Loading...
Loading...
Image-text pairs, instruction tuning, visual QA, cross-modal data, foundation model training data
1,936 datasets
15,282 behavioral annotations of LLM and VLM reasoning traces were collected by neulab. The dataset covers responses from 15 models across 6 benchmarks, with each row containing correctness and a JSON-encoded behavioral annotation. It was last updated on 2026-05-08.
7,000 production records combine sensor, quality, machine, and edge metrics. The dataset is hosted on Kaggle and focuses on industrial IoT and edge computing applications. Its author, organization, and specific collection details are not provided.
MCSBench v1.0 is a diagnostic benchmark for evaluating multimodal large language models. It contains base visual question answering records, reasoning-chain selection records, evidence fields, and image references. The dataset was created by mcsbench and last updated on May 7, 2026.
Search-VL-RL-8K is an open recipe for training frontier multimodal search agents, authored by OpenSearch-VL. The dataset was last updated on May 7, 2026. It likely contains data for training agents using methods like Cold-Start Agentic SFT and Multi-Turn Fatal-Aware GRPO.
KIT-MRT provides a preview sample of the KITScenes Multimodal dataset. The sample contains one representative sequence intended for data format inspection. The preview was last updated on May 6, 2026.
A 16.1 KB review document authored by Liucheng Li, last updated in March 2026. It synthesizes advances in artificial intelligence for gastrointestinal medicine, covering multimodal imaging, digital biomarkers, and real-time monitoring platforms. The document discusses applications in functional GI disorders, inflammatory bowel disease, and GI oncology.
A 15.3 KB DOCX file authored by Liucheng Li, summarizing a review on artificial intelligence applications in gastrointestinal functional assessment. The document synthesizes advances in multimodal imaging, digital biomarkers, and real-time monitoring for GI disorders. It was last updated on March 25, 2026.
WikiVQABench is a human-curated benchmark for knowledge-grounded visual question answering. IBM Research constructed it by systematically combining Wikipedia images, article captions, and structured knowledge from Wikidata. Candidate multiple-choice questions were generated by large language models and then reviewed by human annotators for factual correctness and visual-text consistency.
Search-VL-SFT-36K is a dataset for supervised fine-tuning of frontier multimodal search agents, created by OpenSearch-VL. The dataset was last updated on May 7, 2026. It likely contains data for training agents on multi-turn, fatal-aware tasks with visual tool use.
Blip3O Pretrain Long Caption Parquet is a dataset hosted on HuggingFace by the author LastTransformer. The title suggests it contains data for pretraining vision-language models, likely pairing images with detailed textual descriptions. The dataset was last updated on June 20, 2026.
VAST-AI provides a compact 10-example subset of the AniGen training dataset for generative AI. This sample includes unique raw assets and full cross-modal files across multiple directories like raw, renders, skeleton, and voxels. The dataset was last updated on April 13, 2026.
Wafer VQA Dataset is a multimodal benchmark built on the MixedWM38 wafer-map collection. It provides annotations for wafer map understanding, defect reasoning, and visual question answering. The dataset is organized into two annotation styles: tuple_generation for sequence-level optimization and stepwise_reasoning for supervised fine-tuning.
DocVQA Media Judged is a dataset for document visual question answering, likely containing images of documents paired with questions and answers. It was published by the author merve on the Hugging Face platform and was last updated on June 11, 2026. The dataset's specific scale and content require verification after download.
MolDeTox is a benchmark dataset designed to evaluate toxicity-aware molecular editing capabilities of LLMs and VLMs. It is constructed based on the concept of toxicity cliffs, where structurally similar molecules exhibit opposite toxicity labels. The dataset was created by the MolDeTox organization and was last updated on May 5, 2026.
221.9 MB of multimodal data from Sungwon Jung's 2026 study of emotional contagion in a YouTube live chat during a major political event. The collection includes CSV and JSONL files alongside analysis code in IPYNB and RMD formats. It was published under a CC-BY-4.0 license on figshare.
COCO-ARVQA is an Arabic Visual Question Answering dataset built over images from the MS COCO 2017 train2017 archive. It provides Arabic questions, answers, answer lists, and identifiers linking to COCO images, created by author MouaffakAyoub and last updated on 2026-04-27. The dataset does not redistribute the COCO images themselves, requiring users to obtain the official image archive separately.
VULCA-Bench is a bilingual multicultural art-critique corpus containing 7,236 multimodal samples, with 7,234 including embedded images. It covers eight cultural traditions and uses a schema with 236 cultural dimensions. The dataset was created by author harryHURRY and last updated on April 30, -2026.
NVIDIA's Nemotron Image Training v3 is a collection of image-centric multimodal training data. It is a large-scale, multi-subdataset release where each subset includes standardized conversation JSONL files and a dataset card describing sources, licensing, and media layout. The dataset was last updated on 2026-04-28.
58,320 structured JSON records from a study of image embedded prompt injection vulnerability and defense effectiveness across four vision-language models applied to dental panoramic radiography. The dataset includes 9,720 baseline calls and 48,600 defense calls, with pre-computed analysis tables. It was authored by Babak Saravi and last updated on April 10, 2026.
A self-distilled instruction-following dataset created by HarryMayne. It contains data elicited from four models—Qwen3.5-35B-A3B, Qwen3.5 397B-A17B, GPT-4.1, and Kimi K2.5—using prompts from the Dolma 3 corpus at temperature 1. The dataset was last updated on May 14, 2026.