Fine-Tuning Small Open-Source LLMs to Outperform Large Closed-Source Models by 60% on Specialized Tasks
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Parsed, using Together AI's fine-tuning platform, fine-tuned a Gemma 3 27B open-source model to outperform Claude Sonnet 4 by 60% on a healthcare clinical scribing task, at 10–100x lower inference cost. The approach centers on three pillars: high-quality task-specific data curation, a rigorous multi-dimensional evaluation harness built in collaboration with clinical experts, and iterative SFT followed by reinforcement learning using the evaluation harness as a reward signal. Before fine-tuning, Gemma 3 27B performed 35% worse than Sonnet 4; after fine-tuning on tens of thousands of harness-optimized examples, it dramatically surpassed all tested proprietary models. The evaluation framework decomposes clinical note quality into granular binary checks across dimensions like clinical soundness, source fidelity, and template conformance, enabling reliable RL optimization. The case challenges the assumption that large proprietary models are necessary for complex domain tasks.
