Data Versioning
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
The practice of saving datasets as distinct, retrievable versions so any past state can be restored, compared, or produced as evidence in an audit.
What is Data Versioning?
Data versioning is the practice of saving datasets as distinct, retrievable versions, so any past state can be restored, compared, or produced as evidence. It does for data what version control does for code.
The practice matters most in machine learning, where a model is the product of code plus data. Versioning the code alone reproduces nothing if the training set has changed since.
How Data Versioning Works
Datasets are too large to copy wholesale on every change, so versioning tools store snapshots efficiently: they hash file contents, keep each unique block once, and record which blocks make up each version. Tools built on this idea include DVC, Git LFS, and lakeFS.
Each training run then records the exact data version it used, alongside the code commit and hyperparameters. Reproducing the run later means checking out both.
Applications of Data Versioning
- Reproducibility: retraining a model on the exact dataset that produced it, months or years later.
- Rollback: reverting to a known-good dataset after a bad ingestion or a labeling error.
- Comparison: measuring how a model changes when trained on last quarter’s data versus this quarter’s.
- Governance: proving to an auditor what data a deployed model was trained on. The EU AI Act requires technical documentation for high-risk AI systems that describes their training data; a pinned dataset version is concrete evidence behind that documentation.
Example of Data Versioning
An insurer’s internal audit questions a pricing model deployed 14 months ago: was a prohibited attribute present in its training data? The current customer table is no help – it has gained 14 months of new rows and lost records deleted on request.
Because every training run pinned a dataset version, the team checks out the exact snapshot the model saw, inspects its columns, and reruns the training job to confirm it yields the same model. The audit question takes an afternoon instead of becoming an unanswerable dispute.
Related AI terms: Data Lineage · Data Integrity · Version Control · Model Registry · MLflow
Did you like the Data Versioning gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About