Model Cards and Datasheets

Purpose

Model cards and datasheets are standardized documentation formats that make AI systems more transparent and reviewable. They were introduced by Google and Microsoft researchers and have become widely adopted as a governance best practice.

Model Cards

A model card describes a trained model's intended use, performance characteristics, and limitations.

Model Card Template

Section	Contents
Model name and version	Identifier and version number
Model type	Architecture and technique (e.g., transformer, gradient boosting)
Intended use	Primary use cases the model was designed for
Out-of-scope use	Use cases the model should not be used for
Training data	Description of training data (source, size, date range, preprocessing)
Evaluation data	Description of evaluation data and how it differs from training data
Performance metrics	Key metrics (accuracy, precision, recall, F1, AUC, etc.) with confidence intervals
Disaggregated performance	Performance broken down by relevant demographic or contextual groups
Limitations	Known limitations, failure modes, and edge cases
Ethical considerations	Potential harms, biases, and sensitive use contexts
Recommendations	Guidance for users on responsible deployment

Datasheets for Datasets

A datasheet describes a dataset's provenance, composition, and intended use.

Datasheet Template

Section	Contents
Dataset name and version	Identifier and version
Purpose	Why was this dataset created?
Creator	Who created the dataset and on behalf of whom?
Composition	What data is included? How many instances? What features?
Collection process	How was the data collected? What consent was obtained?
Preprocessing	What cleaning, filtering, or transformation was applied?
Intended use	What tasks is this dataset appropriate for?
Not appropriate for	What tasks should this dataset not be used for?
Distribution	How is the dataset shared? Under what license?
Maintenance	Who maintains the dataset? How are updates handled?
Legal and ethical	Are there legal or ethical concerns? (Privacy, bias, representation)

When to Require These

Artifact	When Required
Model card	All Tier 3 cases. Recommended for Tier 2.
Datasheet	When the model uses a custom or curated dataset. Not required for off-the-shelf vendor models (covered by vendor checklist).

For vendor-procured AI, request equivalent documentation from the vendor as part of the Vendor Checklist.