Topic analysis

Alignment Whack-a-Mole: Finetuning Activates Recall of Copyrighted Books in LLMs

This repository supports the paper 'Alignment whack-a-mole,' demonstrating that finetuning Large Language Models can trigger the recall of copyrighted book excerpts. The codebase includes pipelines for data preprocessing, finetuning models like GPT-4o and DeepSeek, and evaluating verbatim memorization.