Social Bias Evaluation for Large Language Models Requires Prompt Variations

This repository contains the code for social bias evaluation for LLMs using prompt variations. The dataset is BBQ dataset. Prompt variations are comprised of three perspectives.

1. task instruction and prompt
   • see data/template.tsv
2. few-shot examples
   • see data/BBQ_few_shot.jsonl
3. debias-prompts
   • see data/debias_prompts.json

How to Use

You can run experiments with the following command.

Dataset Preparation

Before inference, prepare the variation of 1. task instruction and prompt, 2. few-shot settings. python3 data/convert_format.py

Inference

You can run the inference by each LLM. export PYTHONPATH="$pwd/src:$PYTHONPATH"; python3 src/pred.py --model <model_name> --file <file_name> --debias_prompt <debias_prompt>

• model_name: model checkpoint in huggingface.
• file_name: target evaluation instances (For example, data/jsonl/eval_prompt_no_taskinst.jsonl)
• debias_prompt: debias_prompt key. See the above description. When evaluating without debias-prompts, drop this arg.

Evaluation

You can calculate task performance and social bias of LLMs. python3 evaluation/eval_bbq.py --result_dir <result_folder>