m′MoStudy.
Concept glossary/Training, validation, and test data

Training, validation, and test data.

Training data fits a model, validation data guides model selection, and held-out test data estimates performance after those choices have been made.

Put it into practice.

Fit preprocessing on the training split, then apply the learned transformation to the other splits. A future result or duplicate record must not leak into the features used to predict it.

A common mistake

Repeatedly tuning on the test set makes its score less useful as an independent estimate. Accuracy also needs context about class balance, error costs, and the intended use.

Original independent study explanation. Revised October 2026. Practice assumptions are illustrative.