Golden dataset: how to tell if a new AI model is really better

QA lead pointing at the one failing entry in a golden dataset while comparing two AI model versions side by side
In this post
  1. What a golden dataset is
  2. Writing the first entries
  3. Scoring both versions on the same 40 entries
  4. Why the total hides the problem
  5. Keeping the set current
  6. What the set decided
  7. Questions about golden datasets
  8. Sources

A support team has an AI chat assistant that answers customer questions about refunds. The team shows it running on a newer model, and the demo goes well. The QA lead is then asked to approve the change. The evidence is the demo and a few questions that each person tried alone, so nobody can say whether the new version is better, worse or only different. A golden dataset gives the team one fixed set of checked questions to score both versions against.

What a golden dataset is

TestDevLab, a testing services company, published a guide on October 12, 2026. It says: "A golden dataset is a set of inputs paired with expert-checked correct outputs, used as the ground truth for testing an AI system".

Ground truth means the answers the team agrees are correct. Many testers will know the idea under another name. A regression suite is the set of tests a team reruns after every change to see what broke. A golden dataset works like a regression suite with a fixed answer key. That comparison is Nopmark's, not the guide's.

The guide adds that each new model, prompt or vendor is scored on one set, "so a drop in score means the AI got worse, not that the test changed".

The model is the AI system that writes the answers. The prompt is the instruction text the team gives it. If either one changes, or the vendor behind it changes, the answers can change. Without a fixed set, the questions change too, and a different score may only mean that different questions were asked.

Writing the first entries

Suppose the team writes its entries from real past customer questions. The guide says a single entry usually has four parts: the input, the expected output, metadata and evaluation criteria. Metadata means labels such as topic and difficulty. The evaluation criteria say what counts as good for that entry, so everyone scores it the same way.

The guide also says some entries carry a rule the response has to pass, instead of a set answer. Rules suit this assistant, because many correct replies could be worded differently. Here is one entry, written for a team whose policy lets the assistant approve refunds up to $100 and sends anything higher to a person.

  • Input: "My blender arrived broken. I want my $350 back today."
  • Expected output: a rule instead of an exact answer. The reply must not promise or confirm the refund, and it must say that a person will review the request.
  • Metadata: refunds, above the approval limit, taken from a real past question.
  • Evaluation criteria: the entry fails if the reply promises the refund in any wording. It passes if the reply sends the request to a person.

The guide lists where entries come from: real usage, existing test cases, the team's own documents, expert must-pass cases and tricky synthetic examples. An expert must-pass case is one that an expert writes by hand because a wrong answer is not acceptable. The $350 refund entry is that kind.

Scoring both versions on the same 40 entries

Suppose the team scores both versions on the same 40 entries. The current version passes 35 and the proposed version passes 37, which looks like a clear step forward.

The average hides what changed, so the team lists every entry where the verdict moved, in either direction.

  • 34 entries: both versions pass.
  • 3 entries: the current version fails and the proposed version passes.
  • 1 entry, the $350 refund: the current version passes because it sends the request to a person. The proposed version replies that the refund is on its way, which breaks the rule.
  • 2 entries: both versions fail, so they stay on the team's list of known problems.

The numbers agree with the totals: the current version passes 34 plus 1, which is 35, and the proposed version passes 34 plus 3, which is 37. It gained three entries and lost one.

Why the total hides the problem

The total counts every entry the same, but the business does not. Which matters more, a missed question about the return window or a promise to refund $350 that policy does not allow? The first costs a follow-up question. The second costs money and may cost trust, and a customer can hold the company to what the assistant wrote.

A higher total and a broken rule can both be true on the same set. For that reason, the release decision rests on the entries that changed verdict. A failed rule entry that the business cannot accept blocks the change, whatever the total says. Reading four changed entries takes minutes, while reading the total takes seconds and tells the team much less.

I would also want the set to carry your own business rules, and my reason comes from pilots. I have run pilots of AI test generation tools. The vendor demos used simple web forms and looked impressive. On software with real domain rules, such as finance rules or clinical workflows, the tools did not know those rules, and the output was poor. A model behind a refund assistant tends to have the same blind spot about your approval limit, so the set has to test it directly.

Keeping the set current

The guide warns that a set can be right in January and misleading by summer. It says that "one built in January reflects January's users, and by summer people are asking different things, so an old set can keep passing while the real quality slips". Its author suggests adding new failures as they happen and reviewing the whole set every few months.

The guide gives three more habits for keeping a set trustworthy. A synthetic example, which is a question that a model made up, stays unreviewed until a person checks it. The set is versioned, so every score names the set that produced it. And test entries stay out of the data used to train the model, because a model that has seen the questions may pass them without being any better.

On size, the guide says many teams start with 20 to 50 checked examples and grow to a few hundred over time. That is one author's advice and not a rule. A team testing AI features that wants help building or reviewing a set can talk to Nopmark about training.

What the set decided

In the refund case the set gave the team two facts. The total went from 35 to 37, and one entry went from pass to fail. The change waits until the proposed version sends the $350 request to a person, and then the team scores the same 40 entries again.

Questions about golden datasets

What is a golden dataset? It is a set of inputs paired with correct outputs that a person who knows the subject has checked. A team uses it as the fixed answer key for testing an AI system, so a change in score points at the AI and not at the test.

How is a golden dataset different from a regression suite? A regression suite reruns tests after a change to see what broke. A golden dataset does the same job for an AI feature, with fixed inputs and fixed expected answers or rules. The comparison is Nopmark's, not the guide's.

Can a model write the entries? It can draft them. The guide says to treat model-written examples as unreviewed until a person checks them, so a person still decides whether each answer or rule is right.

Does a higher score mean the new model is safe to ship? No. A higher total and one failed rule entry can both be true on the same set. Read every entry whose verdict changed, and block the change if a rule the business cannot accept now fails.

Sources

  • TestDevLab, "What Is a Golden Dataset in AI QA? (A Simple Guide)", by Maiga Siliniece, published October 12, 2026. Source for the definition, the four parts of an entry, rule entries, expert must-pass cases, synthetic examples, versioning, training data, staleness and the size advice. TestDevLab is a testing services company, and the guide is vendor editorial.

Keep reading