What is Zero-shot / Few-shot Learning

Definition

Zero-shot and few-shot learning describe a model's ability to perform a task with no task-specific training examples (zero-shot) or just a handful provided in the prompt (few-shot), relying on knowledge gained during pre-training rather than dedicated fine-tuning.

« Back to Glossary Index

Performs new tasks without collecting and labelling training data
Adapts instantly via instructions or a few in-prompt examples
Reduces the cost and time of building task-specific models
Makes capable AI accessible for tasks with little available data

Real World Example

Given just two labelled examples in the prompt, a language model correctly classifies the sentiment of new support tickets, performing the task few-shot without any dedicated training run.

FAQs

What is the difference between zero-shot and few-shot?

Zero-shot provides no examples and relies on instructions alone, while few-shot includes a small number of examples in the prompt.

Why is this capability valuable?

It lets models tackle new tasks immediately without gathering labelled data or fine-tuning, saving time and cost.

When is fine-tuning still better?

When a task needs consistent, high-accuracy specialised behaviour that prompting with a few examples cannot reliably achieve.