_penalty_std': 0.009013802744448185, 'reward_repeat_floor_mean': 0.9991541504859924, 'reward_repeat_floor_std': 0.0008856873610056937, 'reward_near_duplicate_penalty_mean': 0.9961925745010376, 'reward_near_duplicate_penalty_std': 0.003039491130039096, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9969475269317627, 'reward_distinct_2_std': 0.008633782155811787, 'reward_judge_quality_mean': 0.2499999850988388, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.15535502135753632, 'reward_total_composite_std': 0.13758698105812073, 'epoch': 0.0}

  2%|▏         | 79/3300 [17:40<10:49:02, 12.09s/it]INFO 04-12 15:43:00 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 80/3300 [17:49<9:56:17, 11.11s/it] 2026-04-12 15:43:08,943 | INFO | train_grpo_train | metrics_logged mode=train step=80

                                                   
{'loss': 0.0313, 'grad_norm': 15.44277572631836, 'learning_rate': 9.760606060606062e-06, 'num_tokens': 177964.0, 'completions/mean_length': 44.75, 'completions/min_length': 36.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5841773748397827, 'rewards/meter/std': 0.4157775640487671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.974759578704834, 'rewards/lexical_plausibility/std': 0.028471704572439194, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.24121344089508057, 'rewards/repeat_penalty/mean': 0.99813312292099, 'rewards/repeat_penalty/std': 0.0023070746101439, 'rewards/repeat_floor/mean': 0.9997199773788452, 'rewards/repeat_floor/std': 0.0003460668376646936, 'rewards/near_duplicate_penalty/mean': 0.99813312292099, 'rewards/near_duplicate_penalty/std': 0.0023070746101439, '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': 0.2804443836212158, 'rewards/total_composite/std': 0.30718445777893066, 'reward': 0.2804443836212158, 'reward_std': 0.30718445777893066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22217977046966553, 'sampling/sampling_logp_difference/max': 1.2089881896972656, 'sampling/importance_sampling_ratio/min': 0.29849913716316223, 'sampling/importance_sampling_ratio/mean': 1.059512734413147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1421422958374023, 'clip_ratio/low_mean': 0.08887456730008125, 'clip_ratio/low_min': 0.08887456730008125, 'clip_ratio/high_mean': 0.10736673697829247, 'clip_ratio/high_max': 0.10736673697829247, 'clip_ratio/region_mean': 0.19624130427837372, 'reward_total_mean': 0.2804443836212158, 'reward_meter_mean': 0.5841773748397827, 'reward_meter_std': 0.4157775640487671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.974759578704834, 'reward_lexical_plausibility_std': 0.028471704572439194, 'reward_repeat_penalty_mean': 0.99813312292099, 'reward_repeat_penalty_std': 0.0023070746101439, 'reward_repeat_floor_mean': 0.9997199773788452, 'reward_repeat_floor_std': 0.0003460668376646936, 'reward_near_duplicate_penalty_mean': 0.99813312292099, 'reward_near_duplicate_penalty_std': 0.0023070746101439, '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': 0.24121344089508057, 'reward_total_composite_mean': 0.2804443836212158, 'reward_total_composite_std': 0.30718445777893066, 'epoch': 0.0}

  2%|▏         | 80/3300 [17:49<9:56:17, 11.11s/it]INFO 04-12 15:43:09 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 81/3300 [17:58<9:33:16, 10.69s/it]2026-04-12 15:43:18,638 | INFO | train_grpo_train | metrics_logged mode=train step=81

                                                   
{'loss': -0.0645, 'grad_norm': 14.706957817077637, 'learning_rate': 9.757575757575758e-06, 'num_tokens': 179701.0, 'completions/mean_length': 44.125, 'completions/min_length': 33.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.5484030246734619, 'rewards/meter/std': 0.44246095418930054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.8968328237533569, 'rewards/lexical_plausibility/std': 0.16266535222530365, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/repeat_penalty/mean': 0.9913541078567505, 'rewards/repeat_penalty/std': 0.012487656436860561, 'rewards/repeat_floor/mean': 0.998703122138977, 'rewards/repeat_floor/std': 0.0018731470918282866, 'rewards/near_duplicate_penalty/mean': 0.9913541078567505, 'rewards/near_duplicate_penalty/std': 0.012487656436860561, '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': 0.2909441292285919, 'rewards/total_composite/std': 0.30240848660469055, 'reward': 0.2909441292285919, 'reward_std': 0.30240851640701294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2205350548028946, 'sampling/sampling_logp_difference/max': 1.7228679656982422, 'sampling/importance_sampling_ratio/min': 0.17855332791805267, 'sampling/importance_sampling_ratio/mean': 1.0267966985702515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7641369998455048, 'clip_ratio/low_mean': 0.09226511046290398, 'clip_ratio/low_min': 0.09226511046290398, 'clip_ratio/high_mean': 0.09743781574070454, 'clip_ratio/high_max': 0.09743781574070454, 'clip_ratio/region_mean': 0.1897029262036085, 'reward_total_mean': 0.2909441292285919, 'reward_meter_mean': 0.5484030246734619, 'reward_meter_std': 0.44246095418930054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.8968328237533569, 'reward_lexical_plausibility_std': 0.16266535222530365, 'reward_repeat_penalty_mean': 0.9913541078567505, 'reward_repeat_penalty_std': 0.012487656436860561, 'reward_repeat_floor_mean': 0.998703122138977, 'reward_repeat_floor_std': 0.0018731470918282866, 'reward_near_duplicate_penalty_mean': 0.9913541078567505, 'reward_near_duplicate_penalty_std': 0.012487656436860561, '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': 0.2909441292285919, 'reward_total_composite_std': 0.30240848660469055, 'epoch': 0.0}

  2%|▏         | 81/3300 [17:58<9:33:16, 10.69s/it]INFO 04-12 15:43:18 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 82/3300 [18:08<9:12:10, 10.30s/it]2026-04-12 15:43:28,022 | INFO | train_grpo_train | metrics_logged mode=train step=82

                                                   
{'loss': 0.0241, 'grad_norm': 16.850309371948242, 'learning_rate': 9.754545454545455e-06, 'num_tokens': 181481.0, 'completions/mean_length': 59.5, 'completions/min_length': 52.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.5, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6107265949249268, 'rewards/meter/std': 0.2725120186805725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.9830554127693176, 'rewards/repeat_penalty/std': 0.03362467139959335, 'rewards/repeat_floor/mean': 0.9980804920196533, 'rewards/repeat_floor/std': 0.003408765187487006, 'rewards/near_duplicate_penalty/mean': 0.9936132431030273, 'rewards/near_duplicate_penalty/std': 0.010421634651720524, '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': 0.17677387595176697, 'rewards/total_composite/std': 0.08098450303077698, 'reward': 0.17677387595176697, 'reward_std': 0.08098450303077698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19317473471164703, 'sampling/sampling_logp_difference/max': 2.1691017150878906, 'sampling/importance_sampling_ratio/min': 0.11428023129701614, 'sampling/importance_sampling_ratio/mean': 0.9689168334007263, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.841888353228569, 'clip_ratio/low_mean': 0.06377946771681309, 'clip_ratio/low_min': 0.06377946771681309, 'clip_ratio/high_mean': 0.09316704235970974, 'clip_ratio/high_max': 0.09316704235970974, 'clip_ratio/region_mean': 0.15694651007652283, 'reward_total_mean': 0.17677387595176697, 'reward_meter_mean': 0.6107265949249268, 'reward_meter_std': 0.2725120186805725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9830554127693176, 'reward_repeat_penalty_std': 0.03362467139959335, 'reward_repeat_floor_mean': 0.9980804920196533, 'reward_repeat_floor_std': 0.003408765187487006, 'reward_near_duplicate_penalty_mean': 0.9936132431030273, 'reward_near_duplicate_penalty_std': 0.010421634651720524, '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': 0.32499998807907104, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.17677387595176697, 'reward_total_composite_std': 0.08098450303077698, 'epoch': 0.0}

  2%|▏         | 82/3300 [18:08<9:12:10, 10.30s/it]INFO 04-12 15:43:28 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 83/3300 [18:18<9:06:32, 10.19s/it]2026-04-12 15:43:37,977 | INFO | train_grpo_train | metrics_logged mode=train step=83

                                                   
{'loss': 0.0495, 'grad_norm': 19.120960235595703, 'learning_rate': 9.751515151515152e-06, 'num_tokens': 183033.0, 'completions/mean_length': 41.0, 'completions/min_length': 37.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5597180128097534, 'rewards/meter/std': 0.4198746383190155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.914467453956604, 'rewards/lexical_plausibility/std': 0.10147204250097275, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.13093072175979614, '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': 0.19074517488479614, 'rewards/total_composite/std': 0.17437240481376648, 'reward': 0.19074517488479614, 'reward_std': 0.17437238991260529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2425851970911026, 'sampling/sampling_logp_difference/max': 2.108829975128174, 'sampling/importance_sampling_ratio/min': 0.23900221288204193, 'sampling/importance_sampling_ratio/mean': 1.015384554862976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.782499670982361, 'clip_ratio/low_mean': 0.1307391026057303, 'clip_ratio/low_min': 0.1307391026057303, 'clip_ratio/high_mean': 0.04645378701388836, 'clip_ratio/high_max': 0.04645378701388836, 'clip_ratio/region_mean': 0.17719288961961865, 'reward_total_mean': 0.19074517488479614, 'reward_meter_mean': 0.5597180128097534, 'reward_meter_std': 0.4198746383190155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.914467453956604, 'reward_lexical_plausibility_std': 0.10147204250097275, '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': 0.3499999940395355, 'reward_judge_quality_std': 0.13093072175979614, 'reward_total_composite_mean': 0.19074517488479614, 'reward_total_composite_std': 0.17437240481376648, 'epoch': 0.0}

  3%|▎         | 83/3300 [18:18<9:06:32, 10.19s/it]INFO 04-12 15:43:38 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 84/3300 [18:28<9:03:45, 10.14s/it]2026-04-12 15:43:48,008 | INFO | train_grpo_train | metrics_logged mode=train step=84

                                                   
{'loss': -0.0664, 'grad_norm': 13.650961875915527, 'learning_rate': 9.74848484848485e-06, 'num_tokens': 184523.0, 'completions/mean_length': 45.25, 'completions/min_length': 29.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.525891900062561, 'rewards/meter/std': 0.45709413290023804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, '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': 0.6599999666213989, 'rewards/judge_quality/std': 0.2805097699165344, 'rewards/repeat_penalty/mean': 0.9931754469871521, 'rewards/repeat_penalty/std': 0.012115638703107834, 'rewards/repeat_floor/mean': 0.9989762902259827, 'rewards/repeat_floor/std': 0.0018173520220443606, 'rewards/near_duplicate_penalty/mean': 0.9931754469871521, 'rewards/near_duplicate_penalty/std': 0.012115638703107834, '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': 0.308012992143631, 'rewards/total_composite/std': 0.2992304563522339, 'reward': 0.308012992143631, 'reward_std': 0.2992304563522339, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1705971211194992, 'sampling/sampling_logp_difference/max': 1.1007606983184814, 'sampling/importance_sampling_ratio/min': 0.33261796832084656, 'sampling/importance_sampling_ratio/mean': 1.0181703567504883, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6049704849720001, 'clip_ratio/low_mean': 0.050076065585017204, 'clip_ratio/low_min': 0.050076065585017204, 'clip_ratio/high_mean': 0.1008609514683485, 'clip_ratio/high_max': 0.1008609514683485, 'clip_ratio/region_mean': 0.1509370170533657, 'reward_total_mean': 0.308012992143631, 'reward_meter_mean': 0.525891900062561, 'reward_meter_std': 0.45709413290023804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, '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': 0.9931754469871521, 'reward_repeat_penalty_std': 0.012115638703107834, 'reward_repeat_floor_mean': 0.9989762902259827, 'reward_repeat_floor_std': 0.0018173520220443606, 'reward_near_duplicate_penalty_mean': 0.9931754469871521, 'reward_near_duplicate_penalty_std': 0.012115638703107834, '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': 0.2805097699165344, 'reward_total_composite_mean': 0.308012992143631, 'reward_total_composite_std': 0.2992304563522339, 'epoch': 0.0}

  3%|▎         | 84/3300 [18:28<9:03:45, 10.14s/it]INFO 04-12 15:43:48 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 85/3300 [18:36<8:36:55,  9.65s/it]2026-04-12 15:43:56,493 | INFO | train_grpo_train | metrics_logged mode=train step=85

                                                   
{'loss': 0.1338, 'grad_norm': 11.762645721435547, 'learning_rate': 9.745454545454547e-06, 'num_tokens': 186131.0, 'completions/mean_length': 50.0, 'completions/min_length': 36.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.4237501919269562, 'rewards/meter/std': 0.3781985640525818, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9809688329696655, 'rewards/lexical_plausibility/std': 0.05382820963859558, 'rewards/judge_quality/mean': 0.5587500333786011, 'rewards/judge_quality/std': 0.30286434292793274, 'rewards/repeat_penalty/mean': 0.9983229041099548, 'rewards/repeat_penalty/std': 0.0031740295235067606, 'rewards/repeat_floor/mean': 0.9997484683990479, 'rewards/repeat_floor/std': 0.00047610781621187925, 'rewards/near_duplicate_penalty/mean': 0.9983229041099548, 'rewards/near_duplicate_penalty/std': 0.0031740295235067606, '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': 0.2398170679807663, 'rewards/total_composite/std': 0.28019580245018005, 'reward': 0.2398170679807663, 'reward_std': 0.28019580245018005, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19141238927841187, 'sampling/sampling_logp_difference/max': 1.9095420837402344, 'sampling/importance_sampling_ratio/min': 0.14814820885658264, 'sampling/importance_sampling_ratio/mean': 1.0343586206436157, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.041260287165642, 'clip_ratio/low_mean': 0.1227434640750289, 'clip_ratio/low_min': 0.1227434640750289, 'clip_ratio/high_mean': 0.04827235825359821, 'clip_ratio/high_max': 0.04827235825359821, 'clip_ratio/region_mean': 0.1710158223286271, 'reward_total_mean': 0.2398170679807663, 'reward_meter_mean': 0.4237501919269562, 'reward_meter_std': 0.3781985640525818, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9809688329696655, 'reward_lexical_plausibility_std': 0.05382820963859558, 'reward_repeat_penalty_mean': 0.9983229041099548, 'reward_repeat_penalty_std': 0.0031740295235067606, 'reward_repeat_floor_mean': 0.9997484683990479, 'reward_repeat_floor_std': 0.00047610781621187925, 'reward_near_duplicate_penalty_mean': 0.9983229041099548, 'reward_near_duplicate_penalty_std': 0.0031740295235067606, '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': 0.30286434292793274, 'reward_total_composite_mean': 0.2398170679807663, 'reward_total_composite_std': 0.28019580245018005, 'epoch': 0.0}

  3%|▎         | 85/3300 [18:36<8:36:55,  9.65s/it]INFO 04-12 15:43:56 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 86/3300 [18:46<8:41:57,  9.74s/it]2026-04-12 15:44:06,466 | INFO | train_grpo_train | metrics_logged mode=train step=86

                                                   
{'loss': 0.1462, 'grad_norm': 14.28665828704834, 'learning_rate': 9.742424242424244e-06, 'num_tokens': 187884.0, 'completions/mean_length': 50.125, 'completions/min_length': 34.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9526747465133667, 'rewards/meter/std': 0.08271235227584839, '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': 0.9904210567474365, 'rewards/repeat_penalty/std': 0.011389861814677715, 'rewards/repeat_floor/mean': 0.9985631704330444, 'rewards/repeat_floor/std': 0.001708473195321858, 'rewards/near_duplicate_penalty/mean': 0.9904210567474365, 'rewards/near_duplicate_penalty/std': 0.011389861814677715, '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': 0.31550276279449463, 'rewards/total_composite/std': 0.12497393041849136, 'reward': 0.31550276279449463, 'reward_std': 0.12497393041849136, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22854943573474884, 'sampling/sampling_logp_difference/max': 1.1131696701049805, 'sampling/importance_sampling_ratio/min': 0.32851600646972656, 'sampling/importance_sampling_ratio/mean': 1.0479401350021362, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.876336634159088, 'clip_ratio/low_mean': 0.04768762364983559, 'clip_ratio/low_min': 0.04768762364983559, 'clip_ratio/high_mean': 0.15912827104330063, 'clip_ratio/high_max': 0.15912827104330063, 'clip_ratio/region_mean': 0.20681589469313622, 'reward_total_mean': 0.31550276279449463, 'reward_meter_mean': 0.9526747465133667, 'reward_meter_std': 0.08271235227584839, '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': 0.9904210567474365, 'reward_repeat_penalty_std': 0.011389861814677715, 'reward_repeat_floor_mean': 0.9985631704330444, 'reward_repeat_floor_std': 0.001708473195321858, 'reward_near_duplicate_penalty_mean': 0.9904210567474365, 'reward_near_duplicate_penalty_std': 0.011389861814677715, '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': 0.31550276279449463, 'reward_total_composite_std': 0.12497393041849136, 'epoch': 0.0}

  3%|▎         | 86/3300 [18:46<8:41:57,  9.74s/it]INFO 04-12 15:44:06 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 87/3300 [18:56<8:47:35,  9.85s/it]2026-04-12 15:44:16,570 | INFO | train_grpo_train | metrics_logged mode=train step=87

                                                   
{'loss': 0.0422, 'grad_norm': 13.136268615722656, 'learning_rate': 9.739393939393941e-06, 'num_tokens': 189556.0, 'completions/mean_length': 53.0, 'completions/min_length': 39.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5176620483398438, 'rewards/meter/std': 0.43811485171318054, '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': 0.9929601550102234, 'rewards/lexical_plausibility/std': 0.014150693081319332, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9975244998931885, 'rewards/repeat_penalty/std': 0.004858794622123241, 'rewards/repeat_floor/mean': 0.9996286630630493, 'rewards/repeat_floor/std': 0.0007288263295777142, 'rewards/near_duplicate_penalty/mean': 0.9975244998931885, 'rewards/near_duplicate_penalty/std': 0.004858794622123241, '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': 0.15177512168884277, 'rewards/total_composite/std': 0.1423334777355194, 'reward': 0.15177512168884277, 'reward_std': 0.1423334777355194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22891774773597717, 'sampling/sampling_logp_difference/max': 1.6756830215454102, 'sampling/importance_sampling_ratio/min': 0.1871802806854248, 'sampling/importance_sampling_ratio/mean': 1.0692352056503296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0857196748256683, 'clip_ratio/low_mean': 0.12169018387794495, 'clip_ratio/low_min': 0.12169018387794495, 'clip_ratio/high_mean': 0.07350949756801128, 'clip_ratio/high_max': 0.07350949756801128, 'clip_ratio/region_mean': 0.19519968144595623, 'reward_total_mean': 0.15177512168884277, 'reward_meter_mean': 0.5176620483398438, 'reward_meter_std': 0.43811485171318054, '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': 0.9929601550102234, 'reward_lexical_plausibility_std': 0.014150693081319332, 'reward_repeat_penalty_mean': 0.9975244998931885, 'reward_repeat_penalty_std': 0.004858794622123241, 'reward_repeat_floor_mean': 0.9996286630630493, 'reward_repeat_floor_std': 0.0007288263295777142, 'reward_near_duplicate_penalty_mean': 0.9975244998931885, 'reward_near_duplicate_penalty_std': 0.004858794622123241, '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': 0.15177512168884277, 'reward_total_composite_std': 0.1423334777355194, 'epoch': 0.0}

  3%|▎         | 87/3300 [18:56<8:47:35,  9.85s/it]INFO 04-12 15:44:16 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 88/3300 [19:07<9:01:00, 10.11s/it]2026-04-12 15:44:27,269 | INFO | train_grpo_train | metrics_logged mode=train step=88

                                                   
{'loss': 0.0892, 'grad_norm': 9.307430267333984, 'learning_rate': 9.736363636363637e-06, 'num_tokens': 191834.0, 'completions/mean_length': 91.75, 'completions/min_length': 72.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.4519146680831909, 'rewards/meter/std': 0.3520658612251282, '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': 0.9498867392539978, 'rewards/lexical_plausibility/std': 0.09293685108423233, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/repeat_penalty/mean': 0.9991486072540283, 'rewards/repeat_penalty/std': 0.001082957023754716, 'rewards/repeat_floor/mean': 0.9998722672462463, 'rewards/repeat_floor/std': 0.0001624451979296282, 'rewards/near_duplicate_penalty/mean': 0.9991486072540283, 'rewards/near_duplicate_penalty/std': 0.001082957023754716, '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': 0.16503232717514038, 'rewards/total_composite/std': 0.17378371953964233, 'reward': 0.16503232717514038, 'reward_std': 0.17378371953964233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24502447247505188, 'sampling/sampling_logp_difference/max': 1.5297584533691406, 'sampling/importance_sampling_ratio/min': 0.21658797562122345, 'sampling/importance_sampling_ratio/mean': 1.049220323562622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.324828267097473, 'clip_ratio/low_mean': 0.12661213986575603, 'clip_ratio/low_min': 0.12661213986575603, 'clip_ratio/high_mean': 0.08353758230805397, 'clip_ratio/high_max': 0.08353758230805397, 'clip_ratio/region_mean': 0.21014972217381, 'reward_total_mean': 0.16503232717514038, 'reward_meter_mean': 0.4519146680831909, 'reward_meter_std': 0.3520658612251282, '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': 0.9498867392539978, 'reward_lexical_plausibility_std': 0.09293685108423233, 'reward_repeat_penalty_mean': 0.9991486072540283, 'reward_repeat_penalty_std': 0.001082957023754716, 'reward_repeat_floor_mean': 0.9998722672462463, 'reward_repeat_floor_std': 0.0001624451979296282, 'reward_near_duplicate_penalty_mean': 0.9991486072540283, 'reward_near_duplicate_penalty_std': 0.001082957023754716, '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': 0.16503232717514038, 'reward_total_composite_std': 0.17378371953964233, 'epoch': 0.0}

  3%|▎         | 88/3300 [19:07<9:01:00, 10.11s/it]INFO 04-12 15:44:27 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 89/3300 [19:16<8:53:31,  9.97s/it]2026-04-12 15:44:36,923 | INFO | train_grpo_train | metrics_logged mode=train step=89

                                                   
{'loss': 0.0534, 'grad_norm': 18.06661605834961, 'learning_rate': 9.733333333333334e-06, 'num_tokens': 193332.0, 'completions/mean_length': 37.25, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5892886519432068, 'rewards/meter/std': 0.35698428750038147, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.993443489074707, 'rewards/repeat_penalty/std': 0.011376902461051941, 'rewards/repeat_floor/mean': 0.999016523361206, 'rewards/repeat_floor/std': 0.0017065367428585887, 'rewards/near_duplicate_penalty/mean': 0.993443489074707, 'rewards/near_duplicate_penalty/std': 0.011376902461051941, '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': 0.33069658279418945, 'rewards/total_composite/std': 0.2661266326904297, 'reward': 0.33069658279418945, 'reward_std': 0.2661266326904297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21617574989795685, 'sampling/sampling_logp_difference/max': 1.4496643543243408, 'sampling/importance_sampling_ratio/min': 0.23464903235435486, 'sampling/importance_sampling_ratio/mean': 1.0127161741256714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8535578101873398, 'clip_ratio/low_mean': 0.07034521922469139, 'clip_ratio/low_min': 0.07034521922469139, 'clip_ratio/high_mean': 0.07747877482324839, 'clip_ratio/high_max': 0.07747877482324839, 'clip_ratio/region_mean': 0.14782399404793978, 'reward_total_mean': 0.33069658279418945, 'reward_meter_mean': 0.5892886519432068, 'reward_meter_std': 0.35698428750038147, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.993443489074707, 'reward_repeat_penalty_std': 0.011376902461051941, 'reward_repeat_floor_mean': 0.999016523361206, 'reward_repeat_floor_std': 0.0017065367428585887, 'reward_near_duplicate_penalty_mean': 0.993443489074707, 'reward_near_duplicate_penalty_std': 0.011376902461051941, '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': 0.33069658279418945, 'reward_total_composite_std': 0.2661266326904297, 'epoch': 0.0}

  3%|▎         | 89/3300 [19:17<8:53:31,  9.97s/it]INFO 04-12 15:44:37 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 90/3300 [19:28<9:16:58, 10.41s/it]2026-04-12 15:44:48,357 | INFO | train_grpo_train | metrics_logged mode=train step=90

                                                   
{'loss': 0.1017, 'grad_norm': 8.156938552856445, 'learning_rate': 9.730303030303031e-06, 'num_tokens': 196043.0, 'completions/mean_length': 143.875, 'completions/min_length': 124.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.875, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.5183090567588806, 'rewards/meter/std': 0.2683314085006714, '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': 0.9977678656578064, 'rewards/lexical_plausibility/std': 0.006313450634479523, 'rewards/judge_quality/mean': 0.3375000059604645, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9818274974822998, 'rewards/repeat_penalty/std': 0.015070362947881222, 'rewards/repeat_floor/mean': 0.9979612827301025, 'rewards/repeat_floor/std': 0.001504645450040698, 'rewards/near_duplicate_penalty/mean': 0.9934290647506714, 'rewards/near_duplicate_penalty/std': 0.002804782474413514, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9882984757423401, 'rewards/distinct_2/std': 0.013034606352448463, 'rewards/total_composite/mean': 0.1824151575565338, 'rewards/total_composite/std': 0.12442703545093536, 'reward': 0.1824151575565338, 'reward_std': 0.12442702800035477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24606093764305115, 'sampling/sampling_logp_difference/max': 1.7278642654418945, 'sampling/importance_sampling_ratio/min': 0.17766344547271729, 'sampling/importance_sampling_ratio/mean': 1.0536800622940063, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9328472316265106, 'clip_ratio/low_mean': 0.08303666766732931, 'clip_ratio/low_min': 0.08303666766732931, 'clip_ratio/high_mean': 0.09808459505438805, 'clip_ratio/high_max': 0.09808459505438805, 'clip_ratio/region_mean': 0.18112126272171736, 'reward_total_mean': 0.1824151575565338, 'reward_meter_mean': 0.5183090567588806, 'reward_meter_std': 0.2683314085006714, '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': 0.9977678656578064, 'reward_lexical_plausibility_std': 0.006313450634479523, 'reward_repeat_penalty_mean': 0.9818274974822998, 'reward_repeat_penalty_std': 0.015070362947881222, 'reward_repeat_floor_mean': 0.9979612827301025, 'reward_repeat_floor_std': 0.001504645450040698, 'reward_near_duplicate_penalty_mean': 0.9934290647506714, 'reward_near_duplicate_penalty_std': 0.002804782474413514, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9882984757423401, 'reward_distinct_2_std': 0.013034606352448463, 'reward_judge_quality_mean': 0.3375000059604645, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.1824151575565338, 'reward_total_composite_std': 0.12442703545093536, 'epoch': 0.0}

  3%|▎         | 90/3300 [19:28<9:16:58, 10.41s/it]INFO 04-12 15:44:48 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 91/3300 [19:40<9:38:01, 10.81s/it]2026-04-12 15:45:00,092 | INFO | train_grpo_train | metrics_logged mode=train step=91

                                                   
{'loss': 0.0859, 'grad_norm': 8.105484008789062, 'learning_rate': 9.727272727272728e-06, 'num_tokens': 198954.0, 'completions/mean_length': 165.875, 'completions/min_length': 149.0, 'completions/max_length': 208.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.875, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 208.0, 'rewards/meter/mean': 0.2851950526237488, 'rewards/meter/std': 0.2575579881668091, 'rewards/count_adherence/mean': 0.9722222089767456, 'rewards/count_adherence/std': 0.07856741547584534, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 0.9916666746139526, 'rewards/count_floor/std': 0.0235702246427536, 'rewards/lexical_plausibility/mean': 0.9704553484916687, 'rewards/lexical_plausibility/std': 0.05629132688045502, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/repeat_penalty/mean': 0.9946459531784058, 'rewards/repeat_penalty/std': 0.002710993867367506, 'rewards/repeat_floor/mean': 0.9991968870162964, 'rewards/repeat_floor/std': 0.0004066347028128803, 'rewards/near_duplicate_penalty/mean': 0.9946459531784058, 'rewards/near_duplicate_penalty/std': 0.002710993867367506, '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': 0.0776951014995575, 'rewards/total_composite/std': 0.07800179719924927, 'reward': 0.0776951014995575, 'reward_std': 0.07800179719924927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23621301352977753, 'sampling/sampling_logp_difference/max': 1.3729228973388672, 'sampling/importance_sampling_ratio/min': 0.25336530804634094, 'sampling/importance_sampling_ratio/mean': 1.0491832494735718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.093779385089874, 'clip_ratio/low_mean': 0.10351813212037086, 'clip_ratio/low_min': 0.10351813212037086, 'clip_ratio/high_mean': 0.09971963427960873, 'clip_ratio/high_max': 0.09971963427960873, 'clip_ratio/region_mean': 0.2032377663999796, 'reward_total_mean': 0.0776951014995575, 'reward_meter_mean': 0.2851950526237488, 'reward_meter_std': 0.2575579881668091, 'reward_count_adherence_mean': 0.9722222089767456, 'reward_count_adherence_std': 0.07856741547584534, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 0.9916666746139526, 'reward_count_floor_std': 0.0235702246427536, 'reward_lexical_plausibility_mean': 0.9704553484916687, 'reward_lexical_plausibility_std': 0.05629132688045502, 'reward_repeat_penalty_mean': 0.9946459531784058, 'reward_repeat_penalty_std': 0.002710993867367506, 'reward_repeat_floor_mean': 0.9991968870162964, 'reward_repeat_floor_std': 0.0004066347028128803, 'reward_near_duplicate_penalty_mean': 0.9946459531784058, 'reward_near_duplicate_penalty_std': 0.002710993867367506, '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': 0.23749999701976776, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.0776951014995575, 'reward_total_composite_std': 0.07800179719924927, 'epoch': 0.0}

  3%|▎         | 91/3300 [19:40<9:38:01, 10.81s/it]INFO 04-12 15:45:00 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 92/3300 [19:52<10:07:03, 11.35s/it]2026-04-12 15:45:12,720 | INFO | train_grpo_train | metrics_logged mode=train step=92

                                                    
{'loss': 0.069, 'grad_norm': 8.861458778381348, 'learning_rate': 9.724242424242426e-06, 'num_tokens': 202043.0, 'completions/mean_length': 145.125, 'completions/min_length': 117.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 145.125, 'completions/min_terminated_length': 117.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.4660497307777405, 'rewards/meter/std': 0.309230774641037, 'rewards/count_adherence/mean': 0.9821428656578064, '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': 0.9946428537368774, 'rewards/count_floor/std': 0.015152297914028168, 'rewards/lexical_plausibility/mean': 0.9504303336143494, 'rewards/lexical_plausibility/std': 0.06441396474838257, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660183429718, 'rewards/repeat_penalty/mean': 0.9935762882232666, 'rewards/repeat_penalty/std': 0.005078553222119808, 'rewards/repeat_floor/mean': 0.999036431312561, 'rewards/repeat_floor/std': 0.0007617836236022413, 'rewards/near_duplicate_penalty/mean': 0.9935762882232666, 'rewards/near_duplicate_penalty/std': 0.005078553222119808, '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': 0.08608467131853104, 'rewards/total_composite/std': 0.05834082514047623, 'reward': 0.08608467131853104, 'reward_std': 0.05834082514047623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23572970926761627, 'sampling/sampling_logp_difference/max': 1.4420547485351562, 'sampling/importance_sampling_ratio/min': 0.23644143342971802, 'sampling/importance_sampling_ratio/mean': 1.0500637292861938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.634262293577194, 'clip_ratio/low_mean': 0.09698570147156715, 'clip_ratio/low_min': 0.09698570147156715, 'clip_ratio/high_mean': 0.13141235895454884, 'clip_ratio/high_max': 0.13141235895454884, 'clip_ratio/region_mean': 0.228398060426116, 'reward_total_mean': 0.08608467131853104, 'reward_meter_mean': 0.4660497307777405, 'reward_meter_std': 0.309230774641037, 'reward_count_adherence_mean': 0.9821428656578064, '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': 0.9946428537368774, 'reward_count_floor_std': 0.015152297914028168, 'reward_lexical_plausibility_mean': 0.9504303336143494, 'reward_lexical_plausibility_std': 0.06441396474838257, 'reward_repeat_penalty_mean': 0.9935762882232666, 'reward_repeat_penalty_std': 0.005078553222119808, 'reward_repeat_floor_mean': 0.999036431312561, 'reward_repeat_floor_std': 0.0007617836236022413, 'reward_near_duplicate_penalty_mean': 0.9935762882232666, 'reward_near_duplicate_penalty_std': 0.005078553222119808, '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': 0.08608467131853104, 'reward_total_composite_std': 0.05834082514047623, 'epoch': 0.0}

  3%|▎         | 92/3300 [19:52<10:07:03, 11.35s/it]INFO 04-12 15:45:12 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 93/3300 [20:03<10:01:39, 11.26s/it]2026-04-12 15:45:23,749 | INFO | train_grpo_train | metrics_logged mode=train step=93

                                                    
{'loss': -0.0625, 'grad_norm': 10.055813789367676, 'learning_rate': 9.721212121212123e-06, 'num_tokens': 204220.0, 'completions/mean_length': 91.125, 'completions/min_length': 74.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.125, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.7445623874664307, 'rewards/meter/std': 0.3468618392944336, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.9762678146362305, 'rewards/repeat_penalty/std': 0.041598014533519745, 'rewards/repeat_floor/mean': 0.9972038865089417, 'rewards/repeat_floor/std': 0.004112606402486563, 'rewards/near_duplicate_penalty/mean': 0.989409327507019, 'rewards/near_duplicate_penalty/std': 0.007115801330655813, 'rewards/opening_diversity/mean': 1.0, 'rewards/opening_diversity/std': 0.0, 'rewards/distinct_2/mean': 0.9865831732749939, 'rewards/distinct_2/std': 0.037948496639728546, 'rewards/total_composite/mean': 0.16710910201072693, 'rewards/total_composite/std': 0.12870369851589203, 'reward': 0.16710910201072693, 'reward_std': 0.12870368361473083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24216312170028687, 'sampling/sampling_logp_difference/max': 1.7295198440551758, 'sampling/importance_sampling_ratio/min': 0.17736954987049103, 'sampling/importance_sampling_ratio/mean': 1.0720044374465942, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.665288805961609, 'clip_ratio/low_mean': 0.1566590964794159, 'clip_ratio/low_min': 0.1566590964794159, 'clip_ratio/high_mean': 0.04208195023238659, 'clip_ratio/high_max': 0.04208195023238659, 'clip_ratio/region_mean': 0.19874104671180248, 'reward_total_mean': 0.16710910201072693, 'reward_meter_mean': 0.7445623874664307, 'reward_meter_std': 0.3468618392944336, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9762678146362305, 'reward_repeat_penalty_std': 0.041598014533519745, 'reward_repeat_floor_mean': 0.9972038865089417, 'reward_repeat_floor_std': 0.004112606402486563, 'reward_near_duplicate_penalty_mean': 0.989409327507019, 'reward_near_duplicate_penalty_std': 0.007115801330655813, 'reward_opening_diversity_mean': 1.0, 'reward_opening_diversity_std': 0.0, 'reward_distinct_2_mean': 0.9865831732749939, 'reward_distinct_2_std': 0.037948496639728546, 'reward_judge_quality_mean': 0.21249999105930328, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.16710910201072693, 'reward_total_composite_std': 0.12870369851589203, 'epoch': 0.0}

  3%|▎         | 93/3300 [20:03<10:01:39, 11.26s/it]INFO 04-12 15:45:23 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 94/3300 [20:12<9:25:56, 10.59s/it] 2026-04-12 15:45:32,791 | INFO | train_grpo_train | metrics_logged mode=train step=94

                                                   
{'loss': 0.1152, 'grad_norm': 19.25832748413086, 'learning_rate': 9.718181818181818e-06, 'num_tokens': 205672.0, 'completions/mean_length': 31.5, 'completions/min_length': 25.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.22589288651943207, 'rewards/meter/std': 0.2675091326236725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9429793357849121, 'rewards/lexical_plausibility/std': 0.13593381643295288, 'rewards/judge_quality/mean': 0.375, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9963842630386353, 'rewards/repeat_penalty/std': 0.010226743295788765, 'rewards/repeat_floor/mean': 0.9994576573371887, 'rewards/repeat_floor/std': 0.0015340187819674611, 'rewards/near_duplicate_penalty/mean': 0.9963842630386353, 'rewards/near_duplicate_penalty/std': 0.010226743295788765, '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': 0.09153544902801514, 'rewards/total_composite/std': 0.12027516216039658, 'reward': 0.09153544902801514, 'reward_std': 0.12027516216039658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.207672581076622, 'sampling/sampling_logp_difference/max': 1.3137211799621582, 'sampling/importance_sampling_ratio/min': 0.26881787180900574, 'sampling/importance_sampling_ratio/mean': 1.03855299949646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4974738359451294, 'clip_ratio/low_mean': 0.12261239439249039, 'clip_ratio/low_min': 0.12261239439249039, 'clip_ratio/high_mean': 0.06354166753590107, 'clip_ratio/high_max': 0.06354166753590107, 'clip_ratio/region_mean': 0.18615406192839146, 'reward_total_mean': 0.09153544902801514, 'reward_meter_mean': 0.22589288651943207, 'reward_meter_std': 0.2675091326236725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9429793357849121, 'reward_lexical_plausibility_std': 0.13593381643295288, 'reward_repeat_penalty_mean': 0.9963842630386353, 'reward_repeat_penalty_std': 0.010226743295788765, 'reward_repeat_floor_mean': 0.9994576573371887, 'reward_repeat_floor_std': 0.0015340187819674611, 'reward_near_duplicate_penalty_mean': 0.9963842630386353, 'reward_near_duplicate_penalty_std': 0.010226743295788765, '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': 0.09153544902801514, 'reward_total_composite_std': 0.12027516216039658, 'epoch': 0.0}

  3%|▎         | 94/3300 [20:12<9:25:56, 10.59s/it]INFO 04-12 15:45:33 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 95/3300 [20:22<9:14:08, 10.37s/it]2026-04-12 15:45:42,655 | INFO | train_grpo_train | metrics_logged mode=train step=95

                                                   
{'loss': 0.1265, 'grad_norm': 19.655981063842773, 'learning_rate': 9.715151515151516e-06, 'num_tokens': 207325.0, 'completions/mean_length': 39.625, 'completions/min_length': 33.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6288896203041077, 'rewards/meter/std': 0.384452223777771, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, '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': 0.9993181824684143, 'rewards/repeat_penalty/std': 0.0019284712616354227, 'rewards/repeat_floor/mean': 0.9998977184295654, 'rewards/repeat_floor/std': 0.00028927490347996354, 'rewards/near_duplicate_penalty/mean': 0.9993181824684143, 'rewards/near_duplicate_penalty/std': 0.0019284712616354227, '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': 0.24521130323410034, 'rewards/total_composite/std': 0.15802384912967682, 'reward': 0.24521130323410034, 'reward_std': 0.15802384912967682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20129792392253876, 'sampling/sampling_logp_difference/max': 0.9452333450317383, 'sampling/importance_sampling_ratio/min': 0.38858890533447266, 'sampling/importance_sampling_ratio/mean': 1.0726001262664795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.982078045606613, 'clip_ratio/low_mean': 0.07321428693830967, 'clip_ratio/low_min': 0.07321428693830967, 'clip_ratio/high_mean': 0.13783274590969086, 'clip_ratio/high_max': 0.13783274590969086, 'clip_ratio/region_mean': 0.21104703284800053, 'reward_total_mean': 0.24521130323410034, 'reward_meter_mean': 0.6288896203041077, 'reward_meter_std': 0.384452223777771, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 1.0, 'reward_lexical_plausibility_std': 0.0, 'reward_repeat_penalty_mean': 0.9993181824684143, 'reward_repeat_penalty_std': 0.0019284712616354227, 'reward_repeat_floor_mean': 0.9998977184295654, 'reward_repeat_floor_std': 0.00028927490347996354, 'reward_near_duplicate_penalty_mean': 0.9993181824684143, 'reward_near_duplicate_penalty_std': 0.0019284712616354227, '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': 0.24521130323410034, 'reward_total_composite_std': 0.15802384912967682, 'epoch': 0.0}

  3%|▎         | 95/3300 [20:22<9:14:08, 10.37s/it]INFO 04-12 15:45:42 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 96/3300 [20:33<9:17:28, 10.44s/it]2026-04-12 15:45:53,249 | INFO | train_grpo_train | metrics_logged mode=train step=96

                                                   
{'loss': 0.05, 'grad_norm': 10.176275253295898, 'learning_rate': 9.712121212121213e-06, 'num_tokens': 209638.0, 'completions/mean_length': 112.125, 'completions/min_length': 68.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.125, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.27093616127967834, 'rewards/meter/std': 0.30527225136756897, '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': 0.7630946040153503, 'rewards/lexical_plausibility/std': 0.19082005321979523, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/repeat_penalty/mean': 0.9934767484664917, 'rewards/repeat_penalty/std': 0.008737701922655106, 'rewards/repeat_floor/mean': 0.9990215301513672, 'rewards/repeat_floor/std': 0.0013106464175507426, 'rewards/near_duplicate_penalty/mean': 0.9934767484664917, 'rewards/near_duplicate_penalty/std': 0.008737701922655106, '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': 0.0771602988243103, 'rewards/total_composite/std': 0.09407317638397217, 'reward': 0.0771602988243103, 'reward_std': 0.09407317638397217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23485730588436127, 'sampling/sampling_logp_difference/max': 1.482388973236084, 'sampling/importance_sampling_ratio/min': 0.22709451615810394, 'sampling/importance_sampling_ratio/mean': 1.0526814460754395, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.172715663909912, 'clip_ratio/low_mean': 0.1263983454555273, 'clip_ratio/low_min': 0.1263983454555273, 'clip_ratio/high_mean': 0.07382924109697342, 'clip_ratio/high_max': 0.07382924109697342, 'clip_ratio/region_mean': 0.20022758655250072, 'reward_total_mean': 0.0771602988243103, 'reward_meter_mean': 0.27093616127967834, 'reward_meter_std': 0.30527225136756897, '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': 0.7630946040153503, 'reward_lexical_plausibility_std': 0.19082005321979523, 'reward_repeat_penalty_mean': 0.9934767484664917, 'reward_repeat_penalty_std': 0.008737701922655106, 'reward_repeat_floor_mean': 0.9990215301513672, 'reward_repeat_floor_std': 0.0013106464175507426, 'reward_near_duplicate_penalty_mean': 0.9934767484664917, 'reward_near_duplicate_penalty_std': 0.008737701922655106, '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': 0.1414213478565216, 'reward_total_composite_mean': 0.0771602988243103, 'reward_total_composite_std': 0.09407317638397217, 'epoch': 0.0}

  3%|▎         | 96/3300 [20:33<9:17:28, 10.44s/it]INFO 04-12 15:45:53 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 97/3300 [20:42<8:49:28,  9.92s/it]2026-04-12 15:46:01,954 | INFO | train_grpo_train | metrics_logged mode=train step=97

                                                   
{'loss': -0.0049, 'grad_norm': 12.639327049255371, 'learning_rate': 9.70909090909091e-06, 'num_tokens': 211384.0, 'completions/mean_length': 49.25, 'completions/min_length': 33.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5977282524108887, 'rewards/meter/std': 0.3484576642513275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/arabic_clean/mean': 1.0, 'rewards/arabic_clean/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/arabic_floor/mean': 1.0, 'rewards/arabic_floor/std': 0.0, 'rewards/count_floor/mean': 1.0, 'rewards/count_floor/std': 0.0, 'rewards/lexical_plausibility/mean': 0.9726052284240723, 'rewards/lexical_plausibility/std': 0.05644887313246727, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.2745385766029358, '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': 0.26195141673088074, 'rewards/total_composite/std': 0.1852988451719284, 'reward': 0.26195141673088074, 'reward_std': 0.1852988302707672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20207305252552032, 'sampling/sampling_logp_difference/max': 1.3922796249389648, 'sampling/importance_sampling_ratio/min': 0.24850815534591675, 'sampling/importance_sampling_ratio/mean': 1.0321518182754517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.122445374727249, 'clip_ratio/low_mean': 0.14189988188445568, 'clip_ratio/low_min': 0.14189988188445568, 'clip_ratio/high_mean': 0.07165093161165714, 'clip_ratio/high_max': 0.07165093161165714, 'clip_ratio/region_mean': 0.21355081349611282, 'reward_total_mean': 0.26195141673088074, 'reward_meter_mean': 0.5977282524108887, 'reward_meter_std': 0.3484576642513275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_arabic_clean_mean': 1.0, 'reward_arabic_clean_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_arabic_floor_mean': 1.0, 'reward_arabic_floor_std': 0.0, 'reward_count_floor_mean': 1.0, 'reward_count_floor_std': 0.0, 'reward_lexical_plausibility_mean': 0.9726052284240723, 'reward_lexical_plausibility_std': 0.05644887313246727, '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': 0.5049999952316284, 'reward_judge_quality_std': 0.2745385766029358, 'reward_total_composite_mean': 0.26195141673088074, 'reward_total_composite_std': 0.1852988451719284, 'epoch': 0.0}

  3%|▎         | 97/3300 [20:42<8:49:28,  9.92s/it]INFO 04-12 15:46:02 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 98/3300 [20:52<8:50:09,  9.93s/it]2026-04-12 15:46:11,923 | INFO | train_grpo_train | metrics_logged mode=train step=98

                                                   
{'loss': 0.0753, 'grad_norm': 9.187956809997559, 'learning_rate': 9.706060606060606e-06, 'num_tokens': 213668.0, 'completions/mean_length': 95.5, 'completions/min_length': 67.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.2505316734313965, 'rewards/meter/std': 0.32169413566589355, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/arabic_clean/mean': 0.875, 'rewards/arabic_clean/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, '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': 0.05669466033577919, 'rewards/lexical_plausibility/mean': 0.9390342235565186, 'rewards/lexical_plausibility/std': 0.0646214485168457, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/repeat_penalty/mean': 0.9948243498802185, 'rewards/repeat_penalty/std': 0.004409181885421276, 'rewards/repeat_floor/mean': 0.9992235898971558, 'rewards/repeat_floor/std': 0.0006613862351514399, 'rewards/near_duplicate_penalty/mean': 0.9948243498802185, 'rewards/near_duplicate_penalty/std': 0.004409181885421276, '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': 0.08780472725629807, 'rewards/total_composite/std': 0.1362478882074356, 'reward': 0.08780472725629807, 'reward_std': 0.13624787330627441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26570001244544983, 'sampling/sampling_logp_difference/max': 1.4997978210449219, 'sampling/importance_sampling_ratio/min': 0.2231752872467041, 'sampling/importance_sampling_ratio/mean': 1.0519211292266846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7513682544231415, 'clip_ratio/low_mean': 0.14081784337759018, 'clip_ratio/low_min': 0.14081784337759018, 'clip_ratio/high_mean': 0.05643996223807335, 'clip_ratio/high_max': 0.05643996223807335, 'clip_ratio/region_mean': 0.19725780561566353, 'reward_total_mean': 0.08780472725629807, 'reward_meter_mean': 0.2505316734313965, 'reward_meter_std': 0.32169413566589355, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.18898223340511322, 'reward_arabic_clean_mean': 0.875, 'reward_arabic_clean_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, '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': 0.05669466033577919, 'reward_lexical_plausibility_mean': 0.9390342235565186, 'reward_lexical_plausibility_std': 0.0646214485168457, 'reward_repeat_penalty_mean': 0.9948243498802185, 'reward_repeat_penalty_std': 0.004409181885421276, 'reward_repeat_floor_mean': 0.9992235898971558, 'reward_repeat_floor_std': 0.0006613862351514399, 'reward_near_duplicate_penalty_mean': 0.9948243498802185, 'reward_near_duplicate_penalty_std': 0.004409181885421276, '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': 0.2749999761581421, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.08780472725629807, 'reward_total_composite_std': 0.1362478882074356, 'epoch': 0.0}

  3%|▎         | 98/3300 [20:52<8:50:09,  9.93s/it]INFO 04-12 15:46:12 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 99/3300 [21:01<8:39:21,  9.73s/it]2026-04-12 15:46:21,193 | INFO | train_grpo_train | metrics_logged mode=train step=99

                                                   
{'loss': 0.1105, 'grad_norm': 12.105984687805176, 'learning_rate': 9.703030303030305e-06, 'num_tokens': 215618.0, 'completions/mean_length': 75.75, 'completions/min_length': 60.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.75, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.641942024230957, 'rewards/meter/std': 0.37689656019210815, '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': 0.995192289352417, 'rewards/lexical_plausibility/std': 0.013598216697573662, 'rewards/judge_quality/mean': 0.2750000059604645, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/repeat_penalty/mean': 0.9947735071182251, 'rewards/repeat_penalty/std': 0.002478251699358225, 'rewards/repeat_floor/mean': 0.9992160201072693, 'rewards/repeat_floor/std': 0.0003717463696375489, 'rewards/near_duplicate_penalty/mean': 0.9947735071182251, 'rewards/near_duplicate_penalty/std': 0.002478251699358225, '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': 0.1920507550239563, 'rewards/total_composite/std': 0.1380826085805893, 'reward': 0.1920507550239563, 'reward_std': 0.1380826085805893, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25587889552116394, 'sampling/sampling_logp_difference/max': 1.5802116394042969, 'sampling/importance_sampling_ratio/min': 0.20593149960041046, 'sampling/importance_sampling_ratio/mean': 1.0554982423782349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.312505006790161, 'clip_ratio/low_mean': 0.1088887695223093, 'clip_ratio/low_min': 0.1088887695223093, 'clip_ratio/high_mean': 0.11645336635410786, 'clip_ratio/high_max': 0.11645336635410786, 'clip_ratio/region_mean': 0.22534213587641716, 'reward_total_mean': 0.1920507550239563, 'reward_meter_mean': 0.641942024230957, 'reward_meter_std': 0.37689656019210815, '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': 0.995192289352417, 'reward_lexical_plausibility_std': 0.013598216697573662, 'reward_repeat_penalty_mean': 0.9947735071182251, 'reward_repeat_penalty_std': 0.002478251699358225, 'reward_repeat_floor_mean': 0.9992160201072693, 'reward_repeat_floor_std': 0.0003717463696375489, 'reward_near_duplicate_penalty_mean': 0.9947735071182251, 'reward_near_duplicate_penalty_std': 0.002478251699358225, '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': 0.1920507550239563, 'reward_total_composite_std': 0.1380826085805893, 'epoch': 0.0}

  3%|▎         | 99/3300 [21:01<8:39:21,  9.73s/it]INFO 04-12 15:46:21 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 100/3300 [21:11<8:41:35,  9.78s/it]2026-04-12 15:46:31,245 | INFO | train_grpo_train | metrics_logged mode=train step=100

                                                    
{'loss': 0.1118, 'grad_norm': 14.011872291564941, 'learning_rate': 9.7e-06, 'num_tokens': 217289.0, 'completions/mean_length': 54.875, 'completions/min_length': 45.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.14609289169311523, 'rewards/meter/std': 0.1894940733909607, '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': 0.48000001907348633, 'rewards/judge_quality/std': 0.29871153831481934, 'rewards/repeat_penalty/mean': 0.9940087795257568, 'rewards/repeat_penalty/std': 0.0035288461949676275, 'rewards/repeat_floor/mean': 0.9991012811660767, 'rewards/repeat_floor/std': 0.0005293278954923153, 'rewards/near_duplicate_penalty/mean': 0.9940087795257568, 'rewards/near_duplicate_penalty/std': 0.0035288461949676275, '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': 0.049753956496715546, 'rewards/total_composite/std': 0.04388604313135147, 'reward': 0.049753956496715546, 'reward_std': 0.04388603940606117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24366313219070435, 'sampling/sampling_logp_difference/max': 1.695225715637207, 'sampling/importance_sampling_ratio/min': 0.22670167684555054, 'sampling/importance_sampling_ratio/mean': 1.0452406406402588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5091373920440674, 'clip_ratio/low_mean': 0.11318430211395025, 'clip_ratio/low_min': 0.11318430211395025, 'clip_ratio/high_mean': 0.07480035535991192, 'clip_ratio/high_max': 0.07480035535991192, 'clip_ratio/region_mean': 0.18798465747386217, 'reward_total_mean': 0.049753956496715546, 'reward_meter_mean': 0.14609289169311523, 'reward_meter_std': 0.1894940733909607, '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': 0.9940087795257568, 'reward_repeat_penalty_std': 0.0035288461949676275, 'reward_repeat_floor_mean': 0.9991012811660767, 'reward_repeat_floor_std': 0.0005293278954923153, 'reward_near_duplicate_penalty_mean': 0.9940087795257568, 'reward_near_duplicate_penalty_std': 0.0035288461949676275, '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': 0.48000001907348633, 'reward_judge_quality_std': 0.29871153831481934, 'reward_total_composite_mean': 0.049753956496715546, 'reward_total_composite_std': 0.04388604313135147, 'epoch': 0.0}

  3%|▎         | 100/3300 [21:11<8:41:35,  9.78s/it]INFO 04-12 15:46:31 [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 [00:18<01:04,  6.49s/it]

 31%|███       | 4/13 [00:29<01:11,  7.94s/it]

 38%|███▊      | 5/13 [00:39<01:10,  8.77s/it]

 46%|████▌     | 6/13 [01:00<01:30, 12.86s/it]

 54%|█████▍    | 7/13 [01:13<01:16, 12.77s/it]

 62%|██████▏   | 8/13 [01:25<01:02, 12.56s/it]

 69%|██████▉   | 9/13 [01:36<00:48, 12.09s/it]

 77%|███████▋  | 10/13 [01:44<00:32, 10.95s/it]

 85%|████████▍ | 11/13 [01:56<00:22, 11.18s/it]

 92%|█████████▏| 12/13 [02:08<00:11, 11.56s/it]

100%|██████████| 13/13 [02:20<00:00, 11.72s/it]2026-04-12 15:49:05,410 | INFO | train_grpo_train | metrics_logged mode=eval step=100

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 154.1501, 'eval_samples_per_second': 0.675, 'eval_steps_per_second': 0.084, 'eval_num_tokens': 217289.0, 'eval_completions/mean_length': 159.80769230769232, 'eval_completions/min_length': 33.46153846153846, 'eval_completions/max_length': 416.3076923076923, 'eval_completions/clipped_ratio': 0.0673076923076923, 'eval_completions/mean_terminated_length': 134.21841137225812, 'eval_completions/min_terminated_length': 33.46153846153846, 'eval_completions/max_terminated_length': 292.0769230769231, 'eval_rewards/meter/mean': 0.2705841820973616, 'eval_rewards/meter/std': 0.2909646320801515, 'eval_rewards/count_adherence/mean': 0.9342383558933551, 'eval_rewards/count_adherence/std': 0.0986164828332571, 'eval_rewards/arabic_clean/mean': 0.8942307692307693, 'eval_rewards/arabic_clean/std': 0.26159183107889616, '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': 0.9802715090604929, 'eval_rewards/count_floor/std': 0.02958494883317214, 'eval_rewards/lexical_plausibility/mean': 0.9585024301822369, 'eval_rewards/lexical_plausibility/std': 0.07200933190492484, 'eval_rewards/judge_quality/mean': 0.28634615586354184, 'eval_rewards/judge_quality/std': 0.17425813984412414, 'eval_rewards/repeat_penalty/mean': 0.9783954391112695, 'eval_rewards/repeat_penalty/std': 0.03855727448199804, 'eval_rewards/repeat_floor/mean': 0.9976500364450308, 'eval_rewards/repeat_floor/std': 0.0037385166148081995, 'eval_rewards/near_duplicate_penalty/mean': 0.9903955184496366, 'eval_rewards/near_duplicate_penalty/std': 0.015220703539223624, 'eval_rewards/opening_diversity/mean': 0.9927884615384616, 'eval_rewards/opening_diversity/std': 0.020397310646680687, 'eval_rewards/distinct_2/mean': 0.9947044849395752, 'eval_rewards/distinct_2/std': 0.012070114581057658, 'eval_rewards/total_composite/mean': 0.07959950514710866, 'eval_rewards/total_composite/std': 0.1032704974596317, 'eval_reward': 0.07959950514710866, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1744753672526433, 'eval_sampling/sampling_logp_difference/max': 1.2611371553861177, 'eval_sampling/importance_sampling_ratio/min': 0.2868912804585237, 'eval_sampling/importance_sampling_ratio/mean': 1.058729887008667, 'eval_sampling/importance_sampling_ratio/max': 1.7361209667645967, 'eval_entropy': 3.3903161195608287, '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': 0.07959950514710866, 'eval_reward_meter_mean': 0.2705841820973616, 'eval_reward_meter_std': 0.2909646320801515, 'eval_reward_count_adherence_mean': 0.9342383558933551, 'eval_reward_count_adherence_std': 0.0986164828332571, 'eval_reward_arabic_clean_mean': 0.8942307692307693, 'eval_reward_arabic_clean_std': 0.26159183107889616, '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': 0.9802715090604929, 'eval_reward_count_floor_std': 0.02958494883317214, 'eval_reward_lexical_plausibility_mean': 0.9585024301822369, 'eval_reward_lexical_plausibility_std': 0.07200933190492484, 'eval_reward_repeat_penalty_mean': 0.9783954391112695, 'eval_reward_repeat_penalty_std': 0.03855727448199804, 'eval_reward_repeat_floor_mean': 0.9976500364450308, 'eval_reward_repeat_floor_std': 0.0037385166148081995, 'eval_reward_near_duplicate_penalty_mean': 0.9903955184496366, 'eval_reward_near_duplicate_penalty_std': 0.015220703539223624, 'eval_reward_opening_diversity_mean': 0.9927884615384616, 'eval_reward_opening_diversity_std': 0.020397310646680687, 'eval_reward_distinct_2_mean': 0.9947044849395752, 'eval_reward_distinct_2_std': 0.012070114581057658, 'eval_reward_judge_quality_mean': 0.28634615586354184, 'eval_reward_judge_quality_std': 0.17425813984412414, 'eval_reward_total_composite_mean': 0.07959950514710866, 'eval_reward_total_composite_std': 0.1032704974596317, 'epoch': 0.0}

  3%|▎         | 100/3300 [23:45<8:41:35,  9.78s/it]

100%|██████████| 13/13 [02:20<00:00, 11.72s/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(