Dataset · facebook
70M image-text pairs from Facebook (PMD)
facebook/pmd
A large multimodal corpus of 70 million English image-text pairs from well-known open databases.
What's inside
The dataset is a collection of 70 million English image-text pairs featuring 68 million unique images. The corpus is compiled from multiple well-known open sources, including COCO, Conceptual Captions (and 12M), WIT, Localized Narratives, RedCaps, SBU Captions, Visual Genome, and a portion of YFCC100M. The dataset was first introduced in the research paper on the FLAVA model.
What you can do with it
- Pre-training multimodal models (VLMs) and encoders like CLIP
- Training image captioning systems
- Building cross-modal image text-retrieval systems
- Tasks
- Image → text
- Languages
- English
- Size
- 10M–100M rows
- Downloads this month
- 462
- Likes
- 41
- Access
- on request on Hugging Face
Models trained on it
- HuggingFaceM4/idefics-80b-instructText generation 6.4K
- HuggingFaceM4/idefics2-8bImage text → text 5.4K
- HuggingFaceM4/idefics-9bText generation 3.3K
- HuggingFaceM4/idefics-9b-instructText generation 1.2K
- HuggingFaceM4/idefics2-8b-chattyImage text → text 538
- HuggingFaceM4/idefics-80bText generation 223
- HuggingFaceM4/idefics2-8b-baseImage text → text 216
- HuggingFaceM4/idefics2-8b-AWQImage text → text 80
How to load it
from datasets import load_dataset
ds = load_dataset("facebook/pmd")