OBELICS: 141 Million Interleaved Image-Text Web Documents

OBELICS is a massive, curated collection of 141 million English web documents containing 115 billion text tokens and 353 million images. The documents feature interleaved text paragraphs and images, extracted from Common Crawl dumps. It was created by HuggingFaceM4 and released in August 2023.

Use Cases

• Training multimodal AI models based on interleaved image-text web documents.
• Pretraining language models on the 115 billion text tokens extracted from web pages.
• Developing models for tasks requiring joint understanding of images and surrounding text.
• Analyzing the structure and content of modern web documents for research.

Strengths

• Massive scale with 141 million documents and 353 million images.
• Curated collection with documented filtering steps described in the associated paper.
• Contains 115 billion text tokens for large-scale language modeling.

Limitations

• Column-level documentation and sample data are unavailable, requiring manual inspection after download.
• Row count and file formats are unknown, which may limit suitability assessment.
• Last updated 2023-08-22; freshness should be verified.

Provenance

Source

Common Crawl dumps.

Collection Method

Extracted and filtered from web documents.

Time Range

Data collected between February 2020 and February 2023.

Freshness

Last updated 2023-08-22 20:50:09.