Comment by bassie123

I'm pretty sure that's how they score in the training phase. Lower token output = higher score.