Low-Confidence Gold: Refining Low-Confidence Samples for Efficient Instruction Tuning