/FairThinking

Primary LanguagePythonMIT LicenseMIT

FairThinking

TL;DR

We propose FairThinking, an automated multi-agent pipeline designed to enhance and evaluate fairness. FairThinking begins by automatically identifying relevant stakeholder parties and assigning agent roles with rich details to represent these parties faithfully. Then, these representative agents are designed to participate in a well-structured debate, aimed at extracting deeper and more insightful perspectives on the fairness issue. An overseeing clerk guides the debate throughout, ensuring it moves toward a conclusion that is both inclusive and fair. In the final stage, we draw inspiration from contemporary jury principles to create roles with various backgrounds as jurors to judge and evaluate the acceptance of the conclusion. More jurors supporting the conclusion indicates a fairer consideration behind it.

Repo Structure

- FairThinking_code/   
    - demo
    - utils
    - demo.py
    - baseline.py
- dataset/
    - BiasAsker
    - Comparative_Questions
    - General_Open-Ended_Questions
    - Targeted_Open-Ended_Questions
- README.md
- requirements.txt

Setup

FairThinking is implemented on Python 3.10. To install all dependencies, please get into this directory and run the following command.

pip install -r requirements.txt

To conduct experiments on GPT-3.5/GPT-4, you need to add your Openai key here.

To conduct experiments on Llama2/Mistral, you first need to download the model repository, and then set the directory of the model here. In addition, to load and infer the open-source LLM, you can install this official repository first to prepare the necessary environment.

Usage

FairThinking is very easy to use. We have provided a demo, you can execute this script in the installed Python environment. We prepare some collected questions in this directory, the complete dataset will be released later.

This script has 8 parameters:

  1. topic_path assigns the question you want FairThinking to answer. It should be a question.pkl or topic.csv in the demo repository.
  2. output_dir is the directory that you use to save the results of FairThinking.
  3. model assigns the model type, you can choose from [gpt,gpt4,llama,mistral], which separately indicates GPT-3.5-Turbo, GPT-4-Turbo, Llama 2, Mistral.
  4. jury is the jury number you want to set, which is defaulted to be 6.
  5. config_path, attribute, gpu_list, and repeat do not require to config.

Dataset

We provide the whole dataset in this directory. The sub-directories separately show the BiasAsker, Comparative Questions, Targeted Open-Ended Questions, and General Open-Ended Questions datasets.

  1. question.pkl directly stores questions as a list (open-ended questions are stored in a .csv file).
  2. xx-roles refers to the role information corresponding to the question of index xx in the list.
    1. The attribute_list.pkl stores the role information of the debaters.
    2. The jury_list.pkl stores the role information of the jurors, which is consistent with the demo.