• Nvidia GPU with 32 or 48 GB memory
• python 3.10
• Pytorch 2.0

Before you use the code, please download eval_data.json from the source data section of our online paper.

• demo.py: demo of the attack
• eval_utils_counterfact.py: defined h
• gen.py: Generate adversarial completions on various contextual prompts.
• prob.py: Compute the probability of a completion given a prompt.

This project is licensed under the Apache License - see the LICENSE.md file for details

• Official implementation of GPT-J model: https://github.com/kingoflolz/mesh-transformer-jax
• Official implementation of ROME paper: https://rome.baulab.info/