A few dozen to a few hundred representative questions, each with the expected answer or the criteria for a good answer. It is replayed on every prompt or model change to detect regressions.
A good golden dataset covers frequent cases, edge cases and trick cases, and it grows with every bug found in production.