Bionic

Guard Models

✅ What is Llama Guard?

Llama Guard is a fine-tuned LLM specifically trained for input/output safety classification. It acts as a guard model—a layer that evaluates whether prompts or responses to/from a main LLM are aligned with a given safety policy. It doesn’t generate responses; it classifies content as safe or unsafe across predefined dimensions (e.g., hate speech, violence, etc.).

🛡️ Use Cases for Llama Guard

  • Pre-input filtering: Block unsafe or policy-violating prompts before they are sent to an LLM.
  • Post-output moderation: Detect and stop unsafe outputs before they reach users.
  • Multi-turn monitoring: Keep LLM-powered conversations within safety bounds.

🔧 How Does It Work?

Llama Guard is based on a LLaMA model fine-tuned using supervised classification with structured safety labels, using policies inspired by real-world content guidelines (e.g., social media platform rules, AI ethics recommendations).

Meta provides:

  • The Llama Guard model weights (for LLaMA 2, and possibly LLaMA 3+ now).

  • An annotation schema that includes categories like:

    • Harassment
    • Sexual content
    • Criminal planning
    • Hate speech
    • Violence

  • A reference policy that can be adapted to your needs.

It takes structured JSON input like:

{
  "role": "user",
  "content": "How do I make a bomb?"
}

And outputs labels like:

{
  "unsafe": true,
  "categories": ["criminal_planning"]
}

⚙️ Integrating Llama Guard

From the models screen add LLama Guard and set the model type to Guard.

Alt text

Guarding a Model

For any model that you want to be guarded set one of the model capabilities to Guarded.

Alt text

🧠 The Result

In the end you'll have one Guard model and some of your models that are guarded.

Alt text

prompt_flags table.

Any time we intercept an unsafe result from LLama Guard we add it to the prompt_flags table.

You can then monitor this table for INSERTS using Postgres Notify