Few-Shot Learning

Few-shot learning is when you show a language model a handful of examples of a task (2-5 examples) and it learns to perform the task on new inputs. "Here are three examples of converting requirements to user stories. Now convert this requirement." The model generalizes from the examples.

Few-shot learning is one of the most powerful practical techniques for using AI. It lets you teach the model custom formats, styles, and tasks on the fly without any training or fine-tuning. It works because large language models have learned general patterns that can be activated by examples.

Brown et al., "Language Models are Few-Shot Learners" (GPT-3) - https://arxiv.org/abs/2005.14165

Learn Prompting: Few-shot prompting - https://learnprompting.org/