10.3 Data and model versioning
You can trace a prediction back to the data that trained it.
Data and model versioning extends Git's guarantees to the things Git cannot hold, so a prediction can be traced to the exact data and weights behind it, with schema contracts catching upstream changes early. It sits beside experiment tracking as the other half of reproducibility. The silent failure it prevents: code versioned while datasets drift means the same pipeline quietly stops meaning the same thing, and nobody can say when.
Work through these
DVC or lakeFS for data
Version control for data, which ordinary source control handles badly because of size. Two tools are covered, both storing pointers alongside the code.
Dataset cards and schema contracts
A written description of a dataset, and a declared shape it must conform to. The schema is what catches a silently changed upstream column before it reaches a model.
Model versioning and lineage
Recording which data and which code produced which model, so a prediction can be traced back. This is what a compliance review actually asks for.
Reproducibility guarantees and their limits
Exact reproducibility has real limits, from hardware differences to library versions, and stating more certainty than you have is worse than naming the limit.
Sign in to keep your progress.
Free resources
We haven't checked most of these for screen reader use yet.
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.