Debiasing Large Vision-Language Models by Ablating Protected Attribute Representations