What has it got to do with deduplication? I'm talking about crafting some kind of alternative (not necessarily duplicate) data. I agree some kind of post data collection cleaning/filtering of the data before training could potentially catch it. But maybe not!