COYO-700M
Large-scale Image-Text Pair Dataset
COYO-700M is a large-scale image-text pair dataset released by Kakao Brain, containing 747 million pairs of rigorously filtered images and ALT text descriptions. The data is extracted from CommonCrawl, undergoes multiple screenings, and is widely used in visual language model research such as CLIP and text-to-image generation.
Dataset Highlights
High-quality large-scale image-text pairs data, supporting cutting-edge research in visual language models
Large-scale Coverage
Includes 747 million carefully selected image-text pairs, far exceeding most similar public datasets, providing ample samples for large-scale visual language pre-training.
Strict Filtering
Undergoes multiple screenings including URL safety checks, image size and format validation, text length and language detection, NSFW filtering, ensuring reliable data quality.
High Diversity
Data sourced from CommonCrawl web scraping, covering a wide range of thematic areas and visual scenes, possessing rich semantic diversity and distribution characteristics.
CLIP Benchmark
Performs excellently in ViT model training, with performance comparable to or even surpassing LAION-400M, making it a high-quality data choice for CLIP training.
Open and Reproducible
The complete data processing workflow is documented in detail, licensed under CC BY 4.0, supporting free use for academic research and commercial applications.
Kakao Brain Quality
Released by Kakao Brain, a leading AI research laboratory in Asia, the data quality and processing workflow have been rigorously validated and are trustworthy.
Applicable Scenarios
From visual language pre-training to cross-modal retrieval, covering various research directions
CLIP Training
Used for contrastive visual language pre-training, learning cross-modal alignment representations of images and text
Text-to-Image Model
Training generative image models to achieve generation from text descriptions to high-quality images
Zero-Shot Classification
No task-specific labeled data required, directly using pre-trained models for image classification
Image-Text Retrieval
Building a cross-modal search and matching system to achieve bidirectional retrieval of images and text
Data Preview
The following is a JSONL format example of the COYO-700M dataset, with each line containing a record of an image-text pair
{"id": 0, "url": "https://example.com/photo1.jpg", "text": "Beautiful cherry blossom trees lining a peaceful river in spring", "width": 1200, "height": 800, "similarity": 0.298}
{"id": 1, "url": "https://example.com/photo2.jpg", "text": "Professional workspace with dual monitors and ergonomic setup", "width": 1920, "height": 1280, "similarity": 0.315}
{"id": 2, "url": "https://example.com/photo3.jpg", "text": "Traditional Korean temple architecture against mountain backdrop", "width": 1600, "height": 1067, "similarity": 0.327}
3 Steps to Get Started Quickly
Quickly start your multimodal research project from browsing to training
Browse Dataset
View the details of the COYO-700M dataset on the Ace Data Cloud platform, and learn about field descriptions, data scale, and licensing agreements.
Download Data
Download the metadata shard files in Parquet format, and then use the img2dataset tool to batch obtain the original image data.
Load and Train
Load the image-text pair data into your multimodal training pipeline and start CLIP pre-training or text-to-image model training.
Start Exploring COYO-700M Data
747 million high-quality image-text pairs, open license, available now. Whether you are researching visual language models or training generative AI, this dataset is an ideal choice.