Self-trained chessbot
A small transformer that learned chess only by reading game records, one character at a time. Nobody told it the rules: every move it plays, it writes.
Promote to
Loading the model…
About this bot
- It was trained from scratch on 3 million Lichess games between players rated 1800+, by predicting the next character of the move list. It has 10.7M (Base) or 25.2M (Large) parameters. The fine-tuned versions saw 594k extra long games.
- A move counts only if it writes exactly the notation of a legal move. Otherwise the attempt is crossed out and it tries again, up to five more times; if all fail, its highest-scored legal move is played and marked
*. - The top five come from scoring every legal move by how likely the model is to write it.
Code, results and the story of how it was built: github.com/maelritouet/self-trained-chessbot.