Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs