Datasets:
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
- 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, a dataset which contains scraped images from Wikipedia along with their descriptions. WikiMedia released WIT-Base in September 2021, 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, in accordance with WIT-Base's identical license.
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
: adict
feature.image_url
: astr
feature.embedding
: alist
feature.metadata_url
: astr
feature.original_height
: anint
orNaN
feature.original_width
: anint
orNaN
feature.mime_type
: astr
orNone
feature.caption_attribution_description
: astr
orNone
feature.page_url
: astr
feature.attribution_passes_lang_id
: abool
orNone
feature.caption_alt_text_description
: astr
orNone
feature.caption_reference_description
: astr
orNone
feature.caption_title_and_reference_description
: astr
orNone
feature.context_page_description
: astr
orNone
feature.context_section_description
: astr
orNone
feature.hierarchical_section_title
: astr
feature.is_main_image
: abool
orNone
feature.page_changed_recently
: abool
orNone
feature.page_title
: astr
feature.section_title
: astr
orNone
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 dataset, which in turn comes from Google's WIT dataset.
Additional Information
Dataset Curators
Dan Saattrup Nielsen from the The Alexandra Institute curated this dataset.
Licensing Information
The dataset is licensed under the CC BY-SA 4.0 license.