EvalPrompt: Analyzing Large Language Models for Self-Diagnosis

To better understand the capability of using Large Language Models (LLMs) for self-diagnosis, we prompted ChatGPT-4.0 with medical questions from the United States Medical Licensing Exam (USMLE). This repository contains the initial set of USMLE questions, along with the corresponding ChatGPT-4.0 responses which can then be evaluated by humans.

Contents of Repository

This repository contains 3 major files across 2 experiments. Experiment 1 assesses the correctness and clarity of ChatGPT-4.0 and Experiment 2 re-prompts ChatGPT-4.0 through sentence dropout by using the correctly answered questions from the first experiment.

1. generate_responses.ipynb: The notebook used to generate LLM responses for a specific experiment.
2. output.xlsx: The final ChatGPT-4.0 responses.
3. questions.xlsx: The initial USMLE questions.

To directly use the results of this repository, please download the output.xlsx Excel files from Experiments 1 and 2. Please refer to Data File Contents for details on the contents of the data files.

If you would like to replicate the experiment and reproduce the results, please follow Replicate the Experiment to begin setting up your local environment.

Data File Contents

Experiment 1 contains a single ChatGPT-4.0 response for each question, while Experiment 2 contains multiple answers for each question through sentence dropout.

Experiment 1

Column Name	Description
usmle_1_q_num	An ID referring to each Step 1 question. Any numbers missing are questions that were dropped.
usmle_1_q	The full multiple-choice textual question from the USMLE Step 1 test.
correct_answer	The correct answer to the multiple-choice question.
usmle_1_a	ChatGPT's answer to the open-ended version of the question.

Experiment 2

Column Name	Description
usmle_1_q_num	An ID referring to each Step 1 question. Any numbers missing are questions that were dropped.
usmle_1_q	The full multiple-choice textual question from the USMLE Step 1 test.
correct_answer	The correct answer to the multiple-choice question.
non_expert_match	The assessor categorization from all nonexperts, where C = correct, PC = partially correct, I = incorrect, and A = ambiguous. For example, CCC represents 3 assessors each categorizing ChatGPT-4.0's response to the particular question as correct.
expert_match	The assessor categorization from all experts. This is similar to the non_expert_match column.
usmle_1_a_*	Multiple columns each with the same prefix, representing one ChatGPT-4.0 response for each sentence. For example, usmle_1_a_0 represents ChatGPT-4.0's answer to the open-ended question with the first sentence being dropped.

Replicate the Experiment

1. Create a local virtual environment

   python3 -m venv .venv
   

2. Activate the virtual environment

   source .venv/bin/activate
   

3. Install dependencies

   pip install -r requirements.txt
   

4. Attain your API key from OpenAI. You may refer to this guide to locate your OpenAI key.
5. Create an .env file, with the key: OPENAI_API_KEY

   touch .env
   

6. Open the generate_responses.ipynb in the experiment_1 folder, and run each cell of code.
7. Have humans evaluate the output.xlsx file from Experiment 1 to determine which questions the LLM answered correctly and clearly.
8. Update the non_expert_match and expert_match columns from the questions.xlsx file from Experiment 2 with the results from the previous step.
9. Open the generate_responses.ipynb in the experiment_2 folder, and run each cell of code.
10. Have humans evaluate the output.xlsx file from Experiment 2 to evaluate the robustness of the LLM by determining which questions the LLM continued to answer correctly.