MiniGuard-v0.1: Compact Content Safety Classifier

MiniGuard-v0.1, developed by prem-research, is a highly efficient 0.6 billion parameter content safety classifier based on Qwen3-0.6B. It is engineered to classify both user prompts and LLM responses, identifying unsafe content and specifying violated safety categories.

Key Capabilities & Differentiators

• High Accuracy, Low Parameters: Achieves ~99% of Nemotron-Guard-8B's benchmark accuracy with 13x fewer parameters (0.6B vs 8B).
• Cost-Effective: Offers significant cost savings, estimated at 67% per 1 million requests compared to Nemotron-Guard-8B-v3, while outperforming larger specialized models like Qwen3Guard-8B on production data.
• Drop-in Replacement: Uses the same prompt template as nvidia/Llama-3.1-Nemotron-Safety-Guard-8B-v3, allowing for easy integration.
• Targeted Training: Utilizes targeted synthetic data, "Think SFT" (distilling reasoning traces), and Top-K Model Soup for enhanced out-of-distribution generalization.
• Output Format: Provides safety assessments in a structured JSON format, indicating "User Safety", "Response Safety", and "Safety Categories".

Use Cases

MiniGuard-v0.1 is ideal for applications requiring efficient and accurate content moderation for both user inputs and AI-generated responses. It is particularly well-suited for scenarios where minimizing inference cost and latency is critical, without significantly compromising safety performance.