BiMediX2-8B Technical Documentation

Introduction

BiMediX2 is a language model designed for bio-medical applications, developed by the Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), UAE. It specializes in processing diverse medical modalities and is tailored for English text.

Architecture

The model is based on the meta-llama/Llama-3.1-8B-Instruct architecture and is designed to handle image-text-to-text tasks. Key areas of focus include biology, healthcare, and medical applications.

Training

BiMediX2 was trained to function as a Language Model for Medical Modalities (LMM). It aims to assist in various medical tasks, though it has limitations typical of language models, such as potential for hallucinations and biases.

Guide: Running Locally

To run BiMediX2 locally, follow these steps:

1. Clone the Repository: Access the BiMediX2 repository on GitHub and clone it to your local machine.
2. Install Dependencies: Ensure all necessary packages and libraries are installed, as listed in the repository.
3. Download Model Weights: Obtain the model weights, available from the Hugging Face model hub.
4. Execute the Model: Use the provided scripts to run the model on your data.

For optimal performance, consider using a cloud GPU service like AWS, Google Cloud, or Azure.

License

BiMediX2 is released under the CC BY-NC-SA 4.0 license. This model is intended for research use only and is not suitable for clinical or commercial applications. Users should verify outputs with healthcare professionals and use the model responsibly. For more details, consult the license file.