DuReader (robust)

该数据集是一个大规模的中文机器阅读理解基准数据集，旨在评估MRC模型在处理过度敏感、过度稳定和泛化问题方面的能力。测试集分为两个版本（Test1和Test2），并且不公开答案标签，以避免性能被高估。数据集的规模如下：训练集包含15,000个样本；开发集包含1,400个样本；测试集的各个子集分别包含1,300、1,300、800和1,600个样本。该数据集的任务是机器阅读理解。

This dataset is a large-scale Chinese machine reading comprehension (MRC) benchmark dataset designed to evaluate the performance of MRC models in handling over-sensitivity, over-stability, and generalization challenges. The test set is split into two variants (Test1 and Test2), and the answer labels are not publicly released to avoid overestimating model performance. The dataset has the following scale: the training set contains 15,000 samples; the development set contains 1,400 samples; each subset of the test set includes 1,300, 1,300, 800, and 1,600 samples respectively. The core task of this dataset is machine reading comprehension.

搜集汇总

背景与挑战

背景概述

DuReader (robust) 是一个高质量的中文机器阅读理解数据集，专注于真实应用场景，旨在评估和提升模型在鲁棒性方面的性能，包括过敏感性、过稳定性和泛化能力。该数据集源于2020年语言与智能技术竞赛，由百度、中国中文信息学会和中国计算机学会联合推出，用于促进学术界和工业界在机器阅读理解领域的技术交流和发展。