Because the model doesn't overfit and that is good.

You don't want your model to mimic your trainning data. Otherwise, it would performe poorly with new data.