carlwharris / cog-bias-med-LLMs

Addressing common clinical biases in medical language models

cog-bias-med-LLMS

Addressing common clinical biases in medical language models. Released as an arXiv preprint. Data available via Google Drive link here.

Installation

Prerequisites

The USMLE Question Processor requires several external libraries. These dependencies are listed in requirements.txt. To install them, follow these steps:

1. Clone the repository or download the source code.

2. Navigate to the directory containing requirements.txt.

3. Run the following command in your terminal or command prompt:

   pip install -r requirements.txt
   

Models

We include the following models in our implementation:

• OpenAI: gpt-3.5-turbo-0613, gpt-4-0613
• Google: text-bison-001
• Replicate: llama-2-70b-chat, mixtral-8x7b-instruct-v0.1
• HuggingFace: pmc-llama-13b, meditron-7b, meditron-70b

Examples

Loading the USMLE questions

We've included a copy of the USMLE questions in this repo. You can load a list of each question in json format as follows:

from med_llm_bias import load_usmle_questions
test_list = load_usmle_questions()
train_list = load_usmle_questions(question_set='train')
all_questions = load_usmle_questions(question_set='all)

Generating a biased prompt

from med_llm_bias import load_usmle_questions, USMLEQuestionProcessor
test_list = load_usmle_questions()
q_proc = USMLEQuestionProcessor(model_name=None, bias_type="confirmation", mitigation_strategy="few-shot")
prompt, info = q_proc.generate_full_prompt(test_list[0])

This will return a formatted prompt for using the few-shot mitigation strategy described in the paper, ready to be used as input to the API, as well as relevant information. Importantly, info['answer_idx'] contains the correct answer index (e.g., 'A', 'B', etc.), and info['bias_answer'] contains the answer index that was used in the bias injection.

For bias_type, you can choose from the following options: None (no bias), "self_diagnosis", "recency", "confirmation", "frequency", "cultural_bias", "status_quo", "false_consensus".

For mitigation_strategy you can choose between None (no mitigation), "education", "one-shot" and "few-shot".

Querying the API models

from models import llm_model
model = llm_model("gpt-3.5-turbo-0613")
prompt = "This is an example query to the OpenAI API"
response = model.query_model(prompt)

Configuring API Keys

About API Configuration

Our application uses various external services like OpenAI, Google, and Hugging Face models. To access these services, API keys are required. These keys are stored in api_config.json. You must obtain and configure your own API keys. Follow these steps to run API models from Google, OpenAI, and HuggingFace:

1. Rename .api_config.json to api_config.json: A template has been provided in .api_config.json. Rename this file to api_config.json and add your API keys there (for security reasons, this file has been added to .gitignore so it won't be added to your git repository).

2. Obtain API Keys:

   • OpenAI: Create an account at OpenAI. After logging in, access your API keys section and generate a new key.
   • Google Cloud Services: Go to the Google Cloud Console, create a project, and navigate to the 'APIs & Services' dashboard to get your key.
   • Hugging Face Models: Register at Hugging Face. Go to your profile settings to find your API keys.
   • Replicate: Sign up at Replicate. Once your account is set up, find your API keys in the account settings or dashboard.

3. Update the File:

   • Open api_config.json in a text editor.
   • Replace the placeholder keys with your own keys. For example, change "API_KEY": "sk-0NoVblPZ..." to "API_KEY": "your_actual_api_key_here".
   • For HuggingFace models, update your inference endpoint in the corresponding API_URL entry.