This repo contains the original implementation of our paper "PEMT: Multi-Task Correlation Guided Mixture-of-Experts Enables Parameter-Efficient Transfer Learning".

Environment

Please run the command below to install the dependent libraries.

conda create -n pemt_env python=3.8
conda activate pemt_env
pip install -r requirements.txt

PEMT

PEMT consists of two-stage training: Source Task Training and Target Task Adaptation.

Training

1. Source Task Training: The goal of Stage 1 is to capture the task-specific knowledge of each source task. We fine-tune the PLM on multiple source tasks using adapter-based PEFT methods to obtain Source Task Adapter and Task Description Prompts.

2. Target Task Adaptation: In the second stage, PEMT is guided by the correlation between tasks to utilize the distilled knowledge of all source tasks for adaptation to the downstream target task.

Source Task Training

We use 6 high-resource datasets as the source tasks: MNLI, QNLI, QQP, SST-2, SQuAD and ReCoRD.

cd pemt
bash run_pemt_stage1.sh

Checkpoint

We provide source prompts and adapter weights for the six source tasks. You can download all the checkpoint from:

https://drive.google.com/drive/folders/1s-NHXNJ2XmTEkmaD2dx-sfFd7DegFhh4?usp=sharing

Target Task Adaptation

We use other tasks from four benchmarks as the target task to validate the effectiveness of PEMT under both full data and few-shot settings.

Full data

cd pemt
# remenber to modify the stage1 prompt and adapter loading path 
bash run_pemt_stage2.sh

Few data

cd pemt
# remenber to modify the stage1 prompt and adapter loading path 
bash run_pemt_stage2_few_shot.sh

Result Example on SuperGLUE-CB

***** test metrics *****
  epoch                   =              20.0
  test_accuracy           =           92.8571
  test_average_metrics    = 92.85714285714286
  test_loss               =            0.1278
  test_moe_weight_0       =            0.6418
  test_moe_weight_1       =            0.0734
  test_moe_weight_2       =            0.0514
  test_moe_weight_3       =            0.0623
  test_moe_weight_4       =            0.0425
  test_moe_weight_5       =            0.1285
  test_runtime            =        0:00:00.52
  test_samples_per_second =             53.67

Acknowledgements

The implementations of the baselines are from the ATTEMPT repository. Huge thanks to the contributors of those amazing repositories!

Citation and Contact

If you find this repository helpful, please cite our paper.

@article{lin2024pemt,
  title={PEMT: Multi-Task Correlation Guided Mixture-of-Experts Enables Parameter-Efficient Transfer Learning},
  author={Lin, Zhisheng and Fu, Han and Liu, Chenghao and Li, Zhuo and Sun, Jianling},
  journal={arXiv preprint arXiv:2402.15082},
  year={2024}
}