A focused experiment shows a compact 1.7‑billion‑parameter model outperforming a 235‑billion‑parameter model on structured SOAP note generation for clinical documentation by using reinforcement learning from AI feedback (RLAIF) within Snowflake ML.

Researchers built a synthetic dataset of nearly 20,000 doctor‑patient dialogues across 30 specialties and 400+ conditions to train and evaluate the models, ensuring diverse, clinically realistic scenarios with zero semantic duplicates.

Snowflake’s ML Jobs and secure compute enable an end‑to‑end workflow—from data synthesis to distributed RL training to evaluation—without leaving the platform, illustrating a scalable enterprise pattern: synthesize data, train with RL, and deploy cheaply.

Rather than traditional supervised fine‑tuning, they used GRPO (group relative policy optimization), where for each dialogue the 1.7B policy model generates four SOAP-note candidates and is scored via a deterministic JSON check (S, O, A, P) plus a frozen 8B judge assessing factual accuracy, completeness, and clinical appropriateness.

The study argues that scaling alone is insufficient for high‑stakes, narrow tasks like clinical documentation; task‑specific optimization yields greater reliability and efficiency at lower cost.

Practically, small, specialized models can run on a single GPU with subsecond latency and substantially lower costs, enabling production healthcare deployments.

The article provides code and training recipes in Snowflake’s sf-samples repository for practitioners to adapt this approach to other domain-specific structured outputs.

Across evaluations, the RL‑trained 1.7B model matched or surpassed the 235B model, with the largest gains in Assessment and Objective sections, and JSON format compliance reached 99.98%.

Evaluation notes acknowledge subjectivity in LLM scoring but emphasize the core claim: optimizing for a defined quality signal via RL lets small models achieve specialized mastery and superior task-specific performance over larger, generic models.