FastEdit

Why we included this project

FastEdit targets a narrow problem: a trained language model that keeps giving the wrong answer on a specific fact. Instead of retraining, you hand it a small JSON file that pairs a prompt and subject with the corrected target, and it rewrites the relevant weights in seconds using the ROME rank-one editing algorithm. It covers a practical set of open models, including LLaMA, LLaMA-2, BLOOM, Falcon, and Baichuan, and the README spells out the GPU memory and per-iteration speed you can expect at 7B and 13B sizes. That makes it a useful reference for anyone experimenting with knowledge injection or trying to reproduce model-editing results without wading through a heavier framework.