Definition
Few-shot learning means adapting with a small number of labeled examples. With LLMs, those examples are often placed in the prompt (in-context learning) without updating weights.
Classic few-shot ML may also fine-tune or use metric learning on tiny datasets.
In simple terms
Showing a new hire three completed tickets and saying "do it like these" — they generalize the pattern without a full training course.
Where you see it
- Prompting an LLM with 3 Somali→English translation pairs.
- Classifying support intents from a few labeled messages in-context.
How it works
1.Select examples
Diverse, clear demonstrations of the task.
2.Format the prompt
Show input→output patterns, then the real query.
3.Generate
Model continues in the demonstrated style.
Why it matters
- Few-shot prompting unlocks new tasks without collecting thousands of labels.
Often confused
Few-shot always beats fine-tuning.
For stable high-volume tasks, fine-tuning or RAG may outperform prompt examples alone.