Dataset · HuggingFaceM4
Large corpus for training open VLMs
HuggingFaceM4/FineVision
17.3M images, 24.3M examples, and 88.9M turns for training modern open vision-language models.
What's inside
FineVision is a massive collection of image-text data containing 17.3M images, 24.3M examples, 88.9M dialogue turns, and 9.5B answer tokens. Built to train state-of-the-art open vision-language models. Loaded via the datasets library with subset selection; further details are available on the HuggingFaceM4 blog.
What you can do with it
- Pre-training and fine-tuning open VLMs
- Training multimodal assistants
- Evaluating models on image-based dialogues
- Size
- 10M–100M rows
- Downloads this month
- 102.4K
- Likes
- 544
Models trained on it
- kyutai/CASA-Qwen2_5-VL-3BImage text → text 12.4K
- lmms-lab/LLaVA-OneVision-1.5-8B-InstructImage text → text 2.6K
- lmms-lab/LLaVA-OneVision-1.5-4B-InstructImage text → text 788
- llm-jp/llm-jp-4-vl-9b-betaEmbeddings 298
- athrael-soju/UltraVBERTVisual document retrieval 281
- nc-ai-consortium/VAETKI-VL-7B-A1BImage text → text 262
- kyutai/CASA-Helium1-VL-2BImage text → text 205
- tokimoa/llm-jp-4-vl-9b-beta-mlx-bf16Image text → text 69
How to load it
from datasets import load_dataset
ds = load_dataset("HuggingFaceM4/FineVision")