File size: 2,646 Bytes
d341e83
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15212df
d341e83
 
 
2c095d4
d341e83
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: apache-2.0
language:
- multilingual
- af
- am
- ar
- as
- azb
- be
- bg
- bm
- bn
- bo
- bs
- ca
- ceb
- cs
- cy
- da
- de
- du
- el
- en
- eo
- es
- et
- eu
- fa
- fi
- fr
- ga
- gd
- gl
- ha
- hi
- hr
- ht
- hu
- id
- ig
- is
- it
- iw
- ja
- jv
- ka
- ki
- kk
- km
- ko
- la
- lb
- ln
- lo
- lt
- lv
- mi
- mr
- ms
- mt
- my
- 'no'
- oc
- pa
- pl
- pt
- qu
- ro
- ru
- sa
- sc
- sd
- sg
- sk
- sl
- sm
- so
- sq
- sr
- ss
- sv
- sw
- ta
- te
- th
- ti
- tl
- tn
- tpi
- tr
- ts
- tw
- uk
- ur
- uz
- vi
- war
- wo
- xh
- yo
- zh
- zu
task_categories:
- image-to-text
tags:
- ocr
size_categories:
- 1M<n<10M
---
# Synthdog Multilingual

<!-- Provide a quick summary of the dataset. -->

The Synthdog dataset created for training in [Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model](https://gregor-ge.github.io/Centurio/).

Using the [official Synthdog code](https://github.com/clovaai/donut/tree/master/synthdog), we created >1 million training samples for improving OCR capabilities in Large Vision-Language Models.

## Dataset Details

We provide the images for download in two `.tar.gz` files. Download and extract them in folders of the same name (so `cat images.tar.gz.* | tar xvzf -C images; tar xvzf images.tar.gz -C images_non_latin`).
The image path in the dataset expects images to be in those respective folders for unique identification.


Every language has the following amount of samples: 500,000 for English, 10,000 for non-Latin scripts, and 5,000 otherwise.

Text is taken from Wikipedia of the respective languages. Font is `GoNotoKurrent-Regular`.


> Note: Right-to-left written scripts (Arabic, Hebrew, ...) are unfortunatly writte correctly right-to-left but also bottom-to-top. We were not able to fix this issue. However, empirical results in Centurio suggest that this data is still helpful for improving model performance.
> 



## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

```
@article{centurio2025,
  author       = {Gregor Geigle and
                  Florian Schneider and
                  Carolin Holtermann and
                  Chris Biemann and
                  Radu Timofte and
                  Anne Lauscher and
                  Goran Glava\v{s}},
  title        = {Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model},
  journal      = {arXiv},
  volume       = {abs/2501.05122},
  year         = {2025},
  url          = {https://arxiv.org/abs/2501.05122},
  eprinttype    = {arXiv},
  eprint       = {2501.05122},
}
```