Fine-Tuning Stability

Codebase for the EMNLP 2022 paper "Improving Stability of Fine-Tuning Pretrained Language Models via Component-Wise Gradient Norm Clipping" [paper]

Author: Chenghao Yang (yangalan1996@gmail.com) (University of Chicago)

Project Supervisor: Xuezhe Ma (xuezhema@isi.edu) (University of Southern California)

Reference

If you use this code as part of any published research, please acknowledge the following paper (it encourages researchers who publish their code!):

@inproceedings{yang-2022-stability,
  author =  "Yang, Chenghao and Ma, Xuezhe",
  title =  "Improving Stability of Fine-Tuning Pretrained Language Models via Component-Wise Gradient Norm Clipping",
  booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
  year =  "2022"
}

Dependency Installation

conda create -p ./env python=3.7
conda activate ./env # the environment position is optional, you can choose whatever places you like to save dependencies. Here I choose ./env for example.
pip install -r requirements.txt
cd ./CustomToolBox
pip install -e .
# you also need to download the dataset (e.g., anli) and model checkpoints, see below

Dataset and Model Downloading

python download_model_and_data_HF.py --model_name_or_path bert-base-uncased --task_name sst2 --cache_dir ./cache

The supported models list can be found here. Here we use bert-base-uncased as an example.

The supported GLUE tasks list is here. (We use sst2 as an example.)

Run Experiments

bash run.sh

You can play with run.sh to experiment with different settings.

Once you have run the experiments, you can compute the standard deviation of the performance and draw the visualization used in the paper by running:

pip install plotly==5.10.0
python compute_variance.py

Detailed Experiment Results

We will upload the experiment results as well as baseline replication results soon. Currently doing some works to move the data from the institutional server. Stay tuned!

Special Acknowledgement

Our work is based on the finding of Marius Mosbach's ICLR 2021 paper on fine-tuning stability for pretrained language models.

We would like to thank Marius Mosbach for his help on baseline replication.

Also check out their wonderful docker-supported codebase!