Hugging Face's logo

Datasets:
SoMarkAI
/
DocParsingBench

Modalities:
Image
Document
Formats:
imagefolder
Size:
1K - 10K
Tags:
document-parsing
document-intelligence
OCR Bench
Doc Parsing
OCR
pdf
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
3

docs: update latest news with evaluation toolkit release

#2
by nkii22 - opened
base: refs/heads/main
←
from: refs/pr/2
Discussion Files changed
+45
-47
Files changed (2) hide show
  1. .gitignore +0 -1
  2. README.md +45 -46
.gitignore CHANGED
@@ -1,3 +1,2 @@
1
  .DS_Store
2
  **/.DS_Store
3
- README.zh.md
 
1
  .DS_Store
2
  **/.DS_Store
 
README.md CHANGED
@@ -1,8 +1,8 @@
1
  ---
2
  license: odc-by
3
  tags:
4
- - document-parsing
5
- - document-intelligence
6
  - OCR Bench
7
  - Doc Parsing
8
  - OCR
@@ -10,7 +10,7 @@ tags:
10
  - markdown
11
  ---
12
 
13
- DocParsingBench is a document intelligence benchmark of 1,400 images, systematically collected and annotated from real business workflows. It is the first dataset to systematically catalogue the document elements most frequently encountered in enterprise settings, covering five major domains: finance, legal, scientific research, manufacturing, and education.
14
 
15
  [![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-DocParsingBench-blue)](https://huggingface.co/datasets/SoMarkAI/DocParsingBench)
16
  [![ModelScope](https://img.shields.io/badge/☁️%20ModelScope-DocParsingBench-purple)](https://modelscope.cn/datasets/SoMark/DocParsingBench)
@@ -19,73 +19,71 @@ DocParsingBench is a document intelligence benchmark of 1,400 images, systematic
19
 
20
 
21
  ---
22
- ## 🆕 Latest Updates
23
 
24
- **[2026.04.17]** [DocParsingBench](https://github.com/SoMarkAI/DocParsingBench) evaluation toolkit release. Unified scoring is now available for the three core elements of document parsing — **text, formulas, and tables** — together with **batch CLI evaluation, segment-level matching, visualization analysis, and leaderboard generation**. 📊
25
-
26
- **[2026.03.09]** DocParsingBench dataset release. The first document parsing dataset built for real-world industry scenarios, covering finance, legal, scientific research, manufacturing, and education. Now available on [Hugging Face](https://huggingface.co/datasets/SoMarkAI/DocParsingBench) and [ModelScope](https://modelscope.cn/datasets/SoMark/DocParsingBench)! 🔥🔥🔥
27
 
28
  ---
29
 
30
- ## 🎯 Key Highlights
31
 
32
- ### ✅ 1. Five Real-World Industry Domains
33
 
34
- Samples were drawn from **real business workflows**, preserving **scan noise, stamp occlusion, and blurred characters**:
35
 
36
- | Domain | Document Types | Characteristics |
37
- |--------|----------------|-----------------|
38
- | **Finance** | Brokerage research reports, listed-company annual reports, prospectuses | Multi-column tables, stamped PDF scans, complex tables |
39
- | **Legal** | Legal documents, contract clauses, industry standards | Standardized headers and footers, dense footnotes, cross-referenced citations |
40
- | **Scientific Research** | Academic papers, programming textbooks, full-text patents | Mixed two- and three-column layouts, formula-heavy text, code blocks, chemical formulas |
41
- | **Manufacturing** | Operating SOPs, forms and receipts | Blurry scans, handwritten fill-ins, QR codes / barcodes |
42
- | **Education** | Textbooks across English, chemistry, and mathematics | Chemical structures, reaction equations, multiple-choice options, fill-in-the-blanks |
43
 
44
 
45
- ### ✅ 2. Full Coverage of Complex Layouts
46
 
47
- - **Single-column**: textbooks, legal documents
48
- - **Two-column**: academic papers, some research reports
49
- - **Three or more columns**: brokerage financial reports, data reports
50
- - **Mixed layouts**: interleaved text and figures, tables spanning columns, nested sidebars
51
 
52
  ---
53
 
54
- ## 📊 Dataset Composition
55
 
56
- | Dimension | Category |
57
- |-----------|----------|
58
- | **Total samples** | 1,400 pages |
59
- | **Languages** | Chinese, English, Chinese-English mixed |
60
- | **Industry distribution** | Finance / Legal / Scientific Research / Manufacturing / Education |
61
- | **Layout distribution** | Single-column / Two-column / Three-column / Mixed |
62
- | **Annotation format** | Markdown |
63
- | **Chemistry annotation** | Follows the [SoMarkdown](https://github.com/SoMarkAI/SoMarkDown) specification, combining SMILES with LaTeX to fully render chemical structures |
64
 
65
  ---
66
 
67
- ## 🏗️ Annotation Philosophy
68
- - **Fully human-annotated, without exception** — the "ground truth" of document parsing should not be something a model guesses. In practice, we found that errors from model pre-annotation are silently accepted by annotators: when shown an existing markdown draft, annotators tend to tweak rather than redraw, and the model's mistakes get frozen into the dataset. Every bounding box is drawn stroke by stroke; every character is typed by hand. It is not the fastest approach, but it is the one that yields the highest quality.
69
- - This was the most time-consuming part of the project. We wrote detailed boundary rules for every class of document element, then pilot-annotated, iterated, and finalized them class by class.
70
- - Every spec went through 3+ rounds of pilot annotation, team-wide training, and consistency testing, keeping inter-annotator disagreement on the same page below 5%. In total, 6 full batches were rejected and re-annotated, with an annotation-to-review time ratio of 1 : 0.8.
71
 
72
 
73
- ## 📈 Evaluation Benchmark
74
 
75
- The companion [**industrial document parsing benchmark**](https://github.com/SoMarkAI/DocParsingBench) is released alongside this dataset — community submissions are welcome! 📊
76
 
77
 
78
- ## ❤️ Acknowledgements
79
 
80
- This is work that demands extraordinary patience. A single complex page mixing chemical formulas and tables can take hours; one boundary disagreement can send an entire batch back for re-annotation, costing hundreds of person-hours. Behind every annotated page is a human eye parsing the document's structure and a human hand confirming the markdown. Thanks to every annotator who saw this through.
81
 
82
- We also thank the [SoMarkdown](https://github.com/SoMarkAI/SoMarkDown) project for the chemical-structure annotation specification that let us express LLM-readable chemistry precisely.
83
 
84
  ---
85
 
86
- ## 📖 Citation
87
 
88
- If you use DocParsingBench in your research, please cite it as follows:
89
 
90
  ```bibtex
91
  @misc{DocParsingBench-2026,
@@ -99,15 +97,15 @@ If you use DocParsingBench in your research, please cite it as follows:
99
 
100
  ---
101
 
102
- ## 📜 License
103
 
104
- This project is released under the **ODC-BY (Open Data Commons Attribution License)** and is open to both academic research and commercial use.
105
 
106
  ---
107
- > 🌟 If you find this dataset useful, please consider giving us a ⭐ on ModelScope / Hugging Face!
108
 
109
 
110
- #### How to Download
111
 
112
  **🤗 Hugging Face**
113
 
@@ -132,3 +130,4 @@ git clone https://huggingface.co/datasets/SoMarkAI/DocParsingBench
132
  # ModelScope
133
  git clone https://www.modelscope.cn/datasets/SoMark/DocParsingBench.git
134
  ```
 
 
1
  ---
2
  license: odc-by
3
  tags:
4
+ - 文档解析
5
+ - 文档智能
6
  - OCR Bench
7
  - Doc Parsing
8
  - OCR
 
10
  - markdown
11
  ---
12
 
13
+ DocParsingBench是一个包含1400张图片的文档智能测评数据集，系统性地采集并标注了来自真实业务流程的文档样本，首次系统性总结企业高频的文档元素，涵盖金融、法律、科研、工业、教育五大领域。
14
 
15
  [![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-DocParsingBench-blue)](https://huggingface.co/datasets/SoMarkAI/DocParsingBench)
16
  [![ModelScope](https://img.shields.io/badge/☁️%20ModelScope-DocParsingBench-purple)](https://modelscope.cn/datasets/SoMark/DocParsingBench)
 
19
 
20
 
21
  ---
22
+ ## 🆕 最新资讯
23
 
24
+ **[2026.03.09]** DocParsingBench v1.0 正式发布！首个**面向真实行业场景**的智能文档解析数据集，涵盖金融、法律、科研、工业、教育五大领域。现已上线 Hugging Face！🔥🔥🔥
 
 
25
 
26
  ---
27
 
28
+ ## 🎯 核心亮点
29
 
30
+ ### ✅ 1. 五大真实行业场景
31
 
32
+ 我们系统性地采集并标注了来自**真实业务流程**的文档样本，保留了**扫描噪声、印章遮挡、模糊字符**:
33
 
34
+ | 行业 | 文档类型 | 场景特征 |
35
+ |------|---------|---------|
36
+ | **金融** | 券商研报、上市公司财报、招股说明书 | 多栏报表、带印章的PDF扫描件、复杂表格 |
37
+ | **法律** | 法律文书、合同条款、行业标准规范 | 页眉页脚规范化、脚注密集、参考文献交叉引用 |
38
+ | **科研** | 学术论文、编程教材、专利全文 | 双栏/三栏混合、公式密集、代码段、化学式 |
39
+ | **工业制造** | 操作SOP、表格单据 | 扫描件模糊、手写填空项、二维码/条形码 |
40
+ | **教育** | 书籍教材、英文、化学、数学 | 化学结构式、化学反应式、选择题选项、填空空白 |
41
 
42
 
43
+ ### ✅ 2. 复杂版面全覆盖
44
 
45
+ - **单栏**：教材、法律文书
46
+ - **双栏**：学术论文、部分研报
47
+ - **三栏及以上**：券商财报、数据报表
48
+ - **混合布局**：图文混排、表格跨栏、侧边栏嵌套
49
 
50
  ---
51
 
52
+ ## 📊 数据构成
53
 
54
+ | 维度 | 分类 |
55
+ |------|------|
56
+ | **总样本量** | 1400页 |
57
+ | **语言** | 中文、英文、中英混合|
58
+ | **行业分布** | 金融 / 法律 / 科研 / 工业 / 教育 |
59
+ | **版面分布** | 单栏 / 双栏 / 三栏 / 混合 |
60
+ | **标注格式** | Markdown
61
+ | **化学标注** | 采用 [SoMarkdown](https://github.com/SoMarkAI/SoMarkDown) 规范，将 SMILES 与 LaTeX 结合进行表达，可完整渲染化学结构式
62
 
63
  ---
64
 
65
+ ## 🏗️ 数据处理与标注理念
66
+ - **坚持全人工标注**，因为文档解析的“正确答案”不该是模型猜出来的。因为我们在实践中发现：模型预标注的误差会被标注员“惯性接受”。标注员面对已有markdown结果时，倾向于微调而非重画，导致模型错误被固化。每一根框线，都是一笔一笔画出来的；每一个字都是通过键盘输入的。这不是效率最高的方式，但这是质量最高的方式。
67
+ - 这是本项目投入时间最长的环节。我们为每一类文档元素制定了详细的边界判定规则，逐类试标、迭代、定稿。
68
+ - 每一类规范都经过3轮以上试标+全员培训+一致性测试，确保不同标注员对同一页面的判定差异<5%。累计退回重标批次6轮，标注-复核工时比1 : 0.8
69
 
70
 
71
+ ## 📈 评测基准（待发布）
72
 
73
+ 我们同期发布的DocParsingBench**行业文档解析评测基准**欢迎社区参与评测！📊
74
 
75
 
76
+ ## ❤️ 致谢
77
 
78
+ 这是一个需要极大耐心的工作。一个复杂页面的化学式+表格混排，可能耗时数小时；一次边界分歧导致的整批次退回，意味着数百小时的重来。每一页标注的背后，都是人眼对文档结构的理解，是人手对markdown的确认。感谢所有标注团队的坚持。
79
 
80
+ 同时感谢 [SoMarkdown](https://github.com/SoMarkAI/SoMarkDown) 项目提供的化学结构式标注规范，使我们能够用精确表达大模型可读的化学内容。
81
 
82
  ---
83
 
84
+ ## 📖 引用
85
 
86
+ 如果您在研究中使用了DocParsingBench，请按如下格式引用：
87
 
88
  ```bibtex
89
  @misc{DocParsingBench-2026,
 
97
 
98
  ---
99
 
100
+ ## 📜 许可证
101
 
102
+ 本项目采用 **ODC-BY (Open Data Commons Attribution License)** 许可证，开放学术研究和商业应用。
103
 
104
  ---
105
+ > 🌟 如果您觉得这个数据集有价值，欢迎在 modelscope/Hugging Face 上给我们 ⭐！
106
 
107
 
108
+ #### 下载方法
109
 
110
  **🤗 Hugging Face**
111
 
 
130
  # ModelScope
131
  git clone https://www.modelscope.cn/datasets/SoMark/DocParsingBench.git
132
  ```
133
+