rome

Why we included this project

ROME is the reference implementation for a NeurIPS 2022 paper on model editing: it pinpoints where a factual association lives inside a transformer and rewrites that fact with a single rank-one update. The causal tracing notebooks let you probe which layers and attention heads actually carry a given statement, and the editing API takes a prompt, a subject, and the target rewrite. It runs on GPT-2 XL and GPT-J, and ships an evaluation harness with baselines for comparing editing methods on the CounterFact suite. If you are reproducing the paper's claims or building your own editing experiments, this is the place to start.