Today i'm releasing chess-autocomplete, a new state-of-the-art for human move matching accuracy in chess. A chess bot that plays like a human, trained on 1.76 billion chess games.

In this blogpost I walk you through the full process from having an idea, creating a dataset, designing evaluations, and finally training a new state-of-the-art transformer model.

The methods used here have close parallels to how LLMs are trained, so if you're curious about them give this a read.

Everything is open-source:

- The datasets

- Training and inference code

- The final weights for all three model sizes

- Training curves

- Model checkpoints at multiple stages of training

PS: Want a challenge? Try beating the model, you can play against it in the blog post.