OBELICS: 141M Interleaved Image-Text Documents with 115B Tokens

141 million interleaved image-text web documents containing 115 billion text tokens and 353 million images comprise the OBELICS collection. Created by Hugging Face and updated in 2024, it serves as a massive open-source resource for multimodal AI development.

Use Cases

• Pre-training multimodal models using the 353M images and 115B text tokens
• Training vision-language models to understand context within interleaved document structures
• Benchmarking large-scale data processing pipelines on 141M web documents

Strengths

• 141 million documents
• 353 million images
• 115 billion text tokens
• Apache-2.0 license

Limitations

• Significant storage and compute requirements due to the 115B token scale
• Inherent noise and potential biases common in web-scraped content

Provenance

Source

Hugging Face

Collection Method

scraped and curated

Freshness

Last updated August 2024

Geography

global