I was wondering if there's a proper way of detokenizing the output tokens, i.e., constructing the sentence back from the tokens? Considering the fact that the word-piece tokenisation introduces lots of #s.