Appendix C: Model Card Template
Standardized Documentation for AI/ML Models
📋 About Model Cards
Model Cards are standardized documents that provide essential information about machine learning models. They were introduced by Google in 2019 and have become an industry best practice for AI transparency and accountability.
Purpose
- Provide transparency about model capabilities and limitations
- Document intended use cases and out-of-scope applications
- Report performance metrics across different demographic groups
- Communicate ethical considerations and known risks
- Enable informed decision-making by downstream users
When to Create
- Before deploying any new AI/ML model to production
- When significant changes are made to existing models
- During periodic model reviews (annually at minimum)
- As part of the AI governance approval process
Standards Alignment
This template aligns with:
- Google Model Cards: Original model card framework
- Hugging Face Model Cards: Extended format for ML community
- EU AI Act Article 11: Technical documentation requirements
- NIST AI RMF: Documentation and transparency practices
1. Model Details
Basic information identifying the model and its provenance.
1.1 Basic Information
Official name/identifier for the model
Semantic version number (major.minor.patch)
Category of machine learning approach
Technical architecture/algorithm used
Internal model registry identifier
1.2 Ownership & Dates
Team or individual responsible for the model
Individuals who developed the model
Contact for questions about the model
License for model use (internal, proprietary, open source)
1.3 Related Resources
Links to additional documentation, code repositories, etc.
How to cite this model
2. Intended Use
Description of appropriate use cases and contexts for this model.
2.1 Primary Intended Uses
Describe the main intended applications of this model
Who are the intended users of this model?
Where and how will the model be deployed?
2.2 Out-of-Scope Uses
Uses for which this model was not designed or should not be used
Potential ways the model could be misused and safeguards in place
2.3 Human Oversight Requirements
Specific scenarios that require human intervention
3. Training Data
Information about the data used to train the model.
3.1 Data Sources
List all data sources used for training
How was the training data collected?
Time period covered by training data
3.2 Data Statistics
List the most important input features
3.3 Data Privacy & Sensitivity
Types of personal data included in training
How is training data managed after model training?
3.4 Data Representativeness & Bias
Describe any known biases or limitations in the training data
Representation of different groups in training data
| Demographic Category | Group | Percentage |
|---|---|---|
4. Performance Metrics
Quantitative evaluation of model performance.
4.1 Overall Performance
How was the model evaluated?
| Metric | Value | Threshold | Notes |
|---|---|---|---|
Decision threshold used for predictions
4.2 Disaggregated Performance (Fairness Metrics)
Performance breakdown across demographic or protected groups
| Group | Sample Size | AUC-ROC | Precision | Recall | Selection Rate |
|---|---|---|---|---|---|
Ratio of selection rates between groups (should be ≥0.8 for four-fifths rule)
4.3 Operational Metrics
5. Limitations & Risks
Known limitations, potential risks, and mitigation strategies.
5.1 Known Limitations*
5.2 Risks & Mitigations
| Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|
5.3 Failure Modes
What happens when the model fails or produces low-confidence results?
6. Ethical Considerations
Ethical analysis and responsible AI compliance.
6.1 Ethical Review
6.2 Ethical Considerations
6.3 Environmental Impact
7. Maintenance & Updates
Ongoing monitoring, maintenance, and update procedures.
7.1 Monitoring Plan
| Metric | Warning Threshold | Critical Threshold | Action |
|---|---|---|---|
7.2 Update Schedule
7.3 Version History
| Version | Date | Changes | Author |
|---|---|---|---|
8. Approvals & Sign-off
Required Approvals
| Role | Name | Date | Status |
|---|---|---|---|
| Model Owner | |||
| Technical Reviewer | |||
| Data Steward | |||
| RAI Council Representative | |||
| Executive Sponsor (High-Risk only) |
📝 Model Card Best Practices
- Be specific: Avoid vague language. Quantify where possible.
- Be honest: Document limitations and failures, not just successes.
- Update regularly: Model cards are living documents. Update with each significant change.
- Consider your audience: Write for both technical and non-technical readers.
- Include context: Performance numbers without context are not meaningful.
- Test disaggregated metrics: Always evaluate across relevant demographic groups.