data versioning
In plain wordsKeeping a record of every version of the data used to train and test AI models, so that any past result can be traced back to exactly the data that produced it.
The practice and tooling for recording immutable, identifiable versions of datasets — training data, labels, features and evaluation sets — so that every model can be linked to the exact data that produced it, earlier states can be restored, and changes between versions can be inspected. Because datasets are often too large for ordinary source control, tools store content-addressed snapshots or pointers in Git while the data lives in object storage (DVC, Data Version Control), add Git-like branches and commits over data lakes (lakeFS), or rely on table formats with built-in 'time travel' (Delta Lake, Apache Iceberg); model and dataset hubs version data as Git repositories with large-file storage. Data versioning supports reproducibility, debugging of regressions, audit and regulatory documentation, and deletion requests, by recording which models were trained before a person's data was removed.
Tag the dataset version in the training config so we can reproduce the March model exactly.
Origin
Adapted from software version control to machine learning in the late 2010s, when teams found that versioning code alone could not reproduce models; DVC, first released by Dmitry Petrov in 2017, popularized the practice, and lakeFS and lakehouse table formats extended it to data lakes around 2020.
Related
- experiment trackingexperiment tracking, n.: Keeping an automatic logbook of every model-training attempt — settings, data, code and results — so a team can compare attempts, repeat the good ones and see what changed.
- MLOpsMLOps, n.: The working habits and tools that keep machine-learning models dependable inside real products once the experiment is over: automated testing, release, monitoring and retraining; the factory and maintenance crew around a model, not the model itself.
- data provenancedata provenance, n.: The documented history of a piece of data: where it came from, who created and changed it, under what permissions, and how it was processed before reaching an AI system.
- model registrymodel registry, n.: A central catalog where a team stores every version of its trained models, with records of how each was made and which version is approved for use.
- golden datasetgolden dataset, n.: A small, carefully checked set of example questions with known good answers, used as a trusted yardstick to test an AI system every time it changes.
- data pipelinedata pipeline, n.: The chain of automated steps that collects raw data, cleans and transforms it, and delivers it in a usable form — the factory line that turns messy inputs into training material.