What is a golden dataset?

A golden dataset is a set of realistic inputs paired with reference answers that domain experts have reviewed, which you run your AI system against to judge its outputs. Microsoft’s guidance for Copilot quality assurance defines it as “a collection of realistic customer questions along with expert answers.”

It is a test set, not a training set. The same guidance says golden datasets are not used as source data to train or fine-tune a model or to inject context into a prompt. They are used to “assess the quality of the answers that come back from the LLM.” For the wider practice, see LLM evaluation.

What goes into one

A datapoint needs an input and something to compare the output against, which ground truth supplies. HoneyHive’s dataset docs split each datapoint into input, ground truth, chat history and metadata, where ground truth is “the expected or ideal outputs or reference answers for a given input.”

Two sources agree on what the inputs should look like:

• Anthropic advises designing evals that “mirror your real-world task distribution” and adds: “Don’t forget to factor in edge cases!” Its list includes irrelevant input, overly long input and ambiguous cases where even humans would struggle to agree.
• OpenAI’s guide asks for “typical cases, edge cases, and adversarial cases” and warns against datasets that “don’t faithfully reproduce production traffic patterns.”

How to build one

Microsoft’s guidance gives four steps:

1. Write the questions users would realistically ask, without starting from the grounding data. Each question should stand alone with enough context to answer.
2. Optionally have an LLM draft answers as a starting point. Experts must review them.
3. Have a domain expert write or correct each answer so it is accurate and complete.
4. Optionally have a second expert do a QA pass.

Expert labeling costs time. OpenAI’s guide notes that expert annotation “is expensive and time-consuming,” and says to calibrate automated scoring against human judgments before trusting it, scaling up an LLM judge only once it “consistently agrees with human annotations.” See LLM-as-a-judge and human evaluation. When several people label, inter-annotator agreement shows whether they label alike.

How many items

Published guidance does not agree, and the answers depend on the system.

Source	Size guidance
Microsoft’s guidance	About 150 question and answer pairs, certainly not fewer than 100
HoneyHive’s cold start post	20 to 50 hand-reviewed examples was the traditional start for a narrow agent, and the number keeps climbing as agents get more general
Anthropic’s eval guide	Prefers volume: more questions with slightly lower signal automated grading beat fewer hand-graded ones
OpenAI’s guide	No general rule

The sources differ because they describe different trade-offs. A small hand-reviewed set is cheap and trusted. A large auto-graded set catches more variation but depends on the grader.

The cold start problem

A golden dataset built from production traces needs a system in production, and shipping seems to need evals first. HoneyHive’s post by Sunny Bakhda in July 2026 describes this loop: production needs eval coverage, coverage needs curated datasets, and datasets need production traces.

Its advice is to start with a small hand-reviewed set, use synthetic cases only to bootstrap, and get to production “as fast as you safely can” with enough visibility to learn. Citing Hamel Husain and Shreya Shankar, it says synthetic data “generates the inputs you already thought of,” while real users bring the ones you didn’t. OpenAI’s guide gives the same direction: “Log as you develop so you can mine your logs for good eval cases.” Once real failures show up, error analysis tells you which ones to add.

Keeping it useful

The sources above do not prescribe a maintenance schedule. Microsoft’s guidance does say that if quality keeps drifting, the golden dataset may be out of date and should be revised. Treat these as practices to adapt, not rules:

• Add real failures from production as they appear, and replace synthetic items with real ones.
• Keep the set fixed while you compare two versions of a prompt or model, so score changes come from the change. Prompt versioning covers tracking which version produced which run.
• Keep it out of any fine-tuning data, as Microsoft’s guidance says golden datasets are for assessing answers only.

Building one in HoneyHive

HoneyHive’s Curate from Traces page describes filtering sessions, model events or spans in Traces and choosing Add to Dataset. Its use-case table includes “capture successful interactions as golden test cases.” Each curated datapoint keeps a linked_event field that points back to the original trace. You can also upload JSON, JSONL or CSV files to start before any traces exist.

For regression checks, the CI regression docs describe a baseline run produced by evaluate() over a dataset, which a pull request run is compared against with per-metric deltas.

FAQ

What is a golden dataset?

A set of realistic inputs paired with expert-reviewed reference answers, used to score an AI system's outputs. Microsoft's guidance defines it as "a collection of realistic customer questions along with expert answers."

How big should a golden dataset be?

Sources disagree. Microsoft's guidance says to aim for about 150 question and answer pairs and no fewer than 100, while HoneyHive's cold start post calls 20 to 50 hand-reviewed examples the traditional starting point for a narrow agent.

Is a golden dataset training data?

No. Microsoft's guidance says golden datasets are not used as source data to train or fine-tune a model, only to assess the quality of its answers.

Can I build a golden dataset with synthetic data?

For a start, yes. HoneyHive's cold start post calls synthetic cases "a cold-start tool, not a substitute for production signal", so replace them with real traces as they arrive. See synthetic test data for evals.

Can an LLM write the reference answers?

It can draft them, but experts must review them. Microsoft's guidance lists LLM drafts as an optional step that experts must check before use.