Problem with that is that taking away even specific parts of the dataset can have a large impact of performance as a whole… Like when they removed NSFW from an image generator dataset and suddenly it sucked at drawing bodies in general

I don’t get it

“hold my beer” regardless of if I’m actually holding a beer

Ah yes the good old webp format which checks notes just got assigned a CVE 10(out of 10) score. (The implementation that is used in nearly everything anyways)