Lessons from the trenches on reproducible evaluation of language models