This just adds a variable to flavor tagging. Nothing else should change.
We found that our flavor tagging networks were a bit overtrained on the ttbar dataset we normally use. We're considering working around this by adding a per-jet hash and training several networks, so we can avoid ever applying the tagger to a jet we trained on.
This is adding the first part: the hash. It mixes some hit and constituent info with the event number to get something pretty random.