Ai2 Dolma: 3T token open corpus for language model pretraining (2023)