Fine-tuning with 100 examples: when small datasets win
How to know if you need 100, 1,000 or 100,000 training examples for your use case.
Conventional wisdom says fine-tuning needs thousands of examples. For style and format transfer in 2025, that's no longer true. A hundred carefully chosen examples can move a base model meaningfully toward your tone, your output schema and your domain vocabulary.
What matters more than quantity is consistency. Every example must follow the exact same input/output contract. One inconsistent label can drag the whole run in a direction you didn't intend. Curate ruthlessly; reject anything ambiguous.
Skip fine-tuning entirely if your problem is fact recall, novelty handling, or anything where the right answer changes over time. Those are RAG problems. Fine-tuning makes the model speak your language, not know your facts.