avewright/syzygy-model

99M Syzygy expert (squares64)

Endgame specialist. Same 99M squares64 architecture as avewright/chess-transformer-100m-squares64, finetuned on avewright/chess-soft-syzygy.

This file is best-val latest.pt at syzygy-FT step 500 (2026-09-13 19:46 UTC). Train loss ~0.7616. Holdout hard CE ~0.9568.

This repo is not the generalist incumbent or the puzzle expert.

Training

• Warm start: public 99M latest.pt (weights only).
• Split: position-hash 80/20 (seed 274). Train 398846; eval 99555.
• One-hot on DTZ-best WDL move (soft_alpha=0). One epoch (745 steps, bs=536).
• Value masked (value_valid=0). DTZ is not mate.
• Best holdout was step 500 (hard CE 0.9568). Steps 600/700 were worse.

Files

• latest.pt
• step_000500.pt
• model_config.json
• train.log
• pack.json