FairLLM: Mitigating Age-Related Bias in Large Language Models

This archive is distributed in association with the INFORMS Journal on Computing under the MIT License.

The software and data in this repository are a snapshot of the software and data that were used in the research reported on in the paper Mitigating Age-Related Bias in Large Language Models: Strategies for Responsible AI Development by Zhuang Liu, Shiyao Qian, Shuirong Cao, and Tianyu Shi.

Cite

To cite the contents of this repository, please cite both the paper and this repo, using their respective DOIs.

https://doi.org/10.1287/ijoc.2024.0645

https://doi.org/10.1287/ijoc.2024.0645.cd

Below is the BibTex for citing this snapshot of the repository.

@misc{liu2024ijocCode,
  author =        {Zhuang Liu, Shiyao Qian, Shuirong Cao, Tianyu Shi},
  publisher =     {INFORMS Journal on Computing},
  title =         {{Mitigating Age-Related Bias in Large Language Models: Strategies for Responsible AI Development}},
  year =          {2025},
  doi =           {10.1287/ijoc.2024.0645.cd},
  url =           {https://github.com/INFORMSJoC/2024.0645},
  note =          {Available for download at https://github.com/INFORMSJoC/2024.0645},
}  

Overview

FairLLM is a project aimed at reducing age-related bias in large language models (LLMs). As LLMs continue to be widely applied across various domains, ensuring their fairness and inclusivity has become crucial. FairLLM introduces two innovative bias mitigation strategies: Self-BMIL (Self-Bias Mitigation in-the-loop) and Coop-BMIL (Cooperative Bias Mitigation in-the-loop), along with an Empathetic Perspective Exchange strategy. These approaches reduce bias in model outputs through self-reflection, collaborative debate, and perspective transformation, thereby enhancing the fairness and inclusivity of the models.

Project Structure

FairLLM/
├── configs/              # Configuration files for experiments
│   ├── base_config.yaml  # Base configuration settings
│   ├── experiment_configs/ # Experiment-specific configurations
│   └── model_configs/    # Model-specific parameter settings
│   └── training_configs/ # training parameter settings
├── data/                 # Dataset storage
│   ├── bias_datasets/    # Datasets related to bias
│   └── synthetic_data/   # Tools for generating synthetic data
├── docs/                 # Project documentation
│   ├── API_REFERENCE.md  # API documentation
│   ├── ARCHITECTURE.md   # System architecture overview
│   └── FAIRNESS_PROTOCOL.md # Fairness protocol guidelines
├── requirements/         # Project dependencies and environment setup
│   ├── docker/           # Docker-related files
│   │   └── k8s/          # Kubernetes configurations
│   └── requirements-dev.txt # Development environment dependencies
├── results/              # Experiment results and analysis
│   ├── analyzer.py       # Result analysis tools
│   └── experiment_controller.py # Experiment management tools
├── scripts/              # Various scripts for project tasks
│   ├── evaluation_pipeline.py # Evaluation pipeline script
│   ├── model_service.sh  # Script for deploying model services
│   └── train_selfbmil.sh # Script for training with Self-BMIL
└── src/                  # Source code
    ├── agents/           # Agent modules for different strategies
    ├── data/             # Data processing utilities
    ├── evaluation/       # Evaluation metrics and tools
    ├── models/           # Model definitions and architectures
    ├── training/         # Model training definitions
    └── utils/            # General-purpose utilities

Installation

Prerequisites

• Hardware Requirements:
• NVIDIA A100 GPUs with at least 40GB of VRAM.
• Multi-node GPU cluster with high-speed interconnects (recommended).
• Minimum of 128GB RAM per node (recommended).
• Software Requirements:
• Python 3.8 or later
• Specific configurations of Docker (24.0.0+) and Kubernetes (with kubectl, 1.28+).
• CUDA toolkit version 12.1.1.
• Proprietary optimized versions of PyTorch.
• Transformers library by Hugging Face

Using Docker

To build and run the FairLLM service using Docker: 1. Build the Docker image: bash docker build -t fairllm-service -f requirements/docker/Dockerfile.prod . 2. Run the Docker container:bash docker run -d --name fairllm-container -p 8080:8080 fairllm-service 3. Verify the deployment: ```bash docker ps curl http://localhost:8080/health

Using Kubernetes

To deploy FairLLM to a Kubernetes cluster: 1. Build and tag the Docker image: bash docker build -t registry.internal/fairllm-service:latest -f requirements/docker/Dockerfile.prod . docker push registry.internal/fairllm-service:latest 2. Deploy to Kubernetes:bash kubectl apply -f requirements/docker/k8s/deployment.yaml kubectl apply -f requirements/docker/k8s/service.yaml 3. Verify the deployment: ```bash kubectl get pods kubectl get services

Using Prometheus for Monitoring

To monitor FairLLM using Prometheus: 1. Deploy Prometheus: bash kubectl apply -f https://raw.githubusercontent.com/prometheus-operator/prometheus-operator/main/bundle.yaml kubectl apply -f requirements/docker/k8s/Prometheus_metric.yaml 2. Access Prometheus Dashboard:bash kubectl port-forward svc/prometheus-server 9090:9090 3. Query Metrics: - Access the Prometheus dashboard at http://localhost:9090 - Query metrics like fairllm_bias_score, fairllm_inference_latency_seconds, etc.

Usage

Training

To train the FairLLM model using the Self-BMIL strategy: bash bash scripts/train_selfbmil.sh --config configs/training_configs/rlhf_selfbmil_train.yaml To train using the Coop-BMIL strategy: bash bash scripts/train_coopbmil.sh --config configs/training_configs/adversarial_coopmil.yaml

Evaluation

To evaluate the model's performance and fairness metrics: bash python scripts/evaluation_pipeline.py --config configs/experiment_configs/exp_bias_mitigation.yaml

Deployment

To deploy the model as a service: bash bash scripts/model_service.sh --model_path /path/to/your/model --config_path configs/model_configs/llama3_8b_bmil.yaml

Results

The project includes results from various experiments demonstrating the effectiveness of the proposed bias mitigation strategies. Key results include:

• Figure 1: Comparison of Accuracy, Bias Scores, and Fairness Metrics across different models and confgurations.

• Figure 2: Cross-Dataset Fairness Metric Analysis.

• Figure 3: Impact of Different BMIL Methods on Fairness Metric.

Contributing

We welcome contributions to FairLLM! Please see our contribution guidelines for details on how to contribute.

Support

For any issues or questions related to FairLLM, please open an issue in the GitHub repository.