Golden Datasets

A golden dataset is a trusted, versioned set of inputs, expected behavior, labels, evidence, and risk metadata. It is not necessarily large; its value is that teams can rerun offline evaluation on the same decision cases and know what changed. The MLOps duplicate should usually point here for evaluation design, while evaluation datasets covers storage and pipeline ownership.

Anatomy of a golden row

For a support assistant, one row should identify the user query, source document ID, acceptable-answer criteria, refusal rule, slice tags, severity, reviewer owner, and version. A minimal acceptance contract is:

fieldexample
case_idSUP-legal-0042
input”Can I export customer data to a vendor spreadsheet?”
evidence_idpolicy/privacy/export-controls#2026-03
expected_behaviorrefuse unsafe export; cite approved workflow
riskcritical
slicelegal/privacy

The dataset should deliberately include common cases, rare cases, regressions, and high-severity failures from risk-weighted error taxonomies.

Worked inventory

For a ten-case regression gate, a deliberately curated inventory might look like this:

slicecommonrarehighest risk represented
account11high
billing11medium
legal11critical
returns11medium
security11critical

The risk counts are 3 low, 2 medium, 3 high, and 2 critical, so the critical share is . That would be wrong for population accuracy but right for a regression gate that must exercise rare dangerous behavior.

Caveats

Golden sets become stale when policies, products, source documents, or user behavior change. Keep a blind holdout so teams do not tune directly to the public examples. Use coverage reports to show what the set does and does not claim to represent.

References