Chessformer: A Chess-Playing Transformer Model

We introduce Chessformer: a chess-playing transformer model, trained to play by predicting the next move in human chess games given all the previous moves. The purpose is to show that transformers can learn chess when trained in this way and to train a model with a more human-like playing style than other chess-bots. Chessformer is trained on a dataset of 16 million games, and through this learns to play both legal and strategic moves. In random positions outside the training data the best version of Chessformer plays a legal move 99. 2\% 0. 1\% of the time. Playing against Stockfish our model achieves an Elo rating of 1198 22. Based on an analysis of its moves Chessformer has a playing style very different from both humans and Stockfish, despite being trained on human chess games.