Self-assessment

Select the statement that best describes the team today, one per chapter. The result shows a level for each chapter and the chapters to read first. Answers are stored in this browser only.

1. Purpose Which datasets are published for machine learning?

Not yetNone, or extracts shared on request without a stated purpose.FoundationalA statement of purpose exists and one dataset from a coding or transcription task is published with a decision record.AI-readyDatasets for the main model-assisted tasks are selected by a written rule, each with uses, owner, and approvals recorded.AI-nativeDatasets are produced as a stage of production at each round.

2. Form What form do the files take?

Not yetWhatever the source system exports.FoundationalTidy files with a dictionary that names the label, the features, the identifier, and the group key.AI-readyOpen formats with stable column names across versions, the role of every column recorded, and leakage between features and label checked.AI-nativeFiles and dictionary generated from production with the checks as a pipeline stage.

3. Croissant How is a dataset described for loading?

Not yetIt is not; users read the files.FoundationalA Croissant record is generated from the dictionary and checked.AI-readyThe record carries the splits, passes the reference validator, and is published on the page and in the catalog.AI-nativeThe record is generated and validated at each release and registered in a public repository.

4. Card How is a dataset documented?

Not yetA readme, or nothing.FoundationalA card with the twelve sections and a machine-readable header, checked by script.AI-readyThe representativeness section comes from the report, the card is reviewed by a specialist, and it is published everywhere the dataset is.AI-nativeThe measured sections are regenerated at each release and the written sections are versioned with the data.

5. Representativeness How is coverage measured?

Not yetIt is not; the source survey is assumed representative.FoundationalThe composition is compared with census benchmarks on three or four variables and stated in the card.AI-readyCross-tabulations, thin groups, and label coverage are reported with a minimum count, at each release.AI-nativeA reference model's error by group is published with the dataset.

6. Splits How is the dataset split?

Not yetUsers split it themselves, or records are split at random.FoundationalSplit by group with a recorded seed; the assignment file is published and the leakage check runs.AI-readyTime-based test parts where rounds exist; group distributions per part reported; the test part frozen under the version policy.AI-nativeNew rounds join the training part at release and the test part moves on a stated schedule, with a reference model's scores per version.

7. Licence Under what terms is a dataset released?

Not yetUnstated, or the licence of the source file by default.FoundationalAn open licence chosen on purpose, terms that permit training and ask for attribution, and a recorded disclosure review.AI-readyThe licence URL is in the Croissant record and the catalog record, free text is screened as a release check, and the attribution form for model cards is stated.AI-nativeModels that cite the dataset are tracked and the terms are reviewed yearly.

8. Versions How are versions and integrity maintained?

Not yetFiles are updated in place.FoundationalA version policy, a manifest of checksums verified before publication, and a DOI.AI-readyThe card and the record are versioned with the data, superseded versions stay available, and downloads and citations are counted.AI-nativeReleases are a pipeline stage with manifest, record, card, and DOI generated together, and model cards that cite the dataset are tracked.

0 of 8 answered