Almost every machine learning project starts with the same question: where do I get data that I am allowed to use? The good news is that a handful of free, long-running sources cover most needs, from small tabular tables for learning to web-scale text for language-model experiments. The bad news is that “free to download” is not the same as “free to use for anything,” and quality varies widely. This list describes eight places we would check first, what each is best for, and the one thing to verify before you build on its data. None of this is a ranking; the right source depends on your task.
1. Hugging Face Datasets
The Hugging Face Hub hosts a very large and growing collection of datasets covering text, images, audio, video and tabular data, with a search page you can filter by task, language and size. Most entries include a dataset card describing contents, source and licence, and many can be loaded in a line or two with the datasets library or previewed in the browser. It is the natural first stop for language, speech and vision work, and for benchmark data that papers cite. What to check: community-uploaded datasets differ in quality and documentation, so read the card and the licence field before training.
2. Kaggle Datasets
Kaggle hosts user-contributed and organization-contributed datasets alongside notebooks that show how others have analyzed them, which makes it excellent for learning and quick prototyping. Each dataset page lists a licence, file sizes and discussion, and datasets tied to competitions come with clearly defined tasks. What to check: licences vary from open to restrictive, and some uploads are scraped or derived from other sources, so trace the original provider when the data matters for anything beyond practice.
3. Google Dataset Search
Google Dataset Search is a search engine rather than a host. It indexes dataset descriptions from thousands of repositories, government portals and publishers that use standard metadata, and sends you to the original site. It is the best option when you know the topic (for example air quality, medical imaging or legal text) but not where the data lives. What to check: results link to external sites with their own terms, so follow through to the provider’s licence and citation requirements.
4. Common Crawl
Common Crawl is a nonprofit that publishes an open repository of web-crawl data collected since 2008, free for anyone to use. Its raw archives are a foundation for many text corpora used in language-model research, and the project provides tools and index files for working with them. It is the right source when you need web-scale text and have the engineering capacity to filter it. What to check: raw crawl data is noisy and contains personal and copyrighted material, so deduplication, language filtering and legal review are your responsibility. Our explainer on The Pile shows how a curated corpus tackles these issues.
5. OpenML
OpenML is an open platform for sharing datasets, machine-learning tasks and experiment results. Datasets come with machine-readable metadata, and tasks specify evaluation procedures, so you can compare your model with others’ runs on the same splits. It suits researchers and students who care about reproducible comparisons on tabular and classic benchmarks. What to check: collections are mostly small to medium in size, so it is a poor fit for large-scale deep-learning pretraining.
6. UCI Machine Learning Repository
The UCI Machine Learning Repository is one of the oldest dataset archives in the field, created in 1987 and still maintained by the University of California, Irvine. It offers a catalogue of well-documented, mostly small datasets for classification, regression and clustering, often used in textbooks and courses. It is ideal for teaching, baselines and quick experiments. What to check: because these datasets are so familiar, results on them say little about real-world performance, and some are decades old with outdated assumptions.
7. Data.gov
Data.gov is the U.S. government’s open-data catalogue, listing datasets from federal agencies on topics such as weather, health, transportation, energy and the economy. Much of it is public-domain information that can be reused broadly, and it is a strong source for real-world, domain-specific projects and for civic analysis. Many other countries and cities run similar portals, which you can find through search. What to check: formats and documentation vary by agency, and some datasets are updated irregularly or are summaries rather than raw records.
8. Zenodo
Zenodo is a general-purpose research repository operated by CERN. Researchers deposit datasets, code and papers there and receive a persistent DOI, so entries can be cited reliably. It is a good place to find the data behind a specific paper or niche scientific datasets that never reach larger platforms. What to check: quality and licensing are set by each depositor, and there is no central curation, so confirm the licence and read the accompanying paper.
How to vet any dataset before you use it
Whatever the source, a few questions save trouble later. Who collected the data, how and why? What licence applies, and does it allow your use, including commercial use? Does it contain personal data, and were people informed? Which groups or situations are missing? Is there a documented split between training and test data? The practice of writing down these answers is known as a datasheet for datasets, proposed in a 2018 paper by Timnit Gebru and colleagues, and our dataset datasheet checklist turns it into a reading routine. If you are still building foundations, our free learning path pairs well with the beginner-friendly sources above.
What to watch next
- Licence and provenance fields becoming more standardized across dataset hubs.
- Wider use of machine-readable dataset documentation in search results.
- Changes to crawl and scraping rules that affect what web data remains openly available.
- New government and academic portals publishing AI-ready data.
Key takeaways
- Hugging Face and Kaggle are the fastest starting points for modern ML data, with dataset cards or licence fields to read first.
- Google Dataset Search finds data hosted elsewhere; Zenodo and OpenML suit reproducible research.
- Common Crawl offers web-scale text but needs heavy filtering and legal care.
- UCI and Data.gov cover classic teaching sets and real-world public data respectively.
- Always check licence, provenance and personal data before training on any dataset.
Frequently asked questions
Is free data always free to use commercially?
No. “Free to download” and “free to use commercially” are different. Read the licence on each dataset and check whether it restricts commercial use or requires attribution.
Which source is best for a beginner?
Kaggle and the UCI repository are friendly starting points because datasets are small, documented and often come with example notebooks.
How do I find data for a niche topic?
Try Google Dataset Search first, then Zenodo, since both surface specialized datasets that larger hubs may not list.
