We use the mamba package manager. Install Miniforge

mamba env create -f environment.yml
conda activate deed

• The code for editing, extracting the representations, and detection can be found under ./deed/.

• Here is an example for editing and extracting the representations with ROME and GPT-J: python deed.py --model gpt-j-6B --editing_method ROME --instances_per_class 1000 --starting_index_other 1010 --results_dir ./results-1k/ --gpu 0 --paraphrased_prompts --ds zsre

• Only MEND, ROME, and MEMIT are supported. For MALMEN, please refer to the original implementation.

• For MEND, we used the hypernetworks from ROME

• For evaluation, you can run this script (after adapting the direcotries): python eval.py --classifier ada --setting both

This repository is based on EasyEdit ACL2024 (forked on Feb 6th 2024)