Train, Validation, and Test Splits

The three common splits have different jobs.

Training data updates parameters. Validation data guides decisions during development. Test data is held back for a final estimate after model choices are mostly fixed.

trainvalidationtestupdates parametersguides decisionsfinal check
Train, validation, and test splits answer different questions.

If the test set is used repeatedly to choose models, it stops being a clean final check. It becomes part of development.

The split discipline is not bureaucracy. It protects you from believing a model generalizes because you accidentally tuned it to the evaluation set.

Exercise: Which split updates parameters?

Enter 1 for train, 2 for validation, or 3 for test: this split is used for parameter updates.

Compute it first, then check your number.

Exercise: Final estimate

Enter 1 for train, 2 for validation, or 3 for test: this split should be saved for the final estimate.

Compute it first, then check your number.