BTW I’ve heard from researchers from a big GAFAM (can’t name) that they actually use gpt to do this on the ground truth and spot that correction and do a second pass of human labelling on the ground truth to have better dataset.