Here are the instructions for getting started with this repository

Data are scraped from datasets. To download data, first make sure that the version of datasets is 1.4.0 (which, if you installed dependencies in requirements.txt, is NOT). then run the script in preprocess.sh

To generate random or #%_correct data by dataset, run the following:

python create_data.py --variant {random|0_correct|25_correct|50_correct|75_correct} --dataset {dataset}

Alternatively, if you want to generate a variant of all datasets in a config.json file under config, run:

python create_data.py --variant {random|0_correct|25_correct|50_correct|75_correct} --task {config_name}

This will create the corresponding datasets, and a new config.json file under config.
For the mechanistic part of the paper, you will need to create the random variant of data with k=-1 (save as many train samples as possible to sample). An example of the parametric knowledge retrieval datasets (which is referred to as "function_vectors_original") would look like:

python create_data.py --variant {random|0_correct|25_correct|50_correct|75_correct} --task function_vectors_original --k -1

1. Behaviour experiments (Section 4): this script runs on the parametric knowledge retrieval datasets.
2. Extracting FVs for all heads (Section 5.1): this script runs on the parametric knowledge retrieval datasets.
3. Identifying head importance via AIE (Section 5.1): this script runs on the parametric knowledge retrieval datasets.
4. Steering top FV heads (Section 5.3): this script runs on the parametric knowledge retrieval datasets.
5. Ablating top FV heads (Section 5.3): this script runs on the parametric knowledge retrieval datasets.