INFO 04-12 22:11:52 [__init__.py:216] Automatically detected platform cuda.
2026-04-12 22:11:54,382 | INFO | train_grpo_train | mode=train

Filter:   0%|          | 0/109070 [00:00<?, ? examples/s]
Filter:   1%|          | 1000/109070 [00:00<00:38, 2790.32 examples/s]
Filter:   2%|▏         | 2000/109070 [00:00<00:36, 2897.00 examples/s]
Filter:   3%|▎         | 3000/109070 [00:01<00:36, 2942.15 examples/s]
Filter:   4%|▎         | 4000/109070 [00:01<00:36, 2892.65 examples/s]
Filter:   5%|▍         | 5000/109070 [00:01<00:36, 2881.07 examples/s]
Filter:   6%|▌         | 6000/109070 [00:02<00:35, 2903.15 examples/s]
Filter:   6%|▋         | 7000/109070 [00:02<00:34, 2942.03 examples/s]
Filter:   7%|▋         | 8000/109070 [00:02<00:34, 2953.45 examples/s]
Filter:   8%|▊         | 9000/109070 [00:03<00:33, 2977.04 examples/s]
Filter:   9%|▉         | 10000/109070 [00:03<00:33, 2967.51 examples/s]
Filter:  10%|█         | 11000/109070 [00:03<00:33, 2958.38 examples/s]
Filter:  11%|█         | 12000/109070 [00:04<00:32, 2952.45 examples/s]
Filter:  12%|█▏        | 13000/109070 [00:04<00:32, 2932.64 examples/s]
Filter:  13%|█▎        | 14000/109070 [00:04<00:32, 2918.90 examples/s]
Filter:  14%|█▍        | 15000/109070 [00:05<00:31, 2961.99 examples/s]
Filter:  15%|█▍        | 16000/109070 [00:05<00:31, 2976.37 examples/s]
Filter:  16%|█▌        | 17000/109070 [00:05<00:31, 2966.57 examples/s]
Filter:  17%|█▋        | 18000/109070 [00:06<00:30, 2937.87 examples/s]
Filter:  17%|█▋        | 19000/109070 [00:06<00:30, 2907.35 examples/s]
Filter:  18%|█▊        | 20000/109070 [00:06<00:30, 2894.40 examples/s]
Filter:  19%|█▉        | 21000/109070 [00:07<00:30, 2886.18 examples/s]
Filter:  20%|██        | 22000/109070 [00:07<00:30, 2889.05 examples/s]
Filter:  21%|██        | 23000/109070 [00:07<00:29, 2881.68 examples/s]
Filter:  22%|██▏       | 24000/109070 [00:08<00:29, 2893.15 examples/s]
Filter:  23%|██▎       | 25000/109070 [00:08<00:29, 2885.85 examples/s]
Filter:  24%|██▍       | 26000/109070 [00:08<00:28, 2886.42 examples/s]
Filter:  25%|██▍       | 27000/109070 [00:09<00:28, 2872.39 examples/s]
Filter:  26%|██▌       | 28000/109070 [00:09<00:28, 2882.78 examples/s]
Filter:  27%|██▋       | 29000/109070 [00:09<00:27, 2883.44 examples/s]
Filter:  28%|██▊       | 30000/109070 [00:10<00:27, 2881.50 examples/s]
Filter:  28%|██▊       | 31000/109070 [00:10<00:27, 2884.66 examples/s]
Filter:  29%|██▉       | 32000/109070 [00:10<00:26, 2874.63 examples/s]
Filter:  30%|███       | 33000/109070 [00:11<00:26, 2891.33 examples/s]
Filter:  31%|███       | 34000/109070 [00:11<00:25, 2890.44 examples/s]
Filter:  32%|███▏      | 35000/109070 [00:12<00:25, 2896.44 examples/s]
Filter:  33%|███▎      | 36000/109070 [00:12<00:25, 2886.20 examples/s]
Filter:  34%|███▍      | 37000/109070 [00:12<00:24, 2889.07 examples/s]
Filter:  35%|███▍      | 38000/109070 [00:13<00:24, 2880.90 examples/s]
Filter:  36%|███▌      | 39000/109070 [00:13<00:24, 2886.64 examples/s]
Filter:  37%|███▋      | 40000/109070 [00:13<00:23, 2906.26 examples/s]
Filter:  38%|███▊      | 41000/109070 [00:14<00:23, 2911.57 examples/s]
Filter:  39%|███▊      | 42000/109070 [00:14<00:23, 2905.92 examples/s]
Filter:  39%|███▉      | 43000/109070 [00:14<00:22, 2916.11 examples/s]
Filter:  40%|████      | 44000/109070 [00:15<00:22, 2927.30 examples/s]
Filter:  41%|████▏     | 45000/109070 [00:15<00:21, 2939.16 examples/s]
Filter:  42%|████▏     | 46000/109070 [00:15<00:21, 2951.11 examples/s]
Filter:  43%|████▎     | 47000/109070 [00:16<00:21, 2938.57 examples/s]
Filter:  44%|████▍     | 48000/109070 [00:16<00:20, 2918.01 examples/s]
Filter:  45%|████▍     | 49000/109070 [00:16<00:20, 2906.67 examples/s]
Filter:  46%|████▌     | 50000/109070 [00:17<00:20, 2932.86 examples/s]
Filter:  47%|████▋     | 51000/109070 [00:17<00:19, 2946.14 examples/s]
Filter:  48%|████▊     | 52000/109070 [00:17<00:19, 2925.77 examples/s]
Filter:  49%|████▊     | 53000/109070 [00:18<00:19, 2910.93 examples/s]
Filter:  50%|████▉     | 54000/109070 [00:18<00:19, 2895.51 examples/s]
Filter:  50%|█████     | 55000/109070 [00:18<00:18, 2887.94 examples/s]
Filter:  51%|█████▏    | 56000/109070 [00:19<00:18, 2906.43 examples/s]
Filter:  52%|█████▏    | 57000/109070 [00:19<00:17, 2911.80 examples/s]
Filter:  53%|█████▎    | 58000/109070 [00:19<00:17, 2901.67 examples/s]
Filter:  54%|█████▍    | 59000/109070 [00:20<00:17, 2915.99 examples/s]
Filter:  55%|█████▌    | 60000/109070 [00:20<00:16, 2939.65 examples/s]
Filter:  56%|█████▌    | 61000/109070 [00:20<00:16, 2915.23 examples/s]
Filter:  57%|█████▋    | 62000/109070 [00:21<00:16, 2924.53 examples/s]
Filter:  58%|█████▊    | 63000/109070 [00:21<00:15, 2914.21 examples/s]
Filter:  59%|█████▊    | 64000/109070 [00:21<00:15, 2902.76 examples/s]
Filter:  60%|█████▉    | 65000/109070 [00:22<00:15, 2895.31 examples/s]
Filter:  61%|██████    | 66000/109070 [00:22<00:14, 2891.08 examples/s]
Filter:  61%|██████▏   | 67000/109070 [00:23<00:14, 2888.04 examples/s]
Filter:  62%|██████▏   | 68000/109070 [00:23<00:14, 2892.57 examples/s]
Filter:  63%|██████▎   | 69000/109070 [00:23<00:13, 2914.25 examples/s]
Filter:  64%|██████▍   | 70000/109070 [00:24<00:13, 2964.11 examples/s]
Filter:  65%|██████▌   | 71000/109070 [00:24<00:12, 2972.28 examples/s]
Filter:  66%|██████▌   | 72000/109070 [00:24<00:12, 2974.70 examples/s]
Filter:  67%|██████▋   | 73000/109070 [00:25<00:12, 2968.47 examples/s]
Filter:  68%|██████▊   | 74000/109070 [00:25<00:11, 2946.01 examples/s]
Filter:  69%|██████▉   | 75000/109070 [00:25<00:11, 2941.65 examples/s]
Filter:  70%|██████▉   | 76000/109070 [00:26<00:11, 2956.88 examples/s]
Filter:  71%|███████   | 77000/109070 [00:26<00:10, 2922.82 examples/s]
Filter:  72%|███████▏  | 78000/109070 [00:26<00:10, 2954.28 examples/s]
Filter:  72%|███████▏  | 79000/109070 [00:27<00:10, 2978.47 examples/s]
Filter:  73%|███████▎  | 80000/109070 [00:27<00:09, 2975.15 examples/s]
Filter:  74%|███████▍  | 81000/109070 [00:27<00:09, 2984.77 examples/s]
Filter:  75%|███████▌  | 82000/109070 [00:28<00:09, 2981.58 examples/s]
Filter:  76%|███████▌  | 83000/109070 [00:28<00:08, 2981.68 examples/s]
Filter:  77%|███████▋  | 84000/109070 [00:28<00:08, 2970.77 examples/s]
Filter:  78%|███████▊  | 85000/109070 [00:29<00:08, 2955.47 examples/s]
Filter:  79%|███████▉  | 86000/109070 [00:29<00:07, 2942.32 examples/s]
Filter:  80%|███████▉  | 87000/109070 [00:29<00:07, 2964.49 examples/s]
Filter:  81%|████████  | 88000/109070 [00:30<00:07, 2983.23 examples/s]
Filter:  82%|████████▏ | 89000/109070 [00:30<00:06, 2975.49 examples/s]
Filter:  83%|████████▎ | 90000/109070 [00:31<00:08, 2309.53 examples/s]
Filter:  83%|████████▎ | 91000/109070 [00:31<00:07, 2476.75 examples/s]
Filter:  84%|████████▍ | 92000/109070 [00:31<00:06, 2590.31 examples/s]
Filter:  85%|████████▌ | 93000/109070 [00:32<00:05, 2685.23 examples/s]
Filter:  86%|████████▌ | 94000/109070 [00:32<00:05, 2773.45 examples/s]
Filter:  87%|████████▋ | 95000/109070 [00:32<00:04, 2851.27 examples/s]
Filter:  88%|████████▊ | 96000/109070 [00:33<00:04, 2921.56 examples/s]
Filter:  89%|████████▉ | 97000/109070 [00:33<00:04, 2963.13 examples/s]
Filter:  90%|████████▉ | 98000/109070 [00:33<00:03, 2983.76 examples/s]
Filter:  91%|█████████ | 99000/109070 [00:34<00:03, 2988.55 examples/s]
Filter:  92%|█████████▏| 100000/109070 [00:34<00:03, 2960.68 examples/s]
Filter:  93%|█████████▎| 101000/109070 [00:34<00:02, 2937.28 examples/s]
Filter:  94%|█████████▎| 102000/109070 [00:35<00:02, 2923.66 examples/s]
Filter:  94%|█████████▍| 103000/109070 [00:35<00:02, 2913.11 examples/s]
Filter:  95%|█████████▌| 104000/109070 [00:35<00:01, 2900.46 examples/s]
Filter:  96%|█████████▋| 105000/109070 [00:36<00:01, 2890.36 examples/s]
Filter:  97%|█████████▋| 106000/109070 [00:36<00:01, 2887.36 examples/s]
Filter:  98%|█████████▊| 107000/109070 [00:36<00:00, 2885.96 examples/s]
Filter:  99%|█████████▉| 108000/109070 [00:37<00:00, 2885.88 examples/s]
Filter: 100%|█████████▉| 109000/109070 [00:37<00:00, 2952.73 examples/s]
Filter: 100%|██████████| 109070/109070 [00:37<00:00, 2904.18 examples/s]

Fetching 128 files:   0%|          | 0/128 [00:00<?, ?it/s]
Fetching 128 files: 100%|██████████| 128/128 [00:00<00:00, 14467.01it/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,  1.08it/s]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:01<00:00,  1.07it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.13it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.12it/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-12 22:14:08 [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 22:14:09 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-12 22:14:09 [model.py:1510] Using max model len 1536
INFO 04-12 22:14:09 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-12 22:14:09 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-12 22:14:09 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-12 22:14:09 [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 22:14:10 [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 22:14:11 [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 22:14:11 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-12 22:14:11 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-12 22:14:11 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-12 22:14:11 [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,  2.05it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.75it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.62it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.67it/s]

INFO 04-12 22:14:13 [default_loader.py:267] Loading weights took 1.99 seconds
INFO 04-12 22:14:14 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.423672 seconds
INFO 04-12 22:14:19 [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 22:14:19 [backends.py:559] Dynamo bytecode transform time: 4.88 s
INFO 04-12 22:14:23 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.454 s
INFO 04-12 22:14:24 [monitor.py:34] torch.compile takes 4.88 s in total
INFO 04-12 22:14:25 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-12 22:14:25 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-12 22:14:25 [kv_cache_utils.py:1091] Maximum concurrency for 1,536 tokens per request: 34.80x

Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):   0%|          | 0/5 [00:00<?, ?it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  40%|████      | 2/5 [00:00<00:00, 19.87it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  80%|████████  | 4/5 [00:00<00:00, 19.74it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 16.98it/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, 23.29it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 24.20it/s]
INFO 04-12 22:14:26 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-12 22:14:26 [core.py:210] init engine (profile, create kv cache, warmup model) took 12.44 seconds
INFO 04-12 22:14:27 [llm.py:306] Supported_tasks: ('generate',)
2026-04-12 22:14:27,110 | INFO | train_grpo_train | dataset_id=Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1 train_size=9955 eval_size=80
2026-04-12 22:14:27,110 | INFO | train_grpo_train | source_dataset_id=Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits test_size=80 hard_diagnostic_size=3126
2026-04-12 22:14:27,110 | INFO | train_grpo_train | active_rewards=['meter', 'count_adherence', 'hard_gate', 'judge_quality', 'repeat_soft', 'total_composite']
2026-04-12 22:14:27,110 | INFO | train_grpo_train | reward_weights={'meter': 0.0, 'count_adherence': 0.0, 'hard_gate': 0.0, 'judge_quality': 0.0, 'repeat_soft': 0.0, 'total_composite': 1.0}
2026-04-12 22:14:27,110 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo-short1k-no-trio-v2
2026-04-12 22:14:27,110 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/cap_1000/selected_manifest.csv
2026-04-12 22:14:27,111 | INFO | train_grpo_train | hard_diagnostic_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/hard_diagnostic_cap_256/selected_manifest.csv
2026-04-12 22:14:27,111 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-12 22:14:27,111 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-12T22:14:08Z", "resume_mode": "fresh", "requested_spec": "", "result": "fresh", "reason": "explicit_fresh", "local_resume_path": null, "remote_repo": null, "remote_prefix": null, "compatibility": "unchecked", "config_fingerprint": "002103f46be89ed6311e30cbbf1952cf127698400c47c48f018bf0a5f0eb1c65"}
2026-04-12 22:14:27,111 | 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 22:14:27 [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<6:14:48,  6.82s/it]2026-04-12 22:14:34,355 | INFO | train_grpo_train | metrics_logged mode=train step=1

                                                  
{'loss': 0.1289, 'grad_norm': 24.25541877746582, 'learning_rate': 1e-05, 'num_tokens': 1670.0, 'completions/mean_length': 41.75, 'completions/min_length': 29.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6725290417671204, 'rewards/meter/std': 0.40340036153793335, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957022070884705, 'rewards/repeat_soft/std': 0.00409209867939353, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.09941796213388443, 'rewards/total_composite/mean': 0.6763333082199097, 'rewards/total_composite/std': 0.20289066433906555, 'reward': 0.6763333082199097, 'reward_std': 0.20289064943790436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22603146731853485, 'sampling/sampling_logp_difference/max': 1.5626206398010254, 'sampling/importance_sampling_ratio/min': 0.20958609879016876, 'sampling/importance_sampling_ratio/mean': 1.013262391090393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.01923768222332, 'clip_ratio/low_mean': 0.08834622986614704, 'clip_ratio/low_min': 0.08834622986614704, 'clip_ratio/high_mean': 0.15415428206324577, 'clip_ratio/high_max': 0.15415428206324577, 'clip_ratio/region_mean': 0.24250051192939281, 'reward_total_mean': 0.6763333082199097, 'reward_meter_mean': 0.6725290417671204, 'reward_meter_std': 0.40340036153793335, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957022070884705, 'reward_repeat_soft_std': 0.00409209867939353, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.09941796213388443, 'reward_total_composite_mean': 0.6763333082199097, 'reward_total_composite_std': 0.20289066433906555, 'epoch': 0.0}

  0%|          | 1/3300 [00:06<6:14:48,  6.82s/it]INFO 04-12 22:14:34 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 2/3300 [00:14<6:27:46,  7.05s/it]2026-04-12 22:14:41,560 | INFO | train_grpo_train | metrics_logged mode=train step=2

                                                  
{'loss': 0.0733, 'grad_norm': 9.701464653015137, 'learning_rate': 9.996969696969698e-06, 'num_tokens': 4111.0, 'completions/mean_length': 141.125, 'completions/min_length': 106.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.125, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.667778730392456, 'rewards/meter/std': 0.3386266231536865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9944028854370117, 'rewards/repeat_soft/std': 0.00460322480648756, 'rewards/judge_quality/mean': 0.6575000286102295, 'rewards/judge_quality/std': 0.2133910059928894, 'rewards/total_composite/mean': 0.7471907138824463, 'rewards/total_composite/std': 0.150190070271492, 'reward': 0.7471907138824463, 'reward_std': 0.1501900553703308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2034754604101181, 'sampling/sampling_logp_difference/max': 1.743311882019043, 'sampling/importance_sampling_ratio/min': 0.19116829335689545, 'sampling/importance_sampling_ratio/mean': 1.0195162296295166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9704072922468185, 'clip_ratio/low_mean': 0.05871324986219406, 'clip_ratio/low_min': 0.05871324986219406, 'clip_ratio/high_mean': 0.128573689609766, 'clip_ratio/high_max': 0.128573689609766, 'clip_ratio/region_mean': 0.18728693947196007, 'reward_total_mean': 0.7471907138824463, 'reward_meter_mean': 0.667778730392456, 'reward_meter_std': 0.3386266231536865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9944028854370117, 'reward_repeat_soft_std': 0.00460322480648756, 'reward_judge_quality_mean': 0.6575000286102295, 'reward_judge_quality_std': 0.2133910059928894, 'reward_total_composite_mean': 0.7471907138824463, 'reward_total_composite_std': 0.150190070271492, 'epoch': 0.0}

  0%|          | 2/3300 [00:14<6:27:46,  7.05s/it]INFO 04-12 22:14:41 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 3/3300 [00:19<5:52:57,  6.42s/it]2026-04-12 22:14:47,230 | INFO | train_grpo_train | metrics_logged mode=train step=3

                                                  
{'loss': 0.0512, 'grad_norm': 21.948942184448242, 'learning_rate': 9.993939393939395e-06, 'num_tokens': 5819.0, 'completions/mean_length': 36.5, 'completions/min_length': 24.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.34533801674842834, 'rewards/meter/std': 0.3437061011791229, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9806680679321289, 'rewards/repeat_soft/std': 0.02873978205025196, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.5868439078330994, 'rewards/total_composite/std': 0.20601670444011688, 'reward': 0.5868439078330994, 'reward_std': 0.20601673424243927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.319444864988327, 'sampling/sampling_logp_difference/max': 2.2889962196350098, 'sampling/importance_sampling_ratio/min': 0.10136815905570984, 'sampling/importance_sampling_ratio/mean': 0.9693852066993713, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3289805501699448, 'clip_ratio/low_mean': 0.15689394064247608, 'clip_ratio/low_min': 0.15689394064247608, 'clip_ratio/high_mean': 0.07748956605792046, 'clip_ratio/high_max': 0.07748956605792046, 'clip_ratio/region_mean': 0.23438350670039654, 'reward_total_mean': 0.5868439078330994, 'reward_meter_mean': 0.34533801674842834, 'reward_meter_std': 0.3437061011791229, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9806680679321289, 'reward_repeat_soft_std': 0.02873978205025196, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.5868439078330994, 'reward_total_composite_std': 0.20601670444011688, 'epoch': 0.0}

  0%|          | 3/3300 [00:19<5:52:57,  6.42s/it]INFO 04-12 22:14:47 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 4/3300 [00:26<5:55:16,  6.47s/it]2026-04-12 22:14:53,766 | INFO | train_grpo_train | metrics_logged mode=train step=4

                                                  
{'loss': 0.0029, 'grad_norm': 17.112354278564453, 'learning_rate': 9.990909090909093e-06, 'num_tokens': 7420.0, 'completions/mean_length': 43.125, 'completions/min_length': 35.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6681925058364868, 'rewards/meter/std': 0.34428924322128296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9990890026092529, 'rewards/repeat_soft/std': 0.00257677398622036, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.2377273440361023, 'rewards/total_composite/mean': 0.7650955319404602, 'rewards/total_composite/std': 0.1345381885766983, 'reward': 0.7650955319404602, 'reward_std': 0.1345381736755371, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21286356449127197, 'sampling/sampling_logp_difference/max': 1.6740360260009766, 'sampling/importance_sampling_ratio/min': 0.18748880922794342, 'sampling/importance_sampling_ratio/mean': 1.0222951173782349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9584360122680664, 'clip_ratio/low_mean': 0.08466819114983082, 'clip_ratio/low_min': 0.08466819114983082, 'clip_ratio/high_mean': 0.13993905391544104, 'clip_ratio/high_max': 0.13993905391544104, 'clip_ratio/region_mean': 0.22460724506527185, 'reward_total_mean': 0.7650955319404602, 'reward_meter_mean': 0.6681925058364868, 'reward_meter_std': 0.34428924322128296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9990890026092529, 'reward_repeat_soft_std': 0.00257677398622036, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.2377273440361023, 'reward_total_composite_mean': 0.7650955319404602, 'reward_total_composite_std': 0.1345381885766983, 'epoch': 0.0}

  0%|          | 4/3300 [00:26<5:55:16,  6.47s/it]INFO 04-12 22:14:53 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 5/3300 [00:33<6:03:14,  6.61s/it]2026-04-12 22:15:00,641 | INFO | train_grpo_train | metrics_logged mode=train step=5

                                                  
{'loss': 0.0895, 'grad_norm': 9.455750465393066, 'learning_rate': 9.987878787878788e-06, 'num_tokens': 9686.0, 'completions/mean_length': 106.25, 'completions/min_length': 71.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.25, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.5534956455230713, 'rewards/meter/std': 0.35001081228256226, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9964775443077087, 'rewards/repeat_soft/std': 0.0026533447671681643, 'rewards/judge_quality/mean': 0.7099999785423279, 'rewards/judge_quality/std': 0.24213339388370514, 'rewards/total_composite/mean': 0.7070332765579224, 'rewards/total_composite/std': 0.182778000831604, 'reward': 0.7070332765579224, 'reward_std': 0.1827780157327652, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19017860293388367, 'sampling/sampling_logp_difference/max': 1.9287948608398438, 'sampling/importance_sampling_ratio/min': 0.14532321691513062, 'sampling/importance_sampling_ratio/mean': 1.021316409111023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3343858867883682, 'clip_ratio/low_mean': 0.0882453341037035, 'clip_ratio/low_min': 0.0882453341037035, 'clip_ratio/high_mean': 0.0963182132691145, 'clip_ratio/high_max': 0.0963182132691145, 'clip_ratio/region_mean': 0.184563547372818, 'reward_total_mean': 0.7070332765579224, 'reward_meter_mean': 0.5534956455230713, 'reward_meter_std': 0.35001081228256226, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9964775443077087, 'reward_repeat_soft_std': 0.0026533447671681643, 'reward_judge_quality_mean': 0.7099999785423279, 'reward_judge_quality_std': 0.24213339388370514, 'reward_total_composite_mean': 0.7070332765579224, 'reward_total_composite_std': 0.182778000831604, 'epoch': 0.0}

  0%|          | 5/3300 [00:33<6:03:14,  6.61s/it]INFO 04-12 22:15:00 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 6/3300 [00:40<6:11:26,  6.77s/it]2026-04-12 22:15:07,738 | INFO | train_grpo_train | metrics_logged mode=train step=6

                                                  
{'loss': 0.1401, 'grad_norm': 9.252427101135254, 'learning_rate': 9.984848484848485e-06, 'num_tokens': 12284.0, 'completions/mean_length': 125.75, 'completions/min_length': 85.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.75, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.7442054748535156, 'rewards/meter/std': 0.36003756523132324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9989632368087769, 'rewards/repeat_soft/std': 0.0006836142274551094, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.7295387983322144, 'rewards/total_composite/std': 0.1836954951286316, 'reward': 0.7295387983322144, 'reward_std': 0.1836954802274704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20029257237911224, 'sampling/sampling_logp_difference/max': 1.7579917907714844, 'sampling/importance_sampling_ratio/min': 0.17239071428775787, 'sampling/importance_sampling_ratio/mean': 1.0128437280654907, 'sampling/importance_sampling_ratio/max': 1.9965848922729492, 'entropy': 2.4172347486019135, 'clip_ratio/low_mean': 0.07391603104770184, 'clip_ratio/low_min': 0.07391603104770184, 'clip_ratio/high_mean': 0.1392985973507166, 'clip_ratio/high_max': 0.1392985973507166, 'clip_ratio/region_mean': 0.21321462839841843, 'reward_total_mean': 0.7295387983322144, 'reward_meter_mean': 0.7442054748535156, 'reward_meter_std': 0.36003756523132324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9989632368087769, 'reward_repeat_soft_std': 0.0006836142274551094, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.7295387983322144, 'reward_total_composite_std': 0.1836954951286316, 'epoch': 0.0}

  0%|          | 6/3300 [00:40<6:11:26,  6.77s/it]INFO 04-12 22:15:07 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 7/3300 [00:47<6:15:13,  6.84s/it]2026-04-12 22:15:14,683 | INFO | train_grpo_train | metrics_logged mode=train step=7

                                                  
{'loss': -0.0576, 'grad_norm': 13.074331283569336, 'learning_rate': 9.981818181818183e-06, 'num_tokens': 14661.0, 'completions/mean_length': 113.125, 'completions/min_length': 84.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.125, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.7239573001861572, 'rewards/meter/std': 0.26627305150032043, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9978160858154297, 'rewards/repeat_soft/std': 0.001829512882977724, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.7315623760223389, 'rewards/total_composite/std': 0.12214145064353943, 'reward': 0.7315623760223389, 'reward_std': 0.12214145064353943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2251056730747223, 'sampling/sampling_logp_difference/max': 1.424403190612793, 'sampling/importance_sampling_ratio/min': 0.24065205454826355, 'sampling/importance_sampling_ratio/mean': 1.03545081615448, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.566889613866806, 'clip_ratio/low_mean': 0.07710661552846432, 'clip_ratio/low_min': 0.07710661552846432, 'clip_ratio/high_mean': 0.1532231941819191, 'clip_ratio/high_max': 0.1532231941819191, 'clip_ratio/region_mean': 0.23032980971038342, 'reward_total_mean': 0.7315623760223389, 'reward_meter_mean': 0.7239573001861572, 'reward_meter_std': 0.26627305150032043, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9978160858154297, 'reward_repeat_soft_std': 0.001829512882977724, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.7315623760223389, 'reward_total_composite_std': 0.12214145064353943, 'epoch': 0.0}

  0%|          | 7/3300 [00:47<6:15:13,  6.84s/it]INFO 04-12 22:15:14 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 8/3300 [00:52<5:49:19,  6.37s/it]2026-04-12 22:15:20,044 | INFO | train_grpo_train | metrics_logged mode=train step=8

                                                  
{'loss': 0.0692, 'grad_norm': 37.71578598022461, 'learning_rate': 9.97878787878788e-06, 'num_tokens': 16079.0, 'completions/mean_length': 21.25, 'completions/min_length': 16.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.8553460836410522, 'rewards/meter/std': 0.32826167345046997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9081118702888489, 'rewards/repeat_soft/std': 0.10070746392011642, 'rewards/judge_quality/mean': 0.65625, 'rewards/judge_quality/std': 0.29621124267578125, 'rewards/total_composite/mean': 0.8225919008255005, 'rewards/total_composite/std': 0.1914392113685608, 'reward': 0.8225919008255005, 'reward_std': 0.1914391964673996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20113249123096466, 'sampling/sampling_logp_difference/max': 1.2836437225341797, 'sampling/importance_sampling_ratio/min': 0.2770260274410248, 'sampling/importance_sampling_ratio/mean': 1.0172845125198364, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6074406504631042, 'clip_ratio/low_mean': 0.06774115934967995, 'clip_ratio/low_min': 0.06774115934967995, 'clip_ratio/high_mean': 0.07261904887855053, 'clip_ratio/high_max': 0.07261904887855053, 'clip_ratio/region_mean': 0.14036020822823048, 'reward_total_mean': 0.8225919008255005, 'reward_meter_mean': 0.8553460836410522, 'reward_meter_std': 0.32826167345046997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9081118702888489, 'reward_repeat_soft_std': 0.10070746392011642, 'reward_judge_quality_mean': 0.65625, 'reward_judge_quality_std': 0.29621124267578125, 'reward_total_composite_mean': 0.8225919008255005, 'reward_total_composite_std': 0.1914392113685608, 'epoch': 0.0}

  0%|          | 8/3300 [00:52<5:49:19,  6.37s/it]INFO 04-12 22:15:20 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 9/3300 [01:00<6:14:36,  6.83s/it]2026-04-12 22:15:27,891 | INFO | train_grpo_train | metrics_logged mode=train step=9

                                                  
{'loss': 0.0029, 'grad_norm': 9.873559951782227, 'learning_rate': 9.975757575757577e-06, 'num_tokens': 18924.0, 'completions/mean_length': 149.625, 'completions/min_length': 105.0, 'completions/max_length': 190.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 149.625, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 190.0, 'rewards/meter/mean': 0.8943005800247192, 'rewards/meter/std': 0.12937459349632263, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9977504014968872, 'rewards/repeat_soft/std': 0.0031607486307621002, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.713111937046051, 'rewards/total_composite/std': 0.29207462072372437, 'reward': 0.713111937046051, 'reward_std': 0.29207465052604675, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21067966520786285, 'sampling/sampling_logp_difference/max': 1.7020273208618164, 'sampling/importance_sampling_ratio/min': 0.18231353163719177, 'sampling/importance_sampling_ratio/mean': 1.0195698738098145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.251196339726448, 'clip_ratio/low_mean': 0.01953125, 'clip_ratio/low_min': 0.01953125, 'clip_ratio/high_mean': 0.2094412874430418, 'clip_ratio/high_max': 0.2094412874430418, 'clip_ratio/region_mean': 0.2289725374430418, 'reward_total_mean': 0.713111937046051, 'reward_meter_mean': 0.8943005800247192, 'reward_meter_std': 0.12937459349632263, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9977504014968872, 'reward_repeat_soft_std': 0.0031607486307621002, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.713111937046051, 'reward_total_composite_std': 0.29207462072372437, 'epoch': 0.0}

  0%|          | 9/3300 [01:00<6:14:36,  6.83s/it]INFO 04-12 22:15:28 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 10/3300 [01:06<6:04:38,  6.65s/it]2026-04-12 22:15:34,140 | INFO | train_grpo_train | metrics_logged mode=train step=10

                                                   
{'loss': 0.0833, 'grad_norm': 25.13707733154297, 'learning_rate': 9.972727272727274e-06, 'num_tokens': 20493.0, 'completions/mean_length': 34.125, 'completions/min_length': 25.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6848095059394836, 'rewards/meter/std': 0.4096945524215698, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9840165376663208, 'rewards/repeat_soft/std': 0.029735958203673363, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6806565523147583, 'rewards/total_composite/std': 0.34453243017196655, 'reward': 0.6806565523147583, 'reward_std': 0.34453243017196655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21705979108810425, 'sampling/sampling_logp_difference/max': 1.3225932121276855, 'sampling/importance_sampling_ratio/min': 0.2664434611797333, 'sampling/importance_sampling_ratio/mean': 1.0199761390686035, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.476714864373207, 'clip_ratio/low_mean': 0.0765726137906313, 'clip_ratio/low_min': 0.0765726137906313, 'clip_ratio/high_mean': 0.11742378678172827, 'clip_ratio/high_max': 0.11742378678172827, 'clip_ratio/region_mean': 0.19399640057235956, 'reward_total_mean': 0.6806565523147583, 'reward_meter_mean': 0.6848095059394836, 'reward_meter_std': 0.4096945524215698, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9840165376663208, 'reward_repeat_soft_std': 0.029735958203673363, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6806565523147583, 'reward_total_composite_std': 0.34453243017196655, 'epoch': 0.0}

  0%|          | 10/3300 [01:06<6:04:38,  6.65s/it]INFO 04-12 22:15:34 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 11/3300 [01:12<5:50:00,  6.39s/it]2026-04-12 22:15:39,923 | INFO | train_grpo_train | metrics_logged mode=train step=11

                                                   
{'loss': 0.1787, 'grad_norm': 18.680042266845703, 'learning_rate': 9.96969696969697e-06, 'num_tokens': 22069.0, 'completions/mean_length': 47.0, 'completions/min_length': 28.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.6532547473907471, 'rewards/meter/std': 0.3094608187675476, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9933717846870422, 'rewards/repeat_soft/std': 0.007688070647418499, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.7180517911911011, 'rewards/total_composite/std': 0.15922589600086212, 'reward': 0.7180517911911011, 'reward_std': 0.15922589600086212, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20565736293792725, 'sampling/sampling_logp_difference/max': 0.9617314338684082, 'sampling/importance_sampling_ratio/min': 0.3822305202484131, 'sampling/importance_sampling_ratio/mean': 1.0575566291809082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3583329617977142, 'clip_ratio/low_mean': 0.09450654499232769, 'clip_ratio/low_min': 0.09450654499232769, 'clip_ratio/high_mean': 0.12003943882882595, 'clip_ratio/high_max': 0.12003943882882595, 'clip_ratio/region_mean': 0.21454598382115364, 'reward_total_mean': 0.7180517911911011, 'reward_meter_mean': 0.6532547473907471, 'reward_meter_std': 0.3094608187675476, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9933717846870422, 'reward_repeat_soft_std': 0.007688070647418499, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.7180517911911011, 'reward_total_composite_std': 0.15922589600086212, 'epoch': 0.0}

  0%|          | 11/3300 [01:12<5:50:00,  6.39s/it]INFO 04-12 22:15:40 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 12/3300 [01:17<5:34:13,  6.10s/it]2026-04-12 22:15:45,369 | INFO | train_grpo_train | metrics_logged mode=train step=12

                                                   
{'loss': 0.0339, 'grad_norm': 28.459325790405273, 'learning_rate': 9.966666666666667e-06, 'num_tokens': 23560.0, 'completions/mean_length': 26.375, 'completions/min_length': 16.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6143099069595337, 'rewards/meter/std': 0.47742435336112976, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9266142845153809, 'rewards/repeat_soft/std': 0.07501126825809479, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.09941794723272324, 'rewards/total_composite/mean': 0.637225866317749, 'rewards/total_composite/std': 0.23131398856639862, 'reward': 0.637225866317749, 'reward_std': 0.23131398856639862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2075686752796173, 'sampling/sampling_logp_difference/max': 1.0368385314941406, 'sampling/importance_sampling_ratio/min': 0.3545738756656647, 'sampling/importance_sampling_ratio/mean': 1.0258712768554688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.740403950214386, 'clip_ratio/low_mean': 0.06502757407724857, 'clip_ratio/low_min': 0.06502757407724857, 'clip_ratio/high_mean': 0.10130646172910929, 'clip_ratio/high_max': 0.10130646172910929, 'clip_ratio/region_mean': 0.16633403580635786, 'reward_total_mean': 0.637225866317749, 'reward_meter_mean': 0.6143099069595337, 'reward_meter_std': 0.47742435336112976, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9266142845153809, 'reward_repeat_soft_std': 0.07501126825809479, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.09941794723272324, 'reward_total_composite_mean': 0.637225866317749, 'reward_total_composite_std': 0.23131398856639862, 'epoch': 0.0}

  0%|          | 12/3300 [01:17<5:34:13,  6.10s/it]INFO 04-12 22:15:45 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 13/3300 [01:23<5:26:44,  5.96s/it]2026-04-12 22:15:51,099 | INFO | train_grpo_train | metrics_logged mode=train step=13

                                                   
{'loss': -0.0121, 'grad_norm': 21.01743507385254, 'learning_rate': 9.963636363636364e-06, 'num_tokens': 25175.0, 'completions/mean_length': 44.875, 'completions/min_length': 30.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7244622707366943, 'rewards/meter/std': 0.4260440170764923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9886103868484497, 'rewards/repeat_soft/std': 0.017639046534895897, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.6614029407501221, 'rewards/total_composite/std': 0.2924328148365021, 'reward': 0.6614029407501221, 'reward_std': 0.2924328148365021, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21056580543518066, 'sampling/sampling_logp_difference/max': 1.903547763824463, 'sampling/importance_sampling_ratio/min': 0.1796707957983017, 'sampling/importance_sampling_ratio/mean': 1.0298173427581787, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.938667967915535, 'clip_ratio/low_mean': 0.06033376231789589, 'clip_ratio/low_min': 0.06033376231789589, 'clip_ratio/high_mean': 0.13869474548846483, 'clip_ratio/high_max': 0.13869474548846483, 'clip_ratio/region_mean': 0.19902850780636072, 'reward_total_mean': 0.6614029407501221, 'reward_meter_mean': 0.7244622707366943, 'reward_meter_std': 0.4260440170764923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9886103868484497, 'reward_repeat_soft_std': 0.017639046534895897, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.6614029407501221, 'reward_total_composite_std': 0.2924328148365021, 'epoch': 0.0}

  0%|          | 13/3300 [01:23<5:26:44,  5.96s/it]INFO 04-12 22:15:51 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 14/3300 [01:30<5:50:32,  6.40s/it]2026-04-12 22:15:58,432 | INFO | train_grpo_train | metrics_logged mode=train step=14

                                                   
{'loss': 0.1867, 'grad_norm': 9.992451667785645, 'learning_rate': 9.960606060606062e-06, 'num_tokens': 27778.0, 'completions/mean_length': 131.375, 'completions/min_length': 89.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.8847721815109253, 'rewards/meter/std': 0.2510175108909607, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9979761838912964, 'rewards/repeat_soft/std': 0.0023726883810013533, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.21836651861667633, 'rewards/total_composite/mean': 0.8200700283050537, 'rewards/total_composite/std': 0.16812320053577423, 'reward': 0.8200700283050537, 'reward_std': 0.16812318563461304, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2101648896932602, 'sampling/sampling_logp_difference/max': 1.9604809284210205, 'sampling/importance_sampling_ratio/min': 0.14079070091247559, 'sampling/importance_sampling_ratio/mean': 1.0326483249664307, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4273063838481903, 'clip_ratio/low_mean': 0.05659673362970352, 'clip_ratio/low_min': 0.05659673362970352, 'clip_ratio/high_mean': 0.15316884219646454, 'clip_ratio/high_max': 0.15316884219646454, 'clip_ratio/region_mean': 0.20976557582616806, 'reward_total_mean': 0.8200700283050537, 'reward_meter_mean': 0.8847721815109253, 'reward_meter_std': 0.2510175108909607, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9979761838912964, 'reward_repeat_soft_std': 0.0023726883810013533, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.21836651861667633, 'reward_total_composite_mean': 0.8200700283050537, 'reward_total_composite_std': 0.16812320053577423, 'epoch': 0.0}

  0%|          | 14/3300 [01:30<5:50:32,  6.40s/it]INFO 04-12 22:15:58 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 15/3300 [01:36<5:40:04,  6.21s/it]2026-04-12 22:16:04,205 | INFO | train_grpo_train | metrics_logged mode=train step=15

                                                   
{'loss': 0.0293, 'grad_norm': 16.65377426147461, 'learning_rate': 9.957575757575757e-06, 'num_tokens': 29486.0, 'completions/mean_length': 48.5, 'completions/min_length': 25.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.4189527928829193, 'rewards/meter/std': 0.3443324863910675, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9987918138504028, 'rewards/repeat_soft/std': 0.00322684645652771, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.5490054488182068, 'rewards/total_composite/std': 0.2890552580356598, 'reward': 0.5490054488182068, 'reward_std': 0.2890552580356598, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19479593634605408, 'sampling/sampling_logp_difference/max': 2.340357780456543, 'sampling/importance_sampling_ratio/min': 0.09629317373037338, 'sampling/importance_sampling_ratio/mean': 1.0266320705413818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6842268258333206, 'clip_ratio/low_mean': 0.10162054747343063, 'clip_ratio/low_min': 0.10162054747343063, 'clip_ratio/high_mean': 0.11235708184540272, 'clip_ratio/high_max': 0.11235708184540272, 'clip_ratio/region_mean': 0.21397762931883335, 'reward_total_mean': 0.5490054488182068, 'reward_meter_mean': 0.4189527928829193, 'reward_meter_std': 0.3443324863910675, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9987918138504028, 'reward_repeat_soft_std': 0.00322684645652771, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.5490054488182068, 'reward_total_composite_std': 0.2890552580356598, 'epoch': 0.0}

  0%|          | 15/3300 [01:36<5:40:04,  6.21s/it]INFO 04-12 22:16:04 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 16/3300 [01:42<5:33:31,  6.09s/it]2026-04-12 22:16:10,026 | INFO | train_grpo_train | metrics_logged mode=train step=16

                                                   
{'loss': -0.0032, 'grad_norm': 16.885639190673828, 'learning_rate': 9.954545454545456e-06, 'num_tokens': 31057.0, 'completions/mean_length': 45.375, 'completions/min_length': 27.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.43825337290763855, 'rewards/meter/std': 0.4620993435382843, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9892818331718445, 'rewards/repeat_soft/std': 0.01204143650829792, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5848921537399292, 'rewards/total_composite/std': 0.20833450555801392, 'reward': 0.5848921537399292, 'reward_std': 0.20833447575569153, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27420946955680847, 'sampling/sampling_logp_difference/max': 2.7227797508239746, 'sampling/importance_sampling_ratio/min': 0.06569189578294754, 'sampling/importance_sampling_ratio/mean': 1.0040487051010132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9974622428417206, 'clip_ratio/low_mean': 0.07798896264284849, 'clip_ratio/low_min': 0.07798896264284849, 'clip_ratio/high_mean': 0.11831947416067123, 'clip_ratio/high_max': 0.11831947416067123, 'clip_ratio/region_mean': 0.19630843680351973, 'reward_total_mean': 0.5848921537399292, 'reward_meter_mean': 0.43825337290763855, 'reward_meter_std': 0.4620993435382843, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9892818331718445, 'reward_repeat_soft_std': 0.01204143650829792, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5848921537399292, 'reward_total_composite_std': 0.20833450555801392, 'epoch': 0.0}

  0%|          | 16/3300 [01:42<5:33:31,  6.09s/it]INFO 04-12 22:16:10 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 17/3300 [01:48<5:30:38,  6.04s/it]2026-04-12 22:16:15,951 | INFO | train_grpo_train | metrics_logged mode=train step=17

                                                   
{'loss': 0.081, 'grad_norm': 18.933856964111328, 'learning_rate': 9.951515151515152e-06, 'num_tokens': 32706.0, 'completions/mean_length': 47.125, 'completions/min_length': 31.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5159446001052856, 'rewards/meter/std': 0.4337084889411926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9935091733932495, 'rewards/repeat_soft/std': 0.0063369860872626305, 'rewards/judge_quality/mean': 0.5475000143051147, 'rewards/judge_quality/std': 0.14320316910743713, 'rewards/total_composite/mean': 0.6457759737968445, 'rewards/total_composite/std': 0.2059926986694336, 'reward': 0.6457759737968445, 'reward_std': 0.2059926986694336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21604323387145996, 'sampling/sampling_logp_difference/max': 1.6152591705322266, 'sampling/importance_sampling_ratio/min': 0.19883912801742554, 'sampling/importance_sampling_ratio/mean': 1.0173397064208984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0722051858901978, 'clip_ratio/low_mean': 0.11042729578912258, 'clip_ratio/low_min': 0.11042729578912258, 'clip_ratio/high_mean': 0.11535557918250561, 'clip_ratio/high_max': 0.11535557918250561, 'clip_ratio/region_mean': 0.2257828749716282, 'reward_total_mean': 0.6457759737968445, 'reward_meter_mean': 0.5159446001052856, 'reward_meter_std': 0.4337084889411926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9935091733932495, 'reward_repeat_soft_std': 0.0063369860872626305, 'reward_judge_quality_mean': 0.5475000143051147, 'reward_judge_quality_std': 0.14320316910743713, 'reward_total_composite_mean': 0.6457759737968445, 'reward_total_composite_std': 0.2059926986694336, 'epoch': 0.0}

  1%|          | 17/3300 [01:48<5:30:38,  6.04s/it]INFO 04-12 22:16:16 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 18/3300 [01:54<5:29:33,  6.02s/it]2026-04-12 22:16:21,933 | INFO | train_grpo_train | metrics_logged mode=train step=18

                                                   
{'loss': -0.0019, 'grad_norm': 17.924867630004883, 'learning_rate': 9.948484848484849e-06, 'num_tokens': 34138.0, 'completions/mean_length': 30.0, 'completions/min_length': 19.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.583729088306427, 'rewards/meter/std': 0.350315123796463, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.10507649928331375, 'rewards/total_composite/mean': 0.631553053855896, 'rewards/total_composite/std': 0.17265585064888, 'reward': 0.631553053855896, 'reward_std': 0.17265585064888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20634080469608307, 'sampling/sampling_logp_difference/max': 1.4612393379211426, 'sampling/importance_sampling_ratio/min': 0.2319486439228058, 'sampling/importance_sampling_ratio/mean': 1.0241836309432983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7248151153326035, 'clip_ratio/low_mean': 0.10481186397373676, 'clip_ratio/low_min': 0.10481186397373676, 'clip_ratio/high_mean': 0.12524540536105633, 'clip_ratio/high_max': 0.12524540536105633, 'clip_ratio/region_mean': 0.2300572693347931, 'reward_total_mean': 0.631553053855896, 'reward_meter_mean': 0.583729088306427, 'reward_meter_std': 0.350315123796463, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.10507649928331375, 'reward_total_composite_mean': 0.631553053855896, 'reward_total_composite_std': 0.17265585064888, 'epoch': 0.0}

  1%|          | 18/3300 [01:54<5:29:33,  6.02s/it]INFO 04-12 22:16:22 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 19/3300 [02:00<5:33:19,  6.10s/it]2026-04-12 22:16:28,193 | INFO | train_grpo_train | metrics_logged mode=train step=19

                                                   
{'loss': -0.0696, 'grad_norm': 14.65036392211914, 'learning_rate': 9.945454545454546e-06, 'num_tokens': 35798.0, 'completions/mean_length': 52.5, 'completions/min_length': 32.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8749610185623169, 'rewards/meter/std': 0.2477712482213974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9977357387542725, 'rewards/repeat_soft/std': 0.002847629599273205, 'rewards/judge_quality/mean': 0.5550000071525574, 'rewards/judge_quality/std': 0.33393755555152893, 'rewards/total_composite/mean': 0.8100060224533081, 'rewards/total_composite/std': 0.1606670320034027, 'reward': 0.8100060224533081, 'reward_std': 0.1606670320034027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20267684757709503, 'sampling/sampling_logp_difference/max': 1.836294412612915, 'sampling/importance_sampling_ratio/min': 0.15940703451633453, 'sampling/importance_sampling_ratio/mean': 1.0041242837905884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5146896243095398, 'clip_ratio/low_mean': 0.041744403541088104, 'clip_ratio/low_min': 0.041744403541088104, 'clip_ratio/high_mean': 0.1479868022724986, 'clip_ratio/high_max': 0.1479868022724986, 'clip_ratio/region_mean': 0.1897312058135867, 'reward_total_mean': 0.8100060224533081, 'reward_meter_mean': 0.8749610185623169, 'reward_meter_std': 0.2477712482213974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9977357387542725, 'reward_repeat_soft_std': 0.002847629599273205, 'reward_judge_quality_mean': 0.5550000071525574, 'reward_judge_quality_std': 0.33393755555152893, 'reward_total_composite_mean': 0.8100060224533081, 'reward_total_composite_std': 0.1606670320034027, 'epoch': 0.0}

  1%|          | 19/3300 [02:00<5:33:19,  6.10s/it]INFO 04-12 22:16:28 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 20/3300 [02:07<5:46:43,  6.34s/it]2026-04-12 22:16:35,112 | INFO | train_grpo_train | metrics_logged mode=train step=20

                                                   
{'loss': 0.1452, 'grad_norm': 12.367934226989746, 'learning_rate': 9.942424242424244e-06, 'num_tokens': 38169.0, 'completions/mean_length': 109.375, 'completions/min_length': 66.0, 'completions/max_length': 165.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.7363971471786499, 'rewards/meter/std': 0.274208128452301, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973340034484863, 'rewards/repeat_soft/std': 0.002601567655801773, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.6832370758056641, 'rewards/total_composite/std': 0.1437705010175705, 'reward': 0.6832370758056641, 'reward_std': 0.1437705010175705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20733380317687988, 'sampling/sampling_logp_difference/max': 1.9372377395629883, 'sampling/importance_sampling_ratio/min': 0.14410145580768585, 'sampling/importance_sampling_ratio/mean': 1.012389898300171, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.014356479048729, 'clip_ratio/low_mean': 0.09246364235877991, 'clip_ratio/low_min': 0.09246364235877991, 'clip_ratio/high_mean': 0.11669936217367649, 'clip_ratio/high_max': 0.11669936217367649, 'clip_ratio/region_mean': 0.2091630045324564, 'reward_total_mean': 0.6832370758056641, 'reward_meter_mean': 0.7363971471786499, 'reward_meter_std': 0.274208128452301, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973340034484863, 'reward_repeat_soft_std': 0.002601567655801773, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.6832370758056641, 'reward_total_composite_std': 0.1437705010175705, 'epoch': 0.0}

  1%|          | 20/3300 [02:07<5:46:43,  6.34s/it]INFO 04-12 22:16:35 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 21/3300 [02:13<5:37:30,  6.18s/it]2026-04-12 22:16:40,898 | INFO | train_grpo_train | metrics_logged mode=train step=21

                                                   
{'loss': 0.0954, 'grad_norm': 33.46713638305664, 'learning_rate': 9.939393939393939e-06, 'num_tokens': 39729.0, 'completions/mean_length': 36.0, 'completions/min_length': 26.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.24060703814029694, 'rewards/meter/std': 0.28769800066947937, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9997790455818176, 'rewards/repeat_soft/std': 0.0006249745492823422, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.5978760719299316, 'rewards/total_composite/std': 0.15713509917259216, 'reward': 0.5978760719299316, 'reward_std': 0.15713508427143097, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2637569010257721, 'sampling/sampling_logp_difference/max': 2.1834075450897217, 'sampling/importance_sampling_ratio/min': 0.11265699565410614, 'sampling/importance_sampling_ratio/mean': 1.0002814531326294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2168454974889755, 'clip_ratio/low_mean': 0.18306511640548706, 'clip_ratio/low_min': 0.18306511640548706, 'clip_ratio/high_mean': 0.03896103985607624, 'clip_ratio/high_max': 0.03896103985607624, 'clip_ratio/region_mean': 0.2220261562615633, 'reward_total_mean': 0.5978760719299316, 'reward_meter_mean': 0.24060703814029694, 'reward_meter_std': 0.28769800066947937, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9997790455818176, 'reward_repeat_soft_std': 0.0006249745492823422, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.5978760719299316, 'reward_total_composite_std': 0.15713509917259216, 'epoch': 0.0}

  1%|          | 21/3300 [02:13<5:37:30,  6.18s/it]INFO 04-12 22:16:41 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 22/3300 [02:19<5:31:15,  6.06s/it]2026-04-12 22:16:46,699 | INFO | train_grpo_train | metrics_logged mode=train step=22

                                                   
{'loss': 0.0519, 'grad_norm': 14.068692207336426, 'learning_rate': 9.936363636363638e-06, 'num_tokens': 41479.0, 'completions/mean_length': 59.75, 'completions/min_length': 39.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9757623672485352, 'rewards/meter/std': 0.020018933340907097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9967460632324219, 'rewards/repeat_soft/std': 0.005034734960645437, 'rewards/judge_quality/mean': 0.6024999618530273, 'rewards/judge_quality/std': 0.2521762549877167, 'rewards/total_composite/mean': 0.8695176839828491, 'rewards/total_composite/std': 0.07778604328632355, 'reward': 0.8695176839828491, 'reward_std': 0.07778605818748474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20508892834186554, 'sampling/sampling_logp_difference/max': 1.620192527770996, 'sampling/importance_sampling_ratio/min': 0.19786059856414795, 'sampling/importance_sampling_ratio/mean': 1.0207918882369995, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.071342244744301, 'clip_ratio/low_mean': 0.10649552661925554, 'clip_ratio/low_min': 0.10649552661925554, 'clip_ratio/high_mean': 0.07769765146076679, 'clip_ratio/high_max': 0.07769765146076679, 'clip_ratio/region_mean': 0.18419317808002234, 'reward_total_mean': 0.8695176839828491, 'reward_meter_mean': 0.9757623672485352, 'reward_meter_std': 0.020018933340907097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9967460632324219, 'reward_repeat_soft_std': 0.005034734960645437, 'reward_judge_quality_mean': 0.6024999618530273, 'reward_judge_quality_std': 0.2521762549877167, 'reward_total_composite_mean': 0.8695176839828491, 'reward_total_composite_std': 0.07778604328632355, 'epoch': 0.0}

  1%|          | 22/3300 [02:19<5:31:15,  6.06s/it]INFO 04-12 22:16:46 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 23/3300 [02:24<5:25:01,  5.95s/it]2026-04-12 22:16:52,388 | INFO | train_grpo_train | metrics_logged mode=train step=23

                                                   
{'loss': 0.0951, 'grad_norm': 25.26443099975586, 'learning_rate': 9.933333333333334e-06, 'num_tokens': 43076.0, 'completions/mean_length': 39.625, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.41748276352882385, 'rewards/meter/std': 0.4426417350769043, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921599626541138, 'rewards/repeat_soft/std': 0.009428190067410469, 'rewards/judge_quality/mean': 0.5612500309944153, 'rewards/judge_quality/std': 0.1968638300895691, 'rewards/total_composite/mean': 0.6054582595825195, 'rewards/total_composite/std': 0.19344164431095123, 'reward': 0.6054582595825195, 'reward_std': 0.19344164431095123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1857699751853943, 'sampling/sampling_logp_difference/max': 1.0207924842834473, 'sampling/importance_sampling_ratio/min': 0.36030930280685425, 'sampling/importance_sampling_ratio/mean': 1.0526957511901855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6756559312343597, 'clip_ratio/low_mean': 0.11134098190814257, 'clip_ratio/low_min': 0.11134098190814257, 'clip_ratio/high_mean': 0.04910453222692013, 'clip_ratio/high_max': 0.04910453222692013, 'clip_ratio/region_mean': 0.1604455141350627, 'reward_total_mean': 0.6054582595825195, 'reward_meter_mean': 0.41748276352882385, 'reward_meter_std': 0.4426417350769043, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921599626541138, 'reward_repeat_soft_std': 0.009428190067410469, 'reward_judge_quality_mean': 0.5612500309944153, 'reward_judge_quality_std': 0.1968638300895691, 'reward_total_composite_mean': 0.6054582595825195, 'reward_total_composite_std': 0.19344164431095123, 'epoch': 0.0}

  1%|          | 23/3300 [02:24<5:25:01,  5.95s/it]INFO 04-12 22:16:52 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 24/3300 [02:30<5:26:20,  5.98s/it]2026-04-12 22:16:58,427 | INFO | train_grpo_train | metrics_logged mode=train step=24

                                                   
{'loss': 0.0564, 'grad_norm': 18.268474578857422, 'learning_rate': 9.930303030303031e-06, 'num_tokens': 44791.0, 'completions/mean_length': 62.375, 'completions/min_length': 42.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.634819746017456, 'rewards/meter/std': 0.4580250084400177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9998894929885864, 'rewards/repeat_soft/std': 0.00031247673905454576, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7179077863693237, 'rewards/total_composite/std': 0.2593117356300354, 'reward': 0.7179077863693237, 'reward_std': 0.2593117356300354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21290218830108643, 'sampling/sampling_logp_difference/max': 1.9016647338867188, 'sampling/importance_sampling_ratio/min': 0.1493198424577713, 'sampling/importance_sampling_ratio/mean': 1.0068492889404297, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.353201448917389, 'clip_ratio/low_mean': 0.06295787636190653, 'clip_ratio/low_min': 0.06295787636190653, 'clip_ratio/high_mean': 0.107818647287786, 'clip_ratio/high_max': 0.107818647287786, 'clip_ratio/region_mean': 0.17077652364969254, 'reward_total_mean': 0.7179077863693237, 'reward_meter_mean': 0.634819746017456, 'reward_meter_std': 0.4580250084400177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9998894929885864, 'reward_repeat_soft_std': 0.00031247673905454576, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7179077863693237, 'reward_total_composite_std': 0.2593117356300354, 'epoch': 0.0}

  1%|          | 24/3300 [02:30<5:26:20,  5.98s/it]INFO 04-12 22:16:58 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 25/3300 [02:38<5:51:14,  6.44s/it]2026-04-12 22:17:05,930 | INFO | train_grpo_train | metrics_logged mode=train step=25

                                                   
{'loss': 0.0743, 'grad_norm': 8.258381843566895, 'learning_rate': 9.927272727272728e-06, 'num_tokens': 47647.0, 'completions/mean_length': 148.0, 'completions/min_length': 129.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.0, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.5918176174163818, 'rewards/meter/std': 0.33095410466194153, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976328611373901, 'rewards/repeat_soft/std': 0.0029239931609481573, 'rewards/judge_quality/mean': 0.7024999856948853, 'rewards/judge_quality/std': 0.24294327199459076, 'rewards/total_composite/mean': 0.7193312048912048, 'rewards/total_composite/std': 0.1843627542257309, 'reward': 0.7193312048912048, 'reward_std': 0.1843627244234085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16058434545993805, 'sampling/sampling_logp_difference/max': 2.274415969848633, 'sampling/importance_sampling_ratio/min': 0.10285695642232895, 'sampling/importance_sampling_ratio/mean': 1.0136536359786987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0220863968133926, 'clip_ratio/low_mean': 0.06519535928964615, 'clip_ratio/low_min': 0.06519535928964615, 'clip_ratio/high_mean': 0.06505231745541096, 'clip_ratio/high_max': 0.06505231745541096, 'clip_ratio/region_mean': 0.1302476767450571, 'reward_total_mean': 0.7193312048912048, 'reward_meter_mean': 0.5918176174163818, 'reward_meter_std': 0.33095410466194153, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976328611373901, 'reward_repeat_soft_std': 0.0029239931609481573, 'reward_judge_quality_mean': 0.7024999856948853, 'reward_judge_quality_std': 0.24294327199459076, 'reward_total_composite_mean': 0.7193312048912048, 'reward_total_composite_std': 0.1843627542257309, 'epoch': 0.0}

  1%|          | 25/3300 [02:38<5:51:14,  6.44s/it]INFO 04-12 22:17:06 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 26/3300 [02:44<5:46:32,  6.35s/it]2026-04-12 22:17:12,084 | INFO | train_grpo_train | metrics_logged mode=train step=26

                                                   
{'loss': -0.1294, 'grad_norm': 24.321762084960938, 'learning_rate': 9.924242424242425e-06, 'num_tokens': 49143.0, 'completions/mean_length': 30.0, 'completions/min_length': 18.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9500245451927185, 'rewards/meter/std': 0.13721725344657898, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.7997610569000244, 'rewards/total_composite/std': 0.06174775958061218, 'reward': 0.7997610569000244, 'reward_std': 0.06174774840474129, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0748242437839508, 'sampling/sampling_logp_difference/max': 1.1188440322875977, 'sampling/importance_sampling_ratio/min': 0.32665717601776123, 'sampling/importance_sampling_ratio/mean': 0.9947167038917542, 'sampling/importance_sampling_ratio/max': 1.8977136611938477, 'entropy': 0.4834756925702095, 'clip_ratio/low_mean': 0.02777777798473835, 'clip_ratio/low_min': 0.02777777798473835, 'clip_ratio/high_mean': 0.030977724120020866, 'clip_ratio/high_max': 0.030977724120020866, 'clip_ratio/region_mean': 0.058755502104759216, 'reward_total_mean': 0.7997610569000244, 'reward_meter_mean': 0.9500245451927185, 'reward_meter_std': 0.13721725344657898, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.7997610569000244, 'reward_total_composite_std': 0.06174775958061218, 'epoch': 0.0}

  1%|          | 26/3300 [02:44<5:46:32,  6.35s/it]INFO 04-12 22:17:12 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 27/3300 [02:51<6:03:23,  6.66s/it]2026-04-12 22:17:19,472 | INFO | train_grpo_train | metrics_logged mode=train step=27

                                                   
{'loss': -0.0948, 'grad_norm': 7.565678596496582, 'learning_rate': 9.921212121212121e-06, 'num_tokens': 51958.0, 'completions/mean_length': 154.875, 'completions/min_length': 95.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.875, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.611936628818512, 'rewards/meter/std': 0.3855309784412384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9979890584945679, 'rewards/repeat_soft/std': 0.0016030868282541633, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5971717834472656, 'rewards/total_composite/std': 0.28371596336364746, 'reward': 0.5971717834472656, 'reward_std': 0.28371599316596985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21463504433631897, 'sampling/sampling_logp_difference/max': 2.0930118560791016, 'sampling/importance_sampling_ratio/min': 0.12331517040729523, 'sampling/importance_sampling_ratio/mean': 1.0371588468551636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4359234273433685, 'clip_ratio/low_mean': 0.07277752831578255, 'clip_ratio/low_min': 0.07277752831578255, 'clip_ratio/high_mean': 0.1326075941324234, 'clip_ratio/high_max': 0.1326075941324234, 'clip_ratio/region_mean': 0.20538512244820595, 'reward_total_mean': 0.5971717834472656, 'reward_meter_mean': 0.611936628818512, 'reward_meter_std': 0.3855309784412384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9979890584945679, 'reward_repeat_soft_std': 0.0016030868282541633, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5971717834472656, 'reward_total_composite_std': 0.28371596336364746, 'epoch': 0.0}

  1%|          | 27/3300 [02:51<6:03:23,  6.66s/it]INFO 04-12 22:17:19 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 28/3300 [02:57<5:43:26,  6.30s/it]2026-04-12 22:17:24,920 | INFO | train_grpo_train | metrics_logged mode=train step=28

                                                   
{'loss': 0.0946, 'grad_norm': 23.491594314575195, 'learning_rate': 9.918181818181818e-06, 'num_tokens': 53496.0, 'completions/mean_length': 30.25, 'completions/min_length': 20.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.5963934063911438, 'rewards/meter/std': 0.4609495997428894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9303162097930908, 'rewards/repeat_soft/std': 0.06846698373556137, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.644158661365509, 'rewards/total_composite/std': 0.20341292023658752, 'reward': 0.644158661365509, 'reward_std': 0.20341289043426514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22669285535812378, 'sampling/sampling_logp_difference/max': 1.899033546447754, 'sampling/importance_sampling_ratio/min': 0.14971324801445007, 'sampling/importance_sampling_ratio/mean': 1.014909267425537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6166850328445435, 'clip_ratio/low_mean': 0.0859464481472969, 'clip_ratio/low_min': 0.0859464481472969, 'clip_ratio/high_mean': 0.08879269333556294, 'clip_ratio/high_max': 0.08879269333556294, 'clip_ratio/region_mean': 0.17473914148285985, 'reward_total_mean': 0.644158661365509, 'reward_meter_mean': 0.5963934063911438, 'reward_meter_std': 0.4609495997428894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9303162097930908, 'reward_repeat_soft_std': 0.06846698373556137, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.644158661365509, 'reward_total_composite_std': 0.20341292023658752, 'epoch': 0.0}

  1%|          | 28/3300 [02:57<5:43:26,  6.30s/it]INFO 04-12 22:17:25 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 29/3300 [03:04<6:02:12,  6.64s/it]2026-04-12 22:17:32,372 | INFO | train_grpo_train | metrics_logged mode=train step=29

                                                   
{'loss': 0.0661, 'grad_norm': 7.243879795074463, 'learning_rate': 9.915151515151515e-06, 'num_tokens': 56603.0, 'completions/mean_length': 188.375, 'completions/min_length': 179.0, 'completions/max_length': 204.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 188.375, 'completions/min_terminated_length': 179.0, 'completions/max_terminated_length': 204.0, 'rewards/meter/mean': 0.6960962414741516, 'rewards/meter/std': 0.3066192865371704, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9989635348320007, 'rewards/repeat_soft/std': 0.000937586824875325, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.6790146827697754, 'rewards/total_composite/std': 0.14062994718551636, 'reward': 0.6790146827697754, 'reward_std': 0.14062994718551636, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19934551417827606, 'sampling/sampling_logp_difference/max': 1.6785790920257568, 'sampling/importance_sampling_ratio/min': 0.1866389811038971, 'sampling/importance_sampling_ratio/mean': 1.0356954336166382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.226356267929077, 'clip_ratio/low_mean': 0.06664236634969711, 'clip_ratio/low_min': 0.06664236634969711, 'clip_ratio/high_mean': 0.12853810749948025, 'clip_ratio/high_max': 0.12853810749948025, 'clip_ratio/region_mean': 0.19518047384917736, 'reward_total_mean': 0.6790146827697754, 'reward_meter_mean': 0.6960962414741516, 'reward_meter_std': 0.3066192865371704, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9989635348320007, 'reward_repeat_soft_std': 0.000937586824875325, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.6790146827697754, 'reward_total_composite_std': 0.14062994718551636, 'epoch': 0.0}

  1%|          | 29/3300 [03:04<6:02:12,  6.64s/it]INFO 04-12 22:17:32 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 30/3300 [03:11<6:01:34,  6.63s/it]2026-04-12 22:17:38,984 | INFO | train_grpo_train | metrics_logged mode=train step=30

                                                   
{'loss': 0.0748, 'grad_norm': 11.622523307800293, 'learning_rate': 9.912121212121213e-06, 'num_tokens': 58613.0, 'completions/mean_length': 67.25, 'completions/min_length': 48.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.6762853264808655, 'rewards/meter/std': 0.3140008747577667, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974287152290344, 'rewards/repeat_soft/std': 0.003936082124710083, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6988212466239929, 'rewards/total_composite/std': 0.10489797592163086, 'reward': 0.6988212466239929, 'reward_std': 0.10489796847105026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20940160751342773, 'sampling/sampling_logp_difference/max': 2.6506054401397705, 'sampling/importance_sampling_ratio/min': 0.07060845196247101, 'sampling/importance_sampling_ratio/mean': 1.0177046060562134, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0744355469942093, 'clip_ratio/low_mean': 0.08076691627502441, 'clip_ratio/low_min': 0.08076691627502441, 'clip_ratio/high_mean': 0.1163441464304924, 'clip_ratio/high_max': 0.1163441464304924, 'clip_ratio/region_mean': 0.19711106270551682, 'reward_total_mean': 0.6988212466239929, 'reward_meter_mean': 0.6762853264808655, 'reward_meter_std': 0.3140008747577667, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974287152290344, 'reward_repeat_soft_std': 0.003936082124710083, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6988212466239929, 'reward_total_composite_std': 0.10489797592163086, 'epoch': 0.0}

  1%|          | 30/3300 [03:11<6:01:34,  6.63s/it]INFO 04-12 22:17:39 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 31/3300 [03:17<5:55:38,  6.53s/it]2026-04-12 22:17:45,262 | INFO | train_grpo_train | metrics_logged mode=train step=31

                                                   
{'loss': 0.0371, 'grad_norm': 21.77640151977539, 'learning_rate': 9.90909090909091e-06, 'num_tokens': 60623.0, 'completions/mean_length': 79.25, 'completions/min_length': 61.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.3762175440788269, 'rewards/meter/std': 0.3404117226600647, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.995638370513916, 'rewards/repeat_soft/std': 0.0033319119829684496, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.4848466217517853, 'rewards/total_composite/std': 0.30984121561050415, 'reward': 0.4848466217517853, 'reward_std': 0.30984121561050415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2795345187187195, 'sampling/sampling_logp_difference/max': 3.1914567947387695, 'sampling/importance_sampling_ratio/min': 0.041111934930086136, 'sampling/importance_sampling_ratio/mean': 1.0121794939041138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1605252847075462, 'clip_ratio/low_mean': 0.05587913282215595, 'clip_ratio/low_min': 0.05587913282215595, 'clip_ratio/high_mean': 0.14557625632733107, 'clip_ratio/high_max': 0.14557625632733107, 'clip_ratio/region_mean': 0.20145538914948702, 'reward_total_mean': 0.4848466217517853, 'reward_meter_mean': 0.3762175440788269, 'reward_meter_std': 0.3404117226600647, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.995638370513916, 'reward_repeat_soft_std': 0.0033319119829684496, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.4848466217517853, 'reward_total_composite_std': 0.30984121561050415, 'epoch': 0.0}

  1%|          | 31/3300 [03:17<5:55:38,  6.53s/it]INFO 04-12 22:17:45 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 32/3300 [03:23<5:43:44,  6.31s/it]2026-04-12 22:17:51,068 | INFO | train_grpo_train | metrics_logged mode=train step=32

                                                   
{'loss': 0.0928, 'grad_norm': 28.076183319091797, 'learning_rate': 9.906060606060607e-06, 'num_tokens': 62100.0, 'completions/mean_length': 29.625, 'completions/min_length': 22.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.625, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.48696160316467285, 'rewards/meter/std': 0.4289628863334656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9554293155670166, 'rewards/repeat_soft/std': 0.013228065334260464, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.740675687789917, 'rewards/total_composite/std': 0.19279979169368744, 'reward': 0.740675687789917, 'reward_std': 0.19279977679252625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24338388442993164, 'sampling/sampling_logp_difference/max': 2.743772506713867, 'sampling/importance_sampling_ratio/min': 0.06432721018791199, 'sampling/importance_sampling_ratio/mean': 1.0007094144821167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2935088127851486, 'clip_ratio/low_mean': 0.11610591504722834, 'clip_ratio/low_min': 0.11610591504722834, 'clip_ratio/high_mean': 0.10445739328861237, 'clip_ratio/high_max': 0.10445739328861237, 'clip_ratio/region_mean': 0.2205633083358407, 'reward_total_mean': 0.740675687789917, 'reward_meter_mean': 0.48696160316467285, 'reward_meter_std': 0.4289628863334656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9554293155670166, 'reward_repeat_soft_std': 0.013228065334260464, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.740675687789917, 'reward_total_composite_std': 0.19279979169368744, 'epoch': 0.0}

  1%|          | 32/3300 [03:23<5:43:44,  6.31s/it]INFO 04-12 22:17:51 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 33/3300 [03:29<5:37:53,  6.21s/it]2026-04-12 22:17:57,027 | INFO | train_grpo_train | metrics_logged mode=train step=33

                                                   
{'loss': 0.0988, 'grad_norm': 13.971663475036621, 'learning_rate': 9.903030303030305e-06, 'num_tokens': 63936.0, 'completions/mean_length': 65.5, 'completions/min_length': 36.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.36027342081069946, 'rewards/meter/std': 0.43213415145874023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9996891021728516, 'rewards/repeat_soft/std': 0.0006069060764275491, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.5131999254226685, 'rewards/total_composite/std': 0.2780666947364807, 'reward': 0.5131999254226685, 'reward_std': 0.2780666649341583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19844399392604828, 'sampling/sampling_logp_difference/max': 2.652165412902832, 'sampling/importance_sampling_ratio/min': 0.07049839198589325, 'sampling/importance_sampling_ratio/mean': 1.0268425941467285, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8619239330291748, 'clip_ratio/low_mean': 0.07098640222102404, 'clip_ratio/low_min': 0.07098640222102404, 'clip_ratio/high_mean': 0.10280672460794449, 'clip_ratio/high_max': 0.10280672460794449, 'clip_ratio/region_mean': 0.17379312682896852, 'reward_total_mean': 0.5131999254226685, 'reward_meter_mean': 0.36027342081069946, 'reward_meter_std': 0.43213415145874023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9996891021728516, 'reward_repeat_soft_std': 0.0006069060764275491, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.5131999254226685, 'reward_total_composite_std': 0.2780666947364807, 'epoch': 0.0}

  1%|          | 33/3300 [03:29<5:37:53,  6.21s/it]INFO 04-12 22:17:57 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 34/3300 [03:35<5:35:32,  6.16s/it]2026-04-12 22:18:03,095 | INFO | train_grpo_train | metrics_logged mode=train step=34

                                                   
{'loss': 0.1421, 'grad_norm': 17.26323699951172, 'learning_rate': 9.9e-06, 'num_tokens': 65663.0, 'completions/mean_length': 57.875, 'completions/min_length': 38.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.6663380861282349, 'rewards/meter/std': 0.3654095232486725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9954054355621338, 'rewards/repeat_soft/std': 0.006265631411224604, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/total_composite/mean': 0.7185176610946655, 'rewards/total_composite/std': 0.2052834928035736, 'reward': 0.7185176610946655, 'reward_std': 0.2052834928035736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2225874960422516, 'sampling/sampling_logp_difference/max': 1.3435239791870117, 'sampling/importance_sampling_ratio/min': 0.2609245479106903, 'sampling/importance_sampling_ratio/mean': 1.0131149291992188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8193220719695091, 'clip_ratio/low_mean': 0.09162838384509087, 'clip_ratio/low_min': 0.09162838384509087, 'clip_ratio/high_mean': 0.12510230764746666, 'clip_ratio/high_max': 0.12510230764746666, 'clip_ratio/region_mean': 0.21673069149255753, 'reward_total_mean': 0.7185176610946655, 'reward_meter_mean': 0.6663380861282349, 'reward_meter_std': 0.3654095232486725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9954054355621338, 'reward_repeat_soft_std': 0.006265631411224604, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.7185176610946655, 'reward_total_composite_std': 0.2052834928035736, 'epoch': 0.0}

  1%|          | 34/3300 [03:35<5:35:32,  6.16s/it]INFO 04-12 22:18:03 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 35/3300 [03:41<5:32:57,  6.12s/it]2026-04-12 22:18:09,108 | INFO | train_grpo_train | metrics_logged mode=train step=35

                                                   
{'loss': 0.1592, 'grad_norm': 31.29047203063965, 'learning_rate': 9.896969696969699e-06, 'num_tokens': 67253.0, 'completions/mean_length': 33.75, 'completions/min_length': 28.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5175750255584717, 'rewards/meter/std': 0.333164244890213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9875191450119019, 'rewards/repeat_soft/std': 0.016329128295183182, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.211643248796463, 'rewards/total_composite/mean': 0.6191242337226868, 'rewards/total_composite/std': 0.29068228602409363, 'reward': 0.6191242337226868, 'reward_std': 0.29068228602409363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20600245893001556, 'sampling/sampling_logp_difference/max': 2.2265100479125977, 'sampling/importance_sampling_ratio/min': 0.1079043596982956, 'sampling/importance_sampling_ratio/mean': 1.0101490020751953, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9016148820519447, 'clip_ratio/low_mean': 0.06943389028310776, 'clip_ratio/low_min': 0.06943389028310776, 'clip_ratio/high_mean': 0.11182209756225348, 'clip_ratio/high_max': 0.11182209756225348, 'clip_ratio/region_mean': 0.18125598784536123, 'reward_total_mean': 0.6191242337226868, 'reward_meter_mean': 0.5175750255584717, 'reward_meter_std': 0.333164244890213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9875191450119019, 'reward_repeat_soft_std': 0.016329128295183182, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.211643248796463, 'reward_total_composite_mean': 0.6191242337226868, 'reward_total_composite_std': 0.29068228602409363, 'epoch': 0.0}

  1%|          | 35/3300 [03:41<5:32:57,  6.12s/it]INFO 04-12 22:18:09 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 36/3300 [03:49<6:02:00,  6.65s/it]2026-04-12 22:18:17,013 | INFO | train_grpo_train | metrics_logged mode=train step=36

                                                   
{'loss': -0.0159, 'grad_norm': 13.74673080444336, 'learning_rate': 9.893939393939395e-06, 'num_tokens': 68856.0, 'completions/mean_length': 51.375, 'completions/min_length': 38.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.3229061961174011, 'rewards/meter/std': 0.35688337683677673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9967907667160034, 'rewards/repeat_soft/std': 0.003882752498611808, 'rewards/judge_quality/mean': 0.6887500286102295, 'rewards/judge_quality/std': 0.21931307017803192, 'rewards/total_composite/mean': 0.601611852645874, 'rewards/total_composite/std': 0.19675835967063904, 'reward': 0.601611852645874, 'reward_std': 0.19675834476947784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1898266226053238, 'sampling/sampling_logp_difference/max': 1.7012615203857422, 'sampling/importance_sampling_ratio/min': 0.1824532151222229, 'sampling/importance_sampling_ratio/mean': 1.0139268636703491, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4139694049954414, 'clip_ratio/low_mean': 0.10160249006003141, 'clip_ratio/low_min': 0.10160249006003141, 'clip_ratio/high_mean': 0.03077651560306549, 'clip_ratio/high_max': 0.03077651560306549, 'clip_ratio/region_mean': 0.1323790056630969, 'reward_total_mean': 0.601611852645874, 'reward_meter_mean': 0.3229061961174011, 'reward_meter_std': 0.35688337683677673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9967907667160034, 'reward_repeat_soft_std': 0.003882752498611808, 'reward_judge_quality_mean': 0.6887500286102295, 'reward_judge_quality_std': 0.21931307017803192, 'reward_total_composite_mean': 0.601611852645874, 'reward_total_composite_std': 0.19675835967063904, 'epoch': 0.0}

  1%|          | 36/3300 [03:49<6:02:00,  6.65s/it]INFO 04-12 22:18:17 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 37/3300 [03:55<5:52:47,  6.49s/it]2026-04-12 22:18:23,110 | INFO | train_grpo_train | metrics_logged mode=train step=37

                                                   
{'loss': -0.0113, 'grad_norm': 13.515311241149902, 'learning_rate': 9.890909090909092e-06, 'num_tokens': 70697.0, 'completions/mean_length': 59.125, 'completions/min_length': 47.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.765489935874939, 'rewards/meter/std': 0.34022775292396545, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9889501929283142, 'rewards/repeat_soft/std': 0.010796717368066311, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.21685661375522614, 'rewards/total_composite/mean': 0.7452405691146851, 'rewards/total_composite/std': 0.13115821778774261, 'reward': 0.7452405691146851, 'reward_std': 0.13115820288658142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20520399510860443, 'sampling/sampling_logp_difference/max': 1.5714797973632812, 'sampling/importance_sampling_ratio/min': 0.20773755013942719, 'sampling/importance_sampling_ratio/mean': 1.0194767713546753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.044485777616501, 'clip_ratio/low_mean': 0.09386534243822098, 'clip_ratio/low_min': 0.09386534243822098, 'clip_ratio/high_mean': 0.11555166356265545, 'clip_ratio/high_max': 0.11555166356265545, 'clip_ratio/region_mean': 0.20941700600087643, 'reward_total_mean': 0.7452405691146851, 'reward_meter_mean': 0.765489935874939, 'reward_meter_std': 0.34022775292396545, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9889501929283142, 'reward_repeat_soft_std': 0.010796717368066311, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.21685661375522614, 'reward_total_composite_mean': 0.7452405691146851, 'reward_total_composite_std': 0.13115821778774261, 'epoch': 0.0}

  1%|          | 37/3300 [03:55<5:52:47,  6.49s/it]INFO 04-12 22:18:23 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 38/3300 [04:01<5:48:26,  6.41s/it]2026-04-12 22:18:29,339 | INFO | train_grpo_train | metrics_logged mode=train step=38

                                                   
{'loss': 0.0265, 'grad_norm': 11.196772575378418, 'learning_rate': 9.887878787878789e-06, 'num_tokens': 72445.0, 'completions/mean_length': 65.5, 'completions/min_length': 60.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.5, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8452023267745972, 'rewards/meter/std': 0.2854308485984802, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985214471817017, 'rewards/repeat_soft/std': 0.004181958269327879, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.14201989769935608, 'rewards/total_composite/mean': 0.7820681929588318, 'rewards/total_composite/std': 0.09378696978092194, 'reward': 0.7820681929588318, 'reward_std': 0.09378696978092194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16852843761444092, 'sampling/sampling_logp_difference/max': 1.3398032188415527, 'sampling/importance_sampling_ratio/min': 0.2618972063064575, 'sampling/importance_sampling_ratio/mean': 1.028748631477356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.285915844142437, 'clip_ratio/low_mean': 0.017307693138718605, 'clip_ratio/low_min': 0.017307693138718605, 'clip_ratio/high_mean': 0.14909545984119177, 'clip_ratio/high_max': 0.14909545984119177, 'clip_ratio/region_mean': 0.16640315297991037, 'reward_total_mean': 0.7820681929588318, 'reward_meter_mean': 0.8452023267745972, 'reward_meter_std': 0.2854308485984802, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985214471817017, 'reward_repeat_soft_std': 0.004181958269327879, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.14201989769935608, 'reward_total_composite_mean': 0.7820681929588318, 'reward_total_composite_std': 0.09378696978092194, 'epoch': 0.0}

  1%|          | 38/3300 [04:01<5:48:26,  6.41s/it]INFO 04-12 22:18:29 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 39/3300 [04:08<5:55:54,  6.55s/it]2026-04-12 22:18:36,211 | INFO | train_grpo_train | metrics_logged mode=train step=39

                                                   
{'loss': 0.0498, 'grad_norm': 21.736717224121094, 'learning_rate': 9.884848484848486e-06, 'num_tokens': 74160.0, 'completions/mean_length': 56.375, 'completions/min_length': 43.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8037891387939453, 'rewards/meter/std': 0.25627055764198303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9939441680908203, 'rewards/repeat_soft/std': 0.007073741871863604, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.8068495392799377, 'rewards/total_composite/std': 0.14063221216201782, 'reward': 0.8068495392799377, 'reward_std': 0.14063221216201782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18867532908916473, 'sampling/sampling_logp_difference/max': 2.8137412071228027, 'sampling/importance_sampling_ratio/min': 0.05998017266392708, 'sampling/importance_sampling_ratio/mean': 0.9923123121261597, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6779676862061024, 'clip_ratio/low_mean': 0.06264539621770382, 'clip_ratio/low_min': 0.06264539621770382, 'clip_ratio/high_mean': 0.10982005670666695, 'clip_ratio/high_max': 0.10982005670666695, 'clip_ratio/region_mean': 0.17246545292437077, 'reward_total_mean': 0.8068495392799377, 'reward_meter_mean': 0.8037891387939453, 'reward_meter_std': 0.25627055764198303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9939441680908203, 'reward_repeat_soft_std': 0.007073741871863604, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.8068495392799377, 'reward_total_composite_std': 0.14063221216201782, 'epoch': 0.0}

  1%|          | 39/3300 [04:08<5:55:54,  6.55s/it]INFO 04-12 22:18:36 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 40/3300 [04:15<5:52:34,  6.49s/it]2026-04-12 22:18:42,564 | INFO | train_grpo_train | metrics_logged mode=train step=40

                                                   
{'loss': 0.1016, 'grad_norm': 22.00394058227539, 'learning_rate': 9.881818181818182e-06, 'num_tokens': 75789.0, 'completions/mean_length': 40.625, 'completions/min_length': 27.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8645680546760559, 'rewards/meter/std': 0.33847635984420776, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.952668309211731, 'rewards/repeat_soft/std': 0.06325025111436844, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.9103224277496338, 'rewards/total_composite/std': 0.15087567269802094, 'reward': 0.9103224277496338, 'reward_std': 0.15087567269802094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1935538351535797, 'sampling/sampling_logp_difference/max': 2.373173952102661, 'sampling/importance_sampling_ratio/min': 0.09318449348211288, 'sampling/importance_sampling_ratio/mean': 1.0082100629806519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0727061294019222, 'clip_ratio/low_mean': 0.02604166604578495, 'clip_ratio/low_min': 0.02604166604578495, 'clip_ratio/high_mean': 0.16503243800252676, 'clip_ratio/high_max': 0.16503243800252676, 'clip_ratio/region_mean': 0.1910741040483117, 'reward_total_mean': 0.9103224277496338, 'reward_meter_mean': 0.8645680546760559, 'reward_meter_std': 0.33847635984420776, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.952668309211731, 'reward_repeat_soft_std': 0.06325025111436844, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.9103224277496338, 'reward_total_composite_std': 0.15087567269802094, 'epoch': 0.0}

  1%|          | 40/3300 [04:15<5:52:34,  6.49s/it]INFO 04-12 22:18:42 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 41/3300 [04:24<6:34:12,  7.26s/it]2026-04-12 22:18:51,630 | INFO | train_grpo_train | metrics_logged mode=train step=41

                                                   
{'loss': -0.1876, 'grad_norm': 12.826306343078613, 'learning_rate': 9.87878787878788e-06, 'num_tokens': 77469.0, 'completions/mean_length': 50.0, 'completions/min_length': 24.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.6660476922988892, 'rewards/meter/std': 0.40160611271858215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.997750997543335, 'rewards/repeat_soft/std': 0.005148967728018761, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906257808208466, 'rewards/total_composite/mean': 0.6337647438049316, 'rewards/total_composite/std': 0.2952870726585388, 'reward': 0.6337647438049316, 'reward_std': 0.29528704285621643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20901450514793396, 'sampling/sampling_logp_difference/max': 1.7976255416870117, 'sampling/importance_sampling_ratio/min': 0.16569185256958008, 'sampling/importance_sampling_ratio/mean': 1.0211420059204102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4476252645254135, 'clip_ratio/low_mean': 0.09334935806691647, 'clip_ratio/low_min': 0.09334935806691647, 'clip_ratio/high_mean': 0.13066565804183483, 'clip_ratio/high_max': 0.13066565804183483, 'clip_ratio/region_mean': 0.2240150161087513, 'reward_total_mean': 0.6337647438049316, 'reward_meter_mean': 0.6660476922988892, 'reward_meter_std': 0.40160611271858215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.997750997543335, 'reward_repeat_soft_std': 0.005148967728018761, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906257808208466, 'reward_total_composite_mean': 0.6337647438049316, 'reward_total_composite_std': 0.2952870726585388, 'epoch': 0.0}

  1%|          | 41/3300 [04:24<6:34:12,  7.26s/it]INFO 04-12 22:18:51 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 42/3300 [04:30<6:15:27,  6.91s/it]2026-04-12 22:18:57,727 | INFO | train_grpo_train | metrics_logged mode=train step=42

                                                   
{'loss': 0.0467, 'grad_norm': 24.63319969177246, 'learning_rate': 9.875757575757576e-06, 'num_tokens': 79094.0, 'completions/mean_length': 54.125, 'completions/min_length': 34.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8649886250495911, 'rewards/meter/std': 0.33145052194595337, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9943860769271851, 'rewards/repeat_soft/std': 0.008962815627455711, 'rewards/judge_quality/mean': 0.4437499940395355, 'rewards/judge_quality/std': 0.2084595113992691, 'rewards/total_composite/mean': 0.7718085050582886, 'rewards/total_composite/std': 0.18098533153533936, 'reward': 0.7718085050582886, 'reward_std': 0.18098531663417816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2050308734178543, 'sampling/sampling_logp_difference/max': 1.4114322662353516, 'sampling/importance_sampling_ratio/min': 0.24379386007785797, 'sampling/importance_sampling_ratio/mean': 1.0256385803222656, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6599382311105728, 'clip_ratio/low_mean': 0.0513223260641098, 'clip_ratio/low_min': 0.0513223260641098, 'clip_ratio/high_mean': 0.13139697071164846, 'clip_ratio/high_max': 0.13139697071164846, 'clip_ratio/region_mean': 0.18271929677575827, 'reward_total_mean': 0.7718085050582886, 'reward_meter_mean': 0.8649886250495911, 'reward_meter_std': 0.33145052194595337, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9943860769271851, 'reward_repeat_soft_std': 0.008962815627455711, 'reward_judge_quality_mean': 0.4437499940395355, 'reward_judge_quality_std': 0.2084595113992691, 'reward_total_composite_mean': 0.7718085050582886, 'reward_total_composite_std': 0.18098533153533936, 'epoch': 0.0}

  1%|▏         | 42/3300 [04:30<6:15:27,  6.91s/it]INFO 04-12 22:18:57 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 43/3300 [04:37<6:14:26,  6.90s/it]2026-04-12 22:19:04,585 | INFO | train_grpo_train | metrics_logged mode=train step=43

                                                   
{'loss': 0.0083, 'grad_norm': 13.690433502197266, 'learning_rate': 9.872727272727274e-06, 'num_tokens': 80983.0, 'completions/mean_length': 61.125, 'completions/min_length': 50.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7919253706932068, 'rewards/meter/std': 0.27126315236091614, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.984146773815155, 'rewards/repeat_soft/std': 0.01696406863629818, 'rewards/judge_quality/mean': 0.768750011920929, 'rewards/judge_quality/std': 0.21695540845394135, 'rewards/total_composite/mean': 0.8354060649871826, 'rewards/total_composite/std': 0.10415403544902802, 'reward': 0.8354060649871826, 'reward_std': 0.10415403544902802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15504302084445953, 'sampling/sampling_logp_difference/max': 1.9836196899414062, 'sampling/importance_sampling_ratio/min': 0.1375703662633896, 'sampling/importance_sampling_ratio/mean': 1.0255910158157349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.252104938030243, 'clip_ratio/low_mean': 0.06915658712387085, 'clip_ratio/low_min': 0.06915658712387085, 'clip_ratio/high_mean': 0.07177211018279195, 'clip_ratio/high_max': 0.07177211018279195, 'clip_ratio/region_mean': 0.1409286973066628, 'reward_total_mean': 0.8354060649871826, 'reward_meter_mean': 0.7919253706932068, 'reward_meter_std': 0.27126315236091614, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.984146773815155, 'reward_repeat_soft_std': 0.01696406863629818, 'reward_judge_quality_mean': 0.768750011920929, 'reward_judge_quality_std': 0.21695540845394135, 'reward_total_composite_mean': 0.8354060649871826, 'reward_total_composite_std': 0.10415403544902802, 'epoch': 0.0}

  1%|▏         | 43/3300 [04:37<6:14:26,  6.90s/it]INFO 04-12 22:19:04 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 44/3300 [04:44<6:16:35,  6.94s/it]2026-04-12 22:19:11,622 | INFO | train_grpo_train | metrics_logged mode=train step=44

                                                   
{'loss': 0.0866, 'grad_norm': 8.773218154907227, 'learning_rate': 9.869696969696971e-06, 'num_tokens': 83437.0, 'completions/mean_length': 126.75, 'completions/min_length': 103.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.75, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.3680556118488312, 'rewards/meter/std': 0.2608366906642914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973905086517334, 'rewards/repeat_soft/std': 0.0024202882777899504, 'rewards/judge_quality/mean': 0.5074999928474426, 'rewards/judge_quality/std': 0.1642080694437027, 'rewards/total_composite/mean': 0.5676140785217285, 'rewards/total_composite/std': 0.15416587889194489, 'reward': 0.5676140785217285, 'reward_std': 0.1541658639907837, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19027352333068848, 'sampling/sampling_logp_difference/max': 1.7497572898864746, 'sampling/importance_sampling_ratio/min': 0.17381611466407776, 'sampling/importance_sampling_ratio/mean': 1.0400882959365845, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.894649162888527, 'clip_ratio/low_mean': 0.11630215495824814, 'clip_ratio/low_min': 0.11630215495824814, 'clip_ratio/high_mean': 0.07096002250909805, 'clip_ratio/high_max': 0.07096002250909805, 'clip_ratio/region_mean': 0.1872621774673462, 'reward_total_mean': 0.5676140785217285, 'reward_meter_mean': 0.3680556118488312, 'reward_meter_std': 0.2608366906642914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973905086517334, 'reward_repeat_soft_std': 0.0024202882777899504, 'reward_judge_quality_mean': 0.5074999928474426, 'reward_judge_quality_std': 0.1642080694437027, 'reward_total_composite_mean': 0.5676140785217285, 'reward_total_composite_std': 0.15416587889194489, 'epoch': 0.0}

  1%|▏         | 44/3300 [04:44<6:16:35,  6.94s/it]INFO 04-12 22:19:11 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 45/3300 [04:50<6:00:24,  6.64s/it]2026-04-12 22:19:17,574 | INFO | train_grpo_train | metrics_logged mode=train step=45

                                                   
{'loss': -0.0796, 'grad_norm': 12.670366287231445, 'learning_rate': 9.866666666666668e-06, 'num_tokens': 85060.0, 'completions/mean_length': 51.875, 'completions/min_length': 35.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8333461284637451, 'rewards/meter/std': 0.27241066098213196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9904763698577881, 'rewards/repeat_soft/std': 0.016579801216721535, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.8438034057617188, 'rewards/total_composite/std': 0.17045538127422333, 'reward': 0.8438034057617188, 'reward_std': 0.17045536637306213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16479381918907166, 'sampling/sampling_logp_difference/max': 2.0195093154907227, 'sampling/importance_sampling_ratio/min': 0.13272058963775635, 'sampling/importance_sampling_ratio/mean': 1.0289181470870972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5834421515464783, 'clip_ratio/low_mean': 0.09027308598160744, 'clip_ratio/low_min': 0.09027308598160744, 'clip_ratio/high_mean': 0.07856571674346924, 'clip_ratio/high_max': 0.07856571674346924, 'clip_ratio/region_mean': 0.16883880272507668, 'reward_total_mean': 0.8438034057617188, 'reward_meter_mean': 0.8333461284637451, 'reward_meter_std': 0.27241066098213196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9904763698577881, 'reward_repeat_soft_std': 0.016579801216721535, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.8438034057617188, 'reward_total_composite_std': 0.17045538127422333, 'epoch': 0.0}

  1%|▏         | 45/3300 [04:50<6:00:24,  6.64s/it]INFO 04-12 22:19:17 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 46/3300 [04:56<5:55:45,  6.56s/it]2026-04-12 22:19:23,990 | INFO | train_grpo_train | metrics_logged mode=train step=46

                                                   
{'loss': -0.0367, 'grad_norm': 15.993819236755371, 'learning_rate': 9.863636363636364e-06, 'num_tokens': 86639.0, 'completions/mean_length': 41.375, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6189440488815308, 'rewards/meter/std': 0.4612226188182831, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973520040512085, 'rewards/repeat_soft/std': 0.00505313603207469, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404788017273, 'rewards/total_composite/mean': 0.6951350569725037, 'rewards/total_composite/std': 0.24396578967571259, 'reward': 0.6951350569725037, 'reward_std': 0.2439657747745514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22641699016094208, 'sampling/sampling_logp_difference/max': 1.6650819778442383, 'sampling/importance_sampling_ratio/min': 0.18917514383792877, 'sampling/importance_sampling_ratio/mean': 1.0454026460647583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.159736454486847, 'clip_ratio/low_mean': 0.0926593616604805, 'clip_ratio/low_min': 0.0926593616604805, 'clip_ratio/high_mean': 0.11545929498970509, 'clip_ratio/high_max': 0.11545929498970509, 'clip_ratio/region_mean': 0.20811865665018559, 'reward_total_mean': 0.6951350569725037, 'reward_meter_mean': 0.6189440488815308, 'reward_meter_std': 0.4612226188182831, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973520040512085, 'reward_repeat_soft_std': 0.00505313603207469, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404788017273, 'reward_total_composite_mean': 0.6951350569725037, 'reward_total_composite_std': 0.24396578967571259, 'epoch': 0.0}

  1%|▏         | 46/3300 [04:56<5:55:45,  6.56s/it]INFO 04-12 22:19:24 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 47/3300 [05:05<6:38:26,  7.35s/it]2026-04-12 22:19:33,243 | INFO | train_grpo_train | metrics_logged mode=train step=47

                                                   
{'loss': 0.1663, 'grad_norm': 14.117936134338379, 'learning_rate': 9.860606060606061e-06, 'num_tokens': 88311.0, 'completions/mean_length': 60.0, 'completions/min_length': 33.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.5848500728607178, 'rewards/meter/std': 0.310144305229187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9949307441711426, 'rewards/repeat_soft/std': 0.0036206496879458427, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6949256062507629, 'rewards/total_composite/std': 0.19530318677425385, 'reward': 0.6949256062507629, 'reward_std': 0.19530318677425385, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21896903216838837, 'sampling/sampling_logp_difference/max': 1.9479656219482422, 'sampling/importance_sampling_ratio/min': 0.1425638049840927, 'sampling/importance_sampling_ratio/mean': 0.9914653301239014, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5602813884615898, 'clip_ratio/low_mean': 0.0784417437389493, 'clip_ratio/low_min': 0.0784417437389493, 'clip_ratio/high_mean': 0.10587121546268463, 'clip_ratio/high_max': 0.10587121546268463, 'clip_ratio/region_mean': 0.18431295920163393, 'reward_total_mean': 0.6949256062507629, 'reward_meter_mean': 0.5848500728607178, 'reward_meter_std': 0.310144305229187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9949307441711426, 'reward_repeat_soft_std': 0.0036206496879458427, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6949256062507629, 'reward_total_composite_std': 0.19530318677425385, 'epoch': 0.0}

  1%|▏         | 47/3300 [05:05<6:38:26,  7.35s/it]INFO 04-12 22:19:33 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 48/3300 [05:12<6:29:25,  7.18s/it]2026-04-12 22:19:39,947 | INFO | train_grpo_train | metrics_logged mode=train step=48

                                                   
{'loss': 0.1596, 'grad_norm': 25.422317504882812, 'learning_rate': 9.857575757575758e-06, 'num_tokens': 89968.0, 'completions/mean_length': 38.125, 'completions/min_length': 26.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.48131677508354187, 'rewards/meter/std': 0.264057993888855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9779799580574036, 'rewards/repeat_soft/std': 0.028363212943077087, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.651140570640564, 'rewards/total_composite/std': 0.1632310152053833, 'reward': 0.651140570640564, 'reward_std': 0.1632310003042221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24116353690624237, 'sampling/sampling_logp_difference/max': 1.608734130859375, 'sampling/importance_sampling_ratio/min': 0.2001408040523529, 'sampling/importance_sampling_ratio/mean': 1.0072649717330933, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5285140573978424, 'clip_ratio/low_mean': 0.13228978216648102, 'clip_ratio/low_min': 0.13228978216648102, 'clip_ratio/high_mean': 0.0881993044167757, 'clip_ratio/high_max': 0.0881993044167757, 'clip_ratio/region_mean': 0.22048908658325672, 'reward_total_mean': 0.651140570640564, 'reward_meter_mean': 0.48131677508354187, 'reward_meter_std': 0.264057993888855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9779799580574036, 'reward_repeat_soft_std': 0.028363212943077087, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.651140570640564, 'reward_total_composite_std': 0.1632310152053833, 'epoch': 0.0}

  1%|▏         | 48/3300 [05:12<6:29:25,  7.18s/it]INFO 04-12 22:19:40 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 49/3300 [05:19<6:30:49,  7.21s/it]2026-04-12 22:19:47,211 | INFO | train_grpo_train | metrics_logged mode=train step=49

                                                   
{'loss': 0.0595, 'grad_norm': 16.515077590942383, 'learning_rate': 9.854545454545456e-06, 'num_tokens': 92540.0, 'completions/mean_length': 130.5, 'completions/min_length': 119.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.5, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.4894363284111023, 'rewards/meter/std': 0.22659291326999664, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9929240345954895, 'rewards/repeat_soft/std': 0.006053104996681213, 'rewards/judge_quality/mean': 0.4987499713897705, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.5284908413887024, 'rewards/total_composite/std': 0.22537189722061157, 'reward': 0.5284908413887024, 'reward_std': 0.22537189722061157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2464165985584259, 'sampling/sampling_logp_difference/max': 3.279010057449341, 'sampling/importance_sampling_ratio/min': 0.03766552358865738, 'sampling/importance_sampling_ratio/mean': 0.9850641489028931, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1143654063344002, 'clip_ratio/low_mean': 0.050971951335668564, 'clip_ratio/low_min': 0.050971951335668564, 'clip_ratio/high_mean': 0.16969024203717709, 'clip_ratio/high_max': 0.16969024203717709, 'clip_ratio/region_mean': 0.22066219337284565, 'reward_total_mean': 0.5284908413887024, 'reward_meter_mean': 0.4894363284111023, 'reward_meter_std': 0.22659291326999664, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9929240345954895, 'reward_repeat_soft_std': 0.006053104996681213, 'reward_judge_quality_mean': 0.4987499713897705, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.5284908413887024, 'reward_total_composite_std': 0.22537189722061157, 'epoch': 0.0}

  1%|▏         | 49/3300 [05:19<6:30:49,  7.21s/it]INFO 04-12 22:19:47 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 50/3300 [05:26<6:25:21,  7.11s/it]2026-04-12 22:19:54,095 | INFO | train_grpo_train | metrics_logged mode=train step=50

                                                   
{'loss': 0.1871, 'grad_norm': 12.148331642150879, 'learning_rate': 9.851515151515151e-06, 'num_tokens': 94611.0, 'completions/mean_length': 96.875, 'completions/min_length': 60.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.33955442905426025, 'rewards/meter/std': 0.31331250071525574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9961608648300171, 'rewards/repeat_soft/std': 0.005006108433008194, 'rewards/judge_quality/mean': 0.3712499737739563, 'rewards/judge_quality/std': 0.21390503644943237, 'rewards/total_composite/mean': 0.5137906074523926, 'rewards/total_composite/std': 0.15746544301509857, 'reward': 0.5137906074523926, 'reward_std': 0.15746545791625977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2518634796142578, 'sampling/sampling_logp_difference/max': 2.144465446472168, 'sampling/importance_sampling_ratio/min': 0.11713062971830368, 'sampling/importance_sampling_ratio/mean': 1.0296486616134644, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0926555544137955, 'clip_ratio/low_mean': 0.13645275123417377, 'clip_ratio/low_min': 0.13645275123417377, 'clip_ratio/high_mean': 0.07529207319021225, 'clip_ratio/high_max': 0.07529207319021225, 'clip_ratio/region_mean': 0.21174482442438602, 'reward_total_mean': 0.5137906074523926, 'reward_meter_mean': 0.33955442905426025, 'reward_meter_std': 0.31331250071525574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9961608648300171, 'reward_repeat_soft_std': 0.005006108433008194, 'reward_judge_quality_mean': 0.3712499737739563, 'reward_judge_quality_std': 0.21390503644943237, 'reward_total_composite_mean': 0.5137906074523926, 'reward_total_composite_std': 0.15746544301509857, 'epoch': 0.01}

  2%|▏         | 50/3300 [05:26<6:25:21,  7.11s/it]INFO 04-12 22:19:54 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:17,  2.18s/it]

 30%|███       | 3/10 [00:09<00:23,  3.41s/it]

 40%|████      | 4/10 [00:14<00:23,  3.94s/it]

 50%|█████     | 5/10 [00:19<00:22,  4.40s/it]

 60%|██████    | 6/10 [00:24<00:18,  4.53s/it]

 70%|███████   | 7/10 [00:29<00:14,  4.75s/it]

 80%|████████  | 8/10 [00:34<00:09,  4.73s/it]

 90%|█████████ | 9/10 [00:38<00:04,  4.47s/it]

100%|██████████| 10/10 [00:44<00:00,  5.10s/it]2026-04-12 22:20:44,485 | INFO | train_grpo_train | metrics_logged mode=eval step=50

                                                   

                                               
{'eval_loss': nan, 'eval_runtime': 50.3639, 'eval_samples_per_second': 1.588, 'eval_steps_per_second': 0.199, 'eval_num_tokens': 94611.0, 'eval_completions/mean_length': 92.5125, 'eval_completions/min_length': 36.6, 'eval_completions/max_length': 172.9, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 92.5125, 'eval_completions/min_terminated_length': 36.6, 'eval_completions/max_terminated_length': 172.9, 'eval_rewards/meter/mean': 0.622799813747406, 'eval_rewards/meter/std': 0.385142882168293, 'eval_rewards/count_adherence/mean': 0.9683333277702332, 'eval_rewards/count_adherence/std': 0.07128634452819824, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.1414213538169861, 'eval_rewards/repeat_soft/mean': 0.993061774969101, 'eval_rewards/repeat_soft/std': 0.011147955048363656, 'eval_rewards/judge_quality/mean': 0.4986249953508377, 'eval_rewards/judge_quality/std': 0.16713942140340804, 'eval_rewards/total_composite/mean': 0.6409295320510864, 'eval_rewards/total_composite/std': 0.23319732695817946, 'eval_reward': 0.6409295320510864, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.12240503132343292, 'eval_sampling/sampling_logp_difference/max': 1.1358956813812255, 'eval_sampling/importance_sampling_ratio/min': 0.32323764860630033, 'eval_sampling/importance_sampling_ratio/mean': 1.031867229938507, 'eval_sampling/importance_sampling_ratio/max': 1.530899715423584, 'eval_entropy': 1.7908539295196533, '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': 0.6409295320510864, 'eval_reward_meter_mean': 0.622799813747406, 'eval_reward_meter_std': 0.385142882168293, 'eval_reward_count_adherence_mean': 0.9683333277702332, 'eval_reward_count_adherence_std': 0.07128634452819824, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.1414213538169861, 'eval_reward_repeat_soft_mean': 0.993061774969101, 'eval_reward_repeat_soft_std': 0.011147955048363656, 'eval_reward_judge_quality_mean': 0.4986249953508377, 'eval_reward_judge_quality_std': 0.16713942140340804, 'eval_reward_total_composite_mean': 0.6409295320510864, 'eval_reward_total_composite_std': 0.23319732695817946, 'epoch': 0.01}

  2%|▏         | 50/3300 [06:16<6:25:21,  7.11s/it]

100%|██████████| 10/10 [00:44<00:00,  5.10s/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 [06:25<20:31:00, 22.73s/it]2026-04-12 22:20:53,273 | INFO | train_grpo_train | metrics_logged mode=train step=51

                                                    
{'loss': 0.102, 'grad_norm': 15.290033340454102, 'learning_rate': 9.84848484848485e-06, 'num_tokens': 96345.0, 'completions/mean_length': 49.75, 'completions/min_length': 35.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8911745548248291, 'rewards/meter/std': 0.18275241553783417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9963257312774658, 'rewards/repeat_soft/std': 0.005425546783953905, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.8152861595153809, 'rewards/total_composite/std': 0.12008143216371536, 'reward': 0.8152861595153809, 'reward_std': 0.12008143216371536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2254175841808319, 'sampling/sampling_logp_difference/max': 2.007718324661255, 'sampling/importance_sampling_ratio/min': 0.13429473340511322, 'sampling/importance_sampling_ratio/mean': 1.0337834358215332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1199056655168533, 'clip_ratio/low_mean': 0.07456336915493011, 'clip_ratio/low_min': 0.07456336915493011, 'clip_ratio/high_mean': 0.1384337618947029, 'clip_ratio/high_max': 0.1384337618947029, 'clip_ratio/region_mean': 0.21299713104963303, 'reward_total_mean': 0.8152861595153809, 'reward_meter_mean': 0.8911745548248291, 'reward_meter_std': 0.18275241553783417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9963257312774658, 'reward_repeat_soft_std': 0.005425546783953905, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.8152861595153809, 'reward_total_composite_std': 0.12008143216371536, 'epoch': 0.01}

  2%|▏         | 51/3300 [06:25<20:31:00, 22.73s/it]INFO 04-12 22:20:53 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 52/3300 [06:33<16:25:47, 18.21s/it]2026-04-12 22:21:00,929 | INFO | train_grpo_train | metrics_logged mode=train step=52

                                                    
{'loss': 0.1486, 'grad_norm': 10.75352668762207, 'learning_rate': 9.845454545454546e-06, 'num_tokens': 98656.0, 'completions/mean_length': 114.875, 'completions/min_length': 73.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.875, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.7699568867683411, 'rewards/meter/std': 0.3357446789741516, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9915257692337036, 'rewards/repeat_soft/std': 0.004899078514426947, 'rewards/judge_quality/mean': 0.6325000524520874, 'rewards/judge_quality/std': 0.18850921094417572, 'rewards/total_composite/mean': 0.7806956768035889, 'rewards/total_composite/std': 0.20309355854988098, 'reward': 0.7806956768035889, 'reward_std': 0.20309355854988098, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18153929710388184, 'sampling/sampling_logp_difference/max': 1.4434322118759155, 'sampling/importance_sampling_ratio/min': 0.2361159771680832, 'sampling/importance_sampling_ratio/mean': 1.042791724205017, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0450744926929474, 'clip_ratio/low_mean': 0.051473021507263184, 'clip_ratio/low_min': 0.051473021507263184, 'clip_ratio/high_mean': 0.1433755587786436, 'clip_ratio/high_max': 0.1433755587786436, 'clip_ratio/region_mean': 0.1948485802859068, 'reward_total_mean': 0.7806956768035889, 'reward_meter_mean': 0.7699568867683411, 'reward_meter_std': 0.3357446789741516, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9915257692337036, 'reward_repeat_soft_std': 0.004899078514426947, 'reward_judge_quality_mean': 0.6325000524520874, 'reward_judge_quality_std': 0.18850921094417572, 'reward_total_composite_mean': 0.7806956768035889, 'reward_total_composite_std': 0.20309355854988098, 'epoch': 0.01}

  2%|▏         | 52/3300 [06:33<16:25:47, 18.21s/it]INFO 04-12 22:21:01 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 53/3300 [06:39<13:07:06, 14.54s/it]2026-04-12 22:21:06,920 | INFO | train_grpo_train | metrics_logged mode=train step=53

                                                    
{'loss': -0.0155, 'grad_norm': 24.79887580871582, 'learning_rate': 9.842424242424243e-06, 'num_tokens': 100184.0, 'completions/mean_length': 38.0, 'completions/min_length': 28.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8915675282478333, 'rewards/meter/std': 0.14790304005146027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9969451427459717, 'rewards/repeat_soft/std': 0.002551683457568288, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7729640007019043, 'rewards/total_composite/std': 0.33475446701049805, 'reward': 0.7729640007019043, 'reward_std': 0.33475446701049805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21401749551296234, 'sampling/sampling_logp_difference/max': 2.617398500442505, 'sampling/importance_sampling_ratio/min': 0.07299250364303589, 'sampling/importance_sampling_ratio/mean': 0.976485013961792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6175712272524834, 'clip_ratio/low_mean': 0.037664955481886864, 'clip_ratio/low_min': 0.037664955481886864, 'clip_ratio/high_mean': 0.09748073294758797, 'clip_ratio/high_max': 0.09748073294758797, 'clip_ratio/region_mean': 0.13514568842947483, 'reward_total_mean': 0.7729640007019043, 'reward_meter_mean': 0.8915675282478333, 'reward_meter_std': 0.14790304005146027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9969451427459717, 'reward_repeat_soft_std': 0.002551683457568288, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7729640007019043, 'reward_total_composite_std': 0.33475446701049805, 'epoch': 0.01}

  2%|▏         | 53/3300 [06:39<13:07:06, 14.54s/it]INFO 04-12 22:21:07 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 54/3300 [06:47<11:26:45, 12.69s/it]2026-04-12 22:21:15,297 | INFO | train_grpo_train | metrics_logged mode=train step=54

                                                    
{'loss': 0.0579, 'grad_norm': 10.786398887634277, 'learning_rate': 9.83939393939394e-06, 'num_tokens': 102660.0, 'completions/mean_length': 132.5, 'completions/min_length': 94.0, 'completions/max_length': 197.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.5, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.5160332918167114, 'rewards/meter/std': 0.3664347231388092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9905648231506348, 'rewards/repeat_soft/std': 0.006998918484896421, 'rewards/judge_quality/mean': 0.8237500190734863, 'rewards/judge_quality/std': 0.2722361385822296, 'rewards/total_composite/mean': 0.5760136842727661, 'rewards/total_composite/std': 0.37488675117492676, 'reward': 0.5760136842727661, 'reward_std': 0.37488675117492676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21526861190795898, 'sampling/sampling_logp_difference/max': 2.1233134269714355, 'sampling/importance_sampling_ratio/min': 0.11963456869125366, 'sampling/importance_sampling_ratio/mean': 1.02375066280365, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.026296839118004, 'clip_ratio/low_mean': 0.06303927302360535, 'clip_ratio/low_min': 0.06303927302360535, 'clip_ratio/high_mean': 0.13252770714461803, 'clip_ratio/high_max': 0.13252770714461803, 'clip_ratio/region_mean': 0.19556698016822338, 'reward_total_mean': 0.5760136842727661, 'reward_meter_mean': 0.5160332918167114, 'reward_meter_std': 0.3664347231388092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9905648231506348, 'reward_repeat_soft_std': 0.006998918484896421, 'reward_judge_quality_mean': 0.8237500190734863, 'reward_judge_quality_std': 0.2722361385822296, 'reward_total_composite_mean': 0.5760136842727661, 'reward_total_composite_std': 0.37488675117492676, 'epoch': 0.01}

  2%|▏         | 54/3300 [06:47<11:26:45, 12.69s/it]INFO 04-12 22:21:15 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 55/3300 [06:58<10:52:35, 12.07s/it]2026-04-12 22:21:25,932 | INFO | train_grpo_train | metrics_logged mode=train step=55

                                                    
{'loss': 0.09, 'grad_norm': 20.380891799926758, 'learning_rate': 9.836363636363637e-06, 'num_tokens': 104225.0, 'completions/mean_length': 28.625, 'completions/min_length': 24.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6425614356994629, 'rewards/meter/std': 0.46263325214385986, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9575977325439453, 'rewards/repeat_soft/std': 0.013865554705262184, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.7381623983383179, 'rewards/total_composite/std': 0.23571598529815674, 'reward': 0.7381623983383179, 'reward_std': 0.23571597039699554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22299328446388245, 'sampling/sampling_logp_difference/max': 1.5237703323364258, 'sampling/importance_sampling_ratio/min': 0.21788883209228516, 'sampling/importance_sampling_ratio/mean': 1.0329948663711548, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6725586950778961, 'clip_ratio/low_mean': 0.0873684212565422, 'clip_ratio/low_min': 0.0873684212565422, 'clip_ratio/high_mean': 0.12113425973802805, 'clip_ratio/high_max': 0.12113425973802805, 'clip_ratio/region_mean': 0.20850268099457026, 'reward_total_mean': 0.7381623983383179, 'reward_meter_mean': 0.6425614356994629, 'reward_meter_std': 0.46263325214385986, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9575977325439453, 'reward_repeat_soft_std': 0.013865554705262184, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.7381623983383179, 'reward_total_composite_std': 0.23571598529815674, 'epoch': 0.01}

  2%|▏         | 55/3300 [06:58<10:52:35, 12.07s/it]INFO 04-12 22:21:26 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 56/3300 [07:10<10:49:46, 12.02s/it]2026-04-12 22:21:37,886 | INFO | train_grpo_train | metrics_logged mode=train step=56

                                                    
{'loss': 0.044, 'grad_norm': 17.16199493408203, 'learning_rate': 9.833333333333333e-06, 'num_tokens': 105784.0, 'completions/mean_length': 36.875, 'completions/min_length': 27.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6049321889877319, 'rewards/meter/std': 0.4337708652019501, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.6489694714546204, 'rewards/total_composite/std': 0.19477401673793793, 'reward': 0.6489694714546204, 'reward_std': 0.19477401673793793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18272539973258972, 'sampling/sampling_logp_difference/max': 1.233874797821045, 'sampling/importance_sampling_ratio/min': 0.2911621928215027, 'sampling/importance_sampling_ratio/mean': 1.0084818601608276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6532663255929947, 'clip_ratio/low_mean': 0.10378074645996094, 'clip_ratio/low_min': 0.10378074645996094, 'clip_ratio/high_mean': 0.11709216423332691, 'clip_ratio/high_max': 0.11709216423332691, 'clip_ratio/region_mean': 0.22087291069328785, 'reward_total_mean': 0.6489694714546204, 'reward_meter_mean': 0.6049321889877319, 'reward_meter_std': 0.4337708652019501, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.6489694714546204, 'reward_total_composite_std': 0.19477401673793793, 'epoch': 0.01}

  2%|▏         | 56/3300 [07:10<10:49:46, 12.02s/it]INFO 04-12 22:21:38 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 57/3300 [07:20<10:15:58, 11.40s/it]2026-04-12 22:21:47,832 | INFO | train_grpo_train | metrics_logged mode=train step=57

                                                    
{'loss': 0.1098, 'grad_norm': 9.582910537719727, 'learning_rate': 9.830303030303032e-06, 'num_tokens': 108140.0, 'completions/mean_length': 124.5, 'completions/min_length': 98.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.5, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.6479657888412476, 'rewards/meter/std': 0.37889355421066284, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9963884353637695, 'rewards/repeat_soft/std': 0.003934774082154036, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.6608484387397766, 'rewards/total_composite/std': 0.17602038383483887, 'reward': 0.6608484387397766, 'reward_std': 0.17602038383483887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19246143102645874, 'sampling/sampling_logp_difference/max': 2.835052490234375, 'sampling/importance_sampling_ratio/min': 0.05871544033288956, 'sampling/importance_sampling_ratio/mean': 1.0382243394851685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9649572968482971, 'clip_ratio/low_mean': 0.08068174216896296, 'clip_ratio/low_min': 0.08068174216896296, 'clip_ratio/high_mean': 0.1007919292896986, 'clip_ratio/high_max': 0.1007919292896986, 'clip_ratio/region_mean': 0.18147367145866156, 'reward_total_mean': 0.6608484387397766, 'reward_meter_mean': 0.6479657888412476, 'reward_meter_std': 0.37889355421066284, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9963884353637695, 'reward_repeat_soft_std': 0.003934774082154036, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.6608484387397766, 'reward_total_composite_std': 0.17602038383483887, 'epoch': 0.01}

  2%|▏         | 57/3300 [07:20<10:15:58, 11.40s/it]INFO 04-12 22:21:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 58/3300 [07:31<10:09:23, 11.28s/it]2026-04-12 22:21:58,752 | INFO | train_grpo_train | metrics_logged mode=train step=58

                                                    
{'loss': 0.0677, 'grad_norm': 10.142060279846191, 'learning_rate': 9.827272727272729e-06, 'num_tokens': 110448.0, 'completions/mean_length': 105.5, 'completions/min_length': 94.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.5, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.6721632480621338, 'rewards/meter/std': 0.34179237484931946, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9995437860488892, 'rewards/repeat_soft/std': 0.0007127728313207626, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.6385196447372437, 'rewards/total_composite/std': 0.3149482011795044, 'reward': 0.6385196447372437, 'reward_std': 0.3149482011795044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21328410506248474, 'sampling/sampling_logp_difference/max': 1.8783636093139648, 'sampling/importance_sampling_ratio/min': 0.1528400182723999, 'sampling/importance_sampling_ratio/mean': 1.0194169282913208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9980643540620804, 'clip_ratio/low_mean': 0.09210692159831524, 'clip_ratio/low_min': 0.09210692159831524, 'clip_ratio/high_mean': 0.09094057232141495, 'clip_ratio/high_max': 0.09094057232141495, 'clip_ratio/region_mean': 0.1830474939197302, 'reward_total_mean': 0.6385196447372437, 'reward_meter_mean': 0.6721632480621338, 'reward_meter_std': 0.34179237484931946, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9995437860488892, 'reward_repeat_soft_std': 0.0007127728313207626, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.6385196447372437, 'reward_total_composite_std': 0.3149482011795044, 'epoch': 0.01}

  2%|▏         | 58/3300 [07:31<10:09:23, 11.28s/it]INFO 04-12 22:21:58 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 59/3300 [07:38<9:07:24, 10.13s/it] 2026-04-12 22:22:06,281 | INFO | train_grpo_train | metrics_logged mode=train step=59

                                                   
{'loss': -0.021, 'grad_norm': 14.809919357299805, 'learning_rate': 9.824242424242425e-06, 'num_tokens': 112380.0, 'completions/mean_length': 62.5, 'completions/min_length': 54.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.5, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.6479412317276001, 'rewards/meter/std': 0.38575154542922974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9950402975082397, 'rewards/repeat_soft/std': 0.008863254450261593, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.2670440077781677, 'rewards/total_composite/mean': 0.6347871422767639, 'rewards/total_composite/std': 0.33347341418266296, 'reward': 0.6347871422767639, 'reward_std': 0.3334733843803406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2120247334241867, 'sampling/sampling_logp_difference/max': 1.6805639266967773, 'sampling/importance_sampling_ratio/min': 0.1862688958644867, 'sampling/importance_sampling_ratio/mean': 1.0009524822235107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7054464370012283, 'clip_ratio/low_mean': 0.05873700138181448, 'clip_ratio/low_min': 0.05873700138181448, 'clip_ratio/high_mean': 0.13117719814181328, 'clip_ratio/high_max': 0.13117719814181328, 'clip_ratio/region_mean': 0.18991419952362776, 'reward_total_mean': 0.6347871422767639, 'reward_meter_mean': 0.6479412317276001, 'reward_meter_std': 0.38575154542922974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9950402975082397, 'reward_repeat_soft_std': 0.008863254450261593, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.2670440077781677, 'reward_total_composite_mean': 0.6347871422767639, 'reward_total_composite_std': 0.33347341418266296, 'epoch': 0.01}

  2%|▏         | 59/3300 [07:38<9:07:24, 10.13s/it]INFO 04-12 22:22:06 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 60/3300 [07:46<8:37:00,  9.57s/it]2026-04-12 22:22:14,553 | INFO | train_grpo_train | metrics_logged mode=train step=60

                                                   
{'loss': 0.1206, 'grad_norm': 14.03058910369873, 'learning_rate': 9.821212121212122e-06, 'num_tokens': 114047.0, 'completions/mean_length': 63.375, 'completions/min_length': 45.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.682119607925415, 'rewards/meter/std': 0.40137389302253723, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9977132081985474, 'rewards/repeat_soft/std': 0.00379719166085124, 'rewards/judge_quality/mean': 0.7699999809265137, 'rewards/judge_quality/std': 0.22677870094776154, 'rewards/total_composite/mean': 0.7877250909805298, 'rewards/total_composite/std': 0.19343271851539612, 'reward': 0.7877250909805298, 'reward_std': 0.19343270361423492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20363926887512207, 'sampling/sampling_logp_difference/max': 1.5578479766845703, 'sampling/importance_sampling_ratio/min': 0.21058876812458038, 'sampling/importance_sampling_ratio/mean': 1.0375148057937622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7059862613677979, 'clip_ratio/low_mean': 0.0698032584041357, 'clip_ratio/low_min': 0.0698032584041357, 'clip_ratio/high_mean': 0.09936619736254215, 'clip_ratio/high_max': 0.09936619736254215, 'clip_ratio/region_mean': 0.16916945576667786, 'reward_total_mean': 0.7877250909805298, 'reward_meter_mean': 0.682119607925415, 'reward_meter_std': 0.40137389302253723, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9977132081985474, 'reward_repeat_soft_std': 0.00379719166085124, 'reward_judge_quality_mean': 0.7699999809265137, 'reward_judge_quality_std': 0.22677870094776154, 'reward_total_composite_mean': 0.7877250909805298, 'reward_total_composite_std': 0.19343271851539612, 'epoch': 0.01}

  2%|▏         | 60/3300 [07:47<8:37:00,  9.57s/it]INFO 04-12 22:22:14 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 61/3300 [07:54<8:01:08,  8.91s/it]2026-04-12 22:22:21,854 | INFO | train_grpo_train | metrics_logged mode=train step=61

                                                   
{'loss': -0.0091, 'grad_norm': 13.446069717407227, 'learning_rate': 9.81818181818182e-06, 'num_tokens': 115837.0, 'completions/mean_length': 49.75, 'completions/min_length': 32.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4147126078605652, 'rewards/meter/std': 0.41163748502731323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9938614368438721, 'rewards/repeat_soft/std': 0.010955816134810448, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.16974246501922607, 'rewards/total_composite/mean': 0.5863817930221558, 'rewards/total_composite/std': 0.2178468108177185, 'reward': 0.5863817930221558, 'reward_std': 0.2178467959165573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19411690533161163, 'sampling/sampling_logp_difference/max': 1.7225580215454102, 'sampling/importance_sampling_ratio/min': 0.17860867083072662, 'sampling/importance_sampling_ratio/mean': 1.0362242460250854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.983920469880104, 'clip_ratio/low_mean': 0.1204979931935668, 'clip_ratio/low_min': 0.1204979931935668, 'clip_ratio/high_mean': 0.0703886691480875, 'clip_ratio/high_max': 0.0703886691480875, 'clip_ratio/region_mean': 0.1908866623416543, 'reward_total_mean': 0.5863817930221558, 'reward_meter_mean': 0.4147126078605652, 'reward_meter_std': 0.41163748502731323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9938614368438721, 'reward_repeat_soft_std': 0.010955816134810448, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.16974246501922607, 'reward_total_composite_mean': 0.5863817930221558, 'reward_total_composite_std': 0.2178468108177185, 'epoch': 0.01}

  2%|▏         | 61/3300 [07:54<8:01:08,  8.91s/it]INFO 04-12 22:22:22 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 62/3300 [08:02<7:50:32,  8.72s/it]2026-04-12 22:22:30,124 | INFO | train_grpo_train | metrics_logged mode=train step=62

                                                   
{'loss': 0.0554, 'grad_norm': 8.62053108215332, 'learning_rate': 9.815151515151516e-06, 'num_tokens': 118853.0, 'completions/mean_length': 173.0, 'completions/min_length': 155.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 173.0, 'completions/min_terminated_length': 155.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.4597916603088379, 'rewards/meter/std': 0.29323315620422363, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.99494469165802, 'rewards/repeat_soft/std': 0.005377884954214096, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5711507201194763, 'rewards/total_composite/std': 0.12056976556777954, 'reward': 0.5711507201194763, 'reward_std': 0.12056976556777954, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20841319859027863, 'sampling/sampling_logp_difference/max': 2.5847525596618652, 'sampling/importance_sampling_ratio/min': 0.07541473954916, 'sampling/importance_sampling_ratio/mean': 1.0249940156936646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5373004227876663, 'clip_ratio/low_mean': 0.09918828681111336, 'clip_ratio/low_min': 0.09918828681111336, 'clip_ratio/high_mean': 0.11019240505993366, 'clip_ratio/high_max': 0.11019240505993366, 'clip_ratio/region_mean': 0.20938069187104702, 'reward_total_mean': 0.5711507201194763, 'reward_meter_mean': 0.4597916603088379, 'reward_meter_std': 0.29323315620422363, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.99494469165802, 'reward_repeat_soft_std': 0.005377884954214096, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5711507201194763, 'reward_total_composite_std': 0.12056976556777954, 'epoch': 0.01}

  2%|▏         | 62/3300 [08:02<7:50:32,  8.72s/it]INFO 04-12 22:22:30 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 63/3300 [08:09<7:27:02,  8.29s/it]2026-04-12 22:22:37,398 | INFO | train_grpo_train | metrics_logged mode=train step=63

                                                   
{'loss': 0.2431, 'grad_norm': 20.96488380432129, 'learning_rate': 9.812121212121212e-06, 'num_tokens': 120739.0, 'completions/mean_length': 74.75, 'completions/min_length': 57.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.45688575506210327, 'rewards/meter/std': 0.35617753863334656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985731840133667, 'rewards/repeat_soft/std': 0.0012093555415049195, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.6039559245109558, 'rewards/total_composite/std': 0.18145155906677246, 'reward': 0.6039559245109558, 'reward_std': 0.18145155906677246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21291305124759674, 'sampling/sampling_logp_difference/max': 2.4753305912017822, 'sampling/importance_sampling_ratio/min': 0.08413517475128174, 'sampling/importance_sampling_ratio/mean': 1.0076979398727417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8964705485850573, 'clip_ratio/low_mean': 0.128117217682302, 'clip_ratio/low_min': 0.128117217682302, 'clip_ratio/high_mean': 0.04571073269471526, 'clip_ratio/high_max': 0.04571073269471526, 'clip_ratio/region_mean': 0.17382795037701726, 'reward_total_mean': 0.6039559245109558, 'reward_meter_mean': 0.45688575506210327, 'reward_meter_std': 0.35617753863334656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985731840133667, 'reward_repeat_soft_std': 0.0012093555415049195, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.6039559245109558, 'reward_total_composite_std': 0.18145155906677246, 'epoch': 0.01}

  2%|▏         | 63/3300 [08:09<7:27:02,  8.29s/it]INFO 04-12 22:22:37 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 64/3300 [08:21<8:22:54,  9.32s/it]2026-04-12 22:22:49,144 | INFO | train_grpo_train | metrics_logged mode=train step=64

                                                   
{'loss': 0.0561, 'grad_norm': 11.986035346984863, 'learning_rate': 9.809090909090911e-06, 'num_tokens': 123013.0, 'completions/mean_length': 102.25, 'completions/min_length': 84.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.25, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.2554526627063751, 'rewards/meter/std': 0.23822079598903656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9972249269485474, 'rewards/repeat_soft/std': 0.0024571451358497143, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.43968093395233154, 'rewards/total_composite/std': 0.20654380321502686, 'reward': 0.43968093395233154, 'reward_std': 0.20654380321502686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23794835805892944, 'sampling/sampling_logp_difference/max': 1.656651258468628, 'sampling/importance_sampling_ratio/min': 0.1907767653465271, 'sampling/importance_sampling_ratio/mean': 1.025471806526184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1588610261678696, 'clip_ratio/low_mean': 0.10111558996140957, 'clip_ratio/low_min': 0.10111558996140957, 'clip_ratio/high_mean': 0.1057747658342123, 'clip_ratio/high_max': 0.1057747658342123, 'clip_ratio/region_mean': 0.20689035579562187, 'reward_total_mean': 0.43968093395233154, 'reward_meter_mean': 0.2554526627063751, 'reward_meter_std': 0.23822079598903656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9972249269485474, 'reward_repeat_soft_std': 0.0024571451358497143, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.43968093395233154, 'reward_total_composite_std': 0.20654380321502686, 'epoch': 0.01}

  2%|▏         | 64/3300 [08:21<8:22:54,  9.32s/it]INFO 04-12 22:22:49 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 65/3300 [08:29<7:54:10,  8.79s/it]2026-04-12 22:22:56,752 | INFO | train_grpo_train | metrics_logged mode=train step=65

                                                   
{'loss': 0.0258, 'grad_norm': 21.1607723236084, 'learning_rate': 9.806060606060607e-06, 'num_tokens': 124597.0, 'completions/mean_length': 48.0, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8362926840782166, 'rewards/meter/std': 0.15002521872520447, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975625872612, 'rewards/repeat_soft/std': 0.004202974960207939, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.8105879426002502, 'rewards/total_composite/std': 0.09124923497438431, 'reward': 0.8105879426002502, 'reward_std': 0.09124922752380371, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19722096621990204, 'sampling/sampling_logp_difference/max': 4.355945110321045, 'sampling/importance_sampling_ratio/min': 0.012830307707190514, 'sampling/importance_sampling_ratio/mean': 1.0022335052490234, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8046837113797665, 'clip_ratio/low_mean': 0.11510978080332279, 'clip_ratio/low_min': 0.11510978080332279, 'clip_ratio/high_mean': 0.06482380721718073, 'clip_ratio/high_max': 0.06482380721718073, 'clip_ratio/region_mean': 0.17993358802050352, 'reward_total_mean': 0.8105879426002502, 'reward_meter_mean': 0.8362926840782166, 'reward_meter_std': 0.15002521872520447, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975625872612, 'reward_repeat_soft_std': 0.004202974960207939, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.8105879426002502, 'reward_total_composite_std': 0.09124923497438431, 'epoch': 0.01}

  2%|▏         | 65/3300 [08:29<7:54:10,  8.79s/it]INFO 04-12 22:22:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 66/3300 [08:41<8:52:22,  9.88s/it]2026-04-12 22:23:09,187 | INFO | train_grpo_train | metrics_logged mode=train step=66

                                                   
{'loss': 0.7836, 'grad_norm': 12.42125129699707, 'learning_rate': 9.803030303030304e-06, 'num_tokens': 126598.0, 'completions/mean_length': 106.125, 'completions/min_length': 35.0, 'completions/max_length': 435.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 435.0, 'rewards/meter/mean': 0.7745488882064819, 'rewards/meter/std': 0.28584718704223633, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980877637863159, 'rewards/repeat_soft/std': 0.003999420441687107, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.7292307615280151, 'rewards/total_composite/std': 0.18889670073986053, 'reward': 0.7292307615280151, 'reward_std': 0.18889670073986053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19706401228904724, 'sampling/sampling_logp_difference/max': 1.9290742874145508, 'sampling/importance_sampling_ratio/min': 0.14528262615203857, 'sampling/importance_sampling_ratio/mean': 1.0292773246765137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0778553932905197, 'clip_ratio/low_mean': 0.0641922689974308, 'clip_ratio/low_min': 0.0641922689974308, 'clip_ratio/high_mean': 0.12423842307180166, 'clip_ratio/high_max': 0.12423842307180166, 'clip_ratio/region_mean': 0.18843069206923246, 'reward_total_mean': 0.7292307615280151, 'reward_meter_mean': 0.7745488882064819, 'reward_meter_std': 0.28584718704223633, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980877637863159, 'reward_repeat_soft_std': 0.003999420441687107, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.7292307615280151, 'reward_total_composite_std': 0.18889670073986053, 'epoch': 0.01}

  2%|▏         | 66/3300 [08:41<8:52:22,  9.88s/it]INFO 04-12 22:23:09 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 67/3300 [08:49<8:23:33,  9.35s/it]2026-04-12 22:23:17,210 | INFO | train_grpo_train | metrics_logged mode=train step=67

                                                   
{'loss': -0.0138, 'grad_norm': 12.676423072814941, 'learning_rate': 9.800000000000001e-06, 'num_tokens': 128362.0, 'completions/mean_length': 61.5, 'completions/min_length': 57.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9230592250823975, 'rewards/meter/std': 0.17033740878105164, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996062159538269, 'rewards/repeat_soft/std': 0.007730784825980663, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/total_composite/mean': 0.8164828419685364, 'rewards/total_composite/std': 0.10062293708324432, 'reward': 0.8164828419685364, 'reward_std': 0.10062292963266373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16694454848766327, 'sampling/sampling_logp_difference/max': 1.8855005502700806, 'sampling/importance_sampling_ratio/min': 0.15175308287143707, 'sampling/importance_sampling_ratio/mean': 1.0206559896469116, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.132050782442093, 'clip_ratio/low_mean': 0.023706896230578423, 'clip_ratio/low_min': 0.023706896230578423, 'clip_ratio/high_mean': 0.15569907892495394, 'clip_ratio/high_max': 0.15569907892495394, 'clip_ratio/region_mean': 0.17940597515553236, 'reward_total_mean': 0.8164828419685364, 'reward_meter_mean': 0.9230592250823975, 'reward_meter_std': 0.17033740878105164, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996062159538269, 'reward_repeat_soft_std': 0.007730784825980663, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.8164828419685364, 'reward_total_composite_std': 0.10062293708324432, 'epoch': 0.01}

  2%|▏         | 67/3300 [08:49<8:23:33,  9.35s/it]INFO 04-12 22:23:17 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 68/3300 [08:56<7:50:33,  8.74s/it]2026-04-12 22:23:24,524 | INFO | train_grpo_train | metrics_logged mode=train step=68

                                                   
{'loss': -0.0557, 'grad_norm': 17.533666610717773, 'learning_rate': 9.796969696969698e-06, 'num_tokens': 129994.0, 'completions/mean_length': 54.0, 'completions/min_length': 37.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.29206711053848267, 'rewards/meter/std': 0.2812570035457611, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9896067380905151, 'rewards/repeat_soft/std': 0.01452327985316515, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.4468875229358673, 'rewards/total_composite/std': 0.21935208141803741, 'reward': 0.4468875229358673, 'reward_std': 0.21935206651687622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24508477747440338, 'sampling/sampling_logp_difference/max': 2.154849052429199, 'sampling/importance_sampling_ratio/min': 0.11592069268226624, 'sampling/importance_sampling_ratio/mean': 1.0226720571517944, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9609599560499191, 'clip_ratio/low_mean': 0.08957079518586397, 'clip_ratio/low_min': 0.08957079518586397, 'clip_ratio/high_mean': 0.08130989596247673, 'clip_ratio/high_max': 0.08130989596247673, 'clip_ratio/region_mean': 0.1708806911483407, 'reward_total_mean': 0.4468875229358673, 'reward_meter_mean': 0.29206711053848267, 'reward_meter_std': 0.2812570035457611, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9896067380905151, 'reward_repeat_soft_std': 0.01452327985316515, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.4468875229358673, 'reward_total_composite_std': 0.21935208141803741, 'epoch': 0.01}

  2%|▏         | 68/3300 [08:57<7:50:33,  8.74s/it]INFO 04-12 22:23:24 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 69/3300 [09:09<8:50:10,  9.85s/it]2026-04-12 22:23:36,958 | INFO | train_grpo_train | metrics_logged mode=train step=69

                                                   
{'loss': 0.1552, 'grad_norm': 18.79545783996582, 'learning_rate': 9.793939393939394e-06, 'num_tokens': 132069.0, 'completions/mean_length': 72.375, 'completions/min_length': 50.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.6283866167068481, 'rewards/meter/std': 0.29875946044921875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9875814914703369, 'rewards/repeat_soft/std': 0.009496775455772877, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.6875321865081787, 'rewards/total_composite/std': 0.15890254080295563, 'reward': 0.6875321865081787, 'reward_std': 0.15890255570411682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26377734541893005, 'sampling/sampling_logp_difference/max': 2.8466076850891113, 'sampling/importance_sampling_ratio/min': 0.05804087966680527, 'sampling/importance_sampling_ratio/mean': 1.0076905488967896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4112783074378967, 'clip_ratio/low_mean': 0.07913436740636826, 'clip_ratio/low_min': 0.07913436740636826, 'clip_ratio/high_mean': 0.1484571509063244, 'clip_ratio/high_max': 0.1484571509063244, 'clip_ratio/region_mean': 0.22759151831269264, 'reward_total_mean': 0.6875321865081787, 'reward_meter_mean': 0.6283866167068481, 'reward_meter_std': 0.29875946044921875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9875814914703369, 'reward_repeat_soft_std': 0.009496775455772877, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.6875321865081787, 'reward_total_composite_std': 0.15890254080295563, 'epoch': 0.01}

  2%|▏         | 69/3300 [09:09<8:50:10,  9.85s/it]INFO 04-12 22:23:37 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 70/3300 [09:22<9:36:50, 10.72s/it]2026-04-12 22:23:49,755 | INFO | train_grpo_train | metrics_logged mode=train step=70

                                                   
{'loss': 0.0629, 'grad_norm': 11.336478233337402, 'learning_rate': 9.790909090909093e-06, 'num_tokens': 134316.0, 'completions/mean_length': 108.875, 'completions/min_length': 50.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.2835621237754822, 'rewards/meter/std': 0.3008175492286682, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9945306777954102, 'rewards/repeat_soft/std': 0.003017295151948929, 'rewards/judge_quality/mean': 0.6362500190734863, 'rewards/judge_quality/std': 0.24047201871871948, 'rewards/total_composite/mean': 0.563243567943573, 'rewards/total_composite/std': 0.17085498571395874, 'reward': 0.563243567943573, 'reward_std': 0.17085498571395874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21989120543003082, 'sampling/sampling_logp_difference/max': 1.4525012969970703, 'sampling/importance_sampling_ratio/min': 0.23398429155349731, 'sampling/importance_sampling_ratio/mean': 1.0469709634780884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0914076268672943, 'clip_ratio/low_mean': 0.18281028605997562, 'clip_ratio/low_min': 0.18281028605997562, 'clip_ratio/high_mean': 0.05735824815928936, 'clip_ratio/high_max': 0.05735824815928936, 'clip_ratio/region_mean': 0.24016853421926498, 'reward_total_mean': 0.563243567943573, 'reward_meter_mean': 0.2835621237754822, 'reward_meter_std': 0.3008175492286682, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9945306777954102, 'reward_repeat_soft_std': 0.003017295151948929, 'reward_judge_quality_mean': 0.6362500190734863, 'reward_judge_quality_std': 0.24047201871871948, 'reward_total_composite_mean': 0.563243567943573, 'reward_total_composite_std': 0.17085498571395874, 'epoch': 0.01}

  2%|▏         | 70/3300 [09:22<9:36:50, 10.72s/it]INFO 04-12 22:23:50 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 71/3300 [09:31<9:10:16, 10.22s/it]2026-04-12 22:23:58,787 | INFO | train_grpo_train | metrics_logged mode=train step=71

                                                   
{'loss': 0.2256, 'grad_norm': 17.786706924438477, 'learning_rate': 9.787878787878788e-06, 'num_tokens': 136031.0, 'completions/mean_length': 42.375, 'completions/min_length': 25.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6308169364929199, 'rewards/meter/std': 0.404389351606369, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9952167272567749, 'rewards/repeat_soft/std': 0.004450792912393808, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.7730143070220947, 'rewards/total_composite/std': 0.22425441443920135, 'reward': 0.7730143070220947, 'reward_std': 0.22425441443920135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15945900976657867, 'sampling/sampling_logp_difference/max': 2.320220947265625, 'sampling/importance_sampling_ratio/min': 0.09825187176465988, 'sampling/importance_sampling_ratio/mean': 1.0230482816696167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.073503103107214, 'clip_ratio/low_mean': 0.07358044851571321, 'clip_ratio/low_min': 0.07358044851571321, 'clip_ratio/high_mean': 0.05085150431841612, 'clip_ratio/high_max': 0.05085150431841612, 'clip_ratio/region_mean': 0.12443195283412933, 'reward_total_mean': 0.7730143070220947, 'reward_meter_mean': 0.6308169364929199, 'reward_meter_std': 0.404389351606369, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9952167272567749, 'reward_repeat_soft_std': 0.004450792912393808, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.7730143070220947, 'reward_total_composite_std': 0.22425441443920135, 'epoch': 0.01}

  2%|▏         | 71/3300 [09:31<9:10:16, 10.22s/it]INFO 04-12 22:23:59 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 72/3300 [09:40<8:49:19,  9.84s/it]2026-04-12 22:24:08,327 | INFO | train_grpo_train | metrics_logged mode=train step=72

                                                   
{'loss': 0.0725, 'grad_norm': 12.097882270812988, 'learning_rate': 9.784848484848486e-06, 'num_tokens': 138022.0, 'completions/mean_length': 86.875, 'completions/min_length': 77.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.875, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.24359747767448425, 'rewards/meter/std': 0.2615220844745636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.993793249130249, 'rewards/repeat_soft/std': 0.002918896032497287, 'rewards/judge_quality/mean': 0.6112499833106995, 'rewards/judge_quality/std': 0.15037456154823303, 'rewards/total_composite/mean': 0.5423731803894043, 'rewards/total_composite/std': 0.13441166281700134, 'reward': 0.5423731803894043, 'reward_std': 0.13441164791584015, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18156975507736206, 'sampling/sampling_logp_difference/max': 3.1153430938720703, 'sampling/importance_sampling_ratio/min': 0.04436328634619713, 'sampling/importance_sampling_ratio/mean': 0.9998382329940796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9134444296360016, 'clip_ratio/low_mean': 0.1097878199070692, 'clip_ratio/low_min': 0.1097878199070692, 'clip_ratio/high_mean': 0.030120695009827614, 'clip_ratio/high_max': 0.030120695009827614, 'clip_ratio/region_mean': 0.13990851491689682, 'reward_total_mean': 0.5423731803894043, 'reward_meter_mean': 0.24359747767448425, 'reward_meter_std': 0.2615220844745636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.993793249130249, 'reward_repeat_soft_std': 0.002918896032497287, 'reward_judge_quality_mean': 0.6112499833106995, 'reward_judge_quality_std': 0.15037456154823303, 'reward_total_composite_mean': 0.5423731803894043, 'reward_total_composite_std': 0.13441166281700134, 'epoch': 0.01}

  2%|▏         | 72/3300 [09:40<8:49:19,  9.84s/it]INFO 04-12 22:24:08 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 73/3300 [09:51<9:19:58, 10.41s/it]2026-04-12 22:24:19,769 | INFO | train_grpo_train | metrics_logged mode=train step=73

                                                   
{'loss': 0.0614, 'grad_norm': 18.151248931884766, 'learning_rate': 9.781818181818183e-06, 'num_tokens': 139596.0, 'completions/mean_length': 37.75, 'completions/min_length': 32.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6459850072860718, 'rewards/meter/std': 0.36253878474235535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9962372779846191, 'rewards/repeat_soft/std': 0.00649625901132822, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.6873170137405396, 'rewards/total_composite/std': 0.13498029112815857, 'reward': 0.6873170137405396, 'reward_std': 0.13498029112815857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20438161492347717, 'sampling/sampling_logp_difference/max': 1.8445549011230469, 'sampling/importance_sampling_ratio/min': 0.15809568762779236, 'sampling/importance_sampling_ratio/mean': 1.0211987495422363, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.976985052227974, 'clip_ratio/low_mean': 0.07190257450565696, 'clip_ratio/low_min': 0.07190257450565696, 'clip_ratio/high_mean': 0.08791267033666372, 'clip_ratio/high_max': 0.08791267033666372, 'clip_ratio/region_mean': 0.15981524484232068, 'reward_total_mean': 0.6873170137405396, 'reward_meter_mean': 0.6459850072860718, 'reward_meter_std': 0.36253878474235535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9962372779846191, 'reward_repeat_soft_std': 0.00649625901132822, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.6873170137405396, 'reward_total_composite_std': 0.13498029112815857, 'epoch': 0.01}

  2%|▏         | 73/3300 [09:52<9:19:58, 10.41s/it]INFO 04-12 22:24:20 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 74/3300 [09:59<8:26:48,  9.43s/it]2026-04-12 22:24:26,607 | INFO | train_grpo_train | metrics_logged mode=train step=74

                                                   
{'loss': 0.0588, 'grad_norm': 13.155890464782715, 'learning_rate': 9.77878787878788e-06, 'num_tokens': 141237.0, 'completions/mean_length': 59.125, 'completions/min_length': 51.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6750097274780273, 'rewards/meter/std': 0.36580297350883484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9978744983673096, 'rewards/repeat_soft/std': 0.0031460225582122803, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.33683136105537415, 'rewards/total_composite/mean': 0.6564773321151733, 'rewards/total_composite/std': 0.3408243656158447, 'reward': 0.6564773321151733, 'reward_std': 0.3408243656158447, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15842394530773163, 'sampling/sampling_logp_difference/max': 1.5450174808502197, 'sampling/importance_sampling_ratio/min': 0.2133081555366516, 'sampling/importance_sampling_ratio/mean': 1.0020114183425903, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8815535008907318, 'clip_ratio/low_mean': 0.04728598054498434, 'clip_ratio/low_min': 0.04728598054498434, 'clip_ratio/high_mean': 0.09232382103800774, 'clip_ratio/high_max': 0.09232382103800774, 'clip_ratio/region_mean': 0.13960980158299208, 'reward_total_mean': 0.6564773321151733, 'reward_meter_mean': 0.6750097274780273, 'reward_meter_std': 0.36580297350883484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9978744983673096, 'reward_repeat_soft_std': 0.0031460225582122803, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.33683136105537415, 'reward_total_composite_mean': 0.6564773321151733, 'reward_total_composite_std': 0.3408243656158447, 'epoch': 0.01}

  2%|▏         | 74/3300 [09:59<8:26:48,  9.43s/it]INFO 04-12 22:24:26 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 75/3300 [10:12<9:24:51, 10.51s/it]2026-04-12 22:24:39,640 | INFO | train_grpo_train | metrics_logged mode=train step=75

                                                   
{'loss': -0.0769, 'grad_norm': 9.224102973937988, 'learning_rate': 9.775757575757576e-06, 'num_tokens': 143471.0, 'completions/mean_length': 105.25, 'completions/min_length': 74.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.25, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.7617064714431763, 'rewards/meter/std': 0.3298294246196747, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9947261214256287, 'rewards/repeat_soft/std': 0.006881263107061386, 'rewards/judge_quality/mean': 0.5449999570846558, 'rewards/judge_quality/std': 0.1752549260854721, 'rewards/total_composite/mean': 0.7557405233383179, 'rewards/total_composite/std': 0.17117002606391907, 'reward': 0.7557405233383179, 'reward_std': 0.17117002606391907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20651639997959137, 'sampling/sampling_logp_difference/max': 1.6719465255737305, 'sampling/importance_sampling_ratio/min': 0.1878809928894043, 'sampling/importance_sampling_ratio/mean': 1.0183568000793457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3894700407981873, 'clip_ratio/low_mean': 0.07153031043708324, 'clip_ratio/low_min': 0.07153031043708324, 'clip_ratio/high_mean': 0.1301163136959076, 'clip_ratio/high_max': 0.1301163136959076, 'clip_ratio/region_mean': 0.20164662413299084, 'reward_total_mean': 0.7557405233383179, 'reward_meter_mean': 0.7617064714431763, 'reward_meter_std': 0.3298294246196747, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9947261214256287, 'reward_repeat_soft_std': 0.006881263107061386, 'reward_judge_quality_mean': 0.5449999570846558, 'reward_judge_quality_std': 0.1752549260854721, 'reward_total_composite_mean': 0.7557405233383179, 'reward_total_composite_std': 0.17117002606391907, 'epoch': 0.01}

  2%|▏         | 75/3300 [10:12<9:24:51, 10.51s/it]INFO 04-12 22:24:39 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 76/3300 [10:22<9:16:30, 10.36s/it]2026-04-12 22:24:49,903 | INFO | train_grpo_train | metrics_logged mode=train step=76

                                                   
{'loss': 0.0058, 'grad_norm': 18.25312042236328, 'learning_rate': 9.772727272727273e-06, 'num_tokens': 144964.0, 'completions/mean_length': 32.625, 'completions/min_length': 25.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.708471417427063, 'rewards/meter/std': 0.4398741126060486, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9486144185066223, 'rewards/repeat_soft/std': 0.03010723739862442, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.7481735944747925, 'rewards/total_composite/std': 0.20565913617610931, 'reward': 0.7481735944747925, 'reward_std': 0.20565910637378693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17192530632019043, 'sampling/sampling_logp_difference/max': 0.9567670822143555, 'sampling/importance_sampling_ratio/min': 0.3841327428817749, 'sampling/importance_sampling_ratio/mean': 1.0192999839782715, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8063572198152542, 'clip_ratio/low_mean': 0.06461021397262812, 'clip_ratio/low_min': 0.06461021397262812, 'clip_ratio/high_mean': 0.10132488049566746, 'clip_ratio/high_max': 0.10132488049566746, 'clip_ratio/region_mean': 0.16593509446829557, 'reward_total_mean': 0.7481735944747925, 'reward_meter_mean': 0.708471417427063, 'reward_meter_std': 0.4398741126060486, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9486144185066223, 'reward_repeat_soft_std': 0.03010723739862442, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.7481735944747925, 'reward_total_composite_std': 0.20565913617610931, 'epoch': 0.01}

  2%|▏         | 76/3300 [10:22<9:16:30, 10.36s/it]INFO 04-12 22:24:50 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 77/3300 [10:30<8:40:57,  9.70s/it]2026-04-12 22:24:58,165 | INFO | train_grpo_train | metrics_logged mode=train step=77

                                                   
{'loss': -0.0728, 'grad_norm': 17.3715763092041, 'learning_rate': 9.76969696969697e-06, 'num_tokens': 146633.0, 'completions/mean_length': 52.625, 'completions/min_length': 35.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.2402770221233368, 'rewards/meter/std': 0.31088128685951233, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9913026094436646, 'rewards/repeat_soft/std': 0.009482075460255146, 'rewards/judge_quality/mean': 0.6862499713897705, 'rewards/judge_quality/std': 0.28076112270355225, 'rewards/total_composite/mean': 0.5631299018859863, 'rewards/total_composite/std': 0.1694323569536209, 'reward': 0.5631299018859863, 'reward_std': 0.16943234205245972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1446535289287567, 'sampling/sampling_logp_difference/max': 1.915684700012207, 'sampling/importance_sampling_ratio/min': 0.14724098145961761, 'sampling/importance_sampling_ratio/mean': 1.030705451965332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0392472222447395, 'clip_ratio/low_mean': 0.11356967873871326, 'clip_ratio/low_min': 0.11356967873871326, 'clip_ratio/high_mean': 0.033929698169231415, 'clip_ratio/high_max': 0.033929698169231415, 'clip_ratio/region_mean': 0.14749937690794468, 'reward_total_mean': 0.5631299018859863, 'reward_meter_mean': 0.2402770221233368, 'reward_meter_std': 0.31088128685951233, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9913026094436646, 'reward_repeat_soft_std': 0.009482075460255146, 'reward_judge_quality_mean': 0.6862499713897705, 'reward_judge_quality_std': 0.28076112270355225, 'reward_total_composite_mean': 0.5631299018859863, 'reward_total_composite_std': 0.1694323569536209, 'epoch': 0.01}

  2%|▏         | 77/3300 [10:30<8:40:57,  9.70s/it]INFO 04-12 22:24:58 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 78/3300 [10:38<8:09:31,  9.12s/it]2026-04-12 22:25:05,555 | INFO | train_grpo_train | metrics_logged mode=train step=78

                                                   
{'loss': 0.0969, 'grad_norm': 9.234482765197754, 'learning_rate': 9.766666666666667e-06, 'num_tokens': 149203.0, 'completions/mean_length': 126.25, 'completions/min_length': 107.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.25, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.9073135256767273, 'rewards/meter/std': 0.11808735877275467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976153373718262, 'rewards/repeat_soft/std': 0.0015691010048612952, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.23793382942676544, 'rewards/total_composite/mean': 0.8226776123046875, 'rewards/total_composite/std': 0.1000823825597763, 'reward': 0.8226776123046875, 'reward_std': 0.10008236765861511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19733132421970367, 'sampling/sampling_logp_difference/max': 1.7452077865600586, 'sampling/importance_sampling_ratio/min': 0.17460869252681732, 'sampling/importance_sampling_ratio/mean': 1.0099540948867798, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9953903406858444, 'clip_ratio/low_mean': 0.11173281446099281, 'clip_ratio/low_min': 0.11173281446099281, 'clip_ratio/high_mean': 0.08277627266943455, 'clip_ratio/high_max': 0.08277627266943455, 'clip_ratio/region_mean': 0.19450908713042736, 'reward_total_mean': 0.8226776123046875, 'reward_meter_mean': 0.9073135256767273, 'reward_meter_std': 0.11808735877275467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976153373718262, 'reward_repeat_soft_std': 0.0015691010048612952, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.23793382942676544, 'reward_total_composite_mean': 0.8226776123046875, 'reward_total_composite_std': 0.1000823825597763, 'epoch': 0.01}

  2%|▏         | 78/3300 [10:38<8:09:31,  9.12s/it]INFO 04-12 22:25:05 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 79/3300 [10:44<7:32:18,  8.43s/it]2026-04-12 22:25:12,368 | INFO | train_grpo_train | metrics_logged mode=train step=79

                                                   
{'loss': -0.0129, 'grad_norm': 13.165565490722656, 'learning_rate': 9.763636363636365e-06, 'num_tokens': 151314.0, 'completions/mean_length': 86.875, 'completions/min_length': 71.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.875, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.575776219367981, 'rewards/meter/std': 0.40430617332458496, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9948314428329468, 'rewards/repeat_soft/std': 0.004851632285863161, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.7033324241638184, 'rewards/total_composite/std': 0.19417116045951843, 'reward': 0.7033324241638184, 'reward_std': 0.19417116045951843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20565244555473328, 'sampling/sampling_logp_difference/max': 1.9638829231262207, 'sampling/importance_sampling_ratio/min': 0.1403125375509262, 'sampling/importance_sampling_ratio/mean': 1.03525710105896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8156342655420303, 'clip_ratio/low_mean': 0.09965229965746403, 'clip_ratio/low_min': 0.09965229965746403, 'clip_ratio/high_mean': 0.10430697165429592, 'clip_ratio/high_max': 0.10430697165429592, 'clip_ratio/region_mean': 0.20395927131175995, 'reward_total_mean': 0.7033324241638184, 'reward_meter_mean': 0.575776219367981, 'reward_meter_std': 0.40430617332458496, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9948314428329468, 'reward_repeat_soft_std': 0.004851632285863161, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.7033324241638184, 'reward_total_composite_std': 0.19417116045951843, 'epoch': 0.01}

  2%|▏         | 79/3300 [10:44<7:32:18,  8.43s/it]INFO 04-12 22:25:12 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 80/3300 [10:51<7:02:25,  7.87s/it]2026-04-12 22:25:18,947 | INFO | train_grpo_train | metrics_logged mode=train step=80

                                                   
{'loss': -0.017, 'grad_norm': 13.996071815490723, 'learning_rate': 9.760606060606062e-06, 'num_tokens': 153151.0, 'completions/mean_length': 60.625, 'completions/min_length': 48.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.786838948726654, 'rewards/meter/std': 0.3647293746471405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986039400100708, 'rewards/repeat_soft/std': 0.0022809384390711784, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.11310552060604095, 'rewards/total_composite/mean': 0.7456879615783691, 'rewards/total_composite/std': 0.13831521570682526, 'reward': 0.7456879615783691, 'reward_std': 0.13831521570682526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21806959807872772, 'sampling/sampling_logp_difference/max': 1.4953420162200928, 'sampling/importance_sampling_ratio/min': 0.22417192161083221, 'sampling/importance_sampling_ratio/mean': 1.045628547668457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2772906571626663, 'clip_ratio/low_mean': 0.05426747165620327, 'clip_ratio/low_min': 0.05426747165620327, 'clip_ratio/high_mean': 0.14767173863947392, 'clip_ratio/high_max': 0.14767173863947392, 'clip_ratio/region_mean': 0.20193921029567719, 'reward_total_mean': 0.7456879615783691, 'reward_meter_mean': 0.786838948726654, 'reward_meter_std': 0.3647293746471405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986039400100708, 'reward_repeat_soft_std': 0.0022809384390711784, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.11310552060604095, 'reward_total_composite_mean': 0.7456879615783691, 'reward_total_composite_std': 0.13831521570682526, 'epoch': 0.01}

  2%|▏         | 80/3300 [10:51<7:02:25,  7.87s/it]INFO 04-12 22:25:19 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 81/3300 [11:00<7:23:36,  8.27s/it]2026-04-12 22:25:28,142 | INFO | train_grpo_train | metrics_logged mode=train step=81

                                                   
{'loss': 0.0024, 'grad_norm': 10.145329475402832, 'learning_rate': 9.757575757575758e-06, 'num_tokens': 155821.0, 'completions/mean_length': 123.75, 'completions/min_length': 77.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.75, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.8844622373580933, 'rewards/meter/std': 0.30505692958831787, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9946202039718628, 'rewards/repeat_soft/std': 0.009633982554078102, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.7800325155258179, 'rewards/total_composite/std': 0.14473092555999756, 'reward': 0.7800325155258179, 'reward_std': 0.14473092555999756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20969630777835846, 'sampling/sampling_logp_difference/max': 2.1087703704833984, 'sampling/importance_sampling_ratio/min': 0.12138713151216507, 'sampling/importance_sampling_ratio/mean': 1.0255879163742065, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.578194409608841, 'clip_ratio/low_mean': 0.024572649970650673, 'clip_ratio/low_min': 0.024572649970650673, 'clip_ratio/high_mean': 0.1784006580710411, 'clip_ratio/high_max': 0.1784006580710411, 'clip_ratio/region_mean': 0.20297330804169178, 'reward_total_mean': 0.7800325155258179, 'reward_meter_mean': 0.8844622373580933, 'reward_meter_std': 0.30505692958831787, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9946202039718628, 'reward_repeat_soft_std': 0.009633982554078102, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.7800325155258179, 'reward_total_composite_std': 0.14473092555999756, 'epoch': 0.01}

  2%|▏         | 81/3300 [11:00<7:23:36,  8.27s/it]INFO 04-12 22:25:28 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 82/3300 [11:09<7:25:24,  8.30s/it]2026-04-12 22:25:36,534 | INFO | train_grpo_train | metrics_logged mode=train step=82

                                                   
{'loss': 0.0932, 'grad_norm': 15.969931602478027, 'learning_rate': 9.754545454545455e-06, 'num_tokens': 157487.0, 'completions/mean_length': 59.25, 'completions/min_length': 49.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.5042653679847717, 'rewards/meter/std': 0.4921252727508545, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9970687627792358, 'rewards/repeat_soft/std': 0.004391745664179325, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.6056262850761414, 'rewards/total_composite/std': 0.26372030377388, 'reward': 0.6056262850761414, 'reward_std': 0.26372030377388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21136270463466644, 'sampling/sampling_logp_difference/max': 2.0872573852539062, 'sampling/importance_sampling_ratio/min': 0.1240268275141716, 'sampling/importance_sampling_ratio/mean': 1.0206705331802368, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.811131864786148, 'clip_ratio/low_mean': 0.0915905861184001, 'clip_ratio/low_min': 0.0915905861184001, 'clip_ratio/high_mean': 0.09937021695077419, 'clip_ratio/high_max': 0.09937021695077419, 'clip_ratio/region_mean': 0.1909608030691743, 'reward_total_mean': 0.6056262850761414, 'reward_meter_mean': 0.5042653679847717, 'reward_meter_std': 0.4921252727508545, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9970687627792358, 'reward_repeat_soft_std': 0.004391745664179325, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.6056262850761414, 'reward_total_composite_std': 0.26372030377388, 'epoch': 0.01}

  2%|▏         | 82/3300 [11:09<7:25:24,  8.30s/it]INFO 04-12 22:25:36 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 83/3300 [11:16<7:09:43,  8.01s/it]2026-04-12 22:25:43,869 | INFO | train_grpo_train | metrics_logged mode=train step=83

                                                   
{'loss': 0.0093, 'grad_norm': 13.831247329711914, 'learning_rate': 9.751515151515152e-06, 'num_tokens': 159104.0, 'completions/mean_length': 57.125, 'completions/min_length': 40.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7920023202896118, 'rewards/meter/std': 0.3581661581993103, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975086450576782, 'rewards/repeat_soft/std': 0.003981848247349262, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.2822834253311157, 'rewards/total_composite/mean': 0.7820268869400024, 'rewards/total_composite/std': 0.22069051861763, 'reward': 0.7820268869400024, 'reward_std': 0.22069051861763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18509216606616974, 'sampling/sampling_logp_difference/max': 1.5339620113372803, 'sampling/importance_sampling_ratio/min': 0.21567945182323456, 'sampling/importance_sampling_ratio/mean': 1.0375295877456665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7064232379198074, 'clip_ratio/low_mean': 0.044791667722165585, 'clip_ratio/low_min': 0.044791667722165585, 'clip_ratio/high_mean': 0.13707098085433245, 'clip_ratio/high_max': 0.13707098085433245, 'clip_ratio/region_mean': 0.18186264857649803, 'reward_total_mean': 0.7820268869400024, 'reward_meter_mean': 0.7920023202896118, 'reward_meter_std': 0.3581661581993103, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975086450576782, 'reward_repeat_soft_std': 0.003981848247349262, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.2822834253311157, 'reward_total_composite_mean': 0.7820268869400024, 'reward_total_composite_std': 0.22069051861763, 'epoch': 0.01}

  3%|▎         | 83/3300 [11:16<7:09:43,  8.01s/it]INFO 04-12 22:25:44 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 84/3300 [11:23<6:59:59,  7.84s/it]2026-04-12 22:25:51,331 | INFO | train_grpo_train | metrics_logged mode=train step=84

                                                   
{'loss': -0.0464, 'grad_norm': 18.266530990600586, 'learning_rate': 9.74848484848485e-06, 'num_tokens': 160981.0, 'completions/mean_length': 53.625, 'completions/min_length': 38.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.674973726272583, 'rewards/meter/std': 0.4194721281528473, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9981741905212402, 'rewards/repeat_soft/std': 0.0017913100309669971, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.682930588722229, 'rewards/total_composite/std': 0.18727268278598785, 'reward': 0.682930588722229, 'reward_std': 0.18727268278598785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2229832410812378, 'sampling/sampling_logp_difference/max': 1.926133155822754, 'sampling/importance_sampling_ratio/min': 0.1657637655735016, 'sampling/importance_sampling_ratio/mean': 1.0519226789474487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2065766006708145, 'clip_ratio/low_mean': 0.07867324538528919, 'clip_ratio/low_min': 0.07867324538528919, 'clip_ratio/high_mean': 0.11280264891684055, 'clip_ratio/high_max': 0.11280264891684055, 'clip_ratio/region_mean': 0.19147589430212975, 'reward_total_mean': 0.682930588722229, 'reward_meter_mean': 0.674973726272583, 'reward_meter_std': 0.4194721281528473, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9981741905212402, 'reward_repeat_soft_std': 0.0017913100309669971, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.682930588722229, 'reward_total_composite_std': 0.18727268278598785, 'epoch': 0.01}

  3%|▎         | 84/3300 [11:23<6:59:59,  7.84s/it]INFO 04-12 22:25:51 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 85/3300 [11:29<6:33:34,  7.35s/it]2026-04-12 22:25:57,493 | INFO | train_grpo_train | metrics_logged mode=train step=85

                                                   
{'loss': 0.1016, 'grad_norm': 26.06890106201172, 'learning_rate': 9.745454545454547e-06, 'num_tokens': 162397.0, 'completions/mean_length': 32.0, 'completions/min_length': 15.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5776013731956482, 'rewards/meter/std': 0.4645298421382904, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5890144109725952, 'rewards/total_composite/std': 0.3000996708869934, 'reward': 0.5890144109725952, 'reward_std': 0.3000996708869934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2313683032989502, 'sampling/sampling_logp_difference/max': 1.7472820281982422, 'sampling/importance_sampling_ratio/min': 0.1742469072341919, 'sampling/importance_sampling_ratio/mean': 1.0024120807647705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7302970439195633, 'clip_ratio/low_mean': 0.0708547318354249, 'clip_ratio/low_min': 0.0708547318354249, 'clip_ratio/high_mean': 0.1231324952095747, 'clip_ratio/high_max': 0.1231324952095747, 'clip_ratio/region_mean': 0.1939872270449996, 'reward_total_mean': 0.5890144109725952, 'reward_meter_mean': 0.5776013731956482, 'reward_meter_std': 0.4645298421382904, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5890144109725952, 'reward_total_composite_std': 0.3000996708869934, 'epoch': 0.01}

  3%|▎         | 85/3300 [11:29<6:33:34,  7.35s/it]INFO 04-12 22:25:57 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 86/3300 [11:36<6:26:34,  7.22s/it]2026-04-12 22:26:04,459 | INFO | train_grpo_train | metrics_logged mode=train step=86

                                                   
{'loss': 0.0885, 'grad_norm': 19.318851470947266, 'learning_rate': 9.742424242424244e-06, 'num_tokens': 164496.0, 'completions/mean_length': 83.375, 'completions/min_length': 70.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.5039986968040466, 'rewards/meter/std': 0.3009590804576874, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974728226661682, 'rewards/repeat_soft/std': 0.0023224493488669395, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5930466651916504, 'rewards/total_composite/std': 0.14554795622825623, 'reward': 0.5930466651916504, 'reward_std': 0.14554795622825623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.234708771109581, 'sampling/sampling_logp_difference/max': 2.4151086807250977, 'sampling/importance_sampling_ratio/min': 0.08935762941837311, 'sampling/importance_sampling_ratio/mean': 0.986754834651947, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1506693363189697, 'clip_ratio/low_mean': 0.07777538895606995, 'clip_ratio/low_min': 0.07777538895606995, 'clip_ratio/high_mean': 0.1278723981231451, 'clip_ratio/high_max': 0.1278723981231451, 'clip_ratio/region_mean': 0.20564778707921505, 'reward_total_mean': 0.5930466651916504, 'reward_meter_mean': 0.5039986968040466, 'reward_meter_std': 0.3009590804576874, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974728226661682, 'reward_repeat_soft_std': 0.0023224493488669395, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5930466651916504, 'reward_total_composite_std': 0.14554795622825623, 'epoch': 0.01}

  3%|▎         | 86/3300 [11:36<6:26:34,  7.22s/it]INFO 04-12 22:26:04 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 87/3300 [11:43<6:23:11,  7.16s/it]2026-04-12 22:26:11,419 | INFO | train_grpo_train | metrics_logged mode=train step=87

                                                   
{'loss': 0.047, 'grad_norm': 32.492706298828125, 'learning_rate': 9.739393939393941e-06, 'num_tokens': 165854.0, 'completions/mean_length': 28.75, 'completions/min_length': 25.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8606054782867432, 'rewards/meter/std': 0.29568642377853394, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9594065546989441, 'rewards/repeat_soft/std': 0.008749538101255894, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.8177131414413452, 'rewards/total_composite/std': 0.17395693063735962, 'reward': 0.8177131414413452, 'reward_std': 0.17395691573619843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1784212738275528, 'sampling/sampling_logp_difference/max': 1.7252826690673828, 'sampling/importance_sampling_ratio/min': 0.17812269926071167, 'sampling/importance_sampling_ratio/mean': 1.010380744934082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4293856024742126, 'clip_ratio/low_mean': 0.07695782743394375, 'clip_ratio/low_min': 0.07695782743394375, 'clip_ratio/high_mean': 0.11240384820848703, 'clip_ratio/high_max': 0.11240384820848703, 'clip_ratio/region_mean': 0.18936167564243078, 'reward_total_mean': 0.8177131414413452, 'reward_meter_mean': 0.8606054782867432, 'reward_meter_std': 0.29568642377853394, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9594065546989441, 'reward_repeat_soft_std': 0.008749538101255894, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.8177131414413452, 'reward_total_composite_std': 0.17395693063735962, 'epoch': 0.01}

  3%|▎         | 87/3300 [11:43<6:23:11,  7.16s/it]INFO 04-12 22:26:11 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 88/3300 [11:50<6:10:03,  6.91s/it]2026-04-12 22:26:17,765 | INFO | train_grpo_train | metrics_logged mode=train step=88

                                                   
{'loss': 0.0314, 'grad_norm': 15.27247142791748, 'learning_rate': 9.736363636363637e-06, 'num_tokens': 167652.0, 'completions/mean_length': 63.75, 'completions/min_length': 62.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6508787870407104, 'rewards/meter/std': 0.44921696186065674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9988118410110474, 'rewards/repeat_soft/std': 0.0022364838514477015, 'rewards/judge_quality/mean': 0.5387500524520874, 'rewards/judge_quality/std': 0.24485784769058228, 'rewards/total_composite/mean': 0.7044016122817993, 'rewards/total_composite/std': 0.15757334232330322, 'reward': 0.7044016122817993, 'reward_std': 0.15757334232330322, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20896635949611664, 'sampling/sampling_logp_difference/max': 1.6069178581237793, 'sampling/importance_sampling_ratio/min': 0.20050464570522308, 'sampling/importance_sampling_ratio/mean': 1.016836166381836, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5759567469358444, 'clip_ratio/low_mean': 0.0732837338000536, 'clip_ratio/low_min': 0.0732837338000536, 'clip_ratio/high_mean': 0.13178288005292416, 'clip_ratio/high_max': 0.13178288005292416, 'clip_ratio/region_mean': 0.20506661385297775, 'reward_total_mean': 0.7044016122817993, 'reward_meter_mean': 0.6508787870407104, 'reward_meter_std': 0.44921696186065674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9988118410110474, 'reward_repeat_soft_std': 0.0022364838514477015, 'reward_judge_quality_mean': 0.5387500524520874, 'reward_judge_quality_std': 0.24485784769058228, 'reward_total_composite_mean': 0.7044016122817993, 'reward_total_composite_std': 0.15757334232330322, 'epoch': 0.01}

  3%|▎         | 88/3300 [11:50<6:10:03,  6.91s/it]INFO 04-12 22:26:18 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 89/3300 [11:57<6:14:11,  6.99s/it]2026-04-12 22:26:24,947 | INFO | train_grpo_train | metrics_logged mode=train step=89

                                                   
{'loss': 0.0346, 'grad_norm': 14.53501033782959, 'learning_rate': 9.733333333333334e-06, 'num_tokens': 169728.0, 'completions/mean_length': 94.5, 'completions/min_length': 91.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.5, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.7820674180984497, 'rewards/meter/std': 0.24537409842014313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976636171340942, 'rewards/repeat_soft/std': 0.0012526443460956216, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.765196681022644, 'rewards/total_composite/std': 0.10896959155797958, 'reward': 0.765196681022644, 'reward_std': 0.10896960645914078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14851930737495422, 'sampling/sampling_logp_difference/max': 1.868389368057251, 'sampling/importance_sampling_ratio/min': 0.15437209606170654, 'sampling/importance_sampling_ratio/mean': 1.010309100151062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7215096801519394, 'clip_ratio/low_mean': 0.03587029594928026, 'clip_ratio/low_min': 0.03587029594928026, 'clip_ratio/high_mean': 0.09394399169832468, 'clip_ratio/high_max': 0.09394399169832468, 'clip_ratio/region_mean': 0.12981428764760494, 'reward_total_mean': 0.765196681022644, 'reward_meter_mean': 0.7820674180984497, 'reward_meter_std': 0.24537409842014313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976636171340942, 'reward_repeat_soft_std': 0.0012526443460956216, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.765196681022644, 'reward_total_composite_std': 0.10896959155797958, 'epoch': 0.01}

  3%|▎         | 89/3300 [11:57<6:14:11,  6.99s/it]INFO 04-12 22:26:25 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 90/3300 [12:03<5:59:18,  6.72s/it]2026-04-12 22:26:31,050 | INFO | train_grpo_train | metrics_logged mode=train step=90

                                                   
{'loss': 0.0817, 'grad_norm': 29.59337043762207, 'learning_rate': 9.730303030303031e-06, 'num_tokens': 171248.0, 'completions/mean_length': 33.0, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.418997198343277, 'rewards/meter/std': 0.3585245907306671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9966121912002563, 'rewards/repeat_soft/std': 0.00797121599316597, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.1865811049938202, 'rewards/total_composite/mean': 0.5975849628448486, 'rewards/total_composite/std': 0.19449637830257416, 'reward': 0.5975849628448486, 'reward_std': 0.19449637830257416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25027692317962646, 'sampling/sampling_logp_difference/max': 2.5391225814819336, 'sampling/importance_sampling_ratio/min': 0.07893563061952591, 'sampling/importance_sampling_ratio/mean': 0.9951542019844055, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5199098736047745, 'clip_ratio/low_mean': 0.09931824635714293, 'clip_ratio/low_min': 0.09931824635714293, 'clip_ratio/high_mean': 0.11300505325198174, 'clip_ratio/high_max': 0.11300505325198174, 'clip_ratio/region_mean': 0.21232329960912466, 'reward_total_mean': 0.5975849628448486, 'reward_meter_mean': 0.418997198343277, 'reward_meter_std': 0.3585245907306671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9966121912002563, 'reward_repeat_soft_std': 0.00797121599316597, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.1865811049938202, 'reward_total_composite_mean': 0.5975849628448486, 'reward_total_composite_std': 0.19449637830257416, 'epoch': 0.01}

  3%|▎         | 90/3300 [12:03<5:59:18,  6.72s/it]INFO 04-12 22:26:31 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 91/3300 [12:11<6:21:24,  7.13s/it]2026-04-12 22:26:39,167 | INFO | train_grpo_train | metrics_logged mode=train step=91

                                                   
{'loss': 0.1256, 'grad_norm': 10.04101848602295, 'learning_rate': 9.727272727272728e-06, 'num_tokens': 173561.0, 'completions/mean_length': 116.125, 'completions/min_length': 95.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.125, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.693138837814331, 'rewards/meter/std': 0.36694416403770447, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9948671460151672, 'rewards/repeat_soft/std': 0.005335790105164051, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291432082653046, 'rewards/total_composite/mean': 0.6941492557525635, 'rewards/total_composite/std': 0.18267148733139038, 'reward': 0.6941492557525635, 'reward_std': 0.18267148733139038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21353013813495636, 'sampling/sampling_logp_difference/max': 2.810418128967285, 'sampling/importance_sampling_ratio/min': 0.060179825872182846, 'sampling/importance_sampling_ratio/mean': 1.0399506092071533, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0453978031873703, 'clip_ratio/low_mean': 0.05563131347298622, 'clip_ratio/low_min': 0.05563131347298622, 'clip_ratio/high_mean': 0.15723679400980473, 'clip_ratio/high_max': 0.15723679400980473, 'clip_ratio/region_mean': 0.21286810748279095, 'reward_total_mean': 0.6941492557525635, 'reward_meter_mean': 0.693138837814331, 'reward_meter_std': 0.36694416403770447, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9948671460151672, 'reward_repeat_soft_std': 0.005335790105164051, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291432082653046, 'reward_total_composite_mean': 0.6941492557525635, 'reward_total_composite_std': 0.18267148733139038, 'epoch': 0.01}

  3%|▎         | 91/3300 [12:11<6:21:24,  7.13s/it]INFO 04-12 22:26:39 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 92/3300 [12:19<6:40:09,  7.48s/it]2026-04-12 22:26:47,422 | INFO | train_grpo_train | metrics_logged mode=train step=92

                                                   
{'loss': 0.0532, 'grad_norm': 14.25716781616211, 'learning_rate': 9.724242424242426e-06, 'num_tokens': 175372.0, 'completions/mean_length': 58.375, 'completions/min_length': 41.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.26952993869781494, 'rewards/meter/std': 0.3523905575275421, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9935294389724731, 'rewards/repeat_soft/std': 0.008154787123203278, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.500016450881958, 'rewards/total_composite/std': 0.15747496485710144, 'reward': 0.500016450881958, 'reward_std': 0.15747497975826263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.218789741396904, 'sampling/sampling_logp_difference/max': 1.3333110809326172, 'sampling/importance_sampling_ratio/min': 0.2674177289009094, 'sampling/importance_sampling_ratio/mean': 1.0159591436386108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6216311752796173, 'clip_ratio/low_mean': 0.1285328259691596, 'clip_ratio/low_min': 0.1285328259691596, 'clip_ratio/high_mean': 0.06342516466975212, 'clip_ratio/high_max': 0.06342516466975212, 'clip_ratio/region_mean': 0.19195799063891172, 'reward_total_mean': 0.500016450881958, 'reward_meter_mean': 0.26952993869781494, 'reward_meter_std': 0.3523905575275421, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9935294389724731, 'reward_repeat_soft_std': 0.008154787123203278, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.500016450881958, 'reward_total_composite_std': 0.15747496485710144, 'epoch': 0.01}

  3%|▎         | 92/3300 [12:19<6:40:09,  7.48s/it]INFO 04-12 22:26:47 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 93/3300 [12:26<6:26:19,  7.23s/it]2026-04-12 22:26:54,051 | INFO | train_grpo_train | metrics_logged mode=train step=93

                                                   
{'loss': 0.0745, 'grad_norm': 10.899441719055176, 'learning_rate': 9.721212121212123e-06, 'num_tokens': 177106.0, 'completions/mean_length': 69.75, 'completions/min_length': 53.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.75, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9947660565376282, 'rewards/meter/std': 0.0041707707569003105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986412525177002, 'rewards/repeat_soft/std': 0.002481638453900814, 'rewards/judge_quality/mean': 0.3725000023841858, 'rewards/judge_quality/std': 0.11055056750774384, 'rewards/total_composite/mean': 0.8092588186264038, 'rewards/total_composite/std': 0.033664677292108536, 'reward': 0.8092588186264038, 'reward_std': 0.033664681017398834, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17616569995880127, 'sampling/sampling_logp_difference/max': 1.2230968475341797, 'sampling/importance_sampling_ratio/min': 0.2943173050880432, 'sampling/importance_sampling_ratio/mean': 1.0400642156600952, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3353325724601746, 'clip_ratio/low_mean': 0.038446055725216866, 'clip_ratio/low_min': 0.038446055725216866, 'clip_ratio/high_mean': 0.14749629236757755, 'clip_ratio/high_max': 0.14749629236757755, 'clip_ratio/region_mean': 0.18594234809279442, 'reward_total_mean': 0.8092588186264038, 'reward_meter_mean': 0.9947660565376282, 'reward_meter_std': 0.0041707707569003105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986412525177002, 'reward_repeat_soft_std': 0.002481638453900814, 'reward_judge_quality_mean': 0.3725000023841858, 'reward_judge_quality_std': 0.11055056750774384, 'reward_total_composite_mean': 0.8092588186264038, 'reward_total_composite_std': 0.033664677292108536, 'epoch': 0.01}

  3%|▎         | 93/3300 [12:26<6:26:19,  7.23s/it]INFO 04-12 22:26:54 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 94/3300 [12:33<6:18:57,  7.09s/it]2026-04-12 22:27:00,885 | INFO | train_grpo_train | metrics_logged mode=train step=94

                                                   
{'loss': 0.0462, 'grad_norm': 17.971059799194336, 'learning_rate': 9.718181818181818e-06, 'num_tokens': 178827.0, 'completions/mean_length': 37.125, 'completions/min_length': 32.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.41076311469078064, 'rewards/meter/std': 0.3964845538139343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9856504797935486, 'rewards/repeat_soft/std': 0.010154510848224163, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.6366584300994873, 'rewards/total_composite/std': 0.15718963742256165, 'reward': 0.6366584300994873, 'reward_std': 0.15718962252140045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19394990801811218, 'sampling/sampling_logp_difference/max': 1.2942476272583008, 'sampling/importance_sampling_ratio/min': 0.2741039991378784, 'sampling/importance_sampling_ratio/mean': 1.0088485479354858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3612593337893486, 'clip_ratio/low_mean': 0.08777131512761116, 'clip_ratio/low_min': 0.08777131512761116, 'clip_ratio/high_mean': 0.07181490398943424, 'clip_ratio/high_max': 0.07181490398943424, 'clip_ratio/region_mean': 0.1595862191170454, 'reward_total_mean': 0.6366584300994873, 'reward_meter_mean': 0.41076311469078064, 'reward_meter_std': 0.3964845538139343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9856504797935486, 'reward_repeat_soft_std': 0.010154510848224163, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.6366584300994873, 'reward_total_composite_std': 0.15718963742256165, 'epoch': 0.01}

  3%|▎         | 94/3300 [12:33<6:18:57,  7.09s/it]INFO 04-12 22:27:01 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 95/3300 [12:41<6:30:42,  7.31s/it]2026-04-12 22:27:08,659 | INFO | train_grpo_train | metrics_logged mode=train step=95

                                                   
{'loss': 0.0122, 'grad_norm': 10.587267875671387, 'learning_rate': 9.715151515151516e-06, 'num_tokens': 181305.0, 'completions/mean_length': 126.75, 'completions/min_length': 107.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.75, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.9200213551521301, 'rewards/meter/std': 0.1255701780319214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9956197738647461, 'rewards/repeat_soft/std': 0.003646343480795622, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.21293865144252777, 'rewards/total_composite/mean': 0.8480715751647949, 'rewards/total_composite/std': 0.08985444903373718, 'reward': 0.8480715751647949, 'reward_std': 0.08985444158315659, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18390880525112152, 'sampling/sampling_logp_difference/max': 1.2477107048034668, 'sampling/importance_sampling_ratio/min': 0.2871614396572113, 'sampling/importance_sampling_ratio/mean': 1.0285959243774414, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9829654097557068, 'clip_ratio/low_mean': 0.08566484786570072, 'clip_ratio/low_min': 0.08566484786570072, 'clip_ratio/high_mean': 0.09545699134469032, 'clip_ratio/high_max': 0.09545699134469032, 'clip_ratio/region_mean': 0.18112183921039104, 'reward_total_mean': 0.8480715751647949, 'reward_meter_mean': 0.9200213551521301, 'reward_meter_std': 0.1255701780319214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9956197738647461, 'reward_repeat_soft_std': 0.003646343480795622, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.21293865144252777, 'reward_total_composite_mean': 0.8480715751647949, 'reward_total_composite_std': 0.08985444903373718, 'epoch': 0.01}

  3%|▎         | 95/3300 [12:41<6:30:42,  7.31s/it]INFO 04-12 22:27:08 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 96/3300 [12:47<6:16:59,  7.06s/it]2026-04-12 22:27:15,126 | INFO | train_grpo_train | metrics_logged mode=train step=96

                                                   
{'loss': 0.1148, 'grad_norm': 14.349672317504883, 'learning_rate': 9.712121212121213e-06, 'num_tokens': 183026.0, 'completions/mean_length': 60.125, 'completions/min_length': 41.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.37904417514801025, 'rewards/meter/std': 0.36078208684921265, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9731553792953491, 'rewards/repeat_soft/std': 0.04177224636077881, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.5731354355812073, 'rewards/total_composite/std': 0.18765978515148163, 'reward': 0.5731354355812073, 'reward_std': 0.18765978515148163, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20568451285362244, 'sampling/sampling_logp_difference/max': 1.8774585723876953, 'sampling/importance_sampling_ratio/min': 0.15297840535640717, 'sampling/importance_sampling_ratio/mean': 1.0121736526489258, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8408035784959793, 'clip_ratio/low_mean': 0.09115871600806713, 'clip_ratio/low_min': 0.09115871600806713, 'clip_ratio/high_mean': 0.10624979063868523, 'clip_ratio/high_max': 0.10624979063868523, 'clip_ratio/region_mean': 0.19740850664675236, 'reward_total_mean': 0.5731354355812073, 'reward_meter_mean': 0.37904417514801025, 'reward_meter_std': 0.36078208684921265, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9731553792953491, 'reward_repeat_soft_std': 0.04177224636077881, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.5731354355812073, 'reward_total_composite_std': 0.18765978515148163, 'epoch': 0.01}

  3%|▎         | 96/3300 [12:47<6:16:59,  7.06s/it]INFO 04-12 22:27:15 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 97/3300 [12:54<6:06:52,  6.87s/it]2026-04-12 22:27:21,560 | INFO | train_grpo_train | metrics_logged mode=train step=97

                                                   
{'loss': -0.0658, 'grad_norm': 13.329829216003418, 'learning_rate': 9.70909090909091e-06, 'num_tokens': 184785.0, 'completions/mean_length': 54.875, 'completions/min_length': 35.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.44293856620788574, 'rewards/meter/std': 0.3831910789012909, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9962660074234009, 'rewards/repeat_soft/std': 0.004540997091680765, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.4797763526439667, 'rewards/total_composite/std': 0.24937774240970612, 'reward': 0.4797763526439667, 'reward_std': 0.24937772750854492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18477389216423035, 'sampling/sampling_logp_difference/max': 1.34785795211792, 'sampling/importance_sampling_ratio/min': 0.2597961723804474, 'sampling/importance_sampling_ratio/mean': 1.0322825908660889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9412241727113724, 'clip_ratio/low_mean': 0.07615137845277786, 'clip_ratio/low_min': 0.07615137845277786, 'clip_ratio/high_mean': 0.08887524623423815, 'clip_ratio/high_max': 0.08887524623423815, 'clip_ratio/region_mean': 0.165026624687016, 'reward_total_mean': 0.4797763526439667, 'reward_meter_mean': 0.44293856620788574, 'reward_meter_std': 0.3831910789012909, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9962660074234009, 'reward_repeat_soft_std': 0.004540997091680765, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.4797763526439667, 'reward_total_composite_std': 0.24937774240970612, 'epoch': 0.01}

  3%|▎         | 97/3300 [12:54<6:06:52,  6.87s/it]INFO 04-12 22:27:21 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 98/3300 [13:01<6:11:02,  6.95s/it]2026-04-12 22:27:28,704 | INFO | train_grpo_train | metrics_logged mode=train step=98

                                                   
{'loss': 0.0294, 'grad_norm': 10.771571159362793, 'learning_rate': 9.706060606060606e-06, 'num_tokens': 186626.0, 'completions/mean_length': 66.125, 'completions/min_length': 57.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.6990766525268555, 'rewards/meter/std': 0.38784509897232056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9952908754348755, 'rewards/repeat_soft/std': 0.007525984197854996, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.19167961180210114, 'rewards/total_composite/mean': 0.7069886326789856, 'rewards/total_composite/std': 0.19634383916854858, 'reward': 0.7069886326789856, 'reward_std': 0.19634383916854858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19846424460411072, 'sampling/sampling_logp_difference/max': 1.7335278987884521, 'sampling/importance_sampling_ratio/min': 0.17666007578372955, 'sampling/importance_sampling_ratio/mean': 1.0128724575042725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7142320424318314, 'clip_ratio/low_mean': 0.04686454962939024, 'clip_ratio/low_min': 0.04686454962939024, 'clip_ratio/high_mean': 0.12352950684726238, 'clip_ratio/high_max': 0.12352950684726238, 'clip_ratio/region_mean': 0.17039405647665262, 'reward_total_mean': 0.7069886326789856, 'reward_meter_mean': 0.6990766525268555, 'reward_meter_std': 0.38784509897232056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9952908754348755, 'reward_repeat_soft_std': 0.007525984197854996, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.19167961180210114, 'reward_total_composite_mean': 0.7069886326789856, 'reward_total_composite_std': 0.19634383916854858, 'epoch': 0.01}

  3%|▎         | 98/3300 [13:01<6:11:02,  6.95s/it]INFO 04-12 22:27:28 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 99/3300 [13:07<6:02:13,  6.79s/it]2026-04-12 22:27:35,114 | INFO | train_grpo_train | metrics_logged mode=train step=99

                                                   
{'loss': 0.0833, 'grad_norm': 17.894399642944336, 'learning_rate': 9.703030303030305e-06, 'num_tokens': 188002.0, 'completions/mean_length': 28.0, 'completions/min_length': 21.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.49489158391952515, 'rewards/meter/std': 0.4102099537849426, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9463370442390442, 'rewards/repeat_soft/std': 0.02908223308622837, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5978349447250366, 'rewards/total_composite/std': 0.18156224489212036, 'reward': 0.5978349447250366, 'reward_std': 0.18156225979328156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.172475203871727, 'sampling/sampling_logp_difference/max': 1.023299217224121, 'sampling/importance_sampling_ratio/min': 0.3594072461128235, 'sampling/importance_sampling_ratio/mean': 1.049785852432251, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2047024965286255, 'clip_ratio/low_mean': 0.08318580687046051, 'clip_ratio/low_min': 0.08318580687046051, 'clip_ratio/high_mean': 0.07235023286193609, 'clip_ratio/high_max': 0.07235023286193609, 'clip_ratio/region_mean': 0.1555360397323966, 'reward_total_mean': 0.5978349447250366, 'reward_meter_mean': 0.49489158391952515, 'reward_meter_std': 0.4102099537849426, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9463370442390442, 'reward_repeat_soft_std': 0.02908223308622837, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5978349447250366, 'reward_total_composite_std': 0.18156224489212036, 'epoch': 0.01}

  3%|▎         | 99/3300 [13:07<6:02:13,  6.79s/it]INFO 04-12 22:27:35 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 100/3300 [13:14<6:05:01,  6.84s/it]2026-04-12 22:27:42,082 | INFO | train_grpo_train | metrics_logged mode=train step=100

                                                    
{'loss': 0.0883, 'grad_norm': 10.852298736572266, 'learning_rate': 9.7e-06, 'num_tokens': 190092.0, 'completions/mean_length': 87.25, 'completions/min_length': 70.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.25189706683158875, 'rewards/meter/std': 0.2712648808956146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9939076900482178, 'rewards/repeat_soft/std': 0.004518185276538134, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5112444162368774, 'rewards/total_composite/std': 0.12693306803703308, 'reward': 0.5112444162368774, 'reward_std': 0.1269330531358719, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21233618259429932, 'sampling/sampling_logp_difference/max': 1.5831584930419922, 'sampling/importance_sampling_ratio/min': 0.20532555878162384, 'sampling/importance_sampling_ratio/mean': 1.0322673320770264, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8813354671001434, 'clip_ratio/low_mean': 0.14453337900340557, 'clip_ratio/low_min': 0.14453337900340557, 'clip_ratio/high_mean': 0.053007518872618675, 'clip_ratio/high_max': 0.053007518872618675, 'clip_ratio/region_mean': 0.19754089787602425, 'reward_total_mean': 0.5112444162368774, 'reward_meter_mean': 0.25189706683158875, 'reward_meter_std': 0.2712648808956146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9939076900482178, 'reward_repeat_soft_std': 0.004518185276538134, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5112444162368774, 'reward_total_composite_std': 0.12693306803703308, 'epoch': 0.01}

  3%|▎         | 100/3300 [13:14<6:05:01,  6.84s/it]INFO 04-12 22:27:42 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:16,  2.10s/it]

 30%|███       | 3/10 [00:09<00:23,  3.36s/it]

 40%|████      | 4/10 [00:13<00:21,  3.65s/it]

 50%|█████     | 5/10 [00:19<00:22,  4.44s/it]

 60%|██████    | 6/10 [00:26<00:21,  5.35s/it]

 70%|███████   | 7/10 [00:36<00:20,  6.92s/it]

 80%|████████  | 8/10 [00:41<00:12,  6.17s/it]

 90%|█████████ | 9/10 [00:45<00:05,  5.59s/it]

100%|██████████| 10/10 [00:48<00:00,  4.89s/it]2026-04-12 22:28:35,565 | INFO | train_grpo_train | metrics_logged mode=eval step=100

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 53.4733, 'eval_samples_per_second': 1.496, 'eval_steps_per_second': 0.187, 'eval_num_tokens': 190092.0, 'eval_completions/mean_length': 89.825, 'eval_completions/min_length': 35.7, 'eval_completions/max_length': 160.4, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 89.825, 'eval_completions/min_terminated_length': 35.7, 'eval_completions/max_terminated_length': 160.4, 'eval_rewards/meter/mean': 0.5703477174043655, 'eval_rewards/meter/std': 0.38051448166370394, 'eval_rewards/count_adherence/mean': 0.9891666650772095, 'eval_rewards/count_adherence/std': 0.025364020839333534, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.2230676978826523, 'eval_rewards/repeat_soft/mean': 0.9923222184181213, 'eval_rewards/repeat_soft/std': 0.011972753837471827, 'eval_rewards/judge_quality/mean': 0.48449999690055845, 'eval_rewards/judge_quality/std': 0.16168474704027175, 'eval_rewards/total_composite/mean': 0.595814323425293, 'eval_rewards/total_composite/std': 0.2441583454608917, 'eval_reward': 0.595814323425293, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.13042871057987213, 'eval_sampling/sampling_logp_difference/max': 1.1281907081604003, 'eval_sampling/importance_sampling_ratio/min': 0.3298598065972328, 'eval_sampling/importance_sampling_ratio/mean': 1.0367148160934447, 'eval_sampling/importance_sampling_ratio/max': 1.5091773986816406, 'eval_entropy': 1.9399710893630981, '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': 0.595814323425293, 'eval_reward_meter_mean': 0.5703477174043655, 'eval_reward_meter_std': 0.38051448166370394, 'eval_reward_count_adherence_mean': 0.9891666650772095, 'eval_reward_count_adherence_std': 0.025364020839333534, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.2230676978826523, 'eval_reward_repeat_soft_mean': 0.9923222184181213, 'eval_reward_repeat_soft_std': 0.011972753837471827, 'eval_reward_judge_quality_mean': 0.48449999690055845, 'eval_reward_judge_quality_std': 0.16168474704027175, 'eval_reward_total_composite_mean': 0.595814323425293, 'eval_reward_total_composite_std': 0.2441583454608917, 'epoch': 0.01}

  3%|▎         | 100/3300 [14:08<6:05:01,  6.84s/it]

100%|██████████| 10/10 [00:48<00:00,  4.89s/it]

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

  3%|▎         | 101/3300 [14:19<21:31:50, 24.23s/it]2026-04-12 22:28:46,892 | INFO | train_grpo_train | metrics_logged mode=train step=101

                                                     
{'loss': 0.0307, 'grad_norm': 8.720218658447266, 'learning_rate': 9.696969696969698e-06, 'num_tokens': 192271.0, 'completions/mean_length': 96.375, 'completions/min_length': 83.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.375, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9825379252433777, 'rewards/meter/std': 0.03238074481487274, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9823436737060547, 'rewards/repeat_soft/std': 0.021364767104387283, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.8100014328956604, 'rewards/total_composite/std': 0.02203506976366043, 'reward': 0.8100014328956604, 'reward_std': 0.02203506790101528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19199949502944946, 'sampling/sampling_logp_difference/max': 1.4776219129562378, 'sampling/importance_sampling_ratio/min': 0.2281796783208847, 'sampling/importance_sampling_ratio/mean': 1.0370380878448486, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.443674087524414, 'clip_ratio/low_mean': 0.04417553171515465, 'clip_ratio/low_min': 0.04417553171515465, 'clip_ratio/high_mean': 0.1387271974235773, 'clip_ratio/high_max': 0.1387271974235773, 'clip_ratio/region_mean': 0.18290272913873196, 'reward_total_mean': 0.8100014328956604, 'reward_meter_mean': 0.9825379252433777, 'reward_meter_std': 0.03238074481487274, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9823436737060547, 'reward_repeat_soft_std': 0.021364767104387283, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.8100014328956604, 'reward_total_composite_std': 0.02203506976366043, 'epoch': 0.01}

  3%|▎         | 101/3300 [14:19<21:31:50, 24.23s/it]INFO 04-12 22:28:47 [block_pool.py:378] Successfully reset prefix cache