WestlakeNLP/FigureBench

FigureBench是第一个从长篇幅科学文本生成科学插图的大规模基准数据集。该数据集涵盖了多种文档类型，包括研究论文、调查、技术博客和教科书，旨在为自动科学插图生成研究提供一个具有挑战性和多样性的测试平台。数据集包含3,000个科学插图样本，用于训练、开发和实验目的。关键挑战包括长上下文推理、高文本密度、结构复杂性和视觉丰富性。数据集来源包括Research-14K数据集、arXiv上的AI调查、开放许可的教育平台和技术博客。注释过程经过严格筛选和评估，确保了高质量的数据。数据集结构采用对话格式，包含系统、用户和助理角色，以及插图图像的相对路径。

FigureBench is the first large-scale benchmark for generating scientific illustrations from long-form scientific texts. The dataset is curated to encompass a wide array of document types, including research papers, surveys, technical blogs, and textbooks, establishing a challenging and diverse testbed to spur research in automatic scientific illustration generation. This dataset contains the Development Set with 3,000 scientific illustration samples, designed for training, development, and experimental purposes. Key challenges include long-context reasoning, high text density, structural complexity, and visual richness. The dataset is sourced from multiple high-quality sources, including the Research-14K dataset, AI surveys on arXiv, open-licensed educational platforms, and technical blogs. The annotation process involves rigorous filtering and evaluation to ensure high-quality data. The dataset structure uses a conversational format with system, user, and assistant roles, along with relative paths to the figure images.