Dataset · UCSC-VLAA
1B image-text pairs with re-captions
UCSC-VLAA/Recap-DataComp-1B
Large DataComp-1B image-text corpus with improved captions from LLaVA-1.5-LLaMA3-8B.
What's inside
The dataset contains images and text descriptions; captions were recreated using the LLaVA-1.5-LLaMA3-8B model for better alignment and detail. Based on DataComp-1B, the card lists zero-shot classification, text-retrieval, image-to-text, and text-to-image tasks. Metadata volume ranges from 1B to 10B examples.
What you can do with it
- Training multimodal image-text models
- Evaluating zero-shot classification and retrieval
- Text-to-image and image-to-text research
- Tasks
- Zero shot classification, Text retrieval, Image → text, Text → image
- Size
- 1B–10B rows
- Downloads this month
- 23.5K
- Likes
- 213
Models trained on it
- UCSC-VLAA/ViT-L-16-HTxt-Recap-CLIPZero shot image classification 7K
- zw123/delta_clip_h14_336Image text → text 2.9K
- zw123/delta_clip_l14_224Image text → text 209
- UCSC-VLAA/ViT-L-14-CLIPS-224-Recap-DataComp-1B 36
- UCSC-VLAA/ViT-L-14-CLIPS-Recap-DataComp-1B 32
- UCSC-VLAA/ViT-H-14-CLIPS-224-Recap-DataComp-1B 23
- NikshepShetty/Florence-2-Recap-DataComp 18
- kerzgrr/imagenv1mText → image 10
How to load it
from datasets import load_dataset
ds = load_dataset("UCSC-VLAA/Recap-DataComp-1B")