The core move

Split your data: train, then test.

You train the model on one slice of data and test it on a different slice it never saw. Scoring well on unseen data is the only result that counts — performance on the training data proves nothing.