Hugging Face Datasets is an open-source library for loading, processing, streaming, and sharing datasets used in machine learning and AI. It provides simple loaders for datasets on the Hugging Face Hub as well as local data, with support for text, images, audio, video, PDFs, medical imaging, and many common structured file formats.

The library is built around efficient, reproducible preprocessing and can feed data into ecosystems such as NumPy, Pandas, PyTorch, TensorFlow, JAX, and Polars. Features include dataset mapping and transformation, streaming without downloading an entire corpus, multimodal columns, caching, and integration with the wider Hugging Face model-training workflow.

Datasets is useful for researchers and engineers who need a consistent data layer across experimentation, evaluation, and training. It reduces the custom code normally required to download, parse, transform, and iterate over heterogeneous datasets while making published datasets easier to reuse and share.