Wanglanhuajiaofen/MSMARCO-annotation

该数据集用于论文Training Dense Retrievers with Multiple Positive Passages的研究，主要基于MS MARCO数据集，使用Qwen3-32B模型进行效用标注。它支持密集检索器的多正例优化目标训练，包括列表式和成对损失在对比学习框架下的统一研究。数据集还涉及其他测试集如NQ和hybrid，可用于评估检索性能的鲁棒性和效果。通过理论分析和实验，该数据集旨在探索如何利用多个正例段落提升检索性能，适用于信息检索和自然语言处理任务。

This repository contains the dataset used in our paper: Training Dense Retrievers with Multiple Positive Passages. Using the Qwen3-32B, we annotate utility on the MS MARCO dataset. For NQ and hybrid test set, see Utility_focused_annotation. Our paper presents a systematic study of multi-positive optimization objectives for dense retrieval, unifying representative listwise and pairwise losses under a contrastive learning framework. Through theoretical analysis and extensive experiments on NQ, MS MARCO, and BEIR, we investigate how different objectives leverage multiple positive passages and demonstrate that LSEPair achieves strong robustness and retrieval performance across diverse supervision settings.