Enhancing Character-Level Understanding in LLMs through Token Internal Structure Learning