Implementation guide
OWASP AISVS v1.01
Model Lifecycle Management & Change Control · Pipeline Fine-Tuning
Register each fine-tuning checkpoint
Verify that fine-tuning checkpoints are registered as distinct artifacts.
OWASP AISVS v1.01, requirement 3.5.4
In plain English
Register every fine-tuning checkpoint as its own artifact with its own identifier. Then "the model we tested", "the model we approved" and "the model we deployed" can be shown to be the same file.
What good looks like
- Each checkpoint has a unique identifier, a hash and a registry entry.
- Test results and approvals point to a specific checkpoint.
- Lineage is recorded: which base model, data and earlier checkpoint it came from.
- A written policy says which checkpoints are kept and for how long.
How teams usually do it
- Log checkpoints to the model registry as they are produced, with the run ID and step number.
- Promote by hash. The deployment pipeline takes the hash that appears on the approval record.
- Automatic clean-up of unpromoted checkpoints under the retention policy.
Suggested evidence
- Registry entries for recent checkpoints, with lineage.
- An approval record that names a checkpoint hash, matched to the deployed artifact.
Common mistakes
- Overwriting a file called "latest", so nobody can tell which version passed testing.
- Testing one checkpoint and deploying a slightly later one because training carried on.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.