Efficient training of cross domain LLMs with LoRA and model merging