Problem
Flo Health, a women's health app with about 81 million monthly active users, runs a compound system with several generative components. Each answer must be medically safe and also useful, and one closed foundation model is risky and costly at this scale. Fine-tuning an open model needs labeled data, but a gynecologist costs about $400 to $600 an hour, and users' free-text data is off-limits. The starting model had 34% medical safety and 42% usefulness, and only 16% of answers were both.
Example from the article: asked "why am I spotting?", the model named only a benign or only a serious cause, never a balanced list. This theme became the "balanced causes" rule and then a judge.
Solution in brief
Expert time goes into designing and calibrating LLM judges instead of labeling. The judges score synthetic answers, and the scores pick chosen and rejected answers for supervised and preference fine-tuning of Llama 3.3 70B, following the TAU recipe from Databricks Research.
Questions this case answers
- How to fine-tune an open LLM for a consumer advice product when expert labels are too expensive and user data cannot be used
- How to build and calibrate LLM judges for domain safety rules and combine a safety score with a usefulness score into preference pairs
- How to fine-tune an open LLM for a consumer product using LLM judges instead of expert labels
- How to build and calibrate a set of LLM judges for domain-specific safety rules
Tools and methods
- LLM judges — one per rubric rule, calibrated on doctors' labels
- TAU (Test-time Adaptive Optimization) — sample N responses, score, fine-tune toward the best
- Synthetic prompts from user themes — no user free text
- Iterative refinement and diversity sampling — candidate answers
- DPO and supervised fine-tuning — on Databricks AI Runtime with LLM Foundry and Compose RL
Google or a one-time email code. No password.
