Skip to content

Dataset · laion

220k LVIS images with GPT-4V and Mistral captions

laion/220k-GPT4Vision-captions-from-LIVIS

LVIS images with detailed text descriptions adapted from GPT-4Vision answers for multimodal training.

What's inside

Contains 220,000 image-text pairs based on the well-known LVIS dataset. The original instruction data from LVIS-Instruct4V (generated using GPT-4Vision) was translated by LAION authors into standard description formats via the Mistral-7B-OpenOrca model. The dataset was published in November 2023 and contains between 100 thousand and 1 million records.

What you can do with it

  • Training and fine-tuning multimodal models (VLM) for image detail understanding
  • Training image generators on detailed text prompts
  • Improving image captioning quality
Size
100K–1M rows
Downloads this month
315
Likes
64

How to load it

from datasets import load_dataset

ds = load_dataset("laion/220k-GPT4Vision-captions-from-LIVIS")