README.md

metadata

license: mit
tags:
  - chess
  - transformer
  - recurrent
  - policy
  - middlegame
  - pytorch
library_name: pytorch

99M middlegame specialist (squares64)

Same 99M squares64 architecture as avewright/chess-transformer-100m-squares64, finetuned on 16–26-piece positions (one-hot best first move) from Lichess/chess-position-evaluations via avewright/lichess-middlegame-bestline.

This file is latest.pt at middlegame-FT step 8000 (2026-09-14 17:09 UTC). Train loss ~1.7870. Frozen holdout hard CE ~1.6071.

Not the generalist incumbent, the opening expert, 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 278). Frozen piece-stratified val 8192.
• One-hot PV1 (soft_alpha=0). Pieces 16–26.
• Polar-NorMuon, bs=528. Best disk ckpt at upload: step 8000.

Files

• latest.pt
• step_008000.pt
• model_config.json
• train.log
• pack.json