You bring up a good point. Without seeing their code and training metrics, how do we know that this isn’t some extremely overfitted model?