Datasheet for Datasets

Explicitly states why the dataset was created and its intended tasks. This section answers the fundamental question: 'What gap does this fill?'

• Primary Goal: The specific machine learning task (e.g., object detection, sentiment analysis).
• Funding Sources: Discloses the grant or commercial entity that financed the creation.
• Creator Narrative: A prose summary explaining the need the dataset addresses.

Documents the granular makeup of instances and the lineage of the raw data. This is the 'bill of materials' for the dataset.

• Instance Representation: Defines what a single record is (e.g., an image, a text snippet, a time-series segment).
• Data Sources: Enumerates original sources (e.g., crowdsourcing, web scraping, sensor logs).
• Demographic Breakdown: If the dataset contains people, this details age, gender, and geographic distributions to surface representation gaps.

Describes the mechanism of acquisition with enough detail to assess reliability and potential biases introduced during gathering.

• Acquisition Method: The specific protocol (e.g., mechanical turk task design, API crawler frequency).
• Ethical Review: States whether an Institutional Review Board (IRB) or ethics committee approved the collection.
• Quality Control: Describes validation steps, such as inter-annotator agreement metrics or outlier removal thresholds.

Details the transformations applied to raw data before publication. This is critical for reproducibility.

• Filtering Logic: Criteria for removing instances (e.g., removing non-English text, blurring faces).
• Normalization: Scaling, tokenization, or encoding changes applied.
• Missing Data Strategy: How NULL values or corrupted files were handled (imputation vs. deletion).

Declares safe operational boundaries and warns against dangerous misuse. This section is a direct guardrail for downstream engineers.

• Recommended Uses: The validated, safe applications the dataset supports.
• Out-of-Scope Uses: Explicitly forbids high-risk applications (e.g., 'Not for clinical diagnosis').
• Bias Potential: A candid discussion of known skews that could amplify societal harms if unmitigated.

Covers the legal and logistical aspects of access and the long-term support plan.

• Licensing: The specific legal agreement (e.g., CC-BY-4.0, custom research-only license).
• Access Mechanism: How to obtain it (direct download, API, application process).
• Versioning Protocol: The plan for updates, errata, and deprecation to ensure downstream stability.

A structured transparency document detailing a machine learning model's intended use, evaluation results, and known limitations. While a datasheet documents the training data, a model card documents the trained artifact itself. Standardized by Google Research in 2018, model cards typically include:

• Quantitative performance metrics across disaggregated subgroups
• Ethical considerations and fairness evaluations
• Recommended out-of-scope use cases

A holistic transparency artifact that documents the safety evaluation and operational context of an entire AI system, not just its model or data. System cards extend beyond model cards by covering:

• User interface design and interaction modalities
• Downstream effects and sociotechnical context
• End-to-end system-level testing results
• Human oversight mechanisms and intervention protocols

The complete, verifiable lineage of a dataset tracking its origin, transformation steps, and chain of custody. Key components include:

• Source documentation (original data collection methodology)
• Processing pipeline logs (cleaning, augmentation, filtering)
• Consent and licensing metadata
• Cryptographic hashes for integrity verification Datasheets operationalize provenance by making it human-readable for downstream consumers.

A formal, machine-readable inventory detailing the complete supply chain of an AI system. An AI BOM extends the software BOM concept to include:

• Model architecture and weight provenance
• Training dataset identifiers and datasheet references
• Software dependencies and compute infrastructure
• Ethical review status and conformity assessments AI BOMs provide the structured metadata that datasheets and model cards surface in narrative form.

A method for tracing a model's specific prediction or behavior back to the individual data points or subsets within the training corpus that most influenced it. Datasheets support attribution by documenting:

• Data collection methodologies and sampling strategies
• Known biases and representational gaps
• Preprocessing and augmentation steps This enables practitioners to investigate model failures by interrogating the underlying data sources.

A regulatory remedy requiring an organization to delete a trained model or its associated data products when they were developed using unlawfully collected or improperly processed data. Datasheets serve as a critical defense mechanism by:

• Documenting consent and licensing status upfront
• Establishing a verifiable paper trail of data provenance
• Enabling proactive compliance before training begins This concept is increasingly relevant under GDPR and the EU AI Act.