Skip to content
Dataset By Humans
Licensing & compliance

Datasheets for datasets: what to include, with a template

A datasheet documents where a dataset came from and how it should be used. The sections that matter, a copy-paste template and how it helps with the EU AI Act.

By Dataset By HumansPublished 1 min read

In this article

Short answer: a datasheet is a short document that travels with a dataset and answers seven questions: why it was made, what is in it, how it was collected, how it was cleaned and labelled, what it should and should not be used for, how it is distributed and who maintains it. It comes from the paper "Datasheets for Datasets" (Gebru et al.), and it is the easiest way to document training data for audits, customers and the EU AI Act.

Why datasheets matter now

Buyers of AI systems increasingly ask where the training data came from. The EU AI Act requires providers of general-purpose AI models to publish a summary of their training content, and requires data governance for high-risk systems. A datasheet for every dataset you use turns those requests into copy-and-paste work instead of an investigation.

The seven sections

SectionQuestions it answers
MotivationWhy was the dataset created? Who funded it?
CompositionWhat do records contain? How many? Any personal or sensitive data?
Collection processWho collected it, how, when, where, with what consent?
Preprocessing and labellingWhat was removed, blurred or normalised? Who labelled it?
UsesWhat is it suited for? What should it not be used for?
DistributionUnder which license? How is it delivered?
MaintenanceWho maintains it? How are errors and takedowns handled?

Copy-paste template

# Datasheet: <dataset name> v<version>

## Motivation
- Purpose:
- Created by / funded by:

## Composition
- Record type and count:
- Fields (see SCHEMA.md):
- Languages / regions:
- Personal data: <none | blurred | invented | consented>
- Known gaps and biases:

## Collection process
- Who created the records (roles, not names):
- Method and devices:
- Period and locations (region level):
- Consent and contributor agreements:

## Preprocessing and labelling
- Removed: <duplicates, corrupted files, ...>
- Privacy processing: <face and plate blurring, GPS removal, ...>
- Labels and who produced them:
- Any AI assistance used:
- Quality checks and review rate:

## Uses
- Intended uses:
- Out-of-scope uses:
- Train/test guidance:

## Distribution
- License:
- Delivery format and checksums:

## Maintenance
- Maintainer and contact:
- Takedown and correction process:
- Version history:

Tips for writing one that people read

  • Be specific. "1,240 photos taken on 14 phone models between March and June 2026 in Hanoi and Ho Chi Minh City" beats "diverse images".
  • State the gaps. Saying what the data does not cover builds more trust than leaving it out.
  • Disclose AI assistance. If a model helped with labels or captions, say where.
  • Version it. Update the datasheet whenever records are added or removed.

Every dataset we deliver includes a DATASHEET.md in this structure, plus the license and checksums. See how we make and check every dataset.

datasheet for datasetsdataset documentationeu ai actdata provenance