1 line
11 KiB
JSON
1 line
11 KiB
JSON
{"self_distillation/token_reweight_context_sample_fraction":0.71875,"actor/kl_loss":0,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.9997,"training/global_step":100,"prompt_length/min":165,"num_turns/max":2,"val-aux/sciknoweval/reward/maj@2/std":0.10956288078402857,"rollout_corr/chi2_seq":1.0585334873758256,"response_length/max":616,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.0020220104047945714,"rollout_corr/rollout_is_seq_std":0.002407139865681529,"global_seqlen/balanced_max":19622,"val-aux/sciknoweval/score/std@16":0.10973085235155812,"val-aux/sciknoweval/score/maj@4/mean":0.6991761904761905,"rollout_corr/rollout_ppl":1.2673596912063658,"val-aux/sciknoweval/reward/best@4/std":0.05746985351073585,"rollout_corr/kl":0.0009287315268036167,"val-aux/sciknoweval/reward/std@16":0.10973085235155812,"self_distillation/token_reweight_w_std":1.1881619467263469e+06,"rollout_corr/rollout_is_seq_fraction_low":0,"timing_s/agent_loop/generate_sequences/max":5.5460103284567595,"val-aux/sciknoweval/acc/best@2/std":0.08320072408312382,"self_distillation/token_reweight_active_token_fraction":0.71875,"val-aux/sciknoweval/reward/maj@8/std":0.05779624954713381,"val-aux/sciknoweval/score/worst@2/mean":0.6464904761904762,"critic/rewards/mean":0.6171875,"val-aux/num_turns/max":2,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"val-aux/sciknoweval/score/maj@2/mean":0.6907285714285715,"val-aux/sciknoweval/incorrect_format/best@4/mean":1,"self_distillation/success_sample_fraction":0.71875,"response_length_non_aborted/max":616,"perf/cpu_memory_used_gb":146.29161834716797,"response_length_non_aborted/mean":317.0234375,"training/rollout_probs_diff_mean":0.004920980893075466,"timing_s/adv":0.4910304043442011,"val-core/sciknoweval/acc/mean@16":0.6895833333333333,"val-aux/sciknoweval/acc/maj@4/mean":0.6991761904761905,"val-aux/sciknoweval/acc/maj@2/std":0.10956288078402857,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.9999476190476191,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.9967095238095238,"critic/advantages/mean":-0.0011474530911073089,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.0004966783087710074,"val-core/sciknoweval/acc/best@16/mean":0.8013714285714286,"val-aux/sciknoweval/acc/worst@2/std":0.10098054289944847,"val-aux/sciknoweval/score/worst@16/mean":0.5435285714285715,"val-aux/sciknoweval/acc/worst@2/mean":0.6464904761904762,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"val-aux/num_turns/mean":2,"rollout_corr/log_ppl_diff":0.0009287331777159125,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.0011569664768572288,"val-aux/sciknoweval/acc/maj@8/mean":0.7047571428571429,"val-aux/sciknoweval/reward/maj@16/std":0.03834834240388913,"prompt_length/mean":296.15625,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0002127457895589398,"val-aux/sciknoweval/reward/best@8/mean":0.7858285714285714,"val-aux/sciknoweval/incorrect_format/maj@16/mean":1,"val-aux/sciknoweval/score/maj@8/mean":0.7047571428571429,"critic/advantages/max":0.75,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.9994476190476191,"val-aux/sciknoweval/score/best@8/mean":0.7858285714285714,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.9980238095238095,"training/epoch":1,"training/rollout_probs_diff_max":0.30163663625717163,"val-aux/sciknoweval/acc/worst@4/std":0.08537783027231041,"critic/score/mean":0.6171875,"val-aux/sciknoweval/reward/worst@2/std":0.10098054289944847,"val-aux/sciknoweval/score/worst@8/std":0.06319778055664713,"rollout_corr/rollout_is_seq_mean":1.0000463724136353,"val-aux/sciknoweval/score/best@16/std":0.02391792727193744,"timing_s/update_actor":15.488128297030926,"val-aux/sciknoweval/acc/maj@4/std":0.08403267534041585,"rollout_corr/training_log_ppl":0.23310650100756902,"val-aux/sciknoweval/acc/best@4/mean":0.7644238095238095,"rollout_corr/rollout_is_ratio_fraction_high":1.2321644135226961e-05,"val-aux/sciknoweval/reward/maj@8/mean":0.7047571428571429,"val-aux/sciknoweval/acc/worst@8/mean":0.5691952380952381,"num_turns/mean":2,"val-aux/sciknoweval/reward/best@8/std":0.038120394142420314,"val-aux/sciknoweval/score/maj@16/mean":0.7074904761904762,"val-aux/sciknoweval/score/best@4/std":0.05746985351073585,"rollout_corr/rollout_is_eff_sample_size":0.9983795954951445,"actor/ppo_kl":0.00010989303795483352,"self_distillation/reprompt_sample_fraction":0.71875,"val-aux/sciknoweval/score/worst@8/mean":0.5691952380952381,"val-aux/sciknoweval/acc/std@16":0.10973085235155812,"num_turns/min":2,"val-aux/sciknoweval/score/best@8/std":0.038120394142420314,"timing_s/agent_loop/tool_calls/max":0,"val-aux/sciknoweval/acc/worst@4/mean":0.6048666666666667,"val-aux/sciknoweval/score/maj@16/std":0.03834834240388913,"_timestamp":1.7829775262133236e+09,"response/aborted_ratio":0,"global_seqlen/balanced_min":19621,"self_distillation/feedback_used_fraction":0,"rollout_corr/log_ppl_abs_diff":0.002280793822137639,"val-aux/sciknoweval/acc/best@8/mean":0.7858285714285714,"val-aux/sciknoweval/reward/best@2/mean":0.733895238095238,"rollout_corr/rollout_is_mean":1.0000070333480835,"rollout_corr/rollout_is_seq_fraction_high":0,"response_length_non_aborted/min":122,"val-aux/sciknoweval/incorrect_format/maj@8/mean":1,"val-aux/sciknoweval/score/worst@4/mean":0.6048666666666667,"self_distillation/token_reweight_w_clipped_mean":1.000056762015447,"perf/total_num_tokens":156974,"training/rollout_probs_diff_valid":1,"val-aux/sciknoweval/score/maj@8/std":0.05779624954713381,"response_length/mean":317.0234375,"self_distillation/feedback_available_fraction":0,"val-aux/sciknoweval/score/worst@2/std":0.10098054289944847,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.9999904761904761,"critic/rewards/min":0,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.9988761904761905,"prompt_length/clip_ratio":0,"training/rollout_probs_diff_std":0.013346283696591854,"response_length/min":122,"self_distillation/token_reweight_w_mean":73094.00544022233,"critic/advantages/min":-0.875,"rollout_corr/rollout_is_max":2.1037662029266357,"perf/mfu/actor_infer":0,"_runtime":4270.568982177,"val-core/sciknoweval/acc/best@16/std":0.02391792727193744,"timing_s/agent_loop/slowest/tool_calls":0,"val-aux/sciknoweval/incorrect_format/maj@16/std":0,"val-aux/sciknoweval/reward/worst@2/mean":0.6464904761904762,"global_seqlen/mean":19621.75,"val-aux/sciknoweval/reward/mean@16":0.6895833333333333,"self_distillation/token_reweight_lambda":0.5,"actor/grad_norm":0.25844827108085155,"timing_s/gen":11.436559945344925,"val-aux/sciknoweval/reward/worst@16/mean":0.5435285714285715,"val-aux/sciknoweval/score/best@4/mean":0.7644238095238095,"actor/lr":1e-06,"val-aux/sciknoweval/acc/best@4/std":0.05746985351073585,"self_distillation/empty_target_batch":0.28125,"actor/entropy":0.21838146448135376,"val-aux/sciknoweval/acc/maj@8/std":0.05779624954713381,"rollout_corr/rollout_is_ratio_fraction_low":6.16082179476507e-05,"val-aux/sciknoweval/reward/maj@4/std":0.08403267534041585,"val-aux/sciknoweval/reward/worst@4/mean":0.6048666666666667,"actor/pg_clipfrac":0.0004233177678543143,"val-aux/sciknoweval/reward/best@2/std":0.08320072408312382,"global_seqlen/max":22010,"val-core/sciknoweval/acc/maj@16/mean":0.7074904761904762,"actor/pg_loss":0.0023272527614608407,"val-aux/sciknoweval/acc/best@8/std":0.038120394142420314,"val-aux/sciknoweval/acc/worst@8/std":0.06319778055664713,"self_distillation/token_reweight_correct_sample_fraction":0.6171875,"critic/returns/max":0.75,"timing_s/agent_loop/generate_sequences/mean":3.556732240387646,"val-aux/sciknoweval/score/best@16/mean":0.8013714285714286,"timing_per_token_ms/update_actor":0.0986668384384097,"val-aux/sciknoweval/incorrect_format/maj@8/std":0,"val-aux/sciknoweval/acc/maj@2/mean":0.6907285714285715,"rollout_corr/rollout_is_seq_min":0.9934689998626709,"timing_s/agent_loop/slowest/prompt_length":232,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"response_length_non_aborted/clip_ratio":0,"val-aux/sciknoweval/reward/worst@8/std":0.06319778055664713,"val-core/sciknoweval/acc/maj@16/std":0.03834834240388913,"val-aux/sciknoweval/score/worst@16/std":0.03927314998986781,"val-aux/sciknoweval/incorrect_format/mean@16":0.999702380952381,"timing_s/reward":0.05615238472819328,"val-aux/sciknoweval/reward/best@16/mean":0.8013714285714286,"val-aux/sciknoweval/incorrect_format/best@8/std":0,"timing_s/agent_loop/tool_calls/mean":0,"actor/pg_clipfrac_lower":0,"val-aux/sciknoweval/score/best@2/mean":0.733895238095238,"critic/returns/mean":-0.0011474530911073089,"critic/returns/min":-0.875,"val-aux/sciknoweval/score/maj@2/std":0.10956288078402857,"val-aux/sciknoweval/acc/worst@16/mean":0.5435285714285715,"val-aux/sciknoweval/incorrect_format/best@8/mean":1,"global_seqlen/minmax_diff":4940,"_step":100,"val-aux/sciknoweval/acc/worst@16/std":0.03927314998986781,"perf/max_memory_reserved_gb":131.27734375,"val-aux/sciknoweval/score/best@2/std":0.08320072408312382,"rollout_corr/ppl_ratio":1.0009334608912468,"self_distillation/success_group_fraction":0.71875,"perf/time_per_step":29.756166722625494,"val-aux/sciknoweval/score/worst@4/std":0.08537783027231041,"rollout_corr/rollout_is_std":0.04028838127851486,"val-aux/sciknoweval/acc/best@2/mean":0.733895238095238,"val-aux/num_turns/min":2,"_wandb":{"runtime":4270},"timing_per_token_ms/gen":0.14091722252082264,"timing_s/old_log_prob":2.225490480661392,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.002346295386522406,"timing_s/save_checkpoint":7.179236024618149,"val-aux/sciknoweval/score/mean@16":0.6895833333333333,"timing_s/start_profile":0.00018987245857715607,"val-aux/sciknoweval/reward/best@16/std":0.02391792727193744,"timing_s/stop_profile":0.00012516416609287262,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.0022003864842959084,"perf/throughput":659.4179345379304,"critic/score/min":0,"val-aux/sciknoweval/reward/maj@2/mean":0.6907285714285715,"rollout_corr/log_ppl_diff_max":0.015468984842300415,"critic/score/max":1,"val-aux/sciknoweval/reward/worst@16/std":0.03927314998986781,"global_seqlen/min":17070,"perf/mfu/actor":0.03632957727570209,"self_distillation/token_reweight_w_clip_frac":0.03758277394808829,"critic/rewards/max":1,"val-aux/sciknoweval/reward/maj@16/mean":0.7074904761904762,"val-aux/sciknoweval/reward/worst@4/std":0.08537783027231041,"val-aux/sciknoweval/reward/best@4/mean":0.7644238095238095,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.0015248805759714617,"self_distillation/token_reweight_eps_w":0.20000000000000004,"val-aux/sciknoweval/reward/worst@8/mean":0.5691952380952381,"timing_s/agent_loop/slowest/generate_sequences":5.5460103284567595,"timing_per_token_ms/adv":0.003128100222611395,"val-aux/sciknoweval/reward/maj@4/mean":0.6991761904761905,"response_length/clip_ratio":0,"timing_s/step":29.756166722625494,"rollout_corr/log_ppl_diff_min":-0.010868668556213379,"prompt_length/max":625,"training/rollout_actor_probs_pearson_corr":0.9981935024261475,"rollout_corr/rollout_log_ppl":0.2321777678298531,"timing_s/testing":108.75369282439351,"rollout_corr/rollout_is_min":0.39213550090789795,"timing_s/agent_loop/slowest/response_length":602,"rollout_corr/k3_kl":0.001084987000297133,"timing_s/agent_loop/generate_sequences/min":1.5517499037086964,"timing_s/agent_loop/tool_calls/min":0,"val-aux/sciknoweval/incorrect_format/std@16":0.0011526736149426837,"rollout_corr/chi2_token":0.002497109118849039,"val-aux/sciknoweval/score/maj@4/std":0.08403267534041585,"perf/max_memory_allocated_gb":124.80831956863403,"rollout_corr/training_ppl":1.2685748748481274,"self_distillation/correct_sample_fraction":0.6171875,"rollout_corr/rollout_is_seq_max":1.0090001821517944,"rollout_corr/rollout_is_seq_max_deviation":0.009000182151794434} |