Datasets
Curated, annotated collections of data for model training/testing.
How the laws define this
Regulators consistently treat datasets as structured or curated collections of data used to train, validate, test, or operationalize artificial intelligence models, requiring metadata and provenance information to support governance. Frameworks diverge significantly regarding the specific obligations and operational contexts attached to these collections. The National Strategy in the field of Artificial Intelligence 2024-2027 conditions dataset reuse on FAIR principles and GDPR privacy safeguards. Conversely, the City of Helsinki AI Register focuses strictly on documenting data sources within municipal register entries to enable risk assessment. Positioned as managed assets, datasets in the Concept for the Development of Artificial Intelligence for 2024–2029 are defined as state and partner data subject to quality rules and quantitative delivery targets. Meanwhile, the National Artificial Intelligence Strategy — Working Document for Citizen Participation 2021–2026 emphasizes quality controls and privacy safeguards to ensure reliability and lawful use across AI development and deployment.
Synthesised from the 4 statutory definitions below. Each one is quoted in full, with its source.
How jurisdictions differ
What is distinctive in each definition, binding law in force first. Showing 3 of the 4 instruments that define this term; all are listed under Definitions below.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| Kazakhstan | Concept for the Development of Artificial Intelligence for 2024–2029 | In Force | Adds description of datasets as state/partner data assets maintained to unified formats and quality standards, catalogued for sharing, subject to metadata rules and access arrangements, with quantitative targets for cumulative delivery. |
| Helsinki, Finland | City of Helsinki AI Register | In Force | Limits to collection of training, reference and operational data used by an AI or algorithmic system, including provenance information and descriptions of data sources, to be documented in the register entry to support governance and risk assessment. |
| Peru | National Artificial Intelligence Strategy — Working Document for Citizen Participation 2021–2026 | Draft | Adds description of curated collections of structured or unstructured data used to train, validate or run models, for which metadata, quality controls, provenance and privacy safeguards are described to ensure reliability and lawful use in AI development and deployment. |
What it triggers
- In Helsinki, Finland, the AI Register documents datasets with provenance and descriptions to support governance and risk assessment.source
- In Kazakhstan, datasets are maintained to unified formats and quality standards, catalogued for sharing as state and partner assets with metadata, quality rules and access arrangements.source
Definitions (4)
Curated, annotated collections of data made available for model training, testing and evaluation, maintained to unified formats and quality standards and catalogued for sharing. The Concept defines these as state and partner data assets subject to metadata, quality rules and access arrangements and sets quantitative targets for cumulative dataset delivery.
Curated collections of structured or unstructured data used to train, validate or run models, for which metadata, quality controls, provenance and privacy safeguards are required to ensure reliability and lawful use in AI development and deployment.
The collection of training, reference and operational data used by an AI or algorithmic system, including provenance information and any description of data sources, which must be documented in the register entry to support governance and risk assessment.
Structured collections of data, including associated metadata and provenance, used for training, validation and testing of AI systems; managed to meet FAIR principles and privacy/GDPR safeguards for reuse in research and public-sector deployments.
Not to be confused with
Library of data refers to a structured repository for datasets, while datasets are the collections of data themselves used for model training.
AI-ready data emphasizes data specifically prepared and governed for AI development, whereas datasets more broadly cover curated collections for training with varying governance.
Questions and answers
What documentation is required for datasets in Helsinki's AI systems?
The City of Helsinki AI Register documents training, reference, and operational data, including provenance and descriptions of sources. This supports governance and risk assessment for public AI systems.
What standards apply to datasets under Kazakhstan's AI Concept?
Datasets are maintained to unified formats and quality standards, catalogued for sharing, and subject to metadata, quality rules, and access arrangements as state and partner data assets.
What safeguards are needed for datasets in Peru's AI strategy draft?
Datasets are described with metadata, quality controls, provenance, and privacy safeguards to ensure reliability and lawful use in AI development and deployment.
Related Terms
dataset
Structured collections used for training and validation of models....
Data
Public and private data assets, emphasizing openness and interoperability....
Library of data
Structured dataset repository used for training models....
Public Data Sources
Administrative and publicly available datasets facilitated by the sandbox for AI application development....
AI-ready data
Data prepared and governed for use in AI development....