Few-shot learning teaches a model a new task from only a handful of supervised examples—too few for standard gradient descent. Instead of fine-tuning, a model can receive examples in a prompt and generate the requested output. section 16.1
Translation by example
Provide several English–French pairs, then leave the requested translation incomplete: “sea otter => loutre de mer,” “peppermint => menthe poivrée,” “plush giraffe => peluche girafe,” followed by “cheese =>”. The model infers the task and supplies an answer without fine-tuning. section 16.1
Translate English to French: sea otter => loutre de mer peppermint => menthe poivrée plush giraffe => peluche girafe cheese =>
Inspect the complete translation prompt and its demonstrations before testing the unfinished query. section 16.1
