laion/a3-rl-DCAgent_selfinstruct-naive-sandboxes-2-verified-70-8B
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
3e52217125c5998e2e7177659dc62fb58bef4acf
a3-rl-DCAgent_selfinstruct-…/training_logs/20260604_123153_metrics_report.md
ModelHub XC 3e52217125 初始化项目，由ModelHub XC社区提供模型 
Model: laion/a3-rl-DCAgent_selfinstruct-naive-sandboxes-2-verified-70-8B
Source: Original Platform
2026-08-07 10:27:18 +08:00

22 KiB
Raw Blame History

SkyRL Training Metrics Analysis

Generated from 3 log files

Overview

Log File Total Steps Metric Blocks Final Reward (mean) Final Reward (max) Total Time (s)
job_567548 9 9 0.4008 0.5059 6840.2
job_567550 50 42 0.3437 0.4668 38753.5
job_567551 80 30 0.3892 0.5586 25084.9

Async Metrics

Mean Std Min Max Count
async/discard_rate 0 0 0 0 81
async/discarded_count 0 0 0 0 81
async/effective_batch_groups 64 0 64 64 81
async/effective_batch_samples 512 0 512 512 81
async/staleness_max 5.48148 1.62874 0 8 81
async/staleness_mean 2.52044 0.596524 0 3.2344 81
async/staleness_min 0.135802 0.344713 0 1 81
async/staleness_ratio 0.921498 0.127923 0 1 81

Generate Metrics

Mean Std Min Max Count
generate/avg_num_tokens 6219.84 593.767 3631.05 7508.99 81
generate/avg_tokens_non_zero_rewards 6723.53 885.765 3316.3 8407.4 81
generate/avg_tokens_zero_rewards 5945.13 669.261 3953.26 7989.15 81
generate/max_num_tokens 30304.3 1297.57 25303 31890 81
generate/min_num_tokens 1 0 1 1 81
generate/std_num_tokens 6150.02 623.638 3243.07 7148.25 81

Loss Metrics

Mean Std Min Max Count
loss/avg_final_rewards 0.36687 0.0610739 0.2559 0.5586 81
loss/avg_raw_advantages 0.00432716 0.00984494 -0.0145 0.0461 81
loss/avg_raw_advantages_abs 0.160136 0.0385764 0.0852 0.2488 81

Policy Metrics

Mean Std Min Max Count
policy/final_loss -1.48148e-05 4.50309e-05 -0.0003 0 81
policy/log_ratio_abs_max 0 0 0 0 81
policy/log_ratio_abs_mean 0 0 0 0 81
policy/log_ratio_abs_p99 0 0 0 0 81
policy/log_ratio_abs_pos00 0 0 0 0 81
policy/log_ratio_abs_pos10 0 0 0 0 81
policy/log_ratio_abs_pos20 0 0 0 0 81
policy/log_ratio_abs_pos30 0 0 0 0 81
policy/log_ratio_abs_pos40 0 0 0 0 81
policy/log_ratio_abs_pos50 0 0 0 0 81
policy/log_ratio_abs_pos60 0 0 0 0 81
policy/log_ratio_abs_pos70 0 0 0 0 81
policy/log_ratio_abs_pos80 0 0 0 0 81
policy/log_ratio_abs_pos90 0 0 0 0 81
policy/n_tokens_dp_gt_10pct 0 0 0 0 81
policy/n_tokens_dp_gt_1pct 0 0 0 0 81
policy/n_tokens_dp_gt_50pct 0 0 0 0 81
policy/policy_entropy 0.135762 0.0098146 0.1121 0.1546 81
policy/policy_loss -0.00113086 0.00263399 -0.0202 0 81
policy/policy_lr 0 0 0 0 81
policy/policy_update_steps 1 0 1 1 81
policy/ppo_clip_ratio 0 0 0 0 81
policy/raw_grad_norm 0.0180111 0.00588504 0.0124 0.064 81

Reward Metrics

Mean Std Min Max Count
reward/avg_pass_at_8 0.522458 0.0740283 0.3438 0.6875 81
reward/avg_raw_reward 0.36687 0.0610739 0.2559 0.5586 81

System Metrics

Mean Std Min Max Count
system/process_rss_gb 15.5018 1.58876 10.6846 17.8517 81
system/process_vms_gb 52.6829 1.43311 50.2918 54.9248 81
system/ram_available_gb 620.096 12.2909 588.791 634.891 81
system/ram_percent 27.7247 1.43436 26 31.4 81
system/ram_total_gb 857.968 0 857.968 857.968 81
system/ram_used_gb 237.871 12.2909 223.076 269.176 81

Timing Metrics

Mean Std Min Max Count
timing/compute_advantages_and_returns 0.0666617 0.0122347 0.0519 0.1531 81
timing/convert_to_training_input 4.72614 0.280405 3.6373 5.4807 81
timing/fwd_logprobs_values_reward 49.3127 5.50948 28.9658 61.4679 81
timing/policy_train 290.478 30.592 151.591 354.228 81
timing/run_training 340.145 35.2621 180.815 408.019 81
timing/step 872.574 400.434 507.833 3374.16 81
timing/sync_weights 20.0706 1.03089 17.0362 23.2754 81
timing/train_critic_and_policy 290.765 30.6031 151.737 354.553 81
timing/wait_for_generation_buffer 507.629 406.806 131.502 3072.09 81
timing/cleanup_old_checkpoints 2.02637 7.98054 0.0063 38.5904 40
timing/save_checkpoints 23.1826 62.9803 7.7944 396.822 40
timing/save_hf_model 6.37928 0.425326 5.9537 7.5975 16

Trainer Metrics

Mean Std Min Max Count
trainer/epoch 0 0 0 0 81
trainer/global_step 40.1111 23.3559 1 80 81

Batch_Errors Metrics

Mean Std Min Max Count
batch_errors/total_batches 60.9136 8.52379 42 121 81
batch_errors/total_instances 487.309 68.1903 336 968 81
batch_errors/total_successful 403.728 74.9498 283 939 81
batch_errors/total_failed 19.8642 5.64525 5 32 81
batch_errors/total_masked 79.2716 25.8404 25 134 81
batch_errors/avg_RewardFileNotFoundError 0.531055 0.271125 0.03125 1.42593 81
batch_errors/total_RewardFileNotFoundError 31.8272 15.4934 2 77 81
batch_errors/avg_DaytonaValidationError 0.0672515 0.0494668 0.0106383 0.186441 24
batch_errors/total_DaytonaValidationError 4 2.87417 1 11 24
batch_errors/avg_ContextLengthExceededError 0.779734 0.294334 0.16129 1.55556 79
batch_errors/total_ContextLengthExceededError 46.9873 18.3376 10 98 79
batch_errors/avg_RuntimeError 0.0324855 0.023494 0.015625 0.101695 46
batch_errors/total_RuntimeError 1.97826 1.40616 1 6 46
batch_errors/avg_DaytonaError 0.0612569 0.0447787 0.0106383 0.206897 52
batch_errors/total_DaytonaError 3.65385 2.65599 1 12 52
batch_errors/avg_AgentTimeoutError 0.0810606 0.0876512 0.015625 0.625 66
batch_errors/total_AgentTimeoutError 4.74242 4.54147 1 30 66
batch_errors/avg_VerifierTimeoutError 0.0785312 0.05595 0.015873 0.225806 23
batch_errors/total_VerifierTimeoutError 4.65217 3.2975 1 14 23
batch_errors/avg_VerifierOutputParseError 0.137931 0 0.137931 0.137931 2
batch_errors/total_VerifierOutputParseError 8 0 8 8 2
batch_errors/avg_DaytonaNotFoundError 0.0559985 0.0495709 0.015873 0.145455 9
batch_errors/total_DaytonaNotFoundError 3.22222 2.68225 1 8 9
batch_errors/avg_InvalidChatHistory 0.0198765 0.00788726 0.015625 0.0338983 5
batch_errors/total_InvalidChatHistory 1.2 0.447214 1 2 5
batch_errors/avg_DaytonaAuthenticationError 0.0163291 0.000790025 0.015873 0.0172414 3
batch_errors/total_DaytonaAuthenticationError 1 0 1 1 3
batch_errors/avg_RewardFileEmptyError 0.015873 nan 0.015873 0.015873 1
batch_errors/total_RewardFileEmptyError 1 nan 1 1 1
batch_errors/avg_AddTestsDirError 0.015873 nan 0.015873 0.015873 1
batch_errors/total_AddTestsDirError 1 nan 1 1 1
batch_errors/avg_AgentSetupTimeoutError 0.206955 0.203236 0.016129 0.479167 4
batch_errors/total_AgentSetupTimeoutError 11 9.62635 1 23 4
batch_errors/avg_OSError 0.0164111 0.000760944 0.015873 0.0169492 2
batch_errors/total_OSError 1 0 1 1 2

Training Progression by Log

job_567548

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
1 0.5059 0.6250 0.000000 -0.0000 1358.1 1155.0
2 0.5000 0.6719 0.000000 0.0000 833.5 586.8
3 0.4199 0.5938 0.000000 0.0000 703.1 332.8
4 0.4609 0.6406 0.000000 0.0000 924.0 548.2
5 0.3359 0.4531 0.000000 -0.0001 609.8 230.0
6 0.3613 0.5000 0.000000 -0.0000 528.5 145.4
7 0.3340 0.4844 0.000000 -0.0000 589.5 248.8
8 0.4082 0.5312 0.000000 -0.0000 605.9 228.5
9 0.2812 0.4844 0.000000 0.0000 687.8 372.3

job_567550

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
9 0.3418 0.5556 0.000000 0.0000 3374.2 3072.1
10 0.3867 0.5938 0.000000 0.0000 1168.0 795.3
11 0.2988 0.4531 0.000000 -0.0000 507.8 131.5
12 0.3145 0.4844 0.000000 0.0000 516.4 177.0
13 0.3301 0.4688 0.000000 -0.0000 610.9 302.7
14 0.2559 0.4375 0.000000 -0.0000 839.0 513.9
15 0.3613 0.5000 0.000000 -0.0000 643.1 282.0
16 0.4551 0.5625 0.000000 -0.0000 563.6 251.9
17 0.3164 0.4688 0.000000 -0.0000 640.3 279.4
18 0.3438 0.4531 0.000000 0.0000 762.5 436.3
19 0.3184 0.4688 0.000000 0.0000 836.6 483.8
20 0.2637 0.3750 0.000000 -0.0000 714.5 341.4
21 0.3750 0.5781 0.000000 0.0000 829.4 430.9
22 0.3418 0.5000 0.000000 -0.0000 763.6 362.2
23 0.3262 0.5156 0.000000 0.0000 1049.2 675.6
24 0.2793 0.4219 0.000000 0.0000 704.3 389.0
25 0.3574 0.4844 0.000000 0.0000 868.9 495.4
26 0.3320 0.4688 0.000000 -0.0000 854.4 467.6
27 0.3223 0.5000 0.000000 0.0000 590.1 255.9
28 0.3398 0.4844 0.000000 -0.0000 880.2 456.9
29 0.3613 0.5625 0.000000 0.0000 884.7 516.6
30 0.3281 0.5156 0.000000 0.0000 1129.6 742.6
31 0.2891 0.3750 0.000000 -0.0000 880.0 508.9
32 0.3418 0.5312 0.000000 0.0000 799.6 456.1
33 0.3105 0.4688 0.000000 -0.0000 801.0 438.2
34 0.3457 0.5312 0.000000 -0.0001 946.0 580.7
35 0.4668 0.6875 0.000000 0.0000 848.1 465.5
36 0.3984 0.5625 0.000000 -0.0001 924.9 576.2
37 0.3438 0.5312 0.000000 -0.0000 883.1 530.8
38 0.3926 0.5156 0.000000 -0.0001 1456.7 1081.1
39 0.3887 0.5312 0.000000 -0.0000 854.4 500.4
40 0.3066 0.4531 0.000000 -0.0000 893.6 560.5
41 0.3281 0.4688 0.000000 -0.0003 2041.8 1642.7
42 0.4082 0.5938 0.000000 -0.0001 767.9 466.6
43 0.3691 0.4531 0.000000 -0.0000 869.6 506.1
44 0.3926 0.5781 0.000000 -0.0001 783.8 429.7
45 0.3242 0.4844 0.000000 -0.0000 673.9 300.3
46 0.3359 0.5312 0.000000 -0.0000 1024.1 663.9
47 0.3203 0.4844 0.000000 0.0000 1112.0 706.6
48 0.3535 0.5156 0.000000 -0.0000 928.2 543.8
49 0.4023 0.5938 0.000000 -0.0000 759.6 398.2
50 0.2656 0.3438 0.000000 0.0000 774.0 416.6

job_567551

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
51 0.4785 0.6226 0.000000 -0.0001 2184.4 1845.0
52 0.5586 0.6875 0.000000 -0.0000 1651.9 1245.4
53 0.3867 0.5000 0.000000 0.0000 652.1 292.6
54 0.3047 0.5156 0.000000 0.0000 560.9 186.2
55 0.3281 0.4531 0.000000 0.0000 683.3 332.7
56 0.4629 0.5938 0.000000 -0.0000 854.5 488.2
57 0.4434 0.5781 0.000000 -0.0001 854.4 440.7
58 0.3535 0.5156 0.000000 -0.0000 785.5 392.6
59 0.3613 0.5469 0.000000 -0.0000 643.7 253.1
60 0.3848 0.5469 0.000000 0.0000 674.6 305.4
61 0.3457 0.5156 0.000000 0.0000 684.6 252.5
62 0.3867 0.5312 0.000000 0.0000 624.4 254.2
63 0.3574 0.5938 0.000000 -0.0001 868.7 474.5
64 0.3066 0.4375 0.000000 -0.0000 692.7 263.9
65 0.3711 0.5000 0.000000 -0.0000 634.6 271.4
66 0.3906 0.5469 0.000000 -0.0000 709.9 303.6
67 0.4883 0.6875 0.000000 0.0000 517.5 135.4
68 0.4043 0.6094 0.000000 0.0000 671.0 285.3
69 0.4531 0.6562 0.000000 -0.0000 730.7 348.2
70 0.3359 0.5000 0.000000 -0.0000 928.1 566.1
71 0.3750 0.4688 0.000000 0.0000 809.2 469.7
72 0.4609 0.6562 0.000000 0.0000 775.9 388.3
73 0.3965 0.5625 0.000000 0.0000 736.8 355.8
74 0.2988 0.3594 0.000000 0.0000 639.6 270.0
75 0.3301 0.4531 0.000000 -0.0000 987.1 598.9
76 0.4746 0.5938 0.000000 0.0000 1003.0 623.9
77 0.3789 0.5625 0.000000 -0.0000 827.6 440.6
78 0.3594 0.4688 0.000000 -0.0000 1027.6 645.9
79 0.3223 0.4688 0.000000 -0.0000 864.1 469.0
80 0.3770 0.5156 0.000000 0.0000 806.3 438.0

Timing Analysis

Average Time Breakdown (% of step time)

Component Avg % of Step Time
wait_for_generation_buffer 53.6%
run_training 43.2%
train_critic_and_policy 37.0%
policy_train 36.9%
fwd_logprobs_values_reward 6.3%
save_checkpoints 2.7%
sync_weights 2.6%
save_hf_model 0.8%
convert_to_training_input 0.6%
cleanup_old_checkpoints 0.2%
compute_advantages_and_returns 0.0%

Cross-Log Comparison

Log Avg Reward Pass@8 Step Time (s) Gen Wait Time (s) Avg Tokens Staleness
job_567548 0.4008 0.5538 760.0180 427.5351 5697.2672 2.0174
job_567550 0.3437 0.5021 922.7027 562.6884 6171.1224 2.5595
job_567551 0.3892 0.5416 836.1619 454.5744 6444.8237 2.6167

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 %
job_567548 8.7 0.0 383.7 tok/s 36.5% 91.9%
job_567550 7.6 0.0 347.5 tok/s 33.3% 92.4%
job_567551 9.0 0.0 399.0 tok/s 39.7% 91.1%

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

job_567548

  • Running requests/engine: avg=8.7, max=22
  • Waiting requests: avg=0.0, max=4
  • Generation throughput/engine: avg=383.7 tok/s, max=790.8 tok/s
  • KV cache usage: avg=36.5%
  • Prefix cache hit rate: avg=91.9%
  • ✅ Well-utilized: Engines saturated (waiting > 0)

job_567550

  • Running requests/engine: avg=7.6, max=24
  • Waiting requests: avg=0.0, max=10
  • Generation throughput/engine: avg=347.5 tok/s, max=878.7 tok/s
  • KV cache usage: avg=33.3%
  • Prefix cache hit rate: avg=92.4%
  • ✅ Well-utilized: Engines saturated (waiting > 0)

job_567551

  • Running requests/engine: avg=9.0, max=24
  • Waiting requests: avg=0.0, max=19
  • Generation throughput/engine: avg=399.0 tok/s, max=851.4 tok/s
  • KV cache usage: avg=39.7%
  • Prefix cache hit rate: avg=91.1%
  • ✅ Well-utilized: Engines saturated (waiting > 0)

Trial-Level Analysis (from result.json)

Total trials parsed: 44526

Turn Count Statistics

Metric Value
Mean 11.5
Median 9.0
Std 8.2
Min 1
Max 116
Count 44464

Exception Distribution

Exception Type Count %
No exception 35187 79.0%
ContextLengthExceededError 3811 8.6%
RewardFileNotFoundError 2794 6.3%
AgentTimeoutError 2128 4.8%
DaytonaError 204 0.5%
VerifierTimeoutError 112 0.3%
DaytonaValidationError 103 0.2%
RuntimeError 93 0.2%
AgentSetupTimeoutError 44 0.1%
DaytonaNotFoundError 29 0.1%
VerifierOutputParseError 16 0.0%
DaytonaAuthenticationError 3 0.0%
RewardFileEmptyError 1 0.0%
AddTestsDirError 1 0.0%

Turn Count by Exception Type

Exception Type Mean Turns Median Turns Count
RewardFileEmptyError 21.0 21.0 1
AddTestsDirError 19.0 19.0 1
AgentTimeoutError 17.8 11.0 2128
ContextLengthExceededError 16.4 13.0 3811
VerifierTimeoutError 13.9 11.5 112
RewardFileNotFoundError 13.2 12.0 2794
DaytonaAuthenticationError 12.7 7.0 3
VerifierOutputParseError 11.5 10.0 16
DaytonaError 11.2 8.0 186
RuntimeError 10.7 10.0 93
No exception 10.4 8.0 35187
DaytonaValidationError 8.5 7.0 103
DaytonaNotFoundError 8.0 6.0 29

Turn Count by Outcome

Outcome Mean Turns Median Turns Count
Success 9.8 8.0 16178
Failure 11.6 9.0 20820

Reward Summary

  • Mean reward: 0.4373
  • Success rate: 43.7%
  • Trials with reward data: 36998