Project 7: Capstone — Personal Health Assistant¶
Mentors: To be announced
Problem: No single project so far has combined prompting, RAG, dataset lookup, and fine-tuning into one assistant that's actually grounded in a specific user's own documents.
Context: LoRA fine-tunes a small open model (e.g. Phi-3-mini) on MedQuAD (public medical Q&A pairs), plus a RAG layer over the user's own uploaded documents, served entirely locally. The lightest-weight fine-tuning seen in the track, scaled up to a generative model instead of a classifier.
Goals: Combine everything from Projects 1-6 (prompting, RAG, dataset lookup, and fine-tuning) into one assistant: a base model specialized on medical Q&A via LoRA, grounded further at answer time by whatever documents the user uploaded.
Deliverables: A full chat app with document upload, built in stages: (1) LoRA fine-tune Phi-3-mini on MedQuAD using an open-source LoRA library (e.g. peft), (2) a local RAG layer (Chroma, as in Project 1) over uploaded documents, (3) retrieval of relevant document chunks at query time alongside the fine-tuned model's own knowledge, (4) all wired behind a chat UI, served locally so no document or query leaves the machine. Mirrors Basic Science's capstone (Project 10) in spirit: the project that makes every earlier technique legible as one system.
Showcase: TBD
References:
- A Question-Entailment Approach to Question Answering (BMC Bioinformatics, 2019): the paper behind the MedQuAD dataset this project's LoRA fine-tune trains on.
- MedAlpaca: An Open-Source Collection of Medical Conversational AI Models and Training Data (code): an open-source project that LoRA fine-tunes a small open model on medical Q&A data for local, privacy-preserving deployment, the closest existing analog to this capstone's approach.