Checkmate-5M

A tiny chess move model built for the Chess LLM Arena.

• Architecture: Pythia / GPT-NeoX (6 layers, hidden size 128, 4 heads), trained from scratch.
• Parameters: 4.8M
• Tokenizer: move-level chess tokenizer: one token per SAN move (without +/#), plus the move-number prompt 1. and single SAN characters. Every legal SAN move is a single token.
• Training: the move-selection policy was optimized with reinforcement learning (policy gradient) on millions of simulated games against the models at the top of the arena leaderboard, then distilled into the network.
• Best used as White.

Usage in the arena

Type DedeProGames/Checkmate-5M as the White model in the arena and press Fight!.

Usage with transformers + outlines

import chess, re
import outlines.models as models
from outlines import generate

model = models.transformers("DedeProGames/Checkmate-5M")
board = chess.Board()
legal = "|".join(re.escape(re.sub(r"[+#]", "", board.san(m))) for m in board.legal_moves)
print(generate.regex(model, legal)("1."))