ThakiCloud/Qwen3.8-27B-Human-KO-Enterprise-Boundary
ThakiCloud/Qwen3.8-27B-Human-KO-Enterprise-Boundary is a 27 billion parameter enterprise policy decision model, derived from the Qwen3.8 architecture, specifically designed to classify agent actions. It was trained on contrastive boundary pairs in Korean to decide between seven predefined actions (e.g., ANSWER, CALL_TOOL, REFUSE) before text generation. This model excels at zero-shot English transfer for enterprise policy decisions, demonstrating robust performance despite being supervised exclusively in Korean, and is optimized for single-forward-pass decision-making.
Loading preview...
Overview
ThakiCloud/Qwen3.8-27B-Human-KO-Enterprise-Boundary is a 27 billion parameter model, a boundary-targeted checkpoint from the Qwen3.8 architecture, specifically designed for enterprise policy decision-making. Unlike general-purpose LLMs, this model's primary function is to determine one of seven predefined actions an enterprise agent should take under a written policy, before any text is generated. These actions include ANSWER, ASK_CLARIFICATION, ASK_CONFIRMATION, CALL_TOOL, CALL_MULTIPLE_TOOLS, ESCALATE, and REFUSE.
Key Capabilities
- Enterprise Policy Decision: Specializes in classifying agent actions based on written policies.
- Korean-Supervised, Zero-Shot English: Uniquely trained exclusively on Korean data for enterprise and boundary-targeted supervision, yet demonstrates strong zero-shot transfer capabilities to English policy scenarios without any English fine-tuning.
- Boundary-Targeted Training: Optimized using contrastive boundary pairs to distinguish between critical decision points (e.g., answer vs. call a tool, call a tool vs. ask for confirmation).
- Efficient Decision Making: Designed for single-forward-pass decisions, offering a faster alternative to models that require reasoning latency for unseen policies.
Good For
- Automated Enterprise Agent Action Classification: Ideal for systems requiring agents to make discrete, policy-driven decisions quickly and consistently.
- Korean and English Policy Enforcement: Suitable for applications in both Korean and English-speaking environments where enterprise policies need to be interpreted and acted upon.
- Reducing Latency in Decision Workflows: Beneficial for scenarios where rapid, pre-generation action classification is critical, recovering part of the performance gap from full reasoning with a single forward pass.
- Research into Policy Decision Learning: Provides a valuable resource for studying how enterprise policy decisions can be learned versus reasoned, particularly concerning boundary internalization and cross-lingual transfer.