Short, automatically graded question sets for ranking small open-weight models against each other. Every question is sent separately, and code is run or rendered to grade it.