Loading...
Loading...
Image-text pairs, instruction tuning, visual QA, cross-modal data, foundation model training data
1,956 datasets
i-CIR is a benchmark for instance-level composed image retrieval containing between 100,000 and 1,000,000 records, released by billpsomas in 2024. It facilitates the retrieval of specific, visually indistinguishable objects by combining a reference image with a text-based modification query. The dataset includes a specialized database of visual, textual, and compositional hard negatives to test model precision.
A processed and enhanced version of the H&M Personalized Fashion Recommendations Kaggle competition dataset. The dataset has been cleaned and augmented with pre-computed embeddings and accessible image URLs by Qdrant, last updated in December 2025.
Simuletic's Surveillance VLM Weapon Knife Detection Dataset is an open-source subset of the Simuletic Safety VLM Dataset. It is designed for instruction tuning of Vision Language Models to locate weapons and knives, reason about threats, and avoid false positives. The dataset was last updated on December 17, 2025.
TaiwanVQA is a visual question answering benchmark containing 2,736 original images paired with 5,472 manually designed questions. It is designed to evaluate the capability of vision-language models in recognizing and reasoning about culturally specific content related to Taiwan. The dataset was created by author hhhuang and last updated on December 4, 2025.
AgentVQA is a multi-domain dataset for training and evaluating visual agents. The dataset contains multiple-choice questions based on screenshots, images, and videos across five domains, including GUI interaction and robot manipulation. It was created by AgentVQA and last updated on December 18, 2025.
Safety Multimodal Jailbreaking is a dataset hosted on HuggingFace by author leeeliu. The dataset was last updated on 2026-01-28, suggesting ongoing maintenance. Its title indicates it likely contains examples for testing or bypassing safety measures in multimodal AI systems.
Over 400,000 human preference responses for evaluating the Flux 2 Pro text-to-image model, collected in less than seven hours via the Rapidata Python API. The dataset was created by Rapidata and last updated on December 2, 2025. It includes evaluations across preference, coherence, and alignment categories.
Mercity's LAION Subset is a curated collection of 24,840 high-quality image-caption pairs from the LAION-5B dataset, formatted at 512x512 resolution. It is designed for training Latent Consistency Model (LCM) LoRA adapters on Stable Diffusion v1.5. The dataset, last updated in November 2025, occupies approximately 4.16 GB and is stored in Parquet files.
Multimodal Physiological Stress Dataset is a collection of dynamic stress data from college students, published on Kaggle. The dataset likely contains time-series physiological measurements, though specific columns and sample sizes are not detailed in the provided metadata. Its raw description indicates a focus on student stress levels, but the exact collection methodology and temporal coverage are unknown.
Vision-Language Models are Confused Tourists evaluates the cultural robustness of VLMs, a largely untested dimension crucial for supporting diverse societies. The dataset was created by author patrickamadeus and was last updated in December 2025. It contains image-text pairs designed to test model stability across diverse cultural inputs.
Image caption data likely contains pairs of images and descriptive text. The dataset is hosted on Kaggle, a platform for data science competitions and projects. Specific details on volume, creation method, and update recency are not provided in the metadata.
Chart VQA likely contains images of charts and graphs paired with natural language questions and answers. The dataset is hosted on Kaggle, a platform for data science competitions and projects. Specific details on volume, creation date, and authorship are not provided in the available metadata.
A dataset from Kaggle focusing on college students' career preferences. The raw description suggests it includes psychological and IoT behavioral indicators. The specific scale, collection method, and temporal coverage are not detailed in the provided metadata.
Lidar_multimodal likely contains data from Light Detection and Ranging (LiDAR) sensors combined with other modalities. The dataset is hosted on Kaggle, but its specific content, size, and creation details are not provided. Columns and sample data are unknown.
Kaggle hosts a dataset titled 'Multimodal_skin_lesion'. The dataset likely contains data related to skin lesions, possibly including images and other data types. The author, organization, and specific details are unknown.
Multimodal-skin-lesion-project-code is a dataset published on Kaggle. The dataset likely contains code and data related to skin lesion analysis. Its specific content, size, and authorship require verification after download.
Viet-Chart-VQA-images is a dataset hosted on Kaggle. The title suggests it contains images paired with questions and answers, likely for training or evaluating Visual Question Answering models. The dataset's content, scale, and provenance require verification after download.
A dataset for aligning large language models with human preferences. The dataset is hosted on Kaggle, but its specific size, authorship, and creation date are not provided in the metadata. The content likely contains pairs of instructions and responses with preference rankings.
Pre-computed baseline exposures for the GPT-Neo-1.3B language model. The dataset is hosted on Kaggle and appears to contain metrics related to privacy or model behavior. The specific data format, size, and creation details are not provided in the metadata.
AUTOPILOT VQA Heatmaps likely contains 661 files related to Visual Question Answering, a task combining computer vision and natural language processing. The dataset appears to focus on heatmap visualizations, which are often used to interpret model attention. It is published on Kaggle, but the author, creation date, and specific content details are not provided in the metadata.