sentence-transformers/msmarco-msmarco-distilbert-base-tas-b

MS MARCO with hard negatives from msmarco-distilbert-base-tas-b 数据集是一个基于真实用户搜索查询的大规模信息检索语料库，这些查询来自 Bing 搜索引擎。该数据集包含多种三元组（查询、正面、负面）子集，用于训练 Sentence Transformer 模型。这些子集在每对查询-正面样本的负面样本数量以及数据存储为字符串或ID方面有所不同。该数据集旨在通过提供具有挑战性的负面样本来提高信息检索系统的性能。

MS MARCO with hard negatives from msmarco-distilbert-base-tas-b is a large-scale information retrieval corpus based on real user search queries from the Bing search engine. This dataset contains multiple triplet (query, positive, negative) subsets for training Sentence Transformer models. These subsets differ in the number of negative samples per query-positive sample pair and the storage format of the data, i.e., whether it is saved as strings or IDs. The dataset aims to improve the performance of information retrieval systems by providing challenging negative samples.

原始信息汇总

数据集概述

基本信息

• 语言: 英语
• 多语言性: 单语种
• 数据集大小: 10M < n < 100M
• 任务类别: 特征提取、句子相似度
• 标签: sentence-transformers
• 数据集名称: MS MARCO with hard negatives from msmarco-distilbert-base-tas-b

数据集配置

triplet 配置

• 特征:
  • query: 字符串
  • positive: 字符串
  • negative: 字符串
• 分割:
  • train: 365876392 字节, 502939 样本
• 下载大小: 240229198 字节
• 数据集大小: 365876392 字节

triplet-50 配置

• 特征:
  • query: 字符串
  • positive: 字符串
  • negative_1 至 negative_50: 字符串
• 分割:
  • train: 5257441890 字节, 288020 样本
• 下载大小: 3444062582 字节
• 数据集大小: 5257441890 字节

triplet-50-ids 配置

• 特征:
  • query: 整数
  • positive: 整数
  • negative_1 至 negative_50: 整数
• 分割:
  • train: 119816320 字节, 288020 样本
• 下载大小: 102036560 字节
• 数据集大小: 119816320 字节

triplet-all 配置

• 特征:
  • query: 字符串
  • positive: 字符串
  • negative: 字符串
• 分割:
  • train: 19792167373 字节, 26391481 样本
• 下载大小: 4702308008 字节
• 数据集大小: 19792167373 字节

triplet-all-ids 配置

• 特征:
  • query: 整数
  • positive: 整数
  • negative: 整数
• 分割:
  • train: 633395544 字节, 26391481 样本
• 下载大小: 192368748 字节
• 数据集大小: 633395544 字节

triplet-hard 配置

• 特征:
  • query: 字符串
  • positive: 字符串
  • negative: 字符串
• 分割:
  • train: 11788992227 字节, 15981882 样本
• 下载大小: 3119903058 字节
• 数据集大小: 11788992227 字节

triplet-hard-ids 配置

• 特征:
  • query: 整数
  • positive: 整数
  • negative: 整数
• 分割:
  • train: 383565168 字节, 15981882 样本
• 下载大小: 121870475 字节
• 数据集大小: 383565168 字节

triplet-ids 配置

• 特征:
  • query: 整数
  • positive: 整数
  • negative: 整数
• 分割:
  • train: 12070536 字节, 502939 样本
• 下载大小: 10133540 字节
• 数据集大小: 12070536 字节

数据集子集

triplet 子集

• 列: "query", "positive", "negative"
• 列类型: str, str, str
• 示例: python { "query": "what are the liberal arts?", "positive": liberal arts. 1. the academic course of instruction at a college intended to provide general knowledge and comprising the arts, humanities, natural sciences, and social sciences, as opposed to professional or technical subjects., "negative": The New York State Education Department requires 60 Liberal Arts credits in a Bachelor of Science program and 90 Liberal Arts credits in a Bachelor of Arts program. In the list of course descriptions, courses which are liberal arts for all students are identified by (Liberal Arts) after the course number. }

triplet-ids 子集

• 列: "query", "positive", "negative"
• 列类型: int, int, int
• 示例: python { "query": 571018, "positive": 7349777, "negative": 6948601 }

triplet-all 子集

• 列: "query", "positive", "negative"
• 列类型: str, str, str
• 示例: python { "query": "what are the liberal arts?", "positive": liberal arts. 1. the academic course of instruction at a college intended to provide general knowledge and comprising the arts, humanities, natural sciences, and social sciences, as opposed to professional or technical subjects., "negative": The New York State Education Department requires 60 Liberal Arts credits in a Bachelor of Science program and 90 Liberal Arts credits in a Bachelor of Arts program. In the list of course descriptions, courses which are liberal arts for all students are identified by (Liberal Arts) after the course number. }

triplet-all-ids 子集

• 列: "query", "positive", "negative"
• 列类型: int, int, int
• 示例: python { "query": 571018, "positive": 7349777, "negative": 6948601 }

triplet-hard 子集

• 列: "query", "positive", "negative"
• 列类型: str, str, str
• 示例: python { "query": "what are the liberal arts?", "positive": liberal arts. 1. the academic course of instruction at a college intended to provide general knowledge and comprising the arts, humanities, natural sciences, and social sciences, as opposed to professional or technical subjects., "negative": The New York State Education Department requires 60 Liberal Arts credits in a Bachelor of Science program and 90 Liberal Arts credits in a Bachelor of Arts program. In the list of course descriptions, courses which are liberal arts for all students are identified by (Liberal Arts) after the course number. }

triplet-hard-ids 子集

• 列: "query", "positive", "negative"
• 列类型: int, int, int
• 示例: python { "query": 571018, "positive": 7349777, "negative": 6948601 }

triplet-50 子集

• 列: "query", "positive", negative_1, negative_2, negative_3, negative_4, negative_5, negative_6, negative_7, negative_8, negative_9, negative_10, negative_11, negative_12, negative_13, negative_14, negative_15, negative_16, negative_17, negative_18, negative_19, negative_20, negative_21, negative_22, negative_23, negative_24, negative_25, negative_26, negative_27, negative_28, negative_29, negative_30, negative_31, negative_32, negative_33, negative_34, negative_35, negative_36, negative_37, negative_38, negative_39, negative_40, negative_41, negative_42, negative_43, negative_44, negative_45, negative_46, negative_47, negative_48, negative_49, negative_50
• 列类型: str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str, str
• 示例: python { "query": "what are the liberal arts?", "positive": "liberal arts. 1. the academic course of instruction at a college intended to provide general knowledge and comprising the arts, humanities, natural sciences, and social sciences, as opposed to professional or technical subjects.", "negative_1": "The New York State Education Department requires 60 Liberal Arts credits in a Bachelor of Science program and 90 Liberal Arts credits in a Bachelor of Arts program. In the list of course descriptions, courses which are liberal arts for all students are identified by (Liberal Arts) after the course number.", "negative_2": "What Does it Mean to Study Liberal Arts? A liberal arts major offers a broad overview of the arts, sciences, and humanities. Within the context of a liberal arts degree, you can study modern languages, music, English, anthropology, history, womens studies, psychology, math, political science or many other disciplines.", "negative_3": "What Is Liberal Studies? Liberal studies, also known as liberal arts, comprises a broad exploration of social sciences, natural sciences, humanities, and the arts. If you are interested in a wide-ranging education in humanities, communication, and thinking, read on to find out about the educational and career possibilities in liberal studies.", "negative_4": "You can choose from an array of liberal arts majors. Most of these are offered in the liberal arts departments of colleges that belong to universities and at smaller colleges that are designated as liberal arts institutions.", "negative_5": "Majors. You can choose from an array of liberal arts majors. Most of these are offered in the liberal arts departments of colleges that belong to universities and at smaller colleges that are designated as liberal arts institutions.", "negative_6": "liberal arts. plural noun. Definition of liberal arts for English Language Learners. : areas of study (such as history, language, and literature) that are intended to give you general knowledge rather than to develop specific skills needed for a profession. Nglish: Translation of liberal arts for Spanish speakers Britannica.com: Encyclopedia article about liberal arts.", "negative_7": "Because they award less than 50% of their degrees in engineering, and the rest in liberal arts (sciences). Baccalaureate colleges are a type of Liberal Arts colleges, But offering lesser number of degrees compared to LAC. Its the other way round. A liberal arts college focuses on liberal arts, e.g. sciences, literature, history, sociology, etc. They might offer a few professional degrees (most frequently engineering) as well, but typically the professional majors are well integrated into the liberal arts framework as well.", "negative_8": "A liberal arts college is a four-year institution that focuses on the study of liberal arts. Liberal arts colleges are geared more toward the acquisition of knowledge and less toward specific professions. [MORE: The Path to Higher Education] Graduate school.", "negative_9": "1 BA = Bachelor of Arts degree BS = Bachelor of Science degree. 2 I think the question requires more of an explanation than what the terms BA and BS translate to. 3 B.A. (Bachelor of Arts) A bachelor of arts (B.A.) degree is what is generally called a liberal arts degree. I think the question requires more of an explanation than what the terms BA and BS translate to. 2 B.A. (Bachelor of Arts) A bachelor of arts (B.A.) degree is what is generally called a liberal arts degree.", "negative_10": "West Hills College LemooreAssociate of Arts (A.A.), Liberal Arts and Sciences/Liberal StudiesAssociate of Arts (A.A.), Liberal Arts and Sciences/Liberal Studies. -Student Government President for two years. -Valedictorian. -Alpha Gamma Sigma (Alpha Chi chapter) President/College Relations Liaison.", "negative_11": "You can pursue associate degree in academic area such as business administration, law, arts, engineering, paralegal studies, liberal arts, computer science, and more. Q: What are online associate programs?", "negative_12": "liberal arts definition The areas of learning that cultivate general intellectual ability rather than technical or professional skills. Liberal arts is often used as a synonym for humanities, because literature, languages, history, an" }

搜集汇总

数据集介绍

构建方式

该数据集基于微软构建的大规模信息检索语料库MS MARCO，该语料库源自Bing搜索引擎的真实用户查询。针对每个查询及其对应的正例段落，研究团队利用msmarco-distilbert-base-tas-b模型挖掘出与查询最相似的50个段落作为负例，从而生成三元组结构的数据。数据集提供了多种配置：包含单一负例的基础三元组（triplet）、包含全部50个负例的三元组（triplet-all），以及通过交叉编码器ms-marco-MiniLM-L-6-v2筛选、确保负例与查询相似度低于正例一定阈值的困难三元组（triplet-hard）。此外，还提供了以整数标识符替代文本的ID版本，便于与配套语料库联用。

特点

该数据集的核心特点在于其丰富的负例构建策略与多层次的配置设计。通过使用单一模型挖掘出的50个最相似段落作为负例，数据能够有效模拟信息检索中难以区分的负样本场景。困难三元组配置进一步利用交叉编码器进行精炼，过滤掉与查询相似度过高的负例，从而提升训练样本的判别性难度。数据集提供字符串与ID两种格式，并包含从单负例到多负例的多种子集，灵活适应不同训练需求。所有样本均未去重，保留了原始检索结果的多样性，为句子嵌入模型的对比学习提供了坚实的数据基础。

使用方法

该数据集专为训练Sentence Transformer模型设计，使用时可通过HuggingFace Datasets库加载指定配置，如`load_dataset("sentence-transformers/msmarco-msmarco-distilbert-base-tas-b", "triplet")`。对于文本格式的数据，可直接将查询、正例和负例字段输入三元组损失函数进行训练。ID版本需配合`sentence-transformers/msmarco-corpus`语料库，通过映射整数标识符获取原始文本。用户可根据任务需求选择不同子集：基础三元组适用于快速实验，全部三元组提供更多负例，困难三元组则针对高难度区分场景。数据仅包含训练集，可直接用于模型微调与评估。

背景与挑战

背景概述

在信息检索与自然语言处理领域，如何精准地衡量查询与文档之间的语义相似度始终是核心研究问题之一。MS MARCO数据集由微软研究院基于Bing搜索引擎的真实用户查询构建而成，自2016年发布以来，已成为大规模文本检索与问答系统的标杆性基准。该数据集囊括了百万级真实查询及对应的相关性标注，为深度语义匹配模型的训练与评估提供了坚实的数据基础。在此基础上，sentence-transformers团队于2021年前后推出了msmarco-distilbert-base-tas-b版本的数据集，其核心贡献在于利用13种不同的嵌入模型为每个查询-正例对挖掘出最相似的50个段落作为难负样本，从而显著提升了句向量模型在细粒度语义区分上的训练效率与泛化能力。这一数据集的诞生不仅推动了稠密检索技术的发展，也为后续基于对比学习的句子表征研究奠定了重要基石。

当前挑战

该数据集所面临的挑战主要体现在两个层面。在领域问题层面，信息检索系统需要从海量文本中精准定位与用户意图高度匹配的段落，而真实查询往往存在歧义性、口语化以及上下文缺失等复杂情况，使得语义匹配任务极具难度。传统的基于词频的检索方法难以捕捉深层语义关联，而依赖人工标注的监督学习又受限于标注成本与规模。在数据集构建过程中，挖掘高质量难负样本本身便构成一大技术瓶颈——如何从数十亿候选段落中高效筛选出与查询语义接近但非相关的负例，以避免模型学到过于简单的判别边界，同时保证样本的多样性与代表性，是构建过程中的关键挑战。此外，数据集的规模庞大（逾2600万三元组），存储与处理开销巨大，且不同模型生成的难负样本质量参差不齐，进一步增加了数据清洗与质量控制的复杂性。

常用场景

经典使用场景

在信息检索与文本语义匹配的研究领域中，sentence-transformers/msmarco-msmarco-distilbert-base-tas-b数据集扮演着举足轻重的角色。该数据集源于微软构建的大规模真实用户搜索查询语料库MS MARCO，通过将Bing搜索引擎中的用户查询与相关段落进行配对，并利用预训练模型挖掘出与查询语义高度相似的困难负样本，形成了经典的三元组结构。研究者常借助这一数据集来训练和评估句嵌入模型，尤其是在对比学习框架下，通过拉近查询与正例段落之间的距离、推远与困难负例之间的表征，从而提升模型对文本语义细微差别的捕捉能力。这一场景不仅验证了模型在段落排序与相关性判别上的性能，更成为衡量句向量质量的重要基准。

实际应用

在实际应用层面，该数据集所支撑的句嵌入模型已广泛部署于现代搜索引擎、智能问答系统与推荐引擎中。例如，在搜索引擎的场景下，模型能够将用户输入的自然语言查询实时编码为向量，并在海量文档库中通过近似最近邻搜索快速定位最相关的段落，从而显著提升检索效率与用户体验。在客服机器人或知识库问答系统中，该数据集训练出的模型可以准确匹配用户问题与预定义的答案模板，实现精准的意图识别与响应。此外，在内容推荐领域，通过计算用户历史行为与候选物品的语义相似度，系统能够生成更具个性化的推荐列表，进而提高用户黏性与平台转化率。

衍生相关工作

该数据集的发布催生了一系列具有深远影响的经典工作。其中，基于msmarco-distilbert-base-tas-b模型本身便成为句嵌入研究的里程碑，其采用的双塔架构与对比学习策略被后续诸多工作所继承和优化。此外，利用该数据集生成的困难负样本训练范式，启发了诸如Margin-MSE、MNRL等新型损失函数的设计，这些方法进一步提升了模型在零样本与少样本场景下的鲁棒性。在模型蒸馏方面，该数据集也被用于训练轻量级学生模型，如DistilBERT系列，使得高性能句嵌入模型得以在资源受限的环境中部署。这些衍生工作共同构筑了现代语义检索的技术基石，推动了自然语言处理从学术研究向工业落地的跨越。