Model description
GPT-2 medium finetuned on 8 million chess games (short algebraic notation)
Data: Chess DB + sample from lichess + sample from CCRL
Example context: "1-0 2700 1350 1.e4 e5 2.Nf3 Nc6" (white score-black score white_elo black_elo moves)
Model results
- ELO (measured against Stockfish) ~ 1340
- % legal moves 98.5%
- checkmates in one move (from BigBench benchmark) - 46.5%