RealMattBrewer/oxford-iiit-pet-webdataset

Oxford-IIIT Pet — WebDataset shards with published ratarmount indexes

The Oxford-IIIT Pet dataset — 37 cat and dog breeds, 7,349 photos — packaged as plain, uncompressed WebDataset tar shards that work everywhere tar shards work (🤗 datasets, the dataset viewer, webdataset, PyTorch loaders), plus a published ratarmount-rs index for every shard.

With the published indexes you can mount the whole dataset straight from the Hub and open any image instantly — no download, no extraction, no scanning — or read individual samples by name from Python.

	Train	Test
Samples	3,680	3,669
Shards	3	4
Size	380 MiB	414 MiB
Classes	37 breeds (12 cat, 25 dog)	37 breeds

Quick start

🤗 Datasets

from datasets import load_dataset

ds = load_dataset("RealMattBrewer/oxford-iiit-pet-webdataset", split="train")
sample = ds[0]
sample["jpg"]            # PIL.Image
sample["json"]["breed"]  # "abyssinian"
sample["cls"]            # 0

Streaming works too: load_dataset(..., split="train", streaming=True).

WebDataset

import webdataset as wds

url = "https://huggingface.co/datasets/RealMattBrewer/oxford-iiit-pet-webdataset/resolve/main/train/train-{000000..000002}.tar"
dataset = wds.WebDataset(url, shardshuffle=True).shuffle(1000).decode("pil").to_tuple("jpg", "cls")

Mount it with ratarmount-rs (no download)

ratarmount -r hf://datasets/RealMattBrewer/oxford-iiit-pet-webdataset pets/
ls pets/train/train-000000.tar/ | head
xdg-open pets/train/train-000000.tar/Abyssinian_1.jpg
ratarmount -u pets/

The whole repository mounts in about 0.2 s, pinned to one commit. -r opens each tar shard as a directory on first access, using the published index next to it instead of reading the shard, and every file you open is fetched with HTTP range requests (with an optional NVMe range cache).

To see every training sample in one flat directory, mount the shards together:

R=hf://datasets/RealMattBrewer/oxford-iiit-pet-webdataset
ratarmount $R/train/train-00000{0,1,2}.tar train/   # 3,680 samples, one directory

The mount is ideal for browsing, spot checks and random access by name; for full-speed training epochs, stream whole shards with 🤗 datasets or webdataset as shown above.

Read samples by name from Python (ratarmount-rs library)

import ratarmount_rs

shard = ratarmount_rs.open(
    "hf://datasets/RealMattBrewer/oxford-iiit-pet-webdataset/train/train-000000.tar"
)
jpeg_bytes = shard.read("/Abyssinian_1.jpg")      # one range request, no scan
meta = shard.read("/Abyssinian_1.json")
for entry in shard.iter(order="archive"):          # archive order, paged
    ...

The ratar fsspec protocol gives fsspec-aware libraries the same access:

import fsspec

fs = fsspec.filesystem("ratar", fo="hf://datasets/RealMattBrewer/oxford-iiit-pet-webdataset/train/train-000000.tar")
fs.cat_file("Abyssinian_1.json")

ratarmount-rs hf:// support, published indexes and the Python package are part of the Local AI work and ship with ratarmount-rs 0.2.0. Until that release, build from the repository's main branch.

Dataset structure

Each sample is three files sharing one key — the official Oxford-IIIT image name — so every WebDataset reader groups them automatically:

train-000000.tar
├── Abyssinian_1.jpg    the photo (JPEG)
├── Abyssinian_1.json   {"image_id": "Abyssinian_1", "label": 0, "breed": "abyssinian",
│                        "species": "cat", "split": "train"}
├── Abyssinian_1.cls    0
└── ...

Field	Type	Description
jpg	image	Original photo as published by the authors
cls	int	Breed label, 0–36, in the standard alphabetical order
json.breed	string	Breed name (abyssinian … yorkshire_terrier)
json.species	string	cat or dog
json.image_id	string	Official image name; joins with the trimaps and bounding boxes in the original metadata
json.split	string	train or test (the standard split)

All 37 breeds and their labels

Label	Breed	Label	Breed	Label	Breed
0	abyssinian	13	english_setter	26	ragdoll
1	american_bulldog	14	german_shorthaired	27	russian_blue
2	american_pit_bull_terrier	15	great_pyrenees	28	saint_bernard
3	basset_hound	16	havanese	29	samoyed
4	beagle	17	japanese_chin	30	scottish_terrier
5	bengal	18	keeshond	31	shiba_inu
6	birman	19	leonberger	32	siamese
7	bombay	20	maine_coon	33	sphynx
8	boxer	21	miniature_pinscher	34	staffordshire_bull_terrier
9	british_shorthair	22	newfoundland	35	wheaten_terrier
10	chihuahua	23	persian	36	yorkshire_terrier
11	egyptian_mau	24	pomeranian
12	english_cocker_spaniel	25	pug

Files in this repository

Path	What it is
train/train-00000N.tar, test/test-00000N.tar	WebDataset shards (uncompressed tar, ≤ 128 MiB)
*/<shard>.tar.index.ptr	ratarmount index pointer (ratarmount.index.pointer.v1, sha256 of the index)
*/<shard>.tar.index.<sha256>.sqlite	Published ratarmount index for the shard
*/shards.json	Manifest: per-shard size, sha256, member and sample counts, first/last key

Why published indexes

A tar file has no table of contents: to find one member, a reader normally scans the archive from the start. ratarmount builds a small SQLite index of every member's offset once, and ratarmount --publish-index uploads it next to the shard. Every later user's mount discovers that index, checks it against the shard, and seeks straight to the bytes they need.

Measured from a laptop with empty caches, reading directly from the Hub (medians of three runs on 2026-10-17 with ratarmount-rs's hf-dataset.py benchmark):

Operation	Without an index	With the published index
Mount the whole repository (-r)	—	0.16 s
Open train-000000.tar (128 MiB, 3,954 members) inside that mount	—	1.5 s
Mount that one shard on its own	16.8 s (remote scan)	1.2 s
Read one 125 KiB photo from the open shard	—	0.1 s

How it was built

Reproducible from the source parquet files with ratarmount-rs:

ratarmount shard train/ -o out/train --prefix train --codec none --size 128M --mtime 1792108800
ratarmount shard test/  -o out/test  --prefix test  --codec none --size 128M --mtime 1792108800
ratarmount --publish-index hf://datasets/RealMattBrewer/oxford-iiit-pet-webdataset@main/train/train-000000.tar
# ...one --publish-index per shard

• Source: timm/oxford-iiit-pet (standard labels and splits).
• Samples are stored in byte order of their image names, so shards are grouped by breed — shuffle when training (shard shuffling plus a shuffle buffer, as in the WebDataset example above).
• Four source images were PNG files despite the dataset's JPEG convention (Egyptian_Mau_14, Egyptian_Mau_156, Egyptian_Mau_186 in train, Abyssinian_5 in test); they were re-encoded as JPEG (quality 95) so every sample has a jpg field. All other images are byte-identical to the source.
• Tar members use a fixed timestamp (2026-10-16) and mode 0644; identical input produces identical shards.

License and attribution

The images and annotations are by Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman and C. V. Jawahar (Visual Geometry Group, University of Oxford, and CVIT, IIIT Hyderabad), distributed under the Creative Commons Attribution-ShareAlike 4.0 International license. This repackaging is distributed under the same license. Image copyright remains with the original photographers; see the dataset page for details.

Citation

@InProceedings{parkhi12a,
  author    = "Omkar M. Parkhi and Andrea Vedaldi and Andrew Zisserman and C. V. Jawahar",
  title     = "Cats and Dogs",
  booktitle = "IEEE Conference on Computer Vision and Pattern Recognition",
  year      = "2012",
}