Needle in the Haystack (NITH)

该数据集是一项评估大型语言模型在较大语境（所谓的“干草堆”）中寻找特定“针”句子的回忆任务。该任务测试了大型语言模型在不同语境长度下，完成对位于其中的“针”句子部分表示的能力。该数据集规模超过1100万个标记。任务的类型为回忆任务。

This dataset is a recall task aimed at evaluating the ability of large language models to locate specific "needle" sentences within large contextual corpora, commonly referred to as "haystacks". This task assesses how effectively large language models can perform partial representation of the embedded "needle" sentences across varying context lengths. The dataset consists of over 11 million tokens, and this is a recall-style task.

搜集汇总

背景与挑战

背景概述

该数据集是一个包含100本Project Gutenberg书籍信息的小型文本数据集，每本书包含标题、出版日期、URL和完整文本内容，数据格式为parquet，总大小为24.9 MB。