Glossary · AI engineering and governance
Model versioning
Also known as: Model version control
German: Modellversionierung
In machine learning operations, model versioning is the practice of uniquely identifying and storing each version of a model together with the code, data, configuration and validation results that produced it, so that any deployed model can be traced, reproduced, compared and rolled back.
- Industrial AI
- AI
In one sentence
Model versioning identifies and stores each model version with its code, data and settings so it can be traced, reproduced and rolled back.
Example
The edge device reports that it runs model 3.2.1, which the registry links to the training data snapshot of May, the training script commit and the validation report.
How it applies
- Engineering: A model version is more than a file of weights. Store it with the training data reference, preprocessing, hyperparameters, software dependencies and validation results, typically in a model registry.
- Operation: Know which version runs on which device or line. This is essential for investigating incidents and for Rollback capability.
- Documentation: Include model versions in the plant's or product's Configuration baseline and release notes. Documentation that describes AI behavior should name the model version or version range it applies to.
Model versioning vs. software versioning
Software versioning tracks code changes. Model versioning also has to track data: retraining the same code on new data produces a different model with potentially different behavior. Both need to be linked, so that a model version can be rebuilt exactly.