A reliable benchmark for evaluating training data detection in realistic, large-scale post-training settings