Free Datasets for Practice Projects and Portfolios

Free Datasets for Practice Projects and Portfolios

Free datasets for data science projects come from four reliable places: Kaggle, the UCI Machine Learning Repository, government open-data portals, and Google Dataset Search. Pick based on whether you need something beginner-clean, academically rigorous, real-world messy, or hard to find.

Updated for 2026. Links and free-tier details change — verify before relying on any resource for a deadline.

A portfolio project lives or dies on the quality of its underlying data as much as the quality of its analysis. Picking a dataset almost at random and running the same tutorial everyone else has already run does not demonstrate much. The sources below are grouped by what they are actually good for, so you can match the dataset to the kind of project you are trying to build.

General-Purpose Dataset Hubs

Kaggle Datasets — the largest community-driven dataset hub, with thousands of datasets ranging from beginner tabular data to large real-world corpora. Best for: finding a dataset that already has example notebooks and an active discussion thread attached. Status: free, with per-dataset licenses that vary — check the license tab before publishing derived work.

Google Dataset Search — a search engine, not a host, that indexes datasets from thousands of repositories including government sites, academic archives and Kaggle itself. Best for: finding a dataset on a narrow, specific topic that isn't on Kaggle or UCI. Status: free to search; licensing depends entirely on the original source it links to.

Papers With Code Datasets — a section of the Papers With Code site that links datasets to the academic papers and benchmark leaderboards that use them. Best for: finding the exact dataset behind a specific published machine learning result you want to reproduce. Status: free, links out to original sources.

Academic and Machine Learning Benchmark Sources

UCI Machine Learning Repository — one of the oldest dataset collections online, maintained by the University of California, Irvine, with smaller and generally cleaner datasets built for specific classification, regression or clustering tasks. Best for: learning a specific algorithm end-to-end without fighting messy data first. Status: free, mostly public domain or academic-use licenses.

Hugging Face Datasets — a hub of datasets built primarily for natural language processing and machine learning model training, searchable and filterable by task type. Best for: NLP, text classification and language model fine-tuning projects. Status: free, with per-dataset licenses shown on each dataset card.

COCO (Common Objects in Context) — a large, widely used benchmark dataset of labeled images for object detection, segmentation and captioning, maintained by a consortium of research and industry contributors. Best for: computer vision projects needing richly annotated images rather than raw, unlabeled photos. Status: free for research and most non-commercial use; check the specific license terms before commercial use.

Open Images Dataset (from Google) — a large collection of images with labels, bounding boxes and segmentation masks across thousands of object categories. Best for: computer vision projects at a larger scale than COCO, or needing a broader category set. Status: free, images are licensed CC-BY with per-image attribution available.

ImageNet — the dataset behind the famous ImageNet Large Scale Visual Recognition Challenge that helped kick off the modern deep learning era, still widely used as a benchmark and pretraining source. Best for: image classification projects wanting to compare results against a well-known historical benchmark. Status: free for non-commercial research use with registration; check current terms for commercial use.

Government and Public Data Portals

data.gov — the United States federal government's central open-data portal, covering agriculture, climate, transportation, health, finance and more. Best for: realistic, large-scale civic datasets for analytics and visualization portfolio pieces. Status: free, generally public domain, though check individual agency licensing notes.

Our World in Data — a research project publishing cleaned, well-documented global datasets on topics like population, energy, health and poverty, alongside the charts built from them. Best for: global comparative analysis and dashboard-style visualization projects. Status: free, openly licensed for reuse with attribution.

Eurostat / national statistics offices — most countries publish their national statistics agency's data openly (Eurostat for the EU, equivalents elsewhere). Best for: region-specific economic or demographic analysis. Status: free, licensing varies by country but is typically open for non-commercial and commercial reuse alike.

Domain-Specific and Niche Sources

Kaggle Competitions datasets — beyond the general dataset hub, Kaggle's competition pages each ship with a dedicated training and test dataset, often more rigorously cleaned than community uploads because a competition depends on data quality. Best for: practicing a specific modeling technique against a dataset other people are actively working with, so you can compare your approach to public notebooks. Status: free, competition-specific rules apply to some datasets restricting reuse outside the competition.

FiveThirtyEight's data repository — a public GitHub repository of the datasets behind FiveThirtyEight's data journalism articles, covering sports, politics, economics and culture. Best for: a visualization or storytelling project where the dataset already comes with a published article's worth of context and narrative angle. Status: free, openly licensed for reuse.

World Bank Open Data — global economic and development indicators (GDP, population, health, education) across nearly every country, going back decades. Best for: time-series and cross-country comparison projects. Status: free, open license for reuse including commercial use.

NASA Open Data Portal — a large collection of space, earth science and climate datasets published directly by NASA. Best for: a niche, attention-grabbing portfolio project outside the usual housing-price or titanic-survival territory. Status: free, U.S. government works are generally public domain.

AWS Open Data Registry — a catalog of large public datasets hosted for free access on Amazon S3, spanning genomics, satellite imagery, climate and more. Best for: projects that specifically want to practice working with genuinely large-scale data using cloud tools. Status: free to access; standard AWS data-transfer costs may still apply depending on how you consume the data, so check each dataset's specific access notes.

OpenStreetMap — a free, community-maintained global map dataset covering roads, buildings, points of interest and more, downloadable in bulk or queried through its API. Best for: geospatial and mapping projects that need real-world location data without a paid mapping API. Status: free, licensed under the Open Database License, which permits reuse with attribution and share-alike terms for derivative databases.

World Health Organization (WHO) Global Health Observatory — a free public repository of global health statistics covering disease prevalence, mortality, health system indicators and more across nearly every country. Best for: health-focused analytics or visualization projects with genuine global scope. Status: free, openly licensed for reuse.

Google Public Data / BigQuery public datasets — a collection of large public datasets (including Google Trends data, GitHub's own public repository metadata, and various government datasets) hosted for free querying through BigQuery's sandbox tier. Best for: SQL-based analytics projects on genuinely large datasets without needing to download or host the data yourself. Status: free within BigQuery's sandbox query limits; check current free-tier quota before running a very large query.

Comparing Sources by Size, Domain, and Cleanliness

Use this table as a starting filter, not a final answer — always confirm the current license and size on the source's own page before committing a project to it.

Matching a Dataset to a Specific Project Type

Building a classification model for a resume project? Start at UCI or a well-labeled Kaggle dataset with a clear binary or multi-class target column — a clean label is more important than an exotic topic for demonstrating modeling fundamentals.

Building a dashboard or BI-style visualization piece? Our World in Data and data.gov both provide large, real-world, slightly messier datasets that give you genuine cleaning and aggregation decisions to narrate, which tends to make a stronger visualization portfolio piece than an already-pristine dataset.

Building an NLP or text-classification project? Hugging Face Datasets is built specifically for this, with task-type filters (sentiment, summarization, translation) that save you from repurposing a tabular dataset for a job it wasn't designed for.

Building something with genuine geographic or time-series depth? World Bank Open Data and NASA's portal both support long historical ranges that reward a project built around change over time rather than a single snapshot.

Building a computer vision model? Start with COCO or Open Images rather than scraping your own photos, since both come with expert-quality bounding boxes and labels that would take enormous effort to replicate yourself. ImageNet remains a strong choice specifically if you want a benchmark to compare your model's accuracy against widely published results.

Building a mapping or location-based project? OpenStreetMap provides genuinely global, continuously updated location data that a static CSV dataset can't match for a project involving routing, points of interest, or geographic queries.

Building a large-scale SQL analytics project? Google's BigQuery public datasets let you practice writing real analytical SQL against genuinely large tables without downloading or hosting anything yourself, which is a different (and increasingly common) skill from the local-CSV workflow most beginner tutorials teach.

Reading a Dataset's Documentation Before You Start

Every dataset worth using ships with some form of documentation, even if it's just a short description field on a Kaggle upload page. Read it fully before writing any code. Look specifically for how missing values are represented (blank, NA, -999, a special sentinel value), what unit each numeric column uses, whether the data was collected over a specific date range that might no longer be current, and whether the source explicitly states any known quality caveats.

Skipping this step is the single most common reason a beginner's analysis produces a confidently wrong conclusion — a column silently using -1 to mean "unknown" instead of an actual value will quietly corrupt an average or a model if you don't catch it during initial exploration.

Handling Personal and Sensitive Data Responsibly

Some free datasets, especially those scraped from social platforms, health records, or public government surveys, contain personally identifiable information or sensitive attributes even when technically "public." Before building a project around any dataset involving people, check whether it has been properly anonymized, whether its source explicitly permits redistribution of derived analysis, and whether local regulations (like GDPR for EU-related data) apply to how you can process or publish results.

This matters even for a learning project you never intend to monetize. Publishing an analysis that re-identifies individuals from a "public" dataset, even accidentally through cross-referencing multiple sources, can cause real harm and reflects poorly in a portfolio review regardless of technical skill demonstrated. When in doubt, prefer datasets explicitly built for open research and teaching (UCI, Our World in Data, World Bank) over raw scraped data, since the former have generally already been vetted for these concerns.

A Note on API-Based Data vs. Static Dataset Files

Not every "free dataset" comes as a downloadable file. Sources like Google's BigQuery public datasets, OpenStreetMap, and many government portals expose data through a query interface or API rather than a single CSV or JSON download. This distinction matters for project planning: an API-based source can require handling rate limits, authentication tokens, and pagination, which adds real engineering work beyond the data analysis itself.

For a first data science project, a static downloadable file (from Kaggle or UCI) removes this extra layer of complexity and lets you focus purely on cleaning and analysis. For a later project specifically meant to demonstrate API integration skills, working against a live API-based dataset is a genuine plus, since it shows you can handle the practical realities of pulling live data rather than only working with a pre-packaged file someone else already assembled.

How to Actually Use This List

Start by deciding what kind of project you are building before searching for a dataset, not the other way around. A classification model for a resume project needs a clean, labeled dataset from UCI or a well-documented Kaggle upload. A dashboard or visualization project benefits from a larger, real-world dataset from data.gov or Our World in Data, since messier data gives you more genuine cleaning and storytelling decisions to explain. An NLP project should start at Hugging Face Datasets, where task tags make filtering fast.

Once you find a candidate dataset, spend ten minutes checking it before committing to a full project: read the license, skim the discussion or documentation for known issues, and load the first few hundred rows yourself to sanity-check column types and missing values. This upfront check saves far more time than it costs, especially for a timed portfolio sprint.

Finally, keep a short running note of which datasets you've evaluated and rejected, along with why. This sounds like overhead, but it prevents re-discovering the same overused Kaggle dataset three months later and wasting another ten minutes re-evaluating something you already decided against.

The Five Mistakes

1. Picking the most popular dataset without a plan. The most-used Kaggle datasets (a famous housing price set, a well-known titanic passenger list) are fine for learning but instantly recognizable to any reviewer — plan a distinctive angle or skip them for a portfolio piece.

2. Not checking the license before publishing derived work. Some datasets restrict commercial use or require attribution; check before putting a finished project or its outputs anywhere public.

3. Choosing a dataset too large for your hardware. A multi-gigabyte dataset that cannot load into memory on a personal laptop turns a learning project into an infrastructure problem before any analysis begins.

4. Skipping the data-quality check. Assuming a dataset is clean because it is popular or from a government source leads to broken analysis downstream — always inspect summary statistics yourself first.

5. Never documenting why you chose the dataset. A portfolio project's write-up should explain the dataset's source, its limitations, and why it fit the question you asked — this reasoning is often what a reviewer reads most closely.

6. Treating an API-based data source as identical to a static file download. Building your project plan around a live API without accounting for rate limits, authentication, or pagination can turn a straightforward analysis project into an unplanned engineering exercise partway through.

7. Skipping a privacy check on datasets involving real people. Even "public" data about individuals — scraped social posts, health survey microdata — can carry re-identification risk; verify a dataset's anonymization and redistribution terms before building anything around it, even for a non-commercial learning project.

🔗 Read next: Free APIs to build real portfolio projects with, or go back to the pillar — the complete free developer resources collection.