Nothing Reported It

A test harness for small language models produced misleading results due to hidden bugs in metrics, data loading, and error handling. After fixing five key issu