Design and product work
Golden dataset
Also calledGolden set, Ground truth data, Answer key
A curated set of example inputs paired with expert-approved correct outputs, used as the answer key when evaluating an AI feature.
A golden dataset collects the cases that matter: common requests, tricky edge cases and known failures, each with the answer a domain expert agrees is right. Every change to the AI feature is scored against it.
It needs maintaining like any product asset. New failure modes found in real usage should be added, and answers revisited when the rules of the domain change.
Common mistake
A golden set built only from the cases the team thought of. Sample real requests too, or the set will miss what users actually ask.