uw-math-ai/Math2Vec-embedding-dataset
收藏资源简介:
Math2Vec嵌入数据集是一个用于数学嵌入模型训练的数据集,基于同一数学概念可以有多种表面形式(如非正式自然语言、重新表述、Lean 4类型签名和完整声明)的原则构建。该数据集旨在用于对比嵌入训练,以对齐这些不同视图。它包含三个配置:1) mathlib_views(默认配置),提供133,621个mathlib4概念的多视图表示,每个概念包含最多四个并行表示和三个LLM生成的困难负样本;2) meld,包含270个数学等价但词汇不同的语句对,用于评估嵌入模型是否捕捉数学等价性而非表面相似性;3) meld_distractors,提供541个干扰语句,覆盖18个数学领域,用于增强MELD检索评估的难度。数据集基于mathlib v4.19.0构建,所有自然语言内容均为英语,部分内容由LLM生成并经过质量验证。
The Math2Vec Embedding Dataset is a training dataset for mathematical embedding models, built around the principle that the same mathematical concept can be expressed in many surface forms—informal natural language, alternative rephrasings, Lean 4 type signatures, and Lean 4 full declarations. It is designed for contrastive embedding training that aligns these views. The dataset includes three configs: 1) mathlib_views (default), which provides multi-view representations for 133,621 mathlib4 concepts, each with up to four parallel representations and three LLM-generated hard negatives; 2) meld, which consists of 270 mathematically equivalent but lexically different statement pairs for evaluating whether embedding models capture mathematical equivalence rather than surface similarity; and 3) meld_distractors, which offers 541 distractor statements across 18 mathematical domains to make MELD retrieval evaluation harder. The dataset is based on mathlib v4.19.0, with all natural-language content in English and some content generated by LLMs with quality validation.




