How to meet it
OWASP AISVS v1.01
Model Lifecycle Management & Change Control · Pipeline Fine-Tuning
Verify models used in RLHF
Verify that models used in RLHF fine-tuning are versioned and integrity-verified before use in a training run.
OWASP AISVS v1.01, requirement 3.5.1
In plain English
When fine-tuning with human feedback, version and verify the reward model and reference models before each run, since a swapped or altered model changes what the new model learns.
What good looks like
- Reward and reference models are registered and versioned.
- Their integrity is checked at the start of each run.
- Run records state which versions were used.
How teams usually do it
- Hash checks against the model registry when the training job starts.
- Run metadata stored with the resulting model.
Evidence an auditor would accept
- Training run records listing model versions and hashes.
Common mistakes
- Loading the reward model from a shared path that anyone can overwrite.
Not sure which requirements apply to your system? Take the assessment.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.