I imagine there would be value in not just throwing all of GitHub commits in as training data, but also rating the quality.