Misleading ChartQA Benchmark
收藏资源简介:
Misleading ChartQA Benchmark是一个大规模的多模态数据集,由香港科技大学的研究团队构建,旨在评估大型多模态语言模型在识别和推理误导性图表方面的能力。该数据集包含超过3000个经过精心挑选的示例,涵盖21种误导类型和10种图表类型。每个示例包括标准化的图表代码、CSV数据和带有标签解释的多项选择题。该数据集通过多轮机器学习模型检查和专家人工审核来验证,为研究误导性图表理解提供了一个基础。
Misleading ChartQA Benchmark is a large-scale multimodal dataset constructed by the research team from the Hong Kong University of Science and Technology, designed to evaluate the ability of large multimodal language models to identify and reason about misleading charts. This dataset contains over 3,000 carefully curated examples, covering 21 types of misleading scenarios and 10 categories of charts. Each example includes standardized chart code, CSV data, and multiple-choice questions with annotated explanations. The dataset is validated through multi-round machine learning model checks and expert manual reviews, providing a foundational resource for research on misleading chart comprehension.

- 1Unmasking Deceptive Visuals: Benchmarking Multimodal Large Language Models on Misleading Chart Question Answering香港科技大学 · 2025年



