Why document a dataset
Months later, someone will ask what a model was trained on: a customer, an auditor, a regulator or your own team. A short datasheet written while you build the dataset answers that in minutes instead of weeks.
What to include
- Purpose: what the data is for and what it should represent.
- Sources: where each part came from, when, and under what rights.
- Collection: how it was gathered and by whom.
- Cleaning: which checks ran, how many files each flagged, and what was decided.
- Labelling: guidelines, who labelled, agreement between labellers.
- Splits: how training, validation and test data were separated.
- Personal data: what was found, legal basis, how it was handled.
- Known gaps: settings, groups or classes that are missing or rare.
- Version: a date or hash, and which models used it.
Keep it next to the data
Store the datasheet with the dataset version, and update it when the data changes. An outdated datasheet is worse than none.
Regulation
For high-risk AI systems, Article 10 of the EU AI Act asks providers for data governance covering many of these points. See Article 10 and your training data.