ArxivMIA

该数据集名为ArxivMIA，是一个全新的基准数据集，包含了来自Arxiv网站的计算机科学和数学领域的摘要。它旨在评估在具有挑战性的情境下，预训练数据检测方法的效果。ArxivMIA数据集的特点是平均句子长度较长，每样本平均包含143.1个标记。该数据集特别构建了包含成员数据和非成员数据，非常适合用来测试在RedPajama数据集上预训练的大型语言模型。该数据集规模为2000个样本，任务是对预训练数据检测。

This dataset, named ArxivMIA, is a novel benchmark dataset containing abstracts from the fields of computer science and mathematics on the arXiv website. It is designed to evaluate the performance of pre-training data detection methods in challenging scenarios. ArxivMIA features a relatively long average sentence length, with each sample containing an average of 143.1 tokens. The dataset is specifically constructed with both member and non-member data, making it an ideal testbed for large language models (LLMs) pre-trained on the RedPajama dataset. The dataset comprises 2000 samples in total, with its core task being pre-training data detection.