Data Card

Documents the origin and transformation history of the data. This section answers the question: Where did this data come from?

• Primary Source: The original collector or generating process (e.g., sensor telemetry, user surveys, public records).
• Collection Methodology: The specific mechanism used to gather raw observations, including timeframes and sampling strategies.
• Transformation Graph: A record of all cleaning, normalization, and aggregation steps applied to the raw source before publication.
• Version History: Tracks iterative changes to the dataset itself, linking each version to specific preprocessing code commits.

Describes the internal structure and statistical makeup of the data. This section answers: What is actually in this dataset?

• Feature Descriptions: Semantic definitions, data types (int, float, categorical), and units for every column.
• Missing Value Strategy: Explicitly states the reason for null values (e.g., sensor failure, non-response) and the imputation method used.
• Class Balance: Reports the distribution of labels or categorical variables to warn of skew and potential bias.
• Temporal & Spatial Coverage: Defines the exact time range and geographic boundaries the data represents.

Identifies sensitive attributes and documents the risk mitigation steps taken. This section answers: Can this data harm individuals?

• PII Identification: Flags columns containing direct identifiers (names, emails) or quasi-identifiers (zip code, DOB) that enable re-identification.
• Consent Mechanism: Describes the legal basis for data usage and the notice provided to data subjects.
• Fairness Audit: Reports on demographic parity and representation gaps across protected subgroups.
• Risk Mitigation: Lists applied techniques such as k-anonymity, differential privacy noise injection, or redaction.

Defines the safe operating envelope for the dataset. This section answers: What should I do with this data?

• Suitable Tasks: Enumerates validated use cases (e.g., classification, object detection, trend forecasting).
• Out-of-Scope Applications: Explicitly prohibits dangerous or unethical uses the data was not designed for.
• Known Failure Modes: Documents edge cases where the data distribution breaks down or becomes unreliable.
• Benchmark Performance: Provides baseline model metrics to set realistic expectations for downstream utility.

Clarifies the lifecycle management and access protocols. This section answers: How do I get it and will it change?

• Access Protocol: Specifies the distribution mechanism, such as direct download, API access, or secure enclave querying.
• Licensing: Defines the legal terms of use, including attribution requirements and commercial restrictions.
• Deprecation Policy: States the conditions under which the dataset will be sunset or superseded.
• Contact & Citation: Provides a persistent identifier (DOI) and a point of contact for errata or questions.

Data Cards complement Model Cards, which document trained models. While a Model Card reports a model's accuracy and fairness metrics, a Data Card explains the upstream data that caused those metrics.

• Dependency Chain: A model inherits the biases and gaps of its training data.
• Joint Auditing: Responsible AI requires reviewing the Data Card and the Model Card together to trace errors to their root cause.
• Standard Alignment: Both formats align with frameworks like the About ML annotation standard to ensure interoperability.