Llama Guard is a language-model-based safety classifier. It served as an output classifier after domain fine-tuning degraded safety on sensitive topics; the team retained the adaptation and added the safeguard. They also ran Garak before deployment and before each later fine-tuning cycle, treating evaluation as a gate. section 4.9 section 4.9
Llama Guard
Llama Guard is a language-model-based safety classifier that evaluates content against defined safety categories.
How AI Governance covers it
- How deep it goes
- Describes its role as an output classifier but does not detail its internals or configuration.
- Examples
- Uses degraded safety on sensitive topics after domain fine-tuning as the reason to add the classifier.
- What you can do after
- Apply Llama Guard as an output classifier when fine-tuning weakens safety.
- How it connects
- Links Llama Guard to fine-tuning, safety alignment, and Garak evaluation.
Related concepts
- guardrailkind of · A guardrail is a constraint, control, or monitoring mechanism that limits system behavior to reduce risks and keep actions or outputs within acceptable bounds.
- guard modelkind of · A guard model is a compact classifier that labels prompts or outputs for safety-related properties such as benignness or maliciousness.
This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.
