If given a limited dataset, how would you approach training a language model?