GUIDE · AIunipedia Community
Zero-shot prompting asks a model to perform a task without task demonstrations in the prompt. Few-shot prompting adds examples of the desired input/output behaviour. Examples can clarify a format or label boundary, but they can also introduce an accidental pattern.
Suppose you classify support tickets into delivery, returns and payment. If every example with the word “late” is labelled delivery, the model may learn a shortcut that fails on “my refund is late”.
Build a controlled comparison
- Write a clear instruction-only baseline.
- Add a few representative examples without changing the task definition.
- Evaluate both versions on unseen cases.
- Inspect per-category errors rather than only a total score.
- Repeat after changing example order to test sensitivity.
Keep your demonstration set separate from the evaluation set. Do not choose examples by repeatedly looking at the final test answers. Include ambiguous cases and explain how they should be handled.
Report what actually changed
A useful experiment note includes the model/version, examples, evaluation cases, scoring rule and error breakdown. If the new version improves formatting but reduces classification accuracy, both findings matter.
The goal is not to add the maximum number of examples. It is to make the task clearer at an acceptable context and maintenance cost.
Discussion: What has helped more in your work: better instructions, better examples or better evaluation cases?
Background: Hugging Face LLM course.
An editorial starter discussion from AIunipedia Community, owned and run by Dr Fun and Michael Oracle. Share your own experience and help improve the explanation.
Continue exploring
A Practical Prompt Engineering Template · Temperature vs Top-p: What They Actually Control
Browse: Prompt Engineering · All learning paths
Tip: open a related discussion in a new tab to keep your place here. On a phone, touch and hold the link for your browser’s options.