UofTCSSLab/Maia3-79M

1---2language:3- en4tags:5- chess6- maia7- maia38- chessformer9- move-prediction10- human-ai11- interpretability12---13 14# Maia3-79M15 16Part of the [**Maia3**](https://huggingface.co/collections/UofTCSSLab/maia3) family of transformer models for human chess move prediction. This is the **79M-parameter variant**.17 18For full details (architecture details, training recipe, full evaluation, and ablations) see our paper [*Chessformer: A Unified Architecture for Chess Modeling*](https://openreview.net/forum?id=2ltBRzEHyd) (ICLR 2026).19 20## Model summary21 22- **Family:** Maia-3, human move prediction models built on the **Chessformer** architecture23- **Architecture:** encoder-only transformer with board squares as tokens, augmented by **Geometric Attention Bias (GAB)**, a dynamic positional encoding that adapts to the geometry of chess, and an attention-based source-destination policy head24- **Parameters:** 79M25- **Task:** predicting the move a human player of a given skill level would make from a given position26- **Training data:** Lichess human games, January 2023 – July 202527- **License:** AGPLv328 29## Intended use30 31Maia-3 models predict human chess moves conditioned on player rating. Typical uses include:32 33- Research on human chess modeling and human–AI alignment34- Tools for chess education and entertainment35- Move-suggestion and analysis tools that emulate play at a chosen rating36- Mechanistic interpretability research: the square-token design makes attention patterns and activations directly attributable to board squares37 38Not intended for maximum playing strength. For strong engine play built on the same architecture, see the Chessformer integration into Leela Chess Zero described in the paper.39 40## How to use41 42Maia3-79M is a PyTorch checkpoint trained with the code at [CSSLab/maia3](https://github.com/CSSLab/maia3). Clone that repo, set up the conda environment, and load the checkpoint following the instructions in its README.43 44Architecture hyperparameters for this variant are defined in `ablate_size.sh` in the training repo.45 46## Training47 48- **Data:** Lichess monthly game dumps, January 2023 – July 202549- **Code:** [CSSLab/maia3](https://github.com/CSSLab/maia3)50- **Config:** size ablation row corresponding to 79M parameters in `ablate_size.sh`51 52## Evaluation53 54The Maia-3 family reaches **57.1% move-matching accuracy** on human moves, significantly surpassing the previous state of the art with fewer than a quarter of the parameters. Per-size accuracy curves, scaling analysis, and skill-conditioned breakdowns are reported in the paper.55 56## Citation57 58```bibtex59@inproceedings{monroe2026chessformer,60  title={Chessformer: A Unified Architecture for Chess Modeling},61  author={Daniel Monroe and George Eilender and Philip Chalmers and Zhenwei Tang and Ashton Anderson},62  booktitle={The Fourteenth International Conference on Learning Representations},63  year={2026},64  url={https://openreview.net/forum?id=2ltBRzEHyd}65}66```