View Code? Open in Web Editor NEW

This project forked from circle-hit/sapt

Code for ACL 2024 accepted paper titled "SAPT: A Shared Attention Framework for Parameter-Efficient Continual Learning of Large Language Models"

Shell 0.21% Python 99.79%

sapt's Introduction

SAPT

The official implementation for the ACL 2024 paper SAPT: A Shared Attention Framework for Parameter-Efficient Continual Learning of Large Language Models.

Requirements

Python 3.10.12
PyTorch 2.1.0
Transformers 4.30.2
CUDA 12.2

Preparation

The train/dev/test data from SuperNI and Long Sequence Benchmark is placed in /CL_Benchmark.

And the generated pseudo data points are in /generated_data.

Training

First run gen_script_{benchmark}_{model}.py to obtain the training script.

For example, to implement T5 model on the SuperNI benchmark:

python gen_script_superni_t5.py

Then run the resulting script to start the training process.

Evaluation

To calculate metrics of Average Performance (AP), Forgetting Rate (F.Ra), Forward Transfer (FWT) and Backward Transfer (BWT):

python score.py your_result_path single_result_path

Citation

If you find our work useful for your research, please kindly cite our paper as follows:

@article{zhao2024sapt,
  title={Sapt: A shared attention framework for parameter-efficient continual learning of large language models},
  author={Zhao, Weixiang and Wang, Shilong and Hu, Yulin and Zhao, Yanyan and Qin, Bing and Zhang, Xuanyu and Yang, Qing and Xu, Dongliang and Che, Wanxiang},
  journal={arXiv preprint arXiv:2401.08295},
  year={2024}
}