Selective Forgetting in LLMs

Abstract

Executive Summary This work approaches the problem of selectively forgetting knowledge from a large language model (LLM) for the purposes of safety, copyright, security, or otherwise. Also known as machine unlearning, this entails training a model to forget certain elements of the dataset on which it was trained. Unlearning methods must be evaluated both in terms of the extent to which the information has successfully been forgotten, and the performance of the unlearned model on the remaining (retained) data. We build on the work of TOFU (Task of Fictitious Unlearning) [21], which provides a dataset and benchmark for evaluating unlearning techniques. We create a new, TOFU-inspired question–answer dataset for the task of machine unlearning. The new dataset includes 10,500 question–answer pairs relating to over 1,000 distinct, synthetic entities of several types. Each question–answer pair is tagged with the entities it refers to, with the graph representation of our dataset containing over 2,600 edges between different entities. We perform two experiments with our dataset. The first experiment aims to capture whether the difficulty of forgetting a concept from a LLM depends on its granularity. For example, is unlearning more likely to be successful if forgetting a single book, rather than the book's author (as an author is connected to multiple books)? We find that granularity does not have a tangible effect on model performance in our dataset. There may be a small effect from granularity on the difficulty of forgetting, but this is not statistically significant across our results. Our second experiment explores the knock-on effect of forgetting a relationship between two entities. For example, if unlearning has been run on a model to forget only who wrote a book, but not the book or author themselves, is the model worse at responding to other questions about that book or author? We find that the model performance is lower on questions that contain the entities pertained in the relationship, than on those that do not. 1