Label granularity is the specificity represented by categories or annotations in a supervised fine-tuning dataset. Coarse labels simplify decisions; fine-grained labels preserve nuance but increase annotation difficulty. section 2.4.2
Choose granularity by balancing practical constraints against the desired behavior:
- Use broad categories such as compliant or risky when consistent, operationally simple decisions matter most. section 2.4.2
- Use detailed categories when predictions must distinguish conditions, mixed signals, uncertainty, or reasoning components. section 2.4.2
- Define labeling guidelines and review 50–100 examples with a domain expert before full annotation. section 2.4.2
Investment research sentiment
A coarse scheme labels a note positive or negative. A finer scheme adds neutral, mixed, and uncertain categories, or separates positive earnings sentiment from negative cash-flow sentiment. section 2.4.2












