File size: 4,955 Bytes
a244a6f
152505d
 
 
 
 
 
7177bf7
152505d
7177bf7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a244a6f
152505d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48c9982
 
 
 
 
 
 
 
152505d
 
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
---
tags:
- ocr
- arabic
- document-understanding
- structure-preservation
- computer-vision
pretty_name: Misraj-DocOCR
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: uuid
    dtype: string
  - name: markdown
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 541115359
    num_examples: 400
  download_size: 537036141
  dataset_size: 541115359
---
# Misraj-DocOCR: An Arabic Document OCR Benchmark📄

**Dataset:** [Misraj/Misraj-DocOCR](https://huggingface.co/datasets/Misraj/Misraj-DocOCR)
**Domain:** Arabic Document OCR (text + structure)  
**Size:** 400 expertly verified pages (real + synthetic)  
**Use cases:** OCR, Document Understanding, Markdown/HTML structure preservation  
**Status:** Public 🤝

## ✨ Overview

**Misraj-DocOCR** is a curated, expert-verified benchmark for **Arabic document OCR** with an emphasis on **structure preservation** (Markdown/HTML tables, lists, footnotes, math, watermarks, multi-column, marginalia, etc.). Each page includes high-quality ground truth designed to evaluate both **text fidelity** and **layout/structure fidelity**.

- **Diverse content:** books, reports, forms, scholarly pages, and complex layouts.  
- **Expert-verified ground truth:** human-reviewed for text **and** structure.  
- **Open & reproducible:** intended for fair comparisons and reliable benchmarking.

---

## 📦 Data format

Each example typically includes:
- `uuid`: id of sample
- `image`: page image (PIL-compatible)
- `markdown`: target transcription with structure

### 🔌 Loading

```python
from datasets import load_dataset

ds = load_dataset("Misraj/Misraj-DocOCR")
split = ds["train"]  # or another available split

ex = split[0]
img = ex["image"]  # PIL.Image
gt  = ex.get("markdown") or ex.get("text")
print(gt[:400])
# img.show()  # uncomment in a local environment
```

---

## 🧪 Metrics

We report both **text** and **structure** metrics:

* **Text:** WER ↓, CER ↓, BLEU ↑, ChrF ↑
* **Structure:** **TEDS ↑**, **MARS ↑** (Markdown/HTML structure fidelity)

---

## 🏆 Leaderboard (Misraj-DocOCR)

Best values are **bold**, second-best are <u>underlined</u>.

| Model                         |      WER ↓ |      CER ↓ |      BLEU ↑ |      CHRF ↑ |   TEDS ↑ |       MARS ↑ |
| ----------------------------- | ---------: | ---------: | ----------: | ----------: | -------: | -----------: |
| **Baseer (ours)** |   **0.25** |       0.53 | <u>76.18</u> | <u>87.77</u> |   **66** |   **76.885** |
| Gemini-2.5-pro                | <u>0.37</u> | <u>0.31</u> |   **77.92** |   **89.55** | <u>52</u> | <u>70.775</u> |
| Azure AI Document Intelligence[^azure] |       0.44 |   **0.27** |       62.04 |       82.49 |       42 |       62.245 |
| Dots.ocr                      |       0.50 |       0.40 |       58.16 |       78.41 |       40 |       59.205 |
| Nanonets                      |       0.71 |       0.55 |       42.22 |       67.89 |       37 |       52.445 |
| Qari                          |       0.76 |       0.64 |       38.59 |       64.50 |       21 |       42.750 |
| Qwen2.5-VL-32B                |       0.76 |       0.59 |       37.62 |       62.64 |       41 |       51.820 |
| GPT-5                         |       0.86 |       0.62 |       40.67 |        61.6 |       48 |         54.8 |
| Qwen2.5-VL-3B-Instruct        |       0.87 |       0.71 |       25.39 |       53.42 |       27 |       40.210 |
| Qwen2.5-VL-7B                 |       0.92 |       0.77 |       31.57 |       54.70 |       27 |       40.850 |
| Gemma3-12B                    |       0.96 |       0.80 |       19.75 |       44.53 |       33 |       38.765 |
| Gemma3-4B                     |       1.01 |       0.85 |        9.57 |       31.39 |       28 |       29.695 |
| GPT-4o-mini                   |       1.36 |       1.10 |       22.63 |       47.04 |       26 |        36.52 |
| AIN                           |       1.23 |       1.11 |        1.25 |        2.24 |       21 |       11.620 |
| Aya-vision                    |       1.41 |       1.07 |        2.91 |        9.81 |       26 |       17.905 |

**Highlights:**

* **Baseer (ours)** leads on **WER**, **TEDS**, and **MARS** → strong text & structure fidelity.
* **Gemini-2.5-pro** tops **BLEU/ChrF**; **Azure AI Document Intelligence** attains lowest **CER**.

---

## 📚 How to cite

If you use **Misraj-DocOCR**, please cite:

```bibtex
@misc{hennara2025baseervisionlanguagemodelarabic,
      title={Baseer: A Vision-Language Model for Arabic Document-to-Markdown OCR}, 
      author={Khalil Hennara and Muhammad Hreden and Mohamed Motasim Hamed and Ahmad Bastati and Zeina Aldallal and Sara Chrouf and Safwan AlModhayan},
      year={2025},
      eprint={2509.18174},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.18174}, 
}
```