🐺 COYO-700M: Image-Text Pair Dataset

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 a similar strategy to 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.

More details on the data acquisition process can be found in [our paper] (which will be updated soon).

Updates

• Nov 2022: Release COYO-Labeled-300M
• Aug 2022: Release COYO-700M Dataset

Data Collection Process

• We collected about 10 billion pairs of alt-text and image sources in HTML documents in CommonCrawl from Oct. 2020 to Aug. 2021. and eliminated uninformative pairs through the image and text level filtering process with minimal cost. The following figure outlines our data collection procedure.

https://en.wikipedia.org/wiki/Napoleon

Data Filtering

Image Level

• Included all image formats that Pillow library can decode. (JPEG, WEBP, PNG, BMP, ...)
• Removed images less than 5KB image size.
• Removed images with an aspect ratio greater than 3.0.
• Removed images with min(width, height) < 200.
• Removed images with a score of OpenNSFW2 or GantMan/NSFW higher than 0.5.
• Removed all duplicate images based on the image pHash value from external public datasets.
  • ImageNet-1K/21K, Flickr-30K, MS-COCO, CC-3M, CC-12M

Text Level

• Collected only English text using cld3.
• Replaced consecutive whitespace characters with a single whitespace and removed the whitespace before and after the sentence.
  • e.g. "\n \n Load image into Gallery viewer, valentine's day roses\n \n" → "Load image into Gallery viewer, valentine's day roses"
• Removed texts with a length of 5 or less.
• Removed texts that do not have a noun form.
• Removed texts with less than 3 words or more than 256 words and texts over 1000 in length.
• Removed texts appearing more than 10 times.
  • e.g. “thumbnail for”, “image for”, “picture of”
• Removed texts containing NSFW words collected from profanity_filter, better_profanity, and google_twunter_lol.

Image-Text Level

• Removed duplicated samples based on (image_phash, text).
  • Different text may exist for the same image URL.

Dataset Preview

id	url	text	width	height	image_phash	text_length	word_count	num_tokens_bert	num_tokens_gpt	num_faces	clip_similarity_vitb32	clip_similarity_vitl14	nsfw_score_opennsfw2	nsfw_score_gantman	watermark_score	aesthetic_score_laion_v2
4896263451343		Fishing Fleet (Monterey), California art by Art Riley. HD giclee art prints for sale at CaliforniaWatercolor.com - original California paintings, & premium giclee prints for sale	600	447	bac58374982e0fc7	178	25	39	40	0	0.319336	0.248169	2.54512e-05	0.0293861	0.0406009	7.04812
1425929344479		The Gate by Pete2453	600	347	8374726575bc0f8a	20	4	6	6	0	0.24939	0.203735	6.97374e-06	0.00823276	0.0721415	6.98521
7456063527931		Beautiful Pictures From the Shores of the Mythical Land (42	600	320	949d1fe559e2cc90	59	10	11	14	0	0.290771	0.179321	0.0130615	0.0178628	0.489642	6.94643
3221225511175		contemporary expensive lighting fixtures with minimum lighting	800	499	e5ea35075ab912c6	62	7	7	8	0	0.263916	0.217896	0.000990868	0.0137114	0.0960748	4.57594
5626407855002		Nintendo Co.'s Super Mario is displayed on coffee mugs for sale at the Nintendo World store in New York, U.S., on Friday, May 17, 2013.	2000	1309	9311891e9437f4f3	135	27	37	35	0	0.400878	0.316650	0.00362968	0.0317519	0.0022693	6.324910
1125282207474		FILE PHOTO: A rainbow appears on the Auckland skyline featuring Sky Tower in New Zealand	800	525	85b89c0166ee63be	88	15	16	16	0	0.4453125	0.3505859	2.640485e-05	0.012074	0.0219129	5.294523
1434519186493		A man covers himself with algae as he poses for photographs on a beach in Qingdao, Shandong province on Tuesday, July 23, 2013. -- FILE PHOTO: REUTERS	860	573	f2c48dabbf93810a	150	26	35	36	7	0.4165039	0.3427734	0.025009	0.01608	0.072775	6.833739

Dataset Numbers

	count	ratio
# of image-text pairs	746,972,269	100.00%
# of unique urls	656,114,783	87.84%
# of unique image_phash	579,679,137	77.60%
# of unique text	566,253,888	75.81%

Meta-Attributes

Attributes

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 separated 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

Statistics

• Statistics for numeric columns

	width	height	text_length	word_count	num_tokens_bert	num_tokens_gpt	num_faces
mean	621.78	540.99	68.53	11.13	15.75	17.24	0.60
min	200	200	6	3	1	3	0
max	21449	22507	1000	323	811	1523	736

	watermark_score	clip_similarity_vitb32	clip_similarity_vitl14	aesthetic_score_laion_v2	nsfw_score_opennsfw2
mean	0.178544	0.291266	0.254632	4.769132	0.012903
min	0.0	-0.080871	-0.176269	1.171712	0.0
max	1.0	0.591796	0.581542	8.082607	0.499755

• Image Size
• CLIP Similarity
• Text Length & Word Size
• Watermark score
• Aesthetic score
• Number of faces
• For more detailed statistics on COYO-700M, please see the Data Studio report on COYO-700M.

Getting Started

Download

• You can download the dataset from Huggingface Dataset.
• For more information on downloading the image dataset, please refer to download/README.md.

Usage

• pytorch
• tensorflow

Experiments

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. Since this observation supports the high quality of our dataset, we hope it to be continuously updated with open collaboration. Our pre-trained models and training codes will be released soon along with the technical report.

ALIGN

Model	Data	ImageNet KNN	COCO I2T	COCO T2I
EfficientNet-B7 + BERT-base	ALIGN-1.8B	69.300	55.400	41.700
EfficientNet-B7 + BERT-base	COYO-700M	68.618	59.000	42.419

• Our experiment setup followed ALIGN.
  • We increased the batch size from 16K to 64K and reduced training steps by 1/4 for faster training.

unCLIP (OpenAI DALL·E 2)

A high quality picture of a medieval knight with golden armor	A person with the head of a cat in the style of Andy Warhol

A pencil drawing of an astronaut riding a horse	Goryeo celadon in the shape of darth vader

• We implemented the smaller version of unCLIP to validate the effectiveness of COYO for the text-conditional generation tasks.
• Specifically, we tried to reproduce three components of the original unCLIP: diffusion-based prior, decoder with some modifications, and super-resolution model for upscaling 64x64 into 256x256px.
• Detailed information on our modified version of unCLIP and quantitative analysis would be included in the upcoming technical report.

ViT

Model	Data	ImageNet Validation Top-1 Acc
ViT-L/16	JFT-300M	87.76%
ViT-L/16	COYO-Labeled-300M	87.24%

• We also provide COYO-Labeled-300M by adding machine-generated vision labels to a subset of COYO-700M for comparison with the JFT-300M.
  • We first removed the duplicated images by image_phash.
  • Then, we labeled 300M unique images into 21,841 classes by EfficientNetV2-XL trained with ImageNet-21K dataset.
• Our experiment setup followed ViT.
• We also provide vit pre-training, fine-tuning code for reproducibility with weight files.

Citation

If you apply this dataset to any project and research, please cite our code:

@misc{kakaobrain2022coyo-700m,
  title         = {COYO-700M: Image-Text Pair Dataset},
  author        = {Byeon, Minwoo and Park, Beomhee and Kim, Haecheon and Lee, Sungjun and Baek, Woonhyuk and Kim, Saehoon},
  year          = {2022},
  howpublished  = {\url{https://github.com/kakaobrain/coyo-dataset}},
}

People

• Minwoo Byeon (@mwbyeon)
• Beomhee Park (@beomheepark)
• Haecheon Kim (@HaecheonKim)
• Sungjun Lee (@justhungryman)
• Woonhyuk Baek (@wbaek)
• Saehoon Kim (@saehoonkim)
• and Kakao Brain Large-Scale AI Studio

Disclaimer & Content Warning

The COYO dataset is recommended to be used for research purposes. Kakao Brain tried to construct a "Safe" dataset when building the COYO dataset. (See Data Filtering Section) Kakao Brain is constantly making efforts to create more "Safe" datasets. However, despite these efforts, this large-scale dataset was not hand-picked by humans to avoid the risk due to its very large size (over 700M). Keep in mind that the unscreened nature of the dataset means that the collected images can lead to strongly discomforting and disturbing content for humans. The COYO dataset may contain some inappropriate data, and any problems resulting from such data are the full responsibility of the user who used it. Therefore, it is strongly recommended that this dataset be used only for research, keeping this in mind when using the dataset, and Kakao Brain does not recommend using this dataset as it is without special processing to clear inappropriate data to create commercial products.

License

The COYO dataset of Kakao Brain is licensed under CC-BY-4.0 License. The full license can be found in the LICENSE.cc-by-4.0 file. The dataset includes “Image URL” and “Text” collected from various sites by analyzing Common Crawl data, an open data web crawling project. The collected data (images and text) is subject to the license to which each content belongs.

Obligation to use

While Open Source may be free to use, that does not mean it is free of obligation. To determine whether your intended use of the COYO dataset is suitable for the CC-BY-4.0 license, please consider the license guide. If you violate the license, you may be subject to legal action such as the prohibition of use or claim for damages depending on the use.

Contact

COYO dataset was released as an open source in the hope that it will be helpful to many research institutes and startups for research purposes. We look forward to contacting us from various places who wish to cooperate with us.

[email protected]