Teaching a model chess.

Play against TalNet ↗

I chose Mikhail Tal because I see him as eccentric, unique, bold and innovative. I wanted to see whether a small model could learn some of his move choices from his games.

I worked on this with AI and asked it to teach me the process. We went back and forth as I asked questions and worked out what to try next. The project grew from training a model on my PC into a local browser extension and a standalone chess engine.

The training data

The Tal archive contained 2,431 games. Training started with a broader collection from 18 player archives, then the model was fine-tuned on Tal’s own moves.

Each training example contained a board position, the move played and the final game result. The larger set had about 4.43 million examples. Keeping only Tal’s moves from his games produced another 82,132 examples for fine-tuning.

Positions were stored from the perspective of the player whose turn it was, so the model saw a consistent board orientation for either colour. When choosing a move, it could only pick from legal moves.

Training on my PC

Training ran on Windows 11 with a Radeon RX 7900 XTX, using Python, PyTorch and AMD’s ROCm stack. Getting the Python environment, GPU software and training code to work together was part of the job.

The choice of numerical format made a difference on this PC. The local benchmark notes recorded much slower GPU computation with bf16 than with fp16, so the trainer was changed to prefer fp16 automatically on the GPU. That benchmark measured GPU computation on this setup; it didn’t measure the whole training process.

The trainer saved the best model separately from the checkpoint used to resume a run. Fine-tuning reached 12 epochs, or passes through the training data. The best saved validation result came much earlier, at epoch 2.

Checking the result

The saved Tal checkpoint matched the recorded move in 50.61% of validation examples. That means its first choice was the same move as the one in the game record.

Whole games had been split between training and validation, but positions can repeat across games. A later check found validation positions that also appeared in training. That overlap means the score doesn’t establish how well the model handles entirely unseen positions.

Matching individual moves also doesn’t show whether it plays a good complete game or has picked up Tal’s style. Those questions need separate tests.

The model’s own choices

An early version used Stockfish checks and a handwritten preference for attacking moves alongside the model. Those rules could improve a suggestion, but they also changed it. I wanted to see what the model would choose, including its mistakes.

The extension and standalone player handle this differently:

• Tal Vision’s Astra mode normally keeps the model’s highest-probability legal move. It only consults Stockfish to check alternatives to a repeating position and reports when it intervenes.
• The standalone Python player chooses moves without Stockfish. It normally takes the model’s favourite legal move, avoiding threefold repetition unless its own position estimate suggests accepting the draw.

The match results below belong to the standalone player.

The local extension

The trained model was exported to ONNX for Tal Vision, a Chromium extension. After importing the model, you can paste a chess position and get a move suggestion. The model is stored and run locally. The extension doesn’t submit moves.

The interface needed to handle more than a successful prediction. Tests cover model imports, stale suggestions, finished games and warnings about unverified move history. There are also tests for communication between the standalone engine and a chess client.

A 20-game match

The standalone engine connected through UCI, a protocol used by chess programs. The saved local match log records 20 games against Stockfish at skill level 0. Colours alternated, with 60 seconds on the clock plus one second per move.

That adds up to 9 points out of 20, a 45% score, with no recorded losses on time. It’s a small baseline match against Stockfish’s lowest skill setting. It doesn’t establish an Elo rating or show performance against full-strength Stockfish.

The match runner had a bug that could distort the results: an engine crash counted as a win for its opponent. It was changed to stop the run if an engine process died, so a software failure couldn’t produce a free win.

What still needs work

The model still makes poor moves. The standalone player doesn’t search ahead, and its position estimate was learned from final game results. Endgame play needs work, as does an evaluation set with less overlap with training.