CHAOS
收藏资源简介:
CHAOS(图表分析与异常样本)是一个专门设计用于在现实噪声条件下对多模态语言模型(MLLMs)进行压力测试的第一个基准数据集。它包含了10种视觉扰动和5种文本扰动,每种扰动分为三个严重级别(简单、中等、困难)。数据集总共包含了112,500个经过扰动的图表(每种扰动×3级别×15类型)。CHAOS旨在衡量和了解MLLMs在现实环境变得嘈杂时如何以及为何失败——理想情况下,它们仍能成功。
CHAOS (Chart Analysis and Outlier Samples) is the first benchmark dataset specifically designed for stress-testing multimodal large language models (MLLMs) under realistic noisy conditions. It contains 10 visual perturbations and 5 text perturbations, each categorized into three severity levels: easy, medium, and hard. The dataset totals 112,500 perturbed charts (each perturbation × 3 severity levels × 15 chart types). CHAOS aims to measure and understand how and why MLLMs fail when real-world environments become noisy, ideally enabling such models to still achieve successful performance.
CHAOS Benchmark 数据集概述
基本信息
- 许可证: AGPL-3.0
- 语言: 英语 (en)
- 数据集大小: 6,103,383 字节
- 下载大小: 5,293,393 字节
- 类别大小: 10K < n < 100K
- 标签: chart-analysis, chart-understanding, MLLM, robustness
- 任务类别: visual-question-answering, image-to-text
数据集结构
特征
index: int64image: stringquestion: stringanswer: stringsplit: stringimgname: string
数据拆分
| 拆分名称 | 字节数 | 样本数 |
|---|---|---|
| chaos_vision | 4,559,793 | 60 |
| chaos_text | 1,543,590 | 30 |
数据集内容
- CHAOS-Vision: 75,000 样本
- CHAOS-Text: 37,500 样本
- CHAOS-All: 112,500 样本
下载链接
评估指标
- Relaxed Accuracy (±5% tolerance)
- CHAOS Robustness Score (ℛ)
数据集格式
- 格式: TSV (Tab-Separated Values)
- 列: index, image (Base64编码), question, answer, split, imgname
引用
bibtex @inproceedings{chaos2025, ... } @inproceedings{masry-etal-2022-chartqa, title = "{C}hart{QA}: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning", author = "Masry, Ahmed and Long, Do and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul", booktitle = "Findings of the Association for Computational Linguistics: ACL 2022", month = may, year = "2022", address = "Dublin, Ireland", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2022.findings-acl.177", doi = "10.18653/v1/2022.findings-acl.177", pages = "2263--2279", }
更新记录
| 日期 (YYYY-MM-DD) | 更新内容 |
|---|---|
| 2025-04-29 | 🚀 初始数据集发布 (所有扰动和TSV元数据上传) |




