Efficient Continual Pre-training of LLMs for Low-resource Languages