Dataset · lmms-lab
Training data for the LLaVA-NeXT multimodal model
lmms-lab/LLaVA-NeXT-Data
A mix of images and instructions for SFT training of multimodal models.
What's inside
Contains images and texts used during the instruction tuning stage for LLaVA-NeXT models. The dataset size ranges from 100k to 1 million records. The data is updated in raw format: the archive unpacks into a JSON file and a structured folder with images.
What you can do with it
- SFT fine-tuning of multimodal LLMs for image-based QA
- Training visual analysis and instruction-following systems
- Comparing quality across data mixtures during VLM training
- Size
- 100K–1M rows
- Downloads this month
- 9.5K
- Likes
- 48
Models trained on it
- DeepGlint-AI/llava-mlcd-qwen2.5-7b 130
- rp-yu/Dimple-7BImage text → text 115
- ColorfulAI/OpenOmni-8B-Llama3-Omni 66
- lmms-lab/llama3-llava-next-8b-hf-sae-131k 48
- DeepGlint-AI/ViCToR-LLaVA-SigLIP2-Qwen2.5-7bImage text → text 21
- Inst-IT/LLaVA-Next-Inst-It-Vicuna-7BVideo text → text 20
- junha1125/gaze-attention-qwen3-4b-384-onevisionImage text → text 17
- Inst-IT/LLaVA-Next-Inst-It-Qwen2-7BVideo text → text 13
How to load it
from datasets import load_dataset
ds = load_dataset("lmms-lab/LLaVA-NeXT-Data")