P.S. In case this wasn’t clear: our Super variants are trained on all available data, including our own validation/eval splits, to maximize final performance and, hopefully, generalization.
We ran out of time to properly evaluate the normal variants, but we’re providing them as well for transparency and reproducibility.