Tried it like 3 years ago when the « gzip beats bert at classification » paper came out. Let’s say that it wasn’t great because your model size grows with train set lmao (Same for inference loool)