Hugging Face's logo

Datasets:
kakaobrain
/
coyo-700m

Tasks:
Text-to-Image
Image-to-Text
Zero-Shot Classification
Modalities:
Image
Tabular
Text
Formats:
parquet
Sub-tasks:
image-captioning
Languages:
English
Size:
100M - 1B
ArXiv:
Tags:
image-text pairs
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files
coyo-700m / README.md
brook-park's picture
brook-park
Update README.md (#1)
820cb7b
preview code
|
raw
history blame
11.6 kB
metadata
annotations_creators:
  - no-annotation
language:
  - en
language_creators:
  - other
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: COYO-700M
size_categories:
  - 100M<n<1B
source_datasets:
  - original
tags:
  - image-text pairs
task_categories:
  - text-to-image
  - image-to-text
  - zero-shot-classification
task_ids:
  - image-captioning

Dataset Card for COYO-700M

Table of Contents

Dataset Description

Dataset Summary

COYO-700M is a large-scale dataset that contains 747M image-text pairs as well as many other meta-attributes to increase the usability to train various models. Our dataset follows the similar strategy in previous vision-and-language datasets, collecting many informative pairs of alt-text and its associated image in HTML documents. We expect COYO to be used to train popular large-scale foundation models complementary to other similar datasets. For more details on the data acquisition process, please refer to the technical paper to be released later.

Supported Tasks and Leaderboards

We empirically validated the quality of COYO dataset by re-implementing popular models such as ALIGN, unCLIP, and ViT. We trained these models on COYO-700M or its subsets from scratch, achieving competitive performance to the reported numbers or generated samples in the original papers. Our pre-trained models and training codes will be released soon along with the technical paper.

Languages

The texts in the COYO-700M dataset consist of English.

Dataset Structure

Data Instances

Each instance in COYO-700M represents single image-text pair information with meta-attributes:

{
  'id': 841814333321,
  'url': 'https://blog.dogsof.com/wp-content/uploads/2021/03/Image-from-iOS-5-e1614711641382.jpg',
  'text': 'A Pomsky dog sitting and smiling in field of orange flowers',
  'width': 1000,
  'height': 988,
  'image_phash': 'c9b6a7d8469c1959',
  'text_length': 59,
  'word_count': 11,
  'num_tokens_bert': 13,
  'num_tokens_gpt': 12,
  'num_faces': 0,
  'clip_similarity_vitb32': 0.4296875,
  'clip_similarity_vitl14': 0.35205078125,
  'nsfw_score_opennsfw2': 0.00031447410583496094,
  'nsfw_score_gantman': 0.03298913687467575,
  'watermark_score': 0.1014641746878624,
  'aesthetic_score_laion_v2': 5.435476303100586
}

Data Fields

name type description
id long Unique 64-bit integer ID generated by monotonically_increasing_id()
url string The image URL extracted from the src attribute of the <img> tag
text string The text extracted from the alt attribute of the <img> tag
width integer The width of the image
height integer The height of the image
image_phash string The perceptual hash(pHash) of the image
text_length integer The length of the text
word_count integer The number of words seperated by spaces.
num_tokens_bert integer The number of tokens using BertTokenizer
num_tokens_gpt integer The number of tokens using GPT2TokenizerFast
num_faces integer The number of faces in the image detected by SCRFD
clip_similarity_vitb32 float The cosine similarity between text and image(ViT-B/32) embeddings by OpenAI CLIP
clip_similarity_vitl14 float The cosine similarity between text and image(ViT-L/14) embeddings by OpenAI CLIP
nsfw_score_opennsfw2 float The NSFW score of the image by OpenNSFW2
nsfw_score_gantman float The NSFW score of the image by GantMan/NSFW
watermark_score float The watermark probability of the image by our internal model
aesthetic_score_laion_v2 float The aesthetic score of the image by LAION-Aesthetics-Predictor-V2

Data Splits

Data was not split, since the evaluation was expected to be performed on more widely used downstream task(s).

Dataset Creation

Curation Rationale

Similar to most vision-and-language datasets, our primary goal in the data creation process is to collect many pairs of alt-text and image sources in HTML documents crawled from the web. Therefore, We attempted to eliminate uninformative images or texts with minimal cost and improve our dataset's usability by adding various meta-attributes. Users can use these meta-attributes to sample a subset from COYO-700M and use it to train the desired model. For instance, the num_faces attribute could be used to make a subset like COYO-Faces and develop a privacy-preserving generative model.

Source Data

Initial Data Collection and Normalization

We collected about 10 billion pairs of alt-text and image source in HTML documents in CommonCrawl from Oct. 2020 to Aug. 2021. and eliminated uninformative pairs through the image and/or text level filtering process with minimal cost.

Image Level

  • Include all image formats that Pillow library can decode
  • Less than 5KB image size are dropped
  • Images with aspect ratio is greater than 3.0 are dropped
  • Images with min(width, height) < 200 are dropped
  • Images are dropped if the score of OpenNSFW2 or GantMan/NSFW is higher than 0.5
  • Based on the image pHash value, we removed all duplicate images from external public datasets.
    • ImageNet-1K/21K, Flickr-30K, MS-COCO, CC-3M, CC-12M

Text Level

  • We collected only english text using cld3
  • Consecutive whitespace characters are replaced with a single whitespace and whitespace before and after the sentence are removed
    • e.g. "\n \n Load image into Gallery viewer, valentine&amp;#39;s day roses\n \n" → "Load image into Gallery viewer, valentine&amp;#39;s day roses"
  • Any text with a length of 5 or less has been dropped
  • Text that does not have a noun form has been dropped
  • Text less than 3 words or more than 256 words and text over 1000 words were dropped
  • All texts appearing more than 10 times have been dropped
    • e.g. “thumbnail for”, “image for”, “picture of”

Image-Text Level

  • Based on (image_phash, text), duplicated samples has been dropped
    • Different text may exist for the same image URL.

Who are the source language producers?

Common Crawl is the data source for COYO-700M.

Annotations

Annotation process

The dataset was built in a fully automated process that did not require human annotation.

Who are the annotators?

No human annotation

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

It will be described in a paper to be released soon.

Other Known Limitations

It will be described in a paper to be released soon.

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

[More Information Needed]

Citation Information

[More Information Needed]

Contributions

Thanks to @github-username for adding this dataset.