Skip to content

Dataset · Salesforce

218M image-caption pairs and detailed descriptions from Salesforce

Salesforce/blip3-kale

A large open dataset of 218M images with expanded descriptions enriched with web facts.

What's inside

The open dataset from Salesforce includes 218 million image-text pairs (metadata specifies a 100M-1B range and modalities image, text, tabular). The English descriptions are dense captions enriched with factual knowledge from web sources. The dataset was created for multimodal pre-training research.

What you can do with it

  • Multimodal pre-training of large vision-language models (VLMs)
  • Train systems to generate detailed and factually accurate image descriptions
  • Fine-tune models for image-to-text tasks leveraging external knowledge
Tasks
Image → text
Languages
English
Size
100M–1B rows
Downloads this month
2.7K
Likes
48

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("Salesforce/blip3-kale")