How to Read a Dataset Card

A dataset card documents a dataset. Reading it first saves hours of confusion and helps you avoid misusing data.

Questions to Answer

1. What is one row? A person, a transaction, a day, an image?
2. Where did it come from, and when? Source, collection method and time period.
3. What does each column mean, and in what units?
4. How are missing values recorded? Blanks, NA, ?, -999?
5. How was it labelled, and by whom?
6. What are the known quirks? Header rows, delimiters, encoding, duplicates.
7. Who is represented — and who isn't?
8. What's the licence? Can you use it commercially? Must you cite it?
9. Is it maintained? One-off snapshot or regularly updated?

Look at the Data Too

Documentation is never complete. Preview the first rows — the glitchdata data viewer does this — and check column types, ranges and missing values yourself.

Warning Signs

No licence, no source, no description of collection, vague column names, or claims of being "complete" or "unbiased". Treat such data with caution.

Record What You Use

Note the dataset, version, date retrieved and licence in your project documentation, so results can be reproduced and credited correctly.