Dataset · liuhaotian
CC3M images for LLaVA pre-training
liuhaotian/LLaVA-CC3M-Pretrain-595K
595K image-text pairs from CC3M for feature alignment when training multimodal models.
What's inside
This is a CC-3M subset of 595K image-caption pairs, filtered for more balanced concept coverage. Captions feature synthetic variants from BLIP for reference. Data is in English, intended for the pre-training stage and feature alignment in visual instruction tuning. Format: images with text captions.
What you can do with it
- Pre-training a LLaVA-like multimodal model
- Training the projector for vision and language alignment
- Checking quality of image-text pairs and BLIP captions
- Languages
- English
- Downloads this month
- 1.1K
- Likes
- 190
Models trained on it
- mradermacher/Nanbeige4.1-VLM-GGUFImage text → text 185
- rp-yu/Dimple-7BImage text → text 115
- qresearch/llama-3.1-8B-vision-378Image text → text 81
- qresearch/llama-3-vision-alpha-hfImage text → text 75
- qihoo360/Inner-Adaptor-ArchitectureImage text → text 63
- ytu-ce-cosmos/Turkish-LLaVA-v0.1Image text → text 61
- qihoo360/360VL-8BVisual question answering 59
- qihoo360/360VL-70BVisual question answering 45
How to load it
from datasets import load_dataset
ds = load_dataset("liuhaotian/LLaVA-CC3M-Pretrain-595K")