Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -33,15 +33,15 @@ size_categories:
|
|
33 |
|
34 |
## π Dataset Summary
|
35 |
|
36 |
-
The QARI Markdown Mixed Dataset is a specialized synthetic dataset designed for training Arabic OCR models with a focus on complex document layouts and HTML structure understanding.
|
37 |
This dataset is part of the QARI-OCR project, which achieves state-of-the-art performance in Arabic text recognition.
|
38 |
|
39 |
-
This dataset contains **
|
40 |
- π€ **Full diacritical marks (tashkeel)** support
|
41 |
- π **Mixed font sizes** within documents (headers, body text, annotations)
|
42 |
- π¨ **12 distinct Arabic fonts** ranging from common Naskh to ornate calligraphic styles
|
43 |
- π **Realistic document layouts** with structural HTML tags
|
44 |
-
- πΌοΈ **
|
45 |
|
46 |
## π― Intended Use
|
47 |
|
@@ -55,16 +55,57 @@ This dataset is specifically designed for:
|
|
55 |
|
56 |
| Metric | Value |
|
57 |
|--------|-------|
|
58 |
-
| **Total Images** |
|
59 |
-
| **
|
|
|
|
|
|
|
60 |
| **Font Variety** | 12 Arabic fonts |
|
61 |
| **Font Size Range** | 14px - 100px |
|
62 |
-
| **Degradation Types** | 3 (Clean, Moderate, Heavy) |
|
63 |
| **Diacritics Support** | β
Full tashkeel |
|
64 |
| **HTML Structure** | β
Preserved |
|
65 |
| **Layout Complexity** | β
High (mixed sizes, headers) |
|
66 |
|
|
|
67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
|
70 |
|
|
|
33 |
|
34 |
## π Dataset Summary
|
35 |
|
36 |
+
The QARI v0.3 Markdown Mixed Dataset is a specialized synthetic dataset designed for training Arabic OCR models with a focus on complex document layouts and HTML structure understanding.
|
37 |
This dataset is part of the QARI-OCR project, which achieves state-of-the-art performance in Arabic text recognition.
|
38 |
|
39 |
+
This dataset contains **37,000 synthetically generated Arabic document images** (29.6k train, 3.7k validation, 3.7k test) with corresponding ground truth text in HTML/Markdown format, featuring:
|
40 |
- π€ **Full diacritical marks (tashkeel)** support
|
41 |
- π **Mixed font sizes** within documents (headers, body text, annotations)
|
42 |
- π¨ **12 distinct Arabic fonts** ranging from common Naskh to ornate calligraphic styles
|
43 |
- π **Realistic document layouts** with structural HTML tags
|
44 |
+
- πΌοΈ **Multiple text sources** including Basma2423 and YoussefAnwar Arabic news
|
45 |
|
46 |
## π― Intended Use
|
47 |
|
|
|
55 |
|
56 |
| Metric | Value |
|
57 |
|--------|-------|
|
58 |
+
| **Total Images** | 37,000 |
|
59 |
+
| **Train Set** | 29,600 (80%) |
|
60 |
+
| **Validation Set** | 3,700 (10%) |
|
61 |
+
| **Test Set** | 3,700 (10%) |
|
62 |
+
| **Text Sources** | oddadmix/Basma2423-Text-with-Diacritics-Correction + YoussefAnwar/Arabic-news |
|
63 |
| **Font Variety** | 12 Arabic fonts |
|
64 |
| **Font Size Range** | 14px - 100px |
|
|
|
65 |
| **Diacritics Support** | β
Full tashkeel |
|
66 |
| **HTML Structure** | β
Preserved |
|
67 |
| **Layout Complexity** | β
High (mixed sizes, headers) |
|
68 |
|
69 |
+
## π§ Data Generation Pipeline
|
70 |
|
71 |
+
<div align="center">
|
72 |
+
|
73 |
+
| Stage | Process | Details |
|
74 |
+
|-------|---------|---------|
|
75 |
+
| **1. Text Collection** | Source gathering | Basma2423 (with diacritics) + YoussefAnwar Arabic news |
|
76 |
+
| **2. HTML Templating** | Layout generation | Mixed font sizes, structural elements |
|
77 |
+
| **3. Rendering** | WeasyPrint β PDF β Image | High-quality document rendering |
|
78 |
+
| **4. Degradation** | Synthetic noise | Clean / Moderate / Heavy variants |
|
79 |
+
|
80 |
+
</div>
|
81 |
|
82 |
+
## π Model Performance
|
83 |
+
|
84 |
+
When used to train QARI v0.3, this dataset enables:
|
85 |
+
|
86 |
+
| Metric | Score |
|
87 |
+
|--------|-------|
|
88 |
+
| **Character Error Rate (CER)** | 0.300 |
|
89 |
+
| **Word Error Rate (WER)** | 0.485 |
|
90 |
+
| **BLEU Score** | 0.545 |
|
91 |
+
| **Training Time** | 11 hours |
|
92 |
+
| **COβ Emissions** | 1.88 kg eq. |
|
93 |
+
|
94 |
+
### Key Advantages:
|
95 |
+
- π **Superior layout understanding** compared to plain text models
|
96 |
+
- π·οΈ **HTML tag preservation** for structured document conversion
|
97 |
+
- β‘ **Resource efficient** - 5x less training time than larger datasets
|
98 |
+
- π― **Specialized performance** for document structure tasks
|
99 |
+
|
100 |
+
## Citation
|
101 |
+
|
102 |
+
```markdown
|
103 |
+
@article{wasfy2025qari,
|
104 |
+
title={QARI-OCR: High-Fidelity Arabic Text Recognition through Multimodal Large Language Model Adaptation},
|
105 |
+
author={Wasfy, Ahmed and Nacar, Omer and Elkhateb, Abdelakreem and Reda, Mahmoud and Elshehy, Omar and Ammar, Adel and Boulila, Wadii},
|
106 |
+
journal={arXiv preprint arXiv:2506.02295},
|
107 |
+
year={2025}
|
108 |
+
}
|
109 |
+
```
|
110 |
|
111 |
|