Chess GPT-2 Medium

Fine-tuned GPT-2 Medium (355M parameters) for chess move prediction.

Training Details

• Base model: gpt2-medium (355M parameters)
• Dataset: Lichess/chess-position-evaluations (Stockfish depth≥20)
• Training examples: 1,500,000
• Epochs: 3
• Batch size: 32
• Learning rate: 3e-05
• Max sequence length: 128
• Format: FEN: <fen_string> MOVE: <uci_move>

Usage

The model predicts the best UCI move given a FEN position. Use with constrained decoding (lm-format-enforcer) for 100% legal move rate.

Inference

prompt = "FEN: rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1 MOVE: "