Generalisation

The ultimate goal of any ML model: Making accurate predictions on new, unseen data

To generalize well, the train and test set should follow similar distributions: $P_{\text{train}}\approx P_{\text{test}}$

Spotting overfitting and underfitting

In general, the train loss should have a smaller value than the test one (see Train and test split) but should remain on similar orders of magnitude.