DoctorGPT-An Open-Source Medical Dialogue Model

DoctorGPT is an LLM that can pass the US Medical Licensing Exam. It works offline, it’s cross-platform, & your health data stays private.

DoctorGPT: An Open-Source Medical Dialogue Model

DoctorGPT is an ambitious open-source project that employs a Large Language Model fine-tuned on a medical dialogue dataset using Reinforcement Learning and Constitutional AI. Its primary goal is to pass the US Medical Licensing Exam while maintaining offline operation for patient data privacy. This lightweight model, weighing just 3GB, is designed to fit on various devices, including iOS, Android, and Web platforms.

Features

The training process utilizes Google Colab Pro and takes approximately 24 hours to complete. Detailed usage instructions are provided for iOS and Android versions, enabling easy access and utilization. Furthermore, there’s potential for deploying DoctorGPT as a Flask API, allowing for online learning with human feedback for continuous improvement.

The project recognizes valuable contributions from organizations like Meta and MedAlpaca, among others, indicating collaborative efforts to enhance medical dialogue capabilities.

Built upon Meta’s Llama2 model, DoctorGPT has been meticulously trained on medical dialogue data. Reinforcement Learning and Constitutional AI further enhance its capabilities, enabling it to answer medical questions accurately and effectively. Its compact size makes it suitable for offline use, ensuring data privacy, and allowing users to access medical information without compromising sensitive patient data.

By offering versions for iOS, Android, and the web, DoctorGPT provides a wide range of options for users to access medical information conveniently. With the potential for future deployment as a Flask API and a React front-end for online learning and feedback, DoctorGPT’s impact could extend to continuous improvement through user interaction.

Dependencies

• Numpy (Use matrix math operations)
• PyTorch (Build Deep Learning models)
• Datasets (Access datasets from huggingface hub)
• Huggingface_hub (access huggingface data & models)
• Transformers (Access models from HuggingFace hub)
• Trl (Transformer Reinforcement Learning. And fine-tuning.)
• Bitsandbytes (makes models smaller, aka ‘quantization’)
• Sentencepiece (Byte Pair Encoding scheme aka ‘tokenization’)
• OpenAI (Create synthetic fine-tuning and reward model data)
• TVM (Tensor Virtual Machine, converts onnx model to efficient cross-platform use)
• Peft (Parameter Efficient Fine Tuning, use low rank adaption (LoRa) to fine-tune)
• Onnx (Convert trained model to universal format)

In conclusion

DoctorGPT is a significant open-source initiative that leverages advanced AI techniques to create a capable medical dialogue model. Its focus on accuracy, privacy, and accessibility positions it as a valuable tool for medical professionals, students, and anyone seeking reliable medical information.

References

• Source – Link to the original source of the information (if available)
• Meta – Acknowledgments to Meta for contributions
• MedAlpaca – Acknowledgments to MedAlpaca for contributions