This is the official repository of the paper:
Propagating Knowledge Updates to LMs Through Distillation
Shankar Padmanabhan, Yasumasa Onoe, Michael J.Q. Zhang, Greg Durrett, Eunsol Choi
NeurIPS 2023
This codebase uses Python 3.7.9.
$ conda create -n knowledge_distill -y python=3.7.9
$ conda activate knowledge_distill
(knowledge_distill) $ pip install -r requirements.txt
To run experiments, run a file from the root directory. There are two files, one for Entity Inferences and one for ECBD.
Example:
(knowledge_distill) $ python experiments/gpt_ecbd.py
To choose the editing method, change 'ki_method' in experiments/gpt_ecbd.py or experiments/gpt_entity_inferences.py accordingly.
Compatibility with ROME and MEND will be added soon.
@article{padmanabhan_2023_distill,
title={Propagating Knowledge Updates in LMs Through Distillation},
author={Shankar Padmanabhan and Yasumasa Onoe and Michael J.Q. Zhang and Greg Durrett and Eunsol Choi},
journal={Advances in Neural Information Processing Systems},
volume={36},
year={2023}
}
Please contact shankarpadmanabhan@utexas.edu if you have any questions or concerns.