On the choice of training data for machine learning of geostrophic mesoscale turbulence