Versioning
Tracking model changes over time so you can reproduce results, rollback, and understand what changed between versions.
Think of it like
Git for your model; each commit is a snapshot with metadata and a message.
Example
Model v1.0 trained on 2024 data scores 92% accuracy; v1.1 with 2025 data scores 91%—rollback to v1.0 if needed.
How it actually works
Enables reproducibility: same version + same code + same data = same results. Must version not just weights but also preprocessing, tokenization, and dependencies. Semantic versioning (major.minor.patch) signals compatibility and significance of changes.
For product teams
Speeds incident response (instant rollback) and compliance reporting (audit trail).
For engineers
Artifacts tagged with version string, checksummed, and linked to code commit and dataset version; serving layer selects version at runtime.
Related
- Model Registry — Tracked by.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome