Fine-tuning data needs human judgments when examples lack the labels or rankings a task requires. For example, people can tag legal documents by issue area or rate responses for qualities such as politeness and thoroughness. section 5.3.1
Make annotation useful for fine-tuning by aligning it with the intended task:
- Use labels for the task the model should learn, such as legal-document issue categories. section 5.3.1
- Prioritize accurate, domain-appropriate examples over a larger volume of noisy data; involve domain experts in review where possible. section 5.3.1
- Correct inaccurate examples and filter irrelevant or sensitive information before training. section 5.3.1
