datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Egocentric_10K_Evaluation
Dataset Card for Egocentric_10K_Evaluation
This is a FiftyOne dataset with 30000 samples.
Installation
If you haven't already, install FiftyOne:
pip install -U fiftyone
Usage
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub
# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/Egocentric_10K_Evaluation")
# Launch the App
session = fo.launch_app(dataset)
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/Voxel51/Egocentric_10K_Evaluation.WEIRD
WEIRD
Описание задачи
WEIRD – это расширенная версия подзадачи бинарной классификации оригинального английского бенчмарка WHOOPS!. Датасет оценивает, способна ли мультимодальная модель обнаруживать нарушения здравого смысла в изображениях. Здесь нарушение здравого смысла – это ситуации, противоречащие типичным нормам реальности. Например, пингвины не могут летать, дети не водят автомобили, посетители не накладывают еду официантам, и так далее. В датасете поровну… See the full description on the dataset page: https://huggingface.co/datasets/MERA-evaluation/WEIRD.ppe-benchmark-eval
PPE Benchmark Eval Set (v1)
A held-out, human-verified benchmark for evaluating vision-language models on
personal protective equipment (PPE) detection — specifically hardhat and
safety-vest presence — framed as a VQA-style classification task.
What this is
96 images, balanced 24/24/24/24 across the four hardhat × vest combinations
(yes/yes, yes/no, no/yes, no/no). Sourced from a forked, filtered subset of
the karabuk-university PPE dataset
on Roboflow Universe… See the full description on the dataset page: https://huggingface.co/datasets/khadijah00/ppe-benchmark-eval.japanese-image-classification-evaluation-dataset
recruit-jp/japanese-image-classification-evaluation-dataset
Overview
Developed by: Recruit Co., Ltd.
Dataset type: Image Classification
Language(s): Japanese
LICENSE: CC-BY-4.0
More details are described in our tech blog post.
日本語CLIP学習済みモデルとその評価用データセットの公開
Dataset Details
This dataset is comprised of four image classification tasks related to concepts and things unique to Japan. Specifically, is consists of the following tasks.
jafood101: Image… See the full description on the dataset page: https://huggingface.co/datasets/recruit-jp/japanese-image-classification-evaluation-dataset.evaluation-dataset
DeepSafe Evaluation Dataset
Evaluation set for DeepSafe,
a deepfake detection benchmark.
Tiers
Tier
Samples
Generators
Size
Use
master_eval_small/
198
116
1.7 GB
smoke test, under 2 min
master_eval/
15,454
411
10 GB
the standard benchmark
master_eval_full/
45,954
411
25 GB
complete set
Medium tier composition: 9,954 image, 3,500 audio, 2,000 video.
from huggingface_hub import snapshot_download
snapshot_download("deepsafe/evaluation-dataset"… See the full description on the dataset page: https://huggingface.co/datasets/deepsafe/evaluation-dataset.medical-imaging-model-evaluation-benchmark
医学影像多模态模型评测集（精选示例版）
这是一个面向医学多模态大模型的高质量影像评测集，专门测试模型能否把“看见影像”进一步转化为可解释、可复核、符合临床语境的判断与表达。数据将医学影像与患者描述、病史摘要、检查信息或结构化临床资料配对，覆盖从影像分类、报告生成，到鉴别诊断、治疗方案和胸片质量控制的完整评测链路。
本次公开版本从 2026-07-22 质检通过产物中整理而来，按每个子集最多 50 题进行分层抽样；题量不足 50 的影像质量控制子集完整保留。因此，公开版本包含 5 个任务子集、213 题和 455 个配套影像文件，适合作为医学视觉语言模型的快速对比集、回归测试集和研究教学样例。
数据集亮点
多模态对齐：每条样例同时提供影像和结构化的 question、answer、explanation，支持检查视觉理解、临床语义整合与解释质量。
任务覆盖完整：从“影像是什么”到“如何描述、如何鉴别、如何处置”，并加入真实影像工作流中的胸片质量控制任务。
影像类型丰富：覆盖 X… See the full description on the dataset page: https://huggingface.co/datasets/SHPDRG/medical-imaging-model-evaluation-benchmark.lipika-eval
Lipika eval — Indic font recognition benchmark
The frozen validation set behind loopdesk-ai/lipika
(Indic font recognizer): 6,876 synthetic text crops covering 553 freely-licensed font
families across 13 scripts (Devanagari, Bengali, Gujarati, Gurmukhi, Kannada, Malayalam,
Meetei Mayek, Odia, Ol Chiki, Perso-Arabic, Tamil, Telugu, Latin).
This is the set reported as "synthetic val" in the model card (Lipika v2.4 scores 0.849
family top-1 / 0.977 top-5 / 0.991 script). Use it to… See the full description on the dataset page: https://huggingface.co/datasets/loopdesk-ai/lipika-eval.nanopath-evals
NanoPath evaluation data
This is the immutable data mirror used by NanoPath probe protocol v2. It contains only the exact development records consumed by medarc/nanopath: selected THUNDER training/validation images, prepared development-only slide caches, and the two PathoROB subsets. manifest.json records SHA-256 checksums and binds the snapshot to the checked-in benchmark manifests.
No official THUNDER, HEST, or CPTAC classification test record is included. HEST is absent.… See the full description on the dataset page: https://huggingface.co/datasets/medarc/nanopath-evals.solring-eval
Sol Ring Dataset
(c) 2026, HanClinto Games, LLC
A collection of 307 reference frames for benchmarking Magic: The Gathering card
identification — specifically edition (set) discrimination under real-world
camera conditions.
Purpose
To provide a meaningful, reproducible metric for measuring and comparing the
accuracy of card recognition algorithms, with particular focus on
set / edition identification rather than just card-name recognition.
Theory
In Magic: The… See the full description on the dataset page: https://huggingface.co/datasets/HanClinto/solring-eval.wonders-of-world-images-hf
🌍 Wonders of the World Images 🏛️
¡Bienvenido/a a un viaje visual por las maravillas del mundo!
Este dataset contiene imágenes de 12 maravillas icónicas, listas para que entrenes modelos de visión por computadora, juegues a ser explorador o simplemente disfrutes de la diversidad arquitectónica y natural del planeta.
📦 Estructura del dataset
Clases:
Burj Khalifa
Chichen Itza
Christ the Redeemer
Eiffel Tower
Great Wall of China
Machu Picchu
Pyramids of Giza
Roman… See the full description on the dataset page: https://huggingface.co/datasets/evalverden/wonders-of-world-images-hf.diagram-eval-access-test
Diagram Evaluation Access Test
This public one-image dataset tests the external-access workflow planned for the
Diagram Evaluation project. It is not a research dataset release.
Dataset structure
The repository uses Hugging Face's ImageFolder layout:
data/
train/
metadata.csv
sample-diagram.png
The metadata records the image's provenance, license, and intended use. A production
release can use the same contract with sharded Parquet or WebDataset files.… See the full description on the dataset page: https://huggingface.co/datasets/abhisheklalwani96/diagram-eval-access-test.facepass_eval
FacePass Evaluation Dataset (Real LFW Faces)
This dataset contains real face images from the LFW (Labeled Faces in the Wild) dataset, curated for face recognition evaluation.
⚠️ IMPORTANT: This is the corrected version with actual face photographs (not colored squares).
Key Features
✅ Real faces: Actual photographs of people, not synthetic images✅ Balanced dataset: All individuals have 20+ images✅ Proper splits: 80/20 train/test split per person✅ Standardized: Resized to… See the full description on the dataset page: https://huggingface.co/datasets/besartshyti/facepass_eval.mixlora-eval-data
🚀 MixLoRA Evaluation Data
This dataset is the held-out multimodal evaluation suite used in
Multimodal Instruction Tuning with Conditional Mixture of LoRA (ACL 2024).
It bundles 9 instruction-formatted tasks (mm_tasks/) plus the MME benchmark
(mme/) used to evaluate MixLoRA and baseline models in the paper.
The 9 tasks in mm_tasks/ are the zero-shot / held-out task split from
Vision-Flan. MME is a
separate benchmark, evaluated independently.
Dataset Structure… See the full description on the dataset page: https://huggingface.co/datasets/yingss/mixlora-eval-data.tiktok-techjam-2026-eval
TikTok TechJam 2026 Eval
Held-out demonstration pair used by Seer:
COCO val2017 photographs plus the WildFake DALL·E Advanced (DALL·E 3) subset.
Do not train on this split.
Contents
label
meaning
count
origin
0 / real
photograph
5,000
COCO val2017
1 / fake
AI-generated
8,843
WildFake DALL·E Advanced
Columns: image, label, source, generator, id.
id is the COCO stem for reals, and {session}_{stem} for fakes so duplicate
WildFake basenames stay… See the full description on the dataset page: https://huggingface.co/datasets/glennwuwu/tiktok-techjam-2026-eval.MUMU-Eval-6000
MUMU Eval 6000
This repository contains the 6,000-image source-data evaluation set used for
the Florence-2 and LFM2.5-VL-450M baselines in the MUMU evaluation repository.
It is an independently prepared research split, not an official MUMU Challenge
release.
Splits
Split
Images
Ground truth in manifest
validation
1,000
Yes
test
5,000
Yes
The split contains 2,001 Task A samples, 2,000 Task B samples, and 1,999 Task C
samples. All 6,000 image… See the full description on the dataset page: https://huggingface.co/datasets/JinyuLiu/MUMU-Eval-6000.evaluation
Skill-Aligned Annotation for Text-to-Image Evaluation
Companion dataset for the NeurIPS 2026 paper "Towards Objective Evaluation".
The dataset contains generated images from 7 text-to-image models, evaluated
by 6 human annotators (anonymized) plus an LLM judge across 9 skill-aligned
annotation strategies.
Configs
Config
Rows
Description
images
621
Generated images (621 WebP) with embedded bytes; one row per (prompt_id, generator).
prompts
179
Per-prompt… See the full description on the dataset page: https://huggingface.co/datasets/Skill-Aigned/evaluation.evals-eastrus-vl
evals-eastrus-vl
Independent evaluation dataset for the EstrusVision cattle estrus detection model. Contains ground-truth labels for measuring deployment readiness.
Contents
43 total samples (40 in-domain cattle vulval images, 3 out-of-domain)
Embedded image column (no external file dependencies)
Six symptom ground-truth labels per in-domain sample
out_of_domain flag for rejection testing
notes field with clinical observations
Label distribution (in-domain only)… See the full description on the dataset page: https://huggingface.co/datasets/prapaa/evals-eastrus-vl.piyoshogi-eval
PiyoShogi Eval (paired, 4 devices)
ぴよ将棋の盤面認識モデルの評価用データセット。4機種の実機スクリーンショットを SFEN 単位で束ねた横持ち形式。
対応機種
機種名
識別子 (devices 列)
画面解像度
iPhone 8
iPhone10,1
750 × 1334
iPhone XR
iPhone11,8
828 × 1792
iPhone 15
iPhone15,4
1179 × 2556
iPad Air M3
iPad14,10
1640 × 2360
paired config
1 行 = 1 SFEN、4 機種分の画像を list で持つ。
Column
Type
説明
sfen
string
SFEN形式の局面文字列
hash
string
SFENのSHA-256
type
string
局面ソース種別（現状は全て existing = ぴよ将棋プリセット由来）… See the full description on the dataset page: https://huggingface.co/datasets/ultemica/piyoshogi-eval.sailor-moon-redraw-eval