This repository contains research code for investigations into the mechanistic interpretability of neural networks trained to implemented Boolean circuits.
Alongside example scripts notebooks this repository contains boolean_circuits a python library containing research code.
To install the library clone this repo and run
pip install -e .
.
├── README.md
├── plots/ \\ scripts for plotting specific results
├── notebooks/ \\ notebook examples
├── scripts/ \\ script examples
├── src/
│ └── boolean_circuits/ \\ library source code
│ ├── circuits.py \\ implementation of boolean circuits
│ ├── jax/ \\ jax specific code, models, training etc
│ │ ├── data/ \\ code to simulate other boolean data
│ │ ├── circuits.py \\ jax version of circuits.py (more feature-rich)
│ │ ├── models.py
│ │ ├── probes.py
│ │ ├── sae.py
│ │ └── utils/ \\ misc utilities
│ └── torch/ \\ pytorch specific code
└── tests/ \\ pytest tests
Multitask spare parity problem (MSP) from Michaud et al., (2023).
Code to define simulate Boolean circuits.