A learning signal turns raw data into directional instruction by pairing each input with an output—the intended answer. In supervised fine-tuning, it shows the model what behavior to learn. section 2.4.2
Useful signals are precise, consistent, and intentional. Ambiguous or conflicting labels teach the model those weaknesses, so annotation guidelines and standardization matter as much as model architecture. section 2.4.2
Choose the output form according to the intended use:
- Use compact labels such as "category: utilities" for transaction classification. section 2.4.2
- Use complete sentences when the model must produce readable client summaries. section 2.4.2
- Use machine-readable formats such as JSON when downstream systems require structured output. section 2.4.2



