DLRewrites/unified-reasoning-augmented

这是一个用于研究问答大型语言模型（LLMs）鲁棒性的数据集，基于论文《Beyond Token Noise: Discourse-Level Rewrites Expose Brittleness in QA LLMs》构建。数据集包含来自5个问答基准（ARC-Challenge、BIG-Math、DROP、GPQA Diamond、MMLU-Pro）的原始问题及其增强版本，增强方法包括话语级重写（DLR，如问题重排序、子句反转、简化语言、复杂语言和无关信息添加）、局部词汇扰动（EDA，如随机删除、插入、交换和同义词替换）以及回译（英语→德语→英语）。数据集特征包括原始数据集ID、来源、原始问题、增强问题、答案、选择项列表、增强标签、温度参数、唯一标识符和原始分割信息，旨在评估LLMs在不同类型问题改写下的性能表现和错误模式。

This is a dataset for studying the robustness of question-answering large language models (LLMs), based on the paper Beyond Token Noise: Discourse-Level Rewrites Expose Brittleness in QA LLMs. It contains original and augmented questions from five QA benchmarks (ARC-Challenge, BIG-Math, DROP, GPQA Diamond, MMLU-Pro). Augmentation methods include discourse-level rewrites (DLR, e.g., question reordering, clause reversal, simpler language, harder language, irrelevant info addition), local lexical perturbations (EDA, e.g., random deletion, insertion, swap, synonym replacement), and back-translation (English→German→English). Dataset features include original dataset ID, source, original question, augmented question, answer, choices list, augmentation tag, temperature parameter, unique identifier, and origin split, designed to evaluate LLM performance and error patterns under various types of question rewrites.