Llama Guard 3 8B

Classifies incoming user prompts against a taxonomy of 6 content safety categories (violence, illegal activity, self-harm, sexual content, harassment, and specialized harms) using a fine-tuned Llama 3.1 8B backbone. The model outputs structured safety labels with confidence scores, enabling real-time filtering of unsafe requests before they reach downstream LLMs. Uses instruction-following patterns from Llama 3.1 training combined with safety-specific fine-tuning to distinguish between discussing harmful topics (safe) and requesting harmful actions (unsafe).

Classifies LLM-generated outputs (responses, completions, assistant messages) against the same 6-category safety taxonomy to detect when downstream models produce unsafe content. Operates on the same fine-tuned Llama 3.1 8B architecture but is applied post-generation to catch safety failures in model outputs. Enables real-time detection of jailbreak successes, hallucinated harmful instructions, or unintended unsafe content generation.

Returns safety classifications as structured JSON with per-category confidence scores (typically 0.0-1.0 range) rather than binary pass/fail verdicts, enabling fine-grained safety policy decisions. The model outputs logits or probability distributions across the 6 safety categories, allowing applications to set custom thresholds per category (e.g., stricter on violence, more lenient on political content). Implements a multi-label classification approach where content can be flagged in multiple categories simultaneously.

Classifies content against specialized harm categories beyond standard content policy violations, including CSAM-related content, illegal activities, self-harm, and harassment. The fine-tuning incorporates patterns for detecting nuanced harms (e.g., grooming language, suicide encouragement) that may not be caught by keyword-based or simple pattern-matching approaches. Uses instruction-following capabilities of Llama 3.1 to understand context and intent rather than relying on surface-level text matching.

Supports batch processing of multiple prompts or responses through OpenRouter's API, enabling efficient classification of large volumes of content without per-request overhead. Integrates with OpenRouter's batch API infrastructure to queue, process, and retrieve safety classifications asynchronously, reducing per-request latency and cost for high-volume moderation pipelines. Handles rate limiting, retries, and result aggregation transparently.

Classifies safety across multiple languages using the same fine-tuned Llama 3.1 8B model, leveraging the base model's multilingual capabilities. However, safety fine-tuning is primarily optimized for English, with varying accuracy across other languages depending on training data representation. The model uses cross-lingual transfer learning to extend English safety patterns to other languages, but performance degrades gracefully for low-resource languages or non-Latin scripts.

Integrates with LLM frameworks (LangChain, LlamaIndex, Anthropic SDK, OpenAI SDK) and safety middleware systems through standardized API interfaces. Can be deployed as a prompt guard (pre-LLM) or response filter (post-LLM) in application chains, with built-in support for async/await patterns, error handling, and fallback logic. Supports integration with observability platforms for logging, monitoring, and alerting on safety violations.

Provides safety classifications that can be composed with custom policy rules and business logic to implement application-specific safety policies. The model outputs structured category scores that applications can combine with custom rules (e.g., 'block if violence_score > 0.7 AND user_is_minor', 'warn if harassment_score > 0.5 AND user_is_verified'). Enables policy-as-code approaches where safety decisions are driven by composable rules rather than hard-coded thresholds.