// the find
hiyouga/FastEdit
🩹Editing large language models within 10 seconds⚡
FastEdit is a single-algorithm implementation of ROME (Rank-One Model Editing) for patching specific facts into an LLM's weights without full retraining. It's aimed at researchers who want to reproduce or experiment with knowledge editing on older open LLMs (LLaMA/LLaMA-2, GPT-J, BLOOM, Falcon-7B) rather than anyone looking for a production editing tool.
The edit data format (prompt/subject/target/queries) is simple and well-specified, so you can get a working edit without reading the ROME paper first. Loading, hook-based editing, and generation are split cleanly across mtloader/nethook/rome_main, so the ROME math in compute_u.py and compute_v.py is isolated from the model-specific plumbing. It ships as an installable package (pyfastedit) in addition to the repo clone, and the README backs its claims with actual before/after generations rather than just metrics.
Last commit is August 2023 and only ROME ever got implemented — the MEMIT and NER-based subject extraction items on the TODO list never happened, so the roadmap stalled at v0. Every supported model (LLaMA-2, Baichuan, Falcon-7B) is now a legacy checkpoint; there's no path to Llama 3, Mistral, or Qwen without reimplementing the hparams config yourself. ROME is well known to degrade general model quality after more than a handful of sequential edits, and the README never mentions this or gives a safe edit count. There's no test directory, so a transformers/torch version bump could silently break the weight-hooking logic with no CI to catch it.