I'd bet even that would be only with a some or all of (1) a LOT of training data, (2) a LOT of preprocessing, and (3) using less famous architectures (perhaps r

I'd bet even that would be only with a some or all of (1) a LOT of training data, (2) a LOT of preprocessing, and (3) using less famous architectures (perhaps recursive neural nets). I say so because (among other reasons) in general, current popular ML architectures (like transformers) count processing heirarchical data as one of their weaknesses. For example, there's a theoretical paper proving that self attention blocks (which are central to transformers) cannot solve arbitrarily nested negation(Ie, resolve not(not(not(true))) to true). Practically as well, we see most times that language models have trouble dealing with naturally occurring double negation in language, etc. But CSS/HTML is, I think, very heirarchical.