DoctorGPT

DoctorGPT is an open-source project developed by llSourcell that aims to provide a Large Language Model capable of passing the US Medical Licensing Exam. 

It is based on Meta's Llama2 model, which has been fine-tuned on a Medical Dialogue Dataset and further enhanced using Reinforcement Learning and Constitutional AI. 

The model is designed to be compact, fitting within 3 gigabytes, allowing it to be used offline on various platforms including iOS, Android, and Web without the need for an API.

The project's mission is to offer a personal "doctor" for users, preserving patient confidentiality by operating offline. It can be accessed and used for free. Users are encouraged to contribute to the project through feature additions and improvements via pull requests.

Dependencies for DoctorGPT include libraries like Numpy and PyTorch for matrix math operations and building deep learning models, respectively. It also uses various tools such as HuggingFace's datasets and transformers, as well as Trl (Transformer Reinforcement Learning) for fine-tuning, and Onnx for converting trained models to a universal format.

For training purposes, users can run the provided training notebook either locally or remotely on cloud services like Google Colab Pro, which requires a GPU. The training process involves supervised fine-tuning on custom medical data and reinforcement learning from Constitutional AI Feedback, taking about 24 hours on a paid Google Colab instance.

The model is available for usage through HuggingFace's model hub. There are versions tailored for mobile, including iOS and Android. Installation and usage instructions are provided for each platform. Additionally, there are plans to deploy the model as a Flask API with a React front-end to facilitate online learning through human feedback and reinforcement learning.

The project credits Meta, MedAlpaca, Apache, MLC Chat & OctoML, and aims to offer a versatile, offline, and privacy-preserving medical language model for various platforms.