I guess it is conceivable that you could take 1000 dirty input files and their cleaned output and use this as a training set for a ML-power data cleaning system. But:
Sure. You could and should do that if you can. But, as you say, it'll still only work for one narrow definition of "dirty" and another, equally narrow, definition of "clean".
-It would only work in a narrow domain.
-How do you know if you can trust the results?