MarkrAI/ksafeguard-8b

K-SafeGuard: Korean LLM Safety Moderation

K-SafeGuard (ksafeguard-8b) is a specialized 8 billion parameter moderation classifier developed by Marker-Inc-Korea, built on kakaocorp/kanana-1.5-8b-instruct. Its primary function is to evaluate the safety of Korean LLM interactions, addressing both user prompts and AI responses.

Key Capabilities

• WildGuard 3-task Schema: Classifies interactions based on prompt harmfulness, response refusal, and response harmfulness in a single call.
• Reduced Over-Refusal: Specifically trained to minimize false positives for Korean idiomatic and figurative expressions (e.g., "마약김밥" for delicious kimbap), a common issue with models trained on translated data.
• Flexible Input: Can classify prompt harmfulness alone or perform the full 3-task classification when both prompt and response are provided.
• Performance: Achieves high F1 scores across various Korean safety benchmarks, including WJ (0.983), WG-Prompt (0.954), WG-Refusal (0.960), WG-Resp (0.917), and FR (0.917). It demonstrates the lowest over-refusal rate (FPR 0.124) among compared public guards.

Intended Use Cases

• Content moderation for Korean LLM inputs (prompts) and outputs (responses).
• Diagnosing over-refusal in LLM responses.
• Serving as a classifier within safety filtering pipelines for Korean language applications.

Limitations

• Performance is optimized for Korean; accuracy in other languages is not guaranteed.
• Automated moderation may produce misclassifications and should not replace human review.