The UltraData dataset series: pre-training, SFT and preference data across Chinese, English, multilingual, math and code.