Choose examples that do not skew the answer
In a nutshell: Pick and order few-shot examples so the model learns the task rather than the distribution
Use it when: A prompt carries labelled examples and the model favours one label regardless of the input.
- MIT
- Sources
- 1
- Rules
- 2
- Memos
- 1
- Field
- Prompt engineering
- Authors
- DAIR.AI
Contains 2 Rules
The advice here is to avoid skewing the distribution and instead provide a more balanced number of examples for each label.