thorough work, good stuff.. it even runs a selection-bias analysis against their own benchmark and reports that some tasks that were disproportionately hard for a model. Rare to see a benchmark paper attack itself like that.