Skip to content

Dataset · liuhaotian

LLaVA pretrain: 558K image-text pairs

liuhaotian/LLaVA-Pretrain

A LAION/CC/SBU sub-sample with concept balance and synthetic BLIP captions for LLaVA pre-training.

What's inside

LCS-558K is a LAION/CC/SBU sub-sample filtered for more balanced concept coverage. Captions are attached to the images, including synthetic BLIP captions for reference. The set is designed for the pre-training stage and feature alignment during visual instruction tuning; the language is English.

What you can do with it

  • Pre-train a multimodal model for image and text alignment.
  • Train a projector or LoRA for a LLaVA-like model.
  • Test the visual instruction tuning pipeline.
Languages
English
Downloads this month
5.4K
Likes
237

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("liuhaotian/LLaVA-Pretrain")