| --- |
| license: mit |
| tags: |
| - chess |
| - reinforcement-learning |
| - pytorch |
| --- |
| |
| # exp182 / exp183 — 309M ChessTransformer (A40 run) |
|
|
| ## Models |
|
|
| ### Pretrain (`latest.pt` / `pretrain_step10000.pt`) |
| - **309M** deep-narrow ChessTransformer (96L × 512d, 8 heads) |
| - StrengthenedBoardEncoder + SwiGLU + chess relative bias |
| - Muon pretrain to **step 10,000** / 50k target (batch=192, bf16) |
| - Repo file: uploaded as `pretrain_step10000.pt` and `latest_pretrain.pt` |
| |
| ### RL (`rl_iter010.pt`) |
| - Expert-iteration RL from the step-10k pretrain |
| - Soft MCTS visit targets + hard Stockfish moves (full strength, depth 8) |
| - Checkpoint after **iteration 10** |
|
|
| ## Usage |
| ```python |
| from huggingface_hub import hf_hub_download |
| from chess_inference import load_checkpoint |
| |
| path = hf_hub_download("avewright/exp182-pretrain-309m", "rl_iter010.pt") |
| model = load_checkpoint(path, device="cuda") |
| ``` |
|
|
| Code: https://github.com/avewright/transform |
|
|
