| license: mit | |
| tags: | |
| - chess | |
| - transformer | |
| - recurrent | |
| - policy | |
| - opening | |
| - pytorch | |
| library_name: pytorch | |
| # 99M opening specialist (squares64) | |
| Same **99M** squares64 architecture as | |
| [`avewright/chess-transformer-100m-squares64`](https://huggingface.co/avewright/chess-transformer-100m-squares64), | |
| finetuned on `≥26`-piece positions (one-hot best first move) from | |
| [`Lichess/chess-position-evaluations`](https://huggingface.co/datasets/Lichess/chess-position-evaluations) | |
| via [`avewright/lichess-opening-bestline`](https://huggingface.co/datasets/avewright/lichess-opening-bestline). | |
| This file is **`latest.pt` at opening-FT step 8000** (2026-09-14 14:42 UTC). | |
| Train loss ~1.7585. Frozen holdout hard CE ~1.4045. | |
| Not the generalist incumbent, the puzzle expert, the Syzygy expert, or the endgame expert. | |
| ## Training | |
| - Warm start: public 99M `latest.pt` (weights only), then full resume. | |
| - Split: position-hash 80/20 (seed 277). Frozen piece-stratified val 8192. | |
| - One-hot PV1 (`soft_alpha=0`). Pieces 26–32. | |
| - Polar-NorMuon, bs=528. Best disk ckpt at upload: step 8000. | |
| ## Files | |
| - `latest.pt` | |
| - `step_008000.pt` | |
| - `model_config.json` | |
| - `train.log` | |
| - `pack.json` | |