Golden Dataset

A golden dataset is a curated set of inputs paired with known-correct outputs, used as the reference standard for evaluating an AI model or agent.

Because the expected answers are vetted in advance, a golden dataset lets you measure accuracy objectively and detect regressions between versions.

Keeping a golden dataset held out and confidential prevents agents from being tuned to it, the same principle behind a sealed task bank.