DataComp-12M: A 12 Million Image-Text Pair Subset for Multimodal Model Training

A subset of 12 million image-text pairs from the DataComp-1B-BestPool collection, released by mlfoundations in 2024. The dataset is designed for training image-text models and is licensed under Creative Commons CC-BY-4.0, though individual images retain their original copyrights. It was introduced in the MobileCLIP paper and is reported to yield better model performance than several established benchmarks.

Use Cases

• Training image-text retrieval models based on the 12 million image-url-text samples.
• Benchmarking multimodal model performance against established datasets like CC-12M and YFCC-15M.
• Fine-tuning CLIP-based models for specific downstream tasks using the provided image-text pairs.
• Researching dataset scaling effects and model performance using the curated subset from a larger pool.

Strengths

• Contains 12 million image-text pairs, a substantial scale for training.
• Reported to yield significantly better model performance than CC-12M and YFCC-15M benchmarks.
• Derived from the curated DataComp-1B-BestPool, suggesting a level of quality filtering.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

mlfoundations via Hugging Face Datasets.

Collection Method

Subset of the DataComp-1B-BestPool collection.

Freshness

Last updated 2024-06-26 22:58:01; freshness should be verified.