To estimate model performance more reliably by using multiple train/validation splits.

To make all features categorical.

To remove the need for a test set in every project.

To hide evaluation results from the researcher.