Loading...
Loading...
Image-text pairs, instruction tuning, visual QA, cross-modal data, foundation model training data
1,956 datasets
UniWorld V1 provides between 1,000 and 10,000 image-text pairs sourced from the BLIP3o-60k collection, released by LanguageBind in June 2025. It utilizes Geneval-style annotations to facilitate the training of high-resolution semantic encoders for unified visual understanding and generation.
FaceCaptionHQ-4M is a dataset containing approximately 4 million facial image-text pairs. It was created by OpenFace-CQUPT and was last updated on 2025-06-09. The dataset is a cleaned subset derived from the larger FaceCaption-15M dataset.
A dataset of image captions depicting unsafe or illegal activities, hosted on HuggingFace. The dataset was created by Lenkashell and was last updated on July 16, 2025. The specific content, scale, and structure of the data are not detailed in the available metadata.
A dataset of image captions related to illegal activities, created by Lenkashell and last updated on July 16, 2025. The dataset is hosted on HuggingFace, but its specific content, size, and structure are not detailed. Its intended purpose appears to be for training or evaluating content safety models.
SLAKE contains 642 medical image samples with multi-task annotations, curated by Voxel51 based on research published in 2021 (Arxiv 2102.09542). It provides a specialized dataset for medical visual question answering and computer vision, featuring labels for classification, detection, and segmentation tasks.
ReEdit-Bench is a curated dataset of approximately 1,500 samples for evaluating exemplar-based image editing methods. It was created by tarun-menta and presented in a WACV '25 paper. Each sample contains four images representing an exemplar edit pair.
Examples of harmful and harmless language. It aggregates samples from seven source datasets, including Anthropic/hh-rlhf and allenai/real-toxicity-prompts. The data is available in both Portuguese and English.
400 full-length high-definition talking face videos, split into 81-frame clips and paired with audio embeddings. The dataset was curated by global-optima-research and last updated on June 4, 2025. It is intended for tasks in talking-head generation and multimodal avatar synthesis.
Open-Orca's SlimOrca Dedup is a dataset of 363,000 unique instruction-response examples derived from the SlimOrca collection. It was created by removing RLHF instances and applying minhash and Jaccard similarity techniques for deduplication. The dataset was last updated on Hugging Face on May 19, 2025.
ARGUS is a framework for calculating hallucination and omission costs in free-form video captions. The dataset, created by tomg-group-umd, provides metrics to quantify the degree of hallucinated and omitted content in video-language model outputs. It was last updated on June 10,我们发现了一个问题，您提供的原始描述中包含了中文文本。根据指令，我需要将输入翻译成英文。以下是翻译后的描述，并基于此生成输出。
OlympiadBench contains between 1,000 and 10,000 bilingual scientific problems in mathematics and physics, designed for evaluating AGI reasoning. Created by Hothan and published at ACL 2024, the dataset includes both text-based and multimodal questions in English and Chinese.
A dataset of short, natural Vietnamese dialogues for fine-tuning language models like Mamba, LLaMA, and Gemma. It contains everyday communication, frequently asked questions, and emotional responses, formatted as JSONL and ready for instruction tuning. The dataset was created by hoanghai2110 for the Vietnamese open-source AI community.
PRIMUS is a pioneering collection of open-source datasets for cybersecurity LLM training. The Primus-Reasoning subset contains multiple cybersecurity reasoning tasks sourced from CTI-Bench, including CTI-RCM, CTI-VSP, CTI-ATE, and CTI-MCQ. It was augmented in June 2025 with distilled samples from DeepSeek-R1, incorporating intermediate reasoning steps and final answers.
667 real-world puzzlehunt-style problems curated from Puzzled Pint's Creative Commons archives between 2010 and 2025. The dataset, created by author hzli1202 and last updated on June 10, 2025, is designed as a benchmark to evaluate open-ended, multimodal reasoning in AI models.
Skywork-OR1-RL-Data is a reinforcement learning training dataset containing between 100,000 and 1,000,000 text records released by Skywork in April 2025. The collection features problems categorized by difficulty levels ranging from 0 to 16, calibrated against specific DeepSeek-R1-Distill-Qwen model variants.
PLM-Video-Human is a collection of human-annotated resources for training Vision Language Models, focused on detailed video understanding. The dataset, created by Facebook, was last updated on May 21, -2025. Training tasks include fine-grained open-ended question answering, region-based video captioning, dense captioning, and temporal localization.
MedBookVQA is a multimodal benchmark built from open-access medical textbooks to evaluate general medical AI (GMAI) and multimodal large language models (MLLMs). The dataset was created by slyipae1 and last updated on June 10, 2025. It aims to address the underutilization of structured textbook knowledge for systematic AI evaluation.
PLM-VideoBench is a collection of human-annotated resources for evaluating Vision Language models, focused on detailed video understanding. The dataset includes evaluation data for tasks like FGQA, which probes fine-grained activity understanding through multiple-choice questions. It was authored by Facebook and last updated on May 21, 2025.
Synthetic Visual Genome (SVG) datasets are designed for training Vision-Language Models on scene graph understanding and dense visual relationships. The datasets were created by author jamepark3922 and were last updated on June 11, 2025. They are hosted on the Hugging Face platform.
MemoryBench provides benchmark tasks across spatial memory and action recall categories for robotic manipulation. It serves as the evaluation foundation for the SAM2Act+ framework, focusing on the integration of visual foundation models with memory architectures.