Red Hat Interview Question

What would you do if an ML model performed better on testing data than training data?