BioCite

该数据集名为BioCite，是从BioS数据集中构建的人工合成预训练语料库。它包含了虚构人物的传记，每个传记中都列出了六个关于每个人的不同事实。该数据集旨在模拟真实的预训练数据，其中包括使用模板描述的属性，并且对文档ID进行了精心设计，以挑战模型的归因能力。该数据集的规模为10万份生成的文档，其任务是研究语言模型中的内在来源引用。

This dataset, named BioCite, is a synthetic pre-training corpus constructed from the BioS dataset. It includes biographies of fictional individuals, each containing six distinct facts about the corresponding person. This dataset is designed to simulate real pre-training data, which incorporates attributes described via templates, and its document IDs are meticulously designed to challenge the attribution capabilities of language models. Comprising 100,000 generated documents, the dataset is primarily utilized for investigating intrinsic source attribution in language models.