File size: 5,151 Bytes
8122b92
df71e99
 
f7d1742
df71e99
cb7a651
 
df71e99
 
 
 
 
 
 
 
 
 
8122b92
 
df71e99
8122b92
df71e99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13c8099
df71e99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f7d1742
df71e99
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
---
language:
- en
license: cc-by-nc-sa-4.0
task_categories:
- table-question-answering
- document-question-answering
tags:
- latex
- tables
- arxiv
- scientific-documents
- table-extraction
- document-understanding
pretty_name: arXiv LaTeX Tables 43k
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files: "*.parquet"
---

# arXiv LaTeX Tables 43k

A curated collection of **~43k LaTeX tables** extracted from all arXiv papers published in December 2025, classified by structural complexity for training and evaluating table extraction models.

## Key Specifications

| Aspect | Details |
|--------|---------|
| **Size** | 43,651 tables |
| **Source** | All arXiv papers published in December 2025 |
| **Complexity Classes** | Simple (27,655), Moderate (10,647), Complex (5,349) |
| **Compilability** | All tables compile with `pdflatex` |
| **License Filter** | Only redistributable CC licenses |
| **Format** | Hugging Face Dataset (Parquet) |

## Dataset Structure

```python
{
  "table_id": str,       # Unique ID: "{arxiv_id}_table_{n}"
  "width_pt": float,     # Rendered table width in points
  "height_pt": float,    # Rendered table height in points
  "tabular": str,        # LaTeX tabular source code
  "complexity": str,     # "simple", "moderate", or "complex"
}
```

## Quick Start

```python
from datasets import load_dataset

ds = load_dataset("piushorn/arxiv-latex-tables-43k")
sample = ds['train'][0]
print(sample['tabular'])    # LaTeX source
print(sample['complexity']) # e.g., "simple"
```

Efficient column-selective loading via DuckDB (fetches only needed columns from Parquet):

```python
import duckdb

parquet_url = (
    "https://huggingface.co/datasets/piushorn/arxiv-latex-tables-43k/"
    "resolve/main/data/train-00000-of-00001.parquet"
)
con = duckdb.connect()
rows = con.execute(f"SELECT tabular, complexity FROM read_parquet('{parquet_url}')").fetchall()
```

## Complexity Classification

Tables are classified into three structural complexity levels based on how well their layout maps to Markdown:

| Level | Criteria | Count |
|-------|----------|-------|
| **Simple** | Regular grid, no cell merging, standard lines | 27,655 |
| **Moderate** | Limited merging (header only or horizontal only), partial `\cline` | 10,647 |
| **Complex** | Multi-dimensional merging, nested structures, irregular layouts | 5,349 |

Classification was performed using GPT-5-mini based on structural analysis of the LaTeX source.

## Extraction Pipeline

1. **Source**: All arXiv papers published in December 2025 with redistributable Creative Commons licenses
2. **Extraction**: `tabular` and `tabular*` environments extracted from LaTeX source files
3. **Cleaning**: Citations (`\cite`, `\ref`, etc.) and captions removed; `\href` simplified to link text
4. **Validation**: Each table compiled with `pdflatex` using standard packages (booktabs, multirow, makecell, etc.)
5. **Measurement**: Page dimensions measured with `pdfinfo` to obtain width/height in points
6. **Classification**: Structural complexity classified by GPT-5-mini

### Compilation Template

```latex
\documentclass[varwidth]{standalone}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{booktabs,multirow,makecell,graphicx,array}
\usepackage{amsmath,amssymb}
\usepackage{colortbl}
\usepackage[table]{xcolor}
\usepackage{adjustbox,caption,diagbox}
\usepackage{pifont}
\begin{document}

% tabular environment inserted here

\end{document}
```

Tables that fail to compile with this template are excluded from the dataset.

## Use Cases

- **Table extraction evaluation**: Benchmark document parsers on LaTeX table recovery
- **Synthetic PDF generation**: Render tables into PDFs with known ground truth
- **Table structure recognition**: Train models to understand table layouts
- **Complexity-aware evaluation**: Evaluate models separately on simple vs. complex tables

## Citation

If you use this dataset in your research or project, please cite our paper:

```bibtex
@misc{horn2026benchmarking,
    title = {Benchmarking PDF Parsers on Table Extraction with LLM-based Semantic Evaluation},
    author = {Horn, Pius and Keuper, Janis},
    year = {2026},
    eprint={2603.18652},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url = {https://arxiv.org/abs/2603.18652}
}
```

📄 **Paper:** [arXiv:2603.18652](https://arxiv.org/abs/2603.18652)

### Acknowledgments

This work has been supported by the German Federal Ministry of Research, Technology and Space (BMFTR) in the program "Forschung an Fachhochschulen in Kooperation mit Unternehmen (FH-Kooperativ)" within the joint project **LLMpraxis** under grant 13FH622KX2.

<p align="center">
  <img src="logos/BMFTR_logo.png" alt="BMFTR_logo" width="150" />
  <img src="logos/HAW_logo.png" alt="HAW_logo" width="150" />
</p>

### Licensing Information

**Content License**: Individual tables retain their original arXiv paper licenses (CC BY 4.0, CC BY-SA 4.0, CC BY-NC-SA 4.0, or CC0 1.0).

**Dataset License**: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/)

---

**Source**: arXiv papers, December 2025
**Dataset Created**: March 2026