laion/a3-rl-DCAgent_code-contests-noblock-5-8B
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
main
a3-rl-DCAgent_code-contests…/training_logs/20260525_110326_metrics_report.md
ModelHub XC a508caa68d 初始化项目，由ModelHub XC社区提供模型 
Model: laion/a3-rl-DCAgent_code-contests-noblock-5-8B
Source: Original Platform
2026-08-07 12:29:36 +08:00

22 KiB
Raw Permalink 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)
job_485099 5 5 0.0621 0.1152 23691.2
job_485100 13 9 0.0610 0.1465 39612.5
job_485101 19 7 0.0948 0.1230 31256.0
job_496420 26 8 0.0935 0.2188 40891.0
job_496421 36 10 0.1310 0.2422 42300.2
job_496422 45 9 0.1430 0.2285 42232.9

Async Metrics

Mean Std Min Max Count
async/discard_rate 0 0 0 0 48
async/discarded_count 0 0 0 0 48
async/effective_batch_groups 64 0 64 64 48
async/effective_batch_samples 512 0 512 512 48
async/staleness_max 1.91667 0.794485 0 4 48
async/staleness_mean 0.861646 0.370024 0 1.5625 48
async/staleness_min 0.208333 0.410414 0 1 48
async/staleness_ratio 0.69629 0.320514 0 1 48

Generate Metrics

Mean Std Min Max Count
generate/avg_num_tokens 300.937 132.434 53.543 653.519 48
generate/avg_tokens_non_zero_rewards 3164.85 608.453 2069.75 4650.14 48
generate/avg_tokens_zero_rewards 1 0 1 1 48
generate/max_num_tokens 9563.94 4742.53 4538 31001 48
generate/min_num_tokens 1 0 1 1 48
generate/std_num_tokens 1025.02 261.051 409.607 1779.2 48

Loss Metrics

Mean Std Min Max Count
loss/avg_final_rewards 0.101433 0.0552715 0.0137 0.2422 48
loss/avg_raw_advantages 0.223456 0.116288 0.0251 0.4541 48
loss/avg_raw_advantages_abs 0.223596 0.116357 0.0252 0.4545 48

Policy Metrics

Mean Std Min Max Count
policy/final_loss -0.000310417 0.000183627 -0.0008 -0 48
policy/log_ratio_abs_max 0 0 0 0 48
policy/log_ratio_abs_mean 0 0 0 0 48
policy/log_ratio_abs_p99 0 0 0 0 48
policy/log_ratio_abs_pos00 0 0 0 0 48
policy/log_ratio_abs_pos10 0 0 0 0 48
policy/log_ratio_abs_pos20 0 0 0 0 48
policy/log_ratio_abs_pos30 0 0 0 0 48
policy/log_ratio_abs_pos40 0 0 0 0 48
policy/log_ratio_abs_pos50 0 0 0 0 48
policy/log_ratio_abs_pos60 0 0 0 0 48
policy/log_ratio_abs_pos70 0 0 0 0 48
policy/log_ratio_abs_pos80 0 0 0 0 48
policy/log_ratio_abs_pos90 0 0 0 0 48
policy/n_tokens_dp_gt_10pct 0 0 0 0 48
policy/n_tokens_dp_gt_1pct 0 0 0 0 48
policy/n_tokens_dp_gt_50pct 0 0 0 0 48
policy/policy_entropy 0.00732917 0.0036161 0.0018 0.0182 48
policy/policy_loss -0.0198583 0.0114544 -0.0498 -0.002 48
policy/policy_lr 0 0 0 0 48
policy/policy_update_steps 1 0 1 1 48
policy/ppo_clip_ratio 0 0 0 0 48
policy/raw_grad_norm 0.0109021 0.00505137 0.0028 0.0244 48

Reward Metrics

Mean Std Min Max Count
reward/avg_pass_at_8 0.261562 0.0939986 0.0938 0.4898 48
reward/avg_raw_reward 0.101433 0.0552715 0.0137 0.2422 48

System Metrics

Mean Std Min Max Count
system/process_rss_gb 9.5798 3.43573 5.3781 17.2552 48
system/process_vms_gb 81.1127 5.44727 78.0531 94.4252 48
system/ram_available_gb 606.245 15.2876 574.636 627.878 48
system/ram_percent 29.3396 1.78847 26.8 33 48
system/ram_total_gb 857.969 0.00033958 857.968 857.969 48
system/ram_used_gb 251.724 15.2877 230.09 283.333 48

Timing Metrics

Mean Std Min Max Count
timing/compute_advantages_and_returns 0.0384646 0.0348721 0.0175 0.1923 48
timing/convert_to_training_input 5.4315 26.8172 0.8755 187.298 48
timing/fwd_logprobs_values_reward 7.50122 2.76571 3.0976 15.4184 48
timing/policy_train 59.4245 7.15104 44.4851 76.4064 48
timing/run_training 67.108 9.54666 48.0435 90.3725 48
timing/step 4583 2394.44 351.854 8192.84 48
timing/sync_weights 18.6968 1.59914 17.1829 28.5894 48
timing/train_critic_and_policy 59.5679 7.15635 44.6233 76.5655 48
timing/wait_for_generation_buffer 4491.76 2392.46 271.614 8105.89 48
timing/cleanup_old_checkpoints 4.13983 7.40386 0.0075 24.0683 22
timing/save_checkpoints 15.0911 11.7658 7.3633 44.8581 22
timing/save_hf_model 6.85727 1.37645 5.8497 10.5996 10

Trainer Metrics

Mean Std Min Max Count
trainer/epoch 0 0 0 0 48
trainer/global_step 22.3333 13.0536 1 45 48

Batch_Errors Metrics

Mean Std Min Max Count
batch_errors/total_batches 59.6875 18.9346 2 128 48
batch_errors/total_instances 477.5 151.476 16 1024 48
batch_errors/total_successful 43.8333 25.6634 0 117 48
batch_errors/total_failed 59.0208 18.8617 2 128 48
batch_errors/total_masked 357.688 113.847 15 732 48
batch_errors/avg_DaytonaAuthenticationError 1.00014 0.399696 0.28125 1.76562 48
batch_errors/total_DaytonaAuthenticationError 62.2708 36.1967 1 208 48
batch_errors/avg_AgentEnvironmentTimeoutError 6.00745 0.573894 4.82812 7.5 48
batch_errors/total_AgentEnvironmentTimeoutError 355.104 112.817 15 725 48
batch_errors/avg_DaytonaNotFoundError 0.0269537 0.0189922 0.0144928 0.09375 32
batch_errors/total_DaytonaNotFoundError 1.75 1.34404 1 6 32
batch_errors/avg_VerifierRuntimeError 0.20424 0.127208 0.03125 0.5625 46
batch_errors/total_VerifierRuntimeError 12.5217 8.05188 2 36 46
batch_errors/avg_DaytonaError 0.0224232 0.0110744 0.0147059 0.046875 17
batch_errors/total_DaytonaError 1.52941 0.799816 1 3 17
batch_errors/avg_RuntimeError 0.015625 nan 0.015625 0.015625 1
batch_errors/total_RuntimeError 2 nan 2 2 1
batch_errors/avg_DaytonaValidationError 0.0078125 nan 0.0078125 0.0078125 1
batch_errors/total_DaytonaValidationError 1 nan 1 1 1
batch_errors/avg_ContextLengthExceededError 0.0284136 0.0238451 0.0078125 0.1 13
batch_errors/total_ContextLengthExceededError 1.61538 0.767948 1 3 13
batch_errors/avg_AgentTimeoutError 0.0445227 0.0358926 0.015625 0.1875 27
batch_errors/total_AgentTimeoutError 2.81481 2.30446 1 12 27
batch_errors/avg_AgentSetupTimeoutError 0.0386788 0.0257621 0.0147059 0.078125 6
batch_errors/total_AgentSetupTimeoutError 2.33333 1.75119 1 5 6
batch_errors/avg_EnvironmentStartTimeoutError 0.0153846 nan 0.0153846 0.0153846 1
batch_errors/total_EnvironmentStartTimeoutError 1 nan 1 1 1
batch_errors/avg_VerifierTimeoutError 0.0400638 0.0229369 0.0144928 0.0588235 3
batch_errors/total_VerifierTimeoutError 2.66667 1.52753 1 4 3
batch_errors/avg_OSError 0.0144928 nan 0.0144928 0.0144928 1
batch_errors/total_OSError 1 nan 1 1 1

Training Progression by Log

job_485099

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
1 0.0410 0.1719 0.000000 -0.0003 8192.8 8105.9
2 0.0664 0.2344 0.000000 -0.0002 7137.7 7051.2
3 0.0742 0.2656 0.000000 -0.0001 614.4 526.3
4 0.0137 0.0938 0.000000 -0.0000 7011.6 6938.8
5 0.1152 0.2344 0.000000 -0.0001 734.7 652.9

job_485100

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
5 0.1465 0.3276 0.000000 -0.0005 8132.7 8036.1
6 0.0332 0.1562 0.000000 -0.0003 3465.1 3390.7
7 0.0234 0.1406 0.000000 -0.0002 3987.0 3915.0
8 0.0176 0.0938 0.000000 -0.0001 6656.3 6578.8
9 0.0234 0.1250 0.000000 -0.0001 1287.2 1213.6
10 0.0820 0.2344 0.000000 -0.0005 6403.6 6309.4
11 0.0781 0.1719 0.000000 -0.0002 1527.3 1417.6
12 0.0801 0.2500 0.000000 -0.0004 6140.9 6053.0
13 0.0645 0.2344 0.000000 -0.0002 2012.4 1931.4

job_485101

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
13 0.1230 0.2742 0.000000 -0.0004 8113.5 8023.7
14 0.0371 0.1250 0.000000 -0.0002 6886.3 6818.0
15 0.1191 0.3750 0.000000 -0.0005 791.9 701.5
16 0.1191 0.3125 0.000000 -0.0003 6575.8 6487.1
17 0.0918 0.2500 0.000000 -0.0005 920.2 834.8
18 0.0703 0.2344 0.000000 -0.0002 6610.6 6533.4
19 0.1035 0.2812 0.000000 -0.0001 1357.7 1273.3

job_496420

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
19 0.2188 0.4407 0.000000 -0.0007 8067.2 7961.5
20 0.0801 0.2031 0.000000 -0.0003 3380.6 3297.4
21 0.0996 0.3281 0.000000 -0.0005 4118.8 4027.4
22 0.0703 0.2500 0.000000 -0.0004 5101.1 5011.8
23 0.0938 0.2656 0.000000 -0.0003 2467.0 2383.2
24 0.0781 0.2344 0.000000 -0.0001 6306.5 6217.1
25 0.0547 0.2188 0.000000 -0.0002 4522.1 4243.7
26 0.0527 0.1875 0.000000 -0.0000 6927.7 6847.7

job_496421

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
27 0.2422 0.4898 0.000000 -0.0007 8113.5 8003.3
28 0.0586 0.2500 0.000000 -0.0005 351.9 271.6
29 0.1211 0.2656 0.000000 -0.0004 7096.6 7009.6
30 0.0527 0.1875 0.000000 -0.0001 2113.7 2031.7
31 0.1680 0.3750 0.000000 -0.0004 5328.9 5229.5
32 0.1035 0.2031 0.000000 -0.0002 3341.1 3249.0
33 0.1777 0.3438 0.000000 -0.0004 4248.8 4152.3
34 0.0918 0.1875 0.000000 -0.0002 3744.8 3661.6
35 0.1133 0.2188 0.000000 -0.0002 3951.6 3868.0
36 0.1816 0.2969 0.000000 -0.0003 4009.3 3927.9

job_496422

Step Reward Pass@8 KL Loss Step Time (s) Gen Wait (s)
37 0.1953 0.4444 0.000000 -0.0005 8145.7 8035.3
38 0.0781 0.2344 0.000000 -0.0003 3328.9 3242.5
39 0.1680 0.4375 0.000000 -0.0008 6397.1 6310.3
40 0.1191 0.3281 0.000000 -0.0004 5420.0 5330.6
41 0.1211 0.2812 0.000000 -0.0002 2226.2 2142.9
42 0.1523 0.3281 0.000000 -0.0003 4730.9 4645.8
43 0.2285 0.4688 0.000000 -0.0006 3828.9 3733.7
44 0.1113 0.2656 0.000000 -0.0002 4075.9 3983.0
45 0.1133 0.2344 0.000000 -0.0003 4079.3 3993.7

Timing Analysis

Average Time Breakdown (% of step time)

Component Avg % of Step Time
wait_for_generation_buffer 96.5%
run_training 2.6%
train_critic_and_policy 2.3%
policy_train 2.3%
sync_weights 0.7%
save_checkpoints 0.5%
save_hf_model 0.4%
fwd_logprobs_values_reward 0.3%
cleanup_old_checkpoints 0.2%
convert_to_training_input 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
job_485099 0.0621 0.2000 4738.2360 4655.0062 202.7762 0.8125
job_485100 0.0610 0.1927 4401.3874 4316.1800 213.5232 0.7847
job_485101 0.0948 0.2646 4465.1476 4381.6743 292.8345 0.9732
job_496420 0.0935 0.2660 5111.3807 4998.7086 309.4238 0.8438
job_496421 0.1310 0.2818 4230.0172 4140.4545 357.4424 0.7937
job_496422 0.1430 0.3358 4692.5499 4601.9845 378.8585 0.9705

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_485099 3.8 0.0 175.2 tok/s 16.5% 78.6%
job_485100 3.6 0.0 165.7 tok/s 15.7% 81.1%
job_485101 3.7 0.0 168.8 tok/s 15.6% 80.5%
job_496420 3.2 0.0 145.1 tok/s 13.9% 79.2%
job_496421 3.8 0.0 173.4 tok/s 17.0% 77.6%
job_496422 3.2 0.0 145.2 tok/s 13.7% 74.7%

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_485099

  • Running requests/engine: avg=3.8, max=15
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=175.2 tok/s, max=700.8 tok/s
  • KV cache usage: avg=16.5%
  • Prefix cache hit rate: avg=78.6%
  • ⚠️ Underutilized: Engines starved for requests (0 waiting, avg 3.8 running)
    • Bottleneck is likely upstream (environment execution, not inference)

job_485100

  • Running requests/engine: avg=3.6, max=17
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=165.7 tok/s, max=682.9 tok/s
  • KV cache usage: avg=15.7%
  • Prefix cache hit rate: avg=81.1%
  • ⚠️ Underutilized: Engines starved for requests (0 waiting, avg 3.6 running)
    • Bottleneck is likely upstream (environment execution, not inference)

job_485101

  • Running requests/engine: avg=3.7, max=13
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=168.8 tok/s, max=601.8 tok/s
  • KV cache usage: avg=15.6%
  • Prefix cache hit rate: avg=80.5%
  • ⚠️ Underutilized: Engines starved for requests (0 waiting, avg 3.7 running)
    • Bottleneck is likely upstream (environment execution, not inference)

job_496420

  • Running requests/engine: avg=3.2, max=11
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=145.1 tok/s, max=485.1 tok/s
  • KV cache usage: avg=13.9%
  • Prefix cache hit rate: avg=79.2%
  • ⚠️ Underutilized: Engines starved for requests (0 waiting, avg 3.2 running)
    • Bottleneck is likely upstream (environment execution, not inference)

job_496421

  • Running requests/engine: avg=3.8, max=15
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=173.4 tok/s, max=665.8 tok/s
  • KV cache usage: avg=17.0%
  • Prefix cache hit rate: avg=77.6%
  • ⚠️ Underutilized: Engines starved for requests (0 waiting, avg 3.8 running)
    • Bottleneck is likely upstream (environment execution, not inference)

job_496422

  • Running requests/engine: avg=3.2, max=16
  • Waiting requests: avg=0.0, max=0
  • Generation throughput/engine: avg=145.2 tok/s, max=682.5 tok/s
  • KV cache usage: avg=13.7%
  • Prefix cache hit rate: avg=74.7%
  • ⚠️ Underutilized: Engines starved for requests (0 waiting, avg 3.2 running)
    • Bottleneck is likely upstream (environment execution, not inference)

Trial-Level Analysis (from result.json)

Total trials parsed: 25855

Turn Count Statistics

Metric Value
Mean 1.7
Median 1.0
Std 1.3
Min 1
Max 33
Count 25830

Exception Distribution

Exception Type Count %
AgentEnvironmentTimeoutError 18399 71.2%
DaytonaAuthenticationError 3597 13.9%
No exception 2770 10.7%
VerifierRuntimeError 778 3.0%
AgentTimeoutError 158 0.6%
DaytonaNotFoundError 72 0.3%
DaytonaError 29 0.1%
ContextLengthExceededError 23 0.1%
AgentSetupTimeoutError 17 0.1%
VerifierTimeoutError 8 0.0%
RuntimeError 2 0.0%
DaytonaValidationError 1 0.0%
EnvironmentStartTimeoutError 1 0.0%

Turn Count by Exception Type

Exception Type Mean Turns Median Turns Count
AgentTimeoutError 9.7 8.0 158
No exception 3.2 3.0 2770
VerifierRuntimeError 3.1 3.0 778
VerifierTimeoutError 2.6 2.0 8
DaytonaValidationError 2.0 2.0 1
DaytonaNotFoundError 1.7 2.0 72
DaytonaError 1.6 2.0 24
AgentEnvironmentTimeoutError 1.5 1.0 18399
ContextLengthExceededError 1.4 1.0 23
DaytonaAuthenticationError 1.2 1.0 3597

Turn Count by Outcome

Outcome Mean Turns Median Turns Count
Success 3.2 3.0 4834

Reward Summary

  • Mean reward: 1.0000
  • Success rate: 100.0%
  • Trials with reward data: 4834