What is this:
I created a chatbot that sounds like Dr Gregory House and has knowledge of various medical topics from Wikipedia.
How I made this:
Finetuning: I finetuned a Qwen3 4B model using Unsloth on the cloud using GCP's Compute Engine. I finetuned the model on House's personality by community made transcripts of episodes 3 to 9. I created training data based on these episodes. I also created synthetic training data using these episodes and included them in the finetuning training dataset.
RAG: This chatbot has access to various medical Wikipedia articles. Including categories such as: Medical diagnosis, Diseases and disorders, Rare diseases. I got these articles using Wikipedia's Export feature. I parsed the data using wikiextractor. Then upsert the data into a Pinecone vector database.
Tech stack: I'm using Next.js, React, Tailwind CSS, TypeScript, Python, FastAPI, Vercel, GCP, and Fly.io.
Attribution:
Background from Artsy.net. Taskbar image from Nintendofan12. Gregory House image from. MSN Avatar from FrutigerAero. Font is fs Tahoma 8px. Line, Square, Cross svg icons from freesvgicons.com.