Introduced on day 47
Few-shot prompting
Including a handful of worked input-output examples in the prompt so that the model infers the format and label set without any weight update.
In detail
Demonstrations mainly communicate the output format, the label space and the input distribution; they are weaker at teaching new rules. Accuracy can vary with the choice of examples, their order and formatting, so evaluate several variants and draw examples from a pool disjoint from the held-out evaluation cases.