RAISE: Reinforenced Adaptive Instruction Selection For Large Language Models