Few-shot

Definition

Guiding a model with a handful of examples placed in the prompt, so it follows the pattern without any retraining. A quick way to steer behaviour.

Few-shot prompting shows the model two or three worked examples of the task, then the real input, and it imitates the pattern. It is faster and cheaper than fine-tuning for small adjustments, though fine-tuning wins when examples must be many or consistent.

Also known as

few-shot learning, in-context learning