Best Free Datasets for Data Science Practice in 2026

Practice data is not hard to find; good practice data is. The best dataset for learning is one that matches the skill you want to build, is documented well enough that you can tell what each column means, and comes with a licence that lets you publish what you make. Below, sources are organised by skill, followed by a checklist for judging any dataset you find.

To learn algorithms: small, clean, built in

When the goal is to understand how a model behaves, not to fight with data, start with datasets that ship with your tools. scikit-learn includes small toy datasets such as iris and diabetes, and downloads larger real-world ones, such as California housing, on demand (scikit-learn: dataset loading utilities). They load in one line, have known properties, and appear in thousands of tutorials, so you can compare your results with others.

The UCI Machine Learning Repository has served teaching and research for decades, with a documentation page per dataset covering the variables, how the data was collected and how it has been used (UCI Machine Learning Repository). OpenML adds standardised tasks and published benchmark results for many datasets, which tells you what a reasonable score looks like before you start (OpenML).

The limitation of all three: these datasets are clean and famous. They teach algorithms, not the data work that fills most real jobs, and a portfolio built only on them looks like a course exercise.

To practise cleaning and analysis: public-sector data

Government and international open data is messier and closer to real work: inconsistent names, missing years, footnotes, changing definitions.

• Open Government Data Platform India publishes datasets from ministries and departments, from agriculture and transport to health and education (data.gov.in). Analysing Indian data also makes a portfolio relevant to local employers.
• World Bank Open Data covers development indicators for countries over decades (World Bank Open Data), which suits time comparisons and joining several indicators together.

Expect to spend most of the time on cleaning and reshaping. That is the point: it is the skill employers test.

To practise scale and performance: large trip records

New York City's Taxi and Limousine Commission publishes trip records monthly as Parquet files, with data dictionaries and a zone lookup table (TLC Trip Record Data). A few months add up to tens of millions of rows, which is enough to feel why vectorised code, efficient file formats and query engines matter. The Registry of Open Data on AWS lists much larger public datasets hosted in the cloud, each with its own licence and documentation (Registry of Open Data on AWS), for when you want to practise with cloud tools.

For text, images and audio

Hugging Face hosts a large collection of datasets for natural language processing, speech and vision, each with a dataset card describing its contents, collection method and licence (Hugging Face Datasets). Read the card before using a dataset: it often documents known biases and limitations, which are worth discussing in a project write-up.

Community collections and search engines

Kaggle Datasets has an enormous variety uploaded by users, each with example notebooks (Kaggle Datasets). Quality and licensing vary widely, so check who uploaded the data, where it came from, and what licence is stated. Google Dataset Search searches datasets published across the web (Google Dataset Search), which is the fastest way to find data on a narrow topic.

A checklist before you commit to a dataset

1. Licence. Can you use it, modify it and publish results and derived data? No stated licence means you cannot assume permission.
2. Documentation. Is there a data dictionary: what each column means, its units, how missing values are coded?
3. Provenance. Who collected it, how and when? Scraped data of unknown origin makes a weak foundation.
4. Size and shape. Big enough to be interesting, small enough to explore on your machine; you can sample down later.
5. Question fit. Does it contain what you need to answer a real question, including the target you want to predict?
6. Personal data. Avoid datasets with identifiable people unless they are properly anonymised and licensed for your use.

Turning a dataset into a project

A dataset is not a project; a question is. "Which taxi zones have the longest waits on rainy evenings?" is a project; "analysis of taxi data" is not. Pick a question someone would care about, state it at the top of your notebook, and finish with an answer and its limits. The data science projects for freshers article pairs twelve project ideas with datasets, and how to build a data science portfolio covers presenting the result.

Finding and preparing data is the first step of every project covered in the Data Science course, from its Python and exploratory analysis modules to machine learning and the portfolio work in its career preparation module.

Sources

• Dataset loading utilities (scikit-learn) — scikit-learn documentation
• UCI Machine Learning Repository — UC Irvine
• OpenML — OpenML
• Open Government Data Platform India — Government of India
• World Bank Open Data — World Bank
• TLC Trip Record Data — NYC Taxi and Limousine Commission
• Hugging Face Datasets — Hugging Face
• Kaggle Datasets — Kaggle
• Google Dataset Search — Google
• Registry of Open Data on AWS — AWS