I think his point (unless i've misunderstood) is moreso if a codebase is really the best held out dataset for such a evaluation meant to be wide-sweeping. Does this plot look any or much different if the dataset was more encompassing to include other sorts of text ?
reply