FIND: Function Interpretation and Description Benchmark

A Function Interpretation Benchmark for Evaluating Interpretability Methods (official implementation)

Project Page | Arxiv

Sarah Schwettmannn*, Tamar Rott Shaham*, Joanna Materzynska, Neil Chowdhury, Shuang Li, Jacob Andreas, David Bau, Antonio Torralba.
* equal contribution

This repository is under active development, expect updates!

Setup

Clone this repository:

git clone https://github.com/multimodal-interpretability/FIND
cd FIND

Install dependencies:

pip install -r requirements.txt

Then download and unzip the FIND dataset into ./src/find-dataset:

wget -P ./src/find_dataset/ https://zenodo.org/record/8039658/files/FIND-dataset.zip
unzip ./src/find_dataset/FIND-dataset.zip -d ./src/

We include the dataset structure and examples of 5 functions per category under ./src/find_dataset/

Run interpretations

To run the interpretation, run cd ./src/run_interpretations/ and follow the instructions on the README file. The code will also allow you to add your own interpreter model.

You can also download the full FIND interpretations benchmark and unzip it to ./src/run_interpretations/:

wget -P ./src/run_interpretations https://data.csail.mit.edu/FIND/FIND-interpretations.zip
unzip ./src/run_interpretations/FIND-interpretations.zip -d ./src/run_interpretations/

See interpretation examples at ./src/notebooks/example_interpretations.ipynb

Evaluate interpretations

To evaluate the interpretations, run cd ./src/evaluate_interpretations/ and follow the instructions on the README file.

Generate new functions

To generate a new set of numeric and/or strings functions, run cd ./src/make_functions/ and follow the instructions on the README file.