RoParQ
收藏资源简介:
RoParQ是由首尔国立大学构建的专用于评估大语言模型对转述问题鲁棒性的基准数据集。该数据集包含10,273条高质量多选问答样本,源自MMLU、ARC、CommonsenseQA和MathQA四大权威数据集,通过Gemini和Claude等专有模型生成语义等价的转述变体。数据集构建过程采用严格的预处理流程,包括封闭式问答筛选、问题长度控制及基于法官模型的不一致置信度过滤机制。该数据集主要应用于增强大语言模型的语义不变性理解,旨在解决模型对表面语言模式过度依赖而导致的转述敏感性问题,推动可信人工智能的发展。
RoParQ is a benchmark dataset constructed by Seoul National University specifically for evaluating the robustness of Large Language Models (LLMs) against paraphrased questions. This dataset contains 10,273 high-quality multiple-choice question-answering samples derived from four authoritative datasets: MMLU, ARC, CommonsenseQA, and MathQA. Semantically equivalent paraphrased variants were generated using proprietary models such as Gemini and Claude. The dataset construction process adopts a strict preprocessing pipeline, including closed-ended QA screening, question length control, and an inconsistency confidence filtering mechanism based on judge models. This dataset is primarily applied to enhance the understanding of semantic invariance in LLMs, aiming to resolve the paraphrasing sensitivity issue caused by models' over-reliance on superficial linguistic patterns, and to promote the development of trustworthy artificial intelligence.
RoParQ数据集概述
基本信息
- 许可证: Apache 2.0
- 任务类别: 问答
- 语言: 英语
数据集配置
1. 通用知识配置 (general-knowledge)
数据文件结构
- 训练集: general-knowledge/train-*
- 验证集: general-knowledge/validation-*
- 测试集: general-knowledge/test-*
特征字段
- id: 字符串类型
- source_dataset: 字符串类型
- questions: 字符串列表
- options: 字符串列表
- answer_idx: 整型
- sampled_idxs_list: 整型列表的列表
数据统计
- 训练集: 2,194个样本,1,751,721字节
- 验证集: 470个样本,379,435字节
- 测试集: 470个样本,374,751字节
- 下载大小: 866,662字节
- 数据集总大小: 2,505,907字节
2. 数学推理配置 (math-reasoning)
数据文件结构
- 训练集: math-reasoning/train-*
- 验证集: math-reasoning/validation-*
- 测试集: math-reasoning/test-*
特征字段
- id: 字符串类型
- source_dataset: 字符串类型
- questions: 字符串列表
- options: 字符串列表
- answer_idx: 整型
- sampled_idxs_list: 整型列表的列表
数据统计
- 训练集: 4,998个样本,4,948,523字节
- 验证集: 1,071个样本,1,066,962字节
- 测试集: 1,071个样本,1,058,168字节
- 下载大小: 2,340,396字节
- 数据集总大小: 7,073,653字节

- 1RoParQ: Paraphrase-Aware Alignment of Large Language Models Towards Robustness to Paraphrased Questions首尔国立大学 · 2025年



