Why should training, validation, and test datasets have different roles?

Training data fits model parameters. Validation data guides model selection, feature decisions, and hyperparameters. Test data provides a final assessment after those choices are fixed. Repeatedly using test results to revise the model tu s the test set into another validation set and makes the reported result optimistic. Splitting must reflect how data will arrive in practice, including time and repeated entities. With limited data, cross-validation can improve model selection, but an appropriate final independent evaluation still matters when making a credible deployment claim.