To train the model, users can run the training notebook locally or remotely via a cloud service like Google Colab Pro. The training process requires a GPU and takes approximately 24 hours on a paid instance of Google Colab. For usage, the model can be downloaded onto an iOS or Android device and used for local chatting without an internet connection. The developers also plan to deploy the model as a Flask API with a React front-end for web usage, with the aim of using human feedback to further improve the model's performance through reinforcement learning.
Key takeaways:
- DoctorGPT is a Large Language Model that can pass the US Medical Licensing Exam and is designed to provide everyone their own private doctor. It's free, made for offline usage which preserves patient confidentiality, and it's available on iOS, Android, and Web.
- The model was fine-tuned on a Medical Dialogue Dataset, then further improved using Reinforcement Learning & Constitutional AI. It is only 3 Gigabytes in size, fitting on any local device, eliminating the need to pay an API to use it.
- Training the model requires a GPU and can be done locally or remotely via a cloud service like Google Colab Pro. The total training time for DoctorGPT took 24 hours on a paid instance of Google Colab.
- The model can be used on iOS and Android devices by downloading the respective Machine Learning Compilation Chat Repository, following the installation steps, and adding the latest DoctorGPT model. For web usage, the model can be deployed as a Flask API with a React front-end.