Definition
Plain language
A set of examples held back for the final grade, never used while building or tuning.
As stated in the literature
Held-out split used only for final evaluation; kept hidden from the Researcher to prevent selection on the reported metric.
Also called: test sets
Why it matters: If anyone peeks at or tunes against the test set, the final score stops measuring real ability and starts measuring how well the system was fitted to those particular examples.
For example, a thousand examples are locked away at the start of a project and only opened once, at the very end, to produce the number that gets reported.
Heard on the show
“Meanwhile, factual accuracy on a separate test set barely moves — sixty-nine percent to seventy.”Episode 004 — The Sycophancy Circuit That Survives Alignment Training