This repository provides an implementation of Rank-One Model Editing (ROME) for efficiently locating and editing factual associations within large auto-regressive transformer models. It is targeted at researchers and practitioners in NLP and AI safety who need to modify factual knowledge in pre-trained language models without full retraining.

How It Works

ROME operates by identifying and modifying a low-rank update to the model's weight matrices that specifically targets factual associations. This approach leverages causal tracing to pinpoint the relevant components within the transformer's layers and then applies a rank-one update, making the editing process computationally efficient and precise.

Quick Start & Requirements

• Install via bash ./scripts/setup_conda.sh.
• Requires Conda for dependency management, PyTorch, and CUDA.
• Supports OpenAI's GPT-2 XL (1.5B) and EleutherAI's GPT-J (6B).

Highlighted Details

• Implements Rank-One Model Editing (ROME) for factual association modification.
• Includes notebooks for demonstrating Causal Tracing and ROME.
• Provides scripts for running the full evaluation suite against baselines.
• Offers a framework for integrating and benchmarking new editing methods.

Maintenance & Community

• Actively developed with close monitoring of issues.
• Paper published at NeurIPS 2022.

Licensing & Compatibility

• License details are not explicitly stated in the README, but the project is open-source.
• Currently supports methods editing autoregressive HuggingFace models using the PyTorch backend.

Limitations & Caveats

• Primarily GPU-only.
• Cross-platform compatibility is limited to PyTorch/HuggingFace models, with broader support planned.