rollouts: control-thinking/control-thinking-phase32/rollouts/raw_rollouts.rank-0.jsonl provenance

rollouts/control-thinking/control-thinking-phase32/rollouts/raw_rollouts.rank-0.jsonl.meta.json

@@ -0,0 +1,14 @@
+{
+ "source_repo": "arcadia-impact/scimt-dispatch-rlvr-gemma4-26b-v1-runs",
+ "source_path": "control-thinking/control-thinking-phase32/rollouts/raw_rollouts.rank-0.jsonl",
+ "source_bytes": 81703466,
+ "rows": 1024,
+ "dropped_fields": [
+  "completion_ids",
+  "log_extra",
+  "log_metric",
+  "trainer_state"
+ ],
+ "trainer_state": "TrainerState(epoch=0.020833333333333332, global_step=16, max_steps=32, logging_steps=1, eval_steps=500, save_steps=17, train_batch_size=4, num_train_epochs=1, num_input_tokens_seen=4146766, total_flos=0.0, log_history=[{'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.765625, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3606.0, 'completions/mean_length': 3669.421875, 'completions/mean_terminated_length': 2275.933349609375, 'completions/min_length': 1060.0, 'completions/min_terminated_length': 1060.0, 'entropy': 0.11464092414826155, 'epoch': 0.0013020833333333333, 'frac_reward_zero_std': 0.75, 'grad_norm': 5.118177432450466e-05, 'learning_rate': 1e-05, 'loss': -0.00012268195860087872, 'num_tokens': 294619.0, 'reward': 0.09375, 'reward/selected_zero_std_group_fraction': 0.5, 'reward/zero_std_group_fraction': 0.75, 'reward_components/channel_close_count': 0.234375, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.765625, 'reward_components/format_valid': 0.09375, 'reward_components/native_boundary_valid': 0.234375, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.234375, 'reward_components/parser_unsafe': 0.140625, 'reward_components/parser_valid': 0.09375, 'reward_components/reward': 0.09375, 'reward_components/runs_correct': 0.1875, 'reward_components/runs_total': 1.625, 'reward_components/semantic_correct': 0.09375, 'reward_std': 0.29378482699394226, 'rewards/reward_func/mean': 0.09375, 'rewards/reward_func/std': 0.29378482699394226, 'sampling/importance_sampling_ratio/max': 1.0, 'sampling/importance_sampling_ratio/mean': 0.7739913463592529, 'sampling/importance_sampling_ratio/min': 6.604204116911205e-08, 'sampling/sampling_logp_difference/max': 5.984663963317871, 'sampling/sampling_logp_difference/mean': 0.017266588285565376, 'step': 1, 'step_time': 199.22395840508398}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.484375, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3985.0, 'completions/mean_length': 3057.71875, 'completions/mean_terminated_length': 2082.36376953125, 'completions/min_length': 1040.0, 'completions/min_terminated_length': 1040.0, 'entropy': 0.07956582214683294, 'epoch': 0.0026041666666666665, 'frac_reward_zero_std': 0.75, 'grad_norm': 0.006041001062840223, 'learning_rate': 1e-05, 'loss': 0.008527711033821106, 'num_tokens': 567881.0, 'reward': 0.46875, 'reward/selected_zero_std_group_fraction': 0.5, 'reward/zero_std_group_fraction': 0.75, 'reward_components/channel_close_count': 0.515625, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.484375, 'reward_components/format_valid': 0.484375, 'reward_components/native_boundary_valid': 0.515625, 'reward_components/parser_json': 0.015625, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.5, 'reward_components/parser_unsafe': 0.03125, 'reward_components/parser_valid': 0.484375, 'reward_components/reward': 0.46875, 'reward_components/runs_correct': 0.609375, 'reward_components/runs_total': 1.375, 'reward_components/semantic_correct': 0.46875, 'reward_std': 0.5029674172401428, 'rewards/reward_func/mean': 0.46875, 'rewards/reward_func/std': 0.5029674172401428, 'sampling/importance_sampling_ratio/max': 1.0, 'sampling/importance_sampling_ratio/mean': 0.5096418857574463, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 4.357892990112305, 'sampling/sampling_logp_difference/mean': 0.015825413167476654, 'step': 2, 'step_time': 176.64610660797916}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.296875, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3973.0, 'completions/mean_length': 2763.734375, 'completions/mean_terminated_length': 2201.22216796875, 'completions/min_length': 829.0, 'completions/min_terminated_length': 829.0, 'entropy': 0.14969482272863388, 'epoch': 0.00390625, 'frac_reward_zero_std': 0.5, 'grad_norm': 0.003225702326744795, 'learning_rate': 1e-05, 'loss': -0.009826848283410072, 'num_tokens': 802104.0, 'reward': 0.515625, 'reward/selected_zero_std_group_fraction': 0.3333333333333333, 'reward/zero_std_group_fraction': 0.6666666666666666, 'reward_components/channel_close_count': 0.71875, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.296875, 'reward_components/format_valid': 0.5625, 'reward_components/native_boundary_valid': 0.71875, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.71875, 'reward_components/parser_unsafe': 0.15625, 'reward_components/parser_valid': 0.5625, 'reward_components/reward': 0.515625, 'reward_components/runs_correct': 0.796875, 'reward_components/runs_total': 1.5, 'reward_components/semantic_correct': 0.515625, 'reward_std': 0.5037065148353577, 'rewards/reward_func/mean': 0.515625, 'rewards/reward_func/std': 0.5037065148353577, 'sampling/importance_sampling_ratio/max': 1.0, 'sampling/importance_sampling_ratio/mean': 0.3443005084991455, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 7.6105804443359375, 'sampling/sampling_logp_difference/mean': 0.018400561064481735, 'step': 3, 'step_time': 164.07066006807145}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.46875, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 4038.0, 'completions/mean_length': 3050.71875, 'completions/mean_terminated_length': 2128.411865234375, 'completions/min_length': 764.0, 'completions/min_terminated_length': 764.0, 'entropy': 0.1334073906764388, 'epoch': 0.005208333333333333, 'frac_reward_zero_std': 0.625, 'grad_norm': 0.0029910786543041468, 'learning_rate': 1e-05, 'loss': -0.00887129083275795, 'num_tokens': 1074790.0, 'reward': 0.359375, 'reward/selected_zero_std_group_fraction': 0.3125, 'reward/zero_std_group_fraction': 0.65625, 'reward_components/channel_close_count': 0.53125, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.46875, 'reward_components/format_valid': 0.375, 'reward_components/native_boundary_valid': 0.53125, 'reward_components/parser_json': 0.015625, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.515625, 'reward_components/parser_unsafe': 0.15625, 'reward_components/parser_valid': 0.375, 'reward_components/reward': 0.359375, 'reward_components/runs_correct': 0.546875, 'reward_components/runs_total': 1.625, 'reward_components/semantic_correct': 0.3671875, 'reward_std': 0.4836103618144989, 'rewards/reward_func/mean': 0.359375, 'rewards/reward_func/std': 0.4836103618144989, 'sampling/importance_sampling_ratio/max': 1.071272850036621, 'sampling/importance_sampling_ratio/mean': 0.5123374462127686, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 4.785722255706787, 'sampling/sampling_logp_difference/mean': 0.01855107583105564, 'step': 4, 'step_time': 178.95729677181225}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.4375, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3831.0, 'completions/mean_length': 2997.921875, 'completions/mean_terminated_length': 2143.861083984375, 'completions/min_length': 950.0, 'completions/min_terminated_length': 950.0, 'entropy': 0.09467202052474022, 'epoch': 0.006510416666666667, 'frac_reward_zero_std': 0.5, 'grad_norm': 0.0023277432192116976, 'learning_rate': 1e-05, 'loss': -0.007675642147660255, 'num_tokens': 1331361.0, 'reward': 0.5625, 'reward/selected_zero_std_group_fraction': 0.25, 'reward/zero_std_group_fraction': 0.625, 'reward_components/channel_close_count': 0.5625, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.4375, 'reward_components/format_valid': 0.5625, 'reward_components/native_boundary_valid': 0.5625, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.5625, 'reward_components/parser_unsafe': 0.0, 'reward_components/parser_valid': 0.5625, 'reward_components/reward': 0.5625, 'reward_components/runs_correct': 0.890625, 'reward_components/runs_total': 1.625, 'reward_components/semantic_correct': 0.5625, 'reward_std': 0.5, 'rewards/reward_func/mean': 0.5625, 'rewards/reward_func/std': 0.5, 'sampling/importance_sampling_ratio/max': 2.8180346488952637, 'sampling/importance_sampling_ratio/mean': 0.4982297718524933, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 3.2166824340820312, 'sampling/sampling_logp_difference/mean': 0.013301491737365723, 'step': 5, 'step_time': 174.63352737308014}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.46875, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 4029.0, 'completions/mean_length': 3091.421875, 'completions/mean_terminated_length': 2205.029541015625, 'completions/min_length': 992.0, 'completions/min_terminated_length': 992.0, 'entropy': 0.0794667536392808, 'epoch': 0.0078125, 'frac_reward_zero_std': 0.75, 'grad_norm': 0.00024385291908401996, 'learning_rate': 1e-05, 'loss': -0.0006516455323435366, 'num_tokens': 1596220.0, 'reward': 0.21875, 'reward/selected_zero_std_group_fraction': 0.2916666666666667, 'reward/zero_std_group_fraction': 0.6458333333333334, 'reward_components/channel_close_count': 0.546875, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.46875, 'reward_components/format_valid': 0.21875, 'reward_components/native_boundary_valid': 0.546875, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.546875, 'reward_components/parser_unsafe': 0.3125, 'reward_components/parser_valid': 0.21875, 'reward_components/reward': 0.21875, 'reward_components/runs_correct': 0.4375, 'reward_components/runs_total': 1.5, 'reward_components/semantic_correct': 0.21875, 'reward_std': 0.4166666865348816, 'rewards/reward_func/mean': 0.21875, 'rewards/reward_func/std': 0.4166666865348816, 'sampling/importance_sampling_ratio/max': 1.2318096160888672, 'sampling/importance_sampling_ratio/mean': 0.5213725566864014, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 4.652525424957275, 'sampling/sampling_logp_difference/mean': 0.01827034167945385, 'step': 6, 'step_time': 176.67608823697083}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.328125, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 4018.0, 'completions/mean_length': 2612.953125, 'completions/mean_terminated_length': 1888.6744384765625, 'completions/min_length': 510.0, 'completions/min_terminated_length': 510.0, 'entropy': 0.09812331292778254, 'epoch': 0.009114583333333334, 'frac_reward_zero_std': 0.5, 'grad_norm': 0.007505455520004034, 'learning_rate': 1e-05, 'loss': -0.0021405669394880533, 'num_tokens': 1829241.0, 'reward': 0.53125, 'reward/selected_zero_std_group_fraction': 0.25, 'reward/zero_std_group_fraction': 0.625, 'reward_components/channel_close_count': 0.671875, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.328125, 'reward_components/format_valid': 0.53125, 'reward_components/native_boundary_valid': 0.671875, 'reward_components/parser_json': 0.109375, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.5625, 'reward_components/parser_unsafe': 0.140625, 'reward_components/parser_valid': 0.53125, 'reward_components/reward': 0.53125, 'reward_components/runs_correct': 0.828125, 'reward_components/runs_total': 1.625, 'reward_components/semantic_correct': 0.53125, 'reward_std': 0.5029674172401428, 'rewards/reward_func/mean': 0.53125, 'rewards/reward_func/std': 0.5029674172401428, 'sampling/importance_sampling_ratio/max': 1.0, 'sampling/importance_sampling_ratio/mean': 0.4033518135547638, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 9.102952003479004, 'sampling/sampling_logp_difference/mean': 0.013518767431378365, 'step': 7, 'step_time': 166.27224498591386}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.40625, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 4045.0, 'completions/mean_length': 3009.421875, 'completions/mean_terminated_length': 2265.9736328125, 'completions/min_length': 805.0, 'completions/min_terminated_length': 805.0, 'entropy': 0.11144792195409536, 'epoch': 0.010416666666666666, 'frac_reward_zero_std': 0.5, 'grad_norm': 0.0028186386916786432, 'learning_rate': 1e-05, 'loss': 0.0031182863749563694, 'num_tokens': 2093076.0, 'reward': 0.5, 'reward/selected_zero_std_group_fraction': 0.21875, 'reward/zero_std_group_fraction': 0.609375, 'reward_components/channel_close_count': 0.59375, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.40625, 'reward_components/format_valid': 0.5, 'reward_components/native_boundary_valid': 0.59375, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.59375, 'reward_components/parser_unsafe': 0.09375, 'reward_components/parser_valid': 0.5, 'reward_components/reward': 0.5, 'reward_components/runs_correct': 0.921875, 'reward_components/runs_total': 1.625, 'reward_components/semantic_correct': 0.5, 'reward_std': 0.5039526224136353, 'rewards/reward_func/mean': 0.5, 'rewards/reward_func/std': 0.5039526224136353, 'sampling/importance_sampling_ratio/max': 1.4967069625854492, 'sampling/importance_sampling_ratio/mean': 0.48360365629196167, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 2.9243125915527344, 'sampling/sampling_logp_difference/mean': 0.015845099464058876, 'step': 8, 'step_time': 176.7978098159656}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.53125, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3904.0, 'completions/mean_length': 3180.3125, 'completions/mean_terminated_length': 2142.533447265625, 'completions/min_length': 722.0, 'completions/min_terminated_length': 722.0, 'entropy': 0.11426979303359985, 'epoch': 0.01171875, 'frac_reward_zero_std': 0.75, 'grad_norm': 0.00018392878700979054, 'learning_rate': 1e-05, 'loss': -0.0003445656329859048, 'num_tokens': 2377064.0, 'reward': 0.453125, 'reward/selected_zero_std_group_fraction': 0.25, 'reward/zero_std_group_fraction': 0.625, 'reward_components/channel_close_count': 0.46875, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.53125, 'reward_components/format_valid': 0.453125, 'reward_components/native_boundary_valid': 0.46875, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.46875, 'reward_components/parser_unsafe': 0.015625, 'reward_components/parser_valid': 0.453125, 'reward_components/reward': 0.453125, 'reward_components/runs_correct': 0.734375, 'reward_components/runs_total': 1.625, 'reward_components/semantic_correct': 0.453125, 'reward_std': 0.501733124256134, 'rewards/reward_func/mean': 0.453125, 'rewards/reward_func/std': 0.501733124256134, 'sampling/importance_sampling_ratio/max': 1.9269592761993408, 'sampling/importance_sampling_ratio/mean': 0.6507316827774048, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 3.993579626083374, 'sampling/sampling_logp_difference/mean': 0.015155397355556488, 'step': 9, 'step_time': 183.32763514318503}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.359375, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 4057.0, 'completions/mean_length': 2577.953125, 'completions/mean_terminated_length': 1726.3658447265625, 'completions/min_length': 426.0, 'completions/min_terminated_length': 426.0, 'entropy': 0.14673618786036968, 'epoch': 0.013020833333333334, 'frac_reward_zero_std': 0.875, 'grad_norm': 0.000542615307494998, 'learning_rate': 1e-05, 'loss': -0.001980179687961936, 'num_tokens': 2596197.0, 'reward': 0.359375, 'reward/selected_zero_std_group_fraction': 0.3, 'reward/zero_std_group_fraction': 0.65, 'reward_components/channel_close_count': 0.640625, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.359375, 'reward_components/format_valid': 0.359375, 'reward_components/native_boundary_valid': 0.640625, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.640625, 'reward_components/parser_unsafe': 0.28125, 'reward_components/parser_valid': 0.359375, 'reward_components/reward': 0.359375, 'reward_components/runs_correct': 0.46875, 'reward_components/runs_total': 1.125, 'reward_components/semantic_correct': 0.359375, 'reward_std': 0.4836103618144989, 'rewards/reward_func/mean': 0.359375, 'rewards/reward_func/std': 0.4836103618144989, 'sampling/importance_sampling_ratio/max': 2.0611624717712402, 'sampling/importance_sampling_ratio/mean': 0.46213120222091675, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 2.446122407913208, 'sampling/sampling_logp_difference/mean': 0.017191965132951736, 'step': 10, 'step_time': 157.8165815101238}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.5, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3812.0, 'completions/mean_length': 2687.515625, 'completions/mean_terminated_length': 1279.03125, 'completions/min_length': 447.0, 'completions/min_terminated_length': 447.0, 'entropy': 0.17564184218645096, 'epoch': 0.014322916666666666, 'frac_reward_zero_std': 0.75, 'grad_norm': 0.01946909911930561, 'learning_rate': 1e-05, 'loss': -0.003934633452445269, 'num_tokens': 2827910.0, 'reward': 0.359375, 'reward/selected_zero_std_group_fraction': 0.3181818181818182, 'reward/zero_std_group_fraction': 0.6590909090909091, 'reward_components/channel_close_count': 0.5, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.5, 'reward_components/format_valid': 0.359375, 'reward_components/native_boundary_valid': 0.5, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.5, 'reward_components/parser_unsafe': 0.140625, 'reward_components/parser_valid': 0.359375, 'reward_components/reward': 0.359375, 'reward_components/runs_correct': 0.390625, 'reward_components/runs_total': 1.375, 'reward_components/semantic_correct': 0.359375, 'reward_std': 0.4836103618144989, 'rewards/reward_func/mean': 0.359375, 'rewards/reward_func/std': 0.4836103618144989, 'sampling/importance_sampling_ratio/max': 1.251930594444275, 'sampling/importance_sampling_ratio/mean': 0.6004859805107117, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 17.22214698791504, 'sampling/sampling_logp_difference/mean': 0.020521041005849838, 'step': 11, 'step_time': 161.99184783909004}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.421875, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3902.0, 'completions/mean_length': 2724.078125, 'completions/mean_terminated_length': 1722.946044921875, 'completions/min_length': 694.0, 'completions/min_terminated_length': 694.0, 'entropy': 0.08522575721144676, 'epoch': 0.015625, 'frac_reward_zero_std': 0.75, 'grad_norm': 0.0013244427973404527, 'learning_rate': 1e-05, 'loss': -0.0021439625415951014, 'num_tokens': 3082123.0, 'reward': 0.53125, 'reward/selected_zero_std_group_fraction': 0.3333333333333333, 'reward/zero_std_group_fraction': 0.6666666666666666, 'reward_components/channel_close_count': 0.578125, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.421875, 'reward_components/format_valid': 0.53125, 'reward_components/native_boundary_valid': 0.578125, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.578125, 'reward_components/parser_unsafe': 0.046875, 'reward_components/parser_valid': 0.53125, 'reward_components/reward': 0.53125, 'reward_components/runs_correct': 0.65625, 'reward_components/runs_total': 1.375, 'reward_components/semantic_correct': 0.53125, 'reward_std': 0.5029674172401428, 'rewards/reward_func/mean': 0.53125, 'rewards/reward_func/std': 0.5029674172401428, 'sampling/importance_sampling_ratio/max': 1.247983455657959, 'sampling/importance_sampling_ratio/mean': 0.5126374959945679, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 4.970280647277832, 'sampling/sampling_logp_difference/mean': 0.014084962196648121, 'step': 12, 'step_time': 167.04041529202368}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.328125, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3936.0, 'completions/mean_length': 2455.140625, 'completions/mean_terminated_length': 1653.7906494140625, 'completions/min_length': 611.0, 'completions/min_terminated_length': 611.0, 'entropy': 0.13437950238585472, 'epoch': 0.016927083333333332, 'frac_reward_zero_std': 0.5, 'grad_norm': 0.0015296951169148088, 'learning_rate': 1e-05, 'loss': -0.0026422361843287945, 'num_tokens': 3293524.0, 'reward': 0.46875, 'reward/selected_zero_std_group_fraction': 0.3076923076923077, 'reward/zero_std_group_fraction': 0.6538461538461539, 'reward_components/channel_close_count': 0.671875, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.328125, 'reward_components/format_valid': 0.546875, 'reward_components/native_boundary_valid': 0.671875, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.671875, 'reward_components/parser_unsafe': 0.125, 'reward_components/parser_valid': 0.546875, 'reward_components/reward': 0.46875, 'reward_components/runs_correct': 0.59375, 'reward_components/runs_total': 1.375, 'reward_components/semantic_correct': 0.46875, 'reward_std': 0.5029674172401428, 'rewards/reward_func/mean': 0.46875, 'rewards/reward_func/std': 0.5029674172401428, 'sampling/importance_sampling_ratio/max': 2.9824352264404297, 'sampling/importance_sampling_ratio/mean': 0.4512457251548767, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 10.88036823272705, 'sampling/sampling_logp_difference/mean': 0.015610691159963608, 'step': 13, 'step_time': 157.10858551599085}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.234375, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3142.0, 'completions/mean_length': 2129.546875, 'completions/mean_terminated_length': 1527.5714111328125, 'completions/min_length': 465.0, 'completions/min_terminated_length': 465.0, 'entropy': 0.07177048269659281, 'epoch': 0.018229166666666668, 'frac_reward_zero_std': 0.625, 'grad_norm': 0.0021202724892646074, 'learning_rate': 1e-05, 'loss': -0.008725866675376892, 'num_tokens': 3508663.0, 'reward': 0.65625, 'reward/selected_zero_std_group_fraction': 0.30357142857142855, 'reward/zero_std_group_fraction': 0.6517857142857143, 'reward_components/channel_close_count': 0.765625, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.234375, 'reward_components/format_valid': 0.65625, 'reward_components/native_boundary_valid': 0.765625, 'reward_components/parser_json': 0.015625, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.75, 'reward_components/parser_unsafe': 0.109375, 'reward_components/parser_valid': 0.65625, 'reward_components/reward': 0.65625, 'reward_components/runs_correct': 0.890625, 'reward_components/runs_total': 1.375, 'reward_components/semantic_correct': 0.65625, 'reward_std': 0.4787135720252991, 'rewards/reward_func/mean': 0.65625, 'rewards/reward_func/std': 0.4787135720252991, 'sampling/importance_sampling_ratio/max': 2.031008720397949, 'sampling/importance_sampling_ratio/mean': 0.43108054995536804, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 13.054469108581543, 'sampling/sampling_logp_difference/mean': 0.014828605577349663, 'step': 14, 'step_time': 152.63235335214995}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.421875, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3681.0, 'completions/mean_length': 2637.625, 'completions/mean_terminated_length': 1573.4053955078125, 'completions/min_length': 687.0, 'completions/min_terminated_length': 687.0, 'entropy': 0.10809349920600653, 'epoch': 0.01953125, 'frac_reward_zero_std': 0.625, 'grad_norm': 0.01583734340965748, 'learning_rate': 1e-05, 'loss': -0.0400751456618309, 'num_tokens': 3735583.0, 'reward': 0.578125, 'reward/selected_zero_std_group_fraction': 0.3, 'reward/zero_std_group_fraction': 0.65, 'reward_components/channel_close_count': 0.578125, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.421875, 'reward_components/format_valid': 0.578125, 'reward_components/native_boundary_valid': 0.578125, 'reward_components/parser_json': 0.09375, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.484375, 'reward_components/parser_unsafe': 0.0, 'reward_components/parser_valid': 0.578125, 'reward_components/reward': 0.578125, 'reward_components/runs_correct': 0.703125, 'reward_components/runs_total': 1.5, 'reward_components/semantic_correct': 0.578125, 'reward_std': 0.49776285886764526, 'rewards/reward_func/mean': 0.578125, 'rewards/reward_func/std': 0.49776285886764526, 'sampling/importance_sampling_ratio/max': 2.6422102451324463, 'sampling/importance_sampling_ratio/mean': 0.6450210809707642, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 10.47492504119873, 'sampling/sampling_logp_difference/mean': 0.014680149964988232, 'step': 15, 'step_time': 158.13342921703588}, {'clip_ratio/high_max': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/region_mean': 0.0, 'completions/clipped_ratio': 0.265625, 'completions/max_length': 4096.0, 'completions/max_terminated_length': 3296.0, 'completions/mean_length': 2196.59375, 'completions/mean_terminated_length': 1509.574462890625, 'completions/min_length': 405.0, 'completions/min_terminated_length': 405.0, 'entropy': 0.0970615679398179, 'epoch': 0.020833333333333332, 'frac_reward_zero_std': 0.375, 'grad_norm': 0.0014741194900125265, 'learning_rate': 1e-05, 'loss': -0.00523255905136466, 'num_tokens': 3945221.0, 'reward': 0.609375, 'reward/selected_zero_std_group_fraction': 0.28125, 'reward/zero_std_group_fraction': 0.6328125, 'reward_components/channel_close_count': 0.734375, 'reward_components/channel_open_count': 1.0, 'reward_components/completion_truncated': 0.265625, 'reward_components/format_valid': 0.609375, 'reward_components/native_boundary_valid': 0.734375, 'reward_components/parser_json': 0.0, 'reward_components/parser_labelled_records': 0.0, 'reward_components/parser_natural': 0.734375, 'reward_components/parser_unsafe': 0.125, 'reward_components/parser_valid': 0.609375, 'reward_components/reward': 0.609375, 'reward_components/runs_correct': 0.90625, 'reward_components/runs_total': 1.375, 'reward_components/semantic_correct': 0.609375, 'reward_std': 0.4917473793029785, 'rewards/reward_func/mean': 0.609375, 'rewards/reward_func/std': 0.4917473793029785, 'sampling/importance_sampling_ratio/max': 1.4811298847198486, 'sampling/importance_sampling_ratio/mean': 0.35813361406326294, 'sampling/importance_sampling_ratio/min': 0.0, 'sampling/sampling_logp_difference/max': 4.191697597503662, 'sampling/sampling_logp_difference/mean': 0.01573760434985161, 'step': 16, 'step_time': 151.0304883638164}], best_metric=None, best_global_step=None, best_model_checkpoint=None, is_local_process_zero=True, is_world_process_zero=True, is_hyper_param_search=False, trial_name=None, trial_params=None, stateful_callbacks={'TrainerControl': {'args': {'should_epoch_stop': False, 'should_evaluate': False, 'should_log': False, 'should_save': True, 'should_training_stop': True}, 'attributes': {}}})",
+ "note": "completion_ids removed; regenerate by tokenising completion_raw_text with the gemma4-26b-a4b tokenizer (verified exact on 221 sampled records). log_extra/log_metric were bound-method reprs and held no data."
+}