Validation set
Beginner
Practice questions you keep hidden from the model, so you can test it fairly while it trains.
Novice
Examples held out from training and used only to measure progress, so you can see when the model stops improving on data it hasn’t seen.
Expert
A held-out split, made before any augmentation and deduplicated against training, used for early stopping and hyperparameter choice. Keep a separate final test set you look at rarely.
See also: Overfitting.