Adversarial GLUE (AdvGLUE)
收藏资源简介:
AdvGLUE是一个针对语言模型鲁棒性评估的多任务基准,由伊利诺伊大学厄巴纳-香槟分校、浙江大学和微软公司合作创建。该数据集包含4978条数据,通过系统地应用14种文本对抗攻击方法于GLUE任务,旨在全面探索和评估现代大规模语言模型在各种对抗攻击下的脆弱性。AdvGLUE数据集的构建过程包括对GLUE任务的系统攻击、人工验证以确保高质量的基准,以及对现有语言模型和鲁棒训练方法的彻底评估。该数据集的应用领域主要集中在提高语言模型对复杂对抗攻击的鲁棒性,解决模型在面对精心设计的文本对抗样本时的脆弱性问题。
AdvGLUE is a multi-task benchmark for evaluating the robustness of language models, jointly created by the University of Illinois Urbana-Champaign, Zhejiang University, and Microsoft Corporation. This dataset contains 4,978 instances, which are generated by systematically applying 14 text adversarial attack methods to GLUE tasks, aiming to comprehensively explore and assess the vulnerability of modern large-scale language models under various adversarial attacks. The construction process of the AdvGLUE dataset includes systematic adversarial attacks on GLUE tasks, manual verification to ensure the high quality of the benchmark, and thorough evaluation of existing language models and robust training methods. The main application fields of this dataset focus on improving the robustness of language models against complex adversarial attacks, and addressing the vulnerability of models when facing carefully designed text adversarial examples.

- 1Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models伊利诺伊大学厄巴纳-香槟分校, 浙江大学, 微软公司 · 2022年



