So you are saying there are cheating examples in the training set?

If that's the case, then we can simply clean up the training/fine tuning set and solve this mess.