The goal
Fine-tune a model to be a medical-coding assistant that correctly identifies ICD-10 codes from clinical notes. Same recipe shape works for legal, financial, telecom, manufacturing — any high-stakes specialized domain.
Critical considerations
- Data source: annotated clinical notes with verified ICD-10 codes. Hand-curated by certified coders.
- Data quality: every example reviewed by a domain expert. No synthetic shortcuts on safety-critical data.
- Evaluation: exact-match accuracy on ICD-10 codes, F1 per category, edge-case behavior.
- Safety: include "I'm not confident" examples — the model must learn to express uncertainty rather than guess.