Where to Find Free Datasets to Practice With

1. What makes a dataset worth practising on

Before the explanation: think of a topic you would enjoy analysing. What would one row of that data have to contain for you to be able to say anything at all about it?

A dataset teaches you something when it can carry a question. Three column types are what let it do that.

A date. Without one you cannot ask whether anything changed, which rules out trends, seasonality, growth rates and before-and-after comparisons. That is a large fraction of what analysts are actually paid for.

A category. A region, a product, a department, a species. This is what you group by, and grouping is the move that turns a pile of rows into a comparison. One category column is enough; two lets you cross them.

A number. Something worth summing or averaging: a price, a count, a duration, a temperature. An identifier is not this, even though it is made of digits, and summing an ID column is a mistake with its own guide.

Two more things raise a dataset from usable to good. Enough rows that a group is not tiny, which in practice means thousands rather than dozens, so an average means something. And a second table you could join to, because joining is the skill that separates spreadsheet work from database work.

2. The places worth your time

There are hundreds of lists of dataset sources. Most of the entries are dead, duplicated, or a single file somebody posted once. These are the ones that repay a bookmark.

Notice what the table is really sorted by. The further up you go, the messier and more real the data, and the more it resembles what lands in your inbox on a Tuesday. The further down, the cleaner and more convenient. Both have a place, and knowing which you are picking is the point.

3. Government open data, and why it is the closest thing to real work

Open government data has a quality that curated collections cannot fake: nobody tidied it for a learner. Column names are inconsistent, dates arrive in three formats, categories have been renamed halfway through the series, and a footnote in a separate PDF explains that 2019 is not comparable to 2020.

That sounds like a reason to avoid it. It is the reason to use it. Cleaning, reconciling and documenting are most of an analyst's day, and they cannot be practised on a file where somebody already did them. A portfolio project built on messy public data lets you show the part of the work that actually distinguishes people, which is judgement about what the data can and cannot support. Documenting data limitations is the write-up half of that skill.

One practical tip for these portals. Look for a data dictionary, sometimes called a codebook or a schema, before you download anything. It is the document explaining what each column means and how it was collected. A dataset with one is worth three without, because the alternative is guessing what STATUS_CD = 4 meant to whoever typed it.

4. Kaggle, and the one thing to check there

Kaggle holds an enormous searchable collection and is usually the fastest way to find data on a named subject. You need a free account to download, and the search is far better than most government portals.

The thing to check is where a dataset came from. Anyone can upload, so a Kaggle page can be an original release, a copy of a government file, a copy of somebody else's copy, or a synthetic file somebody generated. Those are very different things to build a portfolio on, and the page does not always make the difference obvious.

Two signals sit on every dataset page and take ten seconds to read. The provenance or source section should name where the data actually came from, ideally with a link you can follow to the original. The licence tells you whether you may republish it, which matters the moment your project goes on the public web. If a page names no source at all, treat it as an exercise file rather than evidence, and do not build a portfolio claim on it.

Say why a synthetic dataset would be a poor choice for a portfolio project before reading on. The reason is that your findings would be about somebody's random number generator rather than about the world, and a reviewer who recognises the file knows that. Practising your technique on it is fine. Presenting a conclusion from it is not.

5. The famous teaching datasets, and what they can and cannot give you

A handful of files appear in nearly every tutorial: the Titanic passenger list, the iris flower measurements, the tips dataset, the Boston housing file. They are famous because they are small, clean, and load in one line, which makes them excellent for demonstrating a function.

What they cannot give you is a project. They are a few hundred rows with no date column, every question anyone might ask has been asked and published thousands of times, and a reviewer has seen the same analysis many times before yours. Using them to learn groupby is efficient. Putting them in a portfolio spends the one slot you have on something that shows nothing about your judgement.

So use them deliberately, for the ten minutes they are good at, and get your project data from section 2. There is a straightforward test: if a search for the dataset name returns a thousand finished analyses, it is a teaching file rather than a portfolio subject.

6. The five-minute screen

Run this before committing to any dataset. Every step is one action and the whole thing fits in five minutes, which is a good trade against a week.

1. Open it and read the column names. Can you find a date, a category and a number? If not, stop here.
2. Count the rows. Under a thousand and most groups will be too small for an average to mean anything.
3. Look for the data dictionary. If there is none, can you honestly say what each column means? Guessing is how a confident wrong finding gets made.
4. Sort each important column and look at both ends. The largest and smallest values expose placeholder numbers, negative quantities, and dates in the year 1900.
5. Count the blanks in your key columns. A category column that is a third empty will quietly drop a third of your rows from every grouped result.
6. Write the question you would answer with it, in one sentence. If you cannot, the dataset is not the problem, but you do not have a project yet.

That last step is the one people skip and it is the one that decides the outcome. A dataset is not a project. Turning a fuzzy question into a measurable definition is the step between them.

Now picture the last file somebody sent you at work. Which of those six checks would it have failed? That is usually the same check that cost you an afternoon later on.

What goes wrong, and the fix

Six that cost people the most time.

The download is an Excel file with the real table starting on row 8. Government portals do this constantly: title, logo, notes, blank row, then the headers. Skip those rows on import rather than deleting them by hand, so the step is repeatable when the file is updated.

The numbers are text. Thousands separators, currency symbols, footnote markers and a stray "N/A" all turn a numeric column into text. Fix it at the import step, and check the total afterwards.

Leading zeros have vanished from zip codes or account numbers. Opening a CSV by double-clicking it in Excel does this silently. The CSV import fix covers the correct route.

The file is far bigger than the page said. Compressed downloads describe their compressed size. A 200 MB download can be several gigabytes on disk. Handling large datasets is the guide for when that happens.

The categories changed partway through the series. A region renamed in 2022 becomes two categories in a grouped result, each covering part of the period. Always list the distinct values of a category column before trusting a grouping, and when the duplicates are messier than a rename, that is entity resolution.

The licence does not allow republishing. Check before your project goes public, not after. Government open data is usually fine; a scraped copy of a commercial source usually is not.