Golden Dataset
also: golden set · eval set
A curated set of inputs with known-good answers, used as the ground truth for evaluating a model or pipeline.
A golden dataset (golden set) is a curated collection of representative inputs paired with known-correct outputs, used as the ground truth to evaluate a model, prompt, or RAG pipeline. It is what turns “seems better” into a measurable regression test for AI systems.
Worked example: before changing a prompt you run it against 200 golden question/answer pairs and compare scores — if accuracy drops, you caught the regression before shipping, just like a unit-test suite. Gotcha: a golden set is only as good as its coverage and freshness — if it misses the cases users actually hit, a green eval means nothing — and it must be kept out of training data, or you are grading the model on its own homework.