If you were tasked with improving the accuracy of a language model, what steps would you take?