petezh / chatgpt-advice

Reproduction code for "Taking Advice from ChatGPT"

Taking Advice from ChatGPT

Last Updated on 05/20/2023

Author: Peter Zhang

Paper link

This study treats GPT models as an advisor in a judge-advisor system. We prompt InstructGPT using CoT on the MMLU benchmark and treat model output as "advice." In our lab study, 118 student participants answer 2,828 questions and recieve a chance to update their answer from the advice. We analyze factors affecting weight on advice and participant confidence in advice answers. This repository contains all of the collected data as well as code to reproduce the tables and figures in the paper.

• Setup
• Data
  • Model Output
  • Survey Responses
• Scripts
• Notebooks
• Outputs

Setup

1. Install packages using the following:

pip install -r requirements.txt

2. Add your OpenAI API key to ask_question.py.
3. Download the MMLU benchmark from the official repository and move it to hendrycks_test

Data

Model Output

• Relevant samples from the benchmark are in the benchmark_samples folder.
• The output used to as advice in the survey is results_scratchpad_0218.csv which is an evaluation on hendrycks_sample_0217.csv.
• The output used in the final evaluation is results_0425.csv.csv which is evaluated on hendrycks_sample_0423.csv.

Survey Responses

• The survey responses are downloaded from Qualtrics and placed in the survey_responses folder.
• The updated survey responses from April 23rd are stored in `responses_0425.csv.

Scripts

The create_dataset.py script creates a sample of the MMLU benchmark. The evaluate.py script supports multiple types of prompting. The templates used in the evaluation script are located in the templates folder.

To reproduce the model evaluation, run cot_eval.sh from the main repository folder.

The clean_responses.py script preprocesses the survey responses and creates features used in the analysis.

Notebooks

• The analyze_model_results.ipynb notebook produces figures displaying LLM performance.
• The `explore_questions.ipynb notebook creates some summary statistics about benchmark questions.
• The analyze_survey_responses.ipynb notebook produces all other figures and tables.
• To create survey questions from benchmark samples, use generate_survey_questions.ipynb.

Outputs

All figures and tables are output to their respective folders.