safety-alignment-llamaguard

原文涵盖的主题

• LlamaGuard - AI Content Moderation
• Quick start
• Common workflows
• Workflow 1: Input filtering (prompt moderation)
• Workflow 2: Output filtering (response moderation)
• Workflow 3: vLLM deployment (fast inference)
• Workflow 4: API endpoint (FastAPI)
• Workflow 5: NeMo Guardrails integration

LlamaGuard is a 7-8B parameter model specialized for content safety classification. Installation: Basic usage: Check user prompts before LLM: Safety categories: S1: Violence & Hate S2: Sexual Content S3: Guns & Illegal Weapons S4: Regulated Substances S5: Suicide & Self-Harm S6: Criminal Planning Check LLM responses before showing to user: Production-ready serving: Throughput: 50-100 requests/sec on single A100 Serve as moderation API: Usage: Use with NVIDIA Guardrails: Use LlamaGuard when: Need pre-trained moderation model Want high accuracy (94-95%) Have GPU resources (7-8B model) Need detailed safety categories Building production LLM apps Model versions: LlamaGuard 1 (7B): Original, 6 categories LlamaGuard 2 (8B): Improved, 6 categories LlamaGuard 3 (8B): Latest (2024), enhanced Use alternatives instead: OpenAI Moderation API: Simpler, API-based, free Perspective API: Google's toxicity detection NeMo Guardrails: More comprehensive safety framework Constitutional AI: Training-time safety Issue: Model access denied Login to HuggingFace: Accept license on model page: https://huggingface.co/meta-llama/LlamaGuard-7b Issue: High latency (>500ms) Use vLLM for 10× speedup: Enable tenso

这里只展示节选，请以完整原文和当前要求为准。