DataComp 12M Image-Text Pair Identifiers

12 million unique identifiers (UIDs) reference a filtered subset of the larger DataComp-1B-BestPool dataset. Apple created this collection to train image-text models that outperform those trained on established benchmarks like CC-12M and YFCC-15M. The dataset card was last updated in February 2025.

Use Cases

• Training contrastive image-text models using the referenced UIDs to fetch corresponding image and caption pairs.
• Benchmarking model performance against results reported in the associated MobileCLIP research paper.
• Filtering and reproducing the specific 12M subset from the larger DataComp-1B-BestPool based on the provided identifiers.

Strengths

• Contains 12 million data points, a substantial scale for training.
• Models trained on this subset show significant performance gains over models trained on CC-12M and YFCC-15M, as cited in the associated paper.

Limitations

• The dataset provides only UIDs, not the actual image or text data, requiring access to the source pool.
• Specific details on image resolution, caption length, and content diversity are not provided in the available description.
• The exact curation method and filtering criteria for selecting these 12M points from the 1B pool are not detailed here.

Provenance

Source

Subset of DataComp-1B-BestPool, curated by Apple.

Collection Method

Filtered selection process from a larger pool of web-scraped image-text pairs; specific methodology referenced in the MobileCLIP paper.

Freshness

Dataset card updated 2025-02-28.