Some critical issues with the SWE-bench dataset