DARA

DARA数据集是由中国科学院软件研究所创建，用于研究文本检索中的对抗性排序攻击防御。该数据集基于MS MARCO段落数据集，通过三种新颖的攻击方法（PRADA, PAT, IDEM）生成了约144,000个对抗性示例，分为训练、验证和测试集。数据集的创建旨在通过提供标准化的评估流程，帮助研究者开发和评估对抗性排序检测方法，以增强神经排序模型在实际应用中的鲁棒性和可信度。

The DARA Dataset was created by the Institute of Software, Chinese Academy of Sciences for research on adversarial ranking attack defense in text retrieval. Based on the MS MARCO Passage Dataset, it generated approximately 144,000 adversarial examples via three novel attack methods (PRADA, PAT, IDEM), and is split into training, validation, and test sets. The dataset aims to provide a standardized evaluation pipeline to help researchers develop and evaluate adversarial ranking detection methods, thereby enhancing the robustness and credibility of neural ranking models in real-world applications.