UKPLab/AITDNA
收藏资源简介:
AITDNA(AI生成文本检测:概念无关)是一个收集自一系列用户研究的人类-AI交互数据集。该数据集包含:1. 每个文本的完整创建信息:原始用户编辑、模型建议、用户查询等。2. 根据论文中描述的不同AI生成文本概念(定义)的每个文本表示。当前支持的概念包括:- 文档级:每个文档一个标签(如果≥50%的令牌是AI生成的,则标记为AI)- 句子级:每个句子一个标签(如果≥50%的令牌是AI生成的,则标记为AI)- 令牌级:每个令牌一个标签- 边界级:通过找到最优分割索引将文本分为N部分(默认N=5)- 跨度级:相同作者身份的字符级跨度(例如,用户:GPUs are speci,AI:alized processors,...)- 基于意图的:基于预定义规则集的句子级标签,指定允许和禁止的用户查询类型。- 基于内容的:基于预定义规则集的句子级标签,指定允许和禁止的模型输出类型。- 基于成员资格的:基于参考人类语料库中N-gram出现情况的令牌级标签(默认N=2,参考人类语料库=数据集中仅人类部分)。语言:英语。许可证:gpl-2.0。
AITDNA (AI Generated Text Detection: Concept-Agnostic) is a human-AI interaction dataset collected from a series of user studies. This dataset contains: 1. Full creation information for each text: original user edits, model suggestions, user queries, etc. 2. Per-text annotations corresponding to the distinct AI-generated text concepts (definitions) described in the accompanying paper. Currently supported concepts include: - Document-level: One label assigned per document, marked as AI if ≥50% of its tokens are AI-generated - Sentence-level: One label assigned per sentence, marked as AI if ≥50% of its tokens are AI-generated - Token-level: One label assigned per token - Boundary-level: Split the text into N segments by identifying optimal segmentation indices (default N=5) - Span-level: Character-level spans with the same author identity (e.g., "User: GPUs are speci", "AI: alized processors", ...) - Intent-based: Sentence-level labels based on a predefined rule set that specifies allowed and prohibited user query types - Content-based: Sentence-level labels based on a predefined rule set that specifies allowed and prohibited model output types - Membership-based: Token-level labels based on the occurrence of N-grams in a reference human corpus (default N=2, reference human corpus = only the human-only portion of the dataset) Language: English License: GPL-2.0.




