Skip to content

Dataset · nvidia

Multimodal data for vision-language models

nvidia/Nemotron-Image-Training-v3

A large collection of image-centric data for training and evaluating vision-language models; subsets in JSONL with descriptions.

What's inside

This is a set of image-centric multimodal training data for vision-language models. The release is structured as a multi-component release: each subset has a standardized conversation JSONL and a card with sources, licenses, and media layouts. The size ranges from 1 to 10 million examples; languages and exact subset composition are not disclosed in the card.

What you can do with it

  • SFT of vision-language models
  • Training image-text-to-text
  • Evaluating VQA
Tasks
Visual question answering, Image text → text
Size
1M–10M rows
Downloads this month
7.4K
Likes
87

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/Nemotron-Image-Training-v3")