We tried to evaluate on data that is as much as possible independent from the training data. So we evaluate on projects outside of github.