Model Versioning & Governance
Track, version, audit, and document ML models across their full lifecycle
Model Governance Fundamentals
In regulated industries (finance, healthcare, insurance), you must be able to answer: which model made this prediction? What data was it trained on? Who approved it for production? When was it last validated? Even outside regulated industries, model governance saves engineering time: you can reproduce any past prediction, safely roll back bad deployments, and audit model changes like code changes.
What to version and track
- **Code**: Git commit hash of training code - **Data**: Hash or version ID of training dataset - **Hyperparameters**: Full config used for the run - **Metrics**: Training and evaluation metrics at every epoch - **Model artifacts**: Serialized model weights and preprocessing objects - **Approval**: Who reviewed and approved the model for production - **Deployment history**: When was each version deployed and retired
Model Card Template
# Model Card: Fraud Detection v3.2
## Model Details
- **Type**: XGBoost classifier
- **Training date**: 2024-03-15
- **Author**: ML Platform Team
- **MLflow run**: runs/abc123
## Intended Use
- **Primary use**: Real-time transaction fraud scoring
- **Out-of-scope**: Account takeover, synthetic identity fraud
## Training Data
- **Source**: transactions_2022_2023 (12M rows)
- **Features**: 47 engineered features (see feature registry)
- **Label**: human-reviewed fraud labels, 0.3% positive rate
## Performance
| Metric | Value | Threshold |
|--------|-------|-----------|
| AUC-ROC | 0.934 | > 0.90 |
| F1 (threshold=0.5) | 0.891 | > 0.85 |
| FPR at 95% recall | 2.1% | < 5% |
## Limitations
- Degrades on transactions > $50,000 (sparse in training data)
- Retraining required if weekly PSI > 0.2