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.