Is a bad LM implementation still useful for research? Shmaybe.