Use source control for machine learning assets

The Proseware diabetes classification model started in a Jupyter notebook on one laptop. When the team grows — data scientists, ML engineers, and reviewers — that approach breaks down. You need a place where everyone works from the same code, can review each other's changes, and can trace every model back to the code that produced it.

What belongs in Git

A Git repository is the right home for the assets that define your model and its training process:

• Training and preprocessing scripts — the Python files that run data preparation and model training
• Job and component definitions — YAML files that describe Azure Machine Learning command jobs and pipeline components
• Environment and dependency files — requirements.txt, conda.yml, or Azure Machine Learning environment definitions
• Configuration files — workspace names, resource group names, compute names, and other references your automation uses

What doesn't belong in Git

Some assets belong in purpose-built services, not in a repository:

• Datasets and raw data — use Azure Machine Learning data assets or Azure Data Lake Storage; data files are large and versioned separately from code
• Trained model artifacts — register named and versioned model assets in Azure Machine Learning
• Training run outputs and logs — Azure Machine Learning stores these with the job
• Secrets and credentials — use GitHub secrets or Azure Key Vault; putting credentials in a repository exposes them to anyone with read access

Why source control matters for machine learning

When each training definition lives in Git, your team works from the same version of the code. A reviewer sees exactly what changed between runs. If a change degrades model performance, you can return to the previous commit. The recorded code and environment definitions also help you reproduce a training run.

This traceability is the foundation for everything that follows — branch-based development, automated validation, and an audit trail that connects every model version to the change that produced it.