Dataset documentation and datasheets

A dataset without documentation is a risk: nobody knows where it came from, what may be done with it, or who it represents badly.

A datasheet — seven headings, following Gebru et al.:

1. Motivation — why was it created, by whom, funded by whom?
2. Composition — what is an instance, how many are there, which fields, is anything missing?
3. Collection — how was it collected, when, by whom, with what consent?
4. Preprocessing — what has been cleaned, filtered, anonymised? Does the raw data still exist?
5. Uses — what has it been used for, what does it not suit?
6. Distribution — the licence, where it is, any restrictions.
7. Maintenance — who is responsible, how are errors reported, how is it versioned?

The heading most often missing — and the most important: «what the dataset does not suit».

Examples of honest limitations:

Collected 2023–2024 in the Stockholm area, 78 % of the annotators were under 30. Spoken forms and dialects from outside the Mälardalen region are underrepresented. The dataset is not to be used for assessing individuals, and not for medical decisions.

That text prevents more misuse than any technical measure.

For personal data there is more: the legal basis, the purpose, the retention period, how deletion is handled (and what happens to a model already trained on the data), and whether the dataset is shared outside the EU/EEA.

Model cards are the equivalent for models: the intended use, performance broken down by group, known limitations, ethical considerations. AI-grafen has its equivalent register in content/sources/ with a licence, a fingerprint and a responsible editor per source — the same idea, machine-readable.