When a production model starts giving wrong answers, rollback depends on knowing what is running, what ran before it, and where the previous checkpoint lives. Versioning records distinct model iterations and their provenance so teams can reproduce decisions, revert deployments, and maintain an audit trail. section 9.1
A fine-tuned model version needs more than a prompt string: record the checkpoint, base model, training-data hash, and evaluation report. Preserve the prompt baseline, few-shot bank, held-out set, and validation harness alongside it so a rollback can restore the earlier approach. section 4.9.3
Table 9.1 Minimum fields in a model version registry entry
|
Field
|
Example
|
Why
|
|
Version tag
|
it-support-v2-20260215-lora-Qwen3-4B-v1
|
Human-readable identifier
|
|
Technique
|
lora, sft,
distill, dpo
|
Tracks the training method
|
|
Base model
|
Qwen/Qwen3-4B-Instruct-2507
|
Required for reproducibility
|
|
Data hash
|
abc123
|
Ties the model to the exact training data
|
|
Checkpoint path
|
chapter06/runs/sft_run1
|
Where to find the weights
|
|
Eval metrics
|
{"overall_f1":
0.72, "safety_pass_rate": 1.0}
|
Performance at registration time
|
|
Status
|
registered,
active, retired
|
Lifecycle state
|
|
Created timestamp
|
2026-02-15T19:36:04Z
|
Audit trail
|
Use these fields as a minimum registry entry; also retain the hyperparameters needed to reproduce the training run. section 9.1.1 section 9.1
A descriptive tag can encode the name, version number, date, technique, base model, and data version: {name}-v{N}-{date}-{technique}-{base_model_short}-{data_version}. The tag helps identify a candidate quickly; the registry entry holds the details and checkpoint location. section 9.1.1
Keep the registry lifecycle explicit:
- Register each new version with its metadata and checkpoint path. section 9.1.2 section 9.1.1
- List versions to see their status, then promote the selected version to active; the replaced version becomes retired. section 9.1.2
- Roll back the most recent promotion when the new version degrades. section 9.1.2
Rolling back a degraded experiment
An IT-support registry records an active SFT baseline with F1 of 0.72 and safety pass rate of 1.0. A DPO experiment fell to 0.58 and 0.75 respectively and was retired after rollback. The entries retain data hashes and checkpoint paths, making the history inspectable. section 9.1.3
Start versioning before the first deployment, even with one model. A JSON file is easy to inspect, diff in a pull request, and commit alongside the code. As the team grows beyond roughly 10–15 model versions or needs multi-team access control, consider a managed registry; the fields and concepts remain useful while the storage backend changes. section 9.1.3 section 9.1.2
Table 9.2 Registry approaches by team scale
|
Approach
|
Best for
|
Tradeoffs
|
|
JSON file (this chapter)
|
Small teams, 1 to 10 models
|
Simple, version-controlled, no
infrastructure
|
|
MLflow / W&B Registry
|
Mid-size teams, 10 to 50 models
|
UI, search, comparison; requires a server
|
|
Vertex AI / SageMaker
|
Large teams, regulated industries
|
Full lifecycle management; cloud lock-in
|
|
DVC / lakeFS
|
Data-heavy pipelines
|
Data versioning; steep learning curve
|
|
Git LFS
|
Large model artifacts
|
Integrated with Git; storage costs scale
|
Compare registry options by team scale and operational tradeoffs before choosing infrastructure. section 9.1.2



































