datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
OpenMath-Vision-CoT-10kVision-OPD-6K
Vision-OPD-6K: Training Data for Vision-OPD
Overview
Vision-OPD proposes a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy, without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use.
Vision-OPD instantiates two conditional policies from the same MLLM:
A crop-conditioned teacher that observes the evidence-centered crop as a privileged… See the full description on the dataset page: https://huggingface.co/datasets/yuanqianhao/Vision-OPD-6K.vstat
VSTAT: Visual State Tracking Benchmark
VSTAT is a video-based benchmark for evaluating the visual state tracking
capability of Multimodal Large Language Models (MLLMs). It contains 834 video
clips paired with 1,500 questions whose answers cannot be inferred from any
single keyframe or short segment.
Dataset Composition
Split
Videos
Questions
synthetic
450
550
self_recorded
80
100
youtube
304
850
Total
834
1,500
Files… See the full description on the dataset page: https://huggingface.co/datasets/nyu-visionx/vstat.Vision-OPD-6K
Vision-OPD-6K: Training Data for Vision-OPD
Training data for Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation.
Overview
Vision-OPD proposes a regional-to-global self-distillation framework that transfers the model's own privileged regional perception to its full-image policy, without external teacher models, ground-truth labels, reward verifiers, or inference-time tool use.
Vision-OPD instantiates two conditional policies from… See the full description on the dataset page: https://huggingface.co/datasets/zwyang6/Vision-OPD-6K.TVQA-Long
Dataset Sources
Repository: https://github.com/Vision-CAIR/MiniGPT4-video
Paper: https://arxiv.org/abs/2407.12679
BibTeX:
@misc{ataallah2024goldfishvisionlanguageunderstandingarbitrarily,
title={Goldfish: Vision-Language Understanding of Arbitrarily Long Videos},
author={Kirolos Ataallah and Xiaoqian Shen and Eslam Abdelrahman and Essam Sleiman and Mingchen Zhuge and Jian Ding and Deyao Zhu and Jürgen Schmidhuber and Mohamed Elhoseiny},
year={2024}… See the full description on the dataset page: https://huggingface.co/datasets/Vision-CAIR/TVQA-Long.calame-pt
CALAME-PT
Context-Aware LAnguage Modeling Evaluation for Portuguese
CALAME-PT is a PT benchmark composed of small texts (contexts) and their respective last words.
These contexts should, in theory, contain enough information so that a human or a model is capable of guessing its last word - without being too specific and/or too ambiguous.
Composition
CALAME-PT is composed of 2 "sets" of data - handwritten and generated.
Handwritten Set: contains 406… See the full description on the dataset page: https://huggingface.co/datasets/NOVA-vision-language/calame-pt.SenseNova-Vision-Corpus-50M-annotationagentic-llm-pretraining-1.7b
Agentic LLM Pretraining Dataset
A pretraining corpus for small language models (1-3B parameters) optimized for agentic tasks. The corpus emphasizes learning to comprehend language, reason, follow instructions, and use tools over memorizing factual knowledge — the assumption is that domain knowledge will be provided at runtime via RAG. The idea is that this could enable much smaller pretraining corpora by omitting the large volumes of text typically needed to memorize facts.… See the full description on the dataset page: https://huggingface.co/datasets/visionscaper/agentic-llm-pretraining-1.7b.Vision_GUI_Assistant
[EMNLP2024] VGA: Vision GUI Assistant - Minimizing Hallucinations through Image-Centric Fine-Tuning
Release
We release our dataset to ensure that everyone can replicate our experimental conclusions.
Directory Description
|-- dataset generate / method(prompts) to generate data
--|-- dataset / data resource
|-- llava training / training code
|-- tuning script / tuing parameters
Setup
Dataset Format
Our dataset follow… See the full description on the dataset page: https://huggingface.co/datasets/zylate/Vision_GUI_Assistant.vision-token-compression-bench
OPTIC-Bench
Optical Text In-Context Benchmark: how reliably do LLMs consume text
delivered as rendered images versus plain text tokens?
In summary, the evaluation reported here finds that optical text compression
is effective only within a narrow and specific envelope. Delivering content
as rendered images genuinely reduces input tokens, by thirteen to
fifty-four per cent depending on the model and the language, but only when
the document is long, the rendering is dense and the… See the full description on the dataset page: https://huggingface.co/datasets/translorentz/vision-token-compression-bench.Chinese-LLaVA-Vision-Instructions本数据集是对于LLaVA的翻译，请从LLaVA dataset下载对应的图片。
百度网盘链接: https://pan.baidu.com/s/1-jgINIkW0MxusmJuSif85w?pwd=q62v
Vision-COTvision-braille-dataset
vision-braille-dataset
Chinese ↔ 通用盲文 (Chinese Braille) parallel training data for the Vision-Braille translation
models. Three corpora ship together in this repo:
Folder
What it is
Rows
Size
cleaned_2345_v3/
Audited + cleaned braille→Chinese training corpus, stages 2→5, with restored 分词连写 word spacing on stage 2
379,152
426 MB
braille_spaced/
Freshly built passage-level corpus, fully word-spaced, across 通用 / 数学 / 医学 / 中医 / 病理学
80,360
208 MB
cleaned_080126_v1/… See the full description on the dataset page: https://huggingface.co/datasets/Violet-yo/vision-braille-dataset.newtphys_gsoPreprocessed Google Scanned Object dataset for NewtPhys simulator. Each object is optimized as 3DGS and its physical properties are estimated with a VLM.
(documentation will be added)
vision-adapter-manifests
vision-adapter-manifests
The 45% agentic / 45% reasoning-doc / 10% conversational SFT mix (114,024 train
rows + 2,328 held-out validation rows) used to train the Vision-Adapter project.
Includes the full cauldron pull from which the mix was sampled.
The image corpus is a separate HF dataset repo
(keypa/vision-adapter-images).
Contents
train_manifest.jsonl — the actual train mixture (45% agentic / 45% doc / 10% conversational). Every row:
{emb:… See the full description on the dataset page: https://huggingface.co/datasets/keypa/vision-adapter-manifests.muse-k2-vision-pilot-20260910
Muse → K2 bridge: first training experiment
Prepared September 10, 2026. This experiment tests whether training a connector
lets the frozen IFM/K2-Horizon-7B decoder use the existing Muse-Glimmer visual
encoder. It does not retrain the vision encoder or K2, and it does not establish
general screenshot, document, natural-image, or visual reasoning capability.
Authorized budget and selected first hardware
The user authorized an initial inexpensive Hugging Face… See the full description on the dataset page: https://huggingface.co/datasets/txgsync/muse-k2-vision-pilot-20260910.Vero-35k-VisionSelector
Vero-35k-VisionSelector (Multi-Domain Balanced Visual Token Pruning Dataset)
This dataset is a balanced ~35.4K multi-domain dataset derived from zlab-princeton/Vero-600k, specifically designed for Vision Token Pruning Pre-training (e.g. VisionSelector LIS) and Decoupled Downstream Task Fine-tuning.
📊 Dataset Overview
Total Samples: 35,400 multi-modal samples
Domain Coverage: 6 core domains across 59 distinct tasks (600 samples per task)
Image Preservation: 100%… See the full description on the dataset page: https://huggingface.co/datasets/muyuho/Vero-35k-VisionSelector.lm-eval-results-Nitral-AI-Eris_PrimeV3.05-Vision-7B-private
Dataset Card for Evaluation run of Nitral-AI/Eris_PrimeV3.05-Vision-7B
Dataset automatically created during the evaluation run of model Nitral-AI/Eris_PrimeV3.05-Vision-7B
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 4 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-Nitral-AI-Eris_PrimeV3.05-Vision-7B-private.MSCOCO_PT-BRtvta-vision
TradingView Technical Analysis Vision Dataset (TVTA-Vision)
A research-grade multimodal dataset for training and evaluating Vision-Language Models on TradingView-style financial chart understanding and structured technical analysis.
Dataset Overview
The TVTA-Vision dataset provides:
Chart images: TradingView-style candlestick charts with realistic market conditions
Structured annotations: Technical indicators, market structure, patterns, support/resistance, trend… See the full description on the dataset page: https://huggingface.co/datasets/ismailtasdelen/tvta-vision.Vision-DeepResearch-Text-Dataunified-math-vision-dataset
Unified Math Vision Dataset Bundle
Generated at: 2025-09-19 17:12:44
This is a unified dataset bundle containing multiple math and vision reasoning datasets.
Dataset Statistics
Total samples: 15858
mathvision: 3344 samples
wemath: 500 samples
mmmu: 415 samples
mathvista: 6141 samples
logicvista: 448 samples
dynamath: 5010 samples
Contents
manifest.jsonl: Complete dataset in JSONL format (1 JSON per line)
manifest.csv: Summary in CSV format
images/: Directory… See the full description on the dataset page: https://huggingface.co/datasets/Haonian/unified-math-vision-dataset.ghana-ag-vision
Ghana Agricultural Vision Catalogue
Part of the Ghana Agricultural Data Commons v0.2.0 public dataset release by Sinuosity Physical Intelligence Lab. Repository payloads include redistributed source data only where licensing and publication gates allow; other entries are registry, provenance, governance or builder metadata.
This dataset is part of the Ghana Agricultural Data Commons from Sinuosity Physical Intelligence Lab.
GhanaAgData is a systematic, publicly documented… See the full description on the dataset page: https://huggingface.co/datasets/sinuosity/ghana-ag-vision.glm-vision-sft-mix
glm-vision-sft-mix
Merged SFT mix used to train fable-glm-vision — a 44.8M vision projector that gives GLM-5.2 image understanding with a frozen backbone, following the recipe by Harry Partridge (original post).
151,439 train / 5,354 val examples. Each row: {"image": <filename>, "conversations": [user, assistant], "source": <subset>}. The user turn contains a literal image-span placeholder followed by the instruction; answers are deliberately SHORT (single actions / single… See the full description on the dataset page: https://huggingface.co/datasets/0xSero/glm-vision-sft-mix.Voluspa_the_Seeresss_Vision_the_Ultimate_Poetic_Rendering
Völuspá the Seeress's Vision the Ultimate Poetic Rendering
Dataset Details
Dataset Description
This dataset weaves the ancient threads of Völuspá, the Seeress's profound vision from the Poetic Edda, into a modern ShareGPT JSONL format. It holds 66 sacred exchanges, each a rune-carved conversation: a mortal seeker requests the recitation of a stanza, and the divine voice responds with the poem's eternal words—from creation's dawn through Ragnarök's… See the full description on the dataset page: https://huggingface.co/datasets/RuneForgeAI/Voluspa_the_Seeresss_Vision_the_Ultimate_Poetic_Rendering.VisionRewardDB-Image-regression
Regression Dataset Quick Start
All image pair info is in regression.jsonl. Use the tag to identify the source.
File Structure
/
├── regression.jsonl # Metadata file with image pairs and source tags
├── regression_IMDB.tar.gz # Compressed file containing images from the ImageRewardDB source
└── README.md # This usage guide
The images associated with the 'ImageRewardDB' tag can be obtained by decompressing the regression_IMDB.tar.gz file. For… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/VisionRewardDB-Image-regression.tad66k-vision-de
TAD66K Vision DE
Deutschsprachiger Vision-Language Datensatz auf Basis von TAD66K.
Strukturierte Bildbeschreibungen als kompaktes Schema-Format,
mit tiefen atmosphärischen Dialogen auf Deutsch.
Primär entwickelt als Trainingsdatensatz für
Kassandra Vision LoRA.
Konzept
Statt Bilder direkt als Text zu beschreiben wird ein kompaktes
maschinenlesbares Schema verwendet (~80 Tokens statt ~180 Fließtext):
F:beach_scene
G:people#8b4513@fg|sand#d2b48c@mg|water#4682b4@bg… See the full description on the dataset page: https://huggingface.co/datasets/sebelsn/tad66k-vision-de.zamai-pashto-vision
ZamAI Pashto Vision
Languages: psLicense: cc-by-4.0Task categories: image-to-text, image-classificationSize categories: 1K<n<10K
Summary
This dataset is part of the ZamAI Pashto data collection. It is intended for image-to-text, image-classification tasks in Pashto.
How to use
from datasets import load_dataset
dataset = load_dataset("tasal9/zamai-pashto-vision")
print(dataset)
Configs
pashto_captions: load with… See the full description on the dataset page: https://huggingface.co/datasets/tasal9/zamai-pashto-vision.LTAC-1400_vision.json🧬 LTAC 1400_vision.json
📝 Opis Projektu
Ten dataset/moduł stanowi fundament Vision Organ (1400) w ramach architektury inteligencji emergentnej S.A.R.A / ALT. Moduł ten odpowiada za multimodalną percepcję, integrację danych wizualnych oraz zdolność systemu do "wyobrażania sobie" i analizowania zasobów zewnętrznych przy użyciu silnika Exa API.
🏗️ Architektura Systemowa
Moduł 1400 nie jest prostym "rozpoznawaczem obrazów". Jest częścią Reaktora Wektorowego, gdzie dane wizualne są konwertowane… See the full description on the dataset page: https://huggingface.co/datasets/Ltac26/LTAC-1400_vision.json.Indonesian-Legal-Vision
Indonesian Legal Vision & Document Understanding Dataset (ID-Legal-Vision)
Repositori ini memuat dataset resmi visual dokumen hukum (Multimodal Document Vision & Visual Document Understanding / VDU) berdaulat Indonesia yang dikurasi, disintesis, dan dipublikasikan secara terbuka oleh Dahono Labs (PT Dahono Consulting Agency).
Dataset ini memuat 1.000 citra dokumen hukum Indonesia beresolusi tinggi (A4 150 DPI) yang dipasangkan dengan anotasi penalaran bertahap (Chain-of-Thought… See the full description on the dataset page: https://huggingface.co/datasets/DahonoLabs/Indonesian-Legal-Vision.