Qwen3.5-4B Sudoku Mask15 Agentic-ESOpt

This repository contains the exact full-parameter checkpoint corresponding to the released Sudoku Mask15 Agentic-ESOpt log in Agentic-ESOpt.

Training and evaluation setting

Base model: Qwen/Qwen3.5-4B
Method: Agentic-ESOpt (ES32)
Generations: 100 (0 through 99)
Population: 32; case batch: 32
Full-parameter update; alpha 5e-4; z-score reward normalization
Exploration sigma: cosine schedule from 7e-4 to 5e-4, zero warmup
Agent horizon: 45 turns; 64 generated tokens per turn
Sampling: temperature 0.7, top-p 0.8, top-k 20, min-p 0.0, presence penalty 1.5, repetition penalty 1.0
Evaluation: 32 held-out Mask15 cases, three repeats
Final evaluation: 17.00/32 = 53.125% (standard deviation 0.025516)

The published replay history uses generations 0–89 from the main cosine run and generations 90–99 from the fourth independent final-ten rerun starting at the generation-89 checkpoint. The included es_replay_export_manifest.json records all 100 applied updates; its reward sequence matches the released log generation by generation.

The model is provided as standard Transformers safetensors weights together with its tokenizer and generation configuration.