Moderate LLM outputs with LlamaGuard

Classify text for 6 safety categories

Production input/output filtering where you need a specialized 7B moderation model instead of general LLM.

• · Conversation turns
• · User prompts or bot responses
• · Safety context

• · Classification (safe/unsafe)
• · Category (S1-S6)
• · Confidence

• · transformers
• · torch
• · vllm (optional)

HuggingFace auth token; 8GB VRAM for 7B model

• · False positives (over-blocking)
• · Category ambiguity (S3 vs. S4)
• · Truncated text loss

• · Meta's specialized safety model
• · 94-95% accuracy on safety benchmarks