Meta Llama Guard
used in
0
stacks
Family of open-weights LLM-based safety classifiers from Meta for moderating prompts and responses.
Fine-tuned Llama models that classify user prompts and model responses as safe or unsafe across hazard categories aligned with the MLCommons taxonomy. The Llama Guard 3 generation ships in 1B, 8B, and 11B-vision variants, covering multilingual text and mixed text-image inputs. Weights are distributed under the Llama Community License and are available for self-hosting or via inference providers.
Strengths
- Open weights, self-hostable, with multiple sizes from 1B to 11B including a vision variant
- Hazard categories aligned with the MLCommons safety taxonomy
- Taxonomy can be adjusted via prompting or fine-tuning for custom policies
- Widely available through hosted inference providers and guardrail toolkits
Tradeoffs
- Requires GPU serving or a paid inference provider; classification adds latency per call
- Llama Community License is not an OSI-approved open-source license and carries usage restrictions
- Default taxonomy may not match application-specific policies without customization