Files
Qwen3-4B-Chemical-GRPO-TR/artifacts/wandb-summary.json
ModelHub XC 1700d223a4 初始化项目,由ModelHub XC社区提供模型
Model: SeongryongJung/Qwen3-4B-Chemical-GRPO-TR
Source: Original Platform
2026-08-09 20:57:18 +08:00

1 line
10 KiB
JSON

{"rollout_corr/rollout_is_std":0.03857421875,"val-aux/sciknoweval/incorrect_format/maj@2/std":0.0023491098553929386,"training/global_step":100,"val-aux/sciknoweval/score/worst@8/std":0.0857140101161578,"val-aux/sciknoweval/incorrect_format/maj@4/mean":0.9998999999999999,"timing_per_token_ms/update_actor":0.050081200018674395,"val-aux/sciknoweval/reward/mean@16":0.6880952380952381,"val-aux/sciknoweval/score/maj@8/std":0.06145786769010523,"actor/grad_norm":0.19856154546141624,"critic/returns/max":0.75,"timing_s/agent_loop/tool_calls/min":0,"timing_s/stop_profile":0.00011307559907436371,"val-aux/sciknoweval/incorrect_format/maj@8/mean":1,"val-aux/sciknoweval/reward/best@2/mean":0.733652380952381,"val-aux/sciknoweval/score/best@8/std":0.03693927335257077,"val-aux/sciknoweval/reward/best@16/std":0.022230541736786433,"timing_s/agent_loop/generate_sequences/mean":4.972953367498121,"critic/score/min":0,"val-aux/num_turns/min":2,"prompt_length/clip_ratio":0,"val-aux/sciknoweval/incorrect_format/worst@16/std":0.004551356214978213,"val-aux/sciknoweval/score/maj@8/mean":0.6996714285714285,"val-aux/sciknoweval/reward/maj@2/mean":0.6873380952380953,"val-aux/sciknoweval/acc/best@2/std":0.08603291821062148,"val-aux/sciknoweval/reward/maj@8/std":0.06145786769010523,"val-aux/sciknoweval/reward/worst@4/std":0.10105952090465425,"val-aux/sciknoweval/incorrect_format/maj@4/std":0.0009636106375423148,"val-aux/sciknoweval/incorrect_format/best@8/mean":1,"actor/pg_clipfrac":8.191586903194548e-05,"val-aux/sciknoweval/score/best@16/mean":0.8012238095238096,"rollout_corr/chi2_token":0.0015719085931777954,"prompt_length/max":625,"val-aux/sciknoweval/reward/best@4/mean":0.7654619047619048,"val-aux/sciknoweval/reward/best@16/mean":0.8012238095238096,"val-aux/sciknoweval/incorrect_format/best@8/std":0,"val-aux/sciknoweval/acc/best@4/std":0.05713592644762481,"val-aux/sciknoweval/score/best@4/mean":0.7654619047619048,"rollout_corr/rollout_is_ratio_fraction_low":5.7171793741872534e-05,"val-aux/sciknoweval/reward/worst@16/std":0.06804166756222384,"val-aux/sciknoweval/incorrect_format/maj@16/std":0,"val-aux/sciknoweval/acc/maj@8/std":0.06145786769010523,"timing_s/agent_loop/generate_sequences/max":8.91172250919044,"val-aux/sciknoweval/score/maj@4/std":0.0900863886286685,"training/epoch":1,"val-aux/sciknoweval/score/best@2/std":0.08603291821062148,"val-aux/sciknoweval/acc/worst@4/std":0.10105952090465425,"val-aux/sciknoweval/score/maj@2/std":0.11903344827555573,"val-core/sciknoweval/acc/best@16/std":0.022230541736786433,"val-core/sciknoweval/acc/maj@16/mean":0.6995095238095238,"rollout_corr/rollout_is_seq_min":0.9923958778381348,"val-aux/sciknoweval/incorrect_format/worst@2/std":0.003080186451482751,"critic/rewards/max":1,"_wandb":{"runtime":4227},"val-aux/sciknoweval/reward/maj@4/std":0.0900863886286685,"perf/total_num_tokens":198254,"rollout_corr/rollout_ppl":1.5293380408547819,"rollout_corr/rollout_is_seq_std":0.0019207203295081854,"val-aux/sciknoweval/acc/std@16":0.11930428433281097,"response_length_non_aborted/min":194,"val-aux/sciknoweval/acc/worst@16/std":0.06804166756222384,"val-aux/sciknoweval/reward/worst@2/mean":0.6426714285714287,"val-aux/sciknoweval/acc/maj@2/mean":0.6873380952380953,"val-aux/sciknoweval/reward/worst@16/mean":0.5196142857142857,"global_seqlen/balanced_max":24784,"critic/score/mean":0.62890625,"rollout_corr/rollout_is_min":0.3767692446708679,"perf/cpu_memory_used_gb":104.02334213256836,"prompt_length/min":165,"perf/time_per_step":27.04090141505003,"training/rollout_probs_diff_std":0.01128289382904768,"global_seqlen/minmax_diff":7976,"actor/pg_clipfrac_lower":0,"rollout_corr/rollout_is_seq_max_deviation":0.007774233818054199,"val-aux/sciknoweval/score/maj@16/mean":0.6995095238095238,"val-core/sciknoweval/acc/maj@16/std":0.046993204377870766,"val-aux/sciknoweval/acc/worst@16/mean":0.5196142857142857,"rollout_corr/training_log_ppl":0.41552202677121386,"val-aux/sciknoweval/incorrect_format/best@2/mean":0.999952380952381,"rollout_corr/k3_kl":0.0007973912350962564,"val-aux/sciknoweval/reward/worst@8/mean":0.5585714285714286,"training/rollout_actor_probs_pearson_corr":0.9990952014923096,"rollout_corr/rollout_is_ratio_fraction_high":8.167398846126162e-06,"val-aux/sciknoweval/reward/maj@16/std":0.046993204377870766,"num_turns/max":2,"response_length/max":1157,"_timestamp":1.782973080828363e+09,"perf/mfu/actor":0.07253216062486065,"val-aux/sciknoweval/acc/worst@2/std":0.11377153742699202,"rollout_corr/kl":0.0008094973310157627,"val-aux/sciknoweval/incorrect_format/worst@4/std":0.004004909923005107,"critic/returns/mean":-0.003304733894765377,"rollout_corr/chi2_seq":1.4306007844861597,"val-aux/sciknoweval/acc/worst@8/std":0.0857140101161578,"timing_s/agent_loop/generate_sequences/min":1.5211810767650604,"val-core/sciknoweval/acc/mean@16":0.6880952380952381,"_step":100,"timing_per_token_ms/gen":0.11866824797057118,"perf/throughput":916.454286032319,"val-aux/sciknoweval/reward/maj@4/mean":0.6969476190476189,"rollout_corr/training_ppl":1.5306207397952676,"val-aux/sciknoweval/incorrect_format/best@16/std":0,"val-aux/sciknoweval/reward/best@8/std":0.03693927335257077,"timing_s/old_log_prob":2.3071191553026438,"val-aux/sciknoweval/score/mean@16":0.6880952380952381,"val-aux/sciknoweval/incorrect_format/worst@4/mean":0.9978047619047619,"rollout_corr/rollout_is_mean":0.9998645186424255,"val-aux/sciknoweval/acc/maj@4/std":0.0900863886286685,"rollout_corr/rollout_is_max":2.2078511714935303,"response_length_non_aborted/mean":478.2734375,"val-aux/sciknoweval/reward/worst@8/std":0.0857140101161578,"rollout_corr/rollout_is_eff_sample_size":0.9985136461377951,"val-aux/sciknoweval/acc/best@8/std":0.03693927335257077,"critic/score/max":1,"training/rollout_probs_diff_valid":1,"critic/advantages/min":-0.75,"val-aux/sciknoweval/score/maj@2/mean":0.6873380952380953,"training/rollout_probs_diff_mean":0.005019628442823887,"val-aux/sciknoweval/incorrect_format/maj@8/std":0,"rollout_corr/rollout_is_seq_fraction_low":0,"val-aux/sciknoweval/acc/maj@2/std":0.11903344827555573,"val-aux/sciknoweval/score/maj@4/mean":0.6969476190476189,"actor/pg_loss":-0.00012778816744685173,"global_seqlen/max":27971,"num_turns/min":2,"val-aux/sciknoweval/score/std@16":0.11930428433281097,"val-aux/sciknoweval/incorrect_format/worst@2/mean":0.9988666666666667,"timing_s/agent_loop/tool_calls/max":0,"timing_s/agent_loop/slowest/tool_calls":0,"critic/advantages/mean":-0.003304733894765377,"actor/ppo_kl":-9.224095602267113e-05,"val-aux/sciknoweval/reward/best@4/std":0.05713592644762481,"rollout_corr/rollout_is_seq_max":1.0077742338180542,"critic/rewards/min":0,"rollout_corr/log_ppl_diff_min":-0.007830023765563965,"val-aux/sciknoweval/acc/best@2/mean":0.733652380952381,"val-aux/sciknoweval/reward/worst@4/mean":0.5992857142857143,"val-aux/sciknoweval/incorrect_format/best@16/mean":1,"val-aux/sciknoweval/incorrect_format/worst@8/std":0.004680790716043936,"timing_s/agent_loop/slowest/prompt_length":190,"perf/max_memory_reserved_gb":129.306640625,"val-aux/sciknoweval/incorrect_format/worst@16/mean":0.9938380952380953,"perf/max_memory_allocated_gb":123.54408025741577,"response_length/min":194,"val-aux/sciknoweval/acc/maj@8/mean":0.6996714285714285,"val-aux/sciknoweval/reward/best@2/std":0.08603291821062148,"critic/returns/min":-0.75,"val-aux/sciknoweval/score/worst@16/std":0.06804166756222384,"rollout_corr/rollout_log_ppl":0.4147125284653157,"val-aux/sciknoweval/acc/best@8/mean":0.7866809523809524,"val-aux/sciknoweval/acc/worst@8/mean":0.5585714285714286,"timing_s/agent_loop/tool_calls/mean":0,"val-aux/sciknoweval/score/worst@2/std":0.11377153742699202,"val-aux/sciknoweval/incorrect_format/mean@16":0.9994047619047619,"rollout_corr/rollout_is_seq_mean":0.999876856803894,"val-aux/sciknoweval/incorrect_format/best@2/std":0.0006717493323650422,"val-aux/sciknoweval/incorrect_format/worst@8/mean":0.9960904761904762,"critic/rewards/mean":0.62890625,"val-aux/sciknoweval/acc/best@4/mean":0.7654619047619048,"timing_s/save_checkpoint":6.6051019467413425,"val-aux/sciknoweval/score/worst@16/mean":0.5196142857142857,"rollout_corr/log_ppl_abs_diff":0.0015918280114419758,"training/rollout_probs_diff_max":0.3175773024559021,"val-aux/sciknoweval/score/best@4/std":0.05713592644762481,"val-aux/sciknoweval/reward/maj@16/mean":0.6995095238095238,"perf/mfu/actor_infer":0,"val-aux/sciknoweval/acc/maj@4/mean":0.6969476190476189,"val-aux/sciknoweval/incorrect_format/best@4/mean":1,"timing_s/update_actor":9.928798228502274,"rollout_corr/rollout_is_seq_fraction_high":0,"val-aux/sciknoweval/score/best@2/mean":0.733652380952381,"critic/advantages/max":0.75,"val-aux/sciknoweval/score/best@8/mean":0.7866809523809524,"timing_s/reward":0.06497454643249512,"val-aux/sciknoweval/score/worst@2/mean":0.6426714285714287,"global_seqlen/min":19995,"val-aux/sciknoweval/reward/std@16":0.11930428433281097,"timing_per_token_ms/adv":0.0006147575128339063,"val-aux/sciknoweval/acc/worst@2/mean":0.6426714285714287,"val-aux/sciknoweval/incorrect_format/std@16":0.0023053472298853674,"val-core/sciknoweval/acc/best@16/mean":0.8012238095238096,"actor/entropy":0.397185742855072,"val-aux/sciknoweval/reward/worst@2/std":0.11377153742699202,"val-aux/sciknoweval/score/maj@16/std":0.046993204377870766,"_runtime":4227.218885724,"timing_s/step":27.04090141505003,"timing_s/agent_loop/slowest/generate_sequences":8.91172250919044,"response/aborted_ratio":0,"timing_s/agent_loop/slowest/response_length":1039,"val-aux/sciknoweval/reward/best@8/mean":0.7866809523809524,"timing_s/gen":14.529502945020795,"response_length_non_aborted/max":1157,"timing_s/testing":126.50542075186968,"val-aux/sciknoweval/incorrect_format/maj@16/mean":1,"actor/kl_loss":0,"val-aux/sciknoweval/score/worst@4/std":0.10105952090465425,"response_length/clip_ratio":0,"val-aux/sciknoweval/score/worst@4/mean":0.5992857142857143,"global_seqlen/balanced_min":24776,"val-aux/sciknoweval/incorrect_format/maj@2/mean":0.9993761904761905,"rollout_corr/ppl_ratio":1.0008116397075355,"response_length/mean":478.2734375,"timing_s/start_profile":0.00015527382493019104,"rollout_corr/log_ppl_diff_max":0.007098793983459473,"val-aux/num_turns/mean":2,"response_length_non_aborted/clip_ratio":0,"prompt_length/mean":296.15625,"val-aux/sciknoweval/score/worst@8/mean":0.5585714285714286,"val-aux/sciknoweval/incorrect_format/best@4/std":0,"val-aux/num_turns/max":2,"val-aux/sciknoweval/reward/maj@8/mean":0.6996714285714285,"actor/lr":1e-06,"val-aux/sciknoweval/score/best@16/std":0.022230541736786433,"global_seqlen/mean":24781.75,"num_turns/mean":2,"val-aux/sciknoweval/acc/worst@4/mean":0.5992857142857143,"timing_s/adv":0.12187813594937325,"rollout_corr/log_ppl_diff":0.0008094983058981597,"val-aux/sciknoweval/reward/maj@2/std":0.11903344827555573}