Datasets

This section provides detailed documentation for all datasets available in stable-datasets.

Overview

stable-datasets provides easy access to a wide variety of datasets for machine learning research, with a focus on stability and reproducibility. Each dataset page includes:

• Example Samples: Visual examples or data snippets from the dataset

• Dataset Details: Number of classes, target types, and data specifications

• Data Structure: Keys and data types returned when accessing the dataset

• Usage Examples: Code snippets showing how to load and use the dataset

• Related Datasets: Links to similar or derived datasets

• Citation: The original paper to cite when using the dataset

Getting Started

All datasets can be loaded using the same consistent API:

from stable_datasets.images.<dataset_module> import <DatasetClass>

# First run will download + prepare cache, then return the split as a HF Dataset
ds = <DatasetClass>(split="train")

# If you omit the split (split=None), you get a DatasetDict with all available splits
ds_all = <DatasetClass>(split=None)

# Access individual examples
sample = ds[0]
print(sample.keys())  # e.g., {"image", "label"}

# Optional: make it PyTorch-friendly
ds_torch = ds.with_format("torch")

Available Datasets

Image Classification Datasets

• Arabic Characters
• CIFAR-10
• CIFAR-100
• CIFAR-10-C
• CIFAR-100-C
• CARS-196
• Describable Textures Dataset (DTD)
• MedMNIST

Note

Documentation is being added progressively, as datasets are ready for usage. Please only use datasets found in the documentation.