Loading...
Loading...
Image-text pairs, instruction tuning, visual QA, cross-modal data, foundation model training data
1,928 datasets
AMALIA-VL-DPO Dataset is a collection of Direct Preference Optimization (DPO) data for the AMALIA-VL project. It contains preference triplets where each row includes a prompt with normalized role-content turns and image placeholders, a chosen assistant response, and a rejected response. The dataset was created by author 'amalia-llm' and was last updated on June 30, 2026.
YTClickbait21K is a human-annotated dataset of 21,238 YouTube videos for clickbait detection research. It includes video metadata like titles, descriptions, and thumbnails, along with binary clickbait labels from three annotators per video. The dataset was uploaded by Md. Minhazul Islam to figshare on April 9, 2026.
ACM packages the Audio-Centric Multimodal benchmark introduced by the paper 'Efficient and High-Fidelity Omni Modality Retrieval'. The dataset contains four HuggingFace subsets, each using a single test split to preserve natural schemas for queries and candidates. It was authored by chuonghm and last updated on July 4, 2026.
SciGenEdit-10K is a public subset of the S1-Omni-Image project, released by the ScienceOne team at the Chinese Academy of Sciences. It is designed for research on scientific image generation, editing, and multi-turn interactions. The dataset page was last updated on 2026-06-25.
ACL 2026 benchmark dataset of 1,000 multimodal social activity scenarios created by adonaivera. It is designed to evaluate generative AI agents' ability to detect and correct unsafe behavior during iterative plan revision. Each scenario contains a natural-language social activity description and an unsafe hourly plan spanning 11 activities from 7 PM to 5 AM.
A large-scale satellite image time series dataset curated for pre-training spatiotemporal foundation models for Earth observation. It contains approximately 1 million satellite images from 100,000 geographic locations, with each location observed across 10 temporal phases over five years. The dataset was introduced in the TiMo paper and is hosted by Hillui on Hugging Face.
Over 439,000 unique images and 581,000 image-text pairs were automatically extracted from more than 20,000 PDF documents in the REGIS collection. The dataset, created by Geologi, focuses on technical documents, theses, and reports from the Oil & Gas and Geosciences domain. It was last updated on 2026-06-21.
47 video appearances by Italian Prime Minister Giorgia Meloni are analyzed through a 74-variable multimodal coding scheme. The dataset, created by Canan Cetin and last updated in 2026, supports a study on communicative domestication, tracking shifts in themes, gestures, and presentation from opposition to office. Analysis reveals significant changes, such as motherhood references collapsing from 43% to 2% of appearances.
LapChole-FOCUS-VQA is a clinically grounded benchmark designed for evaluating long-context video understanding in minimally invasive surgery. The dataset is maintained by the ORena FOCUS Challenge team and was last updated on June 22, 2026. It is a gated resource, with access granted only to challenge participants after manual review.
SAW-Bench evaluates observer-centric situated awareness in multimodal foundation models. The benchmark probes a model's ability to reason about space, motion, and possible actions from an evolving egocentric viewpoint. It was created by ucsbai and last updated on Hugging Face in June 2026.
SpaRRTa-Lego is the real-world counterpart of a synthetic spatial-relation benchmark. Scenes are photographed with toy minifigures and everyday objects for evaluating Visual Foundation Models. The dataset is associated with a paper (arXiv:2601.11729) and code, and was last updated on 2026-06-25.
1.2 GB of power quality measurements recorded during laboratory experiments on high-impedance faults in medium-voltage covered conductors. The data includes RMS voltage and current, harmonics, phase angles, and power metrics, collected using a Hioki PQ3198 analyzer following IEC standards. Author Diogo Biasuz Dahlke published this dataset on figshare in May 2026.
Simultaneously recorded neuronal population activity from the S1 and PPC regions of awake mice. Data was acquired using two-photon calcium imaging during a texture discrimination task by Shuting Han at the University of Zurich. The temporal coverage and dataset size are not specified in the provided metadata.
A methodological framework and computational pipeline for evaluating symbolic and emotional responses to virtual architectural spaces. The framework integrates symbolic modeling of concepts like justice and identity with psychophysiological measures and AI-ready computational inference. It was authored by Jesus Rafael Hechavarria-Hernandez and published on figshare in May 2026.
32,884 records over 52,461 dental images formatted as image-text conversations for LLaVA-style instruction tuning. Curated by Harisundar R, this dataset serves as the training data for the PALL-VLM model. The dataset was last updated on June 12, 2026.
A 2D light-microscopy dataset for cell-nucleus segmentation assembled across many imaging experiments. The collection spans multiple modalities including fluorescence and brightfield histopathology, covering humans, mice, and flies across 30+ experiments. This is the official BBBC038v1 release from the Broad Bioimage Benchmark Collection, hosted on Hugging Face by MedOtter.
Jiang, Yanru provides experimental materials and anonymized participant data supporting the paper 'Informational Cues Mitigate Heuristic Bias in Evaluating Context-Sensitive AI Image Captions'. The data documents two human-subject experiments examining how informational context shapes audience evaluations of AI-generated image captions. The dataset was last updated on July 14, 2026.
HalluGuard-Prefs is a 76,708-entry synthetic preference dataset created by author lrsbrgrn for training models to detect hallucinations in text. It was used to fine-tune the HalluGuard-Qwen3-4B model via Odds Ratio Preference Optimization (ORPO). The dataset was introduced in a paper at the 64th Annual Meeting of the Association and was last updated on Hugging Face in June 2026.
Waveform data from a multimodal dataset of high-impedance faults in medium-voltage covered conductors contains high-resolution electrical waveform recordings from controlled laboratory experiments. Diogo Biasuz Dahlke collected the data using a Hioki MR8741 waveform recorder, sampling synchronized voltage and current at 20 kS/s. The dataset was last updated on 2026-05-05.
CC12M-Camera provides per-image camera parameter annotations for the Conceptual 12M dataset. The annotations cover approximately 10.97 million images and were predicted by the Puffin camera-centric multimodal model. The dataset was created by KangLiao and was last updated on June 24, 2026.