A model asked to categorise support tickets returns the right categories in an inconsistent format, sometimes as a sentence and sometimes as a label. The instruction already describes the required format in detail. What is likely to help most?

• AAdding three worked examples showing input and the exact output format expected.
• BRaising temperature so the model explores other formats.
• CFine tuning the model on ten thousand historical tickets.
• DApplying chain of thought so the model reasons before answering.