Semi-Parametric Editing with a Retrieval-Augmented Counterfactual Model

Code and data for the ICML 2022 paper Memory-based Model Editing at Scale.

See the paper here and the project website here.

Setup

Environment

This codebase uses Python 3.7.9. Other versions may work as well.

Create a virtualenv (pyenv can help with this) and install the dependencies:

$ python -m venv env
$ source env/bin/activate
(env) $ pip install -r requirements.txt

Data

You can download the data needed for this project from this Google Drive link. You just need to unzip the archive into the top-level serac directory.

Running the code

You can run the code with:

(env) $ python -m run +alg=ALG +experiment=EXP +model=MODEL

See the scripts/ directory for examples. ALG may be one of:

• rep [SERAC]
• gtn [MEND; Mitchell et al., 2022]
• enn [Editable Neural Networks; Sinitsin et al., 2020]
• lu [lookup cache baseline]
• ft [fine-tuning baseline]

The EXP argument may be one of:

• zsre [question-answering; must be used with MODEL=t5large]
• fnli [fact-checking; must be used with MODEL=bert-base]
• sent [sentiment editing; must be used with MODEL=blender-small]

Citing the paper

If this repository is useful for your own research, you can cite our work with the following BibTeX entry:

@inproceedings{mitchell2022memory,
    title={Memory-Based Model Editing at Scale},
    author={Mitchell, Eric and Lin, Charles and Bosselut, Antoine and Finn, Chelsea and Manning, Christopher D.},
    booktitle={International Conference on Machine Learning},
    url={https://arxiv.org/pdf/2206.06520.pdf},
    year={2022},
}