README.md

metadata

license: apache-2.0
task_categories:
  - image-to-text
  - visual-question-answering
tags:
  - llava
  - vision-language
  - pretrain
  - multimodal
size_categories:
  - 100K<n<1M

LLaVA-Pretrain Dataset

Pretraining data for LLaVA (Large Language and Vision Assistant).

Description

This dataset contains the pretraining data used in LLaVA training, including:

• blip_laion_cc_sbu_558k.json - Annotation file with 558K image-caption pairs
• images/ - Corresponding images

Usage

from huggingface_hub import snapshot_download

# Download the dataset
snapshot_download(
    repo_id="pppop7/LLaVA-Pretrain",
    repo_type="dataset",
    local_dir="./llava_pretrain"
)

Related Datasets

• pppop7/LLaVA-Instruct-150K - Instruction tuning data

Reference

• LLaVA Official Repository
• LLaVA Paper