File size: 5,537 Bytes
6761d73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
pretty_name: Danish WIT
language:
- da
license:
- cc-by-sa-4.0
size_categories:
- 100K<n<1M
source_datasets:
- wikimedia/wit_base
task_categories:
- image-to-text
- zero-shot-image-classification
- feature-extraction
task_ids:
- image-captioning
---

# Dataset Card for Danish WIT

## Dataset Description

- **Repository:** <https://gist.github.com/saattrupdan/bb6c9c52d9f4b35258db2b2456d31224>
- **Point of Contact:** [Dan Saattrup Nielsen](mailto:[email protected])
- **Size of downloaded dataset files:** 7.5 GB
- **Size of the generated dataset:** 7.8 GB
- **Total amount of disk used:** 15.3 GB

### Dataset Summary

Google presented the Wikipedia Image Text (WIT) dataset in [July
2021](https://dl.acm.org/doi/abs/10.1145/3404835.3463257), a dataset which contains
scraped images from Wikipedia along with their descriptions. WikiMedia released
WIT-Base in [September
2021](https://techblog.wikimedia.org/2021/09/09/the-wikipedia-image-caption-matching-challenge-and-a-huge-release-of-image-data-for-research/),
being a modified version of WIT where they have removed the images with empty
"reference descriptions", as well as removing images where a person's face covers more
than 10% of the image surface, along with inappropriate images that are candidate for
deletion. This dataset is the Danish portion of the WIT-Base dataset, consisting of
roughly 160,000 images with associated Danish descriptions. We release the dataset
under the [CC BY-SA 4.0 license](https://creativecommons.org/licenses/by-sa/4.0/), in
accordance with WIT-Base's [identical
license](https://huggingface.co/datasets/wikimedia/wit_base#licensing-information).


### Supported Tasks and Leaderboards

Training machine learning models for caption generation, zero-shot image classification
and text-image search are the intended tasks for this dataset. No leaderboard is active
at this point.


### Languages

The dataset is available in Danish (`da`).


## Dataset Structure

### Data Instances

- **Size of downloaded dataset files:** 7.5 GB
- **Size of the generated dataset:** 7.8 GB
- **Total amount of disk used:** 15.3 GB

An example from the `train` split looks as follows.
```
{
    "image": {
        "bytes": b"\xff\xd8\xff\xe0\x00\x10JFIF...",
        "path": None
    },
    "image_url": "https://upload.wikimedia.org/wikipedia/commons/4/45/Bispen_-_inside.jpg",
    "embedding": [2.8568285, 2.9562542, 0.33794892, 8.753725, ...],
    "metadata_url": "http://commons.wikimedia.org/wiki/File:Bispen_-_inside.jpg",
    "original_height": 3161,
    "original_width": 2316,
    "mime_type": "image/jpeg",
    "caption_attribution_description": "Kulturhuset Bispen set indefra. Biblioteket er til venstre",
    "page_url": "https://da.wikipedia.org/wiki/Bispen",
    "attribution_passes_lang_id": True,
    "caption_alt_text_description": None,
    "caption_reference_description": "Bispen set indefra fra 1. sal, hvor ....",
    "caption_title_and_reference_description": "Bispen [SEP] Bispen set indefra ...",
    "context_page_description": "Bispen er navnet på det offentlige kulturhus i ...",
    "context_section_description": "Bispen er navnet på det offentlige kulturhus i ...",
    "hierarchical_section_title": "Bispen",
    "is_main_image": True,
    "page_changed_recently": True,
    "page_title": "Bispen",
    "section_title": None
}
```

### Data Fields

The data fields are the same among all splits.

- `image`: a `dict` feature.
- `image_url`: a `str` feature.
- `embedding`: a `list` feature.
- `metadata_url`: a `str` feature.
- `original_height`: an `int` or `NaN` feature.
- `original_width`: an `int` or `NaN` feature.
- `mime_type`: a `str` or `None` feature.
- `caption_attribution_description`: a `str` or `None` feature.
- `page_url`: a `str` feature.
- `attribution_passes_lang_id`: a `bool` or `None` feature.
- `caption_alt_text_description`: a `str` or `None` feature.
- `caption_reference_description`: a `str` or `None` feature.
- `caption_title_and_reference_description`: a `str` or `None` feature.
- `context_page_description`: a `str` or `None` feature.
- `context_section_description`: a `str` or `None` feature.
- `hierarchical_section_title`: a `str` feature.
- `is_main_image`: a `bool` or `None` feature.
- `page_changed_recently`: a `bool` or `None` feature.
- `page_title`: a `str` feature.
- `section_title`: a `str` or `None` feature.

### Data Splits

Roughly 2.60% of the WIT-Base dataset comes from the Danish Wikipedia. We have split
the resulting 168,740 samples into a training set, validation set and testing set of
the following sizes:

|   split | samples |
|---------|--------:|
| train   | 167,460 |
| val     | 256     |
| test    | 1,024   |


## Dataset Creation

### Curation Rationale

It is quite cumbersome to extract the Danish portion of the WIT-Base dataset,
especially as the dataset takes up 333 GB of disk space, so the curation of Danish-WIT
is purely to make it easier to work with the Danish portion of it.

### Source Data

The original data was collected from WikiMedia's
[WIT-Base](https://huggingface.co/datasets/wikimedia/wit_base) dataset, which in turn
comes from Google's [WIT](https://huggingface.co/datasets/google/wit) dataset.


## Additional Information

### Dataset Curators

[Dan Saattrup Nielsen](https://saattrupdan.github.io/) from the [The Alexandra
Institute](https://alexandra.dk/) curated this dataset.

### Licensing Information

The dataset is licensed under the [CC BY-SA 4.0
license](https://creativecommons.org/licenses/by-sa/4.0/).