FineVision
a massive dataset for training multimodal models
2 min read
FineVision is a massive data collection built to train frontier open Vision-Language Models (VLMs, multimodal models that understand both images and text). The dataset packs 17.3 million images, 24.3 million samples, 88.9 million conversation turns, and 9.5 billion answer tokens. Inside, you'll find diverse structured data complete with relevance and visual-dependency ratings. It's an ideal resource for researchers and developers training or benchmarking multimodal AI.
Key features
- Massive scale: 17.3M images and 9.5B tokens for deep training.
- Multi-turn conversations: 88.9M text turns covering detailed descriptions and visual reasoning questions.
- High-quality annotations: includes image-to-text correspondence scores (image_correspondence_ratings).
- Open license: prompt materials are released under CC-BY-4.0.
What you can do
- Train open Vision-Language Models from scratch or fine-tune them for advanced visual analysis.
- Benchmark the multimodal capabilities of language models across complex scientific and everyday datasets.
- Study how high-quality visual datasets impact AI response generation.
Example
Load the dataset for training and analysis using the datasets library:
from datasets import load_dataset, get_dataset_config_names
# Получаем список подмножеств и загружаем первое
available_subsets = get_dataset_config_names('HuggingFaceM4/FineVision')
ds = load_dataset('HuggingFaceM4/FineVision', name=available_subsets[0], split='train', streaming=True)
# Проверяем первый пример
print(ds[0])Read more
DetailsSize: 17.3M images, 24.3M samples, 88.9M turns, 9.5B tokens
Authors: Luis Wiedmann, Orr Zohar, Amir Mahla, and others (HuggingFaceM4)
License: CC-BY-4.0 for prompts; source sub-datasets are governed by their own terms
Paper: arXiv:2510.17269








Comments