Evaluating Deep Learning-based Vulnerability Detection Models on Realistic Datasets