CC12M and ImageNet21K Recap: 25 Million Images with Recaptioned Text

A combined dataset of approximately 25 million images from the ImageNet21K and CC12M collections, recaptioned by the author. The ImageNet21K portion contains about 13 million examples across roughly 19,000 classes, while the CC12M portion contains 12 million images created in 2021. The dataset was uploaded to Hugging Face by user 'gmongaras' in September 2025.

Use Cases

• Train image classification models based on the roughly 19,000 class labels from ImageNet21K.
• Develop vision-language models based on the recaptioned text associated with images.
• Benchmark model performance on a large-scale, multi-source image dataset.
• Fine-tune models for specific visual concepts using the combined 25 million image examples.

Strengths

• Combines two large-scale image collections for a total of approximately 25 million examples.
• The ImageNet21K portion provides a structured hierarchy of roughly 19,000 classes.
• Images have been recaptioned, potentially improving text descriptions for vision-language tasks.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The author notes the ImageNet21K portion only has ~19K classes instead of the expected 21K.

Provenance

Source

ImageNet21K and CC12M datasets, with processing by the author.

Collection Method

The author states they removed low-quality data, recaptioned the datasets, and combined them.

Freshness

Last updated 2025-09-17 18:21:25; freshness should be verified.