1fanj/Chartographer
收藏资源简介:
Chartographer是一个图表推理数据集,用于评估视觉语言模型是否通过视觉推理而非捷径或对图表的先验熟悉度来回答图表问题。每个图表问题家族包含一个上游原始图表、一个重建图表以及十个具有相同Chartographer `chart_id`和`question_id`的种子控制反事实变体。数据集支持视觉问答和问答任务,适用于图表、视觉语言、图表问答和合成数据相关研究,数据规模在1K到10K之间,语言为英语,来源数据集包括ChartMuseum、ChartQA和CharXiv。
Chartographer is a chart reasoning dataset developed to evaluate whether vision-language models (VLMs) answer chart-related questions through visual reasoning rather than relying on shortcuts or prior familiarity with charts. Each chart question family consists of one upstream original chart, one reconstructed chart, and ten seed-controlled counterfactual variants that share identical Chartographer `chart_id` and `question_id` values. This dataset supports both visual question answering (VQA) and general question answering tasks, and is applicable to research in the fields of charts, vision-language models, chart question answering, and synthetic data. With a scale ranging from 1K to 10K, all texts in this dataset are in English, and its source datasets include ChartMuseum, ChartQA, and CharXiv.




