I'm not sure the benchmark is high enough quality that >80% of problems are well-specified & have correct labels tbh. (But I guess this question has been studied for these benchmarks)