Qwen3-30B-A3B Agentic-ESOpt Math (theta20)

This is the full-parameter BF16 checkpoint obtained by deterministically replaying the first 20 Agentic-ESOpt updates (generations 0–19) on Qwen3-30B-A3B.

Reported result

The source evaluation generated 16 samples per problem. The reported Mean4/Pass4 row is the group with sample_index 12–15. Mean4 is the mean score over those four samples; Pass4 is the mean per-problem maximum over those four samples. Exact unrounded values and the source evaluation 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: 20 (generations 0–19)
• 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_theta20.json is the exact history prefix used to construct this checkpoint.

Reproduction files

• repro/es_history_theta20.json: configuration plus the 20 replayed updates
• repro/metrics.json: exact metric values and aggregation details
• repro/dapo_eval16.json: DAPO 16-sample evaluation records
• repro/aime2026_eval16.json: AIME2026 16-sample evaluation records