Understanding the Effect of Model Compression on Social Bias in Large Language Models

Gustavo Gonçalves and Emma Strubell

This is the official Repo for the EMNLP 2023 Paper - "Understanding the Effect of Model Compression on Social Bias in Large Language Models".

We thank the BiasBench authors for making their codebase available, as we use it as a starting point for our work.

Below you will find a short version of the original repo instructions to run the code for our experiments.

Install

mkdir projects && cd projects
git clone https://github.com/gsgoncalves/EMNLP2023_llm_compression_and_social_bias.git
cd EMNLP2023_llm_compression_and_social_bias
python -m pip install -e .

Running on an HPC Cluster

We provide scripts for running all of the experiments presented in the paper on a SLURM cluster in batch_jobs. If you plan to use these scripts, make sure you customize python_job.sh to run the jobs on your cluster. In addition, you will also need to change both the output (-o) and error (-e) paths.

Quickstart: How to develop in this codebase?

The high level structure of this codebase is as follows:

• batch_jobs: contains the scripts to run the experiments on a SLURM cluster and cluster specific configs. (The bash scripts can easily be ran in a single python instance e.g. cli or IDE)
  • The experiment_name folder contains "environment variables" that are called by the batch job scripts. You can register your new models here.
• bias_bench: contains the code for the BiasBench experiments. (The code is mostly unchanged from the original repo)
  • You can register new models in the models.py file.
• experiments: contains the code for the experiments presented in the paper.
  • Here crows.py, seat.py, stereoset.py were adapted to include the quantized and distil models.
• export: contains the code to export the results to tables.
  • Please note you must run the stereoset_evaluation.py script before running the respective export scripts.

We recommend checking the original BiasBench for further references.

Acknowledgements

This repository makes use of code from the following repositories:

• BiasBench
• Towards Debiasing Sentence Representations
• StereoSet: Measuring Stereotypical Bias in Pre-trained Language Models
• CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models
• On Measuring Social Biases in Sentence Encoders
• Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection
• Towards Understanding and Mitigating Social Biases in Language Models

We thank the authors for making their code publicly available.

Citation

 @inproceedings{goncalves-strubell-2023-understanding,
    title = "Understanding the Effect of Model Compression on Social Bias in Large Language Models",
    author = "Gon{\c{c}}alves, Gustavo  and
      Strubell, Emma",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.161",
    pages = "2663--2675",
}