This repo contains tools used to interpret protein language models. viz contains the frontend app for visualizing SAE features. interprot is a python package containing tools for SAE training and interpretation.
cd viz
pnpm install
pnpm run devdocker compose build
docker compose run --rm interprot bash
pytestEach directory under interprot contains a command-line tool. For example, make_viz_files takes in an SAE checkpoint and generates JSON files containing SAE activations used to serve the visualizer. You can run it with
cd interprot
python -m make_viz_files \
--checkpoint-files <path to checkpoint> \
--output-dir <path to output directory where the JSON files will be saved>Or you can install this repo as a package and run it from any directory:
RUN pip install -e .
make_viz_files --checkpoint-files <path to checkpoint> --output-dir <path to output>test1