Assessing and Addressing Gender Bias in Large Language Models

This repository corresponds to the Master's thesis in Artificial Intelligence by Dennis Agafonov, at University of Amsterdam, 2024. This research is part of a project on Bias Identification methods initiated by Rhite and has been guided and supervised by Rhite and the University of Amsterdam.

⚠️ Note on Code Quality

Please note that this code was developed as part of a thesis project and, due to time constraints, has not undergone extensive optimization or formal quality checks. While it serves the primary purpose of supporting the research findings, it may not meet industry standards in terms of performance, maintainability, security or robustness.

Users are welcome to use and explore the code, but we recommend careful consideration and further testing before applying it in any production environment. Contributions for improvements or optimizations are also encouraged.

Introduction

This repository contains the code to run the experiments discussed in the thesis. The code for the bias assessment methods is in the folder Assessment/. The code for the two variants of the disentanglement model is in the folders GRU_DM/ and Transformer_DM/. The code for the debiasing method is in Debiasing.

Requirements

• Python version: 3.11.4
• Torch and CUDA version: 2.2.0+cu121

The file requirements.txt contains the necessary packages. All experiments were conducted using Ubuntu.

Bias Assessment Experiments

Folder Structure:

Assessment/
├── Data/
├── methods.py
├── run_seat.py
├── run_csps.py
├── run_disco.py
├── run_sentiment.py

• Data/: Contains the datasets used for the bias assessment experiments.
• methods.py: Contains the implementations of the assessment methods.
• run_<method>.ipynb: Jupyter notebooks containing the executions of the bias assessment methods across the model and dataset variations.

Disentanglement Model Experiments

T_DM/
├── Data/
├── main_dem.py
├── dem.py
├── train_dem.py
├── eval_dem.py
├── utils.py

• Data/: Contains the datasets used for generating the BoW vectors, training the T-DM and evaluating the T-DM.
• main_dem.py: Pre-processes the BoW vector vocab data, train data and test data. Main file for running the training and evaluation experiments of the T-DM.
• dem.py: Contains the T-DM model, and the two types of feedforward networks/classifiers (D3, D4, D5 and D6).
• train_dem.py: Contains the training code of the T-DM.
• eval_dem.py: Contains the evaluation code of the T-DM.
• utils.py: Contains all the helper functions.

GRU_DM/
├── Data/
├── main_dem.py
├── dem.py*
├── train_dem.py*
├── eval_dem.py
├── utils.py
├── dict_v0.py*
├── dict.dict*

• Data/: Contains the datasets used for generating the BoW vectors, training the DM and evaluating the GRU-DM.
• main_dem.py: Pre-processes the test data. Main file for running the evaluation experiments of the GRU-DM.
• dem.py: Contains the GRU-DM model, and the two types of feedforward networks/classifiers (D3, D4, D5 and D6).
• train_dem.py: Pre-processes the train data and contains the training code of the GRU-DM.
• eval_dem.py: Contains the evaluation code of the GRU-DM.
• utils.py: Contains all the helper functions.
• dict_v0.py: Contains code to generate the dictionary used for the GRU-DM.
• dict.dict: Contains the dictionary of the tokenizer used for GRU-DM.

Images/

• Images/: Stores the t-SNE visualizations of the evaluation of the DM.

Models/

• Models/: Stores the DM after training.

Debiasing Method Experiments

Debiasing/
├── Data/
├── dem.py*
├── discriminators.py
├── train_dialogue.py
├── evaluate_dialogue.ipynb
├── utils.py
├── dict_v0.py*

• Data/: Contains the datasets used for training and evaluating the LLM.
• dem.py: Contains the architecture of the GRU-DM for easy access.
• discriminators.py: Contains the feedforward networks/classifiers (D1, D2).
• train_dialogue.py: Contains the training code of the LLM.
• evaluate_dialogue.ipynb: Contains the BLEU evaluation code of the LLM.
• utils.py: Contains all the helper functions.
• dict_v0.py: Contains code to generate the dictionary used for the GRU-DM.

Models/

• Models/: Stores the fine-tuned LLM after training.

* The code in these files integrates much code that was not made by me, but was retrieved from https://github.com/zgahhblhc/Debiased-Chat