Contrastive Sparse Autoencoders for Interpreting Planning of Chess-Playing Agents

Project page | HF Space | Paper

Scripts for interpreting planning in LeelaChessZero networks.

🔴 Not a stable codebase

Install & Run

This work relies on poetry to manage the dependencies. To install run (with additional demo group for running the demo):

poetry install

Then to run a particular script use:

poetry run python -m scripts.sae_training.train_contrastive

To run the demo you can use the following make shortcut:

make demo

Tooling

See the lczerolens library (still under development) for more agnostic tooling to interpret the Leela Networks.

Contribute

Feel free to open a discussion, an issue or a PR for any question or feedback.

Cite

If you find this work useful please consider citing the associated paper:

@misc{poupart2024contrastivesparseautoencodersinterpreting,
      title={Contrastive Sparse Autoencoders for Interpreting Planning of Chess-Playing Agents},
      author={Yoann Poupart},
      year={2024},
      eprint={2406.04028},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2406.04028},
}