Beware of Calibration Data for Pruning Large Language Models