EditSSS

This is the code attached to the paper published in ACL 2024 Findings: SSS: Editing Factual Knowledge in Language Models Towards Semantic Sparse Space.

Environment Setup

Follow the instructions in EasyEdit to build the environment.

conda create -n edit python=3.9.7
...
pip install -r requirements.txt

Quick Start

SSS supports three main training-based knowledge editing methods, including FT-L, MEND and SERAC. The base model is gpt2-xl-1.5B. You can also add new test model in "./hparams".

Run the baseline methods using:

bash ./scripts/ft.sh
bash ./scripts/mend.sh
bash ./scripts/serac.sh

Run SSS method using:

bash ./scripts/ft_SSS.sh
bash ./scripts/mend_SSS.sh
bash ./scripts/serac_SSS.sh

To test the locality or portability, you can use the command:

# portability
bash ./scripts/test_inverse_relation.sh
bash ./scripts/test_one_hop.sh
bash ./scripts/test_subject.sh

# locality
bash ./scripts/test_locality.sh

Acknowledgement

This repository is built using the EasyEdit codebase.