Why is a test dataset kept separate during model development?
A.To speed up training
Splitting isn't about speed.
B.To measure final performance on data the model has never seen — unbiased evaluation
Correct. Unseen data = honest generalization measure.
C.To reduce storage costs
Storage isn't the reason.
D.To tune hyperparameters
Hyperparameter tuning uses the VALIDATION set — test stays untouched until the end.
Why this answer
The split exists to detect overfitting: train set teaches the model, validation guides tuning, and the held-out test set gives an honest generalization estimate. Evaluating on training data inflates apparent performance.
Correct answer: B · Last reviewed 2026-10-28
Official AWS references
Practice this interactively
Answer without seeing the solution first in the AI/ML Foundations practice session.
Practice AI/ML Foundations