Comment by calibas

It seems like they just grabbing every possible bit of data available, I doubt there's any mechanism to flag which edits are corrections when training.