OmniCorpus CC 210M: 210 Million Image-Text Interleaved Documents

OpenGVLab's OmniCorpus CC 210M dataset contains 210 million image-text interleaved documents filtered from the Common Crawl web corpus. The dataset is designed for large-scale vision-language model training, as described in an ICLR 2025 spotlight paper. It was last updated on the Hugging Face platform in March 2025.

Use Cases

• Train multimodal large language models based on interleaved image-text sequences.
• Benchmark visual question answering models using the described image-text pairs.
• Conduct research on web-scale data curation and filtering for multimodal datasets.
• Fine-tune image-to-text generation models on a large, diverse corpus.

Strengths

• 210 million documents provide substantial scale for model training.
• Sourced from the diverse Common Crawl web corpus, suggesting broad content variety.
• Dataset is associated with peer-reviewed research (ICLR 2025 spotlight).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific data quality and filtering details require inspection.
• Data may reflect geographic, cultural, and temporal biases inherent to its Common Crawl source.

Provenance

Source

Common Crawl, processed by OpenGVLab.

Collection Method

Filtered from the larger OmniCorpus-CC dataset.

Time Range

null

Freshness

Last updated 2025-03-20 12:47:06.

Geography

null