So much of the NLP documentation uses these lovely, pristine datasets that are packaged into libraries and can be loaded in one line. And then I have to spend ages figuring out how to bolt on my mangled, holey data that fell off the back of a truck