Implementation guide
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
RLHF (reinforcement learning from human feedback) uses helper models, usually a reward model that scores outputs and a reference model that keeps the result from drifting too far. Version these models and verify they are unaltered before every run. A swapped reward model changes what the new model learns to want.
What good looks like
- The reward model, reference model and base model are registered with versions and hashes.
- The training job checks each hash against the registry before it starts, and stops on a mismatch.
- The run record names the exact versions used.
- Write access to these models is as tightly controlled as for production models.
How teams usually do it
- A hash or signature check against the model registry in the first step of the training job.
- Loading helper models from the registry by immutable version, never from a shared file path.
- Storing run metadata, including helper model hashes, with the resulting model.
Suggested evidence
- Training run records listing model versions and hashes.
- Job logs showing the integrity check passed.
- Access policy for the helper models.
Common mistakes
- Loading the reward model from a shared folder anyone on the team can overwrite.
- Retraining the reward model between runs without giving it a new version.
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.