circrl 1.0.0

Tools for applying circuits-style interpretability techniques to RL agents.

CircRL

A small library of mech interp tools, primarily focused on interpreting RL policies (though most of the tools are general).

The library has three main components:

• Hooks: Tools for hooking into PyTorch models that provide simple, safe wrappers around PyTorch forward hooks functionality and allow easy caching, patching and arbitrary hook functions.
• Probing: Tools for training linear probes on model activations (or any other data), including sparse probes.
• Rollouts: Tools for running rollouts and collectiong various kinds of data through a unified interface.

Installation

CircRL is available on PyPI, and can be installed with pip:

pip install circrl

Usage

A detailed, self-contained demo of CircRL is available in the CircRL demo notebook.

License

CircRL is licensed under the MIT license.

Citation

If you use CircRL in your research, please cite according to:

@misc{circrl,
  author    = {MacDiarmid, Monte},
  title     = {CircRL},
  year      = {2023},
  url       = {https://github.com/montemac/circrl}
}