TheArchitect256/qwen3.5-2b-triage-master
TheArchitect256/qwen3.5-2b-triage-master is a 2.3 billion parameter Qwen3.5-based model fine-tuned for tool/intent routing. It functions as a specialized router in multi-agent pipelines, classifying user queries to dispatch to the correct downstream tool. This model excels at single-label classification for tool selection, achieving 99.4% accuracy on its evaluation set. It is designed to efficiently direct tasks based on a given schema, making it suitable for orchestrating complex AI workflows.
Loading preview...
Overview
TheArchitect256/qwen3.5-2b-triage-master is a 2.3 billion parameter model, developed by TheArchitect256, specifically fine-tuned for tool/intent routing. Built upon the unsloth/Qwen3.5-2B base, its primary function is to act as a routing node within multi-agent pipelines (e.g., LangGraph, CrewAI). It classifies user queries to determine the appropriate tool from a predefined schema, effectively dispatching tasks to the correct downstream component.
Key Capabilities
- High Accuracy Tool Routing: Achieves 99.4% accuracy on a held-out evaluation set of 500 examples, demonstrating robust performance in selecting the correct tool.
- Efficient Intent Classification: Designed for single-label classification, framing tool selection as a text generation task.
- Optimized Training: Fine-tuned using LoRA (r=16, alpha=16, dropout=0) on approximately 5,000 custom instruction examples covering various routing scenarios with 2-4 tools per prompt. Training was accelerated using Unsloth and TRL SFTTrainer.
Use Cases
This model is ideal for developers building complex AI systems that require intelligent task orchestration. It can be integrated into:
- Multi-agent frameworks: To direct user requests to specialized agents or tools.
- Workflow automation: For classifying user intent and triggering specific actions or API calls.
- Chatbots and virtual assistants: To accurately interpret user needs and route them to the relevant function (e.g., checking stock, calculating shipping).