Model Lineage: The Git History for Your AI

WHY IT EXISTS: Models are not just code; they are the output of code, data, and configuration. Without a record of these inputs, a trained model becomes an un-reproducible, unauditable artifact, creating massive technical debt and regulatory risk. Model lineage was created to solve this by systematically documenting a model's origins.

THE MENTAL MODEL: Think of model lineage as the complete git log for a trained AI artifact. It's the full provenance record that answers the question: "How exactly was this model created, and can we reproduce it?" It tracks every component—data versions, code commits, hyperparameters, and parent models—that contributed to the final product.

HOW IT WORKS: Model lineage is captured by integrating several MLOps tools. Experiment trackers like MLflow or Weights & Biases log hyperparameters, code versions, and evaluation metrics during training. Data versioning tools like DVC or LakeFS pin the exact version of the training dataset. A model registry then stores the final model artifact along with all this collected metadata. For CI/CD, the entire workflow is captured as code in a pipeline, making the process itself part of the lineage.

WHEN TO USE IT: Lineage is critical in three main scenarios. First, for reproducibility, allowing any engineer to recreate the exact same model. Second, for debugging, to trace a model's unexpected behavior back to a specific data change or hyperparameter tweak. Third, for compliance, as regulations like the EU AI Act and GDPR require auditable records of how models are built and what data they were trained on.

WHEN NOT TO USE IT: The overhead of rigorous lineage tracking might be excessive for early-stage, non-critical experiments or disposable proof-of-concept models where reproducibility is not a primary concern. However, as soon as a model shows promise for production, establishing its lineage becomes non-negotiable.

ONE CANONICAL EXAMPLE: A team fine-tunes a foundation model for a customer support chatbot. Their lineage record tracks the specific version of their internal Q&A dataset, the preprocessing script's commit hash, the learning rate and epoch count, the exact base model used, and the final evaluation scores. When a GDPR data deletion request arrives, they use the lineage to confirm if the user's data was in the training set. The lineage for the base model itself, however, remains unknown to them.