What's extremely bad for research is that they didn't even release their train and test sets. It's not hard to grab 40GB of text from the web that comes from l

What's extremely bad for research is that they didn't even release their train and test sets. It's not hard to grab 40GB of text from the web that comes from links with at least +3 votes on reddit. But even if you do that, you won't get the same train/test set, and the same split. So impossible to know if a model is outperforming their model.