Making a model runnable.
WHAT MUST BE PACKAGED
The model artefact The preprocessing applied to inputs The postprocessing applied to outputs The dependency versions Metadata identifying it
WHY PREPROCESSING MUST TRAVEL WITH IT
Inputs at prediction must be transformed identically to training.
WHAT HAPPENS OTHERWISE
The model receives values unlike what it learned from.
WHAT A PIPELINE OBJECT PROVIDES
Preprocessing and model as one unit.
WHAT FORMATS EXIST
Framework-native serialisation Interchange formats, portable across runtimes Compiled formats for specific hardware
WHAT INTERCHANGE FORMATS PROVIDE
Running without the training framework, frequently faster.
WHAT TO BE CAREFUL WITH
Serialisation tied to library versions Formats that execute arbitrary code on loading
WHY THAT SECOND POINT MATTERS
Loading an untrusted model file can execute code.
WHAT TO RECORD IN METADATA
Training data reference Code version Parameters Metrics achieved When it was trained Who trained it
WHAT A MODEL REGISTRY PROVIDES
Versioned storage with that metadata and a promotion path.
WHAT TO NEVER DEPLOY
A model whose provenance is unknown.
