self-trained-chessbot
Small GPTs (10.7M and 25.2M parameters) trained from scratch on 3 million Lichess games. They
are never told the rules of chess: they only learn to predict the next character of a game's
move list (1-0 1.e4 e5 2.Nf3 Nc6 β¦).
- Play it in your browser: https://maelritouet.github.io/self-trained-chessbot/
- Code, training, evaluation and the decision log: https://github.com/maelritouet/self-trained-chessbot
Files
One folder per model, each with model.onnx (fp32) and meta.json (vocabulary, game delimiter,
context size):
| Folder | Model | Size |
|---|---|---|
base/ |
6 layers, 6 heads, 384-dim, 10.7M parameters | 51 MB |
large/ |
8 layers, 8 heads, 512-dim, 25.2M parameters | 112 MB |
ft-base/ |
base, fine-tuned on 594k long games | 51 MB |
ft-large/ |
large, fine-tuned on 594k long games | 112 MB |
The ONNX graph takes idx: int64 (batch, seq) character ids (seq β€ 1024) and returns
logits: float32 (batch, seq, 33). SHA-256 sums are in
web/models.release.json.
How to use
The prompt is a newline, the result the model's side should achieve, a space, and the movetext so
far, ending where the next move starts: "\n1-0 1.e4 e5 2." asks for a white move played to win.
The model writes characters until a space. Its text counts as a move only if it is exactly the
SAN of a legal move; otherwise sample again. The full move function (retries, scoring every legal
move) is chessgpt/player.py (Python) and web/js/player.js (browser) in the GitHub repo.
Training
- Data: rated blitz, rapid and classical games from the Lichess database (October 2026, CC0) where both players are rated 1800+, rewritten as one line per game with the result first. 3M games, 1.26B characters. The fine-tunes used 594k games of 100+ plies from another month.
- Tokens: one per character, 33 symbols.
- Model: GPT-2-style decoder, 1,024-character context.
- Compute: one Kaggle T4 per run, fp16 mixed precision. base: 3.93B tokens in 8.6 h; large: 1.97B tokens in 9.1 h (the same compute budget, the project's controlled experiment); fine-tunes: +413M tokens each.
Results
All measured on data the models never trained on.
| base | large | ft-base | ft-large | |
|---|---|---|---|---|
| Parameters | 10.7M | 25.2M | 10.7M | 25.2M |
| Val loss (nats/char) | 0.386 | 0.366 | 0.393 | 0.374 |
| Legal-move rate | 95.6% | 97.7% | 96.1% | 97.8% |
| Same move as the human | 40.7% | 44.2% | 41.6% | 44.4% |
| Puzzles solved (1,000) | 8.0% | 11.3% | 6.8% | 8.3% |
| Score vs Stockfish Skill 0 / 1 / 2 | 0.06 / 0.00 / 0.03 | 0.17 / 0.02 / 0.00 | 0.09 / 0.02 / 0.01 | 0.08 / 0.06 / 0.02 |
| Forfeits vs Stockfish (of 150 games) | 49 | 39 | 52 | 28 |
- Legal move = the text is exactly the SAN of a legal move, greedy first try, 8,197 positions from 100 held-out games.
- Stockfish = Stockfish 19 at UCI Skill Level 0, 1 and 2, 50,000 nodes per move, 50 games per level. The model gets 6 tries per move and forfeits after 6 illegal outputs.
- Puzzles = 1,000 Lichess puzzles, 200 per rating bucket. Lichess rule: every solution move, greedy, one try each; any checkmating move also counts.
Limitations
- No search and no rules: it plays by pattern, and it gets weaker late in long games, where its legal-move rate drops (85β95% after move 80).
- It is weak: about 0.17 points per game at best against Stockfish's lowest setting.
- Trained on games between 1800+ players on one site in one month; it has no notion of time controls or of who is playing.
License
The model weights are licensed under CC BY 4.0. The training data (the Lichess database) is CC0.