AI, explained
What is a golden dataset?
A golden dataset is the only reason you will ever be able to say whether a change made the system better or just different.
What it is made of
Each case has three parts: the input exactly as it arrives in production, the expected output or the criteria a correct output must satisfy, and a note on why. That third part is what stops the set decaying, because in six months nobody remembers why a borderline case was labelled the way it was, and without the reasoning the label becomes unarguable folklore.
Who writes it
The domain expert, not the engineer. The person who does the job today is the only one who knows that an invoice with two purchase order references means the second one is the amendment, or that a particular phrasing in a support ticket always signals an escalation. Engineers writing their own evaluation cases produce a set the system passes by construction, which is worse than no set at all because it manufactures confidence.
How many cases
Fewer than people expect to start. Fifty to a hundred well-chosen cases catch most regressions on a narrow task, and a small set that actually gets run beats a large set nobody maintains. What matters more than volume is composition:
- Common cases, weighted roughly as they appear in real traffic.
- Known hard cases, the ones that generated complaints or reworks.
- Edge cases with genuinely ambiguous correct answers, where the expected behaviour is to flag rather than to guess.
- Cases where the correct answer is a refusal, because a system that never abstains will confidently answer things it should not.
When to rerun it
- Before and after any prompt change, since prompt edits routinely fix one case and break two.
- On every model version change, including provider-side updates you did not initiate.
- After the retrieval corpus changes materially, because adding documents shifts what gets retrieved for existing queries.
- On a schedule regardless, monthly is usually enough, to catch drift nobody triggered.
Common mistakes
Buying or generating the set defeats the purpose, since the point is that it encodes your specific definition of correct. Letting it leak into development is the other failure, because once engineers optimise directly against the cases, the score stops predicting production behaviour. Keep a holdout portion that is scored but never inspected during development. Treat the set as an asset with an owner and a review date, the same as any other production artefact.
An AI audit includes building the first version of this with the people who do the work.
Frequently asked questions
What is a golden dataset in AI?
A fixed set of production-realistic inputs paired with correct outputs and the reasoning behind each label, used to score an AI system objectively. Prompt changes, model upgrades and retrieval changes are all measured against it, which converts improvement from an impression into a number you can defend.
How many cases does a golden dataset need?
Fifty to a hundred well-chosen cases catch most regressions on a narrow task, and a small maintained set beats a large abandoned one. Composition matters more than size: common cases weighted as in real traffic, known hard cases from past complaints, genuinely ambiguous edge cases, and cases where the correct answer is a refusal.
Can we buy a golden dataset?
Buying or generating one defeats its purpose, because the value is that it encodes your own specific definition of a correct answer. The domain expert who does the job today has to write it. Engineers writing their own cases produce a set the system passes by construction, which manufactures confidence rather than measuring it.
Related
Ready to put AI to work?
Book a discovery audit and we will map the highest-ROI AI agents and automations for your business.
Book a discovery audit →