MaTU-30

All 30 image-classification tasks in one collection, with 1,041,734 images, explicit class labels, train/test membership and source hashes. Prepared for MaTU (paper) and related federated multi-task and model-merging experiments. Every configuration includes embedded image bytes; EuroSAT is the default viewer configuration.

The complete image shards occupy about 20.78 GB. Allow additional space for Hub and Arrow caches. This is a curated, source-based collection; the split differences below mean it is not an exact reproduction of every original MaTU or EMR protocol.

Usage

from datasets import load_dataset

data = load_dataset("vtsouval/MaTU-30", "eurosat")
example = data["train"][0]
image, label = example["image"], example["label"]
class_name = data["train"].features["label"].int2str(label)

Prepare all tasks with the MaTU CLI:

python scripts/prepare_data.py --benchmark thirty --hf-repo vtsouval/MaTU-30

Use --benchmark small for the five-task starter set or --dataset NAME for one configuration. Pin revision="<commit SHA>" in Hugging Face, or --revision <commit SHA> in the CLI, for reproducibility. All images are supplied here; no separate source map is needed.

Configurations

Source links identify the original datasets. The attribution column is abbreviated; complete creators, citations, primary references and conditions are retained in each linked terms record.

Config / source	Classes	Train / test	Original credit	Source terms
animals10n	10	50,000 / 5,000	Hwanjun Song et al.	No named licence
cars	196	8,144 / 8,041	Jonathan Krause et al.	Source terms
cifar10	10	50,000 / 10,000	Alex Krizhevsky et al.	Unknown (publisher card)
cub200	200	9,430 / 2,358	Catherine Wah et al.	Noncommercial research/education
dtd	47	3,760 / 1,880	Mircea Cimpoi et al.	Research use
eurosat	10	21,600 / 5,400	Patrick Helber et al.	MIT
fashionmnist	10	60,000 / 10,000	Han Xiao et al.	MIT
fruits	24	6,231 / 3,110	Mihai Oltean	MIT (historical release)
gtsrb	43	26,640 / 12,630	Johannes Stallkamp et al.	Free use; citation
kenyanfood	13	6,536 / 815	Kaihong Wang et al.	Source terms
landscape	5	10,000 / 500	DeepNets (Kaggle publisher utkarshsaxenadn)	CC0 1.0
mnist	10	60,000 / 10,000	Yann LeCun et al.	MIT (publisher card)
resisc45	45	25,200 / 6,300	Gong Cheng et al.	CC BY-NC 4.0
sun397	397	87,003 / 21,751	Jianxiong Xiao et al.	Research only
vegetables	15	15,000 / 3,000	M. Israk Ahmed et al.	CC BY-SA 4.0
beans	3	1,034 / 128	AI Lab Makerere	MIT (author metadata)
catsdogs	2	19,998 / 5,000	Microsoft Research et al.	Source terms
cifar100	100	50,000 / 10,000	Alex Krizhevsky et al.	Unknown (publisher card)
dogs	120	12,000 / 8,580	Aditya Khosla et al.	ImageNet-derived source
emnist	47	112,800 / 18,800	Gregory Cohen et al.	No named licence
flowers	5	3,453 / 864	Alexander Mamaev	Unknown (publisher metadata)
food101	101	75,750 / 25,250	Lukas Bossard et al.	Unknown (publisher card)
garbage	6	1,768 / 431	Gary Thung, Mindy Yang	MIT
intel	6	14,034 / 3,000	Intel et al.	Original-owner copyright
kvasir_v2	8	6,400 / 1,600	Konstantin Pogorelov et al.	Research/education
mangoleaf	8	3,200 / 800	Sawkat Ali et al.	CC BY-NC 3.0
pets	37	3,680 / 3,669	Omkar M. Parkhi et al.	CC BY-SA 4.0
stl10	10	5,000 / 8,000	Adam Coates et al.	ImageNet-derived source
svhn	10	73,257 / 26,032	Yuval Netzer et al.	Noncommercial
weather	11	5,501 / 1,376	Haixia Xiao	CC0 1.0

Schema and integrity

Every row contains image (embedded bytes, null embedded path), label (ClassLabel), sample_id, source_path, source_sha256 and source_hash_type. Source paths are relative identifiers. Hash types distinguish original file bytes from decoded RGB size and pixel bytes. Native MNIST, CIFAR10, Fashion-MNIST and EMNIST arrays are losslessly encoded as PNG; other files retain their prepared source bytes.

catalog.json provides classes, counts, shard SHA256 values and membership digests. Detailed provenance and split policies are in metadata/<config>.json. All source files were decoded and hash-checked; native pixels were checked before and after PNG encoding. Packing preserves the existing five-field membership digests exactly. A membership digest hashes newline-terminated, sorted-key compact JSON of the five non-image row fields in order.

Split notes

• MNIST, CIFAR10, Fashion-MNIST, EMNIST Balanced and STL10 retain canonical native membership. Stanford Dogs uses official annotation membership; Fruits is a 24-class historical subset with restored author splits.
• GTSRB retains a reduced 26,640-image training corpus. DTD retains 3,760/1,880 rows with an unproven original partition. CUB200 uses a new stratified split.
• Unassigned source pools use stratified 80/20 with seed 42 (scikit-learn 1.1.3). Existing source assignments are retained, validation rows are omitted, and no arbitrary test-set cap is applied.
• CatsDogs excludes two corrupt files, cat/666.jpg and dog/11702.jpg. Final CatsDogs, Dogs and Fruits assignment CSVs are supplied under splits/.
• Exact duplicate content crosses train/test in 16 configurations. Memberships are preserved and overlap counts disclosed per config; these splits are not guaranteed free of evaluation leakage. Byte equality does not detect perceptual duplicates or differently encoded copies.

External source-identity checks have differing coverage, recorded per config. In particular, Weather has matches to author examples and a consistent source chain; MangoLeafBD has matching primary source metadata, but original-archive pixel fingerprints were unavailable. These are not described as complete original-byte verification.

Licensing and credits

The curator has confirmed clearance to redistribute all 30 image configurations. That authorization is distinct from independently documented source licences: there is no blanket MIT licence. Source-specific noncommercial, research, attribution and ShareAlike conditions remain applicable. Terms labelled unknown or unnamed are reported as such; curator authorization is recorded separately in the metadata.

Notable restrictions include noncommercial terms for SVHN, MangoLeafBD (CC BY-NC 3.0) and RESISC45 (CC BY-NC 4.0); research/education restrictions for CUB200, DTD, SUN397 and Kvasir-V2; and CC BY-SA 4.0 terms for Pets and Vegetables. Kvasir-V2 requires its specified publication citation; other purposes, including commercial use, require the source's written permission. Consult each terms record for the complete applicable wording.

Original creators retain their rights. Existing copyright notices, primary-source references and all named creators are retained in licenses/ and licenses/dataset-licenses.json. EuroSAT credits Copernicus Sentinel imagery. Fruits files match a historical MIT release; the current author release uses CC BY-SA 4.0, and both notices are preserved. MNIST's MIT label is attributed specifically to the publisher card; unspecified source terms are not converted to MIT.

This release adds Parquet packaging, hashes and the documented membership changes. It does not resize images; native arrays are encoded losslessly. Some recovered source exports are JPEG derivatives. Please cite MaTU when using this collection and the original dataset papers and creators listed in the per-config metadata.