Hey they said it had a lot of training data, not necessarily high-quality python code training data.