RECALL

RECALL数据集由北京大学计算机科学学院多媒体信息处理国家重点实验室创建，旨在评估大型语言模型对反事实外部知识的鲁棒性。该数据集包含超过47000个样本，通过引入反事实信息到现有知识库中构建，用于测试模型在面对不准确或误导性信息时的表现。数据集涵盖两个任务：问答和文本生成，每个任务都设计了包含反事实信息的上下文，以评估模型在复杂信息环境下的准确性和可靠性。RECALL数据集的应用领域主要集中在提高语言模型在实际应用中的鲁棒性和准确性，特别是在需要处理大量外部信息和知识的场景中。

The RECALL dataset was developed by the State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University, with the goal of evaluating the robustness of large language models (LLMs) against counterfactual external knowledge. This dataset includes over 47,000 samples, which are constructed by introducing counterfactual information into existing knowledge bases, to test the model's performance when confronted with inaccurate or misleading information. The dataset covers two tasks: question answering and text generation. For each task, contexts containing counterfactual information are designed to assess the model's accuracy and reliability in complex information environments. The application scenarios of the RECALL dataset mainly focus on enhancing the robustness and accuracy of language models in real-world applications, particularly in scenarios that require processing large volumes of external information and knowledge.