# Dataset Card

A dataset card records a dataset's origins, composition, collection and labeling decisions, permitted uses, and limitations for informed reuse.

- Type: artifact
- Domain: Engineering & architecture
- Level: Intermediate
- Last reviewed: 2026-09-26
- Original: [Soroush’s Product Wiki](<https://wiki.soroush.uk/topic/dataset-card/>)

## What it means

A dataset card is documentation that helps consumers judge whether a particular dataset version fits a task. Datasheets for datasets provide a closely related, more extensive documentation approach; formats differ, but both describe provenance, composition, preparation, intended uses, distribution, and maintenance. Unlike a database schema, a card explains how the data came to represent the world and where that representation is incomplete.

## Why it matters

Model behavior and evaluation credibility depend on the data used. Product teams can use dataset documentation to challenge sampling gaps, unclear labels, incompatible permissions, and claims of representativeness before investing in training or treating an evaluation score as evidence.

## When to use it

Create or review a card for training, evaluation, or shared research datasets, particularly when ownership changes or data is reused for a new purpose. It complements access and retention controls; documenting a dataset does not make otherwise impermissible reuse acceptable.

## How it works

1. Identify the dataset version, purpose, owner, source populations, collection period, and unit of observation.
2. Describe sampling, collection, consent or permissions, labeling instructions, annotator disagreement, and exclusions.
3. Document fields, missingness, transformations, splits, duplicates, and known coverage gaps.
4. Specify supported and discouraged uses, distribution conditions, access restrictions, and maintenance responsibilities.
5. Review the documentation with data producers and consumers, record unresolved questions, and revise it when the dataset changes.

## A practical example

A PM reviews a dataset card before training a ticket classifier. It reveals that closed tickets were sampled but unresolved escalations were excluded. The team adds an escalation sample and revises its evaluation plan because the original dataset understated difficult cases.

## Where it breaks

Documentation cannot recover unknown provenance or prove data quality. Categories and labels can themselves encode flawed assumptions. Detailed disclosure must avoid exposing personal information, and maintaining the card requires cooperation from data owners.

## Common mistakes

- Equating a field dictionary with complete dataset documentation.
- Claiming a sample is representative without identifying a target population.
- Assuming documentation or public availability grants permission for every reuse.
- Keeping one card while changing dataset contents, labels, or splits silently.

## Visualization

[View the visualization on the website](<https://wiki.soroush.uk/topic/dataset-card/#visualization-dataset-card>)

## Sources

1. [Datasheets for Datasets](<https://arxiv.org/abs/1803.09010>) — Communications of the ACM · Peer-reviewed research
2. [Dataset Cards](<https://huggingface.co/docs/hub/datasets-cards>) — Hugging Face · Official documentation

## Related topics

- [Model Card](<https://wiki.soroush.uk/topic/model-card/>)
- [Machine Learning \(ML\)](<https://wiki.soroush.uk/topic/machine-learning/>)
- [Research Data Management](<https://wiki.soroush.uk/topic/research-data-management/>)
