This project builds on the TOFU paper and codebase (Maini, Feng, Schwarzschild et al., 2024), which published a dataset for benchmarking approximate unlearning methods - techniques for removing knowledge of a concept from a large language model.
We explore two research questions:
-
Does the granularity of the concepts being forgotten impact the quality of forgetting that can be achieved? By granularity here we mean the position in a hierarchy of concepts, for example a book is published by an author, an author writes multiple books, and a publisher publishes multiple authors.
-
Does removing a relationship between two entities reduce the performance of the model on unrelated questions about those entities?
To address these we create a new TOFU-inspired question-answer dataset with entities of different types (publishers, authors, books) and with each question-answer pair in the dataset labelled with the entities it refers to.
You can pip install the dependencies and arcsf library with pip install . in your preferred virtual environment.
We developed the code with Poetry:
-
Install dependencies with Poetry
poetry install
-
Install pre-commit hooks:
poetry run pre-commit install --install-hooks
See the config files and scripts readmes.