Data-Lab / templates /dataset-card-template.md
KrazyKitty's picture
Create DATA LAB private review package
a490f0e verified
|
Raw History Blame Contribute Delete
2.55 kB
metadata
license: <reviewed-spdx-identifier-or-other>
pretty_name: <dataset-display-name>
language:
  - <language-tag>
tags:
  - <purpose-tag>

Dataset Card for <dataset-id>

Template only. Replace every angle-bracket field and remove this notice before release.

Summary

  • Version: <immutable-version>
  • Status: <proposed | approved-for-private-research | released | withdrawn>
  • Maintainer: <responsible-owner>
  • Canonical source manifest: <path-or-immutable-link-and-sha256>
  • Release decision: <approval-record>

Describe what the dataset contains and what evidence supports that description. Do not promote a planned collection or processing pipeline into a released dataset claim.

Intended uses

Supported

  • <specific, evidence-backed research use>

Out of scope

  • <use not evaluated or not permitted>

Source data and provenance

For every source, provide the publisher, canonical URL, version/revision, retrieval date, integrity digest, licence, permission scope, privacy review, consent or lawful basis, and transformation lineage. Link to a machine-readable manifest without exposing private paths or source content.

Processing

Record the exact code revision, configuration, command, dependency lock, input/output hashes, filter versions, exclusion counts, deduplication method, and split assignment.

Composition and splits

Report records, bytes, characters, tokens (with tokenizer version), languages, domains, time range, formats, source mixture, missing values, duplicates, and rejection counts for every split.

Licensing, privacy, and consent

Explain the licence analysis, attribution, redistribution rights, privacy controls, sensitive-data review, consent or other lawful basis, retention, withdrawal, and deletion process.

Quality and validation

Provide scoped measurements, manual-review sampling, contamination checks, known blind spots, and unresolved issues. Avoid unsupported labels such as clean, safe, or representative.

Limitations and risks

Document sampling bias, detector limitations, temporal limits, language/domain gaps, harmful content, privacy residuals, contamination risk, and inappropriate downstream uses.

Reproducibility

Provide immutable source, code, configuration, environment, and artifact identifiers sufficient for an authorised reviewer to reproduce the release.

Changes and withdrawal

Maintain an append-only version history with corrections, removals, licence changes, and downstream notification requirements.