Dataset Datasheet Skill

Models inherit the flaws of their data, and most data debt is invisible because nobody wrote down where the data came from. A datasheet is that record: how the dataset was collected, what's in it, what's missing, and what it should not be used for. It's the difference between a reusable asset and a liability.

Required Inputs

Ask for these only if they aren't already provided:

• Dataset name, version, owner and what it's used for today.
• Motivation — why it was created and for what task.
• Composition — what an instance is, how many, fields/labels, and time range.
• Collection — sources, method (scraped, logged, purchased, annotated), and consent/licensing basis.
• Known issues — gaps, imbalances, label noise, sensitive attributes, duplicates.

Output Format

Datasheet: [dataset] v[version]

Owner: [team] · Created: [date] · License: [license]

1. Motivation — why this dataset exists, the task it serves, and who funded/created it.

2. Composition

• What a single instance represents; total count; the schema (fields, label definitions).
• Class/label balance and key distributions (and notable skews).
• Sensitive attributes present (directly or by proxy), and whether individuals are identifiable.
• Known missing data, duplicates, or noise.

3. Collection process — sources, mechanism (scrape/log/survey/annotation), time window, sampling strategy, and the legal/consent basis (license, ToS, opt-in).

4. Preprocessing / labelling — cleaning, dedup, filtering, and how labels were produced (who annotated, guidelines, inter-annotator agreement).

5. Recommended uses & limits

• Appropriate uses: tasks this data supports well.
• Do not use for: tasks where its biases/gaps would cause harm or invalid results.