andyjliu / persona-steered-generation-bias

Code and data that can be used to replicate results for the ACL 2024 findings paper "Evaluating Large Language Model Biases in Persona-Steered Generation".

persona-steered-generation-bias

Code and data that can be used to replicate results for the ACL 2024 findings paper "Evaluating Large Language Model Biases in Persona-Steered Generation".

Setup

conda create -n persona_steered_generation python==3.11
conda activate persona_steered_generation
pip install -r requirements.txt
git clone https://github.com/tatsu-lab/opinions_qa.git
conda env config vars set OPENAI_API_KEY=<YOUR_KEY_HERE>

Evaluation

src/generate_stances.py can be used to, given an OpinionsQA survey, generate a list of stances that can be passed into a prompt for persona-steered generation; all personas are stored in src/personas_and_topics.json. Note that this script uses the OpenAI API to generate initial descriptions of each persona: we recommend verifying that persona descriptions are similar for both sides of a topic and accurately reflect the personas that they describe. It may be necessary to manually rewriting descriptions before generating data using them to avoid affecting evaluation. Once new data is generated and stored in the persona-tracking-data directory, it can be automatically evaluated for steerability using the src/persona_evaluation.py script. Metric values can then be computed with the src/analyze_experiment.py script.

Replication

The data in the persona-tracking-results directory is sourced by running our experimental analysis scripts on the corresponding data directories. The replicate_figures.ipynb notebook can then be used to replicate all results and figures from our paper.

Acknowledgements

Raw data is sourced from OpinionsQA. The LLM caricature metrics used are modified versions of those found in CoMPosT.