What is chatdoctor?

ChatDoctor is a research project from a team of medical and computer science researchers at several universities. It takes an existing general-purpose AI language model called LLaMA, developed by Meta, and further trains it on medical conversation data so it can respond to patient questions in the style of a doctor. The training data comes from three sources: 100,000 real patient-doctor conversations scraped from HealthCareMagic.com, 10,000 conversations from iCliniq.com, and a smaller synthetic set of 5,000 conversations generated with ChatGPT. Before training on any medical data, the model was first trained on a general conversational dataset from Stanford (called Alpaca) to give it basic language skills. The combination is meant to produce a model that knows how to hold a conversation and also has exposure to how doctors actually phrase diagnoses and recommendations. The project includes two extended variants: one that can look up a structured disease database to ground its answers, and one that can search Wikipedia in real time before responding. These are meant to reduce the chance of the model making things up. The repository provides scripts for fine-tuning the model on your own hardware using either full training or a lighter method called LoRA, which adjusts only a small fraction of the model's parameters and requires less computing power. Inference scripts let you run a local chat session once a model is set up. The authors are clear that the model has not reached reliable clinical accuracy and should not be used in real medical settings. It is presented as a research prototype to study how language models can be adapted for medical question answering, not as a replacement for professional medical advice.