Flo Health Fine-Tunes Open AI Models Using Databricks

Flo Health fine-tuned an open-weight language model using Databricks AI Runtime to handle health queries for its 81 million monthly active users. By implementing a specialized post-training pipeline, the team increased the model's medical safety score from 34% to approximately 61% and its usefulness score from 42% to 76%. The combined score, which measures responses meeting both safety and usefulness standards simultaneously, climbed from 16% to 47%—nearly tripling the initial baseline.

To achieve these results without human-annotated training pairs, Flo built an automated evaluation framework powered by 26 distinct AI judges calibrated against clinician reviews. The setup included 12 medical-safety judges acting as hard constraints, 6 micro judges for voice, grounding, framing, and fine-grained safety, and 8 usefulness criteria covering completeness, relevance, coherence, and tone. Using privacy-first principles, Flo generated more than 15,000 synthetic questions to create preference pairs. The team executed distributed supervised fine-tuning and Direct Preference Optimization on 32 serverless NVIDIA H100 GPUs using PyTorch with Fully Sharded Data Parallel, integrated with LLM Foundry, Compose-RL, and MLflow.

During initial Reinforcement Fine-Tuning runs, the team noted that safety scores plateaued when preference pairs lacked medical contrast. Flo addressed this by introducing graded safety scoring to re-rank candidate responses that passed binary safety checks. Looking ahead, Flo plans to apply this reusable alignment loop to newer open-weight models, including Gemma, Qwen, and Kimi.

For enterprise machine learning practitioners, Flo's approach demonstrates how domain-specific AI judges and synthetic datasets can eliminate reliance on slow, expensive human labeling during alignment. Furthermore, deploying custom fine-tuned models rather than relying on frontier model APIs can reduce operational inference costs by an order of magnitude. Unifying data, experimentation, and distributed GPU orchestration on a platform like Databricks AI Runtime allows engineering teams to rapidly iterate on post-training workflows without manually standing up cluster infrastructure.