hotchpotch/natural-questions-context-relevance

该数据集包含了提取的硬负样本，以及问题和文本跨度之间的相关性分数和重排分数。它基于tomaarsen的自然问题硬负样本数据集，并遵循原始自然问题数据集的许可。数据集适用于训练如OpenProvence等模型。提供了四个子集：nodup_freq2、default、freq1和posratio-0p5，分别适用于不同的训练场景。

This dataset includes extracted hard negatives, along with relevance scores between questions and text spans, and reranker scores. It is based on the tomaarsen/natural-questions-hard-negatives dataset and follows the license of the original Natural Questions dataset. The dataset is suitable for training models such as OpenProvence. It provides four subsets: nodup_freq2, default, freq1, and posratio-0p5, each designed for different training scenarios.