hard20 pts · first answer countsprompt-engineeringfew-shot

Examples with random labels

A 2022 study ("Rethinking the Role of Demonstrations") found that replacing the correct labels in few-shot examples with random labels often hurt classification accuracy only slightly. What does this suggest?

A.Examples teach format and label set, not only the mappingB.Labels in examples are always ignored, so remove themC.Wrong labels improve accuracy and should be preferredD.Few-shot prompting has no effect on model output

Discussion 0

Join free to post in the discussion.

Examples with random labels

Discussion 0

Related questions