Sequential-NIAH
收藏资源简介:
Sequential-NIAH数据集是由腾讯优图实验室创建的,旨在评估大型语言模型处理长文本上下文并从中提取顺序信息的能力。该数据集包含三种类型的针生成管道:合成针、真实针和开放域问答针,涵盖了从8K到128K令牌长度的上下文。它由14,000个样本组成,其中2,000个用于测试。数据集通过将顺序针插入到长文本中,以评估模型在理解和提取长文本中顺序信息方面的性能。
The Sequential-NIAH Dataset was developed by Tencent YouTu Lab to evaluate the capability of large language models (LLMs) to process long-text contexts and extract sequential information therefrom. This dataset includes three types of needle generation pipelines: synthetic needles, real-world needles, and open-domain question answering needles, covering context lengths ranging from 8K to 128K tokens. It consists of 14,000 total samples, among which 2,000 are reserved for testing. The dataset assesses model performance in understanding and extracting sequential information from long texts by inserting sequential needles into lengthy textual contexts.




