Guard models

Guard models are classifiers trained to judge whether a prompt, a response or a tool call breaks a policy. They run on every request beside the main model, so they are kept small and fast.

How they work

A guard model takes one message or tool call and returns a verdict: whether it breaks a policy, and which one. Some learn a fixed set of categories in training. Others read the policy text along with the content, so a rule can be added without retraining.

Why they are small

The check sits in the path of every request, so its latency is added to every answer. Classifying is a narrow job, so a guard model can be far smaller and faster than the model it guards.

How they are kept accurate

  • Labeled data. They are trained on traffic that breaks each policy and traffic that does not. See classification fine-tuning.
  • A larger model. It re-checks the small model's verdicts off the request path, where its slowness delays nobody, and its verdicts become training data. This is a form of knowledge distillation.
  • New attacks. The model is retrained on what red teaming and live traffic turn up.

With White Circle

We train our own guard models: a small one gives verdicts in real time, a larger one confirms them, and the larger model's verdicts are then used to improve the small one. Control analyzes every input and output in real time.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL