kmeng01/rome
A research implementation of Rank-One Model Editing for modifying factual knowledge in GPT-2 XL and GPT-J language models.
Not currently ranked — collecting fresh signals.
star history
This repository implements ROME, a technique for locating and editing factual associations stored in auto-regressive transformer models. It provides causal tracing tools to identify internal representations of specific facts in GPT models, then applies rank-one updates to the MLP layers to alter stored knowledge. The framework supports GPT-2 XL (1.5B) and GPT-J (6B), includes evaluation on the CounterFact dataset, and offers notebook demonstrations for both causal tracing and model editing.
Frequently asked
- What is kmeng01/rome?
- A research implementation of Rank-One Model Editing for modifying factual knowledge in GPT-2 XL and GPT-J language models.
- Is rome open source?
- Yes — kmeng01/rome is open source, released under the MIT license.
- What language is rome written in?
- kmeng01/rome is primarily written in Python.
- How popular is rome?
- kmeng01/rome has 770 stars on GitHub.
- Where can I find rome?
- kmeng01/rome is on GitHub at https://github.com/kmeng01/rome.