“Data labeling companies are purchasing old GitHub repositories of human-written code from before 2022 to use as high-quality training data for their models.”