Dataset · apple
Synthetic descriptions and embeddings from Apple (12M)
apple/DataCompDR-12M
Generated metadata and vector representations for 12 million images from the DataComp subset.
What's inside
The dataset from Apple contains synthetic text descriptions, embeddings, and metadata created by pretrained multimodal models based on a 12-million slice of DataComp-1B in English. The images themselves and original captions are not included—they must be fetched from the mlfoundations/DataComp-12M repository, which shares identifiers (UIDs) by shards.
What you can do with it
- Training text-to-image generation models on improved captions
- Training multimodal encoders and semantic search (image-to-text) systems
- Knowledge distillation from large VLMs using ready-made embeddings
- Tasks
- Text → image, Image → text
- Languages
- English
- Downloads this month
- 1.4K
- Likes
- 39
How to load it
from datasets import load_dataset
ds = load_dataset("apple/DataCompDR-12M")