1 line
11 KiB
JSON
1 line
11 KiB
JSON
{"rollout_corr/log_ppl_diff_min":-0.014117404818534851,"val-aux/sciknoweval/reward/maj@16/mean":0.7674333333333333,"self_distillation/correct_sample_fraction":0.6640625,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.00755187697793512,"timing_s/old_log_prob":2.3089759554713964,"rollout_corr/chi2_token":0.0008595744147896767,"response_length/min":29,"val-aux/sciknoweval/acc/worst@2/mean":0.7324333333333333,"training/epoch":1,"perf/max_memory_reserved_gb":143.0078125,"timing_per_token_ms/adv":0.004661559897271101,"response_length/mean":61.05078125,"self_distillation/empty_target_batch":0.1640625,"val-aux/sciknoweval/score/maj@8/std":0.04659246363692196,"response_length_non_aborted/clip_ratio":0,"val-aux/sciknoweval/reward/best@8/mean":0.8498714285714286,"val-aux/sciknoweval/score/worst@16/std":0.022450981651084227,"val-core/sciknoweval/acc/best@16/std":0.022818983837850748,"val-aux/sciknoweval/score/worst@2/std":0.06710901416436343,"val-aux/sciknoweval/reward/maj@4/std":0.06469916434542414,"prompt_length/min":165,"val-aux/sciknoweval/reward/worst@16/std":0.022450981651084227,"timing_s/agent_loop/slowest/generate_sequences":1.1843877080827951,"timing_s/agent_loop/slowest/prompt_length":359,"val-aux/sciknoweval/incorrect_format/best@8/mean":1,"val-aux/sciknoweval/reward/maj@8/std":0.04659246363692196,"rollout_corr/rollout_is_max":1.6167027950286865,"rollout_corr/rollout_is_mean":1.0001804828643799,"val-aux/sciknoweval/reward/best@16/mean":0.8647904761904761,"val-aux/sciknoweval/score/worst@8/mean":0.6884666666666667,"prompt_length/mean":296.15625,"srpo/correct_sample_fraction":0.6640625,"perf/max_memory_allocated_gb":138.39261770248413,"val-aux/sciknoweval/reward/maj@8/mean":0.7679904761904761,"val-core/sciknoweval/acc/best@16/mean":0.8647904761904761,"val-aux/sciknoweval/reward/best@4/mean":0.828,"val-aux/sciknoweval/acc/best@2/mean":0.7995714285714286,"val-aux/sciknoweval/acc/std@16":0.08180825117524429,"val-aux/sciknoweval/reward/best@8/std":0.03842860858872272,"response_length_non_aborted/mean":61.05078125,"val-aux/sciknoweval/acc/best@8/std":0.03842860858872272,"prompt_length/clip_ratio":0,"timing_s/agent_loop/generate_sequences/min":0.32152901589870453,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0016347712194112146,"srpo/grpo_sample_fraction":0.8203125,"val-aux/sciknoweval/reward/best@2/std":0.07120304368358653,"critic/advantages/min":-0.75,"_step":100,"val-aux/sciknoweval/reward/maj@2/std":0.08209907766512188,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"self_distillation/reprompt_sample_fraction":0.8359375,"val-aux/sciknoweval/reward/worst@8/mean":0.6884666666666667,"val-aux/sciknoweval/score/best@8/mean":0.8498714285714286,"val-aux/sciknoweval/reward/best@16/std":0.022818983837850748,"global_seqlen/min":9718,"critic/score/max":1,"val-aux/sciknoweval/score/maj@2/mean":0.7658571428571429,"actor/entropy":0.07664129883050919,"val-aux/sciknoweval/score/worst@4/mean":0.7069809523809524,"global_seqlen/max":14844,"training/global_step":100,"critic/rewards/min":0,"timing_per_token_ms/gen":0.4833862506069128,"val-aux/num_turns/mean":2,"response_length_non_aborted/max":123,"timing_s/adv":0.4262763448059559,"val-aux/sciknoweval/score/maj@16/mean":0.7674333333333333,"val-aux/sciknoweval/score/worst@4/std":0.04900531157529151,"srpo/routed_token_count":61.05078125,"val-aux/sciknoweval/score/maj@4/mean":0.767504761904762,"val-aux/sciknoweval/acc/worst@4/std":0.04900531157529151,"val-aux/sciknoweval/score/std@16":0.08180825117524429,"actor/pg_loss":-0.056587886293527845,"critic/rewards/max":1,"val-aux/sciknoweval/acc/maj@4/mean":0.767504761904762,"val-aux/sciknoweval/incorrect_format/std@16":0.005316651201494241,"val-aux/sciknoweval/acc/best@2/std":0.07120304368358653,"rollout_corr/rollout_is_seq_std":0.003016919596120715,"srpo/sdpo_loss":0.006881217535010364,"timing_s/agent_loop/slowest/tool_calls":0,"val-aux/sciknoweval/incorrect_format/maj@16/std":0.0002604287412166727,"rollout_corr/chi2_seq":0.05469026346690953,"critic/score/min":0,"rollout_corr/k3_kl":0.00035141411022021965,"global_seqlen/balanced_max":11431,"srpo/grpo_actor/ppo_kl":0,"self_distillation/success_sample_fraction":0.8359375,"timing_s/step":26.25247493945062,"self_distillation/success_group_fraction":0.84375,"timing_s/agent_loop/tool_calls/mean":0,"timing_s/agent_loop/tool_calls/max":0,"val-aux/sciknoweval/reward/best@4/std":0.05458358345213735,"val-aux/sciknoweval/acc/worst@8/std":0.034267210500084945,"val-core/sciknoweval/acc/maj@16/mean":0.7674333333333333,"rollout_corr/rollout_is_eff_sample_size":0.9992250254135918,"timing_per_token_ms/update_actor":0.17307132634888966,"val-aux/sciknoweval/incorrect_format/best@4/std":0.0002127457895589398,"timing_s/agent_loop/slowest/response_length":121,"training/rollout_probs_diff_mean":0.0021808038000017405,"critic/rewards/mean":0.6640625,"val-aux/sciknoweval/reward/worst@2/mean":0.7324333333333333,"val-core/sciknoweval/acc/maj@16/std":0.033146197360321726,"val-aux/sciknoweval/reward/maj@16/std":0.033146197360321726,"val-aux/sciknoweval/acc/maj@2/std":0.08209907766512188,"global_seqlen/minmax_diff":5126,"timing_s/agent_loop/tool_calls/min":0,"rollout_corr/training_log_ppl":0.07520282005225454,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.008456497157340367,"val-aux/sciknoweval/acc/maj@8/std":0.04659246363692196,"critic/advantages/mean":0.0006318382220342755,"rollout_corr/ppl_ratio":1.000278169522062,"val-aux/sciknoweval/reward/worst@8/std":0.034267210500084945,"rollout_corr/rollout_is_seq_max_deviation":0.015818357467651367,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.9860095238095239,"val-aux/sciknoweval/incorrect_format/mean@16":0.9982142857142857,"self_distillation/dynamic_weight_mean":0.17968750419095159,"actor/lr":5e-06,"_runtime":4328.501085711,"val-aux/sciknoweval/acc/maj@8/mean":0.7679904761904761,"val-aux/sciknoweval/acc/maj@2/mean":0.7658571428571429,"val-aux/sciknoweval/reward/worst@4/std":0.04900531157529151,"val-aux/sciknoweval/score/best@8/std":0.03842860858872272,"val-aux/sciknoweval/score/maj@4/std":0.06469916434542414,"val-aux/sciknoweval/score/best@16/std":0.022818983837850748,"critic/returns/max":0.875,"timing_s/testing":97.83039686828852,"val-aux/sciknoweval/reward/std@16":0.08180825117524429,"val-aux/sciknoweval/acc/worst@16/mean":0.6746142857142856,"rollout_corr/rollout_log_ppl":0.07492950143864618,"val-aux/sciknoweval/incorrect_format/best@8/std":0,"srpo/grpo_actor/pg_clipfrac_lower":0,"timing_s/reward":0.04815688915550709,"val-aux/sciknoweval/reward/worst@16/mean":0.6746142857142856,"val-core/sciknoweval/acc/mean@16":0.7660714285714286,"critic/advantages/max":0.875,"training/rollout_actor_probs_pearson_corr":0.9975420832633972,"timing_s/agent_loop/generate_sequences/mean":0.689643530837202,"val-aux/sciknoweval/acc/maj@4/std":0.06469916434542414,"val-aux/sciknoweval/score/worst@8/std":0.034267210500084945,"val-aux/sciknoweval/score/worst@16/mean":0.6746142857142856,"val-aux/sciknoweval/acc/worst@4/mean":0.7069809523809524,"timing_s/start_profile":7.328763604164124e-05,"rollout_corr/log_ppl_abs_diff":0.001995735538230292,"rollout_corr/log_ppl_diff_max":0.012438356876373291,"srpo/sdpo_sample_fraction":0.1796875,"self_distillation/dynamic_weighting_enabled":1,"val-aux/sciknoweval/acc/best@8/mean":0.8498714285714286,"val-aux/sciknoweval/incorrect_format/maj@8/std":0.0015690976922436521,"actor/kl_loss":0,"rollout_corr/rollout_is_std":0.027851298451423645,"training/rollout_probs_diff_std":0.01022244431078434,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.9967047619047619,"training/rollout_probs_diff_max":0.41022980213165283,"rollout_corr/rollout_ppl":1.08108503324911,"val-aux/sciknoweval/score/best@4/mean":0.828,"val-aux/sciknoweval/score/best@16/mean":0.8647904761904761,"val-aux/sciknoweval/incorrect_format/maj@8/mean":0.999804761904762,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.9940714285714285,"val-aux/sciknoweval/acc/worst@2/std":0.06710901416436343,"perf/mfu/actor_infer":0,"num_turns/min":2,"val-aux/sciknoweval/score/mean@16":0.7660714285714286,"prompt_length/max":625,"val-aux/sciknoweval/score/maj@16/std":0.033146197360321726,"rollout_corr/rollout_is_seq_fraction_high":0,"rollout_corr/rollout_is_seq_fraction_low":0,"rollout_corr/training_ppl":1.081399030983448,"perf/cpu_memory_used_gb":149.58447647094727,"rollout_corr/rollout_is_seq_mean":1.0000782012939453,"val-aux/sciknoweval/score/best@4/std":0.05458358345213735,"timing_s/update_actor":15.826507437974215,"srpo/routed_loss":-0.056587886293527845,"num_turns/mean":2,"val-aux/sciknoweval/reward/mean@16":0.7660714285714286,"val-aux/num_turns/max":2,"global_seqlen/mean":11430.625,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.0029882373225892547,"timing_s/agent_loop/generate_sequences/max":1.1843877080827951,"rollout_corr/log_ppl_diff":0.0002733186136083532,"val-aux/sciknoweval/score/maj@2/std":0.08209907766512188,"rollout_corr/rollout_is_min":0.43883800506591797,"val-aux/sciknoweval/acc/worst@8/mean":0.6884666666666667,"timing_s/gen":7.55484371073544,"val-aux/sciknoweval/incorrect_format/best@4/mean":0.9999904761904761,"val-aux/sciknoweval/reward/worst@4/mean":0.7069809523809524,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.006964443486271992,"self_distillation/dynamic_weight_max":1.1291440725326538,"val-aux/sciknoweval/score/best@2/std":0.07120304368358653,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"val-aux/sciknoweval/reward/maj@4/mean":0.767504761904762,"num_turns/max":2,"perf/total_num_tokens":91445,"val-aux/sciknoweval/reward/maj@2/mean":0.7658571428571429,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.005210571230613174,"response_length/max":123,"global_seqlen/balanced_min":11430,"rollout_corr/rollout_is_ratio_fraction_high":0,"actor/grad_norm":0.2391292154788971,"rollout_corr/rollout_is_seq_max":1.0158183574676514,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.9983523809523809,"timing_s/save_checkpoint":6.585983132943511,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.9998142857142858,"val-aux/sciknoweval/score/best@2/mean":0.7995714285714286,"val-aux/sciknoweval/reward/best@2/mean":0.7995714285714286,"self_distillation/feedback_used_fraction":0,"critic/returns/min":-0.75,"val-aux/sciknoweval/incorrect_format/maj@16/mean":0.9999857142857144,"val-aux/sciknoweval/acc/worst@16/std":0.022450981651084227,"self_distillation/feedback_available_fraction":0,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.008910847165478613,"response_length/clip_ratio":0,"rollout_corr/rollout_is_ratio_fraction_low":6.3983621657826e-05,"rollout_corr/kl":0.000273317242802662,"_wandb":{"runtime":4328},"critic/score/mean":0.6640625,"val-aux/num_turns/min":2,"perf/mfu/actor":0.020568645471119033,"val-aux/sciknoweval/acc/best@4/std":0.05458358345213735,"rollout_corr/rollout_is_seq_min":0.9890013337135315,"val-aux/sciknoweval/score/worst@2/mean":0.7324333333333333,"response_length_non_aborted/min":29,"perf/time_per_step":26.25247493945062,"training/rollout_probs_diff_valid":1,"critic/returns/mean":0.0006318382220342755,"val-aux/sciknoweval/acc/best@4/mean":0.828,"srpo/grpo_actor/pg_clipfrac":0,"_timestamp":1.7829782133912632e+09,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.9903190476190475,"response/aborted_ratio":0,"perf/throughput":435.4113288885671,"val-aux/sciknoweval/score/maj@8/mean":0.7679904761904761,"timing_s/stop_profile":0.00013065896928310394,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.9994285714285714,"val-aux/sciknoweval/reward/worst@2/std":0.06710901416436343,"self_distillation/dynamic_weight_entropy_mean":0.007747210298703067,"srpo/grpo_loss":-0.06346910382853821} |