“The current monolithic pre-training process for large models forces trade-offs, such as including more multilingual data at the expense of coding data.”