laion/explore-tis-minp-40-8B
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
cab2418769d25889a612a2ac24f775a1c1d18dda
explore-tis-minp-40-8B/training_logs/20260612_111844_metrics_report.md
ModelHub XC cab2418769 初始化项目，由ModelHub XC社区提供模型 
Model: laion/explore-tis-minp-40-8B
Source: Original Platform
2026-08-09 12:53:53 +08:00

24 KiB
Raw Blame History

SkyRL Training Metrics Analysis

Generated from 6 log files

Overview

Log File Total Steps Metric Blocks Final Reward (mean) Final Reward (max) Total Time (s)
explore-tis-minp_736731 15 15 0.4516 0.7305 19208.7
explore-tis-minp_736732 51 37 0.4319 0.5352 42221.2
explore-tis-minp_736733 80 30 0.4185 0.7500 38719.8
explore-tis-minp_736734 82 1 0.7598 0.7598 1584.8
explore-tis-minp_736735 84 1 0.7676 0.7676 1825.9
explore-tis-minp_736736 86 1 0.7559 0.7559 1769.7

Async Metrics

Mean Std Min Max Count
async/discard_rate 0 0 0 0 85
async/discarded_count 0 0 0 0 85
async/effective_batch_groups 64 0 64 64 85
async/effective_batch_samples 512 0 512 512 85
async/staleness_max 4.17647 1.49743 0 6 85
async/staleness_mean 2.10056 0.724504 0 4.0312 85
async/staleness_min 0.152941 0.607608 0 4 85
async/staleness_ratio 0.870776 0.242465 0 1 85

Generate Metrics

Mean Std Min Max Count
generate/avg_num_tokens 7822.76 929.534 4784.76 9627.61 85
generate/avg_tokens_non_zero_rewards 6819.11 951.696 4644.54 10207.6 85
generate/avg_tokens_zero_rewards 8651.31 1022.18 4091.01 10658.1 85
generate/max_num_tokens 31349.5 976.865 25886 31825 85
generate/min_num_tokens 25.2471 159.939 1 1225 85
generate/std_num_tokens 6933.96 869.491 3897.34 8954.15 85
generate/tis/aligned_tokens 2.84057e+06 376398 1.69952e+06 3.60399e+06 85
generate/tis/alignment_fail_count 116.165 19.3032 57 159 85
generate/tis/exact_match_fraction 0.92684 0.0182654 0.8793 0.9711 85
generate/tis/lcs_fallback_fraction 0 0 0 0 85
generate/tis/lcs_fallback_messages 0 0 0 0 85
generate/tis/unaligned_fraction 0.07316 0.0182654 0.0289 0.1207 85

Loss Metrics

Mean Std Min Max Count
loss/avg_final_rewards 0.442236 0.102011 0.2168 0.7676 85
loss/avg_raw_advantages -0.00178235 0.00772197 -0.0244 0.015 85
loss/avg_raw_advantages_abs 0.110764 0.028583 0.048 0.1834 85

Policy Metrics

Mean Std Min Max Count
policy/final_loss 0 0 -0 0 85
policy/log_ratio_abs_max 0 0 0 0 85
policy/log_ratio_abs_mean 0 0 0 0 85
policy/log_ratio_abs_p99 0 0 0 0 85
policy/log_ratio_abs_pos00 0 0 0 0 85
policy/log_ratio_abs_pos10 0 0 0 0 85
policy/log_ratio_abs_pos20 0 0 0 0 85
policy/log_ratio_abs_pos30 0 0 0 0 85
policy/log_ratio_abs_pos40 0 0 0 0 85
policy/log_ratio_abs_pos50 0 0 0 0 85
policy/log_ratio_abs_pos60 0 0 0 0 85
policy/log_ratio_abs_pos70 0 0 0 0 85
policy/log_ratio_abs_pos80 0 0 0 0 85
policy/log_ratio_abs_pos90 0 0 0 0 85
policy/n_tokens_dp_gt_10pct 0 0 0 0 85
policy/n_tokens_dp_gt_1pct 0 0 0 0 85
policy/n_tokens_dp_gt_50pct 0 0 0 0 85
policy/policy_entropy 0.2491 0.0194304 0.1667 0.3047 85
policy/policy_loss 0 0 -0 0 85
policy/policy_lr 0 0 0 0 85
policy/policy_update_steps 1 0 1 1 85
policy/ppo_clip_ratio 0 0 0 0 85
policy/raw_grad_norm 3.52941e-06 1.85617e-05 0 0.0001 85
policy/rollout_train_prob_diff_mean 1.50068e+09 1.24421e+10 88782.3 1.14652e+11 85
policy/rollout_train_prob_diff_std 2.44935e+12 2.03611e+13 5.29886e+07 1.87599e+14 85
policy/tis/imp_ratio_capped_fraction 1.17647e-06 1.08465e-05 0 0.0001 85
policy/tis/imp_ratio_mean 0.885954 0.0532631 0.6183 0.9963 85
policy/tis/log_ratio_abs_mean 0.0241671 0.00403521 0.0143 0.0325 85

Reward Metrics

Mean Std Min Max Count
reward/avg_pass_at_8 0.561272 0.095766 0.3438 0.8438 85
reward/avg_raw_reward 0.442236 0.102011 0.2168 0.7676 85

System Metrics

Mean Std Min Max Count
system/process_rss_gb 18.0157 2.25757 11.8836 21.9725 85
system/process_vms_gb 54.0291 1.86577 50.618 58.5468 85
system/ram_available_gb 605.463 22.5972 564.167 638.413 85
system/ram_percent 29.4329 2.62784 25.6 34.2 85
system/ram_total_gb 857.968 4.90241e-05 857.967 857.968 85
system/ram_used_gb 252.504 22.5972 219.554 293.8 85

Timing Metrics

Mean Std Min Max Count
timing/compute_advantages_and_returns 0.0807635 0.0653647 0.0609 0.656 85
timing/convert_to_training_input 6.66973 0.487257 5.034 7.7627 85
timing/fwd_logprobs_values_reward 55.6374 5.94803 39.0434 71.5317 85
timing/policy_train 337.597 41.4135 198.948 415.158 85
timing/run_training 393.561 46.5249 239.913 476.128 85
timing/step 1239.18 542.211 344.768 3879.65 85
timing/sync_weights 27.5786 3.06918 21.8599 34.646 85
timing/train_critic_and_policy 337.842 41.4245 199.153 415.465 85
timing/wait_for_generation_buffer 811.358 559.804 5.1531 3578.14 85
timing/cleanup_old_checkpoints 2.27277 6.04356 0.0066 27.2722 43
timing/save_checkpoints 11.7555 5.91758 7.7797 34.7642 43
timing/save_hf_model 7.80862 3.19645 5.8369 14.5803 17

Tis Metrics

Mean Std Min Max Count
tis/batch_skipped_no_logprobs 0 0 0 0 85
tis/skipped_fraction 0 0 0 0 85

Trainer Metrics

Mean Std Min Max Count
trainer/epoch 0.0588235 0.236691 0 1 85
trainer/global_step 41.8588 24.1396 1 86 85

Batch_Errors Metrics

Mean Std Min Max Count
batch_errors/total_batches 60.5294 12.6252 16 127 85
batch_errors/total_instances 484.235 101.001 128 1016 85
batch_errors/total_successful 432.482 95.246 128 965 85
batch_errors/total_failed 19.2706 6.71421 0 38 85
batch_errors/total_masked 47.1059 19.2038 0 96 85
batch_errors/avg_VerifierTimeoutError 0.0967394 0.0724197 0.015873 0.33871 67
batch_errors/total_VerifierTimeoutError 5.91045 4.33006 1 21 67
batch_errors/avg_ContextLengthExceededError 0.728128 0.28348 0.03125 1.67797 82
batch_errors/total_ContextLengthExceededError 44.4146 16.6874 2 99 82
batch_errors/avg_EnvironmentStartTimeoutError 0.291939 0.511936 0.031746 1.20588 5
batch_errors/total_EnvironmentStartTimeoutError 11 16.8375 2 41 5
batch_errors/avg_AgentTimeoutError 0.0550829 0.0856444 0.00980392 0.607143 57
batch_errors/total_AgentTimeoutError 3.22807 4.83668 1 34 57
batch_errors/avg_DaytonaError 0.0345684 0.0314819 0.015625 0.1 8
batch_errors/total_DaytonaError 2 1.92725 1 6 8
batch_errors/avg_InvalidChatHistory 0.0751811 0.0551992 0.00980392 0.253968 71
batch_errors/total_InvalidChatHistory 4.50704 3.18422 1 16 71
batch_errors/avg_RuntimeError 0.015625 nan 0.015625 0.015625 1
batch_errors/total_RuntimeError 1 nan 1 1 1
batch_errors/avg_AgentSetupTimeoutError 0.123331 0.115332 0.0169492 0.245902 3
batch_errors/total_AgentSetupTimeoutError 7.33333 7.0946 1 15 3
batch_errors/avg_DaytonaAuthenticationError 0.0357143 nan 0.0357143 0.0357143 1
batch_errors/total_DaytonaAuthenticationError 2 nan 2 2 1

Training Progression by Log

explore-tis-minp_736731

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
1 0.7305 0.8125 0.000000 -0.0000 3879.6 3578.1
2 0.5117 0.6719 0.000000 -0.0000 1691.7 1295.3
3 0.3125 0.4844 0.000000 0.0000 344.8 5.2
4 0.4238 0.5312 0.000000 0.0000 1115.5 735.3
5 0.4492 0.5469 0.000000 -0.0000 1041.3 630.7
6 0.4570 0.5938 0.000000 0.0000 1269.0 801.9
7 0.4707 0.5469 0.000000 0.0000 1034.8 604.1
8 0.4180 0.6406 0.000000 -0.0000 1185.1 734.7
9 0.4355 0.5156 0.000000 0.0000 1120.5 722.0
10 0.3652 0.5625 0.000000 0.0000 1216.4 765.8
11 0.4609 0.5781 0.000000 0.0000 853.9 466.1
12 0.4180 0.5781 0.000000 -0.0000 1328.4 902.0
13 0.4590 0.5156 0.000000 0.0000 867.1 480.4
14 0.4512 0.6562 0.000000 -0.0000 1247.7 827.7
15 0.4102 0.5625 0.000000 0.0000 1013.1 578.4

explore-tis-minp_736732

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
15 0.4824 0.5357 0.000000 -0.0000 3502.1 3131.3
16 0.4590 0.5469 0.000000 0.0000 1342.5 888.5
17 0.2500 0.3906 0.000000 0.0000 520.5 68.0
18 0.3965 0.5469 0.000000 -0.0000 1408.0 948.2
19 0.4902 0.6406 0.000000 0.0000 1149.5 721.1
20 0.3691 0.5000 0.000000 0.0000 1059.2 621.7
21 0.3574 0.5000 0.000000 0.0000 1096.3 672.4
22 0.3887 0.5000 0.000000 0.0000 1386.6 952.7
23 0.4688 0.6094 0.000000 0.0000 1009.3 608.5
24 0.4258 0.5781 0.000000 -0.0000 1382.8 981.8
25 0.4668 0.5312 0.000000 -0.0000 801.1 424.5
26 0.4805 0.5781 0.000000 -0.0000 773.9 395.0
27 0.5215 0.6406 0.000000 -0.0000 907.8 510.8
28 0.4414 0.6094 0.000000 0.0000 985.0 524.9
29 0.4434 0.5781 0.000000 -0.0000 993.5 603.5
30 0.4043 0.5156 0.000000 0.0000 1127.2 679.8
31 0.3887 0.4844 0.000000 0.0000 1049.7 594.0
32 0.3809 0.5000 0.000000 0.0000 1175.3 764.2
33 0.4648 0.5938 0.000000 0.0000 1139.4 745.0
34 0.3320 0.4531 0.000000 0.0000 1286.7 851.7
35 0.3926 0.5312 0.000000 0.0000 1082.6 654.5
36 0.4883 0.6094 0.000000 -0.0000 1257.0 783.3
37 0.4453 0.5781 0.000000 -0.0000 917.8 487.1
38 0.5215 0.7344 0.000000 0.0000 983.8 575.6
39 0.4453 0.5312 0.000000 0.0000 1108.1 653.2
40 0.4668 0.5781 0.000000 0.0000 1095.7 667.8
41 0.5020 0.6406 0.000000 -0.0000 1160.8 701.5
42 0.4355 0.5156 0.000000 0.0000 1078.5 675.0
43 0.4668 0.5781 0.000000 -0.0000 1029.6 626.0
44 0.3809 0.4531 0.000000 0.0000 951.2 530.5
45 0.4258 0.5625 0.000000 -0.0000 940.1 511.0
46 0.5352 0.6719 0.000000 0.0000 1020.4 588.9
47 0.4590 0.5938 0.000000 0.0000 1061.9 612.4
48 0.5137 0.5625 0.000000 0.0000 887.2 467.1
49 0.3320 0.4375 0.000000 0.0000 1247.9 782.2
50 0.3730 0.4688 0.000000 0.0000 1233.0 751.7
51 0.3828 0.4844 0.000000 -0.0000 1069.2 617.3

explore-tis-minp_736733

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
51 0.4941 0.5789 0.000000 -0.0000 3508.9 3081.7
52 0.3027 0.4219 0.000000 0.0000 1919.1 1439.8
53 0.3047 0.4688 0.000000 0.0000 770.9 260.2
54 0.3984 0.5156 0.000000 -0.0000 1262.6 822.0
55 0.3906 0.5000 0.000000 0.0000 1294.0 839.9
56 0.4004 0.5000 0.000000 0.0000 1192.4 706.6
57 0.3828 0.5156 0.000000 -0.0000 997.6 536.4
58 0.4512 0.5781 0.000000 0.0000 1202.9 731.3
59 0.4219 0.5000 0.000000 -0.0000 1097.7 667.5
60 0.5391 0.6406 0.000000 -0.0000 888.0 455.3
61 0.3984 0.5469 0.000000 -0.0000 1009.8 558.5
62 0.5293 0.6094 0.000000 -0.0000 1031.2 561.3
63 0.4570 0.5312 0.000000 0.0000 948.6 505.8
64 0.4141 0.5312 0.000000 -0.0000 886.4 422.0
65 0.4023 0.4688 0.000000 0.0000 1000.0 530.1
66 0.3496 0.5312 0.000000 0.0000 1109.5 627.0
67 0.4414 0.5781 0.000000 -0.0000 1090.9 606.6
68 0.4414 0.5469 0.000000 -0.0000 1095.0 613.0
69 0.4551 0.5938 0.000000 0.0000 886.6 403.9
70 0.3594 0.4531 0.000000 0.0000 1316.9 824.5
71 0.4648 0.6094 0.000000 0.0000 1442.1 990.6
72 0.4121 0.5156 0.000000 0.0000 1035.3 560.2
73 0.2500 0.3438 0.000000 0.0000 2000.9 1570.8
74 0.4219 0.5312 0.000000 0.0000 1155.8 707.4
75 0.3848 0.5625 0.000000 0.0000 1615.8 1187.4
76 0.3516 0.4844 0.000000 0.0000 868.2 442.2
77 0.3770 0.5000 0.000000 -0.0000 1422.9 971.2
78 0.2168 0.3750 0.000000 -0.0000 1385.8 923.5
79 0.7500 0.8438 0.000000 0.0000 2052.9 1766.5
80 0.5918 0.6875 0.000000 0.0000 1231.2 832.3

explore-tis-minp_736734

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
82 0.7598 0.8281 0.000000 0.0000 1584.8 1301.6

explore-tis-minp_736735

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
84 0.7676 0.8281 0.000000 0.0000 1825.9 1556.5

explore-tis-minp_736736

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
86 0.7559 0.8281 0.000000 -0.0000 1769.7 1461.4

Timing Analysis

Average Time Breakdown (% of step time)

Component Avg % of Step Time
wait_for_generation_buffer 61.1%
run_training 35.8%
train_critic_and_policy 30.8%
policy_train 30.8%
fwd_logprobs_values_reward 5.0%
sync_weights 2.5%
save_checkpoints 0.9%
save_hf_model 0.7%
convert_to_training_input 0.6%
cleanup_old_checkpoints 0.1%
compute_advantages_and_returns 0.0%

Cross-Log Comparison

Log Avg Reward Pass@8 Step Time (s) Gen Wait Time (s) Avg Tokens Staleness
explore-tis-minp_736731 0.4516 0.5865 1280.5809 875.1822 7340.1784 2.0042
explore-tis-minp_736732 0.4319 0.5504 1141.1122 712.7736 7798.1691 2.2454
explore-tis-minp_736733 0.4185 0.5354 1290.6596 838.1885 8339.4156 2.1802
explore-tis-minp_736734 0.7598 0.8281 1584.7922 1301.6032 5398.6328 0.0000
explore-tis-minp_736735 0.7676 0.8281 1825.9347 1556.4501 5063.8457 0.0000
explore-tis-minp_736736 0.7559 0.8281 1769.7243 1461.3722 5654.4453 0.0000

vLLM Inference Engine Analysis

Metrics from vLLM stat loggers (V1LoggingStatLoggerFixed).

Note

: Ray deduplicates similar log messages with [repeated Nx across cluster], so we typically capture stats from one engine per timestamp. The stats shown are per-engine values. Multiply by num_inference_engines for cluster-wide estimates.

Summary by Log (Per-Engine Stats)

Log Avg Running/Engine Avg Waiting/Engine Avg Gen Throughput/Engine Avg KV Cache % Avg Prefix Hit %
explore-tis-minp_736731 8.7 0.0 374.0 tok/s 36.1% 86.3%
explore-tis-minp_736732 9.7 0.1 418.4 tok/s 40.8% 87.8%
explore-tis-minp_736733 8.4 0.0 359.7 tok/s 34.5% 90.5%
explore-tis-minp_736734 8.3 0.0 351.3 tok/s 28.5% 89.3%
explore-tis-minp_736735 7.8 0.0 338.9 tok/s 25.9% 91.0%
explore-tis-minp_736736 7.1 0.0 322.7 tok/s 24.8% 90.4%

Utilization Analysis (Per-Engine)

Key indicators of inference engine utilization:

  • Running requests/engine: Concurrent requests being processed by each engine
  • Waiting requests: Requests queued (0 = engine not saturated, has spare capacity)
  • Generation throughput: Decode tokens/sec per engine
    • 8B model on H100 can do 1000+ tok/s when saturated
    • If seeing <300 tok/s with 0 waiting, engine is starved for requests

explore-tis-minp_736731

  • Running requests/engine: avg=8.7, max=37
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=374.0 tok/s, max=1500.2 tok/s
  • KV cache usage: avg=36.1%
  • Prefix cache hit rate: avg=86.3%
  • ℹ️ Moderate utilization

explore-tis-minp_736732

  • Running requests/engine: avg=9.7, max=35
  • Waiting requests: avg=0.1, max=24
  • Generation throughput/engine: avg=418.4 tok/s, max=1291.3 tok/s
  • KV cache usage: avg=40.8%
  • Prefix cache hit rate: avg=87.8%
  • ✅ Well-utilized: Engines saturated (waiting > 0)

explore-tis-minp_736733

  • Running requests/engine: avg=8.4, max=34
  • Waiting requests: avg=0.0, max=15
  • Generation throughput/engine: avg=359.7 tok/s, max=1226.3 tok/s
  • KV cache usage: avg=34.5%
  • Prefix cache hit rate: avg=90.5%
  • ✅ Well-utilized: Engines saturated (waiting > 0)

explore-tis-minp_736734

  • Running requests/engine: avg=8.3, max=16
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=351.3 tok/s, max=725.5 tok/s
  • KV cache usage: avg=28.5%
  • Prefix cache hit rate: avg=89.3%
  • ℹ️ Moderate utilization

explore-tis-minp_736735

  • Running requests/engine: avg=7.8, max=16
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=338.9 tok/s, max=755.8 tok/s
  • KV cache usage: avg=25.9%
  • Prefix cache hit rate: avg=91.0%
  • ℹ️ Moderate utilization

explore-tis-minp_736736

  • Running requests/engine: avg=7.1, max=19
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=322.7 tok/s, max=917.5 tok/s
  • KV cache usage: avg=24.8%
  • Prefix cache hit rate: avg=90.4%
  • ℹ️ Moderate utilization