One issue is that this doesn't punish the model of the top prediction produces complete nonsense. I can imagine getting good results when fine tuning an existin

One issue is that this doesn't punish the model of the top prediction produces complete nonsense. I can imagine getting good results when fine tuning an existing model though