AI glossary · Basic · Data & evaluation · Safety & alignment
Dataset card
A dataset card is a document that describes where a dataset came from, what it contains, and what it may and may not be used for. It is a 'user manual and ingredients label' attached to data.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Foundations & applications
- Added
- 2026-10-09
How it works
When a dataset is released publicly or shared within a team, the document is written by answering a set list of questions. It usually covers the purpose of collection, sources and collection methods, data composition (number of items, languages, label types), preprocessing steps, known biases and limitations, license, how personal information is handled, and recommended and discouraged uses. Such question-based documents are sometimes called 'Datasheets', and ones organized as cards 'Data Cards'. Data-sharing platforms such as Hugging Face encourage uploading such a card with each dataset repository.
What goes in a dataset card
Source and composition
- Purpose and sources of collection
- Collection and preprocessing methods
- Number of items, languages, label types
Responsible use
- Known biases and limitations
- License
- Personal information handling
- Recommended and discouraged uses
Why it matters
AI models take on the character of the data they are trained on. If the data leans toward a particular language or region, the model leans the same way. But when the people who build the data are different from those who use it, these circumstances are not easily passed along. A dataset card bridges that gap, letting users judge whether the data fits before choosing it. As copyright and privacy issues have grown, it has also become important as material explaining where data came from.
Things to know
Having a card does not mean the data is clean. Because it is a self-written document, parts may be left out or written optimistically. It is also common for the data to change while the card stays the same, passing on outdated information. A similar document attached to models is the 'model card', and you need to look at both to get a picture of what a model learned.
Example
When you open a public dataset page on Hugging Face, the description document on the first screen is the dataset card. There you check which websites or documents it was collected from, under which license it can be used, and which languages it contains, then decide whether you can use it in your project. Companies often keep a similar template in their internal data catalog and require a card to be written when new training data is registered.