IGenBench-Dataset
收藏资源简介:
IGenBench是世界上第一个文本到信息图的基准测试数据集,专注于评估生成的视觉信息图是否在事实上正确、数字上准确以及语义上忠实于输入文本和数据。数据集包含600个样本,共计5,259个问题,平均每个样本有8.77个问题,涵盖10种问题类型和30多种图表类型。每个样本以JSON格式存储,包含ID、参考图像URL、文本到图像提示、图表类型以及评估信息。
IGenBench is the world's first text-to-infographic benchmark dataset, which focuses on evaluating whether the generated visual infographics are factually correct, numerically accurate and semantically faithful to the input text and data. The dataset contains 600 samples, totaling 5,259 questions, with an average of 8.77 questions per sample, covering 10 types of questions and more than 30 chart types. Each sample is stored in JSON format, including ID, reference image URL, text-to-image prompt, chart type and evaluation information.
IGenBench 数据集概述
数据集基本信息
- 数据集名称: IGenBench Dataset
- 主要用途: 文本到信息图生成的可靠性评估
- 核心关注点: 评估生成的信息图在事实准确性、数值精确性以及语义忠实度方面的可靠性
- 任务类别: 文本到图像
- 标签: 信息图、文本到图像
- 许可协议: MIT 许可证
- 规模类别: 少于1K样本
数据集规模与构成
- 总样本数: 600
- 总问题数: 5,259
- 平均每样本问题数: 8.77
- 问题类型数量: 10
- 图表类型数量: 30+
数据格式
每个样本为一个JSON文件,结构如下: json { "id": "样本ID", "reference_image_url": "参考图像URL", "t2i_prompt": "文本到图像提示词", "chart_type": "图表类型", "evaluation": [ { "source": "问题来源(prompt 或 seed)", "ground": "真实情况", "question": "问题", "question_type": "问题类型" } ] }
使用方法
可通过以下代码加载数据集: python from datasets import load_dataset dataset = load_dataset("Brookseeworld/IGenBench-Dataset")
引用信息
如需使用本数据集,请引用: bibtex @misc{tang2026igenbenchbenchmarkingreliabilitytexttoinfographic, title={IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation}, author={Yinghao Tang and Xueding Liu and Boyuan Zhang and Tingfeng Lan and Yupeng Xie and Jiale Lao and Yiyao Wang and Haoxuan Li and Tingting Gao and Bo Pan and Luoxuan Weng and Xiuqi Huang and Minfeng Zhu and Yingchaojie Feng and Yuyu Luo and Wei Chen}, year={2026}, eprint={2601.04498}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2601.04498}, }
相关资源链接
- 论文: https://arxiv.org/abs/2601.04498
- 代码: https://github.com/MisterBrookT/IGenBench
- 项目主页: https://igen-bench.vercel.app/




