Data Cards: The Label on Your Dataset

Imagine if food packaging had no ingredient list. No allergen warnings. No source. That was the state of datasets for decades. In 2018, Timnit Gebru and colleagues published Datasheets for Datasets, arguing that every dataset should ship with structured documentation.

What goes in a data card

1. 1Motivation: why was this dataset created? By whom?
2. 2Composition: what does each row represent? How many rows?
3. 3Collection process: how was it gathered? Was consent obtained?
4. 4Preprocessing: what was cleaned, filtered, or relabeled?
5. 5Uses: what is it good for? What should it NOT be used for?
6. 6Distribution: how can people access it? What is the license?
7. 7Maintenance: who updates it? How do users report issues?

Real examples you can read

• Every Hugging Face dataset has a README.md that functions as a data card
• Google's Data Cards Playbook provides templates for responsible release
• The FineWeb dataset publishes a detailed card describing filtering decisions
• ImageNet's retrospective data card was added years after release to document known biases

A minimal data card template

--- dataset_name: teen_math_homework_2026 version: 1.0 creators: - name: Tendril content team - contact: data@tendril.neural-forge.io license: CC-BY-4.0 languages: [en] size: rows: 12400 bytes: 45_000_000 collection: method: Scraped from public Khan Academy forums date_range: 2022-01 through 2024-12 consent: Public posts; PII removed intended_uses: - Fine-tuning LLMs for math tutoring - Research on student reasoning patterns out_of_scope: - Identifying or de-anonymizing students - Commercial tutoring without human oversight known_biases: - Skews toward US English - Over-represents algebra, under-represents geometry update_schedule: Annual ---

The big idea: a dataset without a data card is a dataset you cannot trust, audit, or use responsibly. Writing data cards is the baseline hygiene of modern ML.