Is the 62% to 13% reduction spread broadly across tasks? Or driven by a subset where the grader models almost always refuse to hack?
Is the 62% to 13% reduction spread broadly across tasks? Or driven by a subset where the grader models almost always refuse to hack?