Documentation and datasheets
Structured docs for datasets and models—**datasheets**, model cards, and system cards—so others know intended use, limits, and risks.
What it is
Structured docs for datasets and models—datasheets, model cards, and system cards—so others know intended use, limits, and risks.
Why it matters
Undocumented datasets become radioactive over time. Cards are how teams communicate honestly.
How it works (plain)
Answer: motivation, composition, collection, preprocessing, uses, distribution, maintenance, and known harms. Keep docs versioned next to the data.
Everyday example
A food label won’t make food healthy—but hiding ingredients is worse.
Try it
Fill a one-page datasheet for a tiny CSV you own.
Myths
- ⚠️ Myth: Cards are only for paper releases.
- ✓ Reality: Internal cards prevent foot-guns across teams.
- ⚠️ Myth: Marketing blurbs count as model cards.
- ✓ Reality: Cards need limits and eval details.
Sources
- Datasheets for Datasets (Gebru et al.) — cite PDF/arXiv when teaching
- Model Cards (Mitchell et al.)
- NIST AI RMF: https://www.nist.gov/itl/ai-risk-management-framework ↗