MVBench benchmark for video understanding by multimodal AIs
A set of 4,000 video questions across 20 tasks to test time and action perception by multimodal models.
- Visual question answering
- Video 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 set of 4,000 video questions across 20 tasks to test time and action perception by multimodal models.
Over 26,000 prompts in English and Chinese with resulting images and metadata.
Images and 3D data of animals with shape, texture, and fur maps for image-to-3D tasks.
Diagnostic generated image set for evaluating diffusion model classification.
From 10k to 100k examples from English scientific articles with ready-made chunks and FAISS indices for verifying citation accuracy.
About 25,000 generated passport-style images for training and validating classifiers.
A multilingual collection of Wikimedia vector diagrams and maps with text coordinate annotations for evaluating image translation.
A large open collection of English image-text pairs for high-quality text-to-image fine-tuning.
An image collection of character outfits from 10 anime games for classification and computer vision.
Vietnamese e-commerce data with product photos, descriptions, user reviews, and review images.
A collection of tens of thousands of real and generated images for training and evaluating AI image detectors.
The largest biology ML dataset: over 10 million photos, texts, and documents for 454,000 taxa of the tree of life.
Formatted version of the GQA visual-text dataset with over 10 million examples for the lmms-eval benchmark.
A multiple-choice VQA test for evaluating plant pathology understanding in VLMs.
Large Conceptual Captions 12M dataset in WebDataset format for pre-training multimodal models.
English video QA dataset for evaluating multimodal neural network video understanding in real time.
1,000 to 10,000 images for training and validating few-shot object counting models across any class.
A large expert VQA dataset for analyzing the evolution, etymology, and structure of ancient Chinese characters.
108,808 sets of source photos, references, and results for training automated color correction algorithms.
Images from 562 European league football matches for object detection and tracking.
Manga illustrations and English/Japanese texts for classification and VQA.
A re-upload of the classic dataset of 2,000 sports photographs from Flickr for pose estimation tasks.
31k hours of image-prompted spontaneous speech across 105 Indian languages, with transcripts for ASR and multimodal tasks.
Dataset with original and logically manipulated images of cats and dogs for evaluating Multimodal LLMs.