Feb 8 14:11 Update README: fix L0 experiment description 13:43 Update README: link UltraData-Math-Generator to HF Space, add UltraData collection URL 13:19 Update README: add UltraData-Math-Parser reference in L0 section 13:15 Update README: rename Code Repository to Source Code 13:11 Fix: use <p> tag without blank lines to avoid HF rendering artifact 13:03 Update README: improve wording, add Quick Start section - Use HTML anchor tags in centered div for better HF rendering
- Lowercase "large language models", "embedding", "chain-of-thought"
- Replace "closed-source" with "proprietary"
- Remove "Level" from shortcoming bullet points
- Unify verb tense to present tense
- Change "scenarios" to "parsing", "hierarchical" to "tiered"
- Add transition sentence before dataset table
- Add Quick Start section with HuggingFace usage examples Feb 7 15:34 Update README: Merge experimental setup into results section 15:24 Update README: Refine L3 description with format repair and textbook-quality standards 15:13 Rename UltraData-Math-L2 to UltraData-Math-L2-preview 14:56 Update README: Add UltraData-Math-L2 subset config 10:44 Update README: Align terminology with paper (MATH500, OpenCompass, Refined Data, etc.) 08:51 Update README: Enhance Full Evaluation Results description and add R-Bench-Math 08:34 Update README: Simplify benchmark list and unify table headers 08:21 Update README: Simplify experimental setup, unify table column order, remove commonsense benchmarks chuyue +1: Cursor Co-authored-by: Cursor <cursoragent@cursor.com> 07:45 Update README: Add UltraData-Math-L2 dataset description chuyue +1: Cursor Co-authored-by: Cursor <cursoragent@cursor.com> 07:29 Update README: Add UltraData-Math-L2 statistics 07:25 Update README: Add Commonsense Reasoning benchmarks description 07:23 Update README: Remove Avg_ZH column from results table 07:06 Update README: Refine experimental results, setup details, and L0 parser comparison Feb 4 07:18 Update README: rename repository references to UltraData-Math 04:04 Update README: add dataset configs for L1 and L3 subsets viewer Jan 21 03:01 fix bold highlighting in evaluation tables Jan 20 13:07 bold best scores in experiment results Jan 19 13:46 language switch 02:46 change name Jan 18 13:51 add README 08:47 add English README Jan 16 05:42 add dataset navigation links for L1 and L3 zcy
05:23 docs: 简化术语格式，移除英文括号注释 zcy
04:02 docs: add complete project documentation and pipeline diagram zcy
03:22 Create README.md # UltraData-Math
<div align="center">
<img src="assets/ultradata-math-logo.png" width="600"/>
</div>
<div align="center">
[📜 技术报告](https://arxiv.org/abs/xxxx.xxxxx) | [📄 MiniCPM 论文](https://huggingface.co/papers/2506.07900) | [💻 代码仓库](https://github.com/openbmb/UltraData-Math) | [🌐 项目主页](https://huggingface.co/collections/openbmb/minicpm-4-6841ab29d180257e940baa9b)
</div>
## 📚 简介
高质量预训练数据对提升大语言模型的数学推理能力至关重要。然而，现有数学预训练数据构建方案存在以下不足：
- **HTML 解析层面**：通用提取器（如 trafilatura、readability）主要面向新闻/文章场景设计，对数学公式等内容缺乏专门处理，常导致公式结构破坏或丢失；同时论坛类页面的数学讨论部分，难以完整提取。
- **数据质量层面**：现有数据集普遍缺乏系统的质量分级机制，高价值数学内容与低质噪声混杂。
- **数据多样性层面**：主流数据集多源自教科书或竞赛题库，缺少真实网页中的数学讨论与应用场景；合成数据格式单一，难以覆盖多轮对话、多风格表达等多样化需求。
针对上述问题，我们提出 ***UltraData-Math***——一个面向数学推理任务的大规模高质量预训练数据集。本数据集基于 [Ultra-Data](xxx)的L0-L4 分级数据处理框架开发，包含四个递进层级：(1) L0 原始数据层，基于 *magic-html* 开发数学解析器，结合 *w3m* 布局保持渲染与多级回退策略，将 MathML、KaTeX、AsciiMath 标准化为 LaTeX 格式；(2) L1 过滤数据层，通过启发式规则清洗噪声并进行文档级去重；(3) L2 精筛数据层，使用闭源大模型标注种子数据并蒸馏至轻量 Embedding 分类器，实现全量语料的高效质量分级；(4) L3 … 
