RECIPE

Source code for EMNLP 2024 (main) paper Lifelong Knowledge Editing for LLMs with Retrieval-Augmented Continuous Prompt Learning.

DEMO

Before training or testing RECIPE

Please set the root directory path and model directory path in utils/global_attrs.py. Then, please place the Huggingface weight file of the language model to be edited into models/ directory.

Train RECIPE

Please run:

python train_recipe.py -mn 'gpt2-xl' -dn 'zsre'  

Checkpoints will be saved in train_records/recipe/gpt2-xl/train_name/checkpoints/. You can view training information in train_records/recipe/gpt2-xl/train_name/logs/ through Tensorboard.

Evaluate RECIPE

Please run:

python test_recipe.py -en 'recipe' -mn 'gpt2-xl' -et 'sequential' -dvc 'cuda:0' -ckpt 'train_records/recipe/gpt2-xl/train_name/checkpoints/a_checkpoint' -dn 'zsre' -edn 1000 

You can check results in eval_results/recipe.

Citation

Please cite our paper if you use RECIPE in your work.

@inproceedings{chen2024recipe,
    title = {Lifelong Knowledge Editing for LLMs with Retrieval-Augmented Continuous Prompt Learning},
    author={Qizhou Chen and Taolin Zhang and Xiaofeng He and Dongyang Li and Chengyu Wang and Longtao Huang and Hui Xue},
    year = 2024,
    booktitle = {EMNLP},
    url = {https://2024.emnlp.org/program/accepted_main_conference/}
}