747M Image-Text Pairs
A large corpus of alt-text and image pairs from HTML for training multimodal models.
- Text → image
- Image → text
- Zero shot classification
- English
Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.
Answered by Gemini 4 · the chat is kept only in this tab
A large corpus of alt-text and image pairs from HTML for training multimodal models.
English multimodal data with bounding box annotations for training visual reasoning models.
The first benchmark for evaluating recognition and parsing of real digital and photographed documents across 6 languages.
Question-answer pairs for MedPix medical images for VQA training and testing.
The first major open dataset for Ukrainian handwritten text recognition spanning over a century.
Millions of English and Chinese examples for training image editing and generation models.
Over 20 thousand safe-rated anime illustrations with detailed tags, emotions, and descriptions in three languages.
A challenging test set for checking visual understanding and visual question answering.
127K query-image pairs for training visual document retrieval.
Large Chinese dataset (1M to 10M) for search, recommendations, and RAG systems.
1.8M chart QA pairs with Chain-of-Thought annotations for VLM training.
Multi-image instruction datasets for training multimodal models.
Environment and execution logs for evaluating AI agents operating a computer alongside a human user.
English geospatial dataset for multi-modal analysis and Earth imagery segmentation.
English multimodal dataset for training models to describe specific regions in images and videos in detail.
Page images from PubMed Central with layout block annotations for document structure recognition.
55 hours of expert desktop app videos with manual annotations for training computer agents.
A classical dataset of individual handwritten text line images in English for OCR training.
Raw corpus of 2.3 million multimodal tasks and reasoning chains (8.8 billion tokens) before quality filtering.
NVIDIA dataset for training and evaluating the Llama-Nemotron VLM multimodal model.
Approximately 900K question-answer pairs (40% image-based, 60% text-based) for boosting multimodal model reasoning logic.
Over 1 million illustrations extracted from digitized historical books from the British Library (1510–1900).
A collection of images and short videos with questions from Scale AI to evaluate models' ability to understand hidden context.
Over 1,200 instructions and screenshots from iOS, Android, macOS, Windows, and the web for evaluating multimodal models.