Files
Qwen3-4B-Biology-GRPO-TR/artifacts/wandb-summary.json
ModelHub XC 7d17112db8 初始化项目,由ModelHub XC社区提供模型
Model: SeongryongJung/Qwen3-4B-Biology-GRPO-TR
Source: Original Platform
2026-10-03 20:15:22 +08:00

1 line
10 KiB
JSON

{"timing_s/testing":66.87851952016354,"response_length_non_aborted/clip_ratio":0,"critic/advantages/max":0.875,"val-aux/sciknoweval/acc/worst@2/mean":0.34018,"val-aux/sciknoweval/score/worst@16/std":0.032643145410378605,"rollout_corr/rollout_is_seq_max":1.0129314661026,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.008749306906094262,"critic/score/min":0,"timing_s/agent_loop/tool_calls/min":0,"timing_s/agent_loop/tool_calls/mean":0,"training/rollout_actor_probs_pearson_corr":0.9986383318901062,"val-aux/sciknoweval/incorrect_format/best@8/std":0,"timing_s/agent_loop/generate_sequences/max":6.653986681252718,"global_seqlen/mean":19150.75,"val-core/sciknoweval/acc/best@16/mean":0.6543800000000001,"prompt_length/clip_ratio":0,"rollout_corr/rollout_is_seq_min":0.9921380877494812,"val-aux/sciknoweval/incorrect_format/maj@8/std":0.002504474311076614,"critic/advantages/min":-0.875,"val-aux/sciknoweval/acc/worst@8/mean":0.21248,"_timestamp":1.782992219032823e+09,"actor/pg_loss":5.068804603070021e-05,"rollout_corr/rollout_is_std":0.040638964623212814,"rollout_corr/rollout_is_seq_max_deviation":0.012931466102600098,"response_length/max":977,"rollout_corr/rollout_is_max":4.838179588317871,"val-aux/sciknoweval/reward/worst@16/std":0.032643145410378605,"val-aux/sciknoweval/reward/worst@8/std":0.07648178927728648,"timing_s/agent_loop/tool_calls/max":0,"val-aux/sciknoweval/reward/best@8/std":0.08028605088755732,"critic/rewards/max":1,"val-aux/sciknoweval/reward/best@2/std":0.1717994688879882,"val-core/sciknoweval/acc/mean@16":0.42625,"response_length/mean":248.5546875,"val-aux/sciknoweval/reward/mean@16":0.42625,"val-aux/sciknoweval/reward/maj@16/mean":0.44676000000000005,"val-aux/sciknoweval/acc/best@8/std":0.08028605088755732,"critic/returns/max":0.875,"_runtime":4362.153699158,"timing_per_token_ms/adv":0.000782728193439709,"val-aux/sciknoweval/score/best@8/std":0.08028605088755732,"val-aux/sciknoweval/reward/best@4/std":0.12262009307652777,"val-aux/sciknoweval/score/maj@8/mean":0.4453000000000001,"response_length_non_aborted/mean":248.5546875,"perf/max_memory_allocated_gb":123.67111110687256,"val-aux/sciknoweval/reward/maj@8/std":0.1155248617088842,"actor/ppo_kl":0.0007291194974783366,"val-aux/sciknoweval/acc/maj@4/mean":0.43562000000000006,"perf/max_memory_reserved_gb":129.48046875,"val-aux/num_turns/max":2,"global_seqlen/max":24171,"training/rollout_probs_diff_mean":0.005042023025453091,"val-aux/sciknoweval/acc/worst@16/mean":0.18634,"val-core/sciknoweval/acc/best@16/std":0.0472167585635409,"val-aux/sciknoweval/reward/maj@8/mean":0.4453000000000001,"perf/cpu_memory_used_gb":103.95842933654785,"global_seqlen/balanced_max":19154,"val-aux/sciknoweval/reward/worst@16/mean":0.18634,"timing_s/agent_loop/generate_sequences/min":0.7739909812808037,"val-aux/sciknoweval/reward/maj@4/std":0.16364220520548403,"val-aux/sciknoweval/reward/best@4/mean":0.5764,"response_length_non_aborted/max":977,"actor/grad_norm":0.496917188167572,"val-aux/sciknoweval/score/worst@8/mean":0.21248,"_wandb":{"runtime":4362},"timing_s/agent_loop/generate_sequences/mean":2.1721716843530885,"val-aux/sciknoweval/reward/best@8/mean":0.62372,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"val-aux/sciknoweval/score/best@16/std":0.0472167585635409,"prompt_length/max":499,"actor/pg_clipfrac_lower":0,"val-aux/sciknoweval/score/maj@16/mean":0.44676000000000005,"val-core/sciknoweval/acc/maj@16/std":0.07340249583082686,"val-aux/sciknoweval/reward/worst@4/std":0.12852696069401104,"actor/pg_clipfrac":0.0010578534675005358,"rollout_corr/training_log_ppl":0.3114306655479595,"val-aux/sciknoweval/reward/best@16/mean":0.6543800000000001,"rollout_corr/rollout_is_eff_sample_size":0.9983516728081235,"val-aux/sciknoweval/incorrect_format/best@8/mean":1,"training/rollout_probs_diff_valid":1,"rollout_corr/kl":0.0015098723376403278,"val-aux/sciknoweval/score/best@16/mean":0.6543800000000001,"rollout_corr/rollout_is_seq_fraction_low":0,"actor/entropy":0.2835403382778168,"rollout_corr/log_ppl_diff_max":0.014475628733634949,"prompt_length/min":180,"val-aux/sciknoweval/reward/maj@2/mean":0.42506,"perf/mfu/actor_infer":0,"timing_s/old_log_prob":2.375061484053731,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.004842664831259046,"val-aux/sciknoweval/score/worst@2/mean":0.34018,"rollout_corr/rollout_is_seq_mean":1.000152587890625,"val-aux/sciknoweval/reward/maj@2/std":0.2031877981200387,"critic/score/mean":0.5234375,"val-aux/sciknoweval/acc/best@8/mean":0.62372,"val-aux/sciknoweval/score/maj@8/std":0.1155248617088842,"timing_s/save_checkpoint":6.8637774139642715,"val-aux/num_turns/min":2,"val-aux/sciknoweval/reward/std@16":0.20340466932397192,"rollout_corr/log_ppl_abs_diff":0.003156350867357105,"val-aux/sciknoweval/acc/worst@8/std":0.07648178927728648,"num_turns/max":2,"val-aux/sciknoweval/acc/worst@4/mean":0.26536,"val-aux/sciknoweval/reward/maj@16/std":0.07340249583082686,"val-aux/sciknoweval/incorrect_format/std@16":0.009682458365518542,"val-core/sciknoweval/acc/maj@16/mean":0.44676000000000005,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"critic/score/max":1,"num_turns/min":2,"timing_s/gen":12.76038683950901,"perf/mfu/actor":0.056309282462256305,"val-aux/sciknoweval/acc/best@2/mean":0.51122,"val-aux/sciknoweval/reward/maj@4/mean":0.43562000000000006,"val-aux/sciknoweval/incorrect_format/mean@16":0.9975,"timing_s/update_actor":9.685975370928645,"rollout_corr/rollout_is_seq_std":0.002947622211650014,"rollout_corr/ppl_ratio":1.0015187887474895,"val-aux/sciknoweval/incorrect_format/maj@16/mean":1,"training/epoch":7,"val-aux/sciknoweval/acc/best@4/std":0.12262009307652777,"training/rollout_probs_diff_max":0.3625403642654419,"val-aux/sciknoweval/score/worst@4/std":0.12852696069401104,"val-aux/sciknoweval/score/maj@4/mean":0.43562000000000006,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0029552123640712146,"val-aux/sciknoweval/acc/best@2/std":0.1717994688879882,"timing_per_token_ms/gen":0.20054041866272215,"perf/total_num_tokens":153206,"timing_s/adv":0.11991865560412407,"training/global_step":100,"val-aux/sciknoweval/reward/worst@2/mean":0.34018,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.9955799999999999,"val-aux/sciknoweval/score/best@4/mean":0.5764,"val-aux/sciknoweval/acc/maj@2/mean":0.42506,"val-aux/sciknoweval/reward/worst@8/mean":0.21248,"val-aux/sciknoweval/acc/maj@8/std":0.1155248617088842,"critic/advantages/mean":-0.003972182981669903,"rollout_corr/k3_kl":0.001307048618286899,"timing_s/reward":0.05503258667886257,"rollout_corr/log_ppl_diff_min":-0.009900987148284912,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.99798,"critic/rewards/min":0,"training/rollout_probs_diff_std":0.012832351960241795,"timing_per_token_ms/update_actor":0.06322190626299652,"val-aux/sciknoweval/score/best@8/mean":0.62372,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.019500890160644296,"global_seqlen/balanced_min":19146,"num_turns/mean":2,"response_length/min":111,"val-aux/sciknoweval/score/maj@2/std":0.2031877981200387,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.98444,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.01634315511370097,"rollout_corr/training_ppl":1.3729312745854259,"timing_s/agent_loop/slowest/response_length":977,"val-aux/sciknoweval/incorrect_format/maj@8/mean":0.99984,"timing_s/agent_loop/slowest/generate_sequences":6.653986681252718,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.9997800000000001,"response_length_non_aborted/min":111,"timing_s/step":25.090152885764837,"val-aux/sciknoweval/acc/worst@16/std":0.032643145410378605,"val-aux/sciknoweval/reward/best@2/mean":0.51122,"val-aux/sciknoweval/acc/std@16":0.20340466932397192,"val-aux/sciknoweval/score/worst@8/std":0.07648178927728648,"timing_s/stop_profile":0.00011190585792064667,"rollout_corr/rollout_is_ratio_fraction_low":4.7147572331596166e-05,"val-aux/sciknoweval/score/maj@16/std":0.07340249583082686,"rollout_corr/rollout_ppl":1.3707687840797007,"_step":100,"critic/returns/mean":-0.003972182981669903,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.9751599999999999,"val-aux/sciknoweval/score/worst@2/std":0.17528803531142753,"global_seqlen/minmax_diff":10977,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.012525805206779563,"val-aux/sciknoweval/reward/worst@2/std":0.17528803531142753,"rollout_corr/chi2_seq":0.866482462733984,"response_length/clip_ratio":0,"val-aux/sciknoweval/acc/worst@4/std":0.12852696069401104,"val-aux/sciknoweval/acc/worst@2/std":0.17528803531142753,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.9993599999999999,"perf/time_per_step":25.090152885764837,"val-aux/sciknoweval/acc/maj@8/mean":0.4453000000000001,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"val-aux/sciknoweval/incorrect_format/maj@16/std":0,"critic/returns/min":-0.875,"val-aux/sciknoweval/score/worst@16/mean":0.18634,"timing_s/agent_loop/slowest/prompt_length":431,"val-aux/sciknoweval/score/best@4/std":0.12262009307652777,"val-aux/sciknoweval/score/best@2/std":0.1717994688879882,"val-aux/num_turns/mean":2,"actor/kl_loss":0,"val-aux/sciknoweval/acc/maj@4/std":0.16364220520548403,"val-aux/sciknoweval/acc/best@4/mean":0.5764,"rollout_corr/chi2_token":0.0023139831610023975,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.019403379135729548,"critic/rewards/mean":0.5234375,"timing_s/agent_loop/slowest/tool_calls":0,"actor/lr":1e-06,"val-aux/sciknoweval/reward/worst@4/mean":0.26536,"val-aux/sciknoweval/acc/maj@2/std":0.2031877981200387,"val-aux/sciknoweval/reward/best@16/std":0.0472167585635409,"val-aux/sciknoweval/score/maj@4/std":0.16364220520548403,"perf/throughput":763.2775331100266,"rollout_corr/rollout_is_mean":1.000214695930481,"rollout_corr/rollout_is_ratio_fraction_high":3.143171488773078e-05,"global_seqlen/min":13194,"rollout_corr/rollout_log_ppl":0.30992078845156357,"rollout_corr/log_ppl_diff":0.0015098770963959396,"rollout_corr/rollout_is_min":0.209492027759552,"timing_s/start_profile":9.883567690849304e-05,"val-aux/sciknoweval/incorrect_format/best@4/mean":1,"rollout_corr/rollout_is_seq_fraction_high":0,"val-aux/sciknoweval/score/best@2/mean":0.51122,"prompt_length/mean":349.90625,"val-aux/sciknoweval/score/std@16":0.20340466932397192,"response/aborted_ratio":0,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.9915,"val-aux/sciknoweval/score/maj@2/mean":0.42506,"val-aux/sciknoweval/score/worst@4/mean":0.26536,"val-aux/sciknoweval/score/mean@16":0.42625}