Skip to main content
All terms
Evaluation

Holdout Set

Data kept separate from training so a model can be tested on unseen examples.

Definition

A holdout set is data deliberately set aside and never used during training, so it can measure how well a model generalizes to unseen examples. Evaluating on held-out data guards against overstating performance, since a model can memorize its training set. Keeping the holdout truly separate is essential, as any leakage into training inflates the resulting scores.