Files
qwen3-4b-tooluse-srpo-ema005/artifacts/wandb-summary.json
ModelHub XC 9aabae373d 初始化项目,由ModelHub XC社区提供模型
Model: SeongryongJung/qwen3-4b-tooluse-srpo-ema005
Source: Original Platform
2026-09-20 16:14:29 +08:00

1 line
11 KiB
JSON

{"perf/throughput":979.2552930338192,"rollout_corr/k3_kl":0.001191004929382622,"actor/pg_loss":-0.018329072030610405,"timing_s/gen":11.10028469748795,"val-aux/tooluse/score/best@2/mean":0.5926029411764705,"val-aux/num_turns/max":2,"rollout_corr/log_ppl_abs_diff":0.003266506794716406,"val-aux/tooluse/acc/maj@4/mean":0.5755735294117647,"val-aux/tooluse/score/maj@4/mean":0.5755735294117647,"val-aux/tooluse/acc/maj@8/mean":0.5769705882352941,"response_length_non_aborted/max":621,"val-aux/tooluse/acc/maj@2/std":0.04960767852121298,"rollout_corr/rollout_is_seq_fraction_low":0,"critic/rewards/min":0,"val-aux/tooluse/score/best@4/mean":0.6068382352941176,"training/rollout_probs_diff_valid":1,"val-aux/tooluse/reward/mean@16":0.5726102941176471,"rollout_corr/rollout_log_ppl":0.08498664719536464,"val-aux/tooluse/score/maj@8/mean":0.5769705882352941,"timing_s/agent_loop/slowest/prompt_length":577,"val-aux/tooluse/incorrect_format/maj@8/std":0.001135690269194578,"val-aux/num_turns/mean":2,"val-aux/tooluse/score/best@4/std":0.0287069371757289,"perf/mfu/actor":0.045203198343943346,"val-aux/tooluse/reward/best@4/mean":0.6068382352941176,"rollout_corr/log_ppl_diff_max":0.035721685737371445,"training/epoch":1,"val-aux/tooluse/incorrect_format/worst@4/std":0,"val-aux/tooluse/acc/worst@4/mean":0.5356617647058824,"val-aux/tooluse/score/std@16":0.04960007010677716,"val-aux/tooluse/reward/worst@8/mean":0.5206617647058823,"val-aux/tooluse/incorrect_format/worst@2/mean":0.00010294117647058824,"perf/max_memory_reserved_gb":152.583984375,"val-aux/tooluse/score/maj@4/std":0.03821000405576706,"timing_s/adv":1.2358614541590214,"rollout_corr/rollout_is_std":0.03342141583561897,"val-aux/tooluse/reward/worst@4/mean":0.5356617647058824,"val-aux/tooluse/acc/worst@4/std":0.0360951572887439,"self_distillation/dynamic_weighting_enabled":1,"val-aux/tooluse/reward/best@4/std":0.0287069371757289,"self_distillation/feedback_used_fraction":0,"val-aux/tooluse/acc/best@8/mean":0.6173235294117647,"val-aux/tooluse/incorrect_format/worst@16/mean":0,"val-aux/tooluse/reward/worst@16/std":0.01768991770749048,"rollout_corr/rollout_is_ratio_fraction_high":0,"val-aux/tooluse/score/maj@16/mean":0.57725,"val-aux/tooluse/incorrect_format/maj@4/mean":0.00042647058823529416,"val-core/tooluse/acc/maj@16/std":0.019721439199833088,"global_seqlen/minmax_diff":12263,"rollout_corr/kl":0.0009711506442058315,"timing_s/agent_loop/tool_calls/min":0,"self_distillation/feedback_available_fraction":0,"val-aux/tooluse/acc/worst@2/std":0.04453765865500009,"timing_s/agent_loop/slowest/response_length":621,"timing_s/agent_loop/generate_sequences/mean":2.140533648460405,"rollout_corr/chi2_seq":0.2624655382242054,"srpo/routed_token_count":116.134765625,"perf/total_num_tokens":357549,"val-aux/tooluse/score/worst@4/std":0.0360951572887439,"timing_s/agent_loop/tool_calls/mean":0,"global_seqlen/balanced_max":44695,"val-aux/tooluse/incorrect_format/maj@2/std":0.0034374036633244977,"val-aux/tooluse/incorrect_format/best@4/std":0.00615057597100357,"timing_s/agent_loop/slowest/tool_calls":0,"timing_s/old_log_prob":4.850968439131975,"val-aux/tooluse/acc/best@2/std":0.03951972692059264,"val-aux/tooluse/incorrect_format/best@8/mean":0.005808823529411765,"critic/score/min":0,"val-aux/tooluse/reward/maj@8/std":0.0264644012974148,"training/rollout_probs_diff_std":0.01246405765414238,"training/rollout_actor_probs_pearson_corr":0.996764600276947,"val-aux/tooluse/reward/best@2/std":0.03951972692059264,"critic/score/max":1,"val-aux/tooluse/incorrect_format/mean@16":0.0009191176470588235,"val-aux/tooluse/acc/best@4/std":0.0287069371757289,"val-aux/tooluse/incorrect_format/best@2/std":0.004880137697302882,"srpo/grpo_actor/pg_clipfrac":0.00016560931271669688,"val-aux/tooluse/acc/best@4/mean":0.6068382352941176,"timing_s/agent_loop/generate_sequences/max":4.9810793828219175,"critic/rewards/mean":0.30859375,"val-aux/tooluse/reward/best@8/std":0.019675115236200497,"val-aux/tooluse/score/worst@8/std":0.02801237697039909,"timing_s/reward":0.12771783582866192,"actor/entropy":0.0914105623960495,"critic/returns/mean":-0.0008787272381596267,"val-aux/tooluse/reward/maj@16/std":0.019721439199833088,"num_turns/min":2,"timing_s/testing":68.99287504144013,"srpo/sdpo_sample_fraction":0.01953125,"srpo/grpo_sample_fraction":0.98046875,"timing_s/agent_loop/tool_calls/max":0,"actor/grad_norm":0.061348553746938705,"val-aux/tooluse/incorrect_format/best@16/std":0.006981051343076578,"val-aux/tooluse/score/best@16/mean":0.6253382352941176,"val-aux/tooluse/acc/worst@16/std":0.01768991770749048,"rollout_corr/log_ppl_diff_min":-0.011336058378219604,"val-aux/tooluse/acc/worst@2/mean":0.5539705882352941,"val-aux/tooluse/reward/std@16":0.04960007010677716,"_wandb":{"runtime":6637},"perf/max_memory_allocated_gb":145.91975688934326,"srpo/grpo_actor/pg_clipfrac_lower":0,"val-aux/tooluse/incorrect_format/worst@8/std":0,"response_length_non_aborted/mean":116.134765625,"val-aux/tooluse/acc/maj@2/mean":0.5734558823529412,"rollout_corr/rollout_is_max":1.6952815055847168,"val-aux/tooluse/score/best@8/mean":0.6173235294117647,"val-aux/tooluse/incorrect_format/best@2/mean":0.0018529411764705882,"val-aux/tooluse/score/worst@16/std":0.01768991770749048,"timing_s/agent_loop/slowest/generate_sequences":4.9810793828219175,"actor/kl_loss":0,"prompt_length/max":950,"response_length/clip_ratio":0,"val-aux/tooluse/reward/maj@4/std":0.03821000405576706,"response_length_non_aborted/min":56,"global_seqlen/min":38840,"val-aux/tooluse/score/maj@8/std":0.0264644012974148,"srpo/grpo_actor/ppo_kl":0.00029400874017770917,"val-aux/tooluse/incorrect_format/best@8/std":0.007188980778711138,"rollout_corr/chi2_token":0.002979815355502069,"srpo/grpo_loss":-0.009760163331520744,"critic/score/mean":0.30859375,"_timestamp":1.7828067206710045e+09,"self_distillation/dynamic_weight_max":1.0602504014968872,"prompt_length/mean":582.203125,"val-aux/tooluse/incorrect_format/maj@16/mean":1.4705882352941177e-05,"val-aux/tooluse/score/best@2/std":0.03951972692059264,"srpo/sdpo_loss":0.0005956273162155412,"val-aux/tooluse/score/worst@2/std":0.04453765865500009,"val-aux/tooluse/score/maj@2/std":0.04960767852121298,"timing_s/stop_profile":0.00012310408055782318,"val-aux/tooluse/acc/maj@8/std":0.0264644012974148,"val-aux/tooluse/incorrect_format/maj@8/mean":8.823529411764706e-05,"actor/lr":5e-06,"rollout_corr/ppl_ratio":1.0009827739559114,"val-aux/tooluse/acc/worst@16/mean":0.5086176470588235,"val-aux/tooluse/score/worst@4/mean":0.5356617647058824,"val-aux/tooluse/reward/worst@4/std":0.0360951572887439,"critic/rewards/max":1,"timing_s/update_actor":28.18702543154359,"response_length/mean":116.134765625,"rollout_corr/rollout_is_seq_min":0.9903393983840942,"rollout_corr/rollout_is_ratio_fraction_low":5.0453239964554086e-05,"rollout_corr/log_ppl_diff":0.0009711508364489418,"rollout_corr/rollout_is_seq_mean":0.9999406933784485,"val-aux/tooluse/reward/worst@2/mean":0.5539705882352941,"timing_per_token_ms/adv":0.0034564813610414836,"self_distillation/dynamic_weight_entropy_mean":0.0010740847683337051,"global_seqlen/balanced_min":44690,"rollout_corr/rollout_is_mean":0.9999153017997742,"critic/advantages/max":0.75,"val-aux/tooluse/score/worst@16/mean":0.5086176470588235,"response/aborted_ratio":0,"val-aux/tooluse/score/worst@8/mean":0.5206617647058823,"val-aux/tooluse/incorrect_format/best@16/mean":0.009661764705882354,"rollout_corr/training_log_ppl":0.08595779803181358,"_step":100,"val-aux/tooluse/reward/maj@2/std":0.04960767852121298,"val-core/tooluse/acc/best@16/mean":0.6253382352941176,"val-aux/tooluse/incorrect_format/maj@4/std":0.0024677417075026505,"perf/mfu/actor_infer":0,"perf/cpu_memory_used_gb":156.69676208496094,"val-aux/tooluse/acc/best@2/mean":0.5926029411764705,"training/rollout_probs_diff_max":0.3016180694103241,"training/rollout_probs_diff_mean":0.003179528983309865,"rollout_corr/rollout_is_seq_max":1.0124610662460327,"timing_s/step":45.64042218402028,"self_distillation/empty_target_batch":0.671875,"rollout_corr/rollout_is_seq_fraction_high":0,"global_seqlen/max":51103,"val-core/tooluse/acc/best@16/std":0.012510408738665772,"val-aux/tooluse/reward/best@16/std":0.012510408738665772,"val-core/tooluse/acc/maj@16/mean":0.57725,"val-aux/tooluse/reward/best@16/mean":0.6253382352941176,"val-aux/tooluse/reward/worst@2/std":0.04453765865500009,"val-aux/tooluse/incorrect_format/std@16":0.00355972734026417,"val-aux/tooluse/reward/best@8/mean":0.6173235294117647,"timing_s/agent_loop/generate_sequences/min":1.2067538425326347,"val-aux/tooluse/acc/best@8/std":0.019675115236200497,"response_length/min":56,"_runtime":6637.054506473,"prompt_length/clip_ratio":0,"val-aux/tooluse/reward/maj@4/mean":0.5755735294117647,"val-aux/tooluse/incorrect_format/maj@2/mean":0.0008529411764705883,"critic/returns/max":0.75,"val-aux/tooluse/reward/worst@16/mean":0.5086176470588235,"val-aux/tooluse/acc/maj@4/std":0.03821000405576706,"val-core/tooluse/acc/mean@16":0.5726102941176471,"timing_per_token_ms/update_actor":0.0788340211594595,"global_seqlen/mean":44693.625,"rollout_corr/rollout_is_seq_max_deviation":0.012461066246032715,"val-aux/tooluse/acc/std@16":0.04960007010677716,"perf/time_per_step":45.64042218402028,"srpo/routed_loss":-0.009164536015305202,"val-aux/tooluse/acc/worst@8/std":0.02801237697039909,"val-aux/tooluse/reward/maj@16/mean":0.57725,"response_length_non_aborted/clip_ratio":0,"val-aux/tooluse/incorrect_format/worst@4/mean":0,"val-aux/tooluse/incorrect_format/worst@16/std":0,"timing_per_token_ms/gen":0.18668176952099613,"num_turns/mean":2,"val-aux/tooluse/incorrect_format/worst@8/mean":0,"val-aux/tooluse/score/best@16/std":0.012510408738665772,"val-aux/tooluse/score/best@8/std":0.019675115236200497,"val-aux/tooluse/reward/best@2/mean":0.5926029411764705,"val-aux/num_turns/min":2,"rollout_corr/training_ppl":1.0909515626262873,"rollout_corr/rollout_is_eff_sample_size":0.9988841362904098,"val-aux/tooluse/score/maj@2/mean":0.5734558823529412,"rollout_corr/rollout_is_seq_std":0.0030712715815752745,"val-aux/tooluse/incorrect_format/maj@16/std":0.00046480825380232674,"val-aux/tooluse/incorrect_format/best@4/mean":0.003323529411764706,"val-aux/tooluse/reward/maj@8/mean":0.5769705882352941,"val-aux/tooluse/acc/worst@8/mean":0.5206617647058823,"critic/advantages/min":-0.875,"critic/returns/min":-0.875,"val-aux/tooluse/reward/maj@2/mean":0.5734558823529412,"rollout_corr/rollout_is_min":0.35305193066596985,"response_length/max":621,"timing_s/start_profile":0.00010925903916358948,"val-aux/tooluse/score/maj@16/std":0.019721439199833088,"prompt_length/min":270,"self_distillation/success_sample_fraction":0.328125,"training/global_step":100,"self_distillation/correct_sample_fraction":0.30859375,"val-aux/tooluse/score/mean@16":0.5726102941176471,"critic/advantages/mean":-0.0008787272381596267,"rollout_corr/rollout_ppl":1.0898592164739966,"timing_s/save_checkpoint":6.699409354478121,"val-aux/tooluse/reward/worst@8/std":0.02801237697039909,"srpo/correct_sample_fraction":0.30859375,"num_turns/max":2,"val-aux/tooluse/incorrect_format/worst@2/std":0.0012260684910055711,"self_distillation/success_group_fraction":0.328125,"self_distillation/reprompt_sample_fraction":0.328125,"val-aux/tooluse/score/worst@2/mean":0.5539705882352941,"self_distillation/dynamic_weight_mean":0.019531250232830644}