leandrodevai/imagewoof-320px-resplit

ImageWoof 320px with Fixed Validation and Test Splits

Dataset Description

This dataset is a reproducible, Parquet-based version of the 320px configuration of frgfm/imagewoof. ImageWoof is a subset of ten dog-breed classes from ImageNet designed to be more difficult than broad-category image-classification benchmarks.

This version is intended for image classification and confidence-calibration experiments. It introduces two changes to the source dataset:

1. It corrects the source ClassLabel metadata so every numeric label matches its ImageNet synset and breed name.
2. It divides the original validation set into fixed, stratified validation and test splits.

The images and numeric labels are inherited unchanged from the source dataset. Only label metadata and split assignment were modified.

Dataset Sources

• Source dataset: frgfm/imagewoof
• Original project: fastai/imagenette
• Image source: ImageNet
• Configuration: 320px
• License declared by the source dataset: Apache License 2.0

Dataset Structure

Each example contains:

{
    "image": PIL.Image.Image,
    "label": int,
}

The label feature is a ClassLabel with ten breed names.

Splits

Split	Examples	Intended use
train	9,025	Model training and training-only preprocessing statistics
validation	1,964	Model selection, early stopping, and calibration fitting
test	1,965	Final evaluation only
Total	12,954

The source training split remains unchanged. The source validation split contained 3,929 examples and was divided using a stratified 50/50 split with seed 42.

Label Metadata Correction

The source dataset stores numeric labels in ImageNet synset order, but its original ClassLabel.names list uses a different order. This causes displayed breed names to disagree with the images even though the underlying image-to-ID assignments are correct.

This version preserves every numeric label and assigns the following corrected metadata:

Label ID	ImageNet synset	Correct class name	Source metadata name
0	n02086240	Shih-Tzu	Australian terrier
1	n02087394	Rhodesian ridgeback	Border terrier
2	n02088364	Beagle	Samoyed
3	n02089973	English foxhound	Beagle
4	n02093754	Border terrier	Shih-Tzu
5	n02096294	Australian terrier	English foxhound
6	n02099601	Golden retriever	Rhodesian ridgeback
7	n02105641	Old English sheepdog	Dingo
8	n02111889	Samoyed	Golden retriever
9	n02115641	Dingo	Old English sheepdog

Class Distribution

ID	Class	Train	Validation	Test	Total
0	Shih-Tzu	941	204	205	1,350
1	Rhodesian ridgeback	942	204	204	1,350
2	Beagle	932	209	209	1,350
3	English foxhound	580	112	112	804
4	Border terrier	949	201	200	1,350
5	Australian terrier	943	203	204	1,350
6	Golden retriever	949	201	200	1,350
7	Old English sheepdog	928	211	211	1,350
8	Samoyed	921	214	215	1,350
9	Dingo	940	205	205	1,350

English foxhound contains fewer examples than the other classes. Users should consider this imbalance when interpreting aggregate classification and calibration metrics.

Dataset Creation

Split Procedure

The fixed held-out splits were created from the source validation split with:

validation_test = source["validation"].train_test_split(
    test_size=0.5,
    seed=42,
    stratify_by_column="label",
)

The resulting validation_test["train"] split became validation, and validation_test["test"] became test. Publishing these assignments as Parquet removes dependence on future library behavior and ensures that all experiments use the same examples.

Preprocessing

The source 320px configuration resizes the shorter side of each image to 320 pixels while preserving its aspect ratio. Images are not guaranteed to be square.

No additional resizing, cropping, augmentation, or pixel normalization was applied while creating this dataset.

Training-Set Channel Statistics

Pixel-weighted RGB statistics were calculated from the complete training split after converting images to RGB and scaling values to [0, 1]:

mean = (0.485513, 0.455452, 0.393252)
std = (0.259905, 0.252752, 0.261519)

These values may be useful when training a model from scratch. Models initialized with pretrained ImageNet weights should generally use the preprocessing and normalization specified by those weights, such as weights.transforms() in torchvision.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "leandrodevai/imagewoof-320px-calibration",
    "320px",
)

train_dataset = dataset["train"]
validation_dataset = dataset["validation"]
test_dataset = dataset["test"]

For calibration experiments:

1. Train the classifier on train.
2. Fit temperature scaling or another post-hoc calibrator on validation.
3. Report classification and calibration metrics once on test.

The test split should not be used for architecture selection, hyperparameter tuning, early stopping, or calibrator fitting.

Intended Uses

This dataset is suitable for:

• supervised dog-breed image classification;
• confidence calibration and reliability-diagram experiments;
• evaluation of expected calibration error, negative log-likelihood, and Brier score;
• post-hoc calibration methods such as temperature scaling;
• controlled comparisons between pretrained and from-scratch training strategies.

Limitations

• The dataset contains only ten dog breeds and does not represent general image classification.
• English foxhound is substantially underrepresented relative to the other classes.
• Images inherit ImageNet collection biases, including correlations between breeds, backgrounds, framing, and photographic style.
• Breed labels describe visual categories and should not be interpreted as guarantees about an individual animal's pedigree.
• The validation and test sets originate from the same source split and should not be interpreted as independent data-collection domains.
• Calibration results on this dataset may not transfer to other datasets, domain shifts, corruptions, or deployment settings.

Personal and Sensitive Information

The primary subjects are dogs. As with the source ImageNet data, some images may incidentally contain people, locations, text, or other contextual information. No additional personal information was added during preparation.

Licensing

The source Hugging Face dataset declares the Apache License 2.0. This derivative preserves that declaration. Users are responsible for verifying that their intended use complies with the source dataset, ImageNet terms, and any rights associated with individual images.

Citation

If this dataset is useful, cite the original ImageWoof project:

@software{Howard_Imagewoof_2019,
  title = {Imagewoof: a subset of 10 classes from Imagenet that aren't so easy to classify},
  author = {Jeremy Howard},
  year = {2019},
  month = {March},
  publisher = {GitHub},
  url = {https://github.com/fastai/imagenette#imagewoof}
}

When reporting experiments, also document that this fixed-split derivative uses a stratified 50/50 division of the original validation split with seed 42 and corrected ImageNet-synset label metadata.