Golden datasets: knowing an AI system works

A golden dataset is a small collection of real inputs, each with the behaviour you expect. Run the system on it before every release and you know whether a change made things better or worse.

Why a small set beats a big guess

Trying a few questions by hand feels like testing, but it changes every time and nobody writes it down. Thirty carefully chosen cases, run the same way every time, give you a number you can compare across releases.

What goes in it

Kind of caseWhy it is there
Common requestsThe bulk of real traffic; the system must handle these well
Edge casesUnusual but valid inputs that break naive systems
Must refuseQuestions with no answer in the data, or that the user may not ask
AdversarialInputs that try to make the system ignore its instructions

Take cases from real data wherever possible: support tickets, search logs or questions users actually asked. Invented examples tend to be easier than the real thing.

Expected behaviour, not exact wording

A model rarely produces the same sentence twice, so compare behaviour rather than text. Did it pick the right category? Did it cite the right document? Did it refuse when it should?

  • Exact checks for things that must match: a category, a number, a tool that must be called.
  • Rule checks for things you can test in code: a citation is present, no personal data appears.
  • A judging model for open answers, scored against a written rubric, and checked against human labels before you trust it.

Agree the threshold

Who decides what passes
The pass threshold is a business decision, so agree it with the customer before launch. Agreeing it after the first complaint turns every result into an argument.

Keep it honest

  • Add every real failure you find in production as a new case
  • Keep some cases aside and never tune against them, so the set cannot be memorised
  • Record the score for every release so you can see the trend
Worth remembering
  • Thirty real cases run the same way beat a hundred tries by hand
  • Check behaviour, not exact wording
  • The threshold is agreed, not invented
Try it yourself
  • Write ten cases for a system you have built: six common, two edge, two that must be refused
  • Decide how each one will be checked: exact, rule or judge

Every expert started right here.