⚗️ Status: This project is still in alpha, and may change without warning.

Cross-Care

Cross-Care is a project that compares the co-occurrences of medical terms with demographic terms in pretraining data and investigates the downstream effects of these co-occurrences on the biases of language models trained on them. The project is divided into three main parts: co-occurrence analysis, logits analysis, and propagation evaluation. The results of these analyses are then visualized in a web application.

Citing

@misc{chen2024crosscare,
      title={Cross-Care: Assessing the Healthcare Implications of Pre-training Data on Language Model Bias}, 
      author={Shan Chen and Jack Gallifant and Mingye Gao and Pedro Moreira and Nikolaj Munch and Ajay Muthukkumar and Arvind Rajan and Jaya Kolluri and Amelia Fiske and Janna Hastings and Hugo Aerts and Brian Anthony and Leo Anthony Celi and William G. La Cava and Danielle S. Bitterman},
      year={2024},
      eprint={2405.05506},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Overview

Cross-Care/
│
├── co_occurrence_generate  # contains the code for co-occurrence analysis
│   ├── README.md           # contains instructions on how to run the co-occurrence analysis
│   
├── co_occurrence_results
│   ├── output/pile/           # Results specific to the Pile dataset
│   
├── logits_generate
│   ├── README.md           # contains instructions on how to run the logits analysis
│
├── logits_results
│   ├── api/               # Results specific to the API model
│   ├── hf/                # Results specific to the Hugging Face model
│   ├── hf_tf/             # Results specific to the Hugging Face model with True/False setting
│
├── propagation_eval
│   ├── README.md           # contains instructions on how to run the propagation evaluation
│
├── web_app
│   ├── README.md           # contains instructions on how to run the web app
│
├── .env                    # contains environment variables including API keys
├── .gitignore              # contains files and directories to be ignored by git
├── LICENSE                 # contains the license for this project
├── README.md               # contains the main documentation for this project

Co-occurrence Analysis

The co-occurrence analysis is performed using ...

• Datasets
  • Those included ...
  • Can be downloaded ...
• Keyword dictionaries
• Windows
• Filter
• Co-occurrence Analysis
• Results are saved in the co_occurrence_results folder.

Details on how to run the co-occurrence analysis can be found in the co_occurrence_generate/README.md file.

Logits Analysis

The logits analysis is performed using...

• models
• api vs hf
• languages, tf vs seq
• round robin templates
• Results are saved in the logits_results folder.

Details on how to run the logits analysis can be found in the logits_generate/README.md file.

Propagation Evaluation

The propagation evaluation is performed using...

• Joining the co-occurrence and logits results
• comparing windows, models, sizes, etc.
• Rank based approach
• Results are saved in the propagation_eval folder.

Details on how to run the analysis for propagation evaluation can be found in the propagation_eval/README.md file.

Web App

The web app is a fully functional web application that allows users to interact with the data and results of the project. The web app is built using Next.js, hosted on Vercel, and the backend is hosted on Heroku. The web app is accessible at https://crosscare.net.

Setting Up Environment

Managing API Keys

To securely manage your API keys, you can use environment variables. Follow these steps to set up the environment variables for this project:

1. Create a .env file:

In the root directory of your project, create a file named .env. Add the following lines to this file:

AZURE_KEY=your_azure_openai_key_here
OPENAI_KEY=your_openai_key_here

Replace your_azure_openai_key_here and your_openai_key_here with your actual API keys.

2. Load the variables in your Python script:

If you need these keys then at the beginning of your Python script, add the following code to load and access the environment variables:

from dotenv import load_dotenv
import os

# Load environment variables from .env file
load_dotenv()

# Access the variables
gpt4v_key = os.getenv('AZURE_KEY')
openai_key = os.getenv('OPENAI_KEY')

With these steps, your API keys will be securely stored in the .env file and easily accessible in your Python script.

Contributing

Contributions to improve the code or add additional functionalities are welcome! Please ensure to follow the existing code structure and comment appropriately.

Acknowledgements

The authors thank Oracle Cloud for computing the co-occurrence calculations for process $The Pile$ and the subset of $RedPajama$.

The authors also acknowledge financial support from the Woods Foundation (DB, SC, HA) NIH (NIH-USA U54CA274516-01A1 (SC, HA, DB), NIH-USA U24CA194354 (HA), NIH-USA U01CA190234 (HA), NIH-USA U01CA209414 (HA), and NIH-USA R35CA22052 (HA), NIH-USA U54 TW012043-01 (JG, LAC), NIH-USA OT2OD032701 (JG, LAC), NIH-USA R01EB017205 (LAC), DS-I Africa U54 TW012043-01 (LAC), Bridge2AI OT2OD032701 (LAC), NSF ITEST 2148451 (LAC) and the European Union - European Research Council (HA: 866504)