Zero-Shot and Few-Shot Methods
Definition
Zero-Shot and Few-Shot methods solve tasks without or with few task-specific examples, respectively.
In prompting, zero-shot means a task description without demonstrations; few-shot adds a few examples in context. This does not automatically change model weights.
Simply explained
Like being explained a new task only or additionally seeing some pattern solutions.
Why it matters in practice
Teams can try out tasks without immediately building their own model training.
Example from business
Hypothetical example: A classification prompt contains three examples for the desired ticket categories.
Limits and pitfalls
- Inappropriate examples can fix errors; the method does not guarantee sufficient quality for every task.
Origin & history
The terms originate from machine learning and became widespread especially for in-context learning with large language models.
Use in mid-sized companies
- Text classification, format specifications, and extraction with few demonstrations.
Frequently asked questions
Is few-shot prompting the same as fine-tuning?
No, examples in the prompt are to be distinguished from weight adjustment.
Sources
- Language Models are Few-Shot Learners · Brown et al.: Language Models are Few-Shot Learners –
Editorially reviewed on: