Skip to content
ML Academy

Introduced on day 51

Golden set

A frozen, versioned set of questions with labelled evidence passages, expected answers, query-type tags and answerable flags, used to compare system versions.

In detail

Label evidence as spans of the source document so the same set scores any chunking. Include unanswerable questions and tag query types so results can be sliced. A golden set loses its value as a fair test once it has been tuned against repeatedly, so keep a held-out part and record how many variants were tried.