Could be that many presumed ML experts don't even know how to train on the evaluation set