Skip to content

Dataset · HuggingFaceH4

LLaVA multimodal instructions for VSFT

HuggingFaceH4/llava-instruct-mix-vsft

Formatted corpus with images and text for visual SFT fine-tuning via TRL.

What's inside

It contains image and text instruction pairs with a volume of 100 thousand to 1 million examples. This is a reformatted version of the llava-instruct-mix dataset prepared by the HuggingFaceH4 team. The data is adapted specifically for visual SFT (VSFT) using the SFT Trainer from the TRL library.

What you can do with it

  • SFT fine-tuning of LLaVA-based vision-language models (VLMs)
  • Testing multimodal training pipelines in the TRL library
Size
100K–1M rows
Downloads this month
1.4K
Likes
50

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HuggingFaceH4/llava-instruct-mix-vsft")