Create a Model Card¶
A model card is like a nutrition label for your machine learning model. It provides transparency.
What goes in a Model Card?¶
- Model Details: Algorithm type (e.g., Random Forest), date created, developer, version.
- Intended Use: What is this built for? (e.g., "Predicting default risk on unsecured personal loans.")
- Out of Scope: What should it NOT be used for? (e.g., "Not intended for business loans or mortgages.")
- Metrics: Performance across different slices (e.g., Accuracy is 88% overall, but 85% for group A and 90% for group B). Disclosing these differences is crucial for fairness.
- Training Data: A brief overview of the dataset. "Trained on 50,000 anonymised loan applications from 2018-2022."
- Ethical Considerations: Known limitations or potential biases. "Historically, younger applicants have fewer data points, leading to higher false-positive rates in the 18-21 age bracket."
KSB Mapping¶
| KSB | Description | How This Addresses It |
|---|---|---|
| S5 | Deployment, value assessment, and ROI | Translating model performance into business impact |
| S6 | Communicate through storytelling and visualisation | Presenting ML results to non-technical stakeholders |
| B4 | Consideration of organisational goals | Framing technical results in terms of business objectives |
| B1 | Inquisitive approach | Exploring creative ways to explain model behaviour |