loading
loading
A curated, hand-verified set of inputs and known-good outputs you trust as ground truth for evals.
It's the answer key your eval harness grades against — a few dozen to a few hundred representative cases (including the nasty edge cases that bit you in prod) with outputs a human signed off on. Quality beats quantity: a tight, correct golden set catches real regressions, while a sloppy one teaches you nothing. Example: 80 support tickets paired with the exact correct routing label, frozen so every prompt change is scored on the same bar.
Plainly
Think of Golden Dataset as a simple recipe for doing the work better. A curated, hand-verified set of inputs and known-good outputs you trust as ground truth for evals.
In practice
Use it when you need a repeatable method instead of guessing from vibes. In practice, define the owner, input, output, and failure mode before you rely on it.