Files
Qwen3-8B-Biology-RLSD-TR/artifacts/wandb-summary.json
ModelHub XC 68be06f5db 初始化项目,由ModelHub XC社区提供模型
Model: SeongryongJung/Qwen3-8B-Biology-RLSD-TR
Source: Original Platform
2026-08-18 21:36:57 +08:00

1 line
11 KiB
JSON

{"rollout_corr/ppl_ratio":1.0011649164371192,"val-aux/sciknoweval/acc/std@16":0.1705123962141389,"timing_s/agent_loop/slowest/tool_calls":0,"val-aux/sciknoweval/score/maj@4/mean":0.55496,"val-aux/sciknoweval/reward/worst@8/std":0.08117633716216308,"val-aux/sciknoweval/reward/best@8/std":0.06916601903650618,"val-aux/sciknoweval/reward/maj@2/mean":0.55058,"val-aux/sciknoweval/acc/maj@8/mean":0.5593400000000001,"val-aux/sciknoweval/score/best@4/mean":0.674,"critic/score/max":1,"_runtime":5519.500528115,"global_seqlen/min":15201,"rollout_corr/rollout_is_seq_fraction_high":0,"val-aux/sciknoweval/incorrect_format/mean@16":0.99875,"timing_s/agent_loop/tool_calls/mean":0,"rollout_corr/kl":0.0011583137813406807,"timing_s/reward":0.05930858105421066,"val-aux/sciknoweval/reward/best@4/mean":0.674,"val-core/sciknoweval/acc/best@16/mean":0.73856,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.008311822904754409,"rollout_corr/rollout_is_seq_max_deviation":0.008629322052001953,"rollout_corr/chi2_token":0.003004896454513073,"timing_s/gen":18.558317134156823,"val-aux/sciknoweval/acc/worst@2/std":0.147132195408043,"val-aux/sciknoweval/reward/maj@16/mean":0.55908,"timing_s/agent_loop/slowest/generate_sequences":11.257305985316634,"val-core/sciknoweval/acc/maj@16/std":0.06533731091763299,"timing_s/agent_loop/generate_sequences/min":1.9916328880935907,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.9977199999999999,"val-aux/sciknoweval/incorrect_format/std@16":0.004841229182759271,"val-aux/sciknoweval/score/best@4/std":0.09758132520114339,"self_distillation/token_reweight_context_sample_fraction":0.83984375,"val-aux/sciknoweval/reward/maj@16/std":0.06533731091763299,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.9875,"timing_s/agent_loop/tool_calls/min":0,"val-aux/sciknoweval/acc/maj@2/mean":0.55058,"rollout_corr/rollout_is_seq_std":0.0027923036832362413,"val-aux/num_turns/min":2,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.009763933633531108,"timing_s/agent_loop/generate_sequences/max":11.257305985316634,"self_distillation/token_reweight_w_clip_frac":0.13672411735751666,"val-aux/sciknoweval/incorrect_format/maj@8/std":0,"val-aux/sciknoweval/reward/worst@16/std":0.048294469414788314,"val-aux/sciknoweval/score/mean@16":0.55,"_step":100,"val-aux/sciknoweval/score/worst@16/std":0.048294469414788314,"val-aux/sciknoweval/score/maj@4/std":0.1354735155461244,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0008935323161475471,"actor/pg_loss":0.007820420083589852,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"critic/returns/min":-0.875,"val-core/sciknoweval/acc/maj@16/mean":0.55908,"_wandb":{"runtime":5519},"self_distillation/token_reweight_lambda":0.5,"actor/kl_loss":0,"val-aux/sciknoweval/reward/best@16/std":0.04218677040162369,"training/rollout_probs_diff_mean":0.005966469645500183,"response_length_non_aborted/clip_ratio":0,"training/rollout_actor_probs_pearson_corr":0.9986874461174011,"val-aux/sciknoweval/acc/best@2/mean":0.61904,"global_seqlen/balanced_min":19240,"rollout_corr/rollout_log_ppl":0.3697861341643147,"num_turns/max":2,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.99972,"timing_s/testing":78.02890569157898,"rollout_corr/rollout_is_min":0.3131374716758728,"prompt_length/min":180,"timing_s/adv":0.669283602386713,"val-aux/sciknoweval/reward/maj@4/mean":0.55496,"val-aux/sciknoweval/reward/worst@8/mean":0.3761,"val-aux/sciknoweval/reward/worst@4/std":0.11484012157584442,"val-aux/sciknoweval/acc/best@2/std":0.1400715293582534,"val-aux/sciknoweval/incorrect_format/maj@16/mean":1,"val-core/sciknoweval/acc/best@16/std":0.04218677040162369,"self_distillation/correct_sample_fraction":0.671875,"val-aux/sciknoweval/acc/maj@8/std":0.09364957204441302,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.9999600000000001,"val-aux/sciknoweval/acc/best@8/mean":0.71124,"val-aux/sciknoweval/score/worst@4/mean":0.4220399999999999,"val-aux/sciknoweval/score/worst@2/mean":0.48004,"val-aux/sciknoweval/reward/worst@16/mean":0.3436,"self_distillation/success_sample_fraction":0.83984375,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"actor/ppo_kl":7.754255021552581e-05,"self_distillation/feedback_available_fraction":0,"rollout_corr/chi2_seq":0.8863912406377494,"actor/pg_clipfrac":0.0010761070025182562,"val-aux/sciknoweval/reward/maj@4/std":0.1354735155461244,"actor/lr":1e-06,"perf/max_memory_allocated_gb":135.0777406692505,"timing_per_token_ms/update_actor":0.13609144157108288,"_timestamp":1.783075772740568e+09,"val-aux/sciknoweval/acc/maj@4/mean":0.55496,"val-aux/sciknoweval/acc/worst@2/mean":0.48004,"timing_s/agent_loop/slowest/prompt_length":427,"val-aux/sciknoweval/score/std@16":0.1705123962141389,"val-core/sciknoweval/acc/mean@16":0.55,"val-aux/sciknoweval/incorrect_format/best@8/std":0,"val-aux/sciknoweval/score/maj@2/mean":0.55058,"rollout_corr/k3_kl":0.0012635472568831574,"val-aux/sciknoweval/score/worst@4/std":0.11484012157584442,"val-aux/sciknoweval/reward/maj@8/std":0.09364957204441302,"prompt_length/clip_ratio":0,"num_turns/min":2,"timing_s/stop_profile":0.00021155737340450287,"val-aux/sciknoweval/score/maj@8/mean":0.5593400000000001,"critic/advantages/mean":-0.005199081264436245,"timing_s/start_profile":0.00020011886954307556,"timing_s/step":43.38669454306364,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.009682458365518542,"val-aux/sciknoweval/reward/worst@2/std":0.147132195408043,"val-aux/sciknoweval/reward/worst@2/mean":0.48004,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.00635622529493724,"num_turns/mean":2,"response_length/mean":255.078125,"val-aux/sciknoweval/score/worst@8/mean":0.3761,"val-aux/sciknoweval/reward/best@2/std":0.1400715293582534,"val-aux/sciknoweval/incorrect_format/best@4/mean":1,"val-aux/sciknoweval/acc/worst@16/std":0.048294469414788314,"rollout_corr/rollout_ppl":1.4546512076631188,"self_distillation/token_reweight_w_std":679798.1475089006,"timing_per_token_ms/gen":0.2842008749488028,"val-aux/sciknoweval/acc/best@4/mean":0.674,"perf/mfu/actor":0.04707397319538395,"rollout_corr/rollout_is_max":2.4048268795013428,"critic/rewards/min":0,"perf/throughput":446.20822590632366,"training/global_step":100,"rollout_corr/rollout_is_ratio_fraction_low":1.53139353642473e-05,"global_seqlen/balanced_max":20182,"val-aux/sciknoweval/acc/worst@16/mean":0.3436,"val-aux/sciknoweval/incorrect_format/best@8/mean":1,"critic/returns/mean":-0.005199081264436245,"val-aux/sciknoweval/acc/maj@4/std":0.1354735155461244,"timing_s/agent_loop/slowest/response_length":1625,"rollout_corr/rollout_is_std":0.042716238647699356,"rollout_corr/rollout_is_ratio_fraction_high":4.59418079117313e-05,"self_distillation/token_reweight_w_mean":45580.425170551986,"val-aux/num_turns/mean":2,"val-aux/sciknoweval/score/best@8/mean":0.71124,"rollout_corr/log_ppl_abs_diff":0.002755315537797287,"val-aux/sciknoweval/incorrect_format/maj@16/std":0,"prompt_length/mean":349.90625,"val-aux/sciknoweval/score/maj@2/std":0.16941156579650474,"critic/rewards/mean":0.671875,"critic/advantages/max":0.875,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.99894,"val-aux/sciknoweval/reward/std@16":0.1705123962141389,"rollout_corr/training_ppl":1.456413078121841,"val-aux/sciknoweval/reward/mean@16":0.55,"rollout_corr/rollout_is_seq_max":1.008629322052002,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.0023498085028359223,"rollout_corr/log_ppl_diff_max":0.010634154081344604,"rollout_corr/rollout_is_mean":0.9998152852058411,"self_distillation/reprompt_sample_fraction":0.83984375,"val-aux/sciknoweval/acc/best@8/std":0.06916601903650618,"global_seqlen/mean":19359.5,"perf/mfu/actor_infer":0,"self_distillation/token_reweight_eps_w":0.20000000000000004,"actor/pg_clipfrac_lower":0,"response_length_non_aborted/min":150,"val-aux/sciknoweval/score/best@2/mean":0.61904,"val-aux/sciknoweval/reward/maj@8/mean":0.5593400000000001,"val-aux/sciknoweval/acc/worst@8/std":0.08117633716216308,"perf/cpu_memory_used_gb":205.93875122070312,"perf/max_memory_reserved_gb":149.6015625,"perf/total_num_tokens":154876,"val-aux/sciknoweval/score/worst@2/std":0.147132195408043,"val-aux/sciknoweval/acc/best@4/std":0.09758132520114339,"val-aux/sciknoweval/score/worst@8/std":0.08117633716216308,"global_seqlen/minmax_diff":7394,"val-aux/sciknoweval/reward/best@2/mean":0.61904,"training/rollout_probs_diff_valid":1,"rollout_corr/log_ppl_diff":0.0011583174637053162,"self_distillation/token_reweight_w_clipped_mean":0.9963978114537895,"critic/score/min":0,"global_seqlen/max":22595,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.99216,"val-aux/num_turns/max":2,"response_length_non_aborted/mean":255.078125,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.99556,"val-aux/sciknoweval/acc/worst@4/mean":0.4220399999999999,"perf/time_per_step":43.38669454306364,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"timing_s/old_log_prob":2.932212023064494,"val-aux/sciknoweval/acc/worst@8/mean":0.3761,"response_length_non_aborted/max":1625,"timing_s/agent_loop/tool_calls/max":0,"self_distillation/token_reweight_active_token_fraction":0.83984375,"critic/score/mean":0.671875,"training/epoch":7,"actor/entropy":0.3639138340950012,"prompt_length/max":499,"response/aborted_ratio":0,"rollout_corr/rollout_is_eff_sample_size":0.9981778743517461,"rollout_corr/log_ppl_diff_min":-0.007962644100189209,"rollout_corr/rollout_is_seq_fraction_low":0,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.004480669592817572,"self_distillation/success_group_fraction":0.84375,"val-aux/sciknoweval/reward/best@16/mean":0.73856,"val-aux/sciknoweval/score/maj@8/std":0.09364957204441302,"timing_s/agent_loop/generate_sequences/mean":3.1975462599875755,"val-aux/sciknoweval/incorrect_format/maj@8/mean":1,"val-aux/sciknoweval/reward/best@4/std":0.09758132520114339,"val-aux/sciknoweval/score/worst@16/mean":0.3436,"val-aux/sciknoweval/score/maj@16/std":0.06533731091763299,"training/rollout_probs_diff_max":0.27729934453964233,"response_length/min":150,"val-aux/sciknoweval/score/best@16/std":0.04218677040162369,"val-aux/sciknoweval/score/best@8/std":0.06916601903650618,"critic/advantages/min":-0.875,"val-aux/sciknoweval/acc/maj@2/std":0.16941156579650474,"rollout_corr/training_log_ppl":0.37094445162802003,"val-aux/sciknoweval/score/maj@16/mean":0.55908,"response_length/clip_ratio":0,"timing_s/update_actor":21.07729810476303,"training/rollout_probs_diff_std":0.01337695773690939,"val-aux/sciknoweval/reward/worst@4/mean":0.4220399999999999,"timing_per_token_ms/adv":0.004321415857761778,"response_length/max":1625,"val-aux/sciknoweval/score/best@2/std":0.1400715293582534,"rollout_corr/rollout_is_seq_min":0.9914166331291199,"val-aux/sciknoweval/reward/best@8/mean":0.71124,"val-aux/sciknoweval/acc/worst@4/std":0.11484012157584442,"self_distillation/token_reweight_correct_sample_fraction":0.671875,"val-aux/sciknoweval/reward/maj@2/std":0.16941156579650474,"critic/rewards/max":1,"actor/grad_norm":0.39938923716545105,"critic/returns/max":0.875,"self_distillation/empty_target_batch":0.16015625,"val-aux/sciknoweval/score/best@16/mean":0.73856,"timing_s/save_checkpoint":12.830345639958978,"rollout_corr/rollout_is_seq_mean":0.9998635053634644,"self_distillation/feedback_used_fraction":0}