Doctor Dignity is an open-source project aiming to provide a private, offline, cross-platform LLM capable of passing the US Medical Licensing Exam. It's designed for users prioritizing data privacy and local execution, offering a free alternative to cloud-based medical AI.

How It Works

This project fine-tunes Meta's 7B parameter Llama2 model using a Medical Dialogue Dataset, enhanced with Reinforcement Learning and Constitutional AI. The model is quantized to 3GB for efficient local deployment across iOS, Android, and Web platforms, leveraging TVM for cross-platform optimization and ONNX for model conversion.

Quick Start & Requirements

• Install dependencies: pip install numpy torch datasets huggingface_hub transformers trl bitsandbytes sentencepiece openai tvm peft onnx
• iOS: Requires cloning the repo, downloading weights, and building the TVM runtime with specific MLC LLM wheels.
• Training: Requires a GPU (e.g., Google Colab Pro) and involves running training.ipynb.
• Usage: Models are available on Hugging Face (llsourcell/medllama2_7b).
• Links: Hugging Face Model, Training Notebook

Highlighted Details

• Passes US Medical Licensing Exam.
• 3GB model size for local, offline use.
• Cross-platform support (iOS, Android, Web).
• Fine-tuned with medical data and RLHF.
• Encourages community contributions via pull requests.

Maintenance & Community

• Project encourages pull requests for improvements.
• Credits Meta, MedAlpaca, Apache, MLC Chat & OctoML.

Licensing & Compatibility

• License details are not explicitly stated in the README, but dependencies include Apache 2.0 licensed components. Compatibility for commercial use or closed-source linking requires clarification.

Limitations & Caveats

The project includes a strong disclaimer stating that advice should not be taken seriously due to its work-in-progress nature, warning of potential serious injury or death. Android and Web app deployments are marked as "TODO".