INFO 04-13 12:43:52 [__init__.py:216] Automatically detected platform cuda.
2026-04-13 12:43:53,675 | INFO | train_grpo_train | mode=train

Fetching 128 files:   0%|          | 0/128 [00:00<?, ?it/s]
Fetching 128 files: 100%|██████████| 128/128 [00:00<00:00, 20083.45it/s]
`torch_dtype` is deprecated! Use `dtype` instead!

Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]
Loading checkpoint shards:  33%|███▎      | 1/3 [00:00<00:01,  1.19it/s]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:01<00:00,  1.21it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.33it/s]
Loading checkpoint shards: 100%|██████████| 3/3 [00:02<00:00,  1.30it/s]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-13 12:45:23 [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 04-13 12:45:23 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-13 12:45:23 [model.py:1510] Using max model len 1536
INFO 04-13 12:45:24 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-13 12:45:24 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-13 12:45:24 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-13 12:45:24 [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 04-13 12:45:25 [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 04-13 12:45:25 [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 04-13 12:45:25 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-13 12:45:26 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-13 12:45:26 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-13 12:45:26 [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,  2.08it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.52it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:02<00:00,  1.39it/s]

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:02<00:00,  1.46it/s]

INFO 04-13 12:45:28 [default_loader.py:267] Loading weights took 2.24 seconds
INFO 04-13 12:45:29 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.714648 seconds
INFO 04-13 12:45:34 [backends.py:548] Using cache directory: /root/.cache/vllm/torch_compile_cache/8b012be8b8/rank_0_0/backbone for vLLM's torch.compile
INFO 04-13 12:45:34 [backends.py:559] Dynamo bytecode transform time: 4.61 s
INFO 04-13 12:45:37 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.374 s
INFO 04-13 12:45:39 [monitor.py:34] torch.compile takes 4.61 s in total
INFO 04-13 12:45:40 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-13 12:45:40 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-13 12:45:40 [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):  60%|██████    | 3/5 [00:00<00:00, 21.86it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 19.85it/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, 27.20it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 27.44it/s]
INFO 04-13 12:45:41 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-13 12:45:41 [core.py:210] init engine (profile, create kv cache, warmup model) took 12.33 seconds
INFO 04-13 12:45:42 [llm.py:306] Supported_tasks: ('generate',)
2026-04-13 12:45:42,026 | 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
2026-04-13 12:45:42,027 | 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
2026-04-13 12:45:42,027 | INFO | train_grpo_train | active_rewards=['meter', 'count_adherence', 'hard_gate', 'judge_quality', 'repeat_soft', 'total_composite']
2026-04-13 12:45:42,027 | 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}
2026-04-13 12:45:42,027 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo-short1k-no-trio-v4
2026-04-13 12:45:42,027 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/cap_1000/selected_manifest.csv
2026-04-13 12:45:42,027 | 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
2026-04-13 12:45:42,027 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-13 12:45:42,027 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-13T12:45:23Z", "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"}
2026-04-13 12:45:42,027 | 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 04-13 12:45:42 [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<6:01:41,  6.58s/it]2026-04-13 12:45:48,978 | INFO | train_grpo_train | metrics_logged mode=train step=1

                                                  
{'loss': 0.1244, 'grad_norm': 20.99456787109375, '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': 0.26375001668930054, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.2045568972826004, 'rewards/total_composite/std': 0.10483554750680923, 'reward': 0.2045568972826004, 'reward_std': 0.10483554750680923, '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': 0.10917956382036209, 'clip_ratio/low_min': 0.10917956382036209, 'clip_ratio/high_mean': 0.12637650407850742, 'clip_ratio/high_max': 0.12637650407850742, 'clip_ratio/region_mean': 0.23555606789886951, 'reward_total_mean': 0.2045568972826004, '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': 0.26375001668930054, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.2045568972826004, 'reward_total_composite_std': 0.10483554750680923, 'epoch': 0.0}

  0%|          | 1/3300 [00:06<6:01:41,  6.58s/it]INFO 04-13 12:45:49 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 2/3300 [00:14<6:35:02,  7.19s/it]2026-04-13 12:45:56,591 | INFO | train_grpo_train | metrics_logged mode=train step=2

                                                  
{'loss': 0.0213, 'grad_norm': 7.792138576507568, 'learning_rate': 9.996969696969698e-06, 'num_tokens': 4134.0, 'completions/mean_length': 144.0, 'completions/min_length': 103.0, 'completions/max_length': 163.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 144.0, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 163.0, 'rewards/meter/mean': 0.6690956950187683, 'rewards/meter/std': 0.369981974363327, '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': 0.9968761205673218, 'rewards/repeat_soft/std': 0.0025736435782164335, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.14172407984733582, 'rewards/total_composite/mean': 0.24503539502620697, 'rewards/total_composite/std': 0.15705519914627075, 'reward': 0.24503539502620697, 'reward_std': 0.15705521404743195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19617469608783722, 'sampling/sampling_logp_difference/max': 2.204803943634033, 'sampling/importance_sampling_ratio/min': 0.11027213931083679, 'sampling/importance_sampling_ratio/mean': 1.0382649898529053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9801125675439835, 'clip_ratio/low_mean': 0.16077206656336784, 'clip_ratio/low_min': 0.16077206656336784, 'clip_ratio/high_mean': 0.04601428098976612, 'clip_ratio/high_max': 0.04601428098976612, 'clip_ratio/region_mean': 0.20678634755313396, 'reward_total_mean': 0.24503539502620697, 'reward_meter_mean': 0.6690956950187683, 'reward_meter_std': 0.369981974363327, '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': 0.9968761205673218, 'reward_repeat_soft_std': 0.0025736435782164335, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.14172407984733582, 'reward_total_composite_mean': 0.24503539502620697, 'reward_total_composite_std': 0.15705519914627075, 'epoch': 0.0}

  0%|          | 2/3300 [00:14<6:35:02,  7.19s/it]INFO 04-13 12:45:56 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 3/3300 [00:20<6:07:40,  6.69s/it]2026-04-13 12:46:02,699 | INFO | train_grpo_train | metrics_logged mode=train step=3

                                                  
{'loss': -0.1053, 'grad_norm': 23.21904754638672, 'learning_rate': 9.993939393939395e-06, 'num_tokens': 5844.0, 'completions/mean_length': 36.75, 'completions/min_length': 27.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6859860420227051, 'rewards/meter/std': 0.3346250057220459, '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': 0.9924981594085693, 'rewards/repeat_soft/std': 0.010377585887908936, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.14302222430706024, 'rewards/total_composite/mean': 0.23595881462097168, 'rewards/total_composite/std': 0.10526535660028458, 'reward': 0.23595881462097168, 'reward_std': 0.10526534914970398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.3009033203125, 'sampling/sampling_logp_difference/max': 4.2564239501953125, 'sampling/importance_sampling_ratio/min': 0.014172895811498165, 'sampling/importance_sampling_ratio/mean': 0.9777875542640686, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2389525398612022, 'clip_ratio/low_mean': 0.1306113675236702, 'clip_ratio/low_min': 0.1306113675236702, 'clip_ratio/high_mean': 0.10717256367206573, 'clip_ratio/high_max': 0.10717256367206573, 'clip_ratio/region_mean': 0.23778393119573593, 'reward_total_mean': 0.23595881462097168, 'reward_meter_mean': 0.6859860420227051, 'reward_meter_std': 0.3346250057220459, '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': 0.9924981594085693, 'reward_repeat_soft_std': 0.010377585887908936, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.14302222430706024, 'reward_total_composite_mean': 0.23595881462097168, 'reward_total_composite_std': 0.10526535660028458, 'epoch': 0.0}

  0%|          | 3/3300 [00:20<6:07:40,  6.69s/it]INFO 04-13 12:46:02 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 4/3300 [00:26<5:50:01,  6.37s/it]2026-04-13 12:46:08,576 | INFO | train_grpo_train | metrics_logged mode=train step=4

                                                  
{'loss': -0.0566, 'grad_norm': 16.3281192779541, 'learning_rate': 9.990909090909093e-06, 'num_tokens': 7454.0, 'completions/mean_length': 44.25, 'completions/min_length': 30.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.4112703204154968, 'rewards/meter/std': 0.37593206763267517, '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': 0.997799277305603, 'rewards/repeat_soft/std': 0.0031309700571000576, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.166770800948143, 'rewards/total_composite/mean': 0.23073405027389526, 'rewards/total_composite/std': 0.11460831761360168, 'reward': 0.23073405027389526, 'reward_std': 0.11460831761360168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20872721076011658, 'sampling/sampling_logp_difference/max': 1.4540247917175293, 'sampling/importance_sampling_ratio/min': 0.23362809419631958, 'sampling/importance_sampling_ratio/mean': 1.02910578250885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8074360936880112, 'clip_ratio/low_mean': 0.08966553024947643, 'clip_ratio/low_min': 0.08966553024947643, 'clip_ratio/high_mean': 0.1273432057350874, 'clip_ratio/high_max': 0.1273432057350874, 'clip_ratio/region_mean': 0.21700873598456383, 'reward_total_mean': 0.23073405027389526, 'reward_meter_mean': 0.4112703204154968, 'reward_meter_std': 0.37593206763267517, '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': 0.997799277305603, 'reward_repeat_soft_std': 0.0031309700571000576, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.166770800948143, 'reward_total_composite_mean': 0.23073405027389526, 'reward_total_composite_std': 0.11460831761360168, 'epoch': 0.0}

  0%|          | 4/3300 [00:26<5:50:01,  6.37s/it]INFO 04-13 12:46:08 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 5/3300 [00:33<6:08:45,  6.71s/it]2026-04-13 12:46:15,898 | INFO | train_grpo_train | metrics_logged mode=train step=5

                                                  
{'loss': 0.0491, 'grad_norm': 8.495488166809082, 'learning_rate': 9.987878787878788e-06, 'num_tokens': 9877.0, 'completions/mean_length': 125.875, 'completions/min_length': 101.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.875, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.5144270658493042, 'rewards/meter/std': 0.4267750382423401, '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': 0.9954259395599365, 'rewards/repeat_soft/std': 0.0038400308694690466, 'rewards/judge_quality/mean': 0.30250000953674316, 'rewards/judge_quality/std': 0.20954032242298126, 'rewards/total_composite/mean': 0.19073012471199036, 'rewards/total_composite/std': 0.1002202183008194, 'reward': 0.19073012471199036, 'reward_std': 0.1002202108502388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18132857978343964, 'sampling/sampling_logp_difference/max': 2.1783599853515625, 'sampling/importance_sampling_ratio/min': 0.11322707682847977, 'sampling/importance_sampling_ratio/mean': 1.0050190687179565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3385392278432846, 'clip_ratio/low_mean': 0.07284662686288357, 'clip_ratio/low_min': 0.07284662686288357, 'clip_ratio/high_mean': 0.11934021860361099, 'clip_ratio/high_max': 0.11934021860361099, 'clip_ratio/region_mean': 0.19218684546649456, 'reward_total_mean': 0.19073012471199036, 'reward_meter_mean': 0.5144270658493042, 'reward_meter_std': 0.4267750382423401, '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': 0.9954259395599365, 'reward_repeat_soft_std': 0.0038400308694690466, 'reward_judge_quality_mean': 0.30250000953674316, 'reward_judge_quality_std': 0.20954032242298126, 'reward_total_composite_mean': 0.19073012471199036, 'reward_total_composite_std': 0.1002202183008194, 'epoch': 0.0}

  0%|          | 5/3300 [00:33<6:08:45,  6.71s/it]INFO 04-13 12:46:16 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 6/3300 [00:40<6:19:25,  6.91s/it]2026-04-13 12:46:23,201 | INFO | train_grpo_train | metrics_logged mode=train step=6

                                                  
{'loss': 0.0912, 'grad_norm': 9.133028030395508, 'learning_rate': 9.984848484848485e-06, 'num_tokens': 12497.0, 'completions/mean_length': 128.5, 'completions/min_length': 84.0, 'completions/max_length': 175.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.5, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 175.0, 'rewards/meter/mean': 0.9546395540237427, 'rewards/meter/std': 0.10463159531354904, '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': 0.998668909072876, 'rewards/repeat_soft/std': 0.001581631600856781, 'rewards/judge_quality/mean': 0.23875001072883606, 'rewards/judge_quality/std': 0.027483763173222542, 'rewards/total_composite/mean': 0.22964996099472046, 'rewards/total_composite/std': 0.026852471753954887, 'reward': 0.22964996099472046, 'reward_std': 0.026852469891309738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2118406891822815, 'sampling/sampling_logp_difference/max': 1.8763961791992188, 'sampling/importance_sampling_ratio/min': 0.15314100682735443, 'sampling/importance_sampling_ratio/mean': 1.0181084871292114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.359899118542671, 'clip_ratio/low_mean': 0.13176005519926548, 'clip_ratio/low_min': 0.13176005519926548, 'clip_ratio/high_mean': 0.054821427911520004, 'clip_ratio/high_max': 0.054821427911520004, 'clip_ratio/region_mean': 0.18658148311078548, 'reward_total_mean': 0.22964996099472046, 'reward_meter_mean': 0.9546395540237427, 'reward_meter_std': 0.10463159531354904, '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': 0.998668909072876, 'reward_repeat_soft_std': 0.001581631600856781, 'reward_judge_quality_mean': 0.23875001072883606, 'reward_judge_quality_std': 0.027483763173222542, 'reward_total_composite_mean': 0.22964996099472046, 'reward_total_composite_std': 0.026852471753954887, 'epoch': 0.0}

  0%|          | 6/3300 [00:40<6:19:25,  6.91s/it]INFO 04-13 12:46:23 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 7/3300 [00:48<6:33:33,  7.17s/it]2026-04-13 12:46:31,004 | INFO | train_grpo_train | metrics_logged mode=train step=7

                                                  
{'loss': 0.0142, 'grad_norm': 11.58638858795166, 'learning_rate': 9.981818181818183e-06, 'num_tokens': 15039.0, 'completions/mean_length': 133.75, 'completions/min_length': 85.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.75, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.45653659105300903, 'rewards/meter/std': 0.416522741317749, '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': 0.9975261092185974, 'rewards/repeat_soft/std': 0.002966292668133974, 'rewards/judge_quality/mean': 0.23499999940395355, 'rewards/judge_quality/std': 0.02777460403740406, 'rewards/total_composite/mean': 0.14990538358688354, 'rewards/total_composite/std': 0.05683733522891998, 'reward': 0.14990538358688354, 'reward_std': 0.056837331503629684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21259059011936188, 'sampling/sampling_logp_difference/max': 1.5976266860961914, 'sampling/importance_sampling_ratio/min': 0.20237626135349274, 'sampling/importance_sampling_ratio/mean': 1.0176150798797607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.273042529821396, 'clip_ratio/low_mean': 0.12764834240078926, 'clip_ratio/low_min': 0.12764834240078926, 'clip_ratio/high_mean': 0.09455128386616707, 'clip_ratio/high_max': 0.09455128386616707, 'clip_ratio/region_mean': 0.22219962626695633, 'reward_total_mean': 0.14990538358688354, 'reward_meter_mean': 0.45653659105300903, 'reward_meter_std': 0.416522741317749, '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': 0.9975261092185974, 'reward_repeat_soft_std': 0.002966292668133974, 'reward_judge_quality_mean': 0.23499999940395355, 'reward_judge_quality_std': 0.02777460403740406, 'reward_total_composite_mean': 0.14990538358688354, 'reward_total_composite_std': 0.05683733522891998, 'epoch': 0.0}

  0%|          | 7/3300 [00:48<6:33:33,  7.17s/it]INFO 04-13 12:46:31 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 8/3300 [00:54<6:13:36,  6.81s/it]2026-04-13 12:46:36,965 | INFO | train_grpo_train | metrics_logged mode=train step=8

                                                  
{'loss': -0.0058, 'grad_norm': 30.844223022460938, 'learning_rate': 9.97878787878788e-06, 'num_tokens': 16480.0, 'completions/mean_length': 24.125, 'completions/min_length': 18.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.905945897102356, 'rewards/meter/std': 0.1110290139913559, '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': 0.9606921672821045, 'rewards/repeat_soft/std': 0.005113361403346062, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2624574899673462, 'rewards/total_composite/mean': 0.6936631202697754, 'rewards/total_composite/std': 0.2499092072248459, 'reward': 0.6936631202697754, 'reward_std': 0.2499092072248459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18403935432434082, 'sampling/sampling_logp_difference/max': 2.2623558044433594, 'sampling/importance_sampling_ratio/min': 0.10410494357347488, 'sampling/importance_sampling_ratio/mean': 1.0264980792999268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5599659457802773, 'clip_ratio/low_mean': 0.06472222320735455, 'clip_ratio/low_min': 0.06472222320735455, 'clip_ratio/high_mean': 0.10283747036010027, 'clip_ratio/high_max': 0.10283747036010027, 'clip_ratio/region_mean': 0.16755969356745481, 'reward_total_mean': 0.6936631202697754, 'reward_meter_mean': 0.905945897102356, 'reward_meter_std': 0.1110290139913559, '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': 0.9606921672821045, 'reward_repeat_soft_std': 0.005113361403346062, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2624574899673462, 'reward_total_composite_mean': 0.6936631202697754, 'reward_total_composite_std': 0.2499092072248459, 'epoch': 0.0}

  0%|          | 8/3300 [00:54<6:13:36,  6.81s/it]INFO 04-13 12:46:37 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 9/3300 [01:02<6:39:30,  7.28s/it]2026-04-13 12:46:45,358 | INFO | train_grpo_train | metrics_logged mode=train step=9

                                                  
{'loss': 0.1804, 'grad_norm': 7.546041488647461, 'learning_rate': 9.975757575757577e-06, 'num_tokens': 19319.0, 'completions/mean_length': 148.875, 'completions/min_length': 98.0, 'completions/max_length': 220.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 148.875, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 220.0, 'rewards/meter/mean': 0.9833512306213379, 'rewards/meter/std': 0.016487499698996544, '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': 0.9962527751922607, 'rewards/repeat_soft/std': 0.0033018917310982943, 'rewards/judge_quality/mean': 0.26750001311302185, 'rewards/judge_quality/std': 0.06840008497238159, 'rewards/total_composite/mean': 0.2634543478488922, 'rewards/total_composite/std': 0.06922971457242966, 'reward': 0.2634543478488922, 'reward_std': 0.06922971457242966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2024044394493103, 'sampling/sampling_logp_difference/max': 1.5941190719604492, 'sampling/importance_sampling_ratio/min': 0.20308735966682434, 'sampling/importance_sampling_ratio/mean': 1.0357457399368286, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.254930019378662, 'clip_ratio/low_mean': 0.08165346831083298, 'clip_ratio/low_min': 0.08165346831083298, 'clip_ratio/high_mean': 0.09651176072657108, 'clip_ratio/high_max': 0.09651176072657108, 'clip_ratio/region_mean': 0.17816522903740406, 'reward_total_mean': 0.2634543478488922, 'reward_meter_mean': 0.9833512306213379, 'reward_meter_std': 0.016487499698996544, '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': 0.9962527751922607, 'reward_repeat_soft_std': 0.0033018917310982943, 'reward_judge_quality_mean': 0.26750001311302185, 'reward_judge_quality_std': 0.06840008497238159, 'reward_total_composite_mean': 0.2634543478488922, 'reward_total_composite_std': 0.06922971457242966, 'epoch': 0.0}

  0%|          | 9/3300 [01:02<6:39:30,  7.28s/it]INFO 04-13 12:46:45 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 10/3300 [01:09<6:33:41,  7.18s/it]2026-04-13 12:46:52,336 | INFO | train_grpo_train | metrics_logged mode=train step=10

                                                   
{'loss': 0.1814, 'grad_norm': 20.30479621887207, 'learning_rate': 9.972727272727274e-06, 'num_tokens': 20869.0, 'completions/mean_length': 31.75, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.75, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7482462525367737, 'rewards/meter/std': 0.40584293007850647, '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': 0.9827924370765686, 'rewards/repeat_soft/std': 0.029328446835279465, 'rewards/judge_quality/mean': 0.2524999976158142, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.21522781252861023, 'rewards/total_composite/std': 0.10106831789016724, 'reward': 0.21522781252861023, 'reward_std': 0.10106831789016724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22601714730262756, 'sampling/sampling_logp_difference/max': 1.5818405151367188, 'sampling/importance_sampling_ratio/min': 0.20559634268283844, 'sampling/importance_sampling_ratio/mean': 1.03934907913208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5930405110120773, 'clip_ratio/low_mean': 0.08459400478750467, 'clip_ratio/low_min': 0.08459400478750467, 'clip_ratio/high_mean': 0.15337539464235306, 'clip_ratio/high_max': 0.15337539464235306, 'clip_ratio/region_mean': 0.23796939942985773, 'reward_total_mean': 0.21522781252861023, 'reward_meter_mean': 0.7482462525367737, 'reward_meter_std': 0.40584293007850647, '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': 0.9827924370765686, 'reward_repeat_soft_std': 0.029328446835279465, 'reward_judge_quality_mean': 0.2524999976158142, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.21522781252861023, 'reward_total_composite_std': 0.10106831789016724, 'epoch': 0.0}

  0%|          | 10/3300 [01:09<6:33:41,  7.18s/it]INFO 04-13 12:46:52 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 11/3300 [01:16<6:22:39,  6.98s/it]2026-04-13 12:46:58,826 | INFO | train_grpo_train | metrics_logged mode=train step=11

                                                   
{'loss': 0.177, 'grad_norm': 13.777490615844727, 'learning_rate': 9.96969696969697e-06, 'num_tokens': 22443.0, 'completions/mean_length': 46.75, 'completions/min_length': 33.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5961973667144775, 'rewards/meter/std': 0.4163079261779785, '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': 0.9969432353973389, 'rewards/repeat_soft/std': 0.004917121026664972, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.14606750011444092, 'rewards/total_composite/mean': 0.2538233697414398, 'rewards/total_composite/std': 0.17784301936626434, 'reward': 0.2538233697414398, 'reward_std': 0.17784300446510315, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2327873557806015, 'sampling/sampling_logp_difference/max': 2.384127616882324, 'sampling/importance_sampling_ratio/min': 0.09216935187578201, 'sampling/importance_sampling_ratio/mean': 1.0264379978179932, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2318453043699265, 'clip_ratio/low_mean': 0.14783670287579298, 'clip_ratio/low_min': 0.14783670287579298, 'clip_ratio/high_mean': 0.05930736102163792, 'clip_ratio/high_max': 0.05930736102163792, 'clip_ratio/region_mean': 0.2071440638974309, 'reward_total_mean': 0.2538233697414398, 'reward_meter_mean': 0.5961973667144775, 'reward_meter_std': 0.4163079261779785, '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': 0.9969432353973389, 'reward_repeat_soft_std': 0.004917121026664972, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.14606750011444092, 'reward_total_composite_mean': 0.2538233697414398, 'reward_total_composite_std': 0.17784301936626434, 'epoch': 0.0}

  0%|          | 11/3300 [01:16<6:22:39,  6.98s/it]INFO 04-13 12:46:59 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 12/3300 [01:22<6:03:57,  6.64s/it]2026-04-13 12:47:04,601 | INFO | train_grpo_train | metrics_logged mode=train step=12

                                                   
{'loss': 0.0963, 'grad_norm': 21.680850982666016, 'learning_rate': 9.966666666666667e-06, 'num_tokens': 23907.0, 'completions/mean_length': 23.0, 'completions/min_length': 15.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.46723121404647827, 'rewards/meter/std': 0.4440421462059021, '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': 0.9610613584518433, 'rewards/repeat_soft/std': 0.004068940877914429, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.2979333698749542, 'rewards/total_composite/mean': 0.4028964340686798, 'rewards/total_composite/std': 0.25024446845054626, 'reward': 0.4028964340686798, 'reward_std': 0.25024449825286865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23249541223049164, 'sampling/sampling_logp_difference/max': 1.4548969268798828, 'sampling/importance_sampling_ratio/min': 0.23342442512512207, 'sampling/importance_sampling_ratio/mean': 1.039197325706482, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.634026363492012, 'clip_ratio/low_mean': 0.12728216219693422, 'clip_ratio/low_min': 0.12728216219693422, 'clip_ratio/high_mean': 0.11459790356457233, 'clip_ratio/high_max': 0.11459790356457233, 'clip_ratio/region_mean': 0.24188006576150656, 'reward_total_mean': 0.4028964340686798, 'reward_meter_mean': 0.46723121404647827, 'reward_meter_std': 0.4440421462059021, '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': 0.9610613584518433, 'reward_repeat_soft_std': 0.004068940877914429, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.2979333698749542, 'reward_total_composite_mean': 0.4028964340686798, 'reward_total_composite_std': 0.25024446845054626, 'epoch': 0.0}

  0%|          | 12/3300 [01:22<6:03:57,  6.64s/it]INFO 04-13 12:47:04 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 13/3300 [01:27<5:49:31,  6.38s/it]2026-04-13 12:47:10,379 | INFO | train_grpo_train | metrics_logged mode=train step=13

                                                   
{'loss': 0.0782, 'grad_norm': 20.452024459838867, 'learning_rate': 9.963636363636364e-06, 'num_tokens': 25545.0, 'completions/mean_length': 47.75, 'completions/min_length': 30.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5705592632293701, 'rewards/meter/std': 0.37744197249412537, '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': 0.9955551624298096, 'rewards/repeat_soft/std': 0.0066154426895082, 'rewards/judge_quality/mean': 0.2524999976158142, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.17487716674804688, 'rewards/total_composite/std': 0.051213208585977554, 'reward': 0.17487716674804688, 'reward_std': 0.05121320113539696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22743971645832062, 'sampling/sampling_logp_difference/max': 2.7528858184814453, 'sampling/importance_sampling_ratio/min': 0.06374364346265793, 'sampling/importance_sampling_ratio/mean': 1.0200012922286987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6435254663228989, 'clip_ratio/low_mean': 0.08463804796338081, 'clip_ratio/low_min': 0.08463804796338081, 'clip_ratio/high_mean': 0.14737549424171448, 'clip_ratio/high_max': 0.14737549424171448, 'clip_ratio/region_mean': 0.2320135422050953, 'reward_total_mean': 0.17487716674804688, 'reward_meter_mean': 0.5705592632293701, 'reward_meter_std': 0.37744197249412537, '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': 0.9955551624298096, 'reward_repeat_soft_std': 0.0066154426895082, 'reward_judge_quality_mean': 0.2524999976158142, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.17487716674804688, 'reward_total_composite_std': 0.051213208585977554, 'epoch': 0.0}

  0%|          | 13/3300 [01:27<5:49:31,  6.38s/it]INFO 04-13 12:47:10 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 14/3300 [01:35<6:01:59,  6.61s/it]2026-04-13 12:47:17,519 | INFO | train_grpo_train | metrics_logged mode=train step=14

                                                   
{'loss': 0.1427, 'grad_norm': 8.394230842590332, 'learning_rate': 9.960606060606062e-06, 'num_tokens': 28140.0, 'completions/mean_length': 130.375, 'completions/min_length': 90.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.375, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.9129445552825928, 'rewards/meter/std': 0.1363978236913681, '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': 0.9947641491889954, 'rewards/repeat_soft/std': 0.008366446942090988, 'rewards/judge_quality/mean': 0.273749977350235, 'rewards/judge_quality/std': 0.09085899591445923, 'rewards/total_composite/mean': 0.2543025612831116, 'rewards/total_composite/std': 0.07832667231559753, 'reward': 0.2543025612831116, 'reward_std': 0.07832667976617813, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21353746950626373, 'sampling/sampling_logp_difference/max': 1.6659069061279297, 'sampling/importance_sampling_ratio/min': 0.18901915848255157, 'sampling/importance_sampling_ratio/mean': 1.0240132808685303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4317821264266968, 'clip_ratio/low_mean': 0.14206395111978054, 'clip_ratio/low_min': 0.14206395111978054, 'clip_ratio/high_mean': 0.05798611044883728, 'clip_ratio/high_max': 0.05798611044883728, 'clip_ratio/region_mean': 0.20005006156861782, 'reward_total_mean': 0.2543025612831116, 'reward_meter_mean': 0.9129445552825928, 'reward_meter_std': 0.1363978236913681, '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': 0.9947641491889954, 'reward_repeat_soft_std': 0.008366446942090988, 'reward_judge_quality_mean': 0.273749977350235, 'reward_judge_quality_std': 0.09085899591445923, 'reward_total_composite_mean': 0.2543025612831116, 'reward_total_composite_std': 0.07832667231559753, 'epoch': 0.0}

  0%|          | 14/3300 [01:35<6:01:59,  6.61s/it]INFO 04-13 12:47:17 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 15/3300 [01:41<5:52:11,  6.43s/it]2026-04-13 12:47:23,542 | INFO | train_grpo_train | metrics_logged mode=train step=15

                                                   
{'loss': 0.1828, 'grad_norm': 19.198348999023438, 'learning_rate': 9.957575757575757e-06, 'num_tokens': 29803.0, 'completions/mean_length': 42.875, 'completions/min_length': 30.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6790435314178467, 'rewards/meter/std': 0.25500166416168213, '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': 0.9930763244628906, 'rewards/repeat_soft/std': 0.009657313115894794, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.18715444207191467, 'rewards/total_composite/mean': 0.30946359038352966, 'rewards/total_composite/std': 0.11527153849601746, 'reward': 0.30946359038352966, 'reward_std': 0.11527153104543686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2140396237373352, 'sampling/sampling_logp_difference/max': 1.9556961059570312, 'sampling/importance_sampling_ratio/min': 0.141465961933136, 'sampling/importance_sampling_ratio/mean': 1.0327602624893188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7300214618444443, 'clip_ratio/low_mean': 0.13212848082184792, 'clip_ratio/low_min': 0.13212848082184792, 'clip_ratio/high_mean': 0.08611111342906952, 'clip_ratio/high_max': 0.08611111342906952, 'clip_ratio/region_mean': 0.21823959425091743, 'reward_total_mean': 0.30946359038352966, 'reward_meter_mean': 0.6790435314178467, 'reward_meter_std': 0.25500166416168213, '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': 0.9930763244628906, 'reward_repeat_soft_std': 0.009657313115894794, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.18715444207191467, 'reward_total_composite_mean': 0.30946359038352966, 'reward_total_composite_std': 0.11527153849601746, 'epoch': 0.0}

  0%|          | 15/3300 [01:41<5:52:11,  6.43s/it]INFO 04-13 12:47:23 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 16/3300 [01:46<5:42:38,  6.26s/it]2026-04-13 12:47:29,400 | INFO | train_grpo_train | metrics_logged mode=train step=16

                                                   
{'loss': 0.0641, 'grad_norm': 16.974000930786133, 'learning_rate': 9.954545454545456e-06, 'num_tokens': 31395.0, 'completions/mean_length': 48.0, 'completions/min_length': 42.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.4550546109676361, 'rewards/meter/std': 0.45469674468040466, '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': 0.9949187636375427, 'rewards/repeat_soft/std': 0.0085137402638793, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.2686706781387329, 'rewards/total_composite/mean': 0.25859498977661133, 'rewards/total_composite/std': 0.23468804359436035, 'reward': 0.25859498977661133, 'reward_std': 0.23468804359436035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2124183177947998, 'sampling/sampling_logp_difference/max': 1.2455048561096191, 'sampling/importance_sampling_ratio/min': 0.2877956032752991, 'sampling/importance_sampling_ratio/mean': 1.040221929550171, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.976992592215538, 'clip_ratio/low_mean': 0.1438023690134287, 'clip_ratio/low_min': 0.1438023690134287, 'clip_ratio/high_mean': 0.06051587499678135, 'clip_ratio/high_max': 0.06051587499678135, 'clip_ratio/region_mean': 0.20431824401021004, 'reward_total_mean': 0.25859498977661133, 'reward_meter_mean': 0.4550546109676361, 'reward_meter_std': 0.45469674468040466, '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': 0.9949187636375427, 'reward_repeat_soft_std': 0.0085137402638793, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.2686706781387329, 'reward_total_composite_mean': 0.25859498977661133, 'reward_total_composite_std': 0.23468804359436035, 'epoch': 0.0}

  0%|          | 16/3300 [01:47<5:42:38,  6.26s/it]INFO 04-13 12:47:29 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 17/3300 [01:52<5:33:02,  6.09s/it]2026-04-13 12:47:35,085 | INFO | train_grpo_train | metrics_logged mode=train step=17

                                                   
{'loss': 0.1538, 'grad_norm': 15.405744552612305, 'learning_rate': 9.951515151515152e-06, 'num_tokens': 33024.0, 'completions/mean_length': 44.625, 'completions/min_length': 33.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.49273881316185, 'rewards/meter/std': 0.3816235363483429, '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': 0.9926319718360901, 'rewards/repeat_soft/std': 0.012840799987316132, 'rewards/judge_quality/mean': 0.3062499761581421, 'rewards/judge_quality/std': 0.09500939399003983, 'rewards/total_composite/mean': 0.19640883803367615, 'rewards/total_composite/std': 0.07236796617507935, 'reward': 0.19640883803367615, 'reward_std': 0.07236796617507935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1962367445230484, 'sampling/sampling_logp_difference/max': 1.4637451171875, 'sampling/importance_sampling_ratio/min': 0.23136815428733826, 'sampling/importance_sampling_ratio/mean': 1.0287688970565796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7023908197879791, 'clip_ratio/low_mean': 0.09323069825768471, 'clip_ratio/low_min': 0.09323069825768471, 'clip_ratio/high_mean': 0.08432095218449831, 'clip_ratio/high_max': 0.08432095218449831, 'clip_ratio/region_mean': 0.17755165044218302, 'reward_total_mean': 0.19640883803367615, 'reward_meter_mean': 0.49273881316185, 'reward_meter_std': 0.3816235363483429, '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': 0.9926319718360901, 'reward_repeat_soft_std': 0.012840799987316132, 'reward_judge_quality_mean': 0.3062499761581421, 'reward_judge_quality_std': 0.09500939399003983, 'reward_total_composite_mean': 0.19640883803367615, 'reward_total_composite_std': 0.07236796617507935, 'epoch': 0.0}

  1%|          | 17/3300 [01:52<5:33:02,  6.09s/it]INFO 04-13 12:47:35 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 18/3300 [01:58<5:25:05,  5.94s/it]2026-04-13 12:47:40,694 | INFO | train_grpo_train | metrics_logged mode=train step=18

                                                   
{'loss': 0.1091, 'grad_norm': 19.118051528930664, 'learning_rate': 9.948484848484849e-06, 'num_tokens': 34458.0, 'completions/mean_length': 30.25, 'completions/min_length': 20.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.669553816318512, 'rewards/meter/std': 0.4602260887622833, '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': 0.27125000953674316, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.20250649750232697, 'rewards/total_composite/std': 0.06498918682336807, 'reward': 0.20250649750232697, 'reward_std': 0.06498918682336807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21079884469509125, 'sampling/sampling_logp_difference/max': 1.7254924774169922, 'sampling/importance_sampling_ratio/min': 0.1780853271484375, 'sampling/importance_sampling_ratio/mean': 1.0141559839248657, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.862677589058876, 'clip_ratio/low_mean': 0.050109206698834896, 'clip_ratio/low_min': 0.050109206698834896, 'clip_ratio/high_mean': 0.13927721232175827, 'clip_ratio/high_max': 0.13927721232175827, 'clip_ratio/region_mean': 0.18938641902059317, 'reward_total_mean': 0.20250649750232697, 'reward_meter_mean': 0.669553816318512, 'reward_meter_std': 0.4602260887622833, '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': 0.27125000953674316, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.20250649750232697, 'reward_total_composite_std': 0.06498918682336807, 'epoch': 0.0}

  1%|          | 18/3300 [01:58<5:25:05,  5.94s/it]INFO 04-13 12:47:40 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 19/3300 [02:04<5:32:29,  6.08s/it]2026-04-13 12:47:47,094 | INFO | train_grpo_train | metrics_logged mode=train step=19

                                                   
{'loss': 0.0388, 'grad_norm': 15.521472930908203, 'learning_rate': 9.945454545454546e-06, 'num_tokens': 36130.0, 'completions/mean_length': 54.0, 'completions/min_length': 29.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4761258363723755, 'rewards/meter/std': 0.4134525954723358, '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': 0.9986214637756348, 'rewards/repeat_soft/std': 0.0037080326583236456, 'rewards/judge_quality/mean': 0.3062499761581421, 'rewards/judge_quality/std': 0.09500939399003983, 'rewards/total_composite/mean': 0.18781903386116028, 'rewards/total_composite/std': 0.05385388433933258, 'reward': 0.18781903386116028, 'reward_std': 0.05385388433933258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20062369108200073, 'sampling/sampling_logp_difference/max': 2.039299964904785, 'sampling/importance_sampling_ratio/min': 0.13011977076530457, 'sampling/importance_sampling_ratio/mean': 1.008468747138977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3630243688821793, 'clip_ratio/low_mean': 0.07949496805667877, 'clip_ratio/low_min': 0.07949496805667877, 'clip_ratio/high_mean': 0.10655375942587852, 'clip_ratio/high_max': 0.10655375942587852, 'clip_ratio/region_mean': 0.1860487274825573, 'reward_total_mean': 0.18781903386116028, 'reward_meter_mean': 0.4761258363723755, 'reward_meter_std': 0.4134525954723358, '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': 0.9986214637756348, 'reward_repeat_soft_std': 0.0037080326583236456, 'reward_judge_quality_mean': 0.3062499761581421, 'reward_judge_quality_std': 0.09500939399003983, 'reward_total_composite_mean': 0.18781903386116028, 'reward_total_composite_std': 0.05385388433933258, 'epoch': 0.0}

  1%|          | 19/3300 [02:04<5:32:29,  6.08s/it]INFO 04-13 12:47:47 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 20/3300 [02:11<5:40:09,  6.22s/it]2026-04-13 12:47:53,648 | INFO | train_grpo_train | metrics_logged mode=train step=20

                                                   
{'loss': 0.0749, 'grad_norm': 10.678475379943848, 'learning_rate': 9.942424242424244e-06, 'num_tokens': 38259.0, 'completions/mean_length': 79.125, 'completions/min_length': 59.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.6724998950958252, 'rewards/meter/std': 0.3194420635700226, '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': 0.9975677728652954, 'rewards/repeat_soft/std': 0.003126713912934065, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.20345285534858704, 'rewards/total_composite/std': 0.09404149651527405, 'reward': 0.20345285534858704, 'reward_std': 0.09404149651527405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25104182958602905, 'sampling/sampling_logp_difference/max': 1.7388010025024414, 'sampling/importance_sampling_ratio/min': 0.17573098838329315, 'sampling/importance_sampling_ratio/mean': 1.0093882083892822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2550461143255234, 'clip_ratio/low_mean': 0.18275271356105804, 'clip_ratio/low_min': 0.18275271356105804, 'clip_ratio/high_mean': 0.06071428768336773, 'clip_ratio/high_max': 0.06071428768336773, 'clip_ratio/region_mean': 0.24346700124442577, 'reward_total_mean': 0.20345285534858704, 'reward_meter_mean': 0.6724998950958252, 'reward_meter_std': 0.3194420635700226, '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': 0.9975677728652954, 'reward_repeat_soft_std': 0.003126713912934065, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.20345285534858704, 'reward_total_composite_std': 0.09404149651527405, 'epoch': 0.0}

  1%|          | 20/3300 [02:11<5:40:09,  6.22s/it]INFO 04-13 12:47:53 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 21/3300 [02:17<5:34:28,  6.12s/it]2026-04-13 12:47:59,529 | INFO | train_grpo_train | metrics_logged mode=train step=21

                                                   
{'loss': 0.1033, 'grad_norm': 28.266448974609375, 'learning_rate': 9.939393939393939e-06, 'num_tokens': 39787.0, 'completions/mean_length': 32.0, 'completions/min_length': 21.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.45389506220817566, 'rewards/meter/std': 0.39092472195625305, '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': 0.9942681193351746, 'rewards/repeat_soft/std': 0.006455856841057539, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.166770800948143, 'rewards/total_composite/mean': 0.26310059428215027, 'rewards/total_composite/std': 0.18096305429935455, 'reward': 0.26310059428215027, 'reward_std': 0.18096303939819336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21730417013168335, 'sampling/sampling_logp_difference/max': 1.5662403106689453, 'sampling/importance_sampling_ratio/min': 0.20882883667945862, 'sampling/importance_sampling_ratio/mean': 1.010674238204956, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1850305646657944, 'clip_ratio/low_mean': 0.08216887293383479, 'clip_ratio/low_min': 0.08216887293383479, 'clip_ratio/high_mean': 0.07137393951416016, 'clip_ratio/high_max': 0.07137393951416016, 'clip_ratio/region_mean': 0.15354281244799495, 'reward_total_mean': 0.26310059428215027, 'reward_meter_mean': 0.45389506220817566, 'reward_meter_std': 0.39092472195625305, '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': 0.9942681193351746, 'reward_repeat_soft_std': 0.006455856841057539, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.166770800948143, 'reward_total_composite_mean': 0.26310059428215027, 'reward_total_composite_std': 0.18096305429935455, 'epoch': 0.0}

  1%|          | 21/3300 [02:17<5:34:28,  6.12s/it]INFO 04-13 12:47:59 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 22/3300 [02:23<5:32:25,  6.08s/it]2026-04-13 12:48:05,530 | INFO | train_grpo_train | metrics_logged mode=train step=22

                                                   
{'loss': 0.0779, 'grad_norm': 14.22166633605957, 'learning_rate': 9.936363636363638e-06, 'num_tokens': 41507.0, 'completions/mean_length': 56.0, 'completions/min_length': 39.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9883540868759155, 'rewards/meter/std': 0.009184768423438072, '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': 0.9959346055984497, 'rewards/repeat_soft/std': 0.005422628950327635, 'rewards/judge_quality/mean': 0.23875001072883606, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.2367658019065857, 'rewards/total_composite/std': 0.014922114089131355, 'reward': 0.2367658019065857, 'reward_std': 0.014922115951776505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20160622894763947, 'sampling/sampling_logp_difference/max': 1.526987075805664, 'sampling/importance_sampling_ratio/min': 0.26236656308174133, 'sampling/importance_sampling_ratio/mean': 1.0414557456970215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.040539041161537, 'clip_ratio/low_mean': 0.07369768805801868, 'clip_ratio/low_min': 0.07369768805801868, 'clip_ratio/high_mean': 0.11720447614789009, 'clip_ratio/high_max': 0.11720447614789009, 'clip_ratio/region_mean': 0.19090216420590878, 'reward_total_mean': 0.2367658019065857, 'reward_meter_mean': 0.9883540868759155, 'reward_meter_std': 0.009184768423438072, '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': 0.9959346055984497, 'reward_repeat_soft_std': 0.005422628950327635, 'reward_judge_quality_mean': 0.23875001072883606, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.2367658019065857, 'reward_total_composite_std': 0.014922114089131355, 'epoch': 0.0}

  1%|          | 22/3300 [02:23<5:32:25,  6.08s/it]INFO 04-13 12:48:05 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 23/3300 [02:28<5:27:36,  6.00s/it]2026-04-13 12:48:11,327 | INFO | train_grpo_train | metrics_logged mode=train step=23

                                                   
{'loss': 0.0537, 'grad_norm': 16.252086639404297, 'learning_rate': 9.933333333333334e-06, 'num_tokens': 43114.0, 'completions/mean_length': 40.875, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6953567266464233, 'rewards/meter/std': 0.3707497715950012, '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': 0.9990596175193787, 'rewards/repeat_soft/std': 0.0017459453083574772, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.2291552722454071, 'rewards/total_composite/std': 0.12831708788871765, 'reward': 0.2291552722454071, 'reward_std': 0.12831710278987885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21670785546302795, 'sampling/sampling_logp_difference/max': 1.927046775817871, 'sampling/importance_sampling_ratio/min': 0.14557749032974243, 'sampling/importance_sampling_ratio/mean': 1.029430866241455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7587376534938812, 'clip_ratio/low_mean': 0.09989950433373451, 'clip_ratio/low_min': 0.09989950433373451, 'clip_ratio/high_mean': 0.10116968676447868, 'clip_ratio/high_max': 0.10116968676447868, 'clip_ratio/region_mean': 0.2010691910982132, 'reward_total_mean': 0.2291552722454071, 'reward_meter_mean': 0.6953567266464233, 'reward_meter_std': 0.3707497715950012, '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': 0.9990596175193787, 'reward_repeat_soft_std': 0.0017459453083574772, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.2291552722454071, 'reward_total_composite_std': 0.12831708788871765, 'epoch': 0.0}

  1%|          | 23/3300 [02:28<5:27:36,  6.00s/it]INFO 04-13 12:48:11 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 24/3300 [02:35<5:38:07,  6.19s/it]2026-04-13 12:48:17,974 | INFO | train_grpo_train | metrics_logged mode=train step=24

                                                   
{'loss': 0.0181, 'grad_norm': 19.029216766357422, 'learning_rate': 9.930303030303031e-06, 'num_tokens': 44808.0, 'completions/mean_length': 59.75, 'completions/min_length': 46.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.45434120297431946, 'rewards/meter/std': 0.4283906817436218, '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': 0.9990161657333374, 'rewards/repeat_soft/std': 0.002018744358792901, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.2081165611743927, 'rewards/total_composite/mean': 0.22358007729053497, 'rewards/total_composite/std': 0.2376146912574768, 'reward': 0.22358007729053497, 'reward_std': 0.2376146912574768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.213044673204422, 'sampling/sampling_logp_difference/max': 1.6430997848510742, 'sampling/importance_sampling_ratio/min': 0.19337967038154602, 'sampling/importance_sampling_ratio/mean': 1.0326234102249146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5965741500258446, 'clip_ratio/low_mean': 0.1360377911478281, 'clip_ratio/low_min': 0.1360377911478281, 'clip_ratio/high_mean': 0.05652702413499355, 'clip_ratio/high_max': 0.05652702413499355, 'clip_ratio/region_mean': 0.19256481528282166, 'reward_total_mean': 0.22358007729053497, 'reward_meter_mean': 0.45434120297431946, 'reward_meter_std': 0.4283906817436218, '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': 0.9990161657333374, 'reward_repeat_soft_std': 0.002018744358792901, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.2081165611743927, 'reward_total_composite_mean': 0.22358007729053497, 'reward_total_composite_std': 0.2376146912574768, 'epoch': 0.0}

  1%|          | 24/3300 [02:35<5:38:07,  6.19s/it]INFO 04-13 12:48:18 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 25/3300 [02:42<5:54:55,  6.50s/it]2026-04-13 12:48:25,198 | INFO | train_grpo_train | metrics_logged mode=train step=25

                                                   
{'loss': 0.0051, 'grad_norm': 7.79266881942749, 'learning_rate': 9.927272727272728e-06, 'num_tokens': 47482.0, 'completions/mean_length': 125.25, 'completions/min_length': 110.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.25, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.38681113719940186, 'rewards/meter/std': 0.3000413477420807, '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': 0.9985337257385254, 'rewards/repeat_soft/std': 0.000780539121478796, 'rewards/judge_quality/mean': 0.3387500047683716, 'rewards/judge_quality/std': 0.20538726449012756, 'rewards/total_composite/mean': 0.20475128293037415, 'rewards/total_composite/std': 0.1320890635251999, 'reward': 0.20475128293037415, 'reward_std': 0.1320890486240387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16238392889499664, 'sampling/sampling_logp_difference/max': 1.6394009590148926, 'sampling/importance_sampling_ratio/min': 0.19409628212451935, 'sampling/importance_sampling_ratio/mean': 1.0219614505767822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1800087988376617, 'clip_ratio/low_mean': 0.09639591071754694, 'clip_ratio/low_min': 0.09639591071754694, 'clip_ratio/high_mean': 0.04710742551833391, 'clip_ratio/high_max': 0.04710742551833391, 'clip_ratio/region_mean': 0.14350333623588085, 'reward_total_mean': 0.20475128293037415, 'reward_meter_mean': 0.38681113719940186, 'reward_meter_std': 0.3000413477420807, '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': 0.9985337257385254, 'reward_repeat_soft_std': 0.000780539121478796, 'reward_judge_quality_mean': 0.3387500047683716, 'reward_judge_quality_std': 0.20538726449012756, 'reward_total_composite_mean': 0.20475128293037415, 'reward_total_composite_std': 0.1320890635251999, 'epoch': 0.0}

  1%|          | 25/3300 [02:42<5:54:55,  6.50s/it]INFO 04-13 12:48:25 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 26/3300 [02:48<5:49:06,  6.40s/it]2026-04-13 12:48:31,352 | INFO | train_grpo_train | metrics_logged mode=train step=26

                                                   
{'loss': -0.1191, 'grad_norm': 17.39537811279297, 'learning_rate': 9.924242424242425e-06, 'num_tokens': 48947.0, 'completions/mean_length': 26.125, 'completions/min_length': 17.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9929069876670837, 'rewards/meter/std': 0.011440408416092396, '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': 0.9495017528533936, 'rewards/repeat_soft/std': 0.024305421859025955, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.393950879573822, 'rewards/total_composite/std': 0.05983329936861992, 'reward': 0.393950879573822, 'reward_std': 0.05983330309391022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0914190411567688, 'sampling/sampling_logp_difference/max': 1.7267370223999023, 'sampling/importance_sampling_ratio/min': 0.17786382138729095, 'sampling/importance_sampling_ratio/mean': 1.0033172369003296, 'sampling/importance_sampling_ratio/max': 1.9017900228500366, 'entropy': 0.5844643115997314, 'clip_ratio/low_mean': 0.029411764815449715, 'clip_ratio/low_min': 0.029411764815449715, 'clip_ratio/high_mean': 0.0750319343060255, 'clip_ratio/high_max': 0.0750319343060255, 'clip_ratio/region_mean': 0.10444369912147522, 'reward_total_mean': 0.393950879573822, 'reward_meter_mean': 0.9929069876670837, 'reward_meter_std': 0.011440408416092396, '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': 0.9495017528533936, 'reward_repeat_soft_std': 0.024305421859025955, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.393950879573822, 'reward_total_composite_std': 0.05983329936861992, 'epoch': 0.0}

  1%|          | 26/3300 [02:48<5:49:06,  6.40s/it]INFO 04-13 12:48:31 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 27/3300 [02:56<6:03:42,  6.67s/it]2026-04-13 12:48:38,650 | INFO | train_grpo_train | metrics_logged mode=train step=27

                                                   
{'loss': 0.0134, 'grad_norm': 10.1174898147583, 'learning_rate': 9.921212121212121e-06, 'num_tokens': 51538.0, 'completions/mean_length': 126.875, 'completions/min_length': 96.0, 'completions/max_length': 174.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.875, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 174.0, 'rewards/meter/mean': 0.5053344368934631, 'rewards/meter/std': 0.41090095043182373, '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': 0.9992033839225769, 'rewards/repeat_soft/std': 0.0007164619164541364, 'rewards/judge_quality/mean': 0.29499998688697815, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.20223358273506165, 'rewards/total_composite/std': 0.10832107067108154, 'reward': 0.20223358273506165, 'reward_std': 0.10832107812166214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2293359339237213, 'sampling/sampling_logp_difference/max': 1.6479358673095703, 'sampling/importance_sampling_ratio/min': 0.1924467384815216, 'sampling/importance_sampling_ratio/mean': 1.0598264932632446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6311583518981934, 'clip_ratio/low_mean': 0.07886148244142532, 'clip_ratio/low_min': 0.07886148244142532, 'clip_ratio/high_mean': 0.13825739920139313, 'clip_ratio/high_max': 0.13825739920139313, 'clip_ratio/region_mean': 0.21711888164281845, 'reward_total_mean': 0.20223358273506165, 'reward_meter_mean': 0.5053344368934631, 'reward_meter_std': 0.41090095043182373, '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': 0.9992033839225769, 'reward_repeat_soft_std': 0.0007164619164541364, 'reward_judge_quality_mean': 0.29499998688697815, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.20223358273506165, 'reward_total_composite_std': 0.10832107067108154, 'epoch': 0.0}

  1%|          | 27/3300 [02:56<6:03:42,  6.67s/it]INFO 04-13 12:48:38 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 28/3300 [03:02<5:55:20,  6.52s/it]2026-04-13 12:48:44,812 | INFO | train_grpo_train | metrics_logged mode=train step=28

                                                   
{'loss': -0.0288, 'grad_norm': 17.79926872253418, 'learning_rate': 9.918181818181818e-06, 'num_tokens': 53101.0, 'completions/mean_length': 33.375, 'completions/min_length': 18.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.759634256362915, 'rewards/meter/std': 0.4249105453491211, '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': 0.94775390625, 'rewards/repeat_soft/std': 0.041864264756441116, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.21796134114265442, 'rewards/total_composite/mean': 0.2884432077407837, 'rewards/total_composite/std': 0.24120306968688965, 'reward': 0.2884432077407837, 'reward_std': 0.24120306968688965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1766754537820816, 'sampling/sampling_logp_difference/max': 1.2298507690429688, 'sampling/importance_sampling_ratio/min': 0.29233619570732117, 'sampling/importance_sampling_ratio/mean': 1.024755835533142, 'sampling/importance_sampling_ratio/max': 1.9423034191131592, 'entropy': 1.6584673970937729, 'clip_ratio/low_mean': 0.14047385100275278, 'clip_ratio/low_min': 0.14047385100275278, 'clip_ratio/high_mean': 0.03227124270051718, 'clip_ratio/high_max': 0.03227124270051718, 'clip_ratio/region_mean': 0.17274509370326996, 'reward_total_mean': 0.2884432077407837, 'reward_meter_mean': 0.759634256362915, 'reward_meter_std': 0.4249105453491211, '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': 0.94775390625, 'reward_repeat_soft_std': 0.041864264756441116, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.21796134114265442, 'reward_total_composite_mean': 0.2884432077407837, 'reward_total_composite_std': 0.24120306968688965, 'epoch': 0.0}

  1%|          | 28/3300 [03:02<5:55:20,  6.52s/it]INFO 04-13 12:48:45 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 29/3300 [03:10<6:15:47,  6.89s/it]2026-04-13 12:48:52,585 | INFO | train_grpo_train | metrics_logged mode=train step=29

                                                   
{'loss': 0.0726, 'grad_norm': 7.879302978515625, 'learning_rate': 9.915151515151515e-06, 'num_tokens': 56052.0, 'completions/mean_length': 168.875, 'completions/min_length': 140.0, 'completions/max_length': 195.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.875, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 195.0, 'rewards/meter/mean': 0.6800090074539185, 'rewards/meter/std': 0.3618488311767578, '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': 0.9975203275680542, 'rewards/repeat_soft/std': 0.0027590380050241947, 'rewards/judge_quality/mean': 0.2150000035762787, 'rewards/judge_quality/std': 0.014142133295536041, 'rewards/total_composite/mean': 0.15685734152793884, 'rewards/total_composite/std': 0.08051047474145889, 'reward': 0.15685734152793884, 'reward_std': 0.08051047474145889, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2022891640663147, 'sampling/sampling_logp_difference/max': 1.8225994110107422, 'sampling/importance_sampling_ratio/min': 0.17849254608154297, 'sampling/importance_sampling_ratio/mean': 1.0238168239593506, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2702658623456955, 'clip_ratio/low_mean': 0.06336063705384731, 'clip_ratio/low_min': 0.06336063705384731, 'clip_ratio/high_mean': 0.13224178925156593, 'clip_ratio/high_max': 0.13224178925156593, 'clip_ratio/region_mean': 0.19560242630541325, 'reward_total_mean': 0.15685734152793884, 'reward_meter_mean': 0.6800090074539185, 'reward_meter_std': 0.3618488311767578, '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': 0.9975203275680542, 'reward_repeat_soft_std': 0.0027590380050241947, 'reward_judge_quality_mean': 0.2150000035762787, 'reward_judge_quality_std': 0.014142133295536041, 'reward_total_composite_mean': 0.15685734152793884, 'reward_total_composite_std': 0.08051047474145889, 'epoch': 0.0}

  1%|          | 29/3300 [03:10<6:15:47,  6.89s/it]INFO 04-13 12:48:52 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 30/3300 [03:16<6:02:24,  6.65s/it]2026-04-13 12:48:58,668 | INFO | train_grpo_train | metrics_logged mode=train step=30

                                                   
{'loss': -0.0567, 'grad_norm': 17.424236297607422, 'learning_rate': 9.912121212121213e-06, 'num_tokens': 57912.0, 'completions/mean_length': 48.5, 'completions/min_length': 34.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6463769674301147, 'rewards/meter/std': 0.40498650074005127, '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': 0.994937539100647, 'rewards/repeat_soft/std': 0.006559374276548624, 'rewards/judge_quality/mean': 0.26999998092651367, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/total_composite/mean': 0.19401989877223969, 'rewards/total_composite/std': 0.06956040859222412, 'reward': 0.19401989877223969, 'reward_std': 0.06956040114164352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22575657069683075, 'sampling/sampling_logp_difference/max': 1.6284923553466797, 'sampling/importance_sampling_ratio/min': 0.19622518122196198, 'sampling/importance_sampling_ratio/mean': 1.0097166299819946, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3254844695329666, 'clip_ratio/low_mean': 0.060775401070714, 'clip_ratio/low_min': 0.060775401070714, 'clip_ratio/high_mean': 0.14530975557863712, 'clip_ratio/high_max': 0.14530975557863712, 'clip_ratio/region_mean': 0.20608515664935112, 'reward_total_mean': 0.19401989877223969, 'reward_meter_mean': 0.6463769674301147, 'reward_meter_std': 0.40498650074005127, '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': 0.994937539100647, 'reward_repeat_soft_std': 0.006559374276548624, 'reward_judge_quality_mean': 0.26999998092651367, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.19401989877223969, 'reward_total_composite_std': 0.06956040859222412, 'epoch': 0.0}

  1%|          | 30/3300 [03:16<6:02:24,  6.65s/it]INFO 04-13 12:48:58 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 31/3300 [03:22<5:55:59,  6.53s/it]2026-04-13 12:49:04,931 | INFO | train_grpo_train | metrics_logged mode=train step=31

                                                   
{'loss': 0.1641, 'grad_norm': 16.771867752075195, 'learning_rate': 9.90909090909091e-06, 'num_tokens': 59861.0, 'completions/mean_length': 71.625, 'completions/min_length': 47.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.23064470291137695, 'rewards/meter/std': 0.14039665460586548, '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': 0.9896365404129028, 'rewards/repeat_soft/std': 0.011847006157040596, 'rewards/judge_quality/mean': 0.3449999988079071, 'rewards/judge_quality/std': 0.14880476891994476, 'rewards/total_composite/mean': 0.172789067029953, 'rewards/total_composite/std': 0.09928283095359802, 'reward': 0.172789067029953, 'reward_std': 0.09928283095359802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2790986895561218, 'sampling/sampling_logp_difference/max': 3.0018956661224365, 'sampling/importance_sampling_ratio/min': 0.0874783918261528, 'sampling/importance_sampling_ratio/mean': 1.0112502574920654, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4137107655405998, 'clip_ratio/low_mean': 0.1759606134146452, 'clip_ratio/low_min': 0.1759606134146452, 'clip_ratio/high_mean': 0.06050531938672066, 'clip_ratio/high_max': 0.06050531938672066, 'clip_ratio/region_mean': 0.23646593280136585, 'reward_total_mean': 0.172789067029953, 'reward_meter_mean': 0.23064470291137695, 'reward_meter_std': 0.14039665460586548, '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': 0.9896365404129028, 'reward_repeat_soft_std': 0.011847006157040596, 'reward_judge_quality_mean': 0.3449999988079071, 'reward_judge_quality_std': 0.14880476891994476, 'reward_total_composite_mean': 0.172789067029953, 'reward_total_composite_std': 0.09928283095359802, 'epoch': 0.0}

  1%|          | 31/3300 [03:22<5:55:59,  6.53s/it]INFO 04-13 12:49:05 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 32/3300 [03:28<5:38:50,  6.22s/it]2026-04-13 12:49:10,422 | INFO | train_grpo_train | metrics_logged mode=train step=32

                                                   
{'loss': 0.0739, 'grad_norm': 26.786174774169922, 'learning_rate': 9.906060606060607e-06, 'num_tokens': 61299.0, 'completions/mean_length': 24.75, 'completions/min_length': 16.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.29178377985954285, 'rewards/meter/std': 0.3586907386779785, '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': 0.961837112903595, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.19486258924007416, 'rewards/total_composite/mean': 0.23482491075992584, 'rewards/total_composite/std': 0.2344757467508316, 'reward': 0.23482491075992584, 'reward_std': 0.2344757318496704, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23777849972248077, 'sampling/sampling_logp_difference/max': 2.170952320098877, 'sampling/importance_sampling_ratio/min': 0.11406894028186798, 'sampling/importance_sampling_ratio/mean': 1.0599104166030884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4571485072374344, 'clip_ratio/low_mean': 0.18879137095063925, 'clip_ratio/low_min': 0.18879137095063925, 'clip_ratio/high_mean': 0.04714912362396717, 'clip_ratio/high_max': 0.04714912362396717, 'clip_ratio/region_mean': 0.23594049457460642, 'reward_total_mean': 0.23482491075992584, 'reward_meter_mean': 0.29178377985954285, 'reward_meter_std': 0.3586907386779785, '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': 0.961837112903595, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.19486258924007416, 'reward_total_composite_mean': 0.23482491075992584, 'reward_total_composite_std': 0.2344757467508316, 'epoch': 0.0}

  1%|          | 32/3300 [03:28<5:38:50,  6.22s/it]INFO 04-13 12:49:10 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 33/3300 [03:34<5:37:50,  6.20s/it]2026-04-13 12:49:16,588 | INFO | train_grpo_train | metrics_logged mode=train step=33

                                                   
{'loss': 0.036, 'grad_norm': 17.3479061126709, 'learning_rate': 9.903030303030305e-06, 'num_tokens': 62947.0, 'completions/mean_length': 42.0, 'completions/min_length': 34.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.6314820051193237, 'rewards/meter/std': 0.4792705774307251, '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': 0.9985815286636353, 'rewards/repeat_soft/std': 0.003543524071574211, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.20051582157611847, 'rewards/total_composite/std': 0.11133254319429398, 'reward': 0.20051582157611847, 'reward_std': 0.11133254319429398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20595410466194153, 'sampling/sampling_logp_difference/max': 1.4103212356567383, 'sampling/importance_sampling_ratio/min': 0.24406488239765167, 'sampling/importance_sampling_ratio/mean': 1.0241533517837524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.233390688896179, 'clip_ratio/low_mean': 0.046910145319998264, 'clip_ratio/low_min': 0.046910145319998264, 'clip_ratio/high_mean': 0.13978325575590134, 'clip_ratio/high_max': 0.13978325575590134, 'clip_ratio/region_mean': 0.1866934010758996, 'reward_total_mean': 0.20051582157611847, 'reward_meter_mean': 0.6314820051193237, 'reward_meter_std': 0.4792705774307251, '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': 0.9985815286636353, 'reward_repeat_soft_std': 0.003543524071574211, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.20051582157611847, 'reward_total_composite_std': 0.11133254319429398, 'epoch': 0.0}

  1%|          | 33/3300 [03:34<5:37:50,  6.20s/it]INFO 04-13 12:49:16 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 34/3300 [03:40<5:37:32,  6.20s/it]2026-04-13 12:49:22,781 | INFO | train_grpo_train | metrics_logged mode=train step=34

                                                   
{'loss': 0.0314, 'grad_norm': 33.14555740356445, 'learning_rate': 9.9e-06, 'num_tokens': 64491.0, 'completions/mean_length': 35.0, 'completions/min_length': 28.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.28801655769348145, 'rewards/meter/std': 0.3693912923336029, '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': 0.9700682163238525, 'rewards/repeat_soft/std': 0.0731503814458847, 'rewards/judge_quality/mean': 0.2849999964237213, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.14035621285438538, 'rewards/total_composite/std': 0.055986013263463974, 'reward': 0.14035621285438538, 'reward_std': 0.055986013263463974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23407866060733795, 'sampling/sampling_logp_difference/max': 1.9833375215530396, 'sampling/importance_sampling_ratio/min': 0.13760919868946075, 'sampling/importance_sampling_ratio/mean': 1.0082929134368896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.644989162683487, 'clip_ratio/low_mean': 0.0862089991569519, 'clip_ratio/low_min': 0.0862089991569519, 'clip_ratio/high_mean': 0.14916404150426388, 'clip_ratio/high_max': 0.14916404150426388, 'clip_ratio/region_mean': 0.23537304066121578, 'reward_total_mean': 0.14035621285438538, 'reward_meter_mean': 0.28801655769348145, 'reward_meter_std': 0.3693912923336029, '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': 0.9700682163238525, 'reward_repeat_soft_std': 0.0731503814458847, 'reward_judge_quality_mean': 0.2849999964237213, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.14035621285438538, 'reward_total_composite_std': 0.055986013263463974, 'epoch': 0.0}

  1%|          | 34/3300 [03:40<5:37:32,  6.20s/it]INFO 04-13 12:49:23 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 35/3300 [03:46<5:31:07,  6.09s/it]2026-04-13 12:49:28,594 | INFO | train_grpo_train | metrics_logged mode=train step=35

                                                   
{'loss': 0.0637, 'grad_norm': 19.889467239379883, 'learning_rate': 9.896969696969699e-06, 'num_tokens': 66075.0, 'completions/mean_length': 33.0, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.4673800468444824, 'rewards/meter/std': 0.4234526753425598, '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': 0.9926177263259888, 'rewards/repeat_soft/std': 0.006717136595398188, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.18368743360042572, 'rewards/total_composite/mean': 0.28530552983283997, 'rewards/total_composite/std': 0.2261982560157776, 'reward': 0.28530552983283997, 'reward_std': 0.2261982560157776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20924991369247437, 'sampling/sampling_logp_difference/max': 3.056974411010742, 'sampling/importance_sampling_ratio/min': 0.0470297709107399, 'sampling/importance_sampling_ratio/mean': 0.999767541885376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8244521431624889, 'clip_ratio/low_mean': 0.1276165684685111, 'clip_ratio/low_min': 0.1276165684685111, 'clip_ratio/high_mean': 0.02765804622322321, 'clip_ratio/high_max': 0.02765804622322321, 'clip_ratio/region_mean': 0.15527461469173431, 'reward_total_mean': 0.28530552983283997, 'reward_meter_mean': 0.4673800468444824, 'reward_meter_std': 0.4234526753425598, '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': 0.9926177263259888, 'reward_repeat_soft_std': 0.006717136595398188, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.18368743360042572, 'reward_total_composite_mean': 0.28530552983283997, 'reward_total_composite_std': 0.2261982560157776, 'epoch': 0.0}

  1%|          | 35/3300 [03:46<5:31:07,  6.09s/it]INFO 04-13 12:49:28 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 36/3300 [03:52<5:34:18,  6.15s/it]2026-04-13 12:49:34,881 | INFO | train_grpo_train | metrics_logged mode=train step=36

                                                   
{'loss': 0.0973, 'grad_norm': 16.561141967773438, 'learning_rate': 9.893939393939395e-06, 'num_tokens': 67614.0, 'completions/mean_length': 43.375, 'completions/min_length': 31.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4614804983139038, 'rewards/meter/std': 0.4654310643672943, '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': 0.9922767877578735, 'rewards/repeat_soft/std': 0.0168900229036808, 'rewards/judge_quality/mean': 0.38124996423721313, 'rewards/judge_quality/std': 0.12799972295761108, 'rewards/total_composite/mean': 0.2614460587501526, 'rewards/total_composite/std': 0.18071474134922028, 'reward': 0.2614460587501526, 'reward_std': 0.18071474134922028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1762866973876953, 'sampling/sampling_logp_difference/max': 2.2372875213623047, 'sampling/importance_sampling_ratio/min': 0.10674765706062317, 'sampling/importance_sampling_ratio/mean': 1.0038046836853027, 'sampling/importance_sampling_ratio/max': 1.855939269065857, 'entropy': 1.5117772668600082, 'clip_ratio/low_mean': 0.08194054756313562, 'clip_ratio/low_min': 0.08194054756313562, 'clip_ratio/high_mean': 0.051636514253914356, 'clip_ratio/high_max': 0.051636514253914356, 'clip_ratio/region_mean': 0.13357706181704998, 'reward_total_mean': 0.2614460587501526, 'reward_meter_mean': 0.4614804983139038, 'reward_meter_std': 0.4654310643672943, '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': 0.9922767877578735, 'reward_repeat_soft_std': 0.0168900229036808, 'reward_judge_quality_mean': 0.38124996423721313, 'reward_judge_quality_std': 0.12799972295761108, 'reward_total_composite_mean': 0.2614460587501526, 'reward_total_composite_std': 0.18071474134922028, 'epoch': 0.0}

  1%|          | 36/3300 [03:52<5:34:18,  6.15s/it]INFO 04-13 12:49:35 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 37/3300 [03:58<5:32:23,  6.11s/it]2026-04-13 12:49:40,916 | INFO | train_grpo_train | metrics_logged mode=train step=37

                                                   
{'loss': 0.0809, 'grad_norm': 18.028430938720703, 'learning_rate': 9.890909090909092e-06, 'num_tokens': 69309.0, 'completions/mean_length': 40.875, 'completions/min_length': 34.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.39811915159225464, 'rewards/meter/std': 0.47292259335517883, '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': 0.995173454284668, 'rewards/repeat_soft/std': 0.01254533976316452, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.14606750011444092, 'rewards/total_composite/mean': 0.19178056716918945, 'rewards/total_composite/std': 0.11140396445989609, 'reward': 0.19178056716918945, 'reward_std': 0.11140395700931549, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24725963175296783, 'sampling/sampling_logp_difference/max': 1.543391227722168, 'sampling/importance_sampling_ratio/min': 0.21365532279014587, 'sampling/importance_sampling_ratio/mean': 1.062158226966858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5865648090839386, 'clip_ratio/low_mean': 0.14958483539521694, 'clip_ratio/low_min': 0.14958483539521694, 'clip_ratio/high_mean': 0.1176363117992878, 'clip_ratio/high_max': 0.1176363117992878, 'clip_ratio/region_mean': 0.26722114719450474, 'reward_total_mean': 0.19178056716918945, 'reward_meter_mean': 0.39811915159225464, 'reward_meter_std': 0.47292259335517883, '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': 0.995173454284668, 'reward_repeat_soft_std': 0.01254533976316452, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.14606750011444092, 'reward_total_composite_mean': 0.19178056716918945, 'reward_total_composite_std': 0.11140396445989609, 'epoch': 0.0}

  1%|          | 37/3300 [03:58<5:32:23,  6.11s/it]INFO 04-13 12:49:41 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 38/3300 [04:04<5:29:39,  6.06s/it]2026-04-13 12:49:46,867 | INFO | train_grpo_train | metrics_logged mode=train step=38

                                                   
{'loss': -0.0289, 'grad_norm': 11.82272720336914, 'learning_rate': 9.887878787878789e-06, 'num_tokens': 70984.0, 'completions/mean_length': 56.375, 'completions/min_length': 40.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.940577507019043, 'rewards/meter/std': 0.0998968780040741, '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': 0.9988288879394531, 'rewards/repeat_soft/std': 0.0020019880030304193, 'rewards/judge_quality/mean': 0.32999998331069946, 'rewards/judge_quality/std': 0.1928730010986328, 'rewards/total_composite/mean': 0.3193818926811218, 'rewards/total_composite/std': 0.1936800330877304, 'reward': 0.3193818926811218, 'reward_std': 0.1936800181865692, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20165720582008362, 'sampling/sampling_logp_difference/max': 2.2060861587524414, 'sampling/importance_sampling_ratio/min': 0.11013083904981613, 'sampling/importance_sampling_ratio/mean': 1.0116405487060547, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7450886070728302, 'clip_ratio/low_mean': 0.1500273859128356, 'clip_ratio/low_min': 0.1500273859128356, 'clip_ratio/high_mean': 0.032031250186264515, 'clip_ratio/high_max': 0.032031250186264515, 'clip_ratio/region_mean': 0.1820586360991001, 'reward_total_mean': 0.3193818926811218, 'reward_meter_mean': 0.940577507019043, 'reward_meter_std': 0.0998968780040741, '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': 0.9988288879394531, 'reward_repeat_soft_std': 0.0020019880030304193, 'reward_judge_quality_mean': 0.32999998331069946, 'reward_judge_quality_std': 0.1928730010986328, 'reward_total_composite_mean': 0.3193818926811218, 'reward_total_composite_std': 0.1936800330877304, 'epoch': 0.0}

  1%|          | 38/3300 [04:04<5:29:39,  6.06s/it]INFO 04-13 12:49:47 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 39/3300 [04:10<5:30:12,  6.08s/it]2026-04-13 12:49:52,969 | INFO | train_grpo_train | metrics_logged mode=train step=39

                                                   
{'loss': 0.0854, 'grad_norm': 21.60627555847168, 'learning_rate': 9.884848484848486e-06, 'num_tokens': 72664.0, 'completions/mean_length': 52.0, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8593474626541138, 'rewards/meter/std': 0.059598565101623535, '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': 0.9931809902191162, 'rewards/repeat_soft/std': 0.006509596481919289, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.1340509057044983, 'rewards/total_composite/mean': 0.4315245747566223, 'rewards/total_composite/std': 0.1271274983882904, 'reward': 0.4315245747566223, 'reward_std': 0.1271274983882904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16323387622833252, 'sampling/sampling_logp_difference/max': 3.1816110610961914, 'sampling/importance_sampling_ratio/min': 0.04151871055364609, 'sampling/importance_sampling_ratio/mean': 0.991568922996521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.591550350189209, 'clip_ratio/low_mean': 0.09737164154648781, 'clip_ratio/low_min': 0.09737164154648781, 'clip_ratio/high_mean': 0.025629001669585705, 'clip_ratio/high_max': 0.025629001669585705, 'clip_ratio/region_mean': 0.12300064321607351, 'reward_total_mean': 0.4315245747566223, 'reward_meter_mean': 0.8593474626541138, 'reward_meter_std': 0.059598565101623535, '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': 0.9931809902191162, 'reward_repeat_soft_std': 0.006509596481919289, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.1340509057044983, 'reward_total_composite_mean': 0.4315245747566223, 'reward_total_composite_std': 0.1271274983882904, 'epoch': 0.0}

  1%|          | 39/3300 [04:10<5:30:12,  6.08s/it]INFO 04-13 12:49:53 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 40/3300 [04:16<5:30:48,  6.09s/it]2026-04-13 12:49:59,089 | INFO | train_grpo_train | metrics_logged mode=train step=40

                                                   
{'loss': 0.0882, 'grad_norm': 16.42487144470215, 'learning_rate': 9.881818181818182e-06, 'num_tokens': 74308.0, 'completions/mean_length': 42.5, 'completions/min_length': 30.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5914168357849121, 'rewards/meter/std': 0.41509711742401123, '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': 0.9794532060623169, 'rewards/repeat_soft/std': 0.03394440561532974, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.19820624589920044, 'rewards/total_composite/mean': 0.4732854962348938, 'rewards/total_composite/std': 0.3093723654747009, 'reward': 0.4732854962348938, 'reward_std': 0.3093723952770233, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15868167579174042, 'sampling/sampling_logp_difference/max': 1.9446430206298828, 'sampling/importance_sampling_ratio/min': 0.14303827285766602, 'sampling/importance_sampling_ratio/mean': 1.0041155815124512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0093816071748734, 'clip_ratio/low_mean': 0.07824461627751589, 'clip_ratio/low_min': 0.07824461627751589, 'clip_ratio/high_mean': 0.04752552509307861, 'clip_ratio/high_max': 0.04752552509307861, 'clip_ratio/region_mean': 0.1257701413705945, 'reward_total_mean': 0.4732854962348938, 'reward_meter_mean': 0.5914168357849121, 'reward_meter_std': 0.41509711742401123, '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': 0.9794532060623169, 'reward_repeat_soft_std': 0.03394440561532974, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.19820624589920044, 'reward_total_composite_mean': 0.4732854962348938, 'reward_total_composite_std': 0.3093723654747009, 'epoch': 0.0}

  1%|          | 40/3300 [04:16<5:30:48,  6.09s/it]INFO 04-13 12:49:59 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 41/3300 [04:22<5:28:52,  6.05s/it]2026-04-13 12:50:05,064 | INFO | train_grpo_train | metrics_logged mode=train step=41

                                                   
{'loss': 0.1522, 'grad_norm': 14.568224906921387, 'learning_rate': 9.87878787878788e-06, 'num_tokens': 75939.0, 'completions/mean_length': 43.875, 'completions/min_length': 30.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6071714162826538, 'rewards/meter/std': 0.39920130372047424, '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': 0.9951162338256836, 'rewards/repeat_soft/std': 0.00768527016043663, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.21553505957126617, 'rewards/total_composite/mean': 0.2829132378101349, 'rewards/total_composite/std': 0.15281237661838531, 'reward': 0.2829132378101349, 'reward_std': 0.15281237661838531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2150118201971054, 'sampling/sampling_logp_difference/max': 1.5621118545532227, 'sampling/importance_sampling_ratio/min': 0.20969274640083313, 'sampling/importance_sampling_ratio/mean': 1.028202772140503, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5838479697704315, 'clip_ratio/low_mean': 0.10096247587352991, 'clip_ratio/low_min': 0.10096247587352991, 'clip_ratio/high_mean': 0.07940251752734184, 'clip_ratio/high_max': 0.07940251752734184, 'clip_ratio/region_mean': 0.18036499340087175, 'reward_total_mean': 0.2829132378101349, 'reward_meter_mean': 0.6071714162826538, 'reward_meter_std': 0.39920130372047424, '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': 0.9951162338256836, 'reward_repeat_soft_std': 0.00768527016043663, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.21553505957126617, 'reward_total_composite_mean': 0.2829132378101349, 'reward_total_composite_std': 0.15281237661838531, 'epoch': 0.0}

  1%|          | 41/3300 [04:22<5:28:52,  6.05s/it]INFO 04-13 12:50:05 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 42/3300 [04:28<5:27:49,  6.04s/it]2026-04-13 12:50:11,062 | INFO | train_grpo_train | metrics_logged mode=train step=42

                                                   
{'loss': -0.0594, 'grad_norm': 18.592884063720703, 'learning_rate': 9.875757575757576e-06, 'num_tokens': 77446.0, 'completions/mean_length': 39.375, 'completions/min_length': 29.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7009751200675964, 'rewards/meter/std': 0.37204018235206604, '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': 0.9971165657043457, 'rewards/repeat_soft/std': 0.003871307708323002, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.2537715435028076, 'rewards/total_composite/mean': 0.2689000368118286, 'rewards/total_composite/std': 0.2827819287776947, 'reward': 0.2689000368118286, 'reward_std': 0.2827818989753723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2534762918949127, 'sampling/sampling_logp_difference/max': 2.5374059677124023, 'sampling/importance_sampling_ratio/min': 0.07907124608755112, 'sampling/importance_sampling_ratio/mean': 1.0052545070648193, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2823383659124374, 'clip_ratio/low_mean': 0.19048691913485527, 'clip_ratio/low_min': 0.19048691913485527, 'clip_ratio/high_mean': 0.021276595070958138, 'clip_ratio/high_max': 0.021276595070958138, 'clip_ratio/region_mean': 0.2117635142058134, 'reward_total_mean': 0.2689000368118286, 'reward_meter_mean': 0.7009751200675964, 'reward_meter_std': 0.37204018235206604, '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': 0.9971165657043457, 'reward_repeat_soft_std': 0.003871307708323002, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.2537715435028076, 'reward_total_composite_mean': 0.2689000368118286, 'reward_total_composite_std': 0.2827819287776947, 'epoch': 0.0}

  1%|▏         | 42/3300 [04:28<5:27:49,  6.04s/it]INFO 04-13 12:50:11 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 43/3300 [04:35<5:39:02,  6.25s/it]2026-04-13 12:50:17,794 | INFO | train_grpo_train | metrics_logged mode=train step=43

                                                   
{'loss': 0.0871, 'grad_norm': 14.111536026000977, 'learning_rate': 9.872727272727274e-06, 'num_tokens': 79334.0, 'completions/mean_length': 61.0, 'completions/min_length': 28.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.6533817052841187, 'rewards/meter/std': 0.4500085711479187, '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': 0.9895681142807007, 'rewards/repeat_soft/std': 0.012946262955665588, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.19099363684654236, 'rewards/total_composite/mean': 0.3281722664833069, 'rewards/total_composite/std': 0.24415451288223267, 'reward': 0.3281722664833069, 'reward_std': 0.24415451288223267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19619974493980408, 'sampling/sampling_logp_difference/max': 1.7581138610839844, 'sampling/importance_sampling_ratio/min': 0.17236967384815216, 'sampling/importance_sampling_ratio/mean': 1.019789457321167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8827890306711197, 'clip_ratio/low_mean': 0.1028458047658205, 'clip_ratio/low_min': 0.1028458047658205, 'clip_ratio/high_mean': 0.08642019052058458, 'clip_ratio/high_max': 0.08642019052058458, 'clip_ratio/region_mean': 0.1892659952864051, 'reward_total_mean': 0.3281722664833069, 'reward_meter_mean': 0.6533817052841187, 'reward_meter_std': 0.4500085711479187, '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': 0.9895681142807007, 'reward_repeat_soft_std': 0.012946262955665588, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.19099363684654236, 'reward_total_composite_mean': 0.3281722664833069, 'reward_total_composite_std': 0.24415451288223267, 'epoch': 0.0}

  1%|▏         | 43/3300 [04:35<5:39:02,  6.25s/it]INFO 04-13 12:50:18 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 44/3300 [04:42<5:53:04,  6.51s/it]2026-04-13 12:50:24,908 | INFO | train_grpo_train | metrics_logged mode=train step=44

                                                   
{'loss': 0.0959, 'grad_norm': 11.349555015563965, 'learning_rate': 9.869696969696971e-06, 'num_tokens': 81552.0, 'completions/mean_length': 97.25, 'completions/min_length': 65.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.34926965832710266, 'rewards/meter/std': 0.3050661087036133, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984719753265381, 'rewards/repeat_soft/std': 0.001414117286913097, 'rewards/judge_quality/mean': 0.2487499862909317, 'rewards/judge_quality/std': 0.0699872374534607, 'rewards/total_composite/mean': 0.14352548122406006, 'rewards/total_composite/std': 0.08693286031484604, 'reward': 0.14352548122406006, 'reward_std': 0.08693286031484604, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2462064027786255, 'sampling/sampling_logp_difference/max': 1.6767797470092773, 'sampling/importance_sampling_ratio/min': 0.18697510659694672, 'sampling/importance_sampling_ratio/mean': 1.0362701416015625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6043899059295654, 'clip_ratio/low_mean': 0.15450546145439148, 'clip_ratio/low_min': 0.15450546145439148, 'clip_ratio/high_mean': 0.05760752968490124, 'clip_ratio/high_max': 0.05760752968490124, 'clip_ratio/region_mean': 0.21211299113929272, 'reward_total_mean': 0.14352548122406006, 'reward_meter_mean': 0.34926965832710266, 'reward_meter_std': 0.3050661087036133, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.18898223340511322, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984719753265381, 'reward_repeat_soft_std': 0.001414117286913097, 'reward_judge_quality_mean': 0.2487499862909317, 'reward_judge_quality_std': 0.0699872374534607, 'reward_total_composite_mean': 0.14352548122406006, 'reward_total_composite_std': 0.08693286031484604, 'epoch': 0.0}

  1%|▏         | 44/3300 [04:42<5:53:04,  6.51s/it]INFO 04-13 12:50:25 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 45/3300 [04:48<5:45:51,  6.38s/it]2026-04-13 12:50:30,978 | INFO | train_grpo_train | metrics_logged mode=train step=45

                                                   
{'loss': 0.0246, 'grad_norm': 19.27265167236328, 'learning_rate': 9.866666666666668e-06, 'num_tokens': 83060.0, 'completions/mean_length': 37.5, 'completions/min_length': 31.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8940512537956238, 'rewards/meter/std': 0.2674654424190521, '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': 0.9911835193634033, 'rewards/repeat_soft/std': 0.010344765149056911, 'rewards/judge_quality/mean': 0.3737500011920929, 'rewards/judge_quality/std': 0.13721074163913727, 'rewards/total_composite/mean': 0.3571849465370178, 'rewards/total_composite/std': 0.1577400416135788, 'reward': 0.3571849465370178, 'reward_std': 0.1577400267124176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2245570421218872, 'sampling/sampling_logp_difference/max': 1.398192286491394, 'sampling/importance_sampling_ratio/min': 0.24704314768314362, 'sampling/importance_sampling_ratio/mean': 1.0358315706253052, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0312478840351105, 'clip_ratio/low_mean': 0.09685325808823109, 'clip_ratio/low_min': 0.09685325808823109, 'clip_ratio/high_mean': 0.1374898562207818, 'clip_ratio/high_max': 0.1374898562207818, 'clip_ratio/region_mean': 0.2343431143090129, 'reward_total_mean': 0.3571849465370178, 'reward_meter_mean': 0.8940512537956238, 'reward_meter_std': 0.2674654424190521, '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': 0.9911835193634033, 'reward_repeat_soft_std': 0.010344765149056911, 'reward_judge_quality_mean': 0.3737500011920929, 'reward_judge_quality_std': 0.13721074163913727, 'reward_total_composite_mean': 0.3571849465370178, 'reward_total_composite_std': 0.1577400416135788, 'epoch': 0.0}

  1%|▏         | 45/3300 [04:48<5:45:51,  6.38s/it]INFO 04-13 12:50:31 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 46/3300 [04:54<5:38:00,  6.23s/it]2026-04-13 12:50:36,876 | INFO | train_grpo_train | metrics_logged mode=train step=46

                                                   
{'loss': 0.0744, 'grad_norm': 24.522783279418945, 'learning_rate': 9.863636363636364e-06, 'num_tokens': 84627.0, 'completions/mean_length': 39.875, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5732947587966919, 'rewards/meter/std': 0.4071579575538635, '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': 0.9985332489013672, 'rewards/repeat_soft/std': 0.002732667839154601, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.1403057873249054, 'rewards/total_composite/mean': 0.22148951888084412, 'rewards/total_composite/std': 0.06108399108052254, 'reward': 0.22148951888084412, 'reward_std': 0.06108398735523224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2159995138645172, 'sampling/sampling_logp_difference/max': 1.9328289031982422, 'sampling/importance_sampling_ratio/min': 0.14473818242549896, 'sampling/importance_sampling_ratio/mean': 1.0379143953323364, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3000453412532806, 'clip_ratio/low_mean': 0.0911612194031477, 'clip_ratio/low_min': 0.0911612194031477, 'clip_ratio/high_mean': 0.10591182392090559, 'clip_ratio/high_max': 0.10591182392090559, 'clip_ratio/region_mean': 0.1970730433240533, 'reward_total_mean': 0.22148951888084412, 'reward_meter_mean': 0.5732947587966919, 'reward_meter_std': 0.4071579575538635, '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': 0.9985332489013672, 'reward_repeat_soft_std': 0.002732667839154601, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.1403057873249054, 'reward_total_composite_mean': 0.22148951888084412, 'reward_total_composite_std': 0.06108399108052254, 'epoch': 0.0}

  1%|▏         | 46/3300 [04:54<5:38:00,  6.23s/it]INFO 04-13 12:50:37 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 47/3300 [05:00<5:35:03,  6.18s/it]2026-04-13 12:50:42,935 | INFO | train_grpo_train | metrics_logged mode=train step=47

                                                   
{'loss': 0.1602, 'grad_norm': 19.036218643188477, 'learning_rate': 9.860606060606061e-06, 'num_tokens': 86187.0, 'completions/mean_length': 46.0, 'completions/min_length': 33.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5120313167572021, 'rewards/meter/std': 0.4290798306465149, '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': 0.9942026138305664, 'rewards/repeat_soft/std': 0.008400977589190006, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.13845448195934296, 'rewards/total_composite/mean': 0.22351449728012085, 'rewards/total_composite/std': 0.0865553617477417, 'reward': 0.22351449728012085, 'reward_std': 0.0865553542971611, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24033474922180176, 'sampling/sampling_logp_difference/max': 1.5775656700134277, 'sampling/importance_sampling_ratio/min': 0.2064771205186844, 'sampling/importance_sampling_ratio/mean': 1.0289047956466675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0344381779432297, 'clip_ratio/low_mean': 0.10997415706515312, 'clip_ratio/low_min': 0.10997415706515312, 'clip_ratio/high_mean': 0.11138846911489964, 'clip_ratio/high_max': 0.11138846911489964, 'clip_ratio/region_mean': 0.22136262618005276, 'reward_total_mean': 0.22351449728012085, 'reward_meter_mean': 0.5120313167572021, 'reward_meter_std': 0.4290798306465149, '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': 0.9942026138305664, 'reward_repeat_soft_std': 0.008400977589190006, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.13845448195934296, 'reward_total_composite_mean': 0.22351449728012085, 'reward_total_composite_std': 0.0865553617477417, 'epoch': 0.0}

  1%|▏         | 47/3300 [05:00<5:35:03,  6.18s/it]INFO 04-13 12:50:43 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 48/3300 [05:06<5:30:29,  6.10s/it]2026-04-13 12:50:48,839 | INFO | train_grpo_train | metrics_logged mode=train step=48

                                                   
{'loss': 0.088, 'grad_norm': 29.396474838256836, 'learning_rate': 9.857575757575758e-06, 'num_tokens': 87821.0, 'completions/mean_length': 35.25, 'completions/min_length': 27.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.609480619430542, 'rewards/meter/std': 0.26531553268432617, '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': 0.9907858967781067, 'rewards/repeat_soft/std': 0.00978943146765232, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.23217912018299103, 'rewards/total_composite/mean': 0.3059554100036621, 'rewards/total_composite/std': 0.1653279811143875, 'reward': 0.3059554100036621, 'reward_std': 0.1653279811143875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24612577259540558, 'sampling/sampling_logp_difference/max': 2.0493764877319336, 'sampling/importance_sampling_ratio/min': 0.1288152039051056, 'sampling/importance_sampling_ratio/mean': 0.9925991296768188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6261170208454132, 'clip_ratio/low_mean': 0.1605833861976862, 'clip_ratio/low_min': 0.1605833861976862, 'clip_ratio/high_mean': 0.05046432092785835, 'clip_ratio/high_max': 0.05046432092785835, 'clip_ratio/region_mean': 0.21104770712554455, 'reward_total_mean': 0.3059554100036621, 'reward_meter_mean': 0.609480619430542, 'reward_meter_std': 0.26531553268432617, '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': 0.9907858967781067, 'reward_repeat_soft_std': 0.00978943146765232, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.23217912018299103, 'reward_total_composite_mean': 0.3059554100036621, 'reward_total_composite_std': 0.1653279811143875, 'epoch': 0.0}

  1%|▏         | 48/3300 [05:06<5:30:29,  6.10s/it]INFO 04-13 12:50:49 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 49/3300 [05:14<5:56:24,  6.58s/it]2026-04-13 12:50:56,538 | INFO | train_grpo_train | metrics_logged mode=train step=49

                                                   
{'loss': 0.1115, 'grad_norm': 14.390752792358398, 'learning_rate': 9.854545454545456e-06, 'num_tokens': 90330.0, 'completions/mean_length': 122.625, 'completions/min_length': 92.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.625, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.3861793279647827, 'rewards/meter/std': 0.30217719078063965, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9951063394546509, 'rewards/repeat_soft/std': 0.0040963743813335896, 'rewards/judge_quality/mean': 0.28999999165534973, 'rewards/judge_quality/std': 0.10849621146917343, 'rewards/total_composite/mean': 0.14654898643493652, 'rewards/total_composite/std': 0.08626699447631836, 'reward': 0.14654898643493652, 'reward_std': 0.08626698702573776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2900907099246979, 'sampling/sampling_logp_difference/max': 3.003718614578247, 'sampling/importance_sampling_ratio/min': 0.04960227385163307, 'sampling/importance_sampling_ratio/mean': 0.9769321084022522, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6626681834459305, 'clip_ratio/low_mean': 0.10772638954222202, 'clip_ratio/low_min': 0.10772638954222202, 'clip_ratio/high_mean': 0.12409811466932297, 'clip_ratio/high_max': 0.12409811466932297, 'clip_ratio/region_mean': 0.231824504211545, 'reward_total_mean': 0.14654898643493652, 'reward_meter_mean': 0.3861793279647827, 'reward_meter_std': 0.30217719078063965, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9951063394546509, 'reward_repeat_soft_std': 0.0040963743813335896, 'reward_judge_quality_mean': 0.28999999165534973, 'reward_judge_quality_std': 0.10849621146917343, 'reward_total_composite_mean': 0.14654898643493652, 'reward_total_composite_std': 0.08626699447631836, 'epoch': 0.0}

  1%|▏         | 49/3300 [05:14<5:56:24,  6.58s/it]INFO 04-13 12:50:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 50/3300 [05:21<6:08:03,  6.79s/it]2026-04-13 12:51:03,897 | INFO | train_grpo_train | metrics_logged mode=train step=50

                                                   
{'loss': 0.2079, 'grad_norm': 12.737960815429688, 'learning_rate': 9.851515151515151e-06, 'num_tokens': 92296.0, 'completions/mean_length': 83.75, 'completions/min_length': 57.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.276641845703125, 'rewards/meter/std': 0.22296901047229767, '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': 0.9936968088150024, 'rewards/repeat_soft/std': 0.005944508593529463, 'rewards/judge_quality/mean': 0.25999999046325684, 'rewards/judge_quality/std': 0.07010196894407272, 'rewards/total_composite/mean': 0.14282579720020294, 'rewards/total_composite/std': 0.07114465534687042, 'reward': 0.14282579720020294, 'reward_std': 0.07114465534687042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22888238728046417, 'sampling/sampling_logp_difference/max': 1.416872501373291, 'sampling/importance_sampling_ratio/min': 0.24247115850448608, 'sampling/importance_sampling_ratio/mean': 1.0475715398788452, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.287784308195114, 'clip_ratio/low_mean': 0.13792912662029266, 'clip_ratio/low_min': 0.13792912662029266, 'clip_ratio/high_mean': 0.077578354626894, 'clip_ratio/high_max': 0.077578354626894, 'clip_ratio/region_mean': 0.21550748124718666, 'reward_total_mean': 0.14282579720020294, 'reward_meter_mean': 0.276641845703125, 'reward_meter_std': 0.22296901047229767, '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': 0.9936968088150024, 'reward_repeat_soft_std': 0.005944508593529463, 'reward_judge_quality_mean': 0.25999999046325684, 'reward_judge_quality_std': 0.07010196894407272, 'reward_total_composite_mean': 0.14282579720020294, 'reward_total_composite_std': 0.07114465534687042, 'epoch': 0.01}

  2%|▏         | 50/3300 [05:21<6:08:03,  6.79s/it]INFO 04-13 12:51:04 [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,  2.12s/it]

 30%|███       | 3/10 [00:08<00:21,  3.03s/it]

 40%|████      | 4/10 [00:12<00:19,  3.19s/it]

 50%|█████     | 5/10 [00:17<00:19,  3.88s/it]

 60%|██████    | 6/10 [00:26<00:22,  5.75s/it]

 70%|███████   | 7/10 [00:34<00:19,  6.39s/it]

 80%|████████  | 8/10 [00:40<00:12,  6.33s/it]

 90%|█████████ | 9/10 [00:47<00:06,  6.65s/it]

100%|██████████| 10/10 [00:52<00:00,  5.97s/it]2026-04-13 12:52:02,196 | INFO | train_grpo_train | metrics_logged mode=eval step=50

                                                   

                                               
{'eval_loss': nan, 'eval_runtime': 58.292, 'eval_samples_per_second': 1.372, 'eval_steps_per_second': 0.172, 'eval_num_tokens': 92296.0, 'eval_completions/mean_length': 78.7375, 'eval_completions/min_length': 29.7, 'eval_completions/max_length': 145.5, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 78.7375, 'eval_completions/min_terminated_length': 29.7, 'eval_completions/max_terminated_length': 145.5, 'eval_rewards/meter/mean': 0.5631425261497498, 'eval_rewards/meter/std': 0.3836209237575531, 'eval_rewards/count_adherence/mean': 0.9606249868869782, 'eval_rewards/count_adherence/std': 0.07844334654510021, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.2230676978826523, 'eval_rewards/repeat_soft/mean': 0.9909339249134064, 'eval_rewards/repeat_soft/std': 0.015210337797179818, 'eval_rewards/judge_quality/mean': 0.2937499925494194, 'eval_rewards/judge_quality/std': 0.13265312099829316, 'eval_rewards/total_composite/mean': 0.2063873901963234, 'eval_rewards/total_composite/std': 0.15400917865335942, 'eval_reward': 0.2063873901963234, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.14399238526821137, 'eval_sampling/sampling_logp_difference/max': 1.1252060890197755, 'eval_sampling/importance_sampling_ratio/min': 0.32608655989170077, 'eval_sampling/importance_sampling_ratio/mean': 1.0359672784805298, 'eval_sampling/importance_sampling_ratio/max': 1.5399682164192199, 'eval_entropy': 2.1157538652420045, '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': 0.2063873901963234, 'eval_reward_meter_mean': 0.5631425261497498, 'eval_reward_meter_std': 0.3836209237575531, 'eval_reward_count_adherence_mean': 0.9606249868869782, 'eval_reward_count_adherence_std': 0.07844334654510021, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.2230676978826523, 'eval_reward_repeat_soft_mean': 0.9909339249134064, 'eval_reward_repeat_soft_std': 0.015210337797179818, 'eval_reward_judge_quality_mean': 0.2937499925494194, 'eval_reward_judge_quality_std': 0.13265312099829316, 'eval_reward_total_composite_mean': 0.2063873901963234, 'eval_reward_total_composite_std': 0.15400917865335942, 'epoch': 0.01}

  2%|▏         | 50/3300 [06:19<6:08:03,  6.79s/it]

100%|██████████| 10/10 [00:52<00:00,  5.97s/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 [06:28<22:29:40, 24.92s/it]2026-04-13 12:52:11,231 | INFO | train_grpo_train | metrics_logged mode=train step=51

                                                    
{'loss': 0.1536, 'grad_norm': 13.13121223449707, 'learning_rate': 9.84848484848485e-06, 'num_tokens': 94078.0, 'completions/mean_length': 55.75, 'completions/min_length': 35.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.573480486869812, 'rewards/meter/std': 0.29802122712135315, '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': 0.9975300431251526, 'rewards/repeat_soft/std': 0.0033512876834720373, 'rewards/judge_quality/mean': 0.32374998927116394, 'rewards/judge_quality/std': 0.1487027108669281, 'rewards/total_composite/mean': 0.23654907941818237, 'rewards/total_composite/std': 0.13750989735126495, 'reward': 0.23654907941818237, 'reward_std': 0.13750988245010376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19188879430294037, 'sampling/sampling_logp_difference/max': 1.367401123046875, 'sampling/importance_sampling_ratio/min': 0.2547681927680969, 'sampling/importance_sampling_ratio/mean': 1.0396616458892822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1432204842567444, 'clip_ratio/low_mean': 0.08935721637681127, 'clip_ratio/low_min': 0.08935721637681127, 'clip_ratio/high_mean': 0.04642857238650322, 'clip_ratio/high_max': 0.04642857238650322, 'clip_ratio/region_mean': 0.13578578876331449, 'reward_total_mean': 0.23654907941818237, 'reward_meter_mean': 0.573480486869812, 'reward_meter_std': 0.29802122712135315, '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': 0.9975300431251526, 'reward_repeat_soft_std': 0.0033512876834720373, 'reward_judge_quality_mean': 0.32374998927116394, 'reward_judge_quality_std': 0.1487027108669281, 'reward_total_composite_mean': 0.23654907941818237, 'reward_total_composite_std': 0.13750989735126495, 'epoch': 0.01}

  2%|▏         | 51/3300 [06:28<22:29:40, 24.92s/it]INFO 04-13 12:52:11 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 52/3300 [06:36<17:49:52, 19.76s/it]2026-04-13 12:52:18,789 | INFO | train_grpo_train | metrics_logged mode=train step=52

                                                    
{'loss': 0.1116, 'grad_norm': 12.535508155822754, 'learning_rate': 9.845454545454546e-06, 'num_tokens': 96150.0, 'completions/mean_length': 85.0, 'completions/min_length': 70.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.784797191619873, 'rewards/meter/std': 0.23827414214611053, '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': 0.993104100227356, 'rewards/repeat_soft/std': 0.004849093966186047, 'rewards/judge_quality/mean': 0.32374998927116394, 'rewards/judge_quality/std': 0.1487027257680893, 'rewards/total_composite/mean': 0.28121325373649597, 'rewards/total_composite/std': 0.15429997444152832, 'reward': 0.28121325373649597, 'reward_std': 0.15429995954036713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24347689747810364, 'sampling/sampling_logp_difference/max': 1.919011116027832, 'sampling/importance_sampling_ratio/min': 0.1467519998550415, 'sampling/importance_sampling_ratio/mean': 1.0357811450958252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.725496292114258, 'clip_ratio/low_mean': 0.14716156385838985, 'clip_ratio/low_min': 0.14716156385838985, 'clip_ratio/high_mean': 0.07188026420772076, 'clip_ratio/high_max': 0.07188026420772076, 'clip_ratio/region_mean': 0.2190418280661106, 'reward_total_mean': 0.28121325373649597, 'reward_meter_mean': 0.784797191619873, 'reward_meter_std': 0.23827414214611053, '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': 0.993104100227356, 'reward_repeat_soft_std': 0.004849093966186047, 'reward_judge_quality_mean': 0.32374998927116394, 'reward_judge_quality_std': 0.1487027257680893, 'reward_total_composite_mean': 0.28121325373649597, 'reward_total_composite_std': 0.15429997444152832, 'epoch': 0.01}

  2%|▏         | 52/3300 [06:36<17:49:52, 19.76s/it]INFO 04-13 12:52:19 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 53/3300 [06:42<14:15:35, 15.81s/it]2026-04-13 12:52:25,374 | INFO | train_grpo_train | metrics_logged mode=train step=53

                                                    
{'loss': 0.0093, 'grad_norm': 22.459810256958008, 'learning_rate': 9.842424242424243e-06, 'num_tokens': 97682.0, 'completions/mean_length': 38.5, 'completions/min_length': 26.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6416175365447998, 'rewards/meter/std': 0.4357673227787018, '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': 0.9862592816352844, 'rewards/repeat_soft/std': 0.022048698738217354, 'rewards/judge_quality/mean': 0.30250000953674316, 'rewards/judge_quality/std': 0.09808888286352158, 'rewards/total_composite/mean': 0.23128661513328552, 'rewards/total_composite/std': 0.11701280623674393, 'reward': 0.23128661513328552, 'reward_std': 0.11701279133558273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2023971527814865, 'sampling/sampling_logp_difference/max': 2.5468854904174805, 'sampling/importance_sampling_ratio/min': 0.07832523435354233, 'sampling/importance_sampling_ratio/mean': 0.9859792590141296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5731562711298466, 'clip_ratio/low_mean': 0.10033906158059835, 'clip_ratio/low_min': 0.10033906158059835, 'clip_ratio/high_mean': 0.04652255633845925, 'clip_ratio/high_max': 0.04652255633845925, 'clip_ratio/region_mean': 0.1468616179190576, 'reward_total_mean': 0.23128661513328552, 'reward_meter_mean': 0.6416175365447998, 'reward_meter_std': 0.4357673227787018, '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': 0.9862592816352844, 'reward_repeat_soft_std': 0.022048698738217354, 'reward_judge_quality_mean': 0.30250000953674316, 'reward_judge_quality_std': 0.09808888286352158, 'reward_total_composite_mean': 0.23128661513328552, 'reward_total_composite_std': 0.11701280623674393, 'epoch': 0.01}

  2%|▏         | 53/3300 [06:42<14:15:35, 15.81s/it]INFO 04-13 12:52:25 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 54/3300 [06:51<12:09:18, 13.48s/it]2026-04-13 12:52:33,420 | INFO | train_grpo_train | metrics_logged mode=train step=54

                                                    
{'loss': 0.1895, 'grad_norm': 13.117853164672852, 'learning_rate': 9.83939393939394e-06, 'num_tokens': 99998.0, 'completions/mean_length': 112.5, 'completions/min_length': 77.0, 'completions/max_length': 185.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.5, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 185.0, 'rewards/meter/mean': 0.29589852690696716, 'rewards/meter/std': 0.34566956758499146, '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': 0.9958626627922058, 'rewards/repeat_soft/std': 0.00357510126195848, 'rewards/judge_quality/mean': 0.26874998211860657, 'rewards/judge_quality/std': 0.09523466974496841, 'rewards/total_composite/mean': 0.14415407180786133, 'rewards/total_composite/std': 0.06718795001506805, 'reward': 0.14415407180786133, 'reward_std': 0.06718795001506805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22546643018722534, 'sampling/sampling_logp_difference/max': 1.6960043907165527, 'sampling/importance_sampling_ratio/min': 0.18341492116451263, 'sampling/importance_sampling_ratio/mean': 1.0321494340896606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1551734805107117, 'clip_ratio/low_mean': 0.10768418572843075, 'clip_ratio/low_min': 0.10768418572843075, 'clip_ratio/high_mean': 0.1142218578606844, 'clip_ratio/high_max': 0.1142218578606844, 'clip_ratio/region_mean': 0.22190604358911514, 'reward_total_mean': 0.14415407180786133, 'reward_meter_mean': 0.29589852690696716, 'reward_meter_std': 0.34566956758499146, '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': 0.9958626627922058, 'reward_repeat_soft_std': 0.00357510126195848, 'reward_judge_quality_mean': 0.26874998211860657, 'reward_judge_quality_std': 0.09523466974496841, 'reward_total_composite_mean': 0.14415407180786133, 'reward_total_composite_std': 0.06718795001506805, 'epoch': 0.01}

  2%|▏         | 54/3300 [06:51<12:09:18, 13.48s/it]INFO 04-13 12:52:33 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 55/3300 [06:57<10:12:39, 11.33s/it]2026-04-13 12:52:39,726 | INFO | train_grpo_train | metrics_logged mode=train step=55

                                                    
{'loss': 0.037, 'grad_norm': 17.670133590698242, 'learning_rate': 9.836363636363637e-06, 'num_tokens': 101550.0, 'completions/mean_length': 27.0, 'completions/min_length': 19.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.4525359272956848, 'rewards/meter/std': 0.3512803316116333, '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': 0.48125001788139343, 'rewards/judge_quality/std': 0.2924496531486511, 'rewards/total_composite/mean': 0.32188645005226135, 'rewards/total_composite/std': 0.25953197479248047, 'reward': 0.32188645005226135, 'reward_std': 0.2595319449901581, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1912689059972763, 'sampling/sampling_logp_difference/max': 1.9762558937072754, 'sampling/importance_sampling_ratio/min': 0.13858714699745178, 'sampling/importance_sampling_ratio/mean': 1.041345238685608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.816194623708725, 'clip_ratio/low_mean': 0.13259341288357973, 'clip_ratio/low_min': 0.13259341288357973, 'clip_ratio/high_mean': 0.08467167429625988, 'clip_ratio/high_max': 0.08467167429625988, 'clip_ratio/region_mean': 0.2172650871798396, 'reward_total_mean': 0.32188645005226135, 'reward_meter_mean': 0.4525359272956848, 'reward_meter_std': 0.3512803316116333, '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': 0.48125001788139343, 'reward_judge_quality_std': 0.2924496531486511, 'reward_total_composite_mean': 0.32188645005226135, 'reward_total_composite_std': 0.25953197479248047, 'epoch': 0.01}

  2%|▏         | 55/3300 [06:57<10:12:39, 11.33s/it]INFO 04-13 12:52:39 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 56/3300 [07:05<9:25:57, 10.47s/it] 2026-04-13 12:52:48,186 | INFO | train_grpo_train | metrics_logged mode=train step=56

                                                   
{'loss': -0.0197, 'grad_norm': 19.314542770385742, 'learning_rate': 9.833333333333333e-06, 'num_tokens': 103028.0, 'completions/mean_length': 26.75, 'completions/min_length': 19.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.3456041216850281, 'rewards/meter/std': 0.446493923664093, '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': 0.25, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.14225441217422485, 'rewards/total_composite/std': 0.07255525887012482, 'reward': 0.14225441217422485, 'reward_std': 0.07255525887012482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22346875071525574, 'sampling/sampling_logp_difference/max': 1.535238265991211, 'sampling/importance_sampling_ratio/min': 0.21540437638759613, 'sampling/importance_sampling_ratio/mean': 0.9892282485961914, 'sampling/importance_sampling_ratio/max': 1.902657151222229, 'entropy': 1.9158771336078644, 'clip_ratio/low_mean': 0.13819932006299496, 'clip_ratio/low_min': 0.13819932006299496, 'clip_ratio/high_mean': 0.11600378900766373, 'clip_ratio/high_max': 0.11600378900766373, 'clip_ratio/region_mean': 0.2542031090706587, 'reward_total_mean': 0.14225441217422485, 'reward_meter_mean': 0.3456041216850281, 'reward_meter_std': 0.446493923664093, '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': 0.25, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.14225441217422485, 'reward_total_composite_std': 0.07255525887012482, 'epoch': 0.01}

  2%|▏         | 56/3300 [07:05<9:25:57, 10.47s/it]INFO 04-13 12:52:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 57/3300 [07:12<8:29:20,  9.42s/it]2026-04-13 12:52:55,172 | INFO | train_grpo_train | metrics_logged mode=train step=57

                                                   
{'loss': 0.1387, 'grad_norm': 10.624090194702148, 'learning_rate': 9.830303030303032e-06, 'num_tokens': 105143.0, 'completions/mean_length': 94.375, 'completions/min_length': 56.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.7765734195709229, 'rewards/meter/std': 0.2771216928958893, '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': 0.9940077662467957, 'rewards/repeat_soft/std': 0.009384247474372387, 'rewards/judge_quality/mean': 0.28125, 'rewards/judge_quality/std': 0.0882265716791153, 'rewards/total_composite/mean': 0.2469685971736908, 'rewards/total_composite/std': 0.10987038910388947, 'reward': 0.2469685971736908, 'reward_std': 0.10987038165330887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2192666232585907, 'sampling/sampling_logp_difference/max': 1.433286190032959, 'sampling/importance_sampling_ratio/min': 0.25009241700172424, 'sampling/importance_sampling_ratio/mean': 1.0092370510101318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.317024514079094, 'clip_ratio/low_mean': 0.11426695715636015, 'clip_ratio/low_min': 0.11426695715636015, 'clip_ratio/high_mean': 0.09074284695088863, 'clip_ratio/high_max': 0.09074284695088863, 'clip_ratio/region_mean': 0.20500980410724878, 'reward_total_mean': 0.2469685971736908, 'reward_meter_mean': 0.7765734195709229, 'reward_meter_std': 0.2771216928958893, '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': 0.9940077662467957, 'reward_repeat_soft_std': 0.009384247474372387, 'reward_judge_quality_mean': 0.28125, 'reward_judge_quality_std': 0.0882265716791153, 'reward_total_composite_mean': 0.2469685971736908, 'reward_total_composite_std': 0.10987038910388947, 'epoch': 0.01}

  2%|▏         | 57/3300 [07:12<8:29:20,  9.42s/it]INFO 04-13 12:52:55 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 58/3300 [07:19<7:48:45,  8.68s/it]2026-04-13 12:53:02,102 | INFO | train_grpo_train | metrics_logged mode=train step=58

                                                   
{'loss': 0.0149, 'grad_norm': 10.558327674865723, 'learning_rate': 9.827272727272729e-06, 'num_tokens': 107309.0, 'completions/mean_length': 87.75, 'completions/min_length': 63.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.75, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.401488721370697, 'rewards/meter/std': 0.3869341313838959, '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': 0.9952694177627563, 'rewards/repeat_soft/std': 0.00658520869910717, 'rewards/judge_quality/mean': 0.22749999165534973, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.13977158069610596, 'rewards/total_composite/std': 0.061555054038763046, 'reward': 0.13977158069610596, 'reward_std': 0.061555054038763046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23097935318946838, 'sampling/sampling_logp_difference/max': 1.7834727764129639, 'sampling/importance_sampling_ratio/min': 0.1680535227060318, 'sampling/importance_sampling_ratio/mean': 1.0329930782318115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.427265539765358, 'clip_ratio/low_mean': 0.1279610563069582, 'clip_ratio/low_min': 0.1279610563069582, 'clip_ratio/high_mean': 0.09140318259596825, 'clip_ratio/high_max': 0.09140318259596825, 'clip_ratio/region_mean': 0.21936423890292645, 'reward_total_mean': 0.13977158069610596, 'reward_meter_mean': 0.401488721370697, 'reward_meter_std': 0.3869341313838959, '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': 0.9952694177627563, 'reward_repeat_soft_std': 0.00658520869910717, 'reward_judge_quality_mean': 0.22749999165534973, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.13977158069610596, 'reward_total_composite_std': 0.061555054038763046, 'epoch': 0.01}

  2%|▏         | 58/3300 [07:19<7:48:45,  8.68s/it]INFO 04-13 12:53:02 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 59/3300 [07:26<7:15:26,  8.06s/it]2026-04-13 12:53:08,731 | INFO | train_grpo_train | metrics_logged mode=train step=59

                                                   
{'loss': -0.0049, 'grad_norm': 15.45472526550293, 'learning_rate': 9.824242424242425e-06, 'num_tokens': 109189.0, 'completions/mean_length': 56.0, 'completions/min_length': 34.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.6033459901809692, 'rewards/meter/std': 0.41086646914482117, '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': 0.9966139793395996, 'rewards/repeat_soft/std': 0.006910442374646664, 'rewards/judge_quality/mean': 0.30250000953674316, 'rewards/judge_quality/std': 0.20954032242298126, 'rewards/total_composite/mean': 0.24199925363063812, 'rewards/total_composite/std': 0.24158596992492676, 'reward': 0.24199925363063812, 'reward_std': 0.24158596992492676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.224238783121109, 'sampling/sampling_logp_difference/max': 1.9477157592773438, 'sampling/importance_sampling_ratio/min': 0.1425994336605072, 'sampling/importance_sampling_ratio/mean': 1.0293304920196533, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8733589500188828, 'clip_ratio/low_mean': 0.15564925968647003, 'clip_ratio/low_min': 0.15564925968647003, 'clip_ratio/high_mean': 0.05659562349319458, 'clip_ratio/high_max': 0.05659562349319458, 'clip_ratio/region_mean': 0.2122448831796646, 'reward_total_mean': 0.24199925363063812, 'reward_meter_mean': 0.6033459901809692, 'reward_meter_std': 0.41086646914482117, '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': 0.9966139793395996, 'reward_repeat_soft_std': 0.006910442374646664, 'reward_judge_quality_mean': 0.30250000953674316, 'reward_judge_quality_std': 0.20954032242298126, 'reward_total_composite_mean': 0.24199925363063812, 'reward_total_composite_std': 0.24158596992492676, 'epoch': 0.01}

  2%|▏         | 59/3300 [07:26<7:15:26,  8.06s/it]INFO 04-13 12:53:08 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 60/3300 [07:32<6:46:25,  7.53s/it]2026-04-13 12:53:15,009 | INFO | train_grpo_train | metrics_logged mode=train step=60

                                                   
{'loss': 0.0373, 'grad_norm': 17.19390869140625, 'learning_rate': 9.821212121212122e-06, 'num_tokens': 110693.0, 'completions/mean_length': 43.0, 'completions/min_length': 33.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8154430389404297, 'rewards/meter/std': 0.24470537900924683, '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': 0.9848794937133789, 'rewards/repeat_soft/std': 0.03346961736679077, 'rewards/judge_quality/mean': 0.36000001430511475, 'rewards/judge_quality/std': 0.17188036441802979, 'rewards/total_composite/mean': 0.30636218190193176, 'rewards/total_composite/std': 0.13998998701572418, 'reward': 0.30636218190193176, 'reward_std': 0.13999000191688538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20322245359420776, 'sampling/sampling_logp_difference/max': 1.3733539581298828, 'sampling/importance_sampling_ratio/min': 0.2532561123371124, 'sampling/importance_sampling_ratio/mean': 1.0346107482910156, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8409823924303055, 'clip_ratio/low_mean': 0.08293335977941751, 'clip_ratio/low_min': 0.08293335977941751, 'clip_ratio/high_mean': 0.07878378219902515, 'clip_ratio/high_max': 0.07878378219902515, 'clip_ratio/region_mean': 0.16171714197844267, 'reward_total_mean': 0.30636218190193176, 'reward_meter_mean': 0.8154430389404297, 'reward_meter_std': 0.24470537900924683, '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': 0.9848794937133789, 'reward_repeat_soft_std': 0.03346961736679077, 'reward_judge_quality_mean': 0.36000001430511475, 'reward_judge_quality_std': 0.17188036441802979, 'reward_total_composite_mean': 0.30636218190193176, 'reward_total_composite_std': 0.13998998701572418, 'epoch': 0.01}

  2%|▏         | 60/3300 [07:32<6:46:25,  7.53s/it]INFO 04-13 12:53:15 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 61/3300 [07:39<6:31:32,  7.25s/it]2026-04-13 12:53:21,624 | INFO | train_grpo_train | metrics_logged mode=train step=61

                                                   
{'loss': 0.0539, 'grad_norm': 14.639938354492188, 'learning_rate': 9.81818181818182e-06, 'num_tokens': 112447.0, 'completions/mean_length': 45.25, 'completions/min_length': 31.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.42510533332824707, 'rewards/meter/std': 0.3660593330860138, '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': 0.9985795021057129, 'rewards/repeat_soft/std': 0.0020165368914604187, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.1667708158493042, 'rewards/total_composite/mean': 0.22165712714195251, 'rewards/total_composite/std': 0.09564365446567535, 'reward': 0.22165712714195251, 'reward_std': 0.09564363956451416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22174100577831268, 'sampling/sampling_logp_difference/max': 1.689234733581543, 'sampling/importance_sampling_ratio/min': 0.18466079235076904, 'sampling/importance_sampling_ratio/mean': 1.0645029544830322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2769857197999954, 'clip_ratio/low_mean': 0.12808788567781448, 'clip_ratio/low_min': 0.12808788567781448, 'clip_ratio/high_mean': 0.11083407700061798, 'clip_ratio/high_max': 0.11083407700061798, 'clip_ratio/region_mean': 0.23892196267843246, 'reward_total_mean': 0.22165712714195251, 'reward_meter_mean': 0.42510533332824707, 'reward_meter_std': 0.3660593330860138, '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': 0.9985795021057129, 'reward_repeat_soft_std': 0.0020165368914604187, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.1667708158493042, 'reward_total_composite_mean': 0.22165712714195251, 'reward_total_composite_std': 0.09564365446567535, 'epoch': 0.01}

  2%|▏         | 61/3300 [07:39<6:31:32,  7.25s/it]INFO 04-13 12:53:21 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 62/3300 [07:47<6:40:11,  7.42s/it]2026-04-13 12:53:29,421 | INFO | train_grpo_train | metrics_logged mode=train step=62

                                                   
{'loss': 0.0496, 'grad_norm': 10.596074104309082, 'learning_rate': 9.815151515151516e-06, 'num_tokens': 115297.0, 'completions/mean_length': 152.25, 'completions/min_length': 81.0, 'completions/max_length': 193.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.25, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 193.0, 'rewards/meter/mean': 0.3930585980415344, 'rewards/meter/std': 0.19574135541915894, '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': 0.9903420209884644, 'rewards/repeat_soft/std': 0.024503352120518684, 'rewards/judge_quality/mean': 0.2774999737739563, 'rewards/judge_quality/std': 0.09035643190145493, 'rewards/total_composite/mean': 0.16664311289787292, 'rewards/total_composite/std': 0.06804937869310379, 'reward': 0.16664311289787292, 'reward_std': 0.06804937869310379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22418749332427979, 'sampling/sampling_logp_difference/max': 1.676156997680664, 'sampling/importance_sampling_ratio/min': 0.18709158897399902, 'sampling/importance_sampling_ratio/mean': 1.0282597541809082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.824618175625801, 'clip_ratio/low_mean': 0.1227662805467844, 'clip_ratio/low_min': 0.1227662805467844, 'clip_ratio/high_mean': 0.07461014576256275, 'clip_ratio/high_max': 0.07461014576256275, 'clip_ratio/region_mean': 0.19737642630934715, 'reward_total_mean': 0.16664311289787292, 'reward_meter_mean': 0.3930585980415344, 'reward_meter_std': 0.19574135541915894, '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': 0.9903420209884644, 'reward_repeat_soft_std': 0.024503352120518684, 'reward_judge_quality_mean': 0.2774999737739563, 'reward_judge_quality_std': 0.09035643190145493, 'reward_total_composite_mean': 0.16664311289787292, 'reward_total_composite_std': 0.06804937869310379, 'epoch': 0.01}

  2%|▏         | 62/3300 [07:47<6:40:11,  7.42s/it]INFO 04-13 12:53:29 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 63/3300 [07:53<6:31:13,  7.25s/it]2026-04-13 12:53:36,291 | INFO | train_grpo_train | metrics_logged mode=train step=63

                                                   
{'loss': 0.0831, 'grad_norm': 39.91816711425781, 'learning_rate': 9.812121212121212e-06, 'num_tokens': 117048.0, 'completions/mean_length': 57.875, 'completions/min_length': 55.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5123488903045654, 'rewards/meter/std': 0.3575299084186554, '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': 0.9971224069595337, 'rewards/repeat_soft/std': 0.0021660495549440384, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.09304375946521759, 'rewards/total_composite/mean': 0.21182453632354736, 'rewards/total_composite/std': 0.08976436406373978, 'reward': 0.21182453632354736, 'reward_std': 0.08976437151432037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18080253899097443, 'sampling/sampling_logp_difference/max': 2.7697935104370117, 'sampling/importance_sampling_ratio/min': 0.06267494708299637, 'sampling/importance_sampling_ratio/mean': 0.987522542476654, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7751255873590708, 'clip_ratio/low_mean': 0.0875050537288189, 'clip_ratio/low_min': 0.0875050537288189, 'clip_ratio/high_mean': 0.08232656121253967, 'clip_ratio/high_max': 0.08232656121253967, 'clip_ratio/region_mean': 0.16983161494135857, 'reward_total_mean': 0.21182453632354736, 'reward_meter_mean': 0.5123488903045654, 'reward_meter_std': 0.3575299084186554, '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': 0.9971224069595337, 'reward_repeat_soft_std': 0.0021660495549440384, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.09304375946521759, 'reward_total_composite_mean': 0.21182453632354736, 'reward_total_composite_std': 0.08976436406373978, 'epoch': 0.01}

  2%|▏         | 63/3300 [07:53<6:31:13,  7.25s/it]INFO 04-13 12:53:36 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 64/3300 [08:00<6:22:59,  7.10s/it]2026-04-13 12:53:43,039 | INFO | train_grpo_train | metrics_logged mode=train step=64

                                                   
{'loss': 0.0875, 'grad_norm': 14.192243576049805, 'learning_rate': 9.809090909090911e-06, 'num_tokens': 119152.0, 'completions/mean_length': 81.0, 'completions/min_length': 71.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.0, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.5464773178100586, 'rewards/meter/std': 0.28004148602485657, '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': 0.9979227781295776, 'rewards/repeat_soft/std': 0.0025919738691300154, 'rewards/judge_quality/mean': 0.26999998092651367, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/total_composite/mean': 0.18057963252067566, 'rewards/total_composite/std': 0.11420687288045883, 'reward': 0.18057963252067566, 'reward_std': 0.11420686542987823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2471581995487213, 'sampling/sampling_logp_difference/max': 2.2663111686706543, 'sampling/importance_sampling_ratio/min': 0.10369398444890976, 'sampling/importance_sampling_ratio/mean': 1.0343976020812988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.564219117164612, 'clip_ratio/low_mean': 0.10504915751516819, 'clip_ratio/low_min': 0.10504915751516819, 'clip_ratio/high_mean': 0.11415382288396358, 'clip_ratio/high_max': 0.11415382288396358, 'clip_ratio/region_mean': 0.21920298039913177, 'reward_total_mean': 0.18057963252067566, 'reward_meter_mean': 0.5464773178100586, 'reward_meter_std': 0.28004148602485657, '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': 0.9979227781295776, 'reward_repeat_soft_std': 0.0025919738691300154, 'reward_judge_quality_mean': 0.26999998092651367, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.18057963252067566, 'reward_total_composite_std': 0.11420687288045883, 'epoch': 0.01}

  2%|▏         | 64/3300 [08:00<6:22:59,  7.10s/it]INFO 04-13 12:53:43 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 65/3300 [08:07<6:12:20,  6.91s/it]2026-04-13 12:53:49,490 | INFO | train_grpo_train | metrics_logged mode=train step=65

                                                   
{'loss': 0.052, 'grad_norm': 29.54388427734375, 'learning_rate': 9.806060606060607e-06, 'num_tokens': 120695.0, 'completions/mean_length': 42.875, 'completions/min_length': 30.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5739319920539856, 'rewards/meter/std': 0.36332207918167114, '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': 0.9990882873535156, 'rewards/repeat_soft/std': 0.002288078423589468, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.15361711382865906, 'rewards/total_composite/mean': 0.2958294153213501, 'rewards/total_composite/std': 0.16002905368804932, 'reward': 0.2958294153213501, 'reward_std': 0.16002905368804932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2263413965702057, 'sampling/sampling_logp_difference/max': 2.2635364532470703, 'sampling/importance_sampling_ratio/min': 0.1039821058511734, 'sampling/importance_sampling_ratio/mean': 0.9821252822875977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9036289602518082, 'clip_ratio/low_mean': 0.12230175454169512, 'clip_ratio/low_min': 0.12230175454169512, 'clip_ratio/high_mean': 0.046962485648691654, 'clip_ratio/high_max': 0.046962485648691654, 'clip_ratio/region_mean': 0.16926424019038677, 'reward_total_mean': 0.2958294153213501, 'reward_meter_mean': 0.5739319920539856, 'reward_meter_std': 0.36332207918167114, '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': 0.9990882873535156, 'reward_repeat_soft_std': 0.002288078423589468, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.15361711382865906, 'reward_total_composite_mean': 0.2958294153213501, 'reward_total_composite_std': 0.16002905368804932, 'epoch': 0.01}

  2%|▏         | 65/3300 [08:07<6:12:20,  6.91s/it]INFO 04-13 12:53:49 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 66/3300 [08:17<7:09:49,  7.97s/it]2026-04-13 12:53:59,956 | INFO | train_grpo_train | metrics_logged mode=train step=66

                                                   
{'loss': 0.1807, 'grad_norm': 12.807340621948242, 'learning_rate': 9.803030303030304e-06, 'num_tokens': 122245.0, 'completions/mean_length': 49.75, 'completions/min_length': 33.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.6284730434417725, 'rewards/meter/std': 0.36383092403411865, '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': 0.9980244636535645, 'rewards/repeat_soft/std': 0.002252033678814769, 'rewards/judge_quality/mean': 0.3024999797344208, 'rewards/judge_quality/std': 0.09808888286352158, 'rewards/total_composite/mean': 0.23335705697536469, 'rewards/total_composite/std': 0.1112920418381691, 'reward': 0.23335705697536469, 'reward_std': 0.1112920418381691, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22042016685009003, 'sampling/sampling_logp_difference/max': 1.378006935119629, 'sampling/importance_sampling_ratio/min': 0.2520804703235626, 'sampling/importance_sampling_ratio/mean': 1.0496618747711182, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.378784015774727, 'clip_ratio/low_mean': 0.15572327934205532, 'clip_ratio/low_min': 0.15572327934205532, 'clip_ratio/high_mean': 0.048660715110599995, 'clip_ratio/high_max': 0.048660715110599995, 'clip_ratio/region_mean': 0.20438399445265532, 'reward_total_mean': 0.23335705697536469, 'reward_meter_mean': 0.6284730434417725, 'reward_meter_std': 0.36383092403411865, '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': 0.9980244636535645, 'reward_repeat_soft_std': 0.002252033678814769, 'reward_judge_quality_mean': 0.3024999797344208, 'reward_judge_quality_std': 0.09808888286352158, 'reward_total_composite_mean': 0.23335705697536469, 'reward_total_composite_std': 0.1112920418381691, 'epoch': 0.01}

  2%|▏         | 66/3300 [08:17<7:09:49,  7.97s/it]INFO 04-13 12:54:00 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 67/3300 [08:24<6:50:34,  7.62s/it]2026-04-13 12:54:06,749 | INFO | train_grpo_train | metrics_logged mode=train step=67

                                                   
{'loss': 0.198, 'grad_norm': 16.643781661987305, 'learning_rate': 9.800000000000001e-06, 'num_tokens': 123948.0, 'completions/mean_length': 53.875, 'completions/min_length': 31.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.8704103827476501, 'rewards/meter/std': 0.14272016286849976, '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': 0.9991481900215149, 'rewards/repeat_soft/std': 0.0020753606222569942, 'rewards/judge_quality/mean': 0.2449999898672104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.22029772400856018, 'rewards/total_composite/std': 0.08091630041599274, 'reward': 0.22029772400856018, 'reward_std': 0.08091631531715393, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2308141589164734, 'sampling/sampling_logp_difference/max': 1.8956928253173828, 'sampling/importance_sampling_ratio/min': 0.15021422505378723, 'sampling/importance_sampling_ratio/mean': 1.0245487689971924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7529315799474716, 'clip_ratio/low_mean': 0.20236334949731827, 'clip_ratio/low_min': 0.20236334949731827, 'clip_ratio/high_mean': 0.012096773833036423, 'clip_ratio/high_max': 0.012096773833036423, 'clip_ratio/region_mean': 0.2144601233303547, 'reward_total_mean': 0.22029772400856018, 'reward_meter_mean': 0.8704103827476501, 'reward_meter_std': 0.14272016286849976, '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': 0.9991481900215149, 'reward_repeat_soft_std': 0.0020753606222569942, 'reward_judge_quality_mean': 0.2449999898672104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.22029772400856018, 'reward_total_composite_std': 0.08091630041599274, 'epoch': 0.01}

  2%|▏         | 67/3300 [08:24<6:50:34,  7.62s/it]INFO 04-13 12:54:06 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 68/3300 [08:30<6:24:25,  7.14s/it]2026-04-13 12:54:12,758 | INFO | train_grpo_train | metrics_logged mode=train step=68

                                                   
{'loss': 0.0626, 'grad_norm': 17.704111099243164, 'learning_rate': 9.796969696969698e-06, 'num_tokens': 125479.0, 'completions/mean_length': 41.375, 'completions/min_length': 37.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5753023624420166, 'rewards/meter/std': 0.40855905413627625, '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': 0.9715557098388672, 'rewards/repeat_soft/std': 0.0266946442425251, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.17713528871536255, 'rewards/total_composite/std': 0.05704084038734436, 'reward': 0.17713528871536255, 'reward_std': 0.05704084038734436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1867666095495224, 'sampling/sampling_logp_difference/max': 1.1181955337524414, 'sampling/importance_sampling_ratio/min': 0.32686910033226013, 'sampling/importance_sampling_ratio/mean': 1.0235097408294678, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.623171053826809, 'clip_ratio/low_mean': 0.0751488097012043, 'clip_ratio/low_min': 0.0751488097012043, 'clip_ratio/high_mean': 0.08889474952593446, 'clip_ratio/high_max': 0.08889474952593446, 'clip_ratio/region_mean': 0.16404355922713876, 'reward_total_mean': 0.17713528871536255, 'reward_meter_mean': 0.5753023624420166, 'reward_meter_std': 0.40855905413627625, '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': 0.9715557098388672, 'reward_repeat_soft_std': 0.0266946442425251, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.17713528871536255, 'reward_total_composite_std': 0.05704084038734436, 'epoch': 0.01}

  2%|▏         | 68/3300 [08:30<6:24:25,  7.14s/it]INFO 04-13 12:54:12 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 69/3300 [08:36<6:09:15,  6.86s/it]2026-04-13 12:54:18,963 | INFO | train_grpo_train | metrics_logged mode=train step=69

                                                   
{'loss': 0.1341, 'grad_norm': 19.775510787963867, 'learning_rate': 9.793939393939394e-06, 'num_tokens': 127422.0, 'completions/mean_length': 55.875, 'completions/min_length': 48.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6374300718307495, 'rewards/meter/std': 0.2568950951099396, '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': 0.9850781559944153, 'rewards/repeat_soft/std': 0.010231668129563332, 'rewards/judge_quality/mean': 0.42000001668930054, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/total_composite/mean': 0.3325834274291992, 'rewards/total_composite/std': 0.20612144470214844, 'reward': 0.3325834274291992, 'reward_std': 0.20612144470214844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24013565480709076, 'sampling/sampling_logp_difference/max': 3.8115100860595703, 'sampling/importance_sampling_ratio/min': 0.02211475744843483, 'sampling/importance_sampling_ratio/mean': 0.9981706142425537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4976738691329956, 'clip_ratio/low_mean': 0.14025673642754555, 'clip_ratio/low_min': 0.14025673642754555, 'clip_ratio/high_mean': 0.0714642284438014, 'clip_ratio/high_max': 0.0714642284438014, 'clip_ratio/region_mean': 0.21172096487134695, 'reward_total_mean': 0.3325834274291992, 'reward_meter_mean': 0.6374300718307495, 'reward_meter_std': 0.2568950951099396, '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': 0.9850781559944153, 'reward_repeat_soft_std': 0.010231668129563332, 'reward_judge_quality_mean': 0.42000001668930054, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.3325834274291992, 'reward_total_composite_std': 0.20612144470214844, 'epoch': 0.01}

  2%|▏         | 69/3300 [08:36<6:09:15,  6.86s/it]INFO 04-13 12:54:19 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 70/3300 [08:43<6:06:30,  6.81s/it]2026-04-13 12:54:25,658 | INFO | train_grpo_train | metrics_logged mode=train step=70

                                                   
{'loss': -0.0385, 'grad_norm': 18.971607208251953, 'learning_rate': 9.790909090909093e-06, 'num_tokens': 129347.0, 'completions/mean_length': 68.625, 'completions/min_length': 49.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.18168789148330688, 'rewards/meter/std': 0.11182212084531784, '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': 0.9934610724449158, 'rewards/repeat_soft/std': 0.005277810152620077, 'rewards/judge_quality/mean': 0.23375000059604645, 'rewards/judge_quality/std': 0.025599943473935127, 'rewards/total_composite/mean': 0.09653282165527344, 'rewards/total_composite/std': 0.040975313633680344, 'reward': 0.09653282165527344, 'reward_std': 0.040975309908390045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27443817257881165, 'sampling/sampling_logp_difference/max': 1.9071316719055176, 'sampling/importance_sampling_ratio/min': 0.18012750148773193, 'sampling/importance_sampling_ratio/mean': 1.0374891757965088, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7137632966041565, 'clip_ratio/low_mean': 0.05093393288552761, 'clip_ratio/low_min': 0.05093393288552761, 'clip_ratio/high_mean': 0.1748208962380886, 'clip_ratio/high_max': 0.1748208962380886, 'clip_ratio/region_mean': 0.22575482912361622, 'reward_total_mean': 0.09653282165527344, 'reward_meter_mean': 0.18168789148330688, 'reward_meter_std': 0.11182212084531784, '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': 0.9934610724449158, 'reward_repeat_soft_std': 0.005277810152620077, 'reward_judge_quality_mean': 0.23375000059604645, 'reward_judge_quality_std': 0.025599943473935127, 'reward_total_composite_mean': 0.09653282165527344, 'reward_total_composite_std': 0.040975313633680344, 'epoch': 0.01}

  2%|▏         | 70/3300 [08:43<6:06:30,  6.81s/it]INFO 04-13 12:54:25 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 71/3300 [08:49<6:01:06,  6.71s/it]2026-04-13 12:54:32,138 | INFO | train_grpo_train | metrics_logged mode=train step=71

                                                   
{'loss': 0.4253, 'grad_norm': 27.757349014282227, 'learning_rate': 9.787878787878788e-06, 'num_tokens': 130972.0, 'completions/mean_length': 31.125, 'completions/min_length': 25.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9012611508369446, 'rewards/meter/std': 0.2112041562795639, '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': 0.9959396719932556, 'rewards/repeat_soft/std': 0.0019785345066338778, 'rewards/judge_quality/mean': 0.7450000047683716, 'rewards/judge_quality/std': 0.2121320217847824, 'rewards/total_composite/mean': 0.7219642400741577, 'rewards/total_composite/std': 0.23895327746868134, 'reward': 0.7219642400741577, 'reward_std': 0.23895327746868134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1198556199669838, 'sampling/sampling_logp_difference/max': 2.6911134719848633, 'sampling/importance_sampling_ratio/min': 0.06780540198087692, 'sampling/importance_sampling_ratio/mean': 1.0073957443237305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5624960474669933, 'clip_ratio/low_mean': 0.03030303120613098, 'clip_ratio/low_min': 0.03030303120613098, 'clip_ratio/high_mean': 0.06262362655252218, 'clip_ratio/high_max': 0.06262362655252218, 'clip_ratio/region_mean': 0.09292665775865316, 'reward_total_mean': 0.7219642400741577, 'reward_meter_mean': 0.9012611508369446, 'reward_meter_std': 0.2112041562795639, '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': 0.9959396719932556, 'reward_repeat_soft_std': 0.0019785345066338778, 'reward_judge_quality_mean': 0.7450000047683716, 'reward_judge_quality_std': 0.2121320217847824, 'reward_total_composite_mean': 0.7219642400741577, 'reward_total_composite_std': 0.23895327746868134, 'epoch': 0.01}

  2%|▏         | 71/3300 [08:49<6:01:06,  6.71s/it]INFO 04-13 12:54:32 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 72/3300 [08:56<5:58:09,  6.66s/it]2026-04-13 12:54:38,672 | INFO | train_grpo_train | metrics_logged mode=train step=72

                                                   
{'loss': 0.099, 'grad_norm': 15.447766304016113, 'learning_rate': 9.784848484848486e-06, 'num_tokens': 132900.0, 'completions/mean_length': 79.0, 'completions/min_length': 73.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.3447451591491699, 'rewards/meter/std': 0.28595930337905884, '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': 0.9931455850601196, 'rewards/repeat_soft/std': 0.003910026978701353, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.09441548585891724, 'rewards/total_composite/mean': 0.18368127942085266, 'rewards/total_composite/std': 0.09182135760784149, 'reward': 0.18368127942085266, 'reward_std': 0.0918213501572609, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22150319814682007, 'sampling/sampling_logp_difference/max': 3.541834831237793, 'sampling/importance_sampling_ratio/min': 0.02896014228463173, 'sampling/importance_sampling_ratio/mean': 0.9910763502120972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.12739023193717, 'clip_ratio/low_mean': 0.07343300711363554, 'clip_ratio/low_min': 0.07343300711363554, 'clip_ratio/high_mean': 0.08443638868629932, 'clip_ratio/high_max': 0.08443638868629932, 'clip_ratio/region_mean': 0.15786939579993486, 'reward_total_mean': 0.18368127942085266, 'reward_meter_mean': 0.3447451591491699, 'reward_meter_std': 0.28595930337905884, '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': 0.9931455850601196, 'reward_repeat_soft_std': 0.003910026978701353, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.09441548585891724, 'reward_total_composite_mean': 0.18368127942085266, 'reward_total_composite_std': 0.09182135760784149, 'epoch': 0.01}

  2%|▏         | 72/3300 [08:56<5:58:09,  6.66s/it]INFO 04-13 12:54:38 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 73/3300 [09:02<5:52:32,  6.55s/it]2026-04-13 12:54:44,989 | INFO | train_grpo_train | metrics_logged mode=train step=73

                                                   
{'loss': 0.1001, 'grad_norm': 16.453813552856445, 'learning_rate': 9.781818181818183e-06, 'num_tokens': 134478.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6412177681922913, 'rewards/meter/std': 0.36747169494628906, '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': 0.9885964393615723, 'rewards/repeat_soft/std': 0.02141972817480564, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.18715444207191467, 'rewards/total_composite/mean': 0.33021342754364014, 'rewards/total_composite/std': 0.21703536808490753, 'reward': 0.33021342754364014, 'reward_std': 0.21703538298606873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24010418355464935, 'sampling/sampling_logp_difference/max': 1.873922348022461, 'sampling/importance_sampling_ratio/min': 0.15352031588554382, 'sampling/importance_sampling_ratio/mean': 1.0277440547943115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0719871819019318, 'clip_ratio/low_mean': 0.07103462051600218, 'clip_ratio/low_min': 0.07103462051600218, 'clip_ratio/high_mean': 0.0977333365008235, 'clip_ratio/high_max': 0.0977333365008235, 'clip_ratio/region_mean': 0.16876795701682568, 'reward_total_mean': 0.33021342754364014, 'reward_meter_mean': 0.6412177681922913, 'reward_meter_std': 0.36747169494628906, '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': 0.9885964393615723, 'reward_repeat_soft_std': 0.02141972817480564, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.18715444207191467, 'reward_total_composite_mean': 0.33021342754364014, 'reward_total_composite_std': 0.21703536808490753, 'epoch': 0.01}

  2%|▏         | 73/3300 [09:02<5:52:32,  6.55s/it]INFO 04-13 12:54:45 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 74/3300 [09:08<5:47:16,  6.46s/it]2026-04-13 12:54:51,224 | INFO | train_grpo_train | metrics_logged mode=train step=74

                                                   
{'loss': -0.0896, 'grad_norm': 11.147936820983887, 'learning_rate': 9.77878787878788e-06, 'num_tokens': 136049.0, 'completions/mean_length': 50.375, 'completions/min_length': 24.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7489872574806213, 'rewards/meter/std': 0.250081866979599, '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': 0.9957634210586548, 'rewards/repeat_soft/std': 0.005310617852956057, 'rewards/judge_quality/mean': 0.6812499761581421, 'rewards/judge_quality/std': 0.2307402789592743, 'rewards/total_composite/mean': 0.557012677192688, 'rewards/total_composite/std': 0.20911905169487, 'reward': 0.557012677192688, 'reward_std': 0.2091190665960312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13287407159805298, 'sampling/sampling_logp_difference/max': 1.891297698020935, 'sampling/importance_sampling_ratio/min': 0.15087589621543884, 'sampling/importance_sampling_ratio/mean': 1.0008474588394165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.803953543305397, 'clip_ratio/low_mean': 0.08880797028541565, 'clip_ratio/low_min': 0.08880797028541565, 'clip_ratio/high_mean': 0.05626888386905193, 'clip_ratio/high_max': 0.05626888386905193, 'clip_ratio/region_mean': 0.14507685415446758, 'reward_total_mean': 0.557012677192688, 'reward_meter_mean': 0.7489872574806213, 'reward_meter_std': 0.250081866979599, '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': 0.9957634210586548, 'reward_repeat_soft_std': 0.005310617852956057, 'reward_judge_quality_mean': 0.6812499761581421, 'reward_judge_quality_std': 0.2307402789592743, 'reward_total_composite_mean': 0.557012677192688, 'reward_total_composite_std': 0.20911905169487, 'epoch': 0.01}

  2%|▏         | 74/3300 [09:08<5:47:16,  6.46s/it]INFO 04-13 12:54:51 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 75/3300 [09:15<5:52:30,  6.56s/it]2026-04-13 12:54:58,014 | INFO | train_grpo_train | metrics_logged mode=train step=75

                                                   
{'loss': -0.005, 'grad_norm': 11.774303436279297, 'learning_rate': 9.775757575757576e-06, 'num_tokens': 138121.0, 'completions/mean_length': 85.0, 'completions/min_length': 66.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.7358171939849854, 'rewards/meter/std': 0.3659661114215851, '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': 0.9967468976974487, 'rewards/repeat_soft/std': 0.004990824498236179, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.13427051901817322, 'rewards/total_composite/mean': 0.26152580976486206, 'rewards/total_composite/std': 0.10252867639064789, 'reward': 0.26152580976486206, 'reward_std': 0.10252867639064789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23619340360164642, 'sampling/sampling_logp_difference/max': 1.3284320831298828, 'sampling/importance_sampling_ratio/min': 0.2648922801017761, 'sampling/importance_sampling_ratio/mean': 1.0346297025680542, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.883807897567749, 'clip_ratio/low_mean': 0.04873333126306534, 'clip_ratio/low_min': 0.04873333126306534, 'clip_ratio/high_mean': 0.14055721275508404, 'clip_ratio/high_max': 0.14055721275508404, 'clip_ratio/region_mean': 0.18929054401814938, 'reward_total_mean': 0.26152580976486206, 'reward_meter_mean': 0.7358171939849854, 'reward_meter_std': 0.3659661114215851, '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': 0.9967468976974487, 'reward_repeat_soft_std': 0.004990824498236179, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.13427051901817322, 'reward_total_composite_mean': 0.26152580976486206, 'reward_total_composite_std': 0.10252867639064789, 'epoch': 0.01}

  2%|▏         | 75/3300 [09:15<5:52:30,  6.56s/it]INFO 04-13 12:54:58 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 76/3300 [09:21<5:41:31,  6.36s/it]2026-04-13 12:55:03,897 | INFO | train_grpo_train | metrics_logged mode=train step=76

                                                   
{'loss': 0.1574, 'grad_norm': 25.992359161376953, 'learning_rate': 9.772727272727273e-06, 'num_tokens': 139528.0, 'completions/mean_length': 21.875, 'completions/min_length': 17.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6581739187240601, 'rewards/meter/std': 0.44663143157958984, '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': 0.9404481053352356, 'rewards/repeat_soft/std': 0.0446208231151104, 'rewards/judge_quality/mean': 0.3424999713897705, 'rewards/judge_quality/std': 0.20190167427062988, 'rewards/total_composite/mean': 0.24099212884902954, 'rewards/total_composite/std': 0.09951608628034592, 'reward': 0.24099212884902954, 'reward_std': 0.09951609373092651, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17727313935756683, 'sampling/sampling_logp_difference/max': 0.8434514999389648, 'sampling/importance_sampling_ratio/min': 0.4302230179309845, 'sampling/importance_sampling_ratio/mean': 1.0560768842697144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2954431772232056, 'clip_ratio/low_mean': 0.051764706149697304, 'clip_ratio/low_min': 0.051764706149697304, 'clip_ratio/high_mean': 0.11540627013891935, 'clip_ratio/high_max': 0.11540627013891935, 'clip_ratio/region_mean': 0.16717097628861666, 'reward_total_mean': 0.24099212884902954, 'reward_meter_mean': 0.6581739187240601, 'reward_meter_std': 0.44663143157958984, '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': 0.9404481053352356, 'reward_repeat_soft_std': 0.0446208231151104, 'reward_judge_quality_mean': 0.3424999713897705, 'reward_judge_quality_std': 0.20190167427062988, 'reward_total_composite_mean': 0.24099212884902954, 'reward_total_composite_std': 0.09951608628034592, 'epoch': 0.01}

  2%|▏         | 76/3300 [09:21<5:41:31,  6.36s/it]INFO 04-13 12:55:04 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 77/3300 [09:27<5:39:45,  6.33s/it]2026-04-13 12:55:10,150 | INFO | train_grpo_train | metrics_logged mode=train step=77

                                                   
{'loss': 0.1504, 'grad_norm': 17.336942672729492, 'learning_rate': 9.76969696969697e-06, 'num_tokens': 141097.0, 'completions/mean_length': 40.125, 'completions/min_length': 29.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.5049496293067932, 'rewards/meter/std': 0.3687618672847748, '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': 0.9967241883277893, 'rewards/repeat_soft/std': 0.004613413475453854, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.1629198044538498, 'rewards/total_composite/mean': 0.2808726131916046, 'rewards/total_composite/std': 0.2017921358346939, 'reward': 0.2808726131916046, 'reward_std': 0.2017921358346939, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1849796026945114, 'sampling/sampling_logp_difference/max': 1.5303640365600586, 'sampling/importance_sampling_ratio/min': 0.21645687520503998, 'sampling/importance_sampling_ratio/mean': 0.9946142435073853, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4863542169332504, 'clip_ratio/low_mean': 0.11449978314340115, 'clip_ratio/low_min': 0.11449978314340115, 'clip_ratio/high_mean': 0.0906324926763773, 'clip_ratio/high_max': 0.0906324926763773, 'clip_ratio/region_mean': 0.20513227581977844, 'reward_total_mean': 0.2808726131916046, 'reward_meter_mean': 0.5049496293067932, 'reward_meter_std': 0.3687618672847748, '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': 0.9967241883277893, 'reward_repeat_soft_std': 0.004613413475453854, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.1629198044538498, 'reward_total_composite_mean': 0.2808726131916046, 'reward_total_composite_std': 0.2017921358346939, 'epoch': 0.01}

  2%|▏         | 77/3300 [09:27<5:39:45,  6.33s/it]INFO 04-13 12:55:10 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 78/3300 [09:34<5:49:38,  6.51s/it]2026-04-13 12:55:17,096 | INFO | train_grpo_train | metrics_logged mode=train step=78

                                                   
{'loss': 0.068, 'grad_norm': 11.861577987670898, 'learning_rate': 9.766666666666667e-06, 'num_tokens': 143395.0, 'completions/mean_length': 92.25, 'completions/min_length': 61.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.7410093545913696, 'rewards/meter/std': 0.14935754239559174, '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': 0.9929521083831787, 'rewards/repeat_soft/std': 0.010279295966029167, 'rewards/judge_quality/mean': 0.3974999785423279, 'rewards/judge_quality/std': 0.179344043135643, 'rewards/total_composite/mean': 0.33023709058761597, 'rewards/total_composite/std': 0.16106951236724854, 'reward': 0.33023709058761597, 'reward_std': 0.16106951236724854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22217519581317902, 'sampling/sampling_logp_difference/max': 2.2308349609375, 'sampling/importance_sampling_ratio/min': 0.10743868350982666, 'sampling/importance_sampling_ratio/mean': 1.0363174676895142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6303199976682663, 'clip_ratio/low_mean': 0.11399947199970484, 'clip_ratio/low_min': 0.11399947199970484, 'clip_ratio/high_mean': 0.09098470397293568, 'clip_ratio/high_max': 0.09098470397293568, 'clip_ratio/region_mean': 0.20498417597264051, 'reward_total_mean': 0.33023709058761597, 'reward_meter_mean': 0.7410093545913696, 'reward_meter_std': 0.14935754239559174, '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': 0.9929521083831787, 'reward_repeat_soft_std': 0.010279295966029167, 'reward_judge_quality_mean': 0.3974999785423279, 'reward_judge_quality_std': 0.179344043135643, 'reward_total_composite_mean': 0.33023709058761597, 'reward_total_composite_std': 0.16106951236724854, 'epoch': 0.01}

  2%|▏         | 78/3300 [09:34<5:49:38,  6.51s/it]INFO 04-13 12:55:17 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 79/3300 [09:41<5:49:47,  6.52s/it]2026-04-13 12:55:23,623 | INFO | train_grpo_train | metrics_logged mode=train step=79

                                                   
{'loss': -0.0253, 'grad_norm': 13.1495943069458, 'learning_rate': 9.763636363636365e-06, 'num_tokens': 145468.0, 'completions/mean_length': 82.125, 'completions/min_length': 55.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.29229092597961426, 'rewards/meter/std': 0.2564309537410736, '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': 0.9963474273681641, 'rewards/repeat_soft/std': 0.0026311695110052824, 'rewards/judge_quality/mean': 0.22374999523162842, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.10872744768857956, 'rewards/total_composite/std': 0.05904335901141167, 'reward': 0.10872744768857956, 'reward_std': 0.05904335528612137, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20582804083824158, 'sampling/sampling_logp_difference/max': 1.7601685523986816, 'sampling/importance_sampling_ratio/min': 0.17201586067676544, 'sampling/importance_sampling_ratio/mean': 1.030929684638977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0478455275297165, 'clip_ratio/low_mean': 0.13159164413809776, 'clip_ratio/low_min': 0.13159164413809776, 'clip_ratio/high_mean': 0.09389667399227619, 'clip_ratio/high_max': 0.09389667399227619, 'clip_ratio/region_mean': 0.22548831813037395, 'reward_total_mean': 0.10872744768857956, 'reward_meter_mean': 0.29229092597961426, 'reward_meter_std': 0.2564309537410736, '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': 0.9963474273681641, 'reward_repeat_soft_std': 0.0026311695110052824, 'reward_judge_quality_mean': 0.22374999523162842, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.10872744768857956, 'reward_total_composite_std': 0.05904335901141167, 'epoch': 0.01}

  2%|▏         | 79/3300 [09:41<5:49:47,  6.52s/it]INFO 04-13 12:55:23 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 80/3300 [09:47<5:48:57,  6.50s/it]2026-04-13 12:55:30,093 | INFO | train_grpo_train | metrics_logged mode=train step=80

                                                   
{'loss': 0.0505, 'grad_norm': 14.630961418151855, 'learning_rate': 9.760606060606062e-06, 'num_tokens': 147164.0, 'completions/mean_length': 43.0, 'completions/min_length': 29.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8201456665992737, 'rewards/meter/std': 0.3403693437576294, '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': 0.9994844198226929, 'rewards/repeat_soft/std': 0.0014582598814740777, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.14302222430706024, 'rewards/total_composite/mean': 0.2373991310596466, 'rewards/total_composite/std': 0.17332343757152557, 'reward': 0.2373991310596466, 'reward_std': 0.17332343757152557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22331182658672333, 'sampling/sampling_logp_difference/max': 1.3764514923095703, 'sampling/importance_sampling_ratio/min': 0.2524728775024414, 'sampling/importance_sampling_ratio/mean': 1.042371392250061, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.39580474793911, 'clip_ratio/low_mean': 0.1384577490389347, 'clip_ratio/low_min': 0.1384577490389347, 'clip_ratio/high_mean': 0.09441963024437428, 'clip_ratio/high_max': 0.09441963024437428, 'clip_ratio/region_mean': 0.23287737928330898, 'reward_total_mean': 0.2373991310596466, 'reward_meter_mean': 0.8201456665992737, 'reward_meter_std': 0.3403693437576294, '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': 0.9994844198226929, 'reward_repeat_soft_std': 0.0014582598814740777, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.14302222430706024, 'reward_total_composite_mean': 0.2373991310596466, 'reward_total_composite_std': 0.17332343757152557, 'epoch': 0.01}

  2%|▏         | 80/3300 [09:47<5:48:57,  6.50s/it]INFO 04-13 12:55:30 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 81/3300 [09:54<5:55:55,  6.63s/it]2026-04-13 12:55:37,037 | INFO | train_grpo_train | metrics_logged mode=train step=81

                                                   
{'loss': 0.043, 'grad_norm': 11.774923324584961, 'learning_rate': 9.757575757575758e-06, 'num_tokens': 149486.0, 'completions/mean_length': 80.25, 'completions/min_length': 72.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.9092499017715454, 'rewards/meter/std': 0.14935287833213806, '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': 0.9916456341743469, 'rewards/repeat_soft/std': 0.00773578928783536, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09227287769317627, 'rewards/total_composite/mean': 0.3116473853588104, 'rewards/total_composite/std': 0.08176867663860321, 'reward': 0.3116473853588104, 'reward_std': 0.08176867663860321, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2346639186143875, 'sampling/sampling_logp_difference/max': 1.6828041076660156, 'sampling/importance_sampling_ratio/min': 0.23151129484176636, 'sampling/importance_sampling_ratio/mean': 1.060604214668274, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0485707223415375, 'clip_ratio/low_mean': 0.13378658518195152, 'clip_ratio/low_min': 0.13378658518195152, 'clip_ratio/high_mean': 0.09180899523198605, 'clip_ratio/high_max': 0.09180899523198605, 'clip_ratio/region_mean': 0.22559558041393757, 'reward_total_mean': 0.3116473853588104, 'reward_meter_mean': 0.9092499017715454, 'reward_meter_std': 0.14935287833213806, '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': 0.9916456341743469, 'reward_repeat_soft_std': 0.00773578928783536, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09227287769317627, 'reward_total_composite_mean': 0.3116473853588104, 'reward_total_composite_std': 0.08176867663860321, 'epoch': 0.01}

  2%|▏         | 81/3300 [09:54<5:55:55,  6.63s/it]INFO 04-13 12:55:37 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 82/3300 [10:00<5:51:11,  6.55s/it]2026-04-13 12:55:43,383 | INFO | train_grpo_train | metrics_logged mode=train step=82

                                                   
{'loss': 0.1046, 'grad_norm': 18.41388511657715, 'learning_rate': 9.754545454545455e-06, 'num_tokens': 150996.0, 'completions/mean_length': 39.75, 'completions/min_length': 31.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.3590666651725769, 'rewards/meter/std': 0.34565114974975586, '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': 0.9989273548126221, 'rewards/repeat_soft/std': 0.002356387674808502, 'rewards/judge_quality/mean': 0.22374999523162842, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.13183966279029846, 'rewards/total_composite/std': 0.05637011304497719, 'reward': 0.13183966279029846, 'reward_std': 0.05637011677026749, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2514374256134033, 'sampling/sampling_logp_difference/max': 1.5833864212036133, 'sampling/importance_sampling_ratio/min': 0.20527875423431396, 'sampling/importance_sampling_ratio/mean': 1.02720308303833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.521538481116295, 'clip_ratio/low_mean': 0.07123085763305426, 'clip_ratio/low_min': 0.07123085763305426, 'clip_ratio/high_mean': 0.11304819025099277, 'clip_ratio/high_max': 0.11304819025099277, 'clip_ratio/region_mean': 0.18427904788404703, 'reward_total_mean': 0.13183966279029846, 'reward_meter_mean': 0.3590666651725769, 'reward_meter_std': 0.34565114974975586, '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': 0.9989273548126221, 'reward_repeat_soft_std': 0.002356387674808502, 'reward_judge_quality_mean': 0.22374999523162842, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.13183966279029846, 'reward_total_composite_std': 0.05637011304497719, 'epoch': 0.01}

  2%|▏         | 82/3300 [10:00<5:51:11,  6.55s/it]INFO 04-13 12:55:43 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 83/3300 [10:07<5:45:27,  6.44s/it]2026-04-13 12:55:49,582 | INFO | train_grpo_train | metrics_logged mode=train step=83

                                                   
{'loss': 0.1639, 'grad_norm': 14.891986846923828, 'learning_rate': 9.751515151515152e-06, 'num_tokens': 152512.0, 'completions/mean_length': 44.5, 'completions/min_length': 35.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.760381817817688, 'rewards/meter/std': 0.31577998399734497, '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': 0.9976599216461182, 'rewards/repeat_soft/std': 0.004242512863129377, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.018850916996598244, 'rewards/total_composite/mean': 0.18250644207000732, 'rewards/total_composite/std': 0.04083961620926857, 'reward': 0.18250644207000732, 'reward_std': 0.04083961620926857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18133598566055298, 'sampling/sampling_logp_difference/max': 1.790323257446289, 'sampling/importance_sampling_ratio/min': 0.1669062077999115, 'sampling/importance_sampling_ratio/mean': 1.0080831050872803, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5756563022732735, 'clip_ratio/low_mean': 0.059422025457024574, 'clip_ratio/low_min': 0.059422025457024574, 'clip_ratio/high_mean': 0.09980646055191755, 'clip_ratio/high_max': 0.09980646055191755, 'clip_ratio/region_mean': 0.15922848600894213, 'reward_total_mean': 0.18250644207000732, 'reward_meter_mean': 0.760381817817688, 'reward_meter_std': 0.31577998399734497, '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': 0.9976599216461182, 'reward_repeat_soft_std': 0.004242512863129377, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.018850916996598244, 'reward_total_composite_mean': 0.18250644207000732, 'reward_total_composite_std': 0.04083961620926857, 'epoch': 0.01}

  3%|▎         | 83/3300 [10:07<5:45:27,  6.44s/it]INFO 04-13 12:55:49 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 84/3300 [10:13<5:43:05,  6.40s/it]2026-04-13 12:55:55,884 | INFO | train_grpo_train | metrics_logged mode=train step=84

                                                   
{'loss': 0.0318, 'grad_norm': 18.084400177001953, 'learning_rate': 9.74848484848485e-06, 'num_tokens': 154254.0, 'completions/mean_length': 36.75, 'completions/min_length': 33.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.4844000041484833, 'rewards/meter/std': 0.40570446848869324, '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': 0.9953454732894897, 'rewards/repeat_soft/std': 0.006105078384280205, 'rewards/judge_quality/mean': 0.26999998092651367, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/total_composite/mean': 0.1721809208393097, 'rewards/total_composite/std': 0.08328097313642502, 'reward': 0.1721809208393097, 'reward_std': 0.08328097313642502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24323490262031555, 'sampling/sampling_logp_difference/max': 2.1036877632141113, 'sampling/importance_sampling_ratio/min': 0.12200567126274109, 'sampling/importance_sampling_ratio/mean': 1.0523042678833008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.703775256872177, 'clip_ratio/low_mean': 0.12026222702115774, 'clip_ratio/low_min': 0.12026222702115774, 'clip_ratio/high_mean': 0.09044239670038223, 'clip_ratio/high_max': 0.09044239670038223, 'clip_ratio/region_mean': 0.21070462372153997, 'reward_total_mean': 0.1721809208393097, 'reward_meter_mean': 0.4844000041484833, 'reward_meter_std': 0.40570446848869324, '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': 0.9953454732894897, 'reward_repeat_soft_std': 0.006105078384280205, 'reward_judge_quality_mean': 0.26999998092651367, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.1721809208393097, 'reward_total_composite_std': 0.08328097313642502, 'epoch': 0.01}

  3%|▎         | 84/3300 [10:13<5:43:05,  6.40s/it]INFO 04-13 12:55:56 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 85/3300 [10:19<5:32:59,  6.21s/it]2026-04-13 12:56:01,663 | INFO | train_grpo_train | metrics_logged mode=train step=85

                                                   
{'loss': 0.1219, 'grad_norm': 30.80690574645996, 'learning_rate': 9.745454545454547e-06, 'num_tokens': 155560.0, 'completions/mean_length': 18.25, 'completions/min_length': 14.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.25, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7818843722343445, 'rewards/meter/std': 0.324420303106308, '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': 0.9512310028076172, 'rewards/repeat_soft/std': 0.022036533802747726, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.29293006658554077, 'rewards/total_composite/std': 0.12112832814455032, 'reward': 0.29293006658554077, 'reward_std': 0.12112832069396973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2320132553577423, 'sampling/sampling_logp_difference/max': 1.6422712802886963, 'sampling/importance_sampling_ratio/min': 0.19353996217250824, 'sampling/importance_sampling_ratio/mean': 1.038288950920105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9580261558294296, 'clip_ratio/low_mean': 0.07638351526111364, 'clip_ratio/low_min': 0.07638351526111364, 'clip_ratio/high_mean': 0.10431678220629692, 'clip_ratio/high_max': 0.10431678220629692, 'clip_ratio/region_mean': 0.18070029746741056, 'reward_total_mean': 0.29293006658554077, 'reward_meter_mean': 0.7818843722343445, 'reward_meter_std': 0.324420303106308, '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': 0.9512310028076172, 'reward_repeat_soft_std': 0.022036533802747726, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.29293006658554077, 'reward_total_composite_std': 0.12112832814455032, 'epoch': 0.01}

  3%|▎         | 85/3300 [10:19<5:32:59,  6.21s/it]INFO 04-13 12:56:01 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 86/3300 [10:25<5:37:34,  6.30s/it]2026-04-13 12:56:08,171 | INFO | train_grpo_train | metrics_logged mode=train step=86

                                                   
{'loss': 0.0126, 'grad_norm': 16.33660888671875, 'learning_rate': 9.742424242424244e-06, 'num_tokens': 157581.0, 'completions/mean_length': 73.625, 'completions/min_length': 66.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.3446776866912842, 'rewards/meter/std': 0.27230140566825867, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9970155954360962, 'rewards/repeat_soft/std': 0.002982294885441661, 'rewards/judge_quality/mean': 0.2150000035762787, 'rewards/judge_quality/std': 0.014142133295536041, 'rewards/total_composite/mean': 0.09810009598731995, 'rewards/total_composite/std': 0.06941568851470947, 'reward': 0.09810009598731995, 'reward_std': 0.06941568106412888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2284555733203888, 'sampling/sampling_logp_difference/max': 2.1358413696289062, 'sampling/importance_sampling_ratio/min': 0.11814513802528381, 'sampling/importance_sampling_ratio/mean': 0.9742966294288635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.226352721452713, 'clip_ratio/low_mean': 0.08101782202720642, 'clip_ratio/low_min': 0.08101782202720642, 'clip_ratio/high_mean': 0.10740447510033846, 'clip_ratio/high_max': 0.10740447510033846, 'clip_ratio/region_mean': 0.18842229712754488, 'reward_total_mean': 0.09810009598731995, 'reward_meter_mean': 0.3446776866912842, 'reward_meter_std': 0.27230140566825867, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9970155954360962, 'reward_repeat_soft_std': 0.002982294885441661, 'reward_judge_quality_mean': 0.2150000035762787, 'reward_judge_quality_std': 0.014142133295536041, 'reward_total_composite_mean': 0.09810009598731995, 'reward_total_composite_std': 0.06941568851470947, 'epoch': 0.01}

  3%|▎         | 86/3300 [10:25<5:37:34,  6.30s/it]INFO 04-13 12:56:08 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 87/3300 [10:31<5:28:19,  6.13s/it]2026-04-13 12:56:13,903 | INFO | train_grpo_train | metrics_logged mode=train step=87

                                                   
{'loss': 0.0745, 'grad_norm': 29.98630142211914, 'learning_rate': 9.739393939393941e-06, 'num_tokens': 158871.0, 'completions/mean_length': 20.25, 'completions/min_length': 14.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5150893330574036, 'rewards/meter/std': 0.4621303379535675, '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': 0.9594135880470276, 'rewards/repeat_soft/std': 0.005897748749703169, 'rewards/judge_quality/mean': 0.2512499988079071, 'rewards/judge_quality/std': 0.07623975723981857, 'rewards/total_composite/mean': 0.16783645749092102, 'rewards/total_composite/std': 0.12525562942028046, 'reward': 0.16783645749092102, 'reward_std': 0.12525562942028046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2461036890745163, 'sampling/sampling_logp_difference/max': 2.1723361015319824, 'sampling/importance_sampling_ratio/min': 0.1139112040400505, 'sampling/importance_sampling_ratio/mean': 1.0283335447311401, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.04075089097023, 'clip_ratio/low_mean': 0.10652709566056728, 'clip_ratio/low_min': 0.10652709566056728, 'clip_ratio/high_mean': 0.10458851745352149, 'clip_ratio/high_max': 0.10458851745352149, 'clip_ratio/region_mean': 0.21111561311408877, 'reward_total_mean': 0.16783645749092102, 'reward_meter_mean': 0.5150893330574036, 'reward_meter_std': 0.4621303379535675, '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': 0.9594135880470276, 'reward_repeat_soft_std': 0.005897748749703169, 'reward_judge_quality_mean': 0.2512499988079071, 'reward_judge_quality_std': 0.07623975723981857, 'reward_total_composite_mean': 0.16783645749092102, 'reward_total_composite_std': 0.12525562942028046, 'epoch': 0.01}

  3%|▎         | 87/3300 [10:31<5:28:19,  6.13s/it]INFO 04-13 12:56:14 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 88/3300 [10:38<5:35:54,  6.27s/it]2026-04-13 12:56:20,513 | INFO | train_grpo_train | metrics_logged mode=train step=88

                                                   
{'loss': 0.0042, 'grad_norm': 17.625898361206055, 'learning_rate': 9.736363636363637e-06, 'num_tokens': 160484.0, 'completions/mean_length': 40.625, 'completions/min_length': 34.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.25569644570350647, 'rewards/meter/std': 0.21334248781204224, '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': 0.9973743557929993, 'rewards/repeat_soft/std': 0.004422957077622414, 'rewards/judge_quality/mean': 0.22624999284744263, 'rewards/judge_quality/std': 0.02386719174683094, 'rewards/total_composite/mean': 0.11431076377630234, 'rewards/total_composite/std': 0.022492822259664536, 'reward': 0.11431076377630234, 'reward_std': 0.022492822259664536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20674750208854675, 'sampling/sampling_logp_difference/max': 1.886871337890625, 'sampling/importance_sampling_ratio/min': 0.1515451967716217, 'sampling/importance_sampling_ratio/mean': 1.022700548171997, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7788608074188232, 'clip_ratio/low_mean': 0.10440731886774302, 'clip_ratio/low_min': 0.10440731886774302, 'clip_ratio/high_mean': 0.07957121171057224, 'clip_ratio/high_max': 0.07957121171057224, 'clip_ratio/region_mean': 0.18397853057831526, 'reward_total_mean': 0.11431076377630234, 'reward_meter_mean': 0.25569644570350647, 'reward_meter_std': 0.21334248781204224, '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': 0.9973743557929993, 'reward_repeat_soft_std': 0.004422957077622414, 'reward_judge_quality_mean': 0.22624999284744263, 'reward_judge_quality_std': 0.02386719174683094, 'reward_total_composite_mean': 0.11431076377630234, 'reward_total_composite_std': 0.022492822259664536, 'epoch': 0.01}

  3%|▎         | 88/3300 [10:38<5:35:54,  6.27s/it]INFO 04-13 12:56:20 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 89/3300 [10:45<5:46:04,  6.47s/it]2026-04-13 12:56:27,428 | INFO | train_grpo_train | metrics_logged mode=train step=89

                                                   
{'loss': 0.0473, 'grad_norm': 14.281908988952637, 'learning_rate': 9.733333333333334e-06, 'num_tokens': 162532.0, 'completions/mean_length': 91.0, 'completions/min_length': 67.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.5333120822906494, 'rewards/meter/std': 0.3111574351787567, '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': 0.9964537620544434, 'rewards/repeat_soft/std': 0.0021019114647060633, 'rewards/judge_quality/mean': 0.3487499952316284, 'rewards/judge_quality/std': 0.09876921027898788, 'rewards/total_composite/mean': 0.21525812149047852, 'rewards/total_composite/std': 0.12606896460056305, 'reward': 0.21525812149047852, 'reward_std': 0.12606896460056305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16212689876556396, 'sampling/sampling_logp_difference/max': 2.8756399154663086, 'sampling/importance_sampling_ratio/min': 0.056380048394203186, 'sampling/importance_sampling_ratio/mean': 1.006917953491211, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7011619098484516, 'clip_ratio/low_mean': 0.048041378147900105, 'clip_ratio/low_min': 0.048041378147900105, 'clip_ratio/high_mean': 0.08259881474077702, 'clip_ratio/high_max': 0.08259881474077702, 'clip_ratio/region_mean': 0.13064019288867712, 'reward_total_mean': 0.21525812149047852, 'reward_meter_mean': 0.5333120822906494, 'reward_meter_std': 0.3111574351787567, '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': 0.9964537620544434, 'reward_repeat_soft_std': 0.0021019114647060633, 'reward_judge_quality_mean': 0.3487499952316284, 'reward_judge_quality_std': 0.09876921027898788, 'reward_total_composite_mean': 0.21525812149047852, 'reward_total_composite_std': 0.12606896460056305, 'epoch': 0.01}

  3%|▎         | 89/3300 [10:45<5:46:04,  6.47s/it]INFO 04-13 12:56:27 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 90/3300 [10:50<5:35:00,  6.26s/it]2026-04-13 12:56:33,210 | INFO | train_grpo_train | metrics_logged mode=train step=90

                                                   
{'loss': -0.0018, 'grad_norm': 29.664228439331055, 'learning_rate': 9.730303030303031e-06, 'num_tokens': 163974.0, 'completions/mean_length': 23.25, 'completions/min_length': 21.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.25, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5339425802230835, 'rewards/meter/std': 0.3753529191017151, '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': 0.9940358996391296, 'rewards/repeat_soft/std': 0.00873249489814043, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.13845448195934296, 'rewards/total_composite/mean': 0.23664940893650055, 'rewards/total_composite/std': 0.14900387823581696, 'reward': 0.23664940893650055, 'reward_std': 0.14900384843349457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23893314599990845, 'sampling/sampling_logp_difference/max': 1.5105533599853516, 'sampling/importance_sampling_ratio/min': 0.22078776359558105, 'sampling/importance_sampling_ratio/mean': 0.9652258157730103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6168277114629745, 'clip_ratio/low_mean': 0.06684981845319271, 'clip_ratio/low_min': 0.06684981845319271, 'clip_ratio/high_mean': 0.1836034394800663, 'clip_ratio/high_max': 0.1836034394800663, 'clip_ratio/region_mean': 0.250453257933259, 'reward_total_mean': 0.23664940893650055, 'reward_meter_mean': 0.5339425802230835, 'reward_meter_std': 0.3753529191017151, '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': 0.9940358996391296, 'reward_repeat_soft_std': 0.00873249489814043, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.13845448195934296, 'reward_total_composite_mean': 0.23664940893650055, 'reward_total_composite_std': 0.14900387823581696, 'epoch': 0.01}

  3%|▎         | 90/3300 [10:50<5:35:00,  6.26s/it]INFO 04-13 12:56:33 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 91/3300 [10:57<5:41:43,  6.39s/it]2026-04-13 12:56:39,897 | INFO | train_grpo_train | metrics_logged mode=train step=91

                                                   
{'loss': 0.0441, 'grad_norm': 13.200789451599121, 'learning_rate': 9.727272727272728e-06, 'num_tokens': 165909.0, 'completions/mean_length': 68.875, 'completions/min_length': 59.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.5148859024047852, 'rewards/meter/std': 0.34434616565704346, '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': 0.9969649314880371, 'rewards/repeat_soft/std': 0.0031781643629074097, 'rewards/judge_quality/mean': 0.3199999928474426, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.197257399559021, 'rewards/total_composite/std': 0.14142344892024994, 'reward': 0.197257399559021, 'reward_std': 0.14142344892024994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24964481592178345, 'sampling/sampling_logp_difference/max': 2.125476837158203, 'sampling/importance_sampling_ratio/min': 0.1193760335445404, 'sampling/importance_sampling_ratio/mean': 1.0264556407928467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.708129793405533, 'clip_ratio/low_mean': 0.1285439468920231, 'clip_ratio/low_min': 0.1285439468920231, 'clip_ratio/high_mean': 0.0763765349984169, 'clip_ratio/high_max': 0.0763765349984169, 'clip_ratio/region_mean': 0.20492048189044, 'reward_total_mean': 0.197257399559021, 'reward_meter_mean': 0.5148859024047852, 'reward_meter_std': 0.34434616565704346, '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': 0.9969649314880371, 'reward_repeat_soft_std': 0.0031781643629074097, 'reward_judge_quality_mean': 0.3199999928474426, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.197257399559021, 'reward_total_composite_std': 0.14142344892024994, 'epoch': 0.01}

  3%|▎         | 91/3300 [10:57<5:41:43,  6.39s/it]INFO 04-13 12:56:40 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 92/3300 [11:03<5:40:58,  6.38s/it]2026-04-13 12:56:46,247 | INFO | train_grpo_train | metrics_logged mode=train step=92

                                                   
{'loss': -0.0133, 'grad_norm': 19.60318374633789, 'learning_rate': 9.724242424242426e-06, 'num_tokens': 167577.0, 'completions/mean_length': 40.5, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.2892237901687622, 'rewards/meter/std': 0.32355058193206787, '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': 0.9792777299880981, 'rewards/repeat_soft/std': 0.03071063570678234, 'rewards/judge_quality/mean': 0.32625001668930054, 'rewards/judge_quality/std': 0.14831794798374176, 'rewards/total_composite/mean': 0.15861031413078308, 'rewards/total_composite/std': 0.0482286661863327, 'reward': 0.15861031413078308, 'reward_std': 0.048228669911623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23212920129299164, 'sampling/sampling_logp_difference/max': 1.5110788345336914, 'sampling/importance_sampling_ratio/min': 0.22067178785800934, 'sampling/importance_sampling_ratio/mean': 1.0275473594665527, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.950747475028038, 'clip_ratio/low_mean': 0.13828057888895273, 'clip_ratio/low_min': 0.13828057888895273, 'clip_ratio/high_mean': 0.10061880759894848, 'clip_ratio/high_max': 0.10061880759894848, 'clip_ratio/region_mean': 0.2388993864879012, 'reward_total_mean': 0.15861031413078308, 'reward_meter_mean': 0.2892237901687622, 'reward_meter_std': 0.32355058193206787, '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': 0.9792777299880981, 'reward_repeat_soft_std': 0.03071063570678234, 'reward_judge_quality_mean': 0.32625001668930054, 'reward_judge_quality_std': 0.14831794798374176, 'reward_total_composite_mean': 0.15861031413078308, 'reward_total_composite_std': 0.0482286661863327, 'epoch': 0.01}

  3%|▎         | 92/3300 [11:03<5:40:58,  6.38s/it]INFO 04-13 12:56:46 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 93/3300 [11:10<5:41:45,  6.39s/it]2026-04-13 12:56:52,679 | INFO | train_grpo_train | metrics_logged mode=train step=93

                                                   
{'loss': -0.1215, 'grad_norm': 14.864767074584961, 'learning_rate': 9.721212121212123e-06, 'num_tokens': 169145.0, 'completions/mean_length': 49.0, 'completions/min_length': 38.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7403737306594849, 'rewards/meter/std': 0.40326493978500366, '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': 0.9993371367454529, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.24250000715255737, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.20033590495586395, 'rewards/total_composite/std': 0.061576951295137405, 'reward': 0.20033590495586395, 'reward_std': 0.061576951295137405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22364959120750427, 'sampling/sampling_logp_difference/max': 1.231515645980835, 'sampling/importance_sampling_ratio/min': 0.29184991121292114, 'sampling/importance_sampling_ratio/mean': 1.062864065170288, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.15771621465683, 'clip_ratio/low_mean': 0.0690789483487606, 'clip_ratio/low_min': 0.0690789483487606, 'clip_ratio/high_mean': 0.15948813129216433, 'clip_ratio/high_max': 0.15948813129216433, 'clip_ratio/region_mean': 0.22856707964092493, 'reward_total_mean': 0.20033590495586395, 'reward_meter_mean': 0.7403737306594849, 'reward_meter_std': 0.40326493978500366, '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': 0.9993371367454529, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.24250000715255737, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.20033590495586395, 'reward_total_composite_std': 0.061576951295137405, 'epoch': 0.01}

  3%|▎         | 93/3300 [11:10<5:41:45,  6.39s/it]INFO 04-13 12:56:52 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 94/3300 [11:16<5:38:15,  6.33s/it]2026-04-13 12:56:58,861 | INFO | train_grpo_train | metrics_logged mode=train step=94

                                                   
{'loss': 0.0733, 'grad_norm': 20.891071319580078, 'learning_rate': 9.718181818181818e-06, 'num_tokens': 170838.0, 'completions/mean_length': 33.625, 'completions/min_length': 31.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.40905338525772095, 'rewards/meter/std': 0.38333770632743835, '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': 0.9920638799667358, 'rewards/repeat_soft/std': 0.010272251442074776, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.22936499118804932, 'rewards/total_composite/std': 0.12850287556648254, 'reward': 0.22936499118804932, 'reward_std': 0.12850286066532135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23356930911540985, 'sampling/sampling_logp_difference/max': 1.6581082344055176, 'sampling/importance_sampling_ratio/min': 0.19049902260303497, 'sampling/importance_sampling_ratio/mean': 1.0124125480651855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.796781674027443, 'clip_ratio/low_mean': 0.13331238925457, 'clip_ratio/low_min': 0.13331238925457, 'clip_ratio/high_mean': 0.0598262045532465, 'clip_ratio/high_max': 0.0598262045532465, 'clip_ratio/region_mean': 0.1931385938078165, 'reward_total_mean': 0.22936499118804932, 'reward_meter_mean': 0.40905338525772095, 'reward_meter_std': 0.38333770632743835, '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': 0.9920638799667358, 'reward_repeat_soft_std': 0.010272251442074776, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.22936499118804932, 'reward_total_composite_std': 0.12850287556648254, 'epoch': 0.01}

  3%|▎         | 94/3300 [11:16<5:38:15,  6.33s/it]INFO 04-13 12:56:59 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 95/3300 [11:23<5:48:49,  6.53s/it]2026-04-13 12:57:05,859 | INFO | train_grpo_train | metrics_logged mode=train step=95

                                                   
{'loss': 0.2013, 'grad_norm': 10.570873260498047, 'learning_rate': 9.715151515151516e-06, 'num_tokens': 173021.0, 'completions/mean_length': 89.875, 'completions/min_length': 68.0, 'completions/max_length': 132.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.47886359691619873, 'rewards/meter/std': 0.29336100816726685, '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': 0.9957594275474548, 'rewards/repeat_soft/std': 0.0027065465692430735, 'rewards/judge_quality/mean': 0.3525000214576721, 'rewards/judge_quality/std': 0.14300350844860077, 'rewards/total_composite/mean': 0.24157394468784332, 'rewards/total_composite/std': 0.1338033378124237, 'reward': 0.24157394468784332, 'reward_std': 0.1338033378124237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23752599954605103, 'sampling/sampling_logp_difference/max': 1.6505331993103027, 'sampling/importance_sampling_ratio/min': 0.19194753468036652, 'sampling/importance_sampling_ratio/mean': 1.061669945716858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6323418766260147, 'clip_ratio/low_mean': 0.12274500727653503, 'clip_ratio/low_min': 0.12274500727653503, 'clip_ratio/high_mean': 0.08438725769519806, 'clip_ratio/high_max': 0.08438725769519806, 'clip_ratio/region_mean': 0.2071322649717331, 'reward_total_mean': 0.24157394468784332, 'reward_meter_mean': 0.47886359691619873, 'reward_meter_std': 0.29336100816726685, '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': 0.9957594275474548, 'reward_repeat_soft_std': 0.0027065465692430735, 'reward_judge_quality_mean': 0.3525000214576721, 'reward_judge_quality_std': 0.14300350844860077, 'reward_total_composite_mean': 0.24157394468784332, 'reward_total_composite_std': 0.1338033378124237, 'epoch': 0.01}

  3%|▎         | 95/3300 [11:23<5:48:49,  6.53s/it]INFO 04-13 12:57:06 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 96/3300 [11:29<5:41:47,  6.40s/it]2026-04-13 12:57:11,957 | INFO | train_grpo_train | metrics_logged mode=train step=96

                                                   
{'loss': 0.0583, 'grad_norm': 19.531837463378906, 'learning_rate': 9.712121212121213e-06, 'num_tokens': 174523.0, 'completions/mean_length': 32.75, 'completions/min_length': 24.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5699098110198975, 'rewards/meter/std': 0.35725337266921997, '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': 0.9993458390235901, 'rewards/repeat_soft/std': 0.0018502887105569243, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.09635314345359802, 'rewards/total_composite/mean': 0.22203493118286133, 'rewards/total_composite/std': 0.1043752059340477, 'reward': 0.22203493118286133, 'reward_std': 0.1043751984834671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24431182444095612, 'sampling/sampling_logp_difference/max': 1.7511816024780273, 'sampling/importance_sampling_ratio/min': 0.1735687255859375, 'sampling/importance_sampling_ratio/mean': 1.0452316999435425, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4502941220998764, 'clip_ratio/low_mean': 0.12524668499827385, 'clip_ratio/low_min': 0.12524668499827385, 'clip_ratio/high_mean': 0.09808298293501139, 'clip_ratio/high_max': 0.09808298293501139, 'clip_ratio/region_mean': 0.22332966793328524, 'reward_total_mean': 0.22203493118286133, 'reward_meter_mean': 0.5699098110198975, 'reward_meter_std': 0.35725337266921997, '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': 0.9993458390235901, 'reward_repeat_soft_std': 0.0018502887105569243, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.09635314345359802, 'reward_total_composite_mean': 0.22203493118286133, 'reward_total_composite_std': 0.1043752059340477, 'epoch': 0.01}

  3%|▎         | 96/3300 [11:29<5:41:47,  6.40s/it]INFO 04-13 12:57:12 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 97/3300 [11:35<5:33:35,  6.25s/it]2026-04-13 12:57:17,851 | INFO | train_grpo_train | metrics_logged mode=train step=97

                                                   
{'loss': 0.0396, 'grad_norm': 16.809978485107422, 'learning_rate': 9.70909090909091e-06, 'num_tokens': 176099.0, 'completions/mean_length': 32.0, 'completions/min_length': 26.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.3821866512298584, 'rewards/meter/std': 0.34682658314704895, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9970651865005493, 'rewards/repeat_soft/std': 0.004848456010222435, 'rewards/judge_quality/mean': 0.273749977350235, 'rewards/judge_quality/std': 0.09085899591445923, 'rewards/total_composite/mean': 0.14850452542304993, 'rewards/total_composite/std': 0.12945693731307983, 'reward': 0.14850452542304993, 'reward_std': 0.12945693731307983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24849790334701538, 'sampling/sampling_logp_difference/max': 1.1252381801605225, 'sampling/importance_sampling_ratio/min': 0.32457512617111206, 'sampling/importance_sampling_ratio/mean': 1.038015604019165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6717825829982758, 'clip_ratio/low_mean': 0.13812491949647665, 'clip_ratio/low_min': 0.13812491949647665, 'clip_ratio/high_mean': 0.042803031392395496, 'clip_ratio/high_max': 0.042803031392395496, 'clip_ratio/region_mean': 0.18092795088887215, 'reward_total_mean': 0.14850452542304993, 'reward_meter_mean': 0.3821866512298584, 'reward_meter_std': 0.34682658314704895, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9970651865005493, 'reward_repeat_soft_std': 0.004848456010222435, 'reward_judge_quality_mean': 0.273749977350235, 'reward_judge_quality_std': 0.09085899591445923, 'reward_total_composite_mean': 0.14850452542304993, 'reward_total_composite_std': 0.12945693731307983, 'epoch': 0.01}

  3%|▎         | 97/3300 [11:35<5:33:35,  6.25s/it]INFO 04-13 12:57:18 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 98/3300 [11:41<5:34:12,  6.26s/it]2026-04-13 12:57:24,145 | INFO | train_grpo_train | metrics_logged mode=train step=98

                                                   
{'loss': 0.0738, 'grad_norm': 12.655705451965332, 'learning_rate': 9.706060606060606e-06, 'num_tokens': 177756.0, 'completions/mean_length': 43.125, 'completions/min_length': 32.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.3984159827232361, 'rewards/meter/std': 0.4022829234600067, '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': 0.9977775812149048, 'rewards/repeat_soft/std': 0.0038884899113327265, 'rewards/judge_quality/mean': 0.2512499988079071, 'rewards/judge_quality/std': 0.07219763845205307, 'rewards/total_composite/mean': 0.16491495072841644, 'rewards/total_composite/std': 0.1173827275633812, 'reward': 0.16491495072841644, 'reward_std': 0.11738273501396179, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23115411400794983, 'sampling/sampling_logp_difference/max': 2.0975911617279053, 'sampling/importance_sampling_ratio/min': 0.12275175750255585, 'sampling/importance_sampling_ratio/mean': 1.0054800510406494, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1663263738155365, 'clip_ratio/low_mean': 0.07282582484185696, 'clip_ratio/low_min': 0.07282582484185696, 'clip_ratio/high_mean': 0.06583761237561703, 'clip_ratio/high_max': 0.06583761237561703, 'clip_ratio/region_mean': 0.13866343721747398, 'reward_total_mean': 0.16491495072841644, 'reward_meter_mean': 0.3984159827232361, 'reward_meter_std': 0.4022829234600067, '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': 0.9977775812149048, 'reward_repeat_soft_std': 0.0038884899113327265, 'reward_judge_quality_mean': 0.2512499988079071, 'reward_judge_quality_std': 0.07219763845205307, 'reward_total_composite_mean': 0.16491495072841644, 'reward_total_composite_std': 0.1173827275633812, 'epoch': 0.01}

  3%|▎         | 98/3300 [11:41<5:34:12,  6.26s/it]INFO 04-13 12:57:24 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 99/3300 [11:47<5:30:53,  6.20s/it]2026-04-13 12:57:30,207 | INFO | train_grpo_train | metrics_logged mode=train step=99

                                                   
{'loss': 0.2834, 'grad_norm': 23.216094970703125, 'learning_rate': 9.703030303030305e-06, 'num_tokens': 179079.0, 'completions/mean_length': 21.375, 'completions/min_length': 13.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.375, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.3166063725948334, 'rewards/meter/std': 0.41524726152420044, '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': 0.953080415725708, 'rewards/repeat_soft/std': 0.018900122493505478, 'rewards/judge_quality/mean': 0.4599999785423279, 'rewards/judge_quality/std': 0.25712698698043823, 'rewards/total_composite/mean': 0.284821093082428, 'rewards/total_composite/std': 0.2717686891555786, 'reward': 0.284821093082428, 'reward_std': 0.2717686593532562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22873538732528687, 'sampling/sampling_logp_difference/max': 1.3408770561218262, 'sampling/importance_sampling_ratio/min': 0.2616161108016968, 'sampling/importance_sampling_ratio/mean': 0.9866707921028137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9422717690467834, 'clip_ratio/low_mean': 0.1328728124499321, 'clip_ratio/low_min': 0.1328728124499321, 'clip_ratio/high_mean': 0.0923076942563057, 'clip_ratio/high_max': 0.0923076942563057, 'clip_ratio/region_mean': 0.2251805067062378, 'reward_total_mean': 0.284821093082428, 'reward_meter_mean': 0.3166063725948334, 'reward_meter_std': 0.41524726152420044, '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': 0.953080415725708, 'reward_repeat_soft_std': 0.018900122493505478, 'reward_judge_quality_mean': 0.4599999785423279, 'reward_judge_quality_std': 0.25712698698043823, 'reward_total_composite_mean': 0.284821093082428, 'reward_total_composite_std': 0.2717686891555786, 'epoch': 0.01}

  3%|▎         | 99/3300 [11:47<5:30:53,  6.20s/it]INFO 04-13 12:57:30 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 100/3300 [11:54<5:32:43,  6.24s/it]2026-04-13 12:57:36,537 | INFO | train_grpo_train | metrics_logged mode=train step=100

                                                    
{'loss': 0.0057, 'grad_norm': 15.219758987426758, 'learning_rate': 9.7e-06, 'num_tokens': 180938.0, 'completions/mean_length': 58.375, 'completions/min_length': 48.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.3609089255332947, 'rewards/meter/std': 0.27344444394111633, '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': 0.9970800280570984, 'rewards/repeat_soft/std': 0.0024874836672097445, 'rewards/judge_quality/mean': 0.29499998688697815, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.1632038652896881, 'rewards/total_composite/std': 0.048654042184352875, 'reward': 0.1632038652896881, 'reward_std': 0.048654042184352875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23344087600708008, 'sampling/sampling_logp_difference/max': 1.4850358963012695, 'sampling/importance_sampling_ratio/min': 0.2264942228794098, 'sampling/importance_sampling_ratio/mean': 1.0198392868041992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3165549635887146, 'clip_ratio/low_mean': 0.10147346649318933, 'clip_ratio/low_min': 0.10147346649318933, 'clip_ratio/high_mean': 0.09966985508799553, 'clip_ratio/high_max': 0.09966985508799553, 'clip_ratio/region_mean': 0.20114332158118486, 'reward_total_mean': 0.1632038652896881, 'reward_meter_mean': 0.3609089255332947, 'reward_meter_std': 0.27344444394111633, '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': 0.9970800280570984, 'reward_repeat_soft_std': 0.0024874836672097445, 'reward_judge_quality_mean': 0.29499998688697815, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.1632038652896881, 'reward_total_composite_std': 0.048654042184352875, 'epoch': 0.01}

  3%|▎         | 100/3300 [11:54<5:32:43,  6.24s/it]INFO 04-13 12:57:36 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 20%|██        | 2/10 [00:04<00:17,  2.19s/it]

 30%|███       | 3/10 [00:08<00:20,  2.89s/it]

 40%|████      | 4/10 [00:12<00:19,  3.28s/it]

 50%|█████     | 5/10 [00:15<00:17,  3.44s/it]

 60%|██████    | 6/10 [00:19<00:13,  3.49s/it]

 70%|███████   | 7/10 [00:23<00:11,  3.70s/it]

 80%|████████  | 8/10 [00:27<00:07,  3.79s/it]

 90%|█████████ | 9/10 [00:31<00:03,  3.89s/it]

100%|██████████| 10/10 [00:34<00:00,  3.64s/it]2026-04-13 12:58:15,967 | INFO | train_grpo_train | metrics_logged mode=eval step=100

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 39.4212, 'eval_samples_per_second': 2.029, 'eval_steps_per_second': 0.254, 'eval_num_tokens': 180938.0, 'eval_completions/mean_length': 60.3, 'eval_completions/min_length': 26.1, 'eval_completions/max_length': 115.9, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 60.3, 'eval_completions/min_terminated_length': 26.1, 'eval_completions/max_terminated_length': 115.9, 'eval_rewards/meter/mean': 0.5333475530147552, 'eval_rewards/meter/std': 0.35966198444366454, 'eval_rewards/count_adherence/mean': 0.9702083349227906, 'eval_rewards/count_adherence/std': 0.07405876778066159, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.1767766922712326, 'eval_rewards/repeat_soft/mean': 0.9931838035583496, 'eval_rewards/repeat_soft/std': 0.01113832175033167, 'eval_rewards/judge_quality/mean': 0.28999999463558196, 'eval_rewards/judge_quality/std': 0.09390390664339066, 'eval_rewards/total_composite/mean': 0.19018368870019914, 'eval_rewards/total_composite/std': 0.10621417425572872, 'eval_reward': 0.19018368870019914, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15492237582802773, 'eval_sampling/sampling_logp_difference/max': 1.1971155643463134, 'eval_sampling/importance_sampling_ratio/min': 0.30891612619161607, 'eval_sampling/importance_sampling_ratio/mean': 1.0459434866905213, 'eval_sampling/importance_sampling_ratio/max': 1.5796868562698365, 'eval_entropy': 2.415260064601898, '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': 0.19018368870019914, 'eval_reward_meter_mean': 0.5333475530147552, 'eval_reward_meter_std': 0.35966198444366454, 'eval_reward_count_adherence_mean': 0.9702083349227906, 'eval_reward_count_adherence_std': 0.07405876778066159, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.1767766922712326, 'eval_reward_repeat_soft_mean': 0.9931838035583496, 'eval_reward_repeat_soft_std': 0.01113832175033167, 'eval_reward_judge_quality_mean': 0.28999999463558196, 'eval_reward_judge_quality_std': 0.09390390664339066, 'eval_reward_total_composite_mean': 0.19018368870019914, 'eval_reward_total_composite_std': 0.10621417425572872, 'epoch': 0.01}

  3%|▎         | 100/3300 [12:33<5:32:43,  6.24s/it]

100%|██████████| 10/10 [00:34<00:00,  3.64s/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 [12:42<16:51:39, 18.97s/it]2026-04-13 12:58:25,223 | INFO | train_grpo_train | metrics_logged mode=train step=101

                                                     
{'loss': 0.2388, 'grad_norm': 11.822274208068848, 'learning_rate': 9.696969696969698e-06, 'num_tokens': 182977.0, 'completions/mean_length': 78.875, 'completions/min_length': 52.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.9055176973342896, 'rewards/meter/std': 0.1499919444322586, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9967098832130432, 'rewards/repeat_soft/std': 0.0029830147977918386, 'rewards/judge_quality/mean': 0.29499998688697815, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.27612704038619995, 'rewards/total_composite/std': 0.11570548266172409, 'reward': 0.27612704038619995, 'reward_std': 0.11570549011230469, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22361187636852264, 'sampling/sampling_logp_difference/max': 1.6272878646850586, 'sampling/importance_sampling_ratio/min': 0.19646167755126953, 'sampling/importance_sampling_ratio/mean': 1.0615099668502808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.19254332780838, 'clip_ratio/low_mean': 0.11758391372859478, 'clip_ratio/low_min': 0.11758391372859478, 'clip_ratio/high_mean': 0.07190426252782345, 'clip_ratio/high_max': 0.07190426252782345, 'clip_ratio/region_mean': 0.18948817625641823, 'reward_total_mean': 0.27612704038619995, 'reward_meter_mean': 0.9055176973342896, 'reward_meter_std': 0.1499919444322586, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9967098832130432, 'reward_repeat_soft_std': 0.0029830147977918386, 'reward_judge_quality_mean': 0.29499998688697815, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.27612704038619995, 'reward_total_composite_std': 0.11570548266172409, 'epoch': 0.01}

  3%|▎         | 101/3300 [12:42<16:51:39, 18.97s/it]INFO 04-13 12:58:25 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 102/3300 [12:50<13:48:51, 15.55s/it]2026-04-13 12:58:32,785 | INFO | train_grpo_train | metrics_logged mode=train step=102

                                                     
{'loss': 0.0634, 'grad_norm': 12.916412353515625, 'learning_rate': 9.693939393939395e-06, 'num_tokens': 185048.0, 'completions/mean_length': 79.875, 'completions/min_length': 74.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.22276419401168823, 'rewards/meter/std': 0.20111523568630219, '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': 0.9941090941429138, 'rewards/repeat_soft/std': 0.0047340840101242065, 'rewards/judge_quality/mean': 0.24250000715255737, 'rewards/judge_quality/std': 0.031052952632308006, 'rewards/total_composite/mean': 0.11771676689386368, 'rewards/total_composite/std': 0.029656479135155678, 'reward': 0.11771676689386368, 'reward_std': 0.02965647727251053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20215699076652527, 'sampling/sampling_logp_difference/max': 1.4278860092163086, 'sampling/importance_sampling_ratio/min': 0.239815354347229, 'sampling/importance_sampling_ratio/mean': 1.0422965288162231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8850602060556412, 'clip_ratio/low_mean': 0.11844531446695328, 'clip_ratio/low_min': 0.11844531446695328, 'clip_ratio/high_mean': 0.056453270837664604, 'clip_ratio/high_max': 0.056453270837664604, 'clip_ratio/region_mean': 0.17489858530461788, 'reward_total_mean': 0.11771676689386368, 'reward_meter_mean': 0.22276419401168823, 'reward_meter_std': 0.20111523568630219, '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': 0.9941090941429138, 'reward_repeat_soft_std': 0.0047340840101242065, 'reward_judge_quality_mean': 0.24250000715255737, 'reward_judge_quality_std': 0.031052952632308006, 'reward_total_composite_mean': 0.11771676689386368, 'reward_total_composite_std': 0.029656479135155678, 'epoch': 0.01}

  3%|▎         | 102/3300 [12:50<13:48:51, 15.55s/it]INFO 04-13 12:58:33 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 103/3300 [12:56<11:21:03, 12.78s/it]2026-04-13 12:58:39,106 | INFO | train_grpo_train | metrics_logged mode=train step=103

                                                     
{'loss': 0.1125, 'grad_norm': 17.527645111083984, 'learning_rate': 9.690909090909092e-06, 'num_tokens': 186673.0, 'completions/mean_length': 46.125, 'completions/min_length': 37.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8049889802932739, 'rewards/meter/std': 0.28971365094184875, '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': 0.9981218576431274, 'rewards/repeat_soft/std': 0.004964757245033979, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.18715444207191467, 'rewards/total_composite/mean': 0.3515571355819702, 'rewards/total_composite/std': 0.2247399538755417, 'reward': 0.3515571355819702, 'reward_std': 0.2247399389743805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18538333475589752, 'sampling/sampling_logp_difference/max': 2.0367231369018555, 'sampling/importance_sampling_ratio/min': 0.13045549392700195, 'sampling/importance_sampling_ratio/mean': 1.012653112411499, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.654091089963913, 'clip_ratio/low_mean': 0.06139914924278855, 'clip_ratio/low_min': 0.06139914924278855, 'clip_ratio/high_mean': 0.11172189377248287, 'clip_ratio/high_max': 0.11172189377248287, 'clip_ratio/region_mean': 0.17312104301527143, 'reward_total_mean': 0.3515571355819702, 'reward_meter_mean': 0.8049889802932739, 'reward_meter_std': 0.28971365094184875, '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': 0.9981218576431274, 'reward_repeat_soft_std': 0.004964757245033979, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.18715444207191467, 'reward_total_composite_mean': 0.3515571355819702, 'reward_total_composite_std': 0.2247399538755417, 'epoch': 0.01}

  3%|▎         | 103/3300 [12:56<11:21:03, 12.78s/it]INFO 04-13 12:58:39 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 104/3300 [13:02<9:36:48, 10.83s/it] 2026-04-13 12:58:45,377 | INFO | train_grpo_train | metrics_logged mode=train step=104

                                                    
{'loss': -0.0233, 'grad_norm': 18.21098518371582, 'learning_rate': 9.687878787878788e-06, 'num_tokens': 188211.0, 'completions/mean_length': 38.25, 'completions/min_length': 29.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6183866858482361, 'rewards/meter/std': 0.4217718541622162, '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': 0.9947104454040527, 'rewards/repeat_soft/std': 0.00796837080270052, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.2452881783246994, 'rewards/total_composite/std': 0.10636084526777267, 'reward': 0.2452881783246994, 'reward_std': 0.10636083781719208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19570432603359222, 'sampling/sampling_logp_difference/max': 1.4738616943359375, 'sampling/importance_sampling_ratio/min': 0.2290392965078354, 'sampling/importance_sampling_ratio/mean': 1.0237973928451538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7581131011247635, 'clip_ratio/low_mean': 0.11726908572018147, 'clip_ratio/low_min': 0.11726908572018147, 'clip_ratio/high_mean': 0.060981093905866146, 'clip_ratio/high_max': 0.060981093905866146, 'clip_ratio/region_mean': 0.1782501796260476, 'reward_total_mean': 0.2452881783246994, 'reward_meter_mean': 0.6183866858482361, 'reward_meter_std': 0.4217718541622162, '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': 0.9947104454040527, 'reward_repeat_soft_std': 0.00796837080270052, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.2452881783246994, 'reward_total_composite_std': 0.10636084526777267, 'epoch': 0.01}

  3%|▎         | 104/3300 [13:02<9:36:48, 10.83s/it]INFO 04-13 12:58:45 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 105/3300 [13:09<8:29:40,  9.57s/it]2026-04-13 12:58:52,123 | INFO | train_grpo_train | metrics_logged mode=train step=105

                                                    
{'loss': -0.0525, 'grad_norm': 13.55850887298584, 'learning_rate': 9.684848484848487e-06, 'num_tokens': 190277.0, 'completions/mean_length': 74.25, 'completions/min_length': 67.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.47648465633392334, 'rewards/meter/std': 0.3585143983364105, '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': 0.9975783824920654, 'rewards/repeat_soft/std': 0.002411383204162121, 'rewards/judge_quality/mean': 0.2199999988079071, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.14505738019943237, 'rewards/total_composite/std': 0.051263850182294846, 'reward': 0.14505738019943237, 'reward_std': 0.05126384645700455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24152730405330658, 'sampling/sampling_logp_difference/max': 1.9144554138183594, 'sampling/importance_sampling_ratio/min': 0.14742209017276764, 'sampling/importance_sampling_ratio/mean': 1.0519094467163086, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5489106625318527, 'clip_ratio/low_mean': 0.16279510967433453, 'clip_ratio/low_min': 0.16279510967433453, 'clip_ratio/high_mean': 0.08875360898673534, 'clip_ratio/high_max': 0.08875360898673534, 'clip_ratio/region_mean': 0.25154871866106987, 'reward_total_mean': 0.14505738019943237, 'reward_meter_mean': 0.47648465633392334, 'reward_meter_std': 0.3585143983364105, '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': 0.9975783824920654, 'reward_repeat_soft_std': 0.002411383204162121, 'reward_judge_quality_mean': 0.2199999988079071, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.14505738019943237, 'reward_total_composite_std': 0.051263850182294846, 'epoch': 0.01}

  3%|▎         | 105/3300 [13:09<8:29:40,  9.57s/it]INFO 04-13 12:58:52 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 106/3300 [13:16<7:44:58,  8.73s/it]2026-04-13 12:58:58,798 | INFO | train_grpo_train | metrics_logged mode=train step=106

                                                    
{'loss': 0.003, 'grad_norm': 24.08884048461914, 'learning_rate': 9.681818181818182e-06, 'num_tokens': 191708.0, 'completions/mean_length': 26.875, 'completions/min_length': 24.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.875, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.3493647277355194, 'rewards/meter/std': 0.20752744376659393, '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': 0.9945707321166992, 'rewards/repeat_soft/std': 0.008784092962741852, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.20193137228488922, 'rewards/total_composite/std': 0.06303919106721878, 'reward': 0.20193137228488922, 'reward_std': 0.06303918361663818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2212163507938385, 'sampling/sampling_logp_difference/max': 2.582676887512207, 'sampling/importance_sampling_ratio/min': 0.07557143270969391, 'sampling/importance_sampling_ratio/mean': 1.0338249206542969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0336359068751335, 'clip_ratio/low_mean': 0.13796198274940252, 'clip_ratio/low_min': 0.13796198274940252, 'clip_ratio/high_mean': 0.047883063554763794, 'clip_ratio/high_max': 0.047883063554763794, 'clip_ratio/region_mean': 0.18584504630416632, 'reward_total_mean': 0.20193137228488922, 'reward_meter_mean': 0.3493647277355194, 'reward_meter_std': 0.20752744376659393, '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': 0.9945707321166992, 'reward_repeat_soft_std': 0.008784092962741852, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.20193137228488922, 'reward_total_composite_std': 0.06303919106721878, 'epoch': 0.01}

  3%|▎         | 106/3300 [13:16<7:44:58,  8.73s/it]INFO 04-13 12:58:59 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 107/3300 [13:22<7:02:32,  7.94s/it]2026-04-13 12:59:04,883 | INFO | train_grpo_train | metrics_logged mode=train step=107

                                                    
{'loss': 0.171, 'grad_norm': 17.395442962646484, 'learning_rate': 9.67878787878788e-06, 'num_tokens': 193328.0, 'completions/mean_length': 39.5, 'completions/min_length': 29.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4425399899482727, 'rewards/meter/std': 0.3519269824028015, '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': 0.9966769814491272, 'rewards/repeat_soft/std': 0.00410105474293232, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.24182343482971191, 'rewards/total_composite/mean': 0.28937357664108276, 'rewards/total_composite/std': 0.1784743368625641, 'reward': 0.28937357664108276, 'reward_std': 0.1784743368625641, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17070765793323517, 'sampling/sampling_logp_difference/max': 1.2045235633850098, 'sampling/importance_sampling_ratio/min': 0.29983481764793396, 'sampling/importance_sampling_ratio/mean': 0.9965779185295105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6595074906945229, 'clip_ratio/low_mean': 0.08546209614723921, 'clip_ratio/low_min': 0.08546209614723921, 'clip_ratio/high_mean': 0.06928638741374016, 'clip_ratio/high_max': 0.06928638741374016, 'clip_ratio/region_mean': 0.15474848356097937, 'reward_total_mean': 0.28937357664108276, 'reward_meter_mean': 0.4425399899482727, 'reward_meter_std': 0.3519269824028015, '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': 0.9966769814491272, 'reward_repeat_soft_std': 0.00410105474293232, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.24182343482971191, 'reward_total_composite_mean': 0.28937357664108276, 'reward_total_composite_std': 0.1784743368625641, 'epoch': 0.01}

  3%|▎         | 107/3300 [13:22<7:02:32,  7.94s/it]INFO 04-13 12:59:05 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 108/3300 [13:28<6:32:07,  7.37s/it]2026-04-13 12:59:10,926 | INFO | train_grpo_train | metrics_logged mode=train step=108

                                                    
{'loss': 0.0939, 'grad_norm': 25.62848472595215, 'learning_rate': 9.675757575757577e-06, 'num_tokens': 194611.0, 'completions/mean_length': 25.375, 'completions/min_length': 20.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.37729573249816895, 'rewards/meter/std': 0.37614983320236206, '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': 0.9955626726150513, 'rewards/repeat_soft/std': 0.005067089106887579, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.14024850726127625, 'rewards/total_composite/mean': 0.26623427867889404, 'rewards/total_composite/std': 0.15444695949554443, 'reward': 0.26623427867889404, 'reward_std': 0.15444695949554443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26287201046943665, 'sampling/sampling_logp_difference/max': 1.492593765258789, 'sampling/importance_sampling_ratio/min': 0.2247888594865799, 'sampling/importance_sampling_ratio/mean': 1.0210803747177124, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9178846180438995, 'clip_ratio/low_mean': 0.16352272778749466, 'clip_ratio/low_min': 0.16352272778749466, 'clip_ratio/high_mean': 0.08564814925193787, 'clip_ratio/high_max': 0.08564814925193787, 'clip_ratio/region_mean': 0.24917087703943253, 'reward_total_mean': 0.26623427867889404, 'reward_meter_mean': 0.37729573249816895, 'reward_meter_std': 0.37614983320236206, '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': 0.9955626726150513, 'reward_repeat_soft_std': 0.005067089106887579, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.14024850726127625, 'reward_total_composite_mean': 0.26623427867889404, 'reward_total_composite_std': 0.15444695949554443, 'epoch': 0.01}

  3%|▎         | 108/3300 [13:28<6:32:07,  7.37s/it]INFO 04-13 12:59:11 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 109/3300 [13:35<6:18:33,  7.12s/it]2026-04-13 12:59:17,454 | INFO | train_grpo_train | metrics_logged mode=train step=109

                                                    
{'loss': 0.0497, 'grad_norm': 16.82399559020996, 'learning_rate': 9.672727272727274e-06, 'num_tokens': 196165.0, 'completions/mean_length': 34.25, 'completions/min_length': 31.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6299466490745544, 'rewards/meter/std': 0.3046122193336487, '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': 0.9920294284820557, 'rewards/repeat_soft/std': 0.009845644235610962, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.1507066786289215, 'rewards/total_composite/mean': 0.20779313147068024, 'rewards/total_composite/std': 0.1545531302690506, 'reward': 0.20779313147068024, 'reward_std': 0.1545531302690506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.229281947016716, 'sampling/sampling_logp_difference/max': 1.712673306465149, 'sampling/importance_sampling_ratio/min': 0.18038292229175568, 'sampling/importance_sampling_ratio/mean': 1.0576488971710205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3824533075094223, 'clip_ratio/low_mean': 0.13064675964415073, 'clip_ratio/low_min': 0.13064675964415073, 'clip_ratio/high_mean': 0.08631171379238367, 'clip_ratio/high_max': 0.08631171379238367, 'clip_ratio/region_mean': 0.2169584734365344, 'reward_total_mean': 0.20779313147068024, 'reward_meter_mean': 0.6299466490745544, 'reward_meter_std': 0.3046122193336487, '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': 0.9920294284820557, 'reward_repeat_soft_std': 0.009845644235610962, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.1507066786289215, 'reward_total_composite_mean': 0.20779313147068024, 'reward_total_composite_std': 0.1545531302690506, 'epoch': 0.01}

  3%|▎         | 109/3300 [13:35<6:18:33,  7.12s/it]INFO 04-13 12:59:17 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 110/3300 [13:41<6:01:14,  6.79s/it]2026-04-13 12:59:23,493 | INFO | train_grpo_train | metrics_logged mode=train step=110

                                                    
{'loss': 0.0266, 'grad_norm': 15.715481758117676, 'learning_rate': 9.66969696969697e-06, 'num_tokens': 197812.0, 'completions/mean_length': 36.875, 'completions/min_length': 35.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9310295581817627, 'rewards/meter/std': 0.1262810379266739, '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': 0.9877753257751465, 'rewards/repeat_soft/std': 0.014998417347669601, 'rewards/judge_quality/mean': 0.4987499713897705, 'rewards/judge_quality/std': 0.23277442157268524, 'rewards/total_composite/mean': 0.48087847232818604, 'rewards/total_composite/std': 0.2299085557460785, 'reward': 0.48087847232818604, 'reward_std': 0.2299085408449173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19083642959594727, 'sampling/sampling_logp_difference/max': 1.0480821132659912, 'sampling/importance_sampling_ratio/min': 0.3924664855003357, 'sampling/importance_sampling_ratio/mean': 1.069949984550476, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.354040265083313, 'clip_ratio/low_mean': 0.10091430973261595, 'clip_ratio/low_min': 0.10091430973261595, 'clip_ratio/high_mean': 0.04637694172561169, 'clip_ratio/high_max': 0.04637694172561169, 'clip_ratio/region_mean': 0.14729125145822763, 'reward_total_mean': 0.48087847232818604, 'reward_meter_mean': 0.9310295581817627, 'reward_meter_std': 0.1262810379266739, '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': 0.9877753257751465, 'reward_repeat_soft_std': 0.014998417347669601, 'reward_judge_quality_mean': 0.4987499713897705, 'reward_judge_quality_std': 0.23277442157268524, 'reward_total_composite_mean': 0.48087847232818604, 'reward_total_composite_std': 0.2299085557460785, 'epoch': 0.01}

  3%|▎         | 110/3300 [13:41<6:01:14,  6.79s/it]INFO 04-13 12:59:23 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 111/3300 [13:48<6:10:34,  6.97s/it]2026-04-13 12:59:30,885 | INFO | train_grpo_train | metrics_logged mode=train step=111

                                                    
{'loss': 0.0448, 'grad_norm': 17.277462005615234, 'learning_rate': 9.666666666666667e-06, 'num_tokens': 199647.0, 'completions/mean_length': 65.375, 'completions/min_length': 54.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.5294692516326904, 'rewards/meter/std': 0.29245197772979736, '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': 0.9982734322547913, 'rewards/repeat_soft/std': 0.0013606828870251775, 'rewards/judge_quality/mean': 0.3024999797344208, 'rewards/judge_quality/std': 0.09939099848270416, 'rewards/total_composite/mean': 0.20176908373832703, 'rewards/total_composite/std': 0.11734206974506378, 'reward': 0.20176908373832703, 'reward_std': 0.11734206229448318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23416373133659363, 'sampling/sampling_logp_difference/max': 1.5392494201660156, 'sampling/importance_sampling_ratio/min': 0.21454209089279175, 'sampling/importance_sampling_ratio/mean': 1.045576572418213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.953499972820282, 'clip_ratio/low_mean': 0.07657067384570837, 'clip_ratio/low_min': 0.07657067384570837, 'clip_ratio/high_mean': 0.1422189585864544, 'clip_ratio/high_max': 0.1422189585864544, 'clip_ratio/region_mean': 0.21878963243216276, 'reward_total_mean': 0.20176908373832703, 'reward_meter_mean': 0.5294692516326904, 'reward_meter_std': 0.29245197772979736, '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': 0.9982734322547913, 'reward_repeat_soft_std': 0.0013606828870251775, 'reward_judge_quality_mean': 0.3024999797344208, 'reward_judge_quality_std': 0.09939099848270416, 'reward_total_composite_mean': 0.20176908373832703, 'reward_total_composite_std': 0.11734206974506378, 'epoch': 0.01}

  3%|▎         | 111/3300 [13:48<6:10:34,  6.97s/it]INFO 04-13 12:59:31 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 112/3300 [13:54<5:51:36,  6.62s/it]2026-04-13 12:59:36,670 | INFO | train_grpo_train | metrics_logged mode=train step=112

                                                    
{'loss': -0.0488, 'grad_norm': 26.42018699645996, 'learning_rate': 9.663636363636364e-06, 'num_tokens': 201110.0, 'completions/mean_length': 17.875, 'completions/min_length': 14.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5070533156394958, 'rewards/meter/std': 0.4634113907814026, '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': 0.6875, 'rewards/judge_quality/std': 0.36228442192077637, 'rewards/total_composite/mean': 0.44090500473976135, 'rewards/total_composite/std': 0.31895413994789124, 'reward': 0.44090500473976135, 'reward_std': 0.31895413994789124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17994388937950134, 'sampling/sampling_logp_difference/max': 1.7454051971435547, 'sampling/importance_sampling_ratio/min': 0.17457422614097595, 'sampling/importance_sampling_ratio/mean': 1.0285303592681885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4367366135120392, 'clip_ratio/low_mean': 0.11488095764070749, 'clip_ratio/low_min': 0.11488095764070749, 'clip_ratio/high_mean': 0.01666666753590107, 'clip_ratio/high_max': 0.01666666753590107, 'clip_ratio/region_mean': 0.13154762517660856, 'reward_total_mean': 0.44090500473976135, 'reward_meter_mean': 0.5070533156394958, 'reward_meter_std': 0.4634113907814026, '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': 0.6875, 'reward_judge_quality_std': 0.36228442192077637, 'reward_total_composite_mean': 0.44090500473976135, 'reward_total_composite_std': 0.31895413994789124, 'epoch': 0.01}

  3%|▎         | 112/3300 [13:54<5:51:36,  6.62s/it]INFO 04-13 12:59:36 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 113/3300 [14:00<5:46:00,  6.51s/it]2026-04-13 12:59:42,945 | INFO | train_grpo_train | metrics_logged mode=train step=113

                                                    
{'loss': -0.0158, 'grad_norm': 14.468746185302734, 'learning_rate': 9.660606060606061e-06, 'num_tokens': 202798.0, 'completions/mean_length': 55.0, 'completions/min_length': 46.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.2444886863231659, 'rewards/meter/std': 0.281507283449173, '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': 0.9952455759048462, 'rewards/repeat_soft/std': 0.0036034355871379375, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.2020210474729538, 'rewards/total_composite/mean': 0.20411249995231628, 'rewards/total_composite/std': 0.08018417656421661, 'reward': 0.20411249995231628, 'reward_std': 0.08018417656421661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1938779354095459, 'sampling/sampling_logp_difference/max': 2.061924934387207, 'sampling/importance_sampling_ratio/min': 0.1272088587284088, 'sampling/importance_sampling_ratio/mean': 1.0262696743011475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8172344714403152, 'clip_ratio/low_mean': 0.09356675297021866, 'clip_ratio/low_min': 0.09356675297021866, 'clip_ratio/high_mean': 0.11305304430425167, 'clip_ratio/high_max': 0.11305304430425167, 'clip_ratio/region_mean': 0.20661979727447033, 'reward_total_mean': 0.20411249995231628, 'reward_meter_mean': 0.2444886863231659, 'reward_meter_std': 0.281507283449173, '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': 0.9952455759048462, 'reward_repeat_soft_std': 0.0036034355871379375, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.2020210474729538, 'reward_total_composite_mean': 0.20411249995231628, 'reward_total_composite_std': 0.08018417656421661, 'epoch': 0.01}

  3%|▎         | 113/3300 [14:00<5:46:00,  6.51s/it]INFO 04-13 12:59:43 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 114/3300 [14:06<5:42:28,  6.45s/it]2026-04-13 12:59:49,243 | INFO | train_grpo_train | metrics_logged mode=train step=114

                                                    
{'loss': -0.0393, 'grad_norm': 15.590677261352539, 'learning_rate': 9.657575757575758e-06, 'num_tokens': 204317.0, 'completions/mean_length': 35.875, 'completions/min_length': 30.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9072677493095398, 'rewards/meter/std': 0.18171843886375427, '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': 0.9994348883628845, 'rewards/repeat_soft/std': 0.0014808415435254574, 'rewards/judge_quality/mean': 0.29499998688697815, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.28085845708847046, 'rewards/total_composite/std': 0.11390554904937744, 'reward': 0.28085845708847046, 'reward_std': 0.11390554159879684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2342054545879364, 'sampling/sampling_logp_difference/max': 2.456653594970703, 'sampling/importance_sampling_ratio/min': 0.08572132885456085, 'sampling/importance_sampling_ratio/mean': 1.0775035619735718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0509612560272217, 'clip_ratio/low_mean': 0.10094697307795286, 'clip_ratio/low_min': 0.10094697307795286, 'clip_ratio/high_mean': 0.07347696367651224, 'clip_ratio/high_max': 0.07347696367651224, 'clip_ratio/region_mean': 0.1744239367544651, 'reward_total_mean': 0.28085845708847046, 'reward_meter_mean': 0.9072677493095398, 'reward_meter_std': 0.18171843886375427, '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': 0.9994348883628845, 'reward_repeat_soft_std': 0.0014808415435254574, 'reward_judge_quality_mean': 0.29499998688697815, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.28085845708847046, 'reward_total_composite_std': 0.11390554904937744, 'epoch': 0.01}

  3%|▎         | 114/3300 [14:06<5:42:28,  6.45s/it]INFO 04-13 12:59:49 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 115/3300 [14:13<5:44:55,  6.50s/it]2026-04-13 12:59:55,852 | INFO | train_grpo_train | metrics_logged mode=train step=115

                                                    
{'loss': -0.0278, 'grad_norm': 13.686504364013672, 'learning_rate': 9.654545454545456e-06, 'num_tokens': 206390.0, 'completions/mean_length': 71.125, 'completions/min_length': 64.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.4500429630279541, 'rewards/meter/std': 0.3327544927597046, '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': 0.9969667196273804, 'rewards/repeat_soft/std': 0.0027483219746500254, 'rewards/judge_quality/mean': 0.22749999165534973, 'rewards/judge_quality/std': 0.02121320553123951, 'rewards/total_composite/mean': 0.14654693007469177, 'rewards/total_composite/std': 0.05158424749970436, 'reward': 0.14654693007469177, 'reward_std': 0.05158424377441406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.226264089345932, 'sampling/sampling_logp_difference/max': 1.7206354141235352, 'sampling/importance_sampling_ratio/min': 0.1789523959159851, 'sampling/importance_sampling_ratio/mean': 1.0400731563568115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.294025242328644, 'clip_ratio/low_mean': 0.07087157759815454, 'clip_ratio/low_min': 0.07087157759815454, 'clip_ratio/high_mean': 0.11259152367711067, 'clip_ratio/high_max': 0.11259152367711067, 'clip_ratio/region_mean': 0.18346310127526522, 'reward_total_mean': 0.14654693007469177, 'reward_meter_mean': 0.4500429630279541, 'reward_meter_std': 0.3327544927597046, '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': 0.9969667196273804, 'reward_repeat_soft_std': 0.0027483219746500254, 'reward_judge_quality_mean': 0.22749999165534973, 'reward_judge_quality_std': 0.02121320553123951, 'reward_total_composite_mean': 0.14654693007469177, 'reward_total_composite_std': 0.05158424749970436, 'epoch': 0.01}

  3%|▎         | 115/3300 [14:13<5:44:55,  6.50s/it]INFO 04-13 12:59:56 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 116/3300 [14:19<5:38:02,  6.37s/it]2026-04-13 13:00:01,971 | INFO | train_grpo_train | metrics_logged mode=train step=116

                                                    
{'loss': -0.0171, 'grad_norm': 16.382169723510742, 'learning_rate': 9.651515151515153e-06, 'num_tokens': 207902.0, 'completions/mean_length': 34.0, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8620812892913818, 'rewards/meter/std': 0.3167654573917389, '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': 0.9956598281860352, 'rewards/repeat_soft/std': 0.009578917175531387, 'rewards/judge_quality/mean': 0.44874995946884155, 'rewards/judge_quality/std': 0.12088219076395035, 'rewards/total_composite/mean': 0.41245341300964355, 'rewards/total_composite/std': 0.15856164693832397, 'reward': 0.41245341300964355, 'reward_std': 0.15856164693832397, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.188541442155838, 'sampling/sampling_logp_difference/max': 1.8721799850463867, 'sampling/importance_sampling_ratio/min': 0.15378804504871368, 'sampling/importance_sampling_ratio/mean': 1.039787769317627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6071799099445343, 'clip_ratio/low_mean': 0.04992884211242199, 'clip_ratio/low_min': 0.04992884211242199, 'clip_ratio/high_mean': 0.1412046840414405, 'clip_ratio/high_max': 0.1412046840414405, 'clip_ratio/region_mean': 0.19113352615386248, 'reward_total_mean': 0.41245341300964355, 'reward_meter_mean': 0.8620812892913818, 'reward_meter_std': 0.3167654573917389, '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': 0.9956598281860352, 'reward_repeat_soft_std': 0.009578917175531387, 'reward_judge_quality_mean': 0.44874995946884155, 'reward_judge_quality_std': 0.12088219076395035, 'reward_total_composite_mean': 0.41245341300964355, 'reward_total_composite_std': 0.15856164693832397, 'epoch': 0.01}

  4%|▎         | 116/3300 [14:19<5:38:02,  6.37s/it]INFO 04-13 13:00:02 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 117/3300 [14:25<5:37:06,  6.35s/it]2026-04-13 13:00:08,243 | INFO | train_grpo_train | metrics_logged mode=train step=117

                                                    
{'loss': 0.0325, 'grad_norm': 17.374835968017578, 'learning_rate': 9.648484848484849e-06, 'num_tokens': 209420.0, 'completions/mean_length': 38.75, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7884187698364258, 'rewards/meter/std': 0.2108745276927948, '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': 0.9941059947013855, 'rewards/repeat_soft/std': 0.006391895469278097, 'rewards/judge_quality/mean': 0.23499999940395355, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.2026297152042389, 'rewards/total_composite/std': 0.03590220957994461, 'reward': 0.2026297152042389, 'reward_std': 0.03590220585465431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09602572023868561, 'sampling/sampling_logp_difference/max': 1.2067923545837402, 'sampling/importance_sampling_ratio/min': 0.2991553246974945, 'sampling/importance_sampling_ratio/mean': 0.9958279132843018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48540198802948, 'clip_ratio/low_mean': 0.05902969278395176, 'clip_ratio/low_min': 0.05902969278395176, 'clip_ratio/high_mean': 0.043296586256474257, 'clip_ratio/high_max': 0.043296586256474257, 'clip_ratio/region_mean': 0.10232627904042602, 'reward_total_mean': 0.2026297152042389, 'reward_meter_mean': 0.7884187698364258, 'reward_meter_std': 0.2108745276927948, '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': 0.9941059947013855, 'reward_repeat_soft_std': 0.006391895469278097, 'reward_judge_quality_mean': 0.23499999940395355, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.2026297152042389, 'reward_total_composite_std': 0.03590220957994461, 'epoch': 0.01}

  4%|▎         | 117/3300 [14:25<5:37:06,  6.35s/it]INFO 04-13 13:00:08 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 118/3300 [14:33<5:55:17,  6.70s/it]2026-04-13 13:00:15,748 | INFO | train_grpo_train | metrics_logged mode=train step=118

                                                    
{'loss': 0.0367, 'grad_norm': 10.21153450012207, 'learning_rate': 9.645454545454548e-06, 'num_tokens': 211862.0, 'completions/mean_length': 105.25, 'completions/min_length': 90.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.25, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.6786713600158691, 'rewards/meter/std': 0.38632097840309143, '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': 0.9965351223945618, 'rewards/repeat_soft/std': 0.004216012079268694, 'rewards/judge_quality/mean': 0.26999998092651367, 'rewards/judge_quality/std': 0.06989788264036179, 'rewards/total_composite/mean': 0.19293062388896942, 'rewards/total_composite/std': 0.1444544941186905, 'reward': 0.19293062388896942, 'reward_std': 0.1444544941186905, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24396754801273346, 'sampling/sampling_logp_difference/max': 2.569154739379883, 'sampling/importance_sampling_ratio/min': 0.07660026848316193, 'sampling/importance_sampling_ratio/mean': 1.0612008571624756, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2306472063064575, 'clip_ratio/low_mean': 0.0684139784425497, 'clip_ratio/low_min': 0.0684139784425497, 'clip_ratio/high_mean': 0.1381640676409006, 'clip_ratio/high_max': 0.1381640676409006, 'clip_ratio/region_mean': 0.20657804608345032, 'reward_total_mean': 0.19293062388896942, 'reward_meter_mean': 0.6786713600158691, 'reward_meter_std': 0.38632097840309143, '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': 0.9965351223945618, 'reward_repeat_soft_std': 0.004216012079268694, 'reward_judge_quality_mean': 0.26999998092651367, 'reward_judge_quality_std': 0.06989788264036179, 'reward_total_composite_mean': 0.19293062388896942, 'reward_total_composite_std': 0.1444544941186905, 'epoch': 0.01}

  4%|▎         | 118/3300 [14:33<5:55:17,  6.70s/it]INFO 04-13 13:00:16 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 119/3300 [14:39<5:53:40,  6.67s/it]2026-04-13 13:00:22,353 | INFO | train_grpo_train | metrics_logged mode=train step=119

                                                    
{'loss': 0.171, 'grad_norm': 15.081986427307129, 'learning_rate': 9.642424242424243e-06, 'num_tokens': 213532.0, 'completions/mean_length': 42.75, 'completions/min_length': 33.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.39194825291633606, 'rewards/meter/std': 0.38216766715049744, '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': 0.9988094568252563, 'rewards/repeat_soft/std': 0.001973578706383705, 'rewards/judge_quality/mean': 0.26375001668930054, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.15987950563430786, 'rewards/total_composite/std': 0.07312759011983871, 'reward': 0.15987950563430786, 'reward_std': 0.07312759757041931, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2222803384065628, 'sampling/sampling_logp_difference/max': 1.1502761840820312, 'sampling/importance_sampling_ratio/min': 0.3165493309497833, 'sampling/importance_sampling_ratio/mean': 1.0256190299987793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.469740629196167, 'clip_ratio/low_mean': 0.11986222676932812, 'clip_ratio/low_min': 0.11986222676932812, 'clip_ratio/high_mean': 0.12315725162625313, 'clip_ratio/high_max': 0.12315725162625313, 'clip_ratio/region_mean': 0.24301947839558125, 'reward_total_mean': 0.15987950563430786, 'reward_meter_mean': 0.39194825291633606, 'reward_meter_std': 0.38216766715049744, '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': 0.9988094568252563, 'reward_repeat_soft_std': 0.001973578706383705, 'reward_judge_quality_mean': 0.26375001668930054, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.15987950563430786, 'reward_total_composite_std': 0.07312759011983871, 'epoch': 0.01}

  4%|▎         | 119/3300 [14:39<5:53:40,  6.67s/it]INFO 04-13 13:00:22 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 120/3300 [14:47<6:12:44,  7.03s/it]2026-04-13 13:00:30,229 | INFO | train_grpo_train | metrics_logged mode=train step=120

                                                    
{'loss': 0.2572, 'grad_norm': 16.60124969482422, 'learning_rate': 9.63939393939394e-06, 'num_tokens': 215614.0, 'completions/mean_length': 66.25, 'completions/min_length': 49.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.6359738111495972, 'rewards/meter/std': 0.3151750862598419, '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': 0.9953513741493225, 'rewards/repeat_soft/std': 0.003892518114298582, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.08124037832021713, 'rewards/total_composite/mean': 0.25074681639671326, 'rewards/total_composite/std': 0.107984259724617, 'reward': 0.25074681639671326, 'reward_std': 0.107984259724617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24169640243053436, 'sampling/sampling_logp_difference/max': 1.6041224002838135, 'sampling/importance_sampling_ratio/min': 0.20106592774391174, 'sampling/importance_sampling_ratio/mean': 1.0396535396575928, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5456028133630753, 'clip_ratio/low_mean': 0.11237266473472118, 'clip_ratio/low_min': 0.11237266473472118, 'clip_ratio/high_mean': 0.112040089443326, 'clip_ratio/high_max': 0.112040089443326, 'clip_ratio/region_mean': 0.22441275417804718, 'reward_total_mean': 0.25074681639671326, 'reward_meter_mean': 0.6359738111495972, 'reward_meter_std': 0.3151750862598419, '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': 0.9953513741493225, 'reward_repeat_soft_std': 0.003892518114298582, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.08124037832021713, 'reward_total_composite_mean': 0.25074681639671326, 'reward_total_composite_std': 0.107984259724617, 'epoch': 0.01}

  4%|▎         | 120/3300 [14:47<6:12:44,  7.03s/it]INFO 04-13 13:00:30 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 121/3300 [14:55<6:28:41,  7.34s/it]2026-04-13 13:00:38,274 | INFO | train_grpo_train | metrics_logged mode=train step=121

                                                    
{'loss': -0.0114, 'grad_norm': 10.472457885742188, 'learning_rate': 9.636363636363638e-06, 'num_tokens': 218108.0, 'completions/mean_length': 111.75, 'completions/min_length': 93.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.75, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.21533173322677612, 'rewards/meter/std': 0.13730649650096893, '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': 0.9945930242538452, 'rewards/repeat_soft/std': 0.004704488907009363, 'rewards/judge_quality/mean': 0.2199999988079071, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.10761401057243347, 'rewards/total_composite/std': 0.019664164632558823, 'reward': 0.10761401057243347, 'reward_std': 0.019664164632558823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2531386911869049, 'sampling/sampling_logp_difference/max': 1.755446434020996, 'sampling/importance_sampling_ratio/min': 0.17283006012439728, 'sampling/importance_sampling_ratio/mean': 1.0559929609298706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2308313846588135, 'clip_ratio/low_mean': 0.0991344191133976, 'clip_ratio/low_min': 0.0991344191133976, 'clip_ratio/high_mean': 0.10845869779586792, 'clip_ratio/high_max': 0.10845869779586792, 'clip_ratio/region_mean': 0.20759311690926552, 'reward_total_mean': 0.10761401057243347, 'reward_meter_mean': 0.21533173322677612, 'reward_meter_std': 0.13730649650096893, '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': 0.9945930242538452, 'reward_repeat_soft_std': 0.004704488907009363, 'reward_judge_quality_mean': 0.2199999988079071, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.10761401057243347, 'reward_total_composite_std': 0.019664164632558823, 'epoch': 0.01}

  4%|▎         | 121/3300 [14:55<6:28:41,  7.34s/it]INFO 04-13 13:00:38 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 122/3300 [15:03<6:32:26,  7.41s/it]2026-04-13 13:00:45,853 | INFO | train_grpo_train | metrics_logged mode=train step=122

                                                    
{'loss': 0.0792, 'grad_norm': 16.2755184173584, 'learning_rate': 9.633333333333335e-06, 'num_tokens': 219713.0, 'completions/mean_length': 42.625, 'completions/min_length': 34.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7699140906333923, 'rewards/meter/std': 0.3879815936088562, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9996742010116577, 'rewards/repeat_soft/std': 0.0009215619065798819, 'rewards/judge_quality/mean': 0.3449999988079071, 'rewards/judge_quality/std': 0.14880476891994476, 'rewards/total_composite/mean': 0.28357014060020447, 'rewards/total_composite/std': 0.19001367688179016, 'reward': 0.28357014060020447, 'reward_std': 0.19001366198062897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23517604172229767, 'sampling/sampling_logp_difference/max': 1.3021726608276367, 'sampling/importance_sampling_ratio/min': 0.27194032073020935, 'sampling/importance_sampling_ratio/mean': 1.0255141258239746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5738581717014313, 'clip_ratio/low_mean': 0.09402434760704637, 'clip_ratio/low_min': 0.09402434760704637, 'clip_ratio/high_mean': 0.07149197906255722, 'clip_ratio/high_max': 0.07149197906255722, 'clip_ratio/region_mean': 0.1655163266696036, 'reward_total_mean': 0.28357014060020447, 'reward_meter_mean': 0.7699140906333923, 'reward_meter_std': 0.3879815936088562, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9996742010116577, 'reward_repeat_soft_std': 0.0009215619065798819, 'reward_judge_quality_mean': 0.3449999988079071, 'reward_judge_quality_std': 0.14880476891994476, 'reward_total_composite_mean': 0.28357014060020447, 'reward_total_composite_std': 0.19001367688179016, 'epoch': 0.01}

  4%|▎         | 122/3300 [15:03<6:32:26,  7.41s/it]INFO 04-13 13:00:46 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 123/3300 [15:10<6:32:35,  7.41s/it]2026-04-13 13:00:53,279 | INFO | train_grpo_train | metrics_logged mode=train step=123

                                                    
{'loss': 0.0079, 'grad_norm': 25.827007293701172, 'learning_rate': 9.63030303030303e-06, 'num_tokens': 221099.0, 'completions/mean_length': 25.25, 'completions/min_length': 23.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.48179739713668823, 'rewards/meter/std': 0.37739667296409607, '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': 0.9988503456115723, 'rewards/repeat_soft/std': 0.002489580772817135, 'rewards/judge_quality/mean': 0.4987499713897705, 'rewards/judge_quality/std': 0.1769937425851822, 'rewards/total_composite/mean': 0.338995099067688, 'rewards/total_composite/std': 0.18322403728961945, 'reward': 0.338995099067688, 'reward_std': 0.18322403728961945, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2206656038761139, 'sampling/sampling_logp_difference/max': 2.0103414058685303, 'sampling/importance_sampling_ratio/min': 0.13394293189048767, 'sampling/importance_sampling_ratio/mean': 1.0055018663406372, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.435492642223835, 'clip_ratio/low_mean': 0.08452132297679782, 'clip_ratio/low_min': 0.08452132297679782, 'clip_ratio/high_mean': 0.07278680521994829, 'clip_ratio/high_max': 0.07278680521994829, 'clip_ratio/region_mean': 0.1573081281967461, 'reward_total_mean': 0.338995099067688, 'reward_meter_mean': 0.48179739713668823, 'reward_meter_std': 0.37739667296409607, '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': 0.9988503456115723, 'reward_repeat_soft_std': 0.002489580772817135, 'reward_judge_quality_mean': 0.4987499713897705, 'reward_judge_quality_std': 0.1769937425851822, 'reward_total_composite_mean': 0.338995099067688, 'reward_total_composite_std': 0.18322403728961945, 'epoch': 0.01}

  4%|▎         | 123/3300 [15:10<6:32:35,  7.41s/it]INFO 04-13 13:00:53 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 124/3300 [15:16<6:07:36,  6.94s/it]2026-04-13 13:00:59,133 | INFO | train_grpo_train | metrics_logged mode=train step=124

                                                    
{'loss': -0.0173, 'grad_norm': 23.003311157226562, 'learning_rate': 9.627272727272728e-06, 'num_tokens': 222427.0, 'completions/mean_length': 17.0, 'completions/min_length': 13.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.0, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.7263355255126953, 'rewards/meter/std': 0.4465133249759674, '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': 0.961837112903595, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.3387499749660492, 'rewards/judge_quality/std': 0.09538455307483673, 'rewards/total_composite/mean': 0.2906777858734131, 'rewards/total_composite/std': 0.14767445623874664, 'reward': 0.2906777858734131, 'reward_std': 0.14767445623874664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18750518560409546, 'sampling/sampling_logp_difference/max': 0.9341144561767578, 'sampling/importance_sampling_ratio/min': 0.3929336667060852, 'sampling/importance_sampling_ratio/mean': 1.0421078205108643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8559207618236542, 'clip_ratio/low_mean': 0.06732460018247366, 'clip_ratio/low_min': 0.06732460018247366, 'clip_ratio/high_mean': 0.12367216311395168, 'clip_ratio/high_max': 0.12367216311395168, 'clip_ratio/region_mean': 0.19099676329642534, 'reward_total_mean': 0.2906777858734131, 'reward_meter_mean': 0.7263355255126953, 'reward_meter_std': 0.4465133249759674, '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': 0.961837112903595, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.3387499749660492, 'reward_judge_quality_std': 0.09538455307483673, 'reward_total_composite_mean': 0.2906777858734131, 'reward_total_composite_std': 0.14767445623874664, 'epoch': 0.01}

  4%|▍         | 124/3300 [15:16<6:07:36,  6.94s/it]INFO 04-13 13:00:59 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 125/3300 [15:23<6:07:12,  6.94s/it]2026-04-13 13:01:06,055 | INFO | train_grpo_train | metrics_logged mode=train step=125

                                                    
{'loss': 0.0823, 'grad_norm': 19.11264991760254, 'learning_rate': 9.624242424242425e-06, 'num_tokens': 224245.0, 'completions/mean_length': 58.25, 'completions/min_length': 49.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.3744819164276123, 'rewards/meter/std': 0.2976786494255066, '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': 0.9952961206436157, 'rewards/repeat_soft/std': 0.0038182511925697327, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.14262838661670685, 'rewards/total_composite/mean': 0.17342031002044678, 'rewards/total_composite/std': 0.0660971999168396, 'reward': 0.17342031002044678, 'reward_std': 0.0660971999168396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2614648938179016, 'sampling/sampling_logp_difference/max': 2.5867881774902344, 'sampling/importance_sampling_ratio/min': 0.07526137679815292, 'sampling/importance_sampling_ratio/mean': 1.0358235836029053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.239851325750351, 'clip_ratio/low_mean': 0.1010020449757576, 'clip_ratio/low_min': 0.1010020449757576, 'clip_ratio/high_mean': 0.12148526310920715, 'clip_ratio/high_max': 0.12148526310920715, 'clip_ratio/region_mean': 0.22248730808496475, 'reward_total_mean': 0.17342031002044678, 'reward_meter_mean': 0.3744819164276123, 'reward_meter_std': 0.2976786494255066, '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': 0.9952961206436157, 'reward_repeat_soft_std': 0.0038182511925697327, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.14262838661670685, 'reward_total_composite_mean': 0.17342031002044678, 'reward_total_composite_std': 0.0660971999168396, 'epoch': 0.01}

  4%|▍         | 125/3300 [15:23<6:07:12,  6.94s/it]INFO 04-13 13:01:06 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 126/3300 [15:33<6:53:30,  7.82s/it]2026-04-13 13:01:15,921 | INFO | train_grpo_train | metrics_logged mode=train step=126

                                                    
{'loss': 0.2376, 'grad_norm': 19.68045425415039, 'learning_rate': 9.621212121212122e-06, 'num_tokens': 225588.0, 'completions/mean_length': 23.875, 'completions/min_length': 16.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.3367393910884857, 'rewards/meter/std': 0.44271981716156006, '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': 0.9657489061355591, 'rewards/repeat_soft/std': 0.014413577504456043, 'rewards/judge_quality/mean': 0.22624999284744263, 'rewards/judge_quality/std': 0.02386719174683094, 'rewards/total_composite/mean': 0.12879985570907593, 'rewards/total_composite/std': 0.06986033171415329, 'reward': 0.12879985570907593, 'reward_std': 0.06986033171415329, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2064133882522583, 'sampling/sampling_logp_difference/max': 1.1316213607788086, 'sampling/importance_sampling_ratio/min': 0.3225099444389343, 'sampling/importance_sampling_ratio/mean': 1.0287368297576904, 'sampling/importance_sampling_ratio/max': 1.9636597633361816, 'entropy': 2.439403772354126, 'clip_ratio/low_mean': 0.1330265123397112, 'clip_ratio/low_min': 0.1330265123397112, 'clip_ratio/high_mean': 0.07791610900312662, 'clip_ratio/high_max': 0.07791610900312662, 'clip_ratio/region_mean': 0.2109426213428378, 'reward_total_mean': 0.12879985570907593, 'reward_meter_mean': 0.3367393910884857, 'reward_meter_std': 0.44271981716156006, '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': 0.9657489061355591, 'reward_repeat_soft_std': 0.014413577504456043, 'reward_judge_quality_mean': 0.22624999284744263, 'reward_judge_quality_std': 0.02386719174683094, 'reward_total_composite_mean': 0.12879985570907593, 'reward_total_composite_std': 0.06986033171415329, 'epoch': 0.01}

  4%|▍         | 126/3300 [15:33<6:53:30,  7.82s/it]INFO 04-13 13:01:16 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 127/3300 [15:39<6:32:10,  7.42s/it]2026-04-13 13:01:22,411 | INFO | train_grpo_train | metrics_logged mode=train step=127

                                                    
{'loss': 0.0518, 'grad_norm': 21.802650451660156, 'learning_rate': 9.61818181818182e-06, 'num_tokens': 227064.0, 'completions/mean_length': 38.5, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6797225475311279, 'rewards/meter/std': 0.3927367329597473, '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': 0.9996022582054138, 'rewards/repeat_soft/std': 0.001124941511079669, 'rewards/judge_quality/mean': 0.26375001668930054, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.2068263292312622, 'rewards/total_composite/std': 0.06877505034208298, 'reward': 0.2068263292312622, 'reward_std': 0.06877505779266357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23567445576190948, 'sampling/sampling_logp_difference/max': 1.6988935470581055, 'sampling/importance_sampling_ratio/min': 0.1828857660293579, 'sampling/importance_sampling_ratio/mean': 1.067500114440918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.59160253405571, 'clip_ratio/low_mean': 0.04095253814011812, 'clip_ratio/low_min': 0.04095253814011812, 'clip_ratio/high_mean': 0.15336354076862335, 'clip_ratio/high_max': 0.15336354076862335, 'clip_ratio/region_mean': 0.19431607890874147, 'reward_total_mean': 0.2068263292312622, 'reward_meter_mean': 0.6797225475311279, 'reward_meter_std': 0.3927367329597473, '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': 0.9996022582054138, 'reward_repeat_soft_std': 0.001124941511079669, 'reward_judge_quality_mean': 0.26375001668930054, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.2068263292312622, 'reward_total_composite_std': 0.06877505034208298, 'epoch': 0.01}

  4%|▍         | 127/3300 [15:40<6:32:10,  7.42s/it]INFO 04-13 13:01:22 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 128/3300 [15:46<6:20:22,  7.20s/it]2026-04-13 13:01:29,080 | INFO | train_grpo_train | metrics_logged mode=train step=128

                                                    
{'loss': 0.0646, 'grad_norm': 13.482661247253418, 'learning_rate': 9.615151515151517e-06, 'num_tokens': 228989.0, 'completions/mean_length': 67.625, 'completions/min_length': 58.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.3872728943824768, 'rewards/meter/std': 0.32985642552375793, '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': 0.9928202629089355, 'rewards/repeat_soft/std': 0.006111267488449812, 'rewards/judge_quality/mean': 0.3199999928474426, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/total_composite/mean': 0.19855639338493347, 'rewards/total_composite/std': 0.15774224698543549, 'reward': 0.19855639338493347, 'reward_std': 0.15774224698543549, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22792702913284302, 'sampling/sampling_logp_difference/max': 1.5525579452514648, 'sampling/importance_sampling_ratio/min': 0.21170574426651, 'sampling/importance_sampling_ratio/mean': 1.0355592966079712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.207075983285904, 'clip_ratio/low_mean': 0.11547990329563618, 'clip_ratio/low_min': 0.11547990329563618, 'clip_ratio/high_mean': 0.08882211707532406, 'clip_ratio/high_max': 0.08882211707532406, 'clip_ratio/region_mean': 0.20430202037096024, 'reward_total_mean': 0.19855639338493347, 'reward_meter_mean': 0.3872728943824768, 'reward_meter_std': 0.32985642552375793, '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': 0.9928202629089355, 'reward_repeat_soft_std': 0.006111267488449812, 'reward_judge_quality_mean': 0.3199999928474426, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.19855639338493347, 'reward_total_composite_std': 0.15774224698543549, 'epoch': 0.01}

  4%|▍         | 128/3300 [15:46<6:20:22,  7.20s/it]INFO 04-13 13:01:29 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 129/3300 [15:53<6:07:58,  6.96s/it]2026-04-13 13:01:35,502 | INFO | train_grpo_train | metrics_logged mode=train step=129

                                                    
{'loss': 0.0638, 'grad_norm': 10.612187385559082, 'learning_rate': 9.612121212121212e-06, 'num_tokens': 230746.0, 'completions/mean_length': 66.625, 'completions/min_length': 61.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.5210334062576294, 'rewards/meter/std': 0.3286823332309723, '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': 0.9986972808837891, 'rewards/repeat_soft/std': 0.0018164877546951175, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.03207135200500488, 'rewards/total_composite/mean': 0.17536430060863495, 'rewards/total_composite/std': 0.06717904657125473, 'reward': 0.17536430060863495, 'reward_std': 0.06717904657125473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22001968324184418, 'sampling/sampling_logp_difference/max': 1.4705123901367188, 'sampling/importance_sampling_ratio/min': 0.22980770468711853, 'sampling/importance_sampling_ratio/mean': 1.0479635000228882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.956223636865616, 'clip_ratio/low_mean': 0.10533072054386139, 'clip_ratio/low_min': 0.10533072054386139, 'clip_ratio/high_mean': 0.10481791384518147, 'clip_ratio/high_max': 0.10481791384518147, 'clip_ratio/region_mean': 0.21014863438904285, 'reward_total_mean': 0.17536430060863495, 'reward_meter_mean': 0.5210334062576294, 'reward_meter_std': 0.3286823332309723, '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': 0.9986972808837891, 'reward_repeat_soft_std': 0.0018164877546951175, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.03207135200500488, 'reward_total_composite_mean': 0.17536430060863495, 'reward_total_composite_std': 0.06717904657125473, 'epoch': 0.01}

  4%|▍         | 129/3300 [15:53<6:07:58,  6.96s/it]INFO 04-13 13:01:35 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 130/3300 [16:00<6:08:37,  6.98s/it]2026-04-13 13:01:42,514 | INFO | train_grpo_train | metrics_logged mode=train step=130

                                                    
{'loss': -0.0281, 'grad_norm': 11.608769416809082, 'learning_rate': 9.60909090909091e-06, 'num_tokens': 232821.0, 'completions/mean_length': 82.375, 'completions/min_length': 71.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.375, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.3977913558483124, 'rewards/meter/std': 0.32337522506713867, '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': 0.9964189529418945, 'rewards/repeat_soft/std': 0.0026952349580824375, 'rewards/judge_quality/mean': 0.22749999165534973, 'rewards/judge_quality/std': 0.02121320553123951, 'rewards/total_composite/mean': 0.13609255850315094, 'rewards/total_composite/std': 0.04253906011581421, 'reward': 0.13609255850315094, 'reward_std': 0.04253906011581421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24335147440433502, 'sampling/sampling_logp_difference/max': 2.166168212890625, 'sampling/importance_sampling_ratio/min': 0.11461596190929413, 'sampling/importance_sampling_ratio/mean': 1.0427734851837158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3111905455589294, 'clip_ratio/low_mean': 0.09228312410414219, 'clip_ratio/low_min': 0.09228312410414219, 'clip_ratio/high_mean': 0.11335889808833599, 'clip_ratio/high_max': 0.11335889808833599, 'clip_ratio/region_mean': 0.20564202219247818, 'reward_total_mean': 0.13609255850315094, 'reward_meter_mean': 0.3977913558483124, 'reward_meter_std': 0.32337522506713867, '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': 0.9964189529418945, 'reward_repeat_soft_std': 0.0026952349580824375, 'reward_judge_quality_mean': 0.22749999165534973, 'reward_judge_quality_std': 0.02121320553123951, 'reward_total_composite_mean': 0.13609255850315094, 'reward_total_composite_std': 0.04253906011581421, 'epoch': 0.01}

  4%|▍         | 130/3300 [16:00<6:08:37,  6.98s/it]INFO 04-13 13:01:42 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 131/3300 [16:06<5:52:14,  6.67s/it]2026-04-13 13:01:48,463 | INFO | train_grpo_train | metrics_logged mode=train step=131

                                                    
{'loss': 0.0339, 'grad_norm': 24.785491943359375, 'learning_rate': 9.606060606060607e-06, 'num_tokens': 234316.0, 'completions/mean_length': 28.875, 'completions/min_length': 23.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5526943206787109, 'rewards/meter/std': 0.35744738578796387, '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': 0.9920192360877991, 'rewards/repeat_soft/std': 0.006725827232003212, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.13845448195934296, 'rewards/total_composite/mean': 0.26369211077690125, 'rewards/total_composite/std': 0.16711732745170593, 'reward': 0.26369211077690125, 'reward_std': 0.16711732745170593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26848104596138, 'sampling/sampling_logp_difference/max': 4.247650146484375, 'sampling/importance_sampling_ratio/min': 0.014297792688012123, 'sampling/importance_sampling_ratio/mean': 0.9865650534629822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6055498570203781, 'clip_ratio/low_mean': 0.14761552587151527, 'clip_ratio/low_min': 0.14761552587151527, 'clip_ratio/high_mean': 0.0752564100548625, 'clip_ratio/high_max': 0.0752564100548625, 'clip_ratio/region_mean': 0.22287193592637777, 'reward_total_mean': 0.26369211077690125, 'reward_meter_mean': 0.5526943206787109, 'reward_meter_std': 0.35744738578796387, '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': 0.9920192360877991, 'reward_repeat_soft_std': 0.006725827232003212, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.13845448195934296, 'reward_total_composite_mean': 0.26369211077690125, 'reward_total_composite_std': 0.16711732745170593, 'epoch': 0.01}

  4%|▍         | 131/3300 [16:06<5:52:14,  6.67s/it]INFO 04-13 13:01:48 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 132/3300 [16:13<6:09:59,  7.01s/it]2026-04-13 13:01:56,260 | INFO | train_grpo_train | metrics_logged mode=train step=132

                                                    
{'loss': 0.0736, 'grad_norm': 10.571706771850586, 'learning_rate': 9.603030303030304e-06, 'num_tokens': 236652.0, 'completions/mean_length': 89.0, 'completions/min_length': 76.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.47271233797073364, 'rewards/meter/std': 0.3446253836154938, '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': 0.994101881980896, 'rewards/repeat_soft/std': 0.0062068188562989235, 'rewards/judge_quality/mean': 0.23499999940395355, 'rewards/judge_quality/std': 0.02777460403740406, 'rewards/total_composite/mean': 0.15211760997772217, 'rewards/total_composite/std': 0.04623790457844734, 'reward': 0.15211760997772217, 'reward_std': 0.04623790457844734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21054711937904358, 'sampling/sampling_logp_difference/max': 1.3416767120361328, 'sampling/importance_sampling_ratio/min': 0.2614069879055023, 'sampling/importance_sampling_ratio/mean': 1.0391924381256104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.75347238779068, 'clip_ratio/low_mean': 0.09103407897055149, 'clip_ratio/low_min': 0.09103407897055149, 'clip_ratio/high_mean': 0.10133074037730694, 'clip_ratio/high_max': 0.10133074037730694, 'clip_ratio/region_mean': 0.19236481934785843, 'reward_total_mean': 0.15211760997772217, 'reward_meter_mean': 0.47271233797073364, 'reward_meter_std': 0.3446253836154938, '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': 0.994101881980896, 'reward_repeat_soft_std': 0.0062068188562989235, 'reward_judge_quality_mean': 0.23499999940395355, 'reward_judge_quality_std': 0.02777460403740406, 'reward_total_composite_mean': 0.15211760997772217, 'reward_total_composite_std': 0.04623790457844734, 'epoch': 0.01}

  4%|▍         | 132/3300 [16:13<6:09:59,  7.01s/it]INFO 04-13 13:01:56 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 133/3300 [16:20<6:09:46,  7.01s/it]2026-04-13 13:02:03,261 | INFO | train_grpo_train | metrics_logged mode=train step=133

                                                    
{'loss': 0.0575, 'grad_norm': 15.153925895690918, 'learning_rate': 9.600000000000001e-06, 'num_tokens': 238734.0, 'completions/mean_length': 61.25, 'completions/min_length': 52.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.25, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.7094472646713257, 'rewards/meter/std': 0.2717290222644806, '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': 0.99950110912323, 'rewards/repeat_soft/std': 0.0009013467351906002, 'rewards/judge_quality/mean': 0.23499999940395355, 'rewards/judge_quality/std': 0.02777460403740406, 'rewards/total_composite/mean': 0.19090116024017334, 'rewards/total_composite/std': 0.048894003033638, 'reward': 0.19090116024017334, 'reward_std': 0.0488939993083477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22949044406414032, 'sampling/sampling_logp_difference/max': 1.3840246200561523, 'sampling/importance_sampling_ratio/min': 0.25056806206703186, 'sampling/importance_sampling_ratio/mean': 1.056289792060852, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7774215042591095, 'clip_ratio/low_mean': 0.08348813280463219, 'clip_ratio/low_min': 0.08348813280463219, 'clip_ratio/high_mean': 0.11317358072847128, 'clip_ratio/high_max': 0.11317358072847128, 'clip_ratio/region_mean': 0.19666171353310347, 'reward_total_mean': 0.19090116024017334, 'reward_meter_mean': 0.7094472646713257, 'reward_meter_std': 0.2717290222644806, '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': 0.99950110912323, 'reward_repeat_soft_std': 0.0009013467351906002, 'reward_judge_quality_mean': 0.23499999940395355, 'reward_judge_quality_std': 0.02777460403740406, 'reward_total_composite_mean': 0.19090116024017334, 'reward_total_composite_std': 0.048894003033638, 'epoch': 0.01}

  4%|▍         | 133/3300 [16:20<6:09:46,  7.01s/it]INFO 04-13 13:02:03 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 134/3300 [16:28<6:12:12,  7.05s/it]2026-04-13 13:02:10,427 | INFO | train_grpo_train | metrics_logged mode=train step=134

                                                    
{'loss': -0.0098, 'grad_norm': 9.651640892028809, 'learning_rate': 9.596969696969699e-06, 'num_tokens': 241075.0, 'completions/mean_length': 99.625, 'completions/min_length': 92.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.625, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9560881853103638, 'rewards/meter/std': 0.0939415693283081, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9977115988731384, 'rewards/repeat_soft/std': 0.0038488158024847507, 'rewards/judge_quality/mean': 0.2199999988079071, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.18523168563842773, 'rewards/total_composite/std': 0.0759832113981247, 'reward': 0.18523168563842773, 'reward_std': 0.0759832039475441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23185819387435913, 'sampling/sampling_logp_difference/max': 1.5198945999145508, 'sampling/importance_sampling_ratio/min': 0.2187349498271942, 'sampling/importance_sampling_ratio/mean': 1.0410938262939453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7694606482982635, 'clip_ratio/low_mean': 0.04217284731566906, 'clip_ratio/low_min': 0.04217284731566906, 'clip_ratio/high_mean': 0.18183892592787743, 'clip_ratio/high_max': 0.18183892592787743, 'clip_ratio/region_mean': 0.22401177324354649, 'reward_total_mean': 0.18523168563842773, 'reward_meter_mean': 0.9560881853103638, 'reward_meter_std': 0.0939415693283081, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9977115988731384, 'reward_repeat_soft_std': 0.0038488158024847507, 'reward_judge_quality_mean': 0.2199999988079071, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.18523168563842773, 'reward_total_composite_std': 0.0759832113981247, 'epoch': 0.01}

  4%|▍         | 134/3300 [16:28<6:12:12,  7.05s/it]INFO 04-13 13:02:10 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 135/3300 [16:34<6:10:28,  7.02s/it]2026-04-13 13:02:17,380 | INFO | train_grpo_train | metrics_logged mode=train step=135

                                                    
{'loss': 0.0927, 'grad_norm': 10.741291999816895, 'learning_rate': 9.593939393939394e-06, 'num_tokens': 243080.0, 'completions/mean_length': 84.625, 'completions/min_length': 69.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.625, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.5973300933837891, 'rewards/meter/std': 0.3906180262565613, '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': 0.9978720545768738, 'rewards/repeat_soft/std': 0.0020282238256186247, 'rewards/judge_quality/mean': 0.20000000298023224, 'rewards/judge_quality/std': 0.021380895748734474, 'rewards/total_composite/mean': 0.14806291460990906, 'rewards/total_composite/std': 0.05354469642043114, 'reward': 0.14806291460990906, 'reward_std': 0.05354469269514084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23586569726467133, 'sampling/sampling_logp_difference/max': 1.169741153717041, 'sampling/importance_sampling_ratio/min': 0.3260723352432251, 'sampling/importance_sampling_ratio/mean': 1.0564217567443848, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.5219001173973083, 'clip_ratio/low_mean': 0.08202693238854408, 'clip_ratio/low_min': 0.08202693238854408, 'clip_ratio/high_mean': 0.11562965251505375, 'clip_ratio/high_max': 0.11562965251505375, 'clip_ratio/region_mean': 0.19765658490359783, 'reward_total_mean': 0.14806291460990906, 'reward_meter_mean': 0.5973300933837891, 'reward_meter_std': 0.3906180262565613, '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': 0.9978720545768738, 'reward_repeat_soft_std': 0.0020282238256186247, 'reward_judge_quality_mean': 0.20000000298023224, 'reward_judge_quality_std': 0.021380895748734474, 'reward_total_composite_mean': 0.14806291460990906, 'reward_total_composite_std': 0.05354469642043114, 'epoch': 0.01}

  4%|▍         | 135/3300 [16:34<6:10:28,  7.02s/it]INFO 04-13 13:02:17 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 136/3300 [16:42<6:19:00,  7.19s/it]2026-04-13 13:02:24,959 | INFO | train_grpo_train | metrics_logged mode=train step=136

                                                    
{'loss': 0.0637, 'grad_norm': 13.939590454101562, 'learning_rate': 9.590909090909091e-06, 'num_tokens': 244994.0, 'completions/mean_length': 59.25, 'completions/min_length': 46.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.25, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.4887017607688904, 'rewards/meter/std': 0.396069198846817, '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': 0.9978152513504028, 'rewards/repeat_soft/std': 0.003188708098605275, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.21596543490886688, 'rewards/total_composite/mean': 0.27008306980133057, 'rewards/total_composite/std': 0.1957637369632721, 'reward': 0.27008306980133057, 'reward_std': 0.1957637369632721, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21498402953147888, 'sampling/sampling_logp_difference/max': 1.6607599258422852, 'sampling/importance_sampling_ratio/min': 0.18999452888965607, 'sampling/importance_sampling_ratio/mean': 1.046618103981018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5923212319612503, 'clip_ratio/low_mean': 0.11506733112037182, 'clip_ratio/low_min': 0.11506733112037182, 'clip_ratio/high_mean': 0.11947667319327593, 'clip_ratio/high_max': 0.11947667319327593, 'clip_ratio/region_mean': 0.23454400431364775, 'reward_total_mean': 0.27008306980133057, 'reward_meter_mean': 0.4887017607688904, 'reward_meter_std': 0.396069198846817, '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': 0.9978152513504028, 'reward_repeat_soft_std': 0.003188708098605275, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.21596543490886688, 'reward_total_composite_mean': 0.27008306980133057, 'reward_total_composite_std': 0.1957637369632721, 'epoch': 0.01}

  4%|▍         | 136/3300 [16:42<6:19:00,  7.19s/it]INFO 04-13 13:02:25 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 137/3300 [16:48<6:05:12,  6.93s/it]2026-04-13 13:02:31,272 | INFO | train_grpo_train | metrics_logged mode=train step=137

                                                    
{'loss': 0.0725, 'grad_norm': 20.54743194580078, 'learning_rate': 9.587878787878789e-06, 'num_tokens': 246492.0, 'completions/mean_length': 33.25, 'completions/min_length': 29.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.38131028413772583, 'rewards/meter/std': 0.38800758123397827, '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': 0.9944081902503967, 'rewards/repeat_soft/std': 0.00883419718593359, 'rewards/judge_quality/mean': 0.2524999976158142, 'rewards/judge_quality/std': 0.07086203992366791, 'rewards/total_composite/mean': 0.13826440274715424, 'rewards/total_composite/std': 0.049926865845918655, 'reward': 0.13826440274715424, 'reward_std': 0.049926865845918655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2485567033290863, 'sampling/sampling_logp_difference/max': 1.640737771987915, 'sampling/importance_sampling_ratio/min': 0.1938369870185852, 'sampling/importance_sampling_ratio/mean': 1.0484215021133423, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.452331081032753, 'clip_ratio/low_mean': 0.1011201636865735, 'clip_ratio/low_min': 0.1011201636865735, 'clip_ratio/high_mean': 0.05779710691422224, 'clip_ratio/high_max': 0.05779710691422224, 'clip_ratio/region_mean': 0.15891727060079575, 'reward_total_mean': 0.13826440274715424, 'reward_meter_mean': 0.38131028413772583, 'reward_meter_std': 0.38800758123397827, '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': 0.9944081902503967, 'reward_repeat_soft_std': 0.00883419718593359, 'reward_judge_quality_mean': 0.2524999976158142, 'reward_judge_quality_std': 0.07086203992366791, 'reward_total_composite_mean': 0.13826440274715424, 'reward_total_composite_std': 0.049926865845918655, 'epoch': 0.01}

  4%|▍         | 137/3300 [16:48<6:05:12,  6.93s/it]INFO 04-13 13:02:31 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 138/3300 [16:54<5:47:04,  6.59s/it]2026-04-13 13:02:37,060 | INFO | train_grpo_train | metrics_logged mode=train step=138

                                                    
{'loss': 0.0486, 'grad_norm': 25.14579963684082, 'learning_rate': 9.584848484848486e-06, 'num_tokens': 248096.0, 'completions/mean_length': 31.5, 'completions/min_length': 27.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.45216143131256104, 'rewards/meter/std': 0.3622969090938568, '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': 0.9938479661941528, 'rewards/repeat_soft/std': 0.008477476425468922, 'rewards/judge_quality/mean': 0.26750001311302185, 'rewards/judge_quality/std': 0.07497618347406387, 'rewards/total_composite/mean': 0.17517951130867004, 'rewards/total_composite/std': 0.0906360074877739, 'reward': 0.17517951130867004, 'reward_std': 0.0906360000371933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22594934701919556, 'sampling/sampling_logp_difference/max': 1.1099905967712402, 'sampling/importance_sampling_ratio/min': 0.32956206798553467, 'sampling/importance_sampling_ratio/mean': 1.0708420276641846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.285018354654312, 'clip_ratio/low_mean': 0.13230955973267555, 'clip_ratio/low_min': 0.13230955973267555, 'clip_ratio/high_mean': 0.09946177713572979, 'clip_ratio/high_max': 0.09946177713572979, 'clip_ratio/region_mean': 0.23177133686840534, 'reward_total_mean': 0.17517951130867004, 'reward_meter_mean': 0.45216143131256104, 'reward_meter_std': 0.3622969090938568, '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': 0.9938479661941528, 'reward_repeat_soft_std': 0.008477476425468922, 'reward_judge_quality_mean': 0.26750001311302185, 'reward_judge_quality_std': 0.07497618347406387, 'reward_total_composite_mean': 0.17517951130867004, 'reward_total_composite_std': 0.0906360074877739, 'epoch': 0.01}

  4%|▍         | 138/3300 [16:54<5:47:04,  6.59s/it]INFO 04-13 13:02:37 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 139/3300 [17:00<5:36:42,  6.39s/it]2026-04-13 13:02:42,996 | INFO | train_grpo_train | metrics_logged mode=train step=139

                                                    
{'loss': 0.0586, 'grad_norm': 15.347241401672363, 'learning_rate': 9.581818181818181e-06, 'num_tokens': 249650.0, 'completions/mean_length': 43.25, 'completions/min_length': 35.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.512442946434021, 'rewards/meter/std': 0.4792117178440094, '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': 0.9996022582054138, 'rewards/repeat_soft/std': 0.001124941511079669, 'rewards/judge_quality/mean': 0.23499999940395355, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.15505144000053406, 'rewards/total_composite/std': 0.07194910943508148, 'reward': 0.15505144000053406, 'reward_std': 0.07194910943508148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2576281726360321, 'sampling/sampling_logp_difference/max': 1.5772953033447266, 'sampling/importance_sampling_ratio/min': 0.20653294026851654, 'sampling/importance_sampling_ratio/mean': 1.0444402694702148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1368479430675507, 'clip_ratio/low_mean': 0.10965734161436558, 'clip_ratio/low_min': 0.10965734161436558, 'clip_ratio/high_mean': 0.1077451454475522, 'clip_ratio/high_max': 0.1077451454475522, 'clip_ratio/region_mean': 0.21740248706191778, 'reward_total_mean': 0.15505144000053406, 'reward_meter_mean': 0.512442946434021, 'reward_meter_std': 0.4792117178440094, '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': 0.9996022582054138, 'reward_repeat_soft_std': 0.001124941511079669, 'reward_judge_quality_mean': 0.23499999940395355, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.15505144000053406, 'reward_total_composite_std': 0.07194910943508148, 'epoch': 0.01}

  4%|▍         | 139/3300 [17:00<5:36:42,  6.39s/it]INFO 04-13 13:02:43 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 140/3300 [17:07<5:37:49,  6.41s/it]2026-04-13 13:02:49,466 | INFO | train_grpo_train | metrics_logged mode=train step=140

                                                    
{'loss': 0.1773, 'grad_norm': 31.55306053161621, 'learning_rate': 9.57878787878788e-06, 'num_tokens': 251181.0, 'completions/mean_length': 33.375, 'completions/min_length': 25.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5141806602478027, 'rewards/meter/std': 0.3833635151386261, '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': 0.9908744096755981, 'rewards/repeat_soft/std': 0.01445451844483614, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.21022097766399384, 'rewards/total_composite/mean': 0.2127482146024704, 'rewards/total_composite/std': 0.16522666811943054, 'reward': 0.2127482146024704, 'reward_std': 0.16522665321826935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.29307568073272705, 'sampling/sampling_logp_difference/max': 3.3791017532348633, 'sampling/importance_sampling_ratio/min': 0.034078050404787064, 'sampling/importance_sampling_ratio/mean': 0.9967068433761597, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4846941903233528, 'clip_ratio/low_mean': 0.15150748565793037, 'clip_ratio/low_min': 0.15150748565793037, 'clip_ratio/high_mean': 0.06666666641831398, 'clip_ratio/high_max': 0.06666666641831398, 'clip_ratio/region_mean': 0.21817415207624435, 'reward_total_mean': 0.2127482146024704, 'reward_meter_mean': 0.5141806602478027, 'reward_meter_std': 0.3833635151386261, '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': 0.9908744096755981, 'reward_repeat_soft_std': 0.01445451844483614, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.21022097766399384, 'reward_total_composite_mean': 0.2127482146024704, 'reward_total_composite_std': 0.16522666811943054, 'epoch': 0.01}

  4%|▍         | 140/3300 [17:07<5:37:49,  6.41s/it]INFO 04-13 13:02:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 141/3300 [17:16<6:25:15,  7.32s/it]2026-04-13 13:02:58,889 | INFO | train_grpo_train | metrics_logged mode=train step=141

                                                    
{'loss': 0.0703, 'grad_norm': 10.269293785095215, 'learning_rate': 9.575757575757576e-06, 'num_tokens': 253880.0, 'completions/mean_length': 131.375, 'completions/min_length': 114.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.41031938791275024, 'rewards/meter/std': 0.23784169554710388, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9922277927398682, 'rewards/repeat_soft/std': 0.007713355589658022, 'rewards/judge_quality/mean': 0.20749999582767487, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.11861090362071991, 'rewards/total_composite/std': 0.0585237592458725, 'reward': 0.11861090362071991, 'reward_std': 0.0585237555205822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.239677295088768, 'sampling/sampling_logp_difference/max': 1.5946722030639648, 'sampling/importance_sampling_ratio/min': 0.2029750645160675, 'sampling/importance_sampling_ratio/mean': 1.0555769205093384, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2069272100925446, 'clip_ratio/low_mean': 0.11101297475397587, 'clip_ratio/low_min': 0.11101297475397587, 'clip_ratio/high_mean': 0.12742743827402592, 'clip_ratio/high_max': 0.12742743827402592, 'clip_ratio/region_mean': 0.23844041302800179, 'reward_total_mean': 0.11861090362071991, 'reward_meter_mean': 0.41031938791275024, 'reward_meter_std': 0.23784169554710388, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9922277927398682, 'reward_repeat_soft_std': 0.007713355589658022, 'reward_judge_quality_mean': 0.20749999582767487, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.11861090362071991, 'reward_total_composite_std': 0.0585237592458725, 'epoch': 0.01}

  4%|▍         | 141/3300 [17:16<6:25:15,  7.32s/it]INFO 04-13 13:02:59 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 142/3300 [17:22<6:03:51,  6.91s/it]2026-04-13 13:03:04,859 | INFO | train_grpo_train | metrics_logged mode=train step=142

                                                    
{'loss': 0.1096, 'grad_norm': 14.984911918640137, 'learning_rate': 9.572727272727273e-06, 'num_tokens': 255471.0, 'completions/mean_length': 39.875, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8347629904747009, 'rewards/meter/std': 0.2015080600976944, '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': 0.9970236420631409, 'rewards/repeat_soft/std': 0.005287198815494776, 'rewards/judge_quality/mean': 0.2487500011920929, 'rewards/judge_quality/std': 0.0699872374534607, 'rewards/total_composite/mean': 0.20734521746635437, 'rewards/total_composite/std': 0.10832992941141129, 'reward': 0.20734521746635437, 'reward_std': 0.10832992196083069, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23098348081111908, 'sampling/sampling_logp_difference/max': 1.5200281143188477, 'sampling/importance_sampling_ratio/min': 0.21870574355125427, 'sampling/importance_sampling_ratio/mean': 1.059181571006775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3084492683410645, 'clip_ratio/low_mean': 0.07645521871745586, 'clip_ratio/low_min': 0.07645521871745586, 'clip_ratio/high_mean': 0.12622801680117846, 'clip_ratio/high_max': 0.12622801680117846, 'clip_ratio/region_mean': 0.20268323551863432, 'reward_total_mean': 0.20734521746635437, 'reward_meter_mean': 0.8347629904747009, 'reward_meter_std': 0.2015080600976944, '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': 0.9970236420631409, 'reward_repeat_soft_std': 0.005287198815494776, 'reward_judge_quality_mean': 0.2487500011920929, 'reward_judge_quality_std': 0.0699872374534607, 'reward_total_composite_mean': 0.20734521746635437, 'reward_total_composite_std': 0.10832992941141129, 'epoch': 0.01}

  4%|▍         | 142/3300 [17:22<6:03:51,  6.91s/it]INFO 04-13 13:03:05 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 143/3300 [17:29<6:07:14,  6.98s/it]2026-04-13 13:03:11,995 | INFO | train_grpo_train | metrics_logged mode=train step=143

                                                    
{'loss': 0.0808, 'grad_norm': 12.276087760925293, 'learning_rate': 9.56969696969697e-06, 'num_tokens': 257848.0, 'completions/mean_length': 110.125, 'completions/min_length': 96.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.125, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.6634474992752075, 'rewards/meter/std': 0.32899340987205505, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.996049165725708, 'rewards/repeat_soft/std': 0.002863870235159993, 'rewards/judge_quality/mean': 0.2224999964237213, 'rewards/judge_quality/std': 0.027124052867293358, 'rewards/total_composite/mean': 0.16733479499816895, 'rewards/total_composite/std': 0.07748735696077347, 'reward': 0.16733479499816895, 'reward_std': 0.07748734951019287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24100278317928314, 'sampling/sampling_logp_difference/max': 1.3480043411254883, 'sampling/importance_sampling_ratio/min': 0.2597581446170807, 'sampling/importance_sampling_ratio/mean': 1.0553762912750244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7829255759716034, 'clip_ratio/low_mean': 0.09594781510531902, 'clip_ratio/low_min': 0.09594781510531902, 'clip_ratio/high_mean': 0.1016995906829834, 'clip_ratio/high_max': 0.1016995906829834, 'clip_ratio/region_mean': 0.19764740578830242, 'reward_total_mean': 0.16733479499816895, 'reward_meter_mean': 0.6634474992752075, 'reward_meter_std': 0.32899340987205505, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.996049165725708, 'reward_repeat_soft_std': 0.002863870235159993, 'reward_judge_quality_mean': 0.2224999964237213, 'reward_judge_quality_std': 0.027124052867293358, 'reward_total_composite_mean': 0.16733479499816895, 'reward_total_composite_std': 0.07748735696077347, 'epoch': 0.01}

  4%|▍         | 143/3300 [17:29<6:07:14,  6.98s/it]INFO 04-13 13:03:12 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 144/3300 [17:35<5:54:11,  6.73s/it]2026-04-13 13:03:18,154 | INFO | train_grpo_train | metrics_logged mode=train step=144

                                                    
{'loss': 0.0569, 'grad_norm': 16.01355743408203, 'learning_rate': 9.566666666666668e-06, 'num_tokens': 259556.0, 'completions/mean_length': 41.5, 'completions/min_length': 34.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.34600621461868286, 'rewards/meter/std': 0.3382630944252014, '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': 0.9991676211357117, 'rewards/repeat_soft/std': 0.0013059456832706928, 'rewards/judge_quality/mean': 0.2724999785423279, 'rewards/judge_quality/std': 0.0936177596449852, 'rewards/total_composite/mean': 0.14613009989261627, 'rewards/total_composite/std': 0.05449230223894119, 'reward': 0.14613009989261627, 'reward_std': 0.054492298513650894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22496208548545837, 'sampling/sampling_logp_difference/max': 1.772200584411621, 'sampling/importance_sampling_ratio/min': 0.16995856165885925, 'sampling/importance_sampling_ratio/mean': 1.0439956188201904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9176801443099976, 'clip_ratio/low_mean': 0.08852885197848082, 'clip_ratio/low_min': 0.08852885197848082, 'clip_ratio/high_mean': 0.09302842989563942, 'clip_ratio/high_max': 0.09302842989563942, 'clip_ratio/region_mean': 0.18155728187412024, 'reward_total_mean': 0.14613009989261627, 'reward_meter_mean': 0.34600621461868286, 'reward_meter_std': 0.3382630944252014, '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': 0.9991676211357117, 'reward_repeat_soft_std': 0.0013059456832706928, 'reward_judge_quality_mean': 0.2724999785423279, 'reward_judge_quality_std': 0.0936177596449852, 'reward_total_composite_mean': 0.14613009989261627, 'reward_total_composite_std': 0.05449230223894119, 'epoch': 0.01}

  4%|▍         | 144/3300 [17:35<5:54:11,  6.73s/it]INFO 04-13 13:03:18 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 145/3300 [17:42<5:51:32,  6.69s/it]2026-04-13 13:03:24,729 | INFO | train_grpo_train | metrics_logged mode=train step=145

                                                    
{'loss': 0.0307, 'grad_norm': 21.03781509399414, 'learning_rate': 9.563636363636365e-06, 'num_tokens': 261063.0, 'completions/mean_length': 33.375, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.26710832118988037, 'rewards/meter/std': 0.30466634035110474, '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': 0.9970107078552246, 'rewards/repeat_soft/std': 0.003531090682372451, 'rewards/judge_quality/mean': 0.32375001907348633, 'rewards/judge_quality/std': 0.21206046640872955, 'rewards/total_composite/mean': 0.16713756322860718, 'rewards/total_composite/std': 0.12282278388738632, 'reward': 0.16713756322860718, 'reward_std': 0.12282278388738632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23675867915153503, 'sampling/sampling_logp_difference/max': 1.1583709716796875, 'sampling/importance_sampling_ratio/min': 0.3139972984790802, 'sampling/importance_sampling_ratio/mean': 0.9947248101234436, 'sampling/importance_sampling_ratio/max': 1.9616643190383911, 'entropy': 2.8966290950775146, 'clip_ratio/low_mean': 0.11045025661587715, 'clip_ratio/low_min': 0.11045025661587715, 'clip_ratio/high_mean': 0.062420424073934555, 'clip_ratio/high_max': 0.062420424073934555, 'clip_ratio/region_mean': 0.1728706806898117, 'reward_total_mean': 0.16713756322860718, 'reward_meter_mean': 0.26710832118988037, 'reward_meter_std': 0.30466634035110474, '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': 0.9970107078552246, 'reward_repeat_soft_std': 0.003531090682372451, 'reward_judge_quality_mean': 0.32375001907348633, 'reward_judge_quality_std': 0.21206046640872955, 'reward_total_composite_mean': 0.16713756322860718, 'reward_total_composite_std': 0.12282278388738632, 'epoch': 0.01}

  4%|▍         | 145/3300 [17:42<5:51:32,  6.69s/it]INFO 04-13 13:03:24 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 146/3300 [17:48<5:41:06,  6.49s/it]2026-04-13 13:03:30,757 | INFO | train_grpo_train | metrics_logged mode=train step=146

                                                    
{'loss': -0.1402, 'grad_norm': 15.96103286743164, 'learning_rate': 9.56060606060606e-06, 'num_tokens': 262641.0, 'completions/mean_length': 28.25, 'completions/min_length': 5.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.25, 'completions/min_terminated_length': 5.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.21407319605350494, 'rewards/meter/std': 0.17057731747627258, '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': 0.9917581081390381, 'rewards/repeat_soft/std': 0.01039243582636118, 'rewards/judge_quality/mean': 0.26124998927116394, 'rewards/judge_quality/std': 0.10091544687747955, 'rewards/total_composite/mean': 0.12573854625225067, 'rewards/total_composite/std': 0.07961715012788773, 'reward': 0.12573854625225067, 'reward_std': 0.07961715012788773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2117980718612671, 'sampling/sampling_logp_difference/max': 1.503438949584961, 'sampling/importance_sampling_ratio/min': 0.222364142537117, 'sampling/importance_sampling_ratio/mean': 1.0345395803451538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.863342374563217, 'clip_ratio/low_mean': 0.10940010473132133, 'clip_ratio/low_min': 0.10940010473132133, 'clip_ratio/high_mean': 0.06960083544254303, 'clip_ratio/high_max': 0.06960083544254303, 'clip_ratio/region_mean': 0.17900094017386436, 'reward_total_mean': 0.12573854625225067, 'reward_meter_mean': 0.21407319605350494, 'reward_meter_std': 0.17057731747627258, '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': 0.9917581081390381, 'reward_repeat_soft_std': 0.01039243582636118, 'reward_judge_quality_mean': 0.26124998927116394, 'reward_judge_quality_std': 0.10091544687747955, 'reward_total_composite_mean': 0.12573854625225067, 'reward_total_composite_std': 0.07961715012788773, 'epoch': 0.01}

  4%|▍         | 146/3300 [17:48<5:41:06,  6.49s/it]INFO 04-13 13:03:31 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 147/3300 [17:54<5:38:02,  6.43s/it]2026-04-13 13:03:37,069 | INFO | train_grpo_train | metrics_logged mode=train step=147

                                                    
{'loss': -0.0114, 'grad_norm': 23.897436141967773, 'learning_rate': 9.55757575757576e-06, 'num_tokens': 264117.0, 'completions/mean_length': 27.5, 'completions/min_length': 23.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6483590602874756, 'rewards/meter/std': 0.38227975368499756, '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': 0.9969677925109863, 'rewards/repeat_soft/std': 0.005060871597379446, 'rewards/judge_quality/mean': 0.2849999964237213, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.20882880687713623, 'rewards/total_composite/std': 0.11764669418334961, 'reward': 0.20882880687713623, 'reward_std': 0.11764669418334961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27704405784606934, 'sampling/sampling_logp_difference/max': 2.9856948852539062, 'sampling/importance_sampling_ratio/min': 0.050504397600889206, 'sampling/importance_sampling_ratio/mean': 1.015399694442749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8331507593393326, 'clip_ratio/low_mean': 0.04557291604578495, 'clip_ratio/low_min': 0.04557291604578495, 'clip_ratio/high_mean': 0.1767373187467456, 'clip_ratio/high_max': 0.1767373187467456, 'clip_ratio/region_mean': 0.22231023479253054, 'reward_total_mean': 0.20882880687713623, 'reward_meter_mean': 0.6483590602874756, 'reward_meter_std': 0.38227975368499756, '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': 0.9969677925109863, 'reward_repeat_soft_std': 0.005060871597379446, 'reward_judge_quality_mean': 0.2849999964237213, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.20882880687713623, 'reward_total_composite_std': 0.11764669418334961, 'epoch': 0.01}

  4%|▍         | 147/3300 [17:54<5:38:02,  6.43s/it]INFO 04-13 13:03:37 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 148/3300 [18:01<5:46:50,  6.60s/it]2026-04-13 13:03:44,057 | INFO | train_grpo_train | metrics_logged mode=train step=148

                                                    
{'loss': 0.033, 'grad_norm': 10.57125186920166, 'learning_rate': 9.554545454545455e-06, 'num_tokens': 266314.0, 'completions/mean_length': 85.625, 'completions/min_length': 70.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.6430109739303589, 'rewards/meter/std': 0.4566119611263275, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9956058263778687, 'rewards/repeat_soft/std': 0.003198496997356415, 'rewards/judge_quality/mean': 0.2449999898672104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.18349239230155945, 'rewards/total_composite/std': 0.1234927698969841, 'reward': 0.18349239230155945, 'reward_std': 0.1234927624464035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22205880284309387, 'sampling/sampling_logp_difference/max': 1.428999900817871, 'sampling/importance_sampling_ratio/min': 0.23954838514328003, 'sampling/importance_sampling_ratio/mean': 1.0468692779541016, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9796471297740936, 'clip_ratio/low_mean': 0.06434272974729538, 'clip_ratio/low_min': 0.06434272974729538, 'clip_ratio/high_mean': 0.1586184985935688, 'clip_ratio/high_max': 0.1586184985935688, 'clip_ratio/region_mean': 0.22296122834086418, 'reward_total_mean': 0.18349239230155945, 'reward_meter_mean': 0.6430109739303589, 'reward_meter_std': 0.4566119611263275, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9956058263778687, 'reward_repeat_soft_std': 0.003198496997356415, 'reward_judge_quality_mean': 0.2449999898672104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.18349239230155945, 'reward_total_composite_std': 0.1234927698969841, 'epoch': 0.01}

  4%|▍         | 148/3300 [18:01<5:46:50,  6.60s/it]INFO 04-13 13:03:44 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 149/3300 [18:09<5:59:19,  6.84s/it]2026-04-13 13:03:51,465 | INFO | train_grpo_train | metrics_logged mode=train step=149

                                                    
{'loss': -0.0165, 'grad_norm': 8.9459228515625, 'learning_rate': 9.551515151515152e-06, 'num_tokens': 268823.0, 'completions/mean_length': 122.625, 'completions/min_length': 107.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.625, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.8947402238845825, 'rewards/meter/std': 0.14273923635482788, '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': 0.995945930480957, 'rewards/repeat_soft/std': 0.004707190208137035, 'rewards/judge_quality/mean': 0.21000000834465027, 'rewards/judge_quality/std': 0.01851639896631241, 'rewards/total_composite/mean': 0.19634748995304108, 'rewards/total_composite/std': 0.03072374127805233, 'reward': 0.19634748995304108, 'reward_std': 0.03072374500334263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23941372334957123, 'sampling/sampling_logp_difference/max': 1.2456178665161133, 'sampling/importance_sampling_ratio/min': 0.2877630591392517, 'sampling/importance_sampling_ratio/mean': 1.0520843267440796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.761257380247116, 'clip_ratio/low_mean': 0.05636909045279026, 'clip_ratio/low_min': 0.05636909045279026, 'clip_ratio/high_mean': 0.16791298426687717, 'clip_ratio/high_max': 0.16791298426687717, 'clip_ratio/region_mean': 0.22428207471966743, 'reward_total_mean': 0.19634748995304108, 'reward_meter_mean': 0.8947402238845825, 'reward_meter_std': 0.14273923635482788, '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': 0.995945930480957, 'reward_repeat_soft_std': 0.004707190208137035, 'reward_judge_quality_mean': 0.21000000834465027, 'reward_judge_quality_std': 0.01851639896631241, 'reward_total_composite_mean': 0.19634748995304108, 'reward_total_composite_std': 0.03072374127805233, 'epoch': 0.01}

  5%|▍         | 149/3300 [18:09<5:59:19,  6.84s/it]INFO 04-13 13:03:51 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 150/3300 [18:16<6:00:58,  6.88s/it]2026-04-13 13:03:58,422 | INFO | train_grpo_train | metrics_logged mode=train step=150

                                                    
{'loss': 0.0381, 'grad_norm': 14.835107803344727, 'learning_rate': 9.54848484848485e-06, 'num_tokens': 270893.0, 'completions/mean_length': 88.75, 'completions/min_length': 69.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.33144593238830566, 'rewards/meter/std': 0.2663375735282898, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9982151389122009, 'rewards/repeat_soft/std': 0.001292722998186946, 'rewards/judge_quality/mean': 0.2799999713897705, 'rewards/judge_quality/std': 0.09258200228214264, 'rewards/total_composite/mean': 0.12507778406143188, 'rewards/total_composite/std': 0.06719503551721573, 'reward': 0.12507778406143188, 'reward_std': 0.06719504296779633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2821635901927948, 'sampling/sampling_logp_difference/max': 1.908768653869629, 'sampling/importance_sampling_ratio/min': 0.1482628434896469, 'sampling/importance_sampling_ratio/mean': 1.0197124481201172, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3200297951698303, 'clip_ratio/low_mean': 0.1122272964566946, 'clip_ratio/low_min': 0.1122272964566946, 'clip_ratio/high_mean': 0.132456686347723, 'clip_ratio/high_max': 0.132456686347723, 'clip_ratio/region_mean': 0.2446839828044176, 'reward_total_mean': 0.12507778406143188, 'reward_meter_mean': 0.33144593238830566, 'reward_meter_std': 0.2663375735282898, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9982151389122009, 'reward_repeat_soft_std': 0.001292722998186946, 'reward_judge_quality_mean': 0.2799999713897705, 'reward_judge_quality_std': 0.09258200228214264, 'reward_total_composite_mean': 0.12507778406143188, 'reward_total_composite_std': 0.06719503551721573, 'epoch': 0.02}

  5%|▍         | 150/3300 [18:16<6:00:58,  6.88s/it]INFO 04-13 13:03:58 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


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

 20%|██        | 2/10 [00:03<00:13,  1.70s/it]

 30%|███       | 3/10 [00:07<00:18,  2.58s/it]

 40%|████      | 4/10 [00:12<00:20,  3.42s/it]

 50%|█████     | 5/10 [00:16<00:18,  3.61s/it]

 60%|██████    | 6/10 [00:20<00:15,  3.80s/it]

 70%|███████   | 7/10 [00:24<00:11,  3.88s/it]

 80%|████████  | 8/10 [00:28<00:08,  4.13s/it]

 90%|█████████ | 9/10 [00:32<00:04,  4.02s/it]

100%|██████████| 10/10 [00:36<00:00,  3.84s/it]2026-04-13 13:04:39,260 | INFO | train_grpo_train | metrics_logged mode=eval step=150

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 40.827, 'eval_samples_per_second': 1.959, 'eval_steps_per_second': 0.245, 'eval_num_tokens': 270893.0, 'eval_completions/mean_length': 65.85, 'eval_completions/min_length': 25.5, 'eval_completions/max_length': 129.7, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 65.85, 'eval_completions/min_terminated_length': 25.5, 'eval_completions/max_terminated_length': 129.7, 'eval_rewards/meter/mean': 0.572745019197464, 'eval_rewards/meter/std': 0.3921780318021774, 'eval_rewards/count_adherence/mean': 0.9729166686534881, 'eval_rewards/count_adherence/std': 0.07660323455929756, 'eval_rewards/hard_gate/mean': 0.925, 'eval_rewards/hard_gate/std': 0.18771235942840575, 'eval_rewards/repeat_soft/mean': 0.9933066666126251, 'eval_rewards/repeat_soft/std': 0.011911871808115393, 'eval_rewards/judge_quality/mean': 0.26975000351667405, 'eval_rewards/judge_quality/std': 0.09802335640415549, 'eval_rewards/total_composite/mean': 0.1846210852265358, 'eval_rewards/total_composite/std': 0.123112241178751, 'eval_reward': 0.1846210852265358, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.17532754838466644, 'eval_sampling/sampling_logp_difference/max': 1.2397335052490235, 'eval_sampling/importance_sampling_ratio/min': 0.2943241521716118, 'eval_sampling/importance_sampling_ratio/mean': 1.0524425029754638, 'eval_sampling/importance_sampling_ratio/max': 1.6317439317703246, 'eval_entropy': 3.0035820484161375, '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': 0.1846210852265358, 'eval_reward_meter_mean': 0.572745019197464, 'eval_reward_meter_std': 0.3921780318021774, 'eval_reward_count_adherence_mean': 0.9729166686534881, 'eval_reward_count_adherence_std': 0.07660323455929756, 'eval_reward_hard_gate_mean': 0.925, 'eval_reward_hard_gate_std': 0.18771235942840575, 'eval_reward_repeat_soft_mean': 0.9933066666126251, 'eval_reward_repeat_soft_std': 0.011911871808115393, 'eval_reward_judge_quality_mean': 0.26975000351667405, 'eval_reward_judge_quality_std': 0.09802335640415549, 'eval_reward_total_composite_mean': 0.1846210852265358, 'eval_reward_total_composite_std': 0.123112241178751, 'epoch': 0.02}

  5%|▍         | 150/3300 [18:56<6:00:58,  6.88s/it]

100%|██████████| 10/10 [00:36<00:00,  3.84s/it]

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

  5%|▍         | 151/3300 [19:04<17:02:27, 19.48s/it]2026-04-13 13:04:47,310 | INFO | train_grpo_train | metrics_logged mode=train step=151

                                                     
{'loss': 0.0365, 'grad_norm': 25.638601303100586, 'learning_rate': 9.545454545454547e-06, 'num_tokens': 272257.0, 'completions/mean_length': 18.5, 'completions/min_length': 14.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.38172563910484314, 'rewards/meter/std': 0.47442716360092163, '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': 0.2849999666213989, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.17046958208084106, 'rewards/total_composite/std': 0.10838013887405396, 'reward': 0.17046958208084106, 'reward_std': 0.10838013887405396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23300570249557495, 'sampling/sampling_logp_difference/max': 1.059906005859375, 'sampling/importance_sampling_ratio/min': 0.3464883863925934, 'sampling/importance_sampling_ratio/mean': 1.0775567293167114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6863486915826797, 'clip_ratio/low_mean': 0.09908963739871979, 'clip_ratio/low_min': 0.09908963739871979, 'clip_ratio/high_mean': 0.10864662006497383, 'clip_ratio/high_max': 0.10864662006497383, 'clip_ratio/region_mean': 0.20773625746369362, 'reward_total_mean': 0.17046958208084106, 'reward_meter_mean': 0.38172563910484314, 'reward_meter_std': 0.47442716360092163, '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': 0.2849999666213989, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.17046958208084106, 'reward_total_composite_std': 0.10838013887405396, 'epoch': 0.02}

  5%|▍         | 151/3300 [19:04<17:02:27, 19.48s/it]INFO 04-13 13:04:47 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 152/3300 [19:12<13:48:06, 15.78s/it]2026-04-13 13:04:54,461 | INFO | train_grpo_train | metrics_logged mode=train step=152

                                                     
{'loss': 0.0344, 'grad_norm': 8.511669158935547, 'learning_rate': 9.542424242424242e-06, 'num_tokens': 274275.0, 'completions/mean_length': 89.25, 'completions/min_length': 70.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.25, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.8666473031044006, 'rewards/meter/std': 0.2796079218387604, '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': 0.9987285137176514, 'rewards/repeat_soft/std': 0.002673062961548567, 'rewards/judge_quality/mean': 0.2224999964237213, 'rewards/judge_quality/std': 0.027124052867293358, 'rewards/total_composite/mean': 0.17542394995689392, 'rewards/total_composite/std': 0.08328276872634888, 'reward': 0.17542394995689392, 'reward_std': 0.08328276872634888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22037450969219208, 'sampling/sampling_logp_difference/max': 1.5642757415771484, 'sampling/importance_sampling_ratio/min': 0.20923951268196106, 'sampling/importance_sampling_ratio/mean': 1.0545034408569336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6310852766036987, 'clip_ratio/low_mean': 0.04997062310576439, 'clip_ratio/low_min': 0.04997062310576439, 'clip_ratio/high_mean': 0.16189361177384853, 'clip_ratio/high_max': 0.16189361177384853, 'clip_ratio/region_mean': 0.21186423487961292, 'reward_total_mean': 0.17542394995689392, 'reward_meter_mean': 0.8666473031044006, 'reward_meter_std': 0.2796079218387604, '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': 0.9987285137176514, 'reward_repeat_soft_std': 0.002673062961548567, 'reward_judge_quality_mean': 0.2224999964237213, 'reward_judge_quality_std': 0.027124052867293358, 'reward_total_composite_mean': 0.17542394995689392, 'reward_total_composite_std': 0.08328276872634888, 'epoch': 0.02}

  5%|▍         | 152/3300 [19:12<13:48:06, 15.78s/it]INFO 04-13 13:04:54 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 153/3300 [19:17<11:10:38, 12.79s/it]2026-04-13 13:05:00,257 | INFO | train_grpo_train | metrics_logged mode=train step=153

                                                     
{'loss': 0.0429, 'grad_norm': 15.867789268493652, 'learning_rate': 9.539393939393941e-06, 'num_tokens': 275767.0, 'completions/mean_length': 40.5, 'completions/min_length': 39.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7182189226150513, 'rewards/meter/std': 0.36515012383461, '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': 0.9998095035552979, 'rewards/repeat_soft/std': 0.000538805325049907, 'rewards/judge_quality/mean': 0.2562499940395355, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.19799037277698517, 'rewards/total_composite/std': 0.040014926344156265, 'reward': 0.19799037277698517, 'reward_std': 0.040014926344156265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22870279848575592, 'sampling/sampling_logp_difference/max': 1.2410449981689453, 'sampling/importance_sampling_ratio/min': 0.28908199071884155, 'sampling/importance_sampling_ratio/mean': 1.0672976970672607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.826114594936371, 'clip_ratio/low_mean': 0.0574186984449625, 'clip_ratio/low_min': 0.0574186984449625, 'clip_ratio/high_mean': 0.1119742807932198, 'clip_ratio/high_max': 0.1119742807932198, 'clip_ratio/region_mean': 0.1693929792381823, 'reward_total_mean': 0.19799037277698517, 'reward_meter_mean': 0.7182189226150513, 'reward_meter_std': 0.36515012383461, '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': 0.9998095035552979, 'reward_repeat_soft_std': 0.000538805325049907, 'reward_judge_quality_mean': 0.2562499940395355, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.19799037277698517, 'reward_total_composite_std': 0.040014926344156265, 'epoch': 0.02}

  5%|▍         | 153/3300 [19:17<11:10:38, 12.79s/it]INFO 04-13 13:05:00 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 154/3300 [19:24<9:27:36, 10.83s/it] 2026-04-13 13:05:06,505 | INFO | train_grpo_train | metrics_logged mode=train step=154

                                                    
{'loss': 0.0105, 'grad_norm': 16.742393493652344, 'learning_rate': 9.536363636363637e-06, 'num_tokens': 277471.0, 'completions/mean_length': 41.0, 'completions/min_length': 36.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8258713483810425, 'rewards/meter/std': 0.3087295591831207, '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': 0.9977891445159912, 'rewards/repeat_soft/std': 0.00410388084128499, 'rewards/judge_quality/mean': 0.22374999523162842, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.19741278886795044, 'rewards/total_composite/std': 0.04111725091934204, 'reward': 0.19741278886795044, 'reward_std': 0.04111725091934204, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21883884072303772, 'sampling/sampling_logp_difference/max': 0.9221162796020508, 'sampling/importance_sampling_ratio/min': 0.397676557302475, 'sampling/importance_sampling_ratio/mean': 1.059522032737732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4892155528068542, 'clip_ratio/low_mean': 0.03926282096654177, 'clip_ratio/low_min': 0.03926282096654177, 'clip_ratio/high_mean': 0.14135960396379232, 'clip_ratio/high_max': 0.14135960396379232, 'clip_ratio/region_mean': 0.1806224249303341, 'reward_total_mean': 0.19741278886795044, 'reward_meter_mean': 0.8258713483810425, 'reward_meter_std': 0.3087295591831207, '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': 0.9977891445159912, 'reward_repeat_soft_std': 0.00410388084128499, 'reward_judge_quality_mean': 0.22374999523162842, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.19741278886795044, 'reward_total_composite_std': 0.04111725091934204, 'epoch': 0.02}

  5%|▍         | 154/3300 [19:24<9:27:36, 10.83s/it]INFO 04-13 13:05:06 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 155/3300 [19:30<8:11:21,  9.37s/it]2026-04-13 13:05:12,493 | INFO | train_grpo_train | metrics_logged mode=train step=155

                                                    
{'loss': -0.0251, 'grad_norm': 16.524307250976562, 'learning_rate': 9.533333333333334e-06, 'num_tokens': 279181.0, 'completions/mean_length': 37.75, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6799507141113281, 'rewards/meter/std': 0.3370908200740814, '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': 0.9923340082168579, 'rewards/repeat_soft/std': 0.007860827259719372, 'rewards/judge_quality/mean': 0.30250000953674316, 'rewards/judge_quality/std': 0.09808886796236038, 'rewards/total_composite/mean': 0.24224551022052765, 'rewards/total_composite/std': 0.11215207725763321, 'reward': 0.24224551022052765, 'reward_std': 0.1121520847082138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23383033275604248, 'sampling/sampling_logp_difference/max': 1.3716764450073242, 'sampling/importance_sampling_ratio/min': 0.25368133187294006, 'sampling/importance_sampling_ratio/mean': 1.077167272567749, 'sampling/importance_sampling_ratio/max': 1.9759652614593506, 'entropy': 3.1956107914447784, 'clip_ratio/low_mean': 0.14641496166586876, 'clip_ratio/low_min': 0.14641496166586876, 'clip_ratio/high_mean': 0.08069717511534691, 'clip_ratio/high_max': 0.08069717511534691, 'clip_ratio/region_mean': 0.22711213678121567, 'reward_total_mean': 0.24224551022052765, 'reward_meter_mean': 0.6799507141113281, 'reward_meter_std': 0.3370908200740814, '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': 0.9923340082168579, 'reward_repeat_soft_std': 0.007860827259719372, 'reward_judge_quality_mean': 0.30250000953674316, 'reward_judge_quality_std': 0.09808886796236038, 'reward_total_composite_mean': 0.24224551022052765, 'reward_total_composite_std': 0.11215207725763321, 'epoch': 0.02}

  5%|▍         | 155/3300 [19:30<8:11:21,  9.37s/it]INFO 04-13 13:05:12 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 156/3300 [19:36<7:17:43,  8.35s/it]2026-04-13 13:05:18,467 | INFO | train_grpo_train | metrics_logged mode=train step=156

                                                    
{'loss': 0.0109, 'grad_norm': 16.846054077148438, 'learning_rate': 9.530303030303031e-06, 'num_tokens': 280768.0, 'completions/mean_length': 35.375, 'completions/min_length': 33.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.47947975993156433, 'rewards/meter/std': 0.40722402930259705, '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': 0.9959161281585693, 'rewards/repeat_soft/std': 0.006067224312573671, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.19201096892356873, 'rewards/total_composite/std': 0.09286434203386307, 'reward': 0.19201096892356873, 'reward_std': 0.09286433458328247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2250913381576538, 'sampling/sampling_logp_difference/max': 2.343181848526001, 'sampling/importance_sampling_ratio/min': 0.0960216298699379, 'sampling/importance_sampling_ratio/mean': 1.0370150804519653, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5058412104845047, 'clip_ratio/low_mean': 0.1262683430686593, 'clip_ratio/low_min': 0.1262683430686593, 'clip_ratio/high_mean': 0.07929198257625103, 'clip_ratio/high_max': 0.07929198257625103, 'clip_ratio/region_mean': 0.20556032564491034, 'reward_total_mean': 0.19201096892356873, 'reward_meter_mean': 0.47947975993156433, 'reward_meter_std': 0.40722402930259705, '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': 0.9959161281585693, 'reward_repeat_soft_std': 0.006067224312573671, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.19201096892356873, 'reward_total_composite_std': 0.09286434203386307, 'epoch': 0.02}

  5%|▍         | 156/3300 [19:36<7:17:43,  8.35s/it]INFO 04-13 13:05:18 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 157/3300 [19:41<6:39:01,  7.62s/it]2026-04-13 13:05:24,366 | INFO | train_grpo_train | metrics_logged mode=train step=157

                                                    
{'loss': 0.1598, 'grad_norm': 27.04548454284668, 'learning_rate': 9.527272727272729e-06, 'num_tokens': 282162.0, 'completions/mean_length': 21.25, 'completions/min_length': 17.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6387215852737427, 'rewards/meter/std': 0.4759199023246765, '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': 0.9669870734214783, 'rewards/repeat_soft/std': 0.012691427022218704, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.08892211318016052, 'rewards/total_composite/mean': 0.2150147259235382, 'rewards/total_composite/std': 0.118369922041893, 'reward': 0.2150147259235382, 'reward_std': 0.11836991459131241, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2325924038887024, 'sampling/sampling_logp_difference/max': 2.731811761856079, 'sampling/importance_sampling_ratio/min': 0.06510123610496521, 'sampling/importance_sampling_ratio/mean': 1.0563908815383911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7705140113830566, 'clip_ratio/low_mean': 0.07694604247808456, 'clip_ratio/low_min': 0.07694604247808456, 'clip_ratio/high_mean': 0.14142475463449955, 'clip_ratio/high_max': 0.14142475463449955, 'clip_ratio/region_mean': 0.21837079711258411, 'reward_total_mean': 0.2150147259235382, 'reward_meter_mean': 0.6387215852737427, 'reward_meter_std': 0.4759199023246765, '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': 0.9669870734214783, 'reward_repeat_soft_std': 0.012691427022218704, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.08892211318016052, 'reward_total_composite_mean': 0.2150147259235382, 'reward_total_composite_std': 0.118369922041893, 'epoch': 0.02}

  5%|▍         | 157/3300 [19:41<6:39:01,  7.62s/it]INFO 04-13 13:05:24 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 158/3300 [19:47<6:13:47,  7.14s/it]2026-04-13 13:05:30,386 | INFO | train_grpo_train | metrics_logged mode=train step=158

                                                    
{'loss': 0.0504, 'grad_norm': 15.0586519241333, 'learning_rate': 9.524242424242424e-06, 'num_tokens': 283658.0, 'completions/mean_length': 39.0, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9479594230651855, 'rewards/meter/std': 0.08010739833116531, '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': 0.9995237588882446, 'rewards/repeat_soft/std': 0.001279377960599959, 'rewards/judge_quality/mean': 0.29499998688697815, 'rewards/judge_quality/std': 0.1035098284482956, 'rewards/total_composite/mean': 0.28706982731819153, 'rewards/total_composite/std': 0.10837505757808685, 'reward': 0.28706982731819153, 'reward_std': 0.10837505757808685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2167590707540512, 'sampling/sampling_logp_difference/max': 1.1569538116455078, 'sampling/importance_sampling_ratio/min': 0.31444257497787476, 'sampling/importance_sampling_ratio/mean': 1.0546858310699463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4026435911655426, 'clip_ratio/low_mean': 0.12356818560510874, 'clip_ratio/low_min': 0.12356818560510874, 'clip_ratio/high_mean': 0.07027027010917664, 'clip_ratio/high_max': 0.07027027010917664, 'clip_ratio/region_mean': 0.19383845571428537, 'reward_total_mean': 0.28706982731819153, 'reward_meter_mean': 0.9479594230651855, 'reward_meter_std': 0.08010739833116531, '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': 0.9995237588882446, 'reward_repeat_soft_std': 0.001279377960599959, 'reward_judge_quality_mean': 0.29499998688697815, 'reward_judge_quality_std': 0.1035098284482956, 'reward_total_composite_mean': 0.28706982731819153, 'reward_total_composite_std': 0.10837505757808685, 'epoch': 0.02}

  5%|▍         | 158/3300 [19:47<6:13:47,  7.14s/it]INFO 04-13 13:05:30 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 159/3300 [19:53<5:55:54,  6.80s/it]2026-04-13 13:05:36,391 | INFO | train_grpo_train | metrics_logged mode=train step=159

                                                    
{'loss': 0.0715, 'grad_norm': 17.11335563659668, 'learning_rate': 9.521212121212121e-06, 'num_tokens': 285188.0, 'completions/mean_length': 38.25, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9116860628128052, 'rewards/meter/std': 0.20351704955101013, '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': 0.9917818307876587, 'rewards/repeat_soft/std': 0.017315642908215523, 'rewards/judge_quality/mean': 0.30250000953674316, 'rewards/judge_quality/std': 0.14508618414402008, 'rewards/total_composite/mean': 0.27001631259918213, 'rewards/total_composite/std': 0.08081734925508499, 'reward': 0.27001631259918213, 'reward_std': 0.08081734925508499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21486471593379974, 'sampling/sampling_logp_difference/max': 1.2924151420593262, 'sampling/importance_sampling_ratio/min': 0.27460676431655884, 'sampling/importance_sampling_ratio/mean': 1.033375859260559, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1263866424560547, 'clip_ratio/low_mean': 0.14018620643764734, 'clip_ratio/low_min': 0.14018620643764734, 'clip_ratio/high_mean': 0.05878378264605999, 'clip_ratio/high_max': 0.05878378264605999, 'clip_ratio/region_mean': 0.19896998908370733, 'reward_total_mean': 0.27001631259918213, 'reward_meter_mean': 0.9116860628128052, 'reward_meter_std': 0.20351704955101013, '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': 0.9917818307876587, 'reward_repeat_soft_std': 0.017315642908215523, 'reward_judge_quality_mean': 0.30250000953674316, 'reward_judge_quality_std': 0.14508618414402008, 'reward_total_composite_mean': 0.27001631259918213, 'reward_total_composite_std': 0.08081734925508499, 'epoch': 0.02}

  5%|▍         | 159/3300 [19:53<5:55:54,  6.80s/it]INFO 04-13 13:05:36 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 160/3300 [19:59<5:40:56,  6.51s/it]2026-04-13 13:05:42,246 | INFO | train_grpo_train | metrics_logged mode=train step=160

                                                    
{'loss': 0.1024, 'grad_norm': 24.75937271118164, 'learning_rate': 9.518181818181819e-06, 'num_tokens': 286685.0, 'completions/mean_length': 30.125, 'completions/min_length': 26.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.125, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6666136384010315, 'rewards/meter/std': 0.42434483766555786, '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': 0.9896146059036255, 'rewards/repeat_soft/std': 0.011480763554573059, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.15554054081439972, 'rewards/total_composite/mean': 0.36371999979019165, 'rewards/total_composite/std': 0.18872661888599396, 'reward': 0.36371999979019165, 'reward_std': 0.18872661888599396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18248780071735382, 'sampling/sampling_logp_difference/max': 0.9513940811157227, 'sampling/importance_sampling_ratio/min': 0.38620224595069885, 'sampling/importance_sampling_ratio/mean': 1.059270977973938, 'sampling/importance_sampling_ratio/max': 1.7763392925262451, 'entropy': 2.211267575621605, 'clip_ratio/low_mean': 0.07675521913915873, 'clip_ratio/low_min': 0.07675521913915873, 'clip_ratio/high_mean': 0.10903514828532934, 'clip_ratio/high_max': 0.10903514828532934, 'clip_ratio/region_mean': 0.18579036742448807, 'reward_total_mean': 0.36371999979019165, 'reward_meter_mean': 0.6666136384010315, 'reward_meter_std': 0.42434483766555786, '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': 0.9896146059036255, 'reward_repeat_soft_std': 0.011480763554573059, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.15554054081439972, 'reward_total_composite_mean': 0.36371999979019165, 'reward_total_composite_std': 0.18872661888599396, 'epoch': 0.02}

  5%|▍         | 160/3300 [19:59<5:40:56,  6.51s/it]INFO 04-13 13:05:42 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 161/3300 [20:05<5:31:36,  6.34s/it]2026-04-13 13:05:48,171 | INFO | train_grpo_train | metrics_logged mode=train step=161

                                                    
{'loss': 0.2212, 'grad_norm': 24.690906524658203, 'learning_rate': 9.515151515151516e-06, 'num_tokens': 288121.0, 'completions/mean_length': 33.5, 'completions/min_length': 21.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.5782473087310791, 'rewards/meter/std': 0.40190985798835754, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9835919141769409, 'rewards/repeat_soft/std': 0.018582111224532127, 'rewards/judge_quality/mean': 0.25874999165534973, 'rewards/judge_quality/std': 0.06978282332420349, 'rewards/total_composite/mean': 0.15243111550807953, 'rewards/total_composite/std': 0.1106199324131012, 'reward': 0.15243111550807953, 'reward_std': 0.1106199324131012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25054123997688293, 'sampling/sampling_logp_difference/max': 1.4378957748413086, 'sampling/importance_sampling_ratio/min': 0.23742683231830597, 'sampling/importance_sampling_ratio/mean': 1.0478856563568115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.502296507358551, 'clip_ratio/low_mean': 0.09580637328326702, 'clip_ratio/low_min': 0.09580637328326702, 'clip_ratio/high_mean': 0.09887241385877132, 'clip_ratio/high_max': 0.09887241385877132, 'clip_ratio/region_mean': 0.19467878714203835, 'reward_total_mean': 0.15243111550807953, 'reward_meter_mean': 0.5782473087310791, 'reward_meter_std': 0.40190985798835754, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9835919141769409, 'reward_repeat_soft_std': 0.018582111224532127, 'reward_judge_quality_mean': 0.25874999165534973, 'reward_judge_quality_std': 0.06978282332420349, 'reward_total_composite_mean': 0.15243111550807953, 'reward_total_composite_std': 0.1106199324131012, 'epoch': 0.02}

  5%|▍         | 161/3300 [20:05<5:31:36,  6.34s/it]INFO 04-13 13:05:48 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 162/3300 [20:11<5:26:38,  6.25s/it]2026-04-13 13:05:54,201 | INFO | train_grpo_train | metrics_logged mode=train step=162

                                                    
{'loss': 0.0518, 'grad_norm': 22.90196418762207, 'learning_rate': 9.512121212121213e-06, 'num_tokens': 289618.0, 'completions/mean_length': 27.125, 'completions/min_length': 25.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.125, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6658871173858643, 'rewards/meter/std': 0.3142303228378296, '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': 0.9950172305107117, 'rewards/repeat_soft/std': 0.004112296737730503, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.2066052109003067, 'rewards/total_composite/mean': 0.23733368515968323, 'rewards/total_composite/std': 0.14951294660568237, 'reward': 0.23733368515968323, 'reward_std': 0.14951293170452118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22878214716911316, 'sampling/sampling_logp_difference/max': 1.463071346282959, 'sampling/importance_sampling_ratio/min': 0.23152409493923187, 'sampling/importance_sampling_ratio/mean': 1.0413239002227783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9636720418930054, 'clip_ratio/low_mean': 0.21846669167280197, 'clip_ratio/low_min': 0.21846669167280197, 'clip_ratio/high_mean': 0.03846153989434242, 'clip_ratio/high_max': 0.03846153989434242, 'clip_ratio/region_mean': 0.2569282315671444, 'reward_total_mean': 0.23733368515968323, 'reward_meter_mean': 0.6658871173858643, 'reward_meter_std': 0.3142303228378296, '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': 0.9950172305107117, 'reward_repeat_soft_std': 0.004112296737730503, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.2066052109003067, 'reward_total_composite_mean': 0.23733368515968323, 'reward_total_composite_std': 0.14951294660568237, 'epoch': 0.02}

  5%|▍         | 162/3300 [20:11<5:26:38,  6.25s/it]INFO 04-13 13:05:54 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 163/3300 [20:17<5:21:10,  6.14s/it]2026-04-13 13:06:00,105 | INFO | train_grpo_train | metrics_logged mode=train step=163

                                                    
{'loss': 0.0863, 'grad_norm': 17.661081314086914, 'learning_rate': 9.50909090909091e-06, 'num_tokens': 291212.0, 'completions/mean_length': 39.25, 'completions/min_length': 33.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.810309648513794, 'rewards/meter/std': 0.2990211844444275, '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': 0.9958804845809937, 'rewards/repeat_soft/std': 0.007363190874457359, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.17044061422348022, 'rewards/total_composite/mean': 0.3469280004501343, 'rewards/total_composite/std': 0.1944759637117386, 'reward': 0.3469280004501343, 'reward_std': 0.19447597861289978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22683171927928925, 'sampling/sampling_logp_difference/max': 1.625056505203247, 'sampling/importance_sampling_ratio/min': 0.19690056145191193, 'sampling/importance_sampling_ratio/mean': 1.0392217636108398, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.604361653327942, 'clip_ratio/low_mean': 0.11104768421500921, 'clip_ratio/low_min': 0.11104768421500921, 'clip_ratio/high_mean': 0.10663313511759043, 'clip_ratio/high_max': 0.10663313511759043, 'clip_ratio/region_mean': 0.21768081933259964, 'reward_total_mean': 0.3469280004501343, 'reward_meter_mean': 0.810309648513794, 'reward_meter_std': 0.2990211844444275, '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': 0.9958804845809937, 'reward_repeat_soft_std': 0.007363190874457359, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.17044061422348022, 'reward_total_composite_mean': 0.3469280004501343, 'reward_total_composite_std': 0.1944759637117386, 'epoch': 0.02}

  5%|▍         | 163/3300 [20:17<5:21:10,  6.14s/it]INFO 04-13 13:06:00 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 164/3300 [20:24<5:29:43,  6.31s/it]2026-04-13 13:06:06,799 | INFO | train_grpo_train | metrics_logged mode=train step=164

                                                    
{'loss': -0.0344, 'grad_norm': 10.669726371765137, 'learning_rate': 9.506060606060606e-06, 'num_tokens': 293394.0, 'completions/mean_length': 85.75, 'completions/min_length': 78.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.7769610285758972, 'rewards/meter/std': 0.32415905594825745, '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': 0.99647456407547, 'rewards/repeat_soft/std': 0.0025675075594335794, 'rewards/judge_quality/mean': 0.23499999940395355, 'rewards/judge_quality/std': 0.02777460403740406, 'rewards/total_composite/mean': 0.20259374380111694, 'rewards/total_composite/std': 0.06132829189300537, 'reward': 0.20259374380111694, 'reward_std': 0.06132828816771507, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23315180838108063, 'sampling/sampling_logp_difference/max': 1.7441082000732422, 'sampling/importance_sampling_ratio/min': 0.1748007982969284, 'sampling/importance_sampling_ratio/mean': 1.0587011575698853, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6615799367427826, 'clip_ratio/low_mean': 0.04825309570878744, 'clip_ratio/low_min': 0.04825309570878744, 'clip_ratio/high_mean': 0.12102040741592646, 'clip_ratio/high_max': 0.12102040741592646, 'clip_ratio/region_mean': 0.1692735031247139, 'reward_total_mean': 0.20259374380111694, 'reward_meter_mean': 0.7769610285758972, 'reward_meter_std': 0.32415905594825745, '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': 0.99647456407547, 'reward_repeat_soft_std': 0.0025675075594335794, 'reward_judge_quality_mean': 0.23499999940395355, 'reward_judge_quality_std': 0.02777460403740406, 'reward_total_composite_mean': 0.20259374380111694, 'reward_total_composite_std': 0.06132829189300537, 'epoch': 0.02}

  5%|▍         | 164/3300 [20:24<5:29:43,  6.31s/it]INFO 04-13 13:06:07 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 165/3300 [20:30<5:20:16,  6.13s/it]2026-04-13 13:06:12,510 | INFO | train_grpo_train | metrics_logged mode=train step=165

                                                    
{'loss': -0.0044, 'grad_norm': 23.314119338989258, 'learning_rate': 9.503030303030303e-06, 'num_tokens': 294906.0, 'completions/mean_length': 31.0, 'completions/min_length': 26.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6155316233634949, 'rewards/meter/std': 0.40069058537483215, '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': 0.9974690079689026, 'rewards/repeat_soft/std': 0.004686486907303333, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.28638482093811035, 'rewards/total_composite/std': 0.157676562666893, 'reward': 0.28638482093811035, 'reward_std': 0.157676562666893, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2523231506347656, 'sampling/sampling_logp_difference/max': 2.2936909198760986, 'sampling/importance_sampling_ratio/min': 0.10089338570833206, 'sampling/importance_sampling_ratio/mean': 0.9666851162910461, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5693599730730057, 'clip_ratio/low_mean': 0.11127274483442307, 'clip_ratio/low_min': 0.11127274483442307, 'clip_ratio/high_mean': 0.10227416455745697, 'clip_ratio/high_max': 0.10227416455745697, 'clip_ratio/region_mean': 0.21354690939188004, 'reward_total_mean': 0.28638482093811035, 'reward_meter_mean': 0.6155316233634949, 'reward_meter_std': 0.40069058537483215, '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': 0.9974690079689026, 'reward_repeat_soft_std': 0.004686486907303333, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.28638482093811035, 'reward_total_composite_std': 0.157676562666893, 'epoch': 0.02}

  5%|▌         | 165/3300 [20:30<5:20:16,  6.13s/it]INFO 04-13 13:06:12 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 166/3300 [20:36<5:23:19,  6.19s/it]2026-04-13 13:06:18,843 | INFO | train_grpo_train | metrics_logged mode=train step=166

                                                    
{'loss': 0.1268, 'grad_norm': 17.905162811279297, 'learning_rate': 9.5e-06, 'num_tokens': 296577.0, 'completions/mean_length': 44.875, 'completions/min_length': 34.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.33484289050102234, 'rewards/meter/std': 0.37169912457466125, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9986193180084229, 'rewards/repeat_soft/std': 0.002488938393071294, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.2686706781387329, 'rewards/total_composite/mean': 0.22475406527519226, 'rewards/total_composite/std': 0.22406691312789917, 'reward': 0.22475406527519226, 'reward_std': 0.22406691312789917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23967714607715607, 'sampling/sampling_logp_difference/max': 1.3330392837524414, 'sampling/importance_sampling_ratio/min': 0.26367464661598206, 'sampling/importance_sampling_ratio/mean': 1.035905122756958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.881513684988022, 'clip_ratio/low_mean': 0.13186770491302013, 'clip_ratio/low_min': 0.13186770491302013, 'clip_ratio/high_mean': 0.07607466168701649, 'clip_ratio/high_max': 0.07607466168701649, 'clip_ratio/region_mean': 0.20794236660003662, 'reward_total_mean': 0.22475406527519226, 'reward_meter_mean': 0.33484289050102234, 'reward_meter_std': 0.37169912457466125, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9986193180084229, 'reward_repeat_soft_std': 0.002488938393071294, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.2686706781387329, 'reward_total_composite_mean': 0.22475406527519226, 'reward_total_composite_std': 0.22406691312789917, 'epoch': 0.02}

  5%|▌         | 166/3300 [20:36<5:23:19,  6.19s/it]INFO 04-13 13:06:19 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 167/3300 [20:42<5:25:40,  6.24s/it]2026-04-13 13:06:25,190 | INFO | train_grpo_train | metrics_logged mode=train step=167

                                                    
{'loss': 0.0488, 'grad_norm': 13.672354698181152, 'learning_rate': 9.496969696969698e-06, 'num_tokens': 298093.0, 'completions/mean_length': 42.5, 'completions/min_length': 39.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9360833168029785, 'rewards/meter/std': 0.06844120472669601, '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': 0.9947550296783447, 'rewards/repeat_soft/std': 0.010183374397456646, 'rewards/judge_quality/mean': 0.28125, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.26979902386665344, 'rewards/total_composite/std': 0.08546984195709229, 'reward': 0.26979902386665344, 'reward_std': 0.08546984195709229, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23771153390407562, 'sampling/sampling_logp_difference/max': 1.4594011306762695, 'sampling/importance_sampling_ratio/min': 0.2323753982782364, 'sampling/importance_sampling_ratio/mean': 1.0684665441513062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6879286766052246, 'clip_ratio/low_mean': 0.1279240120202303, 'clip_ratio/low_min': 0.1279240120202303, 'clip_ratio/high_mean': 0.0570512842386961, 'clip_ratio/high_max': 0.0570512842386961, 'clip_ratio/region_mean': 0.1849752962589264, 'reward_total_mean': 0.26979902386665344, 'reward_meter_mean': 0.9360833168029785, 'reward_meter_std': 0.06844120472669601, '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': 0.9947550296783447, 'reward_repeat_soft_std': 0.010183374397456646, 'reward_judge_quality_mean': 0.28125, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.26979902386665344, 'reward_total_composite_std': 0.08546984195709229, 'epoch': 0.02}

  5%|▌         | 167/3300 [20:42<5:25:40,  6.24s/it]INFO 04-13 13:06:25 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 168/3300 [20:48<5:21:27,  6.16s/it]2026-04-13 13:06:31,164 | INFO | train_grpo_train | metrics_logged mode=train step=168

                                                    
{'loss': -0.106, 'grad_norm': 21.89627456665039, 'learning_rate': 9.493939393939395e-06, 'num_tokens': 299624.0, 'completions/mean_length': 26.375, 'completions/min_length': 19.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.647320032119751, 'rewards/meter/std': 0.33076751232147217, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9618749022483826, 'rewards/repeat_soft/std': 0.024751795455813408, 'rewards/judge_quality/mean': 0.2524999976158142, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.17939558625221252, 'rewards/total_composite/std': 0.06344479322433472, 'reward': 0.17939558625221252, 'reward_std': 0.06344479322433472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2663666903972626, 'sampling/sampling_logp_difference/max': 1.9853134155273438, 'sampling/importance_sampling_ratio/min': 0.1373375654220581, 'sampling/importance_sampling_ratio/mean': 1.0470249652862549, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.762563407421112, 'clip_ratio/low_mean': 0.15201344713568687, 'clip_ratio/low_min': 0.15201344713568687, 'clip_ratio/high_mean': 0.0846162810921669, 'clip_ratio/high_max': 0.0846162810921669, 'clip_ratio/region_mean': 0.23662972822785378, 'reward_total_mean': 0.17939558625221252, 'reward_meter_mean': 0.647320032119751, 'reward_meter_std': 0.33076751232147217, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9618749022483826, 'reward_repeat_soft_std': 0.024751795455813408, 'reward_judge_quality_mean': 0.2524999976158142, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.17939558625221252, 'reward_total_composite_std': 0.06344479322433472, 'epoch': 0.02}

  5%|▌         | 168/3300 [20:48<5:21:27,  6.16s/it]INFO 04-13 13:06:31 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 169/3300 [20:55<5:28:49,  6.30s/it]2026-04-13 13:06:37,799 | INFO | train_grpo_train | metrics_logged mode=train step=169

                                                    
{'loss': 0.2582, 'grad_norm': 16.440635681152344, 'learning_rate': 9.490909090909092e-06, 'num_tokens': 301292.0, 'completions/mean_length': 47.5, 'completions/min_length': 34.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.5857816934585571, 'rewards/meter/std': 0.43221327662467957, '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': 0.9991960525512695, 'rewards/repeat_soft/std': 0.002273822668939829, 'rewards/judge_quality/mean': 0.29874998331069946, 'rewards/judge_quality/std': 0.10091544687747955, 'rewards/total_composite/mean': 0.213999941945076, 'rewards/total_composite/std': 0.09905601292848587, 'reward': 0.213999941945076, 'reward_std': 0.09905601292848587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23288343846797943, 'sampling/sampling_logp_difference/max': 0.9852371215820312, 'sampling/importance_sampling_ratio/min': 0.37335067987442017, 'sampling/importance_sampling_ratio/mean': 1.0595744848251343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.978996843099594, 'clip_ratio/low_mean': 0.04683104157447815, 'clip_ratio/low_min': 0.04683104157447815, 'clip_ratio/high_mean': 0.12344735581427813, 'clip_ratio/high_max': 0.12344735581427813, 'clip_ratio/region_mean': 0.17027839738875628, 'reward_total_mean': 0.213999941945076, 'reward_meter_mean': 0.5857816934585571, 'reward_meter_std': 0.43221327662467957, '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': 0.9991960525512695, 'reward_repeat_soft_std': 0.002273822668939829, 'reward_judge_quality_mean': 0.29874998331069946, 'reward_judge_quality_std': 0.10091544687747955, 'reward_total_composite_mean': 0.213999941945076, 'reward_total_composite_std': 0.09905601292848587, 'epoch': 0.02}

  5%|▌         | 169/3300 [20:55<5:28:49,  6.30s/it]INFO 04-13 13:06:38 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 170/3300 [21:01<5:23:19,  6.20s/it]2026-04-13 13:06:43,756 | INFO | train_grpo_train | metrics_logged mode=train step=170

                                                    
{'loss': 0.1449, 'grad_norm': 14.68420124053955, 'learning_rate': 9.487878787878788e-06, 'num_tokens': 302793.0, 'completions/mean_length': 40.625, 'completions/min_length': 32.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8819254636764526, 'rewards/meter/std': 0.28798651695251465, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9958971738815308, 'rewards/repeat_soft/std': 0.006846282631158829, 'rewards/judge_quality/mean': 0.30250000953674316, 'rewards/judge_quality/std': 0.14508618414402008, 'rewards/total_composite/mean': 0.25590571761131287, 'rewards/total_composite/std': 0.18720243871212006, 'reward': 0.25590571761131287, 'reward_std': 0.18720243871212006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2317146360874176, 'sampling/sampling_logp_difference/max': 1.2958250045776367, 'sampling/importance_sampling_ratio/min': 0.2736719846725464, 'sampling/importance_sampling_ratio/mean': 1.0399237871170044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.074298083782196, 'clip_ratio/low_mean': 0.15723508596420288, 'clip_ratio/low_min': 0.15723508596420288, 'clip_ratio/high_mean': 0.06250000186264515, 'clip_ratio/high_max': 0.06250000186264515, 'clip_ratio/region_mean': 0.21973508782684803, 'reward_total_mean': 0.25590571761131287, 'reward_meter_mean': 0.8819254636764526, 'reward_meter_std': 0.28798651695251465, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9958971738815308, 'reward_repeat_soft_std': 0.006846282631158829, 'reward_judge_quality_mean': 0.30250000953674316, 'reward_judge_quality_std': 0.14508618414402008, 'reward_total_composite_mean': 0.25590571761131287, 'reward_total_composite_std': 0.18720243871212006, 'epoch': 0.02}

  5%|▌         | 170/3300 [21:01<5:23:19,  6.20s/it]INFO 04-13 13:06:43 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 171/3300 [21:07<5:23:26,  6.20s/it]2026-04-13 13:06:49,968 | INFO | train_grpo_train | metrics_logged mode=train step=171

                                                    
{'loss': 0.0779, 'grad_norm': 16.1843318939209, 'learning_rate': 9.484848484848485e-06, 'num_tokens': 304440.0, 'completions/mean_length': 39.875, 'completions/min_length': 34.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6370018124580383, 'rewards/meter/std': 0.44355839490890503, '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': 0.9982327222824097, 'rewards/repeat_soft/std': 0.004276552703231573, 'rewards/judge_quality/mean': 0.3062500059604645, 'rewards/judge_quality/std': 0.2081165611743927, 'rewards/total_composite/mean': 0.24946370720863342, 'rewards/total_composite/std': 0.23714184761047363, 'reward': 0.24946370720863342, 'reward_std': 0.23714183270931244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22012512385845184, 'sampling/sampling_logp_difference/max': 1.327301025390625, 'sampling/importance_sampling_ratio/min': 0.26519206166267395, 'sampling/importance_sampling_ratio/mean': 1.0278286933898926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7738112658262253, 'clip_ratio/low_mean': 0.15882117487490177, 'clip_ratio/low_min': 0.15882117487490177, 'clip_ratio/high_mean': 0.010135134682059288, 'clip_ratio/high_max': 0.010135134682059288, 'clip_ratio/region_mean': 0.16895630955696106, 'reward_total_mean': 0.24946370720863342, 'reward_meter_mean': 0.6370018124580383, 'reward_meter_std': 0.44355839490890503, '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': 0.9982327222824097, 'reward_repeat_soft_std': 0.004276552703231573, 'reward_judge_quality_mean': 0.3062500059604645, 'reward_judge_quality_std': 0.2081165611743927, 'reward_total_composite_mean': 0.24946370720863342, 'reward_total_composite_std': 0.23714184761047363, 'epoch': 0.02}

  5%|▌         | 171/3300 [21:07<5:23:26,  6.20s/it]INFO 04-13 13:06:50 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 172/3300 [21:13<5:19:24,  6.13s/it]2026-04-13 13:06:55,918 | INFO | train_grpo_train | metrics_logged mode=train step=172

                                                    
{'loss': 0.0646, 'grad_norm': 16.315248489379883, 'learning_rate': 9.481818181818182e-06, 'num_tokens': 305980.0, 'completions/mean_length': 29.5, 'completions/min_length': 26.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.5681661367416382, 'rewards/meter/std': 0.42752423882484436, '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': 0.964336097240448, 'rewards/repeat_soft/std': 0.09348888695240021, 'rewards/judge_quality/mean': 0.32374998927116394, 'rewards/judge_quality/std': 0.1487027257680893, 'rewards/total_composite/mean': 0.25733670592308044, 'rewards/total_composite/std': 0.1935597062110901, 'reward': 0.25733670592308044, 'reward_std': 0.19355972111225128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26032719016075134, 'sampling/sampling_logp_difference/max': 1.1938762664794922, 'sampling/importance_sampling_ratio/min': 0.30304428935050964, 'sampling/importance_sampling_ratio/mean': 1.0414456129074097, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.785316988825798, 'clip_ratio/low_mean': 0.1700533702969551, 'clip_ratio/low_min': 0.1700533702969551, 'clip_ratio/high_mean': 0.09917582757771015, 'clip_ratio/high_max': 0.09917582757771015, 'clip_ratio/region_mean': 0.26922919787466526, 'reward_total_mean': 0.25733670592308044, 'reward_meter_mean': 0.5681661367416382, 'reward_meter_std': 0.42752423882484436, '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': 0.964336097240448, 'reward_repeat_soft_std': 0.09348888695240021, 'reward_judge_quality_mean': 0.32374998927116394, 'reward_judge_quality_std': 0.1487027257680893, 'reward_total_composite_mean': 0.25733670592308044, 'reward_total_composite_std': 0.1935597062110901, 'epoch': 0.02}

  5%|▌         | 172/3300 [21:13<5:19:24,  6.13s/it]INFO 04-13 13:06:56 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 173/3300 [21:19<5:14:07,  6.03s/it]2026-04-13 13:07:01,714 | INFO | train_grpo_train | metrics_logged mode=train step=173

                                                    
{'loss': 0.0064, 'grad_norm': 31.017641067504883, 'learning_rate': 9.47878787878788e-06, 'num_tokens': 307483.0, 'completions/mean_length': 23.875, 'completions/min_length': 19.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.712361216545105, 'rewards/meter/std': 0.3987361788749695, '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': 0.95976322889328, 'rewards/repeat_soft/std': 0.023058203980326653, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.13786433637142181, 'rewards/total_composite/std': 0.04557495191693306, 'reward': 0.13786433637142181, 'reward_std': 0.04557494819164276, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19148781895637512, 'sampling/sampling_logp_difference/max': 2.123509407043457, 'sampling/importance_sampling_ratio/min': 0.11961112171411514, 'sampling/importance_sampling_ratio/mean': 1.0386604070663452, 'sampling/importance_sampling_ratio/max': 1.9924687147140503, 'entropy': 1.7448338121175766, 'clip_ratio/low_mean': 0.061666667461395264, 'clip_ratio/low_min': 0.061666667461395264, 'clip_ratio/high_mean': 0.12374117225408554, 'clip_ratio/high_max': 0.12374117225408554, 'clip_ratio/region_mean': 0.1854078397154808, 'reward_total_mean': 0.13786433637142181, 'reward_meter_mean': 0.712361216545105, 'reward_meter_std': 0.3987361788749695, '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': 0.95976322889328, 'reward_repeat_soft_std': 0.023058203980326653, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.13786433637142181, 'reward_total_composite_std': 0.04557495191693306, 'epoch': 0.02}

  5%|▌         | 173/3300 [21:19<5:14:07,  6.03s/it]INFO 04-13 13:07:01 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 174/3300 [21:25<5:13:26,  6.02s/it]2026-04-13 13:07:07,706 | INFO | train_grpo_train | metrics_logged mode=train step=174

                                                    
{'loss': 0.0264, 'grad_norm': 14.48010540008545, 'learning_rate': 9.475757575757577e-06, 'num_tokens': 309070.0, 'completions/mean_length': 41.375, 'completions/min_length': 35.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5942786931991577, 'rewards/meter/std': 0.27671706676483154, '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': 0.9999185800552368, 'rewards/repeat_soft/std': 0.0002302482316736132, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.018850916996598244, 'rewards/total_composite/mean': 0.1614990532398224, 'rewards/total_composite/std': 0.04365858435630798, 'reward': 0.1614990532398224, 'reward_std': 0.043658580631017685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23007173836231232, 'sampling/sampling_logp_difference/max': 1.1083860397338867, 'sampling/importance_sampling_ratio/min': 0.33009129762649536, 'sampling/importance_sampling_ratio/mean': 1.0569177865982056, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.840572714805603, 'clip_ratio/low_mean': 0.07121522165834904, 'clip_ratio/low_min': 0.07121522165834904, 'clip_ratio/high_mean': 0.10277472622692585, 'clip_ratio/high_max': 0.10277472622692585, 'clip_ratio/region_mean': 0.1739899478852749, 'reward_total_mean': 0.1614990532398224, 'reward_meter_mean': 0.5942786931991577, 'reward_meter_std': 0.27671706676483154, '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': 0.9999185800552368, 'reward_repeat_soft_std': 0.0002302482316736132, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.018850916996598244, 'reward_total_composite_mean': 0.1614990532398224, 'reward_total_composite_std': 0.04365858435630798, 'epoch': 0.02}

  5%|▌         | 174/3300 [21:25<5:13:26,  6.02s/it]INFO 04-13 13:07:07 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 175/3300 [21:31<5:17:34,  6.10s/it]2026-04-13 13:07:13,991 | INFO | train_grpo_train | metrics_logged mode=train step=175

                                                    
{'loss': 0.0116, 'grad_norm': 15.40805721282959, 'learning_rate': 9.472727272727274e-06, 'num_tokens': 310638.0, 'completions/mean_length': 41.0, 'completions/min_length': 30.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.7776926159858704, 'rewards/meter/std': 0.36967718601226807, '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': 0.9996999502182007, 'rewards/repeat_soft/std': 0.0006752270855940878, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23754699528217316, 'rewards/total_composite/mean': 0.4998127818107605, 'rewards/total_composite/std': 0.2754271328449249, 'reward': 0.4998127818107605, 'reward_std': 0.2754271328449249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1785004884004593, 'sampling/sampling_logp_difference/max': 1.1226024627685547, 'sampling/importance_sampling_ratio/min': 0.325431764125824, 'sampling/importance_sampling_ratio/mean': 1.0309224128723145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8760915398597717, 'clip_ratio/low_mean': 0.0728583112359047, 'clip_ratio/low_min': 0.0728583112359047, 'clip_ratio/high_mean': 0.09241350181400776, 'clip_ratio/high_max': 0.09241350181400776, 'clip_ratio/region_mean': 0.16527181304991245, 'reward_total_mean': 0.4998127818107605, 'reward_meter_mean': 0.7776926159858704, 'reward_meter_std': 0.36967718601226807, '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': 0.9996999502182007, 'reward_repeat_soft_std': 0.0006752270855940878, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23754699528217316, 'reward_total_composite_mean': 0.4998127818107605, 'reward_total_composite_std': 0.2754271328449249, 'epoch': 0.02}

  5%|▌         | 175/3300 [21:31<5:17:34,  6.10s/it]INFO 04-13 13:07:14 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 176/3300 [21:37<5:12:38,  6.00s/it]2026-04-13 13:07:19,779 | INFO | train_grpo_train | metrics_logged mode=train step=176

                                                    
{'loss': 0.1406, 'grad_norm': 25.277259826660156, 'learning_rate': 9.469696969696971e-06, 'num_tokens': 312172.0, 'completions/mean_length': 33.75, 'completions/min_length': 27.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7242546081542969, 'rewards/meter/std': 0.250786691904068, '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': 0.9931771755218506, 'rewards/repeat_soft/std': 0.013044352643191814, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.12799972295761108, 'rewards/total_composite/mean': 0.30834031105041504, 'rewards/total_composite/std': 0.11047318577766418, 'reward': 0.30834031105041504, 'reward_std': 0.11047318577766418, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24297761917114258, 'sampling/sampling_logp_difference/max': 2.2212677001953125, 'sampling/importance_sampling_ratio/min': 0.10847151279449463, 'sampling/importance_sampling_ratio/mean': 1.0259636640548706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2932978942990303, 'clip_ratio/low_mean': 0.1337986309081316, 'clip_ratio/low_min': 0.1337986309081316, 'clip_ratio/high_mean': 0.06702178157866001, 'clip_ratio/high_max': 0.06702178157866001, 'clip_ratio/region_mean': 0.2008204124867916, 'reward_total_mean': 0.30834031105041504, 'reward_meter_mean': 0.7242546081542969, 'reward_meter_std': 0.250786691904068, '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': 0.9931771755218506, 'reward_repeat_soft_std': 0.013044352643191814, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.12799972295761108, 'reward_total_composite_mean': 0.30834031105041504, 'reward_total_composite_std': 0.11047318577766418, 'epoch': 0.02}

  5%|▌         | 176/3300 [21:37<5:12:38,  6.00s/it]INFO 04-13 13:07:20 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 177/3300 [21:43<5:13:08,  6.02s/it]2026-04-13 13:07:25,822 | INFO | train_grpo_train | metrics_logged mode=train step=177

                                                    
{'loss': 0.1314, 'grad_norm': 14.810586929321289, 'learning_rate': 9.466666666666667e-06, 'num_tokens': 313753.0, 'completions/mean_length': 46.625, 'completions/min_length': 39.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.45506995916366577, 'rewards/meter/std': 0.43716365098953247, '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': 0.9980074167251587, 'rewards/repeat_soft/std': 0.005443721543997526, 'rewards/judge_quality/mean': 0.2449999898672104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.16408401727676392, 'rewards/total_composite/std': 0.11723174899816513, 'reward': 0.16408401727676392, 'reward_std': 0.11723174154758453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24596381187438965, 'sampling/sampling_logp_difference/max': 1.5142285823822021, 'sampling/importance_sampling_ratio/min': 0.3284386396408081, 'sampling/importance_sampling_ratio/mean': 1.0515755414962769, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1561138927936554, 'clip_ratio/low_mean': 0.11140831606462598, 'clip_ratio/low_min': 0.11140831606462598, 'clip_ratio/high_mean': 0.09245426766574383, 'clip_ratio/high_max': 0.09245426766574383, 'clip_ratio/region_mean': 0.2038625837303698, 'reward_total_mean': 0.16408401727676392, 'reward_meter_mean': 0.45506995916366577, 'reward_meter_std': 0.43716365098953247, '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': 0.9980074167251587, 'reward_repeat_soft_std': 0.005443721543997526, 'reward_judge_quality_mean': 0.2449999898672104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.16408401727676392, 'reward_total_composite_std': 0.11723174899816513, 'epoch': 0.02}

  5%|▌         | 177/3300 [21:43<5:13:08,  6.02s/it]INFO 04-13 13:07:26 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 178/3300 [21:50<5:35:55,  6.46s/it]2026-04-13 13:07:33,305 | INFO | train_grpo_train | metrics_logged mode=train step=178

                                                    
{'loss': 0.0457, 'grad_norm': 8.597575187683105, 'learning_rate': 9.463636363636364e-06, 'num_tokens': 316168.0, 'completions/mean_length': 103.875, 'completions/min_length': 94.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.875, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.8420461416244507, 'rewards/meter/std': 0.28812357783317566, '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': 0.995198130607605, 'rewards/repeat_soft/std': 0.0031533981673419476, 'rewards/judge_quality/mean': 0.25999999046325684, 'rewards/judge_quality/std': 0.07010196894407272, 'rewards/total_composite/mean': 0.23466189205646515, 'rewards/total_composite/std': 0.0839463621377945, 'reward': 0.23466189205646515, 'reward_std': 0.0839463621377945, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24356208741664886, 'sampling/sampling_logp_difference/max': 1.4635021686553955, 'sampling/importance_sampling_ratio/min': 0.23142436146736145, 'sampling/importance_sampling_ratio/mean': 1.0635943412780762, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.91947278380394, 'clip_ratio/low_mean': 0.12700920179486275, 'clip_ratio/low_min': 0.12700920179486275, 'clip_ratio/high_mean': 0.05360471457242966, 'clip_ratio/high_max': 0.05360471457242966, 'clip_ratio/region_mean': 0.1806139163672924, 'reward_total_mean': 0.23466189205646515, 'reward_meter_mean': 0.8420461416244507, 'reward_meter_std': 0.28812357783317566, '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': 0.995198130607605, 'reward_repeat_soft_std': 0.0031533981673419476, 'reward_judge_quality_mean': 0.25999999046325684, 'reward_judge_quality_std': 0.07010196894407272, 'reward_total_composite_mean': 0.23466189205646515, 'reward_total_composite_std': 0.0839463621377945, 'epoch': 0.02}

  5%|▌         | 178/3300 [21:50<5:35:55,  6.46s/it]INFO 04-13 13:07:33 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 179/3300 [21:57<5:34:14,  6.43s/it]2026-04-13 13:07:39,659 | INFO | train_grpo_train | metrics_logged mode=train step=179

                                                    
{'loss': 0.0504, 'grad_norm': 12.027117729187012, 'learning_rate': 9.460606060606061e-06, 'num_tokens': 318206.0, 'completions/mean_length': 78.75, 'completions/min_length': 73.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.75, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5722213983535767, 'rewards/meter/std': 0.2216475009918213, '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': 0.9970940351486206, 'rewards/repeat_soft/std': 0.00415117247030139, 'rewards/judge_quality/mean': 0.2150000035762787, 'rewards/judge_quality/std': 0.014142133295536041, 'rewards/total_composite/mean': 0.12795740365982056, 'rewards/total_composite/std': 0.05583568662405014, 'reward': 0.12795740365982056, 'reward_std': 0.05583568289875984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22791191935539246, 'sampling/sampling_logp_difference/max': 1.2603139877319336, 'sampling/importance_sampling_ratio/min': 0.2835649847984314, 'sampling/importance_sampling_ratio/mean': 1.0541539192199707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9370436668395996, 'clip_ratio/low_mean': 0.044417669996619225, 'clip_ratio/low_min': 0.044417669996619225, 'clip_ratio/high_mean': 0.13506676629185677, 'clip_ratio/high_max': 0.13506676629185677, 'clip_ratio/region_mean': 0.179484436288476, 'reward_total_mean': 0.12795740365982056, 'reward_meter_mean': 0.5722213983535767, 'reward_meter_std': 0.2216475009918213, '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': 0.9970940351486206, 'reward_repeat_soft_std': 0.00415117247030139, 'reward_judge_quality_mean': 0.2150000035762787, 'reward_judge_quality_std': 0.014142133295536041, 'reward_total_composite_mean': 0.12795740365982056, 'reward_total_composite_std': 0.05583568662405014, 'epoch': 0.02}

  5%|▌         | 179/3300 [21:57<5:34:14,  6.43s/it]INFO 04-13 13:07:39 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 180/3300 [22:03<5:34:52,  6.44s/it]2026-04-13 13:07:46,133 | INFO | train_grpo_train | metrics_logged mode=train step=180

                                                    
{'loss': 0.0455, 'grad_norm': 10.89418888092041, 'learning_rate': 9.457575757575759e-06, 'num_tokens': 320448.0, 'completions/mean_length': 82.25, 'completions/min_length': 73.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.25, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.21472598612308502, 'rewards/meter/std': 0.22956499457359314, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996832549571991, 'rewards/repeat_soft/std': 0.001779663492925465, 'rewards/judge_quality/mean': 0.2199999988079071, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.10347628593444824, 'rewards/total_composite/std': 0.03403228893876076, 'reward': 0.10347628593444824, 'reward_std': 0.03403228521347046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25185060501098633, 'sampling/sampling_logp_difference/max': 1.3912534713745117, 'sampling/importance_sampling_ratio/min': 0.2487632930278778, 'sampling/importance_sampling_ratio/mean': 1.0535268783569336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4406895339488983, 'clip_ratio/low_mean': 0.14273057878017426, 'clip_ratio/low_min': 0.14273057878017426, 'clip_ratio/high_mean': 0.07947134785354137, 'clip_ratio/high_max': 0.07947134785354137, 'clip_ratio/region_mean': 0.22220192663371563, 'reward_total_mean': 0.10347628593444824, 'reward_meter_mean': 0.21472598612308502, 'reward_meter_std': 0.22956499457359314, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996832549571991, 'reward_repeat_soft_std': 0.001779663492925465, 'reward_judge_quality_mean': 0.2199999988079071, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.10347628593444824, 'reward_total_composite_std': 0.03403228893876076, 'epoch': 0.02}

  5%|▌         | 180/3300 [22:03<5:34:52,  6.44s/it]INFO 04-13 13:07:46 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 181/3300 [22:09<5:29:31,  6.34s/it]2026-04-13 13:07:52,238 | INFO | train_grpo_train | metrics_logged mode=train step=181

                                                    
{'loss': 0.1036, 'grad_norm': 16.209243774414062, 'learning_rate': 9.454545454545456e-06, 'num_tokens': 322063.0, 'completions/mean_length': 41.875, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6623384952545166, 'rewards/meter/std': 0.3999437093734741, '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': 0.9959326982498169, 'rewards/repeat_soft/std': 0.0059642670676112175, 'rewards/judge_quality/mean': 0.3075000047683716, 'rewards/judge_quality/std': 0.12464234232902527, 'rewards/total_composite/mean': 0.25208374857902527, 'rewards/total_composite/std': 0.14597024023532867, 'reward': 0.25208374857902527, 'reward_std': 0.14597022533416748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2273944914340973, 'sampling/sampling_logp_difference/max': 1.3746209144592285, 'sampling/importance_sampling_ratio/min': 0.2529354691505432, 'sampling/importance_sampling_ratio/mean': 1.0527184009552002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.160714775323868, 'clip_ratio/low_mean': 0.09416915010660887, 'clip_ratio/low_min': 0.09416915010660887, 'clip_ratio/high_mean': 0.08993902429938316, 'clip_ratio/high_max': 0.08993902429938316, 'clip_ratio/region_mean': 0.18410817440599203, 'reward_total_mean': 0.25208374857902527, 'reward_meter_mean': 0.6623384952545166, 'reward_meter_std': 0.3999437093734741, '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': 0.9959326982498169, 'reward_repeat_soft_std': 0.0059642670676112175, 'reward_judge_quality_mean': 0.3075000047683716, 'reward_judge_quality_std': 0.12464234232902527, 'reward_total_composite_mean': 0.25208374857902527, 'reward_total_composite_std': 0.14597024023532867, 'epoch': 0.02}

  5%|▌         | 181/3300 [22:09<5:29:31,  6.34s/it]INFO 04-13 13:07:52 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 182/3300 [22:15<5:26:29,  6.28s/it]2026-04-13 13:07:58,387 | INFO | train_grpo_train | metrics_logged mode=train step=182

                                                    
{'loss': 0.1077, 'grad_norm': 23.082223892211914, 'learning_rate': 9.451515151515153e-06, 'num_tokens': 323623.0, 'completions/mean_length': 31.0, 'completions/min_length': 27.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.6646571159362793, 'rewards/meter/std': 0.35609716176986694, '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': 0.998947024345398, 'rewards/repeat_soft/std': 0.0015329680172726512, 'rewards/judge_quality/mean': 0.2800000011920929, 'rewards/judge_quality/std': 0.08976158499717712, 'rewards/total_composite/mean': 0.2163580358028412, 'rewards/total_composite/std': 0.11960655450820923, 'reward': 0.2163580358028412, 'reward_std': 0.11960654705762863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24264128506183624, 'sampling/sampling_logp_difference/max': 1.9742622375488281, 'sampling/importance_sampling_ratio/min': 0.1388637274503708, 'sampling/importance_sampling_ratio/mean': 1.0092780590057373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5921870172023773, 'clip_ratio/low_mean': 0.09682540036737919, 'clip_ratio/low_min': 0.09682540036737919, 'clip_ratio/high_mean': 0.10542328143492341, 'clip_ratio/high_max': 0.10542328143492341, 'clip_ratio/region_mean': 0.2022486818023026, 'reward_total_mean': 0.2163580358028412, 'reward_meter_mean': 0.6646571159362793, 'reward_meter_std': 0.35609716176986694, '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': 0.998947024345398, 'reward_repeat_soft_std': 0.0015329680172726512, 'reward_judge_quality_mean': 0.2800000011920929, 'reward_judge_quality_std': 0.08976158499717712, 'reward_total_composite_mean': 0.2163580358028412, 'reward_total_composite_std': 0.11960655450820923, 'epoch': 0.02}

  6%|▌         | 182/3300 [22:15<5:26:29,  6.28s/it]INFO 04-13 13:07:58 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 183/3300 [22:21<5:18:55,  6.14s/it]2026-04-13 13:08:04,190 | INFO | train_grpo_train | metrics_logged mode=train step=183

                                                    
{'loss': 0.0646, 'grad_norm': 23.999311447143555, 'learning_rate': 9.448484848484849e-06, 'num_tokens': 325260.0, 'completions/mean_length': 31.625, 'completions/min_length': 29.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7429101467132568, 'rewards/meter/std': 0.30495375394821167, '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': 0.9971073269844055, 'rewards/repeat_soft/std': 0.005529557354748249, 'rewards/judge_quality/mean': 0.3062499761581421, 'rewards/judge_quality/std': 0.09500939399003983, 'rewards/total_composite/mean': 0.24475175142288208, 'rewards/total_composite/std': 0.07109037786722183, 'reward': 0.24475175142288208, 'reward_std': 0.07109037041664124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17223596572875977, 'sampling/sampling_logp_difference/max': 2.227447986602783, 'sampling/importance_sampling_ratio/min': 0.10780319571495056, 'sampling/importance_sampling_ratio/mean': 1.04634690284729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0528384000062943, 'clip_ratio/low_mean': 0.11872574500739574, 'clip_ratio/low_min': 0.11872574500739574, 'clip_ratio/high_mean': 0.04085249127820134, 'clip_ratio/high_max': 0.04085249127820134, 'clip_ratio/region_mean': 0.15957823628559709, 'reward_total_mean': 0.24475175142288208, 'reward_meter_mean': 0.7429101467132568, 'reward_meter_std': 0.30495375394821167, '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': 0.9971073269844055, 'reward_repeat_soft_std': 0.005529557354748249, 'reward_judge_quality_mean': 0.3062499761581421, 'reward_judge_quality_std': 0.09500939399003983, 'reward_total_composite_mean': 0.24475175142288208, 'reward_total_composite_std': 0.07109037786722183, 'epoch': 0.02}

  6%|▌         | 183/3300 [22:21<5:18:55,  6.14s/it]INFO 04-13 13:08:04 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 184/3300 [22:27<5:14:55,  6.06s/it]2026-04-13 13:08:10,081 | INFO | train_grpo_train | metrics_logged mode=train step=184

                                                    
{'loss': 0.111, 'grad_norm': 15.967321395874023, 'learning_rate': 9.445454545454546e-06, 'num_tokens': 326825.0, 'completions/mean_length': 39.625, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5608224868774414, 'rewards/meter/std': 0.357774555683136, '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': 0.9991849064826965, 'rewards/repeat_soft/std': 0.001608764985576272, 'rewards/judge_quality/mean': 0.25999999046325684, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.18913134932518005, 'rewards/total_composite/std': 0.08285278081893921, 'reward': 0.18913134932518005, 'reward_std': 0.08285277336835861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2188105583190918, 'sampling/sampling_logp_difference/max': 1.127385139465332, 'sampling/importance_sampling_ratio/min': 0.32387906312942505, 'sampling/importance_sampling_ratio/mean': 1.0331828594207764, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.165031999349594, 'clip_ratio/low_mean': 0.1111050397157669, 'clip_ratio/low_min': 0.1111050397157669, 'clip_ratio/high_mean': 0.10008255392313004, 'clip_ratio/high_max': 0.10008255392313004, 'clip_ratio/region_mean': 0.21118759363889694, 'reward_total_mean': 0.18913134932518005, 'reward_meter_mean': 0.5608224868774414, 'reward_meter_std': 0.357774555683136, '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': 0.9991849064826965, 'reward_repeat_soft_std': 0.001608764985576272, 'reward_judge_quality_mean': 0.25999999046325684, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.18913134932518005, 'reward_total_composite_std': 0.08285278081893921, 'epoch': 0.02}

  6%|▌         | 184/3300 [22:27<5:14:55,  6.06s/it]INFO 04-13 13:08:10 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 185/3300 [22:33<5:12:15,  6.01s/it]2026-04-13 13:08:15,979 | INFO | train_grpo_train | metrics_logged mode=train step=185

                                                    
{'loss': 0.1385, 'grad_norm': 20.29409408569336, 'learning_rate': 9.442424242424243e-06, 'num_tokens': 328275.0, 'completions/mean_length': 29.25, 'completions/min_length': 18.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5379096269607544, 'rewards/meter/std': 0.43841618299484253, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.5345224738121033, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9771194458007812, 'rewards/repeat_soft/std': 0.01737549714744091, 'rewards/judge_quality/mean': 0.2762500047683716, 'rewards/judge_quality/std': 0.09179751574993134, 'rewards/total_composite/mean': 0.17947977781295776, 'rewards/total_composite/std': 0.1414765566587448, 'reward': 0.17947977781295776, 'reward_std': 0.1414765566587448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2503550946712494, 'sampling/sampling_logp_difference/max': 1.2705016136169434, 'sampling/importance_sampling_ratio/min': 0.2806907892227173, 'sampling/importance_sampling_ratio/mean': 1.0244457721710205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.178879290819168, 'clip_ratio/low_mean': 0.12165411468595266, 'clip_ratio/low_min': 0.12165411468595266, 'clip_ratio/high_mean': 0.11699196230620146, 'clip_ratio/high_max': 0.11699196230620146, 'clip_ratio/region_mean': 0.23864607699215412, 'reward_total_mean': 0.17947977781295776, 'reward_meter_mean': 0.5379096269607544, 'reward_meter_std': 0.43841618299484253, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.5345224738121033, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9771194458007812, 'reward_repeat_soft_std': 0.01737549714744091, 'reward_judge_quality_mean': 0.2762500047683716, 'reward_judge_quality_std': 0.09179751574993134, 'reward_total_composite_mean': 0.17947977781295776, 'reward_total_composite_std': 0.1414765566587448, 'epoch': 0.02}

  6%|▌         | 185/3300 [22:33<5:12:15,  6.01s/it]INFO 04-13 13:08:16 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 186/3300 [22:39<5:15:27,  6.08s/it]2026-04-13 13:08:22,206 | INFO | train_grpo_train | metrics_logged mode=train step=186

                                                    
{'loss': 0.0903, 'grad_norm': 16.462566375732422, 'learning_rate': 9.43939393939394e-06, 'num_tokens': 330206.0, 'completions/mean_length': 58.375, 'completions/min_length': 49.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.27843326330184937, 'rewards/meter/std': 0.245342418551445, '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': 0.9970526695251465, 'rewards/repeat_soft/std': 0.00237673451192677, 'rewards/judge_quality/mean': 0.22624999284744263, 'rewards/judge_quality/std': 0.02875388041138649, 'rewards/total_composite/mean': 0.09679193049669266, 'rewards/total_composite/std': 0.05075918138027191, 'reward': 0.09679193049669266, 'reward_std': 0.05075918138027191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24853521585464478, 'sampling/sampling_logp_difference/max': 2.083409309387207, 'sampling/importance_sampling_ratio/min': 0.12450501322746277, 'sampling/importance_sampling_ratio/mean': 1.0228904485702515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.215415120124817, 'clip_ratio/low_mean': 0.11344599910080433, 'clip_ratio/low_min': 0.11344599910080433, 'clip_ratio/high_mean': 0.11360197700560093, 'clip_ratio/high_max': 0.11360197700560093, 'clip_ratio/region_mean': 0.22704797610640526, 'reward_total_mean': 0.09679193049669266, 'reward_meter_mean': 0.27843326330184937, 'reward_meter_std': 0.245342418551445, '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': 0.9970526695251465, 'reward_repeat_soft_std': 0.00237673451192677, 'reward_judge_quality_mean': 0.22624999284744263, 'reward_judge_quality_std': 0.02875388041138649, 'reward_total_composite_mean': 0.09679193049669266, 'reward_total_composite_std': 0.05075918138027191, 'epoch': 0.02}

  6%|▌         | 186/3300 [22:39<5:15:27,  6.08s/it]INFO 04-13 13:08:22 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 187/3300 [22:47<5:33:29,  6.43s/it]2026-04-13 13:08:29,450 | INFO | train_grpo_train | metrics_logged mode=train step=187

                                                    
{'loss': 0.2065, 'grad_norm': 6.505162239074707, 'learning_rate': 9.436363636363636e-06, 'num_tokens': 332657.0, 'completions/mean_length': 108.375, 'completions/min_length': 55.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.375, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.5066986680030823, 'rewards/meter/std': 0.2979062497615814, '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': 0.9961463809013367, 'rewards/repeat_soft/std': 0.0023854782339185476, 'rewards/judge_quality/mean': 0.2449999898672104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.16829688847064972, 'rewards/total_composite/std': 0.08634115755558014, 'reward': 0.16829688847064972, 'reward_std': 0.08634115010499954, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23701831698417664, 'sampling/sampling_logp_difference/max': 1.195444107055664, 'sampling/importance_sampling_ratio/min': 0.30256953835487366, 'sampling/importance_sampling_ratio/mean': 1.079443335533142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9278726279735565, 'clip_ratio/low_mean': 0.14010917581617832, 'clip_ratio/low_min': 0.14010917581617832, 'clip_ratio/high_mean': 0.05738133378326893, 'clip_ratio/high_max': 0.05738133378326893, 'clip_ratio/region_mean': 0.19749050959944725, 'reward_total_mean': 0.16829688847064972, 'reward_meter_mean': 0.5066986680030823, 'reward_meter_std': 0.2979062497615814, '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': 0.9961463809013367, 'reward_repeat_soft_std': 0.0023854782339185476, 'reward_judge_quality_mean': 0.2449999898672104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.16829688847064972, 'reward_total_composite_std': 0.08634115755558014, 'epoch': 0.02}

  6%|▌         | 187/3300 [22:47<5:33:29,  6.43s/it]INFO 04-13 13:08:29 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 188/3300 [22:53<5:33:06,  6.42s/it]2026-04-13 13:08:35,860 | INFO | train_grpo_train | metrics_logged mode=train step=188

                                                    
{'loss': 0.2508, 'grad_norm': 15.58817195892334, 'learning_rate': 9.433333333333335e-06, 'num_tokens': 334275.0, 'completions/mean_length': 45.25, 'completions/min_length': 35.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7768089771270752, 'rewards/meter/std': 0.3373166024684906, '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': 0.9966387748718262, 'rewards/repeat_soft/std': 0.003940577618777752, 'rewards/judge_quality/mean': 0.3199999928474426, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.2800449728965759, 'rewards/total_composite/std': 0.1464841514825821, 'reward': 0.2800449728965759, 'reward_std': 0.1464841365814209, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.218722403049469, 'sampling/sampling_logp_difference/max': 1.2261724472045898, 'sampling/importance_sampling_ratio/min': 0.2934134602546692, 'sampling/importance_sampling_ratio/mean': 1.0738236904144287, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.576609343290329, 'clip_ratio/low_mean': 0.10117842815816402, 'clip_ratio/low_min': 0.10117842815816402, 'clip_ratio/high_mean': 0.0984889529645443, 'clip_ratio/high_max': 0.0984889529645443, 'clip_ratio/region_mean': 0.19966738112270832, 'reward_total_mean': 0.2800449728965759, 'reward_meter_mean': 0.7768089771270752, 'reward_meter_std': 0.3373166024684906, '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': 0.9966387748718262, 'reward_repeat_soft_std': 0.003940577618777752, 'reward_judge_quality_mean': 0.3199999928474426, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.2800449728965759, 'reward_total_composite_std': 0.1464841514825821, 'epoch': 0.02}

  6%|▌         | 188/3300 [22:53<5:33:06,  6.42s/it]INFO 04-13 13:08:36 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 189/3300 [22:59<5:22:04,  6.21s/it]2026-04-13 13:08:41,580 | INFO | train_grpo_train | metrics_logged mode=train step=189

                                                    
{'loss': 0.1027, 'grad_norm': 25.442771911621094, 'learning_rate': 9.43030303030303e-06, 'num_tokens': 335768.0, 'completions/mean_length': 30.625, 'completions/min_length': 25.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.625, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.19121405482292175, 'rewards/meter/std': 0.16347219049930573, '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': 0.993026852607727, 'rewards/repeat_soft/std': 0.006410476751625538, 'rewards/judge_quality/mean': 0.2199999988079071, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.10411349684000015, 'rewards/total_composite/std': 0.02334478311240673, 'reward': 0.10411349684000015, 'reward_std': 0.02334478311240673, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25707754492759705, 'sampling/sampling_logp_difference/max': 1.7045478820800781, 'sampling/importance_sampling_ratio/min': 0.18185459077358246, 'sampling/importance_sampling_ratio/mean': 0.9856594204902649, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8737333565950394, 'clip_ratio/low_mean': 0.10483114421367645, 'clip_ratio/low_min': 0.10483114421367645, 'clip_ratio/high_mean': 0.13011138327419758, 'clip_ratio/high_max': 0.13011138327419758, 'clip_ratio/region_mean': 0.23494252748787403, 'reward_total_mean': 0.10411349684000015, 'reward_meter_mean': 0.19121405482292175, 'reward_meter_std': 0.16347219049930573, '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': 0.993026852607727, 'reward_repeat_soft_std': 0.006410476751625538, 'reward_judge_quality_mean': 0.2199999988079071, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.10411349684000015, 'reward_total_composite_std': 0.02334478311240673, 'epoch': 0.02}

  6%|▌         | 189/3300 [22:59<5:22:04,  6.21s/it]INFO 04-13 13:08:41 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 190/3300 [23:05<5:23:35,  6.24s/it]2026-04-13 13:08:47,896 | INFO | train_grpo_train | metrics_logged mode=train step=190

                                                    
{'loss': 0.0027, 'grad_norm': 18.011791229248047, 'learning_rate': 9.427272727272728e-06, 'num_tokens': 337457.0, 'completions/mean_length': 42.125, 'completions/min_length': 31.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.44638946652412415, 'rewards/meter/std': 0.3409777283668518, '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': 0.9952579736709595, 'rewards/repeat_soft/std': 0.008636967279016972, 'rewards/judge_quality/mean': 0.23999999463558197, 'rewards/judge_quality/std': 0.07406560331583023, 'rewards/total_composite/mean': 0.14922666549682617, 'rewards/total_composite/std': 0.056265976279973984, 'reward': 0.14922666549682617, 'reward_std': 0.056265972554683685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23144479095935822, 'sampling/sampling_logp_difference/max': 1.2533469200134277, 'sampling/importance_sampling_ratio/min': 0.2855474650859833, 'sampling/importance_sampling_ratio/mean': 1.0466233491897583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6538505852222443, 'clip_ratio/low_mean': 0.10494689829647541, 'clip_ratio/low_min': 0.10494689829647541, 'clip_ratio/high_mean': 0.1201979462057352, 'clip_ratio/high_max': 0.1201979462057352, 'clip_ratio/region_mean': 0.22514484450221062, 'reward_total_mean': 0.14922666549682617, 'reward_meter_mean': 0.44638946652412415, 'reward_meter_std': 0.3409777283668518, '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': 0.9952579736709595, 'reward_repeat_soft_std': 0.008636967279016972, 'reward_judge_quality_mean': 0.23999999463558197, 'reward_judge_quality_std': 0.07406560331583023, 'reward_total_composite_mean': 0.14922666549682617, 'reward_total_composite_std': 0.056265976279973984, 'epoch': 0.02}

  6%|▌         | 190/3300 [23:05<5:23:35,  6.24s/it]INFO 04-13 13:08:48 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 191/3300 [23:12<5:39:35,  6.55s/it]2026-04-13 13:08:55,175 | INFO | train_grpo_train | metrics_logged mode=train step=191

                                                    
{'loss': 0.1239, 'grad_norm': 9.168624877929688, 'learning_rate': 9.424242424242425e-06, 'num_tokens': 339804.0, 'completions/mean_length': 108.375, 'completions/min_length': 90.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.375, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.6746603846549988, 'rewards/meter/std': 0.3454626798629761, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9948081970214844, 'rewards/repeat_soft/std': 0.004163301084190607, 'rewards/judge_quality/mean': 0.2449999898672104, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.17189136147499084, 'rewards/total_composite/std': 0.08835437148809433, 'reward': 0.17189136147499084, 'reward_std': 0.08835437148809433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24277395009994507, 'sampling/sampling_logp_difference/max': 1.1603548526763916, 'sampling/importance_sampling_ratio/min': 0.31337496638298035, 'sampling/importance_sampling_ratio/mean': 1.060469627380371, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.501090317964554, 'clip_ratio/low_mean': 0.07497364282608032, 'clip_ratio/low_min': 0.07497364282608032, 'clip_ratio/high_mean': 0.13471388444304466, 'clip_ratio/high_max': 0.13471388444304466, 'clip_ratio/region_mean': 0.20968752726912498, 'reward_total_mean': 0.17189136147499084, 'reward_meter_mean': 0.6746603846549988, 'reward_meter_std': 0.3454626798629761, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9948081970214844, 'reward_repeat_soft_std': 0.004163301084190607, 'reward_judge_quality_mean': 0.2449999898672104, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.17189136147499084, 'reward_total_composite_std': 0.08835437148809433, 'epoch': 0.02}

  6%|▌         | 191/3300 [23:12<5:39:35,  6.55s/it]INFO 04-13 13:08:55 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 192/3300 [23:18<5:28:29,  6.34s/it]2026-04-13 13:09:01,019 | INFO | train_grpo_train | metrics_logged mode=train step=192

                                                    
{'loss': 0.1073, 'grad_norm': 26.564390182495117, 'learning_rate': 9.421212121212122e-06, 'num_tokens': 341388.0, 'completions/mean_length': 27.0, 'completions/min_length': 22.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.0, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.5459522604942322, 'rewards/meter/std': 0.4349660873413086, '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': 0.9664831757545471, 'rewards/repeat_soft/std': 0.08932723104953766, 'rewards/judge_quality/mean': 0.3700000047683716, 'rewards/judge_quality/std': 0.1414213478565216, 'rewards/total_composite/mean': 0.24714788794517517, 'rewards/total_composite/std': 0.11071964353322983, 'reward': 0.24714788794517517, 'reward_std': 0.11071963608264923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2507804334163666, 'sampling/sampling_logp_difference/max': 2.2007923126220703, 'sampling/importance_sampling_ratio/min': 0.11071540415287018, 'sampling/importance_sampling_ratio/mean': 1.0464229583740234, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2436086535453796, 'clip_ratio/low_mean': 0.10971423983573914, 'clip_ratio/low_min': 0.10971423983573914, 'clip_ratio/high_mean': 0.10494063422083855, 'clip_ratio/high_max': 0.10494063422083855, 'clip_ratio/region_mean': 0.21465487405657768, 'reward_total_mean': 0.24714788794517517, 'reward_meter_mean': 0.5459522604942322, 'reward_meter_std': 0.4349660873413086, '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': 0.9664831757545471, 'reward_repeat_soft_std': 0.08932723104953766, 'reward_judge_quality_mean': 0.3700000047683716, 'reward_judge_quality_std': 0.1414213478565216, 'reward_total_composite_mean': 0.24714788794517517, 'reward_total_composite_std': 0.11071964353322983, 'epoch': 0.02}

  6%|▌         | 192/3300 [23:18<5:28:29,  6.34s/it]INFO 04-13 13:09:01 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 193/3300 [23:27<6:07:49,  7.10s/it]2026-04-13 13:09:09,902 | INFO | train_grpo_train | metrics_logged mode=train step=193

                                                    
{'loss': 0.6028, 'grad_norm': 12.519133567810059, 'learning_rate': 9.418181818181818e-06, 'num_tokens': 343477.0, 'completions/mean_length': 83.125, 'completions/min_length': 45.0, 'completions/max_length': 292.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 292.0, 'rewards/meter/mean': 0.6665387153625488, 'rewards/meter/std': 0.38474684953689575, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9970675110816956, 'rewards/repeat_soft/std': 0.0038372064009308815, 'rewards/judge_quality/mean': 0.23999999463558197, 'rewards/judge_quality/std': 0.08485281467437744, 'rewards/total_composite/mean': 0.19229841232299805, 'rewards/total_composite/std': 0.11471903324127197, 'reward': 0.19229841232299805, 'reward_std': 0.11471902579069138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24479778110980988, 'sampling/sampling_logp_difference/max': 1.2485408782958984, 'sampling/importance_sampling_ratio/min': 0.2869231402873993, 'sampling/importance_sampling_ratio/mean': 1.0734835863113403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.8438830971717834, 'clip_ratio/low_mean': 0.07805478479713202, 'clip_ratio/low_min': 0.07805478479713202, 'clip_ratio/high_mean': 0.10036757402122021, 'clip_ratio/high_max': 0.10036757402122021, 'clip_ratio/region_mean': 0.17842235881835222, 'reward_total_mean': 0.19229841232299805, 'reward_meter_mean': 0.6665387153625488, 'reward_meter_std': 0.38474684953689575, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9970675110816956, 'reward_repeat_soft_std': 0.0038372064009308815, 'reward_judge_quality_mean': 0.23999999463558197, 'reward_judge_quality_std': 0.08485281467437744, 'reward_total_composite_mean': 0.19229841232299805, 'reward_total_composite_std': 0.11471903324127197, 'epoch': 0.02}

  6%|▌         | 193/3300 [23:27<6:07:49,  7.10s/it]INFO 04-13 13:09:10 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 194/3300 [23:38<7:09:19,  8.29s/it]2026-04-13 13:09:20,972 | INFO | train_grpo_train | metrics_logged mode=train step=194

                                                    
{'loss': 0.3784, 'grad_norm': 5.710408687591553, 'learning_rate': 9.415151515151515e-06, 'num_tokens': 345504.0, 'completions/mean_length': 160.375, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 110.14286041259766, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 407.0, 'rewards/meter/mean': 0.3032819330692291, 'rewards/meter/std': 0.31939348578453064, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.45806270837783813, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9990205764770508, 'rewards/repeat_soft/std': 0.00227516982704401, 'rewards/judge_quality/mean': 0.2524999976158142, 'rewards/judge_quality/std': 0.14694508910179138, 'rewards/total_composite/mean': 0.14139440655708313, 'rewards/total_composite/std': 0.13582220673561096, 'reward': 0.14139440655708313, 'reward_std': 0.13582219183444977, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23204858601093292, 'sampling/sampling_logp_difference/max': 1.1267271041870117, 'sampling/importance_sampling_ratio/min': 0.32409223914146423, 'sampling/importance_sampling_ratio/mean': 1.0642539262771606, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.8016223311424255, 'clip_ratio/low_mean': 0.03341209935024381, 'clip_ratio/low_min': 0.03341209935024381, 'clip_ratio/high_mean': 0.08239117078483105, 'clip_ratio/high_max': 0.08239117078483105, 'clip_ratio/region_mean': 0.11580327013507485, 'reward_total_mean': 0.14139440655708313, 'reward_meter_mean': 0.3032819330692291, 'reward_meter_std': 0.31939348578453064, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.45806270837783813, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9990205764770508, 'reward_repeat_soft_std': 0.00227516982704401, 'reward_judge_quality_mean': 0.2524999976158142, 'reward_judge_quality_std': 0.14694508910179138, 'reward_total_composite_mean': 0.14139440655708313, 'reward_total_composite_std': 0.13582220673561096, 'epoch': 0.02}

  6%|▌         | 194/3300 [23:38<7:09:19,  8.29s/it]INFO 04-13 13:09:21 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 195/3300 [23:44<6:35:38,  7.65s/it]2026-04-13 13:09:27,103 | INFO | train_grpo_train | metrics_logged mode=train step=195

                                                    
{'loss': 0.1807, 'grad_norm': 11.035167694091797, 'learning_rate': 9.412121212121212e-06, 'num_tokens': 347212.0, 'completions/mean_length': 54.5, 'completions/min_length': 38.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.21689274907112122, 'rewards/meter/std': 0.2749010920524597, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974724054336548, 'rewards/repeat_soft/std': 0.004348366055637598, 'rewards/judge_quality/mean': 0.22749999165534973, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.09759880602359772, 'rewards/total_composite/std': 0.047488220036029816, 'reward': 0.09759880602359772, 'reward_std': 0.04748821631073952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25424426794052124, 'sampling/sampling_logp_difference/max': 1.341461181640625, 'sampling/importance_sampling_ratio/min': 0.26146334409713745, 'sampling/importance_sampling_ratio/mean': 1.063415288925171, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7377142012119293, 'clip_ratio/low_mean': 0.17252291180193424, 'clip_ratio/low_min': 0.17252291180193424, 'clip_ratio/high_mean': 0.06174089200794697, 'clip_ratio/high_max': 0.06174089200794697, 'clip_ratio/region_mean': 0.2342638038098812, 'reward_total_mean': 0.09759880602359772, 'reward_meter_mean': 0.21689274907112122, 'reward_meter_std': 0.2749010920524597, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974724054336548, 'reward_repeat_soft_std': 0.004348366055637598, 'reward_judge_quality_mean': 0.22749999165534973, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.09759880602359772, 'reward_total_composite_std': 0.047488220036029816, 'epoch': 0.02}

  6%|▌         | 195/3300 [23:44<6:35:38,  7.65s/it]INFO 04-13 13:09:27 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 196/3300 [23:50<6:03:28,  7.03s/it]2026-04-13 13:09:32,684 | INFO | train_grpo_train | metrics_logged mode=train step=196

                                                    
{'loss': 0.0586, 'grad_norm': 20.3045711517334, 'learning_rate': 9.40909090909091e-06, 'num_tokens': 348558.0, 'completions/mean_length': 19.25, 'completions/min_length': 15.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.5030122995376587, 'rewards/meter/std': 0.4006531834602356, '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': 0.47749999165534973, 'rewards/judge_quality/std': 0.27027764916419983, 'rewards/total_composite/mean': 0.3441108763217926, 'rewards/total_composite/std': 0.25904008746147156, 'reward': 0.3441108763217926, 'reward_std': 0.25904008746147156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18175098299980164, 'sampling/sampling_logp_difference/max': 1.2526583671569824, 'sampling/importance_sampling_ratio/min': 0.28574416041374207, 'sampling/importance_sampling_ratio/mean': 1.0194385051727295, 'sampling/importance_sampling_ratio/max': 1.8192806243896484, 'entropy': 2.041719302535057, 'clip_ratio/low_mean': 0.12170269526541233, 'clip_ratio/low_min': 0.12170269526541233, 'clip_ratio/high_mean': 0.029495615046471357, 'clip_ratio/high_max': 0.029495615046471357, 'clip_ratio/region_mean': 0.1511983103118837, 'reward_total_mean': 0.3441108763217926, 'reward_meter_mean': 0.5030122995376587, 'reward_meter_std': 0.4006531834602356, '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': 0.47749999165534973, 'reward_judge_quality_std': 0.27027764916419983, 'reward_total_composite_mean': 0.3441108763217926, 'reward_total_composite_std': 0.25904008746147156, 'epoch': 0.02}

  6%|▌         | 196/3300 [23:50<6:03:28,  7.03s/it]INFO 04-13 13:09:32 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 197/3300 [23:56<5:45:20,  6.68s/it]2026-04-13 13:09:38,551 | INFO | train_grpo_train | metrics_logged mode=train step=197

                                                    
{'loss': 0.1945, 'grad_norm': 13.562896728515625, 'learning_rate': 9.406060606060607e-06, 'num_tokens': 349978.0, 'completions/mean_length': 30.5, 'completions/min_length': 16.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8292700052261353, 'rewards/meter/std': 0.32876673340797424, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9738503098487854, 'rewards/repeat_soft/std': 0.021921547129750252, 'rewards/judge_quality/mean': 0.45499998331069946, 'rewards/judge_quality/std': 0.3196873664855957, 'rewards/total_composite/mean': 0.4093332290649414, 'rewards/total_composite/std': 0.3494712710380554, 'reward': 0.4093332290649414, 'reward_std': 0.34947124123573303, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22632241249084473, 'sampling/sampling_logp_difference/max': 1.3101892471313477, 'sampling/importance_sampling_ratio/min': 0.26976901292800903, 'sampling/importance_sampling_ratio/mean': 1.0688990354537964, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.687138594686985, 'clip_ratio/low_mean': 0.12821621261537075, 'clip_ratio/low_min': 0.12821621261537075, 'clip_ratio/high_mean': 0.06647727452218533, 'clip_ratio/high_max': 0.06647727452218533, 'clip_ratio/region_mean': 0.19469348713755608, 'reward_total_mean': 0.4093332290649414, 'reward_meter_mean': 0.8292700052261353, 'reward_meter_std': 0.32876673340797424, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9738503098487854, 'reward_repeat_soft_std': 0.021921547129750252, 'reward_judge_quality_mean': 0.45499998331069946, 'reward_judge_quality_std': 0.3196873664855957, 'reward_total_composite_mean': 0.4093332290649414, 'reward_total_composite_std': 0.3494712710380554, 'epoch': 0.02}

  6%|▌         | 197/3300 [23:56<5:45:20,  6.68s/it]INFO 04-13 13:09:38 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 198/3300 [24:02<5:34:15,  6.47s/it]2026-04-13 13:09:44,520 | INFO | train_grpo_train | metrics_logged mode=train step=198

                                                    
{'loss': 0.0457, 'grad_norm': 18.57405662536621, 'learning_rate': 9.403030303030304e-06, 'num_tokens': 351830.0, 'completions/mean_length': 48.5, 'completions/min_length': 40.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.3581885099411011, 'rewards/meter/std': 0.3801507353782654, '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': 0.993811845779419, 'rewards/repeat_soft/std': 0.005281467922031879, 'rewards/judge_quality/mean': 0.3799999952316284, 'rewards/judge_quality/std': 0.2112547606229782, 'rewards/total_composite/mean': 0.2376668006181717, 'rewards/total_composite/std': 0.22460903227329254, 'reward': 0.2376668006181717, 'reward_std': 0.22460903227329254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.214599147439003, 'sampling/sampling_logp_difference/max': 3.266845226287842, 'sampling/importance_sampling_ratio/min': 0.038126517087221146, 'sampling/importance_sampling_ratio/mean': 0.9915326833724976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.411556027829647, 'clip_ratio/low_mean': 0.1614802423864603, 'clip_ratio/low_min': 0.1614802423864603, 'clip_ratio/high_mean': 0.05551740154623985, 'clip_ratio/high_max': 0.05551740154623985, 'clip_ratio/region_mean': 0.21699764393270016, 'reward_total_mean': 0.2376668006181717, 'reward_meter_mean': 0.3581885099411011, 'reward_meter_std': 0.3801507353782654, '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': 0.993811845779419, 'reward_repeat_soft_std': 0.005281467922031879, 'reward_judge_quality_mean': 0.3799999952316284, 'reward_judge_quality_std': 0.2112547606229782, 'reward_total_composite_mean': 0.2376668006181717, 'reward_total_composite_std': 0.22460903227329254, 'epoch': 0.02}

  6%|▌         | 198/3300 [24:02<5:34:15,  6.47s/it]INFO 04-13 13:09:44 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 199/3300 [24:07<5:24:38,  6.28s/it]2026-04-13 13:09:50,372 | INFO | train_grpo_train | metrics_logged mode=train step=199

                                                    
{'loss': 0.3595, 'grad_norm': 22.297731399536133, 'learning_rate': 9.4e-06, 'num_tokens': 353200.0, 'completions/mean_length': 26.25, 'completions/min_length': 16.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5058358907699585, 'rewards/meter/std': 0.4252464175224304, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9765625, 'rewards/repeat_soft/std': 0.019408106803894043, 'rewards/judge_quality/mean': 0.25999999046325684, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.15322190523147583, 'rewards/total_composite/std': 0.0743454247713089, 'reward': 0.15322190523147583, 'reward_std': 0.0743454247713089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2133527547121048, 'sampling/sampling_logp_difference/max': 0.93658447265625, 'sampling/importance_sampling_ratio/min': 0.391964316368103, 'sampling/importance_sampling_ratio/mean': 1.0662285089492798, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.532737761735916, 'clip_ratio/low_mean': 0.08266637660562992, 'clip_ratio/low_min': 0.08266637660562992, 'clip_ratio/high_mean': 0.07500000111758709, 'clip_ratio/high_max': 0.07500000111758709, 'clip_ratio/region_mean': 0.157666377723217, 'reward_total_mean': 0.15322190523147583, 'reward_meter_mean': 0.5058358907699585, 'reward_meter_std': 0.4252464175224304, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9765625, 'reward_repeat_soft_std': 0.019408106803894043, 'reward_judge_quality_mean': 0.25999999046325684, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.15322190523147583, 'reward_total_composite_std': 0.0743454247713089, 'epoch': 0.02}

  6%|▌         | 199/3300 [24:07<5:24:38,  6.28s/it]INFO 04-13 13:09:50 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 200/3300 [24:18<6:31:23,  7.58s/it]2026-04-13 13:10:00,968 | INFO | train_grpo_train | metrics_logged mode=train step=200

                                                    
{'loss': 0.731, 'grad_norm': 10.606621742248535, 'learning_rate': 9.396969696969697e-06, 'num_tokens': 355647.0, 'completions/mean_length': 131.875, 'completions/min_length': 73.0, 'completions/max_length': 451.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 451.0, 'rewards/meter/mean': 0.7083677053451538, 'rewards/meter/std': 0.39610084891319275, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9976977705955505, 'rewards/repeat_soft/std': 0.002274994971230626, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.05249149724841118, 'rewards/total_composite/mean': 0.1362566351890564, 'rewards/total_composite/std': 0.07544992119073868, 'reward': 0.1362566351890564, 'reward_std': 0.07544991374015808, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22925913333892822, 'sampling/sampling_logp_difference/max': 1.3327980041503906, 'sampling/importance_sampling_ratio/min': 0.2637382745742798, 'sampling/importance_sampling_ratio/mean': 1.0662697553634644, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 5.329273879528046, 'clip_ratio/low_mean': 0.04947514459490776, 'clip_ratio/low_min': 0.04947514459490776, 'clip_ratio/high_mean': 0.11767296493053436, 'clip_ratio/high_max': 0.11767296493053436, 'clip_ratio/region_mean': 0.16714810952544212, 'reward_total_mean': 0.1362566351890564, 'reward_meter_mean': 0.7083677053451538, 'reward_meter_std': 0.39610084891319275, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.18898223340511322, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9976977705955505, 'reward_repeat_soft_std': 0.002274994971230626, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.05249149724841118, 'reward_total_composite_mean': 0.1362566351890564, 'reward_total_composite_std': 0.07544992119073868, 'epoch': 0.02}

  6%|▌         | 200/3300 [24:18<6:31:23,  7.58s/it]INFO 04-13 13:10:01 [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 [00:07<00:29,  3.71s/it]

 30%|███       | 3/10 [00:16<00:41,  5.87s/it]

 40%|████      | 4/10 [00:19<00:28,  4.82s/it]

 50%|█████     | 5/10 [00:22<00:21,  4.39s/it]

 60%|██████    | 6/10 [00:31<00:23,  5.88s/it]

 70%|███████   | 7/10 [00:36<00:16,  5.45s/it]

 80%|████████  | 8/10 [00:40<00:09,  4.88s/it]

 90%|█████████ | 9/10 [00:48<00:06,  6.04s/it]

100%|██████████| 10/10 [00:52<00:00,  5.34s/it]2026-04-13 13:10:58,989 | INFO | train_grpo_train | metrics_logged mode=eval step=200

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 58.0055, 'eval_samples_per_second': 1.379, 'eval_steps_per_second': 0.172, 'eval_num_tokens': 355647.0, 'eval_completions/mean_length': 97.65, 'eval_completions/min_length': 27.2, 'eval_completions/max_length': 271.4, 'eval_completions/clipped_ratio': 0.0375, 'eval_completions/mean_terminated_length': 83.01964492797852, 'eval_completions/min_terminated_length': 27.2, 'eval_completions/max_terminated_length': 213.9, 'eval_rewards/meter/mean': 0.3731417119503021, 'eval_rewards/meter/std': 0.32877095341682433, 'eval_rewards/count_adherence/mean': 0.9172916769981384, 'eval_rewards/count_adherence/std': 0.16572578102350236, 'eval_rewards/hard_gate/mean': 0.85, 'eval_rewards/hard_gate/std': 0.28198386132717135, 'eval_rewards/repeat_soft/mean': 0.9944182991981506, 'eval_rewards/repeat_soft/std': 0.008176561235450209, 'eval_rewards/judge_quality/mean': 0.2240000009536743, 'eval_rewards/judge_quality/std': 0.09404514394700528, 'eval_rewards/total_composite/mean': 0.126047283411026, 'eval_rewards/total_composite/std': 0.08404805548489094, 'eval_reward': 0.126047283411026, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1952221006155014, 'eval_sampling/sampling_logp_difference/max': 1.064872646331787, 'eval_sampling/importance_sampling_ratio/min': 0.34764412939548495, 'eval_sampling/importance_sampling_ratio/mean': 1.060666012763977, 'eval_sampling/importance_sampling_ratio/max': 1.7729849934577941, 'eval_entropy': 4.640206694602966, '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': 0.126047283411026, 'eval_reward_meter_mean': 0.3731417119503021, 'eval_reward_meter_std': 0.32877095341682433, 'eval_reward_count_adherence_mean': 0.9172916769981384, 'eval_reward_count_adherence_std': 0.16572578102350236, 'eval_reward_hard_gate_mean': 0.85, 'eval_reward_hard_gate_std': 0.28198386132717135, 'eval_reward_repeat_soft_mean': 0.9944182991981506, 'eval_reward_repeat_soft_std': 0.008176561235450209, 'eval_reward_judge_quality_mean': 0.2240000009536743, 'eval_reward_judge_quality_std': 0.09404514394700528, 'eval_reward_total_composite_mean': 0.126047283411026, 'eval_reward_total_composite_std': 0.08404805548489094, 'epoch': 0.02}

  6%|▌         | 200/3300 [25:16<6:31:23,  7.58s/it]

100%|██████████| 10/10 [00:52<00:00,  5.34s/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(