metadata

annotations_creators:
  - crowdsourced
language:
  - tr
language_creators:
  - crowdsourced
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: laion2B-multi-turkish-subset
size_categories:
  - 10M<n<100M
task_categories:
  - text-to-image
  - image-to-text

Dataset Card for laion2B-multi-turkish-subset

Dataset Description

Homepage: laion-5b
Huggingface: laion/laion2B-multi
Point of Contact: mcemilg

Dataset Summary

LAION-5B is a large scale openly accessible image-text dataset contains text from multiple languages. This is a Turkish subset data of laion/laion2B-multi. It's compatible to be used with image2dataset to fetch the images at scale.

Data Structure

DatasetDict({
    train: Dataset({
        features: ['SAMPLE_ID', 'URL', 'TEXT', 'HEIGHT', 'WIDTH', 'LICENSE', 'LANGUAGE', 'NSFW', 'similarity'],
        num_rows: 34638627
    })
})

{
   'SAMPLE_ID': Value(dtype='int64', id=None),
   'URL': Value(dtype='string', id=None),
   'TEXT': Value(dtype='string', id=None),
   'HEIGHT': Value(dtype='int64', id=None),
   'WIDTH': Value(dtype='int64', id=None),
   'LICENSE': Value(dtype='string', id=None),
   'LANGUAGE': Value(dtype='string', id=None),
   'NSFW': Value(dtype='string', id=None),
   'similarity': Value(dtype='float64', id=None)
}

Notes

The data was basically processed to drop non-Turkish and irrelevant texts before published. Both FastText and langdetect libraries were used to identify if the text is Turkish or not. The cleaning process can be summarized as follows:

replace """ with empty str
remove URLs in texts
Drop if both FastText and LangDetect are highly confident with there is no Turkish in text.
Drop empty text fields.

License

CC-BY-4.0