Qwen3-30B-A3B Agentic-ESOpt DocVQA (theta40)
This is the full-parameter BF16 checkpoint obtained by deterministically replaying all 40 Agentic-ESOpt updates (generations 0–39) on Qwen3-30B-A3B.
Reported result
| Metric | Mean4 | Pass4 |
|---|---|---|
| DocVQA ANLS | 0.3653 | 0.5078 |
| DocVQA Accuracy | 39.0 | 53.0 |
The source evaluation generated 16 samples per problem. The reported Mean4/Pass4 row is the group with sample_index 12–15. ANLS Mean4 is the mean ANLS over those samples; ANLS Pass4 is the mean per-problem maximum ANLS. Accuracy Mean4 is exact-match accuracy over all four samples; Accuracy Pass4 is the fraction of problems with at least one exact match. Exact unrounded values and source records are included under repro/.
Replay configuration
- Base model:
Qwen/Qwen3-30B-A3B - Parameter scope: full (30,532,122,624 floating-point parameters)
- Replayed updates: 40 (generations 0–39)
- Population: 16
- Alpha: 0.0005
- Reward normalization: z-score (
ddof=0,eps=1e-8) - Training sigma schedule: cosine, 0.001 to 0.0005
- Export dtype: BF16
es_replay_export_manifest.json records every replayed generation and verifies that the recomputed normalization weights exactly equal the weights stored in the history. repro/es_history_theta40.json is the exact history used to construct this checkpoint.
Reproduction files
repro/es_history_theta40.json: configuration plus all 40 replayed updatesrepro/metrics.json: exact metric values and aggregation detailsrepro/docvqa_eval16.json: DocVQA 16-sample evaluation records