I mostly directed it into making training corpuses for classical classification models instead. "oh I need 10k ground truth images to train on" *cracks knuckles* "better get tagging then!"
there are widely-used open training datasets with fewer entries than the ones I have produced by myself by hand. I find it moderately alarming.
(the knuckle cracking isn't for dramatic effect, I just have fucked up joints and can't type if I don't crack them relatively frequently)