SemanticSeg
收藏资源简介:
SemanticSeg是一个语义分割数据集,源自论文《Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation》,旨在训练名为segmenter的模型以提升文本块注意力的泛化能力。该数据集包含约16个类别,每个类别至少提供2,000个实例,总数据规模在1万到10万样本之间,具体每个类别的实例数在1,980至4,821不等。数据涵盖多样化的文本来源,包括书籍章节、长指令、短段落、聊天历史、教科书章节、数学文本、学术论文、原始书籍、问答数据、教育网页、维基百科以及多种编程语言(如综合代码、Python、C、Java和Shell)的代码片段。每个数据样本包含三个字段:txt_marker(带有候选切割点的文本字符串)、chunk_id(分割块的边界标识)和chunk_plain_text(块的纯文本内容)。数据集附有每个类别的切割率(Cut rate),以帮助模型学习不同的分割模式,主要用于文本生成任务中的语义分割训练,适用于自然语言处理领域,特别是需要文本结构分析和块注意力机制的应用场景。
SemanticSeg is a semantic segmentation dataset derived from the paper Towards Generalization of Block Attention via Automatic Segmentation and Block Distillation, designed to train a model named segmenter to enhance the generalization ability of block attention. It contains approximately 16 categories, with at least 2,000 instances per category, and a total data scale ranging from 10,000 to 100,000 samples, with instance counts per category varying from 1,980 to 4,821. The dataset covers diverse text sources, including book chapters, long instructions, short paragraphs, chat histories, textbook chapters, mathematical texts, academic papers, raw books, Q&A data, educational web pages, Wikipedia, and code snippets in multiple programming languages (such as comprehensive code, Python, C, Java, and Shell). Each data sample includes three fields: txt_marker (text string with candidate cut points), chunk_id (boundary identifier for segmented blocks), and chunk_plain_text (plain text content of the block). The dataset is accompanied by a cut rate for each category to assist models in learning different segmentation patterns, and it is primarily used for semantic segmentation training in text generation tasks, applicable to the natural language processing field, especially for scenarios requiring text structure analysis and block attention mechanisms.




