Dataset · mlfoundations
Billion-scale image-text pair metadata
mlfoundations/datacomp_1b
Metadata and image URLs with text descriptions exceeding one billion records for multimodal learning.
What's inside
Contains between 1 billion and 10 billion records structured as tables, text, and image URLs (the actual images must be downloaded separately and are governed by their own copyrights). The metadata is distributed under the permissive CC-BY-4.0 license. The dataset is designed for large-scale data curation and filtering when building multimodal models.
What you can do with it
- Pre-train CLIP-level models and multimodal encoders
- Research web data filtering and deduplication algorithms
- Build custom image-text subsets for specialized tasks
- Size
- 1B–10B rows
- Downloads this month
- 33.8K
- Likes
- 54
Models trained on it
- zw123/delta_clip_h14_336Image text → text 2.9K
- UCSC-VLAA/ViT-bigG-14-CLIPA-336-datacomp1BZero shot image classification 439
- UCSC-VLAA/ViT-bigG-14-CLIPA-datacomp1BZero shot image classification 436
- UCSC-VLAA/ViT-H-14-CLIPA-336-datacomp1BZero shot image classification 225
- UCSC-VLAA/ViT-L-14-CLIPA-datacomp1BZero shot image classification 221
- zw123/delta_clip_l14_224Image text → text 209
- UCSC-VLAA/ViT-H-14-CLIPA-datacomp1BZero shot image classification 98
- UCSC-VLAA/ViT-L-14-CLIPA-336-datacomp1BZero shot image classification 93
How to load it
from datasets import load_dataset
ds = load_dataset("mlfoundations/datacomp_1b")