This repository contains the code for the experiments of our VLM circuit overlap paper and the project website.
- data: Contains generated prompt csv files for each model and task. These files can also be generated automatically by the scripts and code. Each task directory also contains an images.tar.gz file that hosts all images on git-lfs.
- data_generation: Contains scripts to generate the images (that are found in the data folder).
- third_party: Contains a fork of the TransformerLens library with edits we made to make it support VLMs (specifically the VLMs we analyzed as well as Llama3.2).
- All script files (
script_.*.py) contain the code for running the experiments described in the paper as GPU jobs. - Other files contain processes used in experiments (attribution patching, faithfulness evaluations, alignment between modalities, circuit overlap measurements, utility functions for each task, etc).
docscontains code for the project website.