Stanford human preferences dataset of 385k collective human preferences over text.