Amazing work, I've been following this for a while. Have you considered putting this into an arxiv whitepaper describing the model + tricks? I've wanted to cite this a couple times, but have had to resort to citing the github repo.