Skip to content

A Daily Lexicon of Trustworthy Data

006·36

eval set

/ɪˈvæl sɛt/ - n.

1 [colloq.] The exam the model is allowed to study before every retake.Keep. Punchy.This is the problem.

Working definition

2. A fixed, labeled collection of cases used to measure a model's quality consistently across versions.

Evidence

See also

benchmark saturationEveryone got an A, so the test resigned.
eval harnessA scoreboard everyone trusts and no one has read the rules of.
model evaluationThe benchmark the model passes, chosen after seeing which benchmark it passes.