Skip to content

Dataset · mlfoundations

DataComp Image-Text Pair Metadata and Links

mlfoundations/datacomp_pools

A collection of metadata and image URLs with text descriptions for building and filtering training datasets.

What's inside

The repository contains metadata files and direct URL links to image-text pairs from the well-known DataComp project rather than the images themselves. The metadata is distributed under the CC-BY-4.0 license, while copyright for individual images belongs to their respective owners. For downloading and working with the pools, the authors refer to the official project repository.

What you can do with it

  • Building and curating large datasets for training CLIP-scale models
  • Experiments assessing the impact of data quality on multimodal model training
  • Creating custom image-text pair datasets using specified criteria
Downloads this month
45.1K
Likes
26

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("mlfoundations/datacomp_pools")