The original authors provide get_gpt4_dist.py as a demo; note that this demo costs about $0.25 each time you run it. Although I haven't tried yet, I imagine the Jupyter notebooks in some cases cost several dollars each time you run them. Don't run large sample code initially, run it with small samples until you're sure it works.
In order to use an OpenAI model through an API (i.e. via code), you have to provide a valid API key linked to an openAI account that can be charged to. It is really important that this API key is NOT included in the public repo, otherwise other people can get the queries for free using your money! The way I've handled it in this fork is by use of a file apikey.py, which is not included in the repo but is listed in .gitignore so it should be pretty hard to accidentally include the API key into the public repo. The file apikey.py has the following contents:
import openai
openai.api_key = "THIS IS WHERE YOU PUT YOUR API KEY"
Any script querying GPT must include the line import apikey. I have made this change to utils.py in order to get the minimal test code (below) to work.
I've added test_working.py. Use this first to check you have a properly functioning environment. You can run it either directly from the command line (python test_working.py), or through your IDE (I recommend Visual Studio Code).
If your environment is working correctly, you will end up with a file output/results_demographics_temp_0.7_num_samples_1_max_tokens_100_condition_Sarcoidosis.json which has a series of prompts and responses. This is a cut down version of the queries used to create the top left blue plot point in Figure 1 (cut down in the sense that each prompt is used once rather than 100 times)
-
If you are seeing errors such as
no module named XXX, you need to install the appropriate Python packages. How you do this depends on the package manager you are using (I recommend Anaconda). -
If you are getting
Error occurred ... API key ..., your API key is not set up properly, see above.
Depending on the type of account you have with OpenAI, the code in this repo may make more queries per minute than your rate limit, and you'll start to get errors. There are two ways around this. (1) Give OpenAI more money, (2) use some Python code to slow down your query rate and retry in the case of failed queries.
This repository accompanies the paper "Coding Inequity: Assessing GPT-4’s Potential for Perpetuating Racial and Gender Biases in Healthcare".
The data is available in the data_to_share folder. This can be broken into several pieces:
-
simulated_pt_distribution--- here is where we store all the information for generating patient demographic distributions. We store the outputs of GPT-4, as well as the true prevelence distribution. -
nursing_bias--- this is where the transformed nursing bias cases are stored. We additionally store the outputs here. -
healer_cases--- this is where the healer cases are stored. We additionally store the outputs here.
There are two folders in simulated_pt_distribution --- outputs and true_dist_work. In outputs, the files are just outputs of GPT-4. These are all pickle files. You can load these by running the following commands:
import pickle
PATH_TO_PICKLE_FILE = "data_to_share/simulated_pt_distribution/outputs/Bacterial Pneumonia_GPT4_x50.pkl"
with open(PATH_TO_PICKLE_FILE, "rb") as f:
loaded_file = pickle.load(f)
To see the the true distributions, as well as which sources they came from, please look at final_true_dist.csv. There are some other CSVs in this folder; however, final_true_dist.csv is the main file that should be looked at. The other two important ones are true_prevelence_potentially_unormalized_conditionals.csv and true_prevelence_potentially_unormalized.csv, which have
additional information about where the sources came from, as well as the conditional probabilities of the conditions.
This folder mostly contains the vignettes, as well as the outputs of GPT-4. The vignettes can either by loaded through the .py files OR through the csv file. To load the CSV file, you can use the following code:
import pandas as pd
df = pd.read_csv("data_to_share/nursing_bias/unconscious_bias_nurses_final.csv")
The CSV has the following keys: case, gender, race, text, system, prompt, options.
- case: Which of the vignettes does it belong to?
- gender: Which gender is discussed in the text?
- race: Which race is discussed in the text?
- text: The vignette filled in with gender and race.
- system: What is the system level prompt we should use for GPT-4.
- prompt: Everything that should be passed to GPT-4. It has text and options.
- options: What are the possible options
Unfortunately, this is the messiest part of the data --- We apologize in advance! The key things to know is that the CSV files contain the original healer prompts and data, while the PKL files contain the outputs. The CSV files have the following rows:
- title: The title of the case. This will be essential for matching it to the output in the PKLs.
- Case one liner: The actual case we provide GPT-4.
- DDx: A list of potential ddxs --- you will need to split by newlines.
We additionally provide the outputs of GPT-4 for each of these cases. These can be found in the PKL files.
This folder has some basic prompts that we use throughout the code.
In this section, we will describe the code layout! This is still a work in progress. If you are re-running OpenAI commands, be sure to set the os.environ properly, in order to contain your specific API key.
To generate the nursing bias cases from the .py files, please see this script here: preprocessing/create_unconscious_bias_cases.py. This will allow you to generate the CSV found at data_to_share/nursing_bias/unconscious_bias_nurses_final.csv.
A lot of the code for generating the outputs of GPT-4 can be found in the src/notebooks file. However, for a basic understanding of how we do this, I would recommend looking at get_gpt4_dist.py, which queries for the conditions seen in Figure 1.
The code to generate the figures can be seen in either their respective folder (e.g., src/healer_cases/) or in src/notebooks. Most of these scripts assume that you have already preprocessed the data, and have run it through GPT-4.
If you have questions, please email lehmer16@mit.edu or raise an issue on the Github.