Offset Unlearning for Large Language Models

TL;DR

This paper introduces { extdelta}-Unlearning, a novel method for unlearning sensitive data in black-box large language models by learning logit offsets through smaller models, addressing ethical concerns while preserving performance.

Contribution

The paper presents { extdelta}-Unlearning, a versatile offset unlearning framework applicable to black-box LLMs that does not require internal model access or data retention.

Findings

Effectively unlearns target data from black-box LLMs.

Maintains or improves performance on general tasks.

Compatible with various unlearning algorithms.

Abstract

Despite the strong capabilities of Large Language Models (LLMs) to acquire knowledge from their training corpora, the memorization of sensitive information in the corpora such as copyrighted, biased, and private content has led to ethical and legal concerns. In response to these challenges, unlearning has emerged as a potential remedy for LLMs affected by problematic training data. However, previous unlearning techniques are either not applicable to black-box LLMs due to required access to model internal weights, or violate data protection principles by retaining sensitive data for inference-time correction. We propose {\delta}-Unlearning, an offset unlearning framework for black-box LLMs. Instead of tuning the black-box LLM itself, {\delta}-Unlearning learns the logit offset needed for unlearning by contrasting the logits from a pair of smaller models. Experiments demonstrate that {\delta}- Unlearning can effectively unlearn target data while maintaining similar or even stronger performance on general out-of-forget-scope tasks. {\delta}-Unlearning also effectively incorporates different unlearning algorithms, making our approach a versatile solution to adapting various existing unlearning algorithms to black-box LLMs.