GPT-2's dataset used every reddit post with more than 3 net votes as training data, IIRC. Seems likely they've iterated on that approach since.