tau-emotion-bench/Tracer-Low-4B
Tracer-Low-4B is a 4 billion parameter interactive tool-using agent developed by tau-emotion-bench, fine-tuned from Qwen/Qwen3-4B. It specializes in handling multi-turn tool-use dialogues with emotionally expressive users across customer service domains. The model reasons using a block before each reply or tool call and supports a 32,768 token context length, making it suitable for robust agentic applications requiring nuanced user interaction.
Loading preview...
Tracer-Low-4B: Interactive Tool-Using Agent for Emotional User Interactions
Tracer-Low-4B is a 4 billion parameter interactive tool-using agent, fine-tuned from Qwen/Qwen3-4B. Developed by tau-emotion-bench, this model is specifically designed to handle multi-turn tool-use dialogues with emotionally expressive users, making it particularly effective in customer service scenarios.
Key Capabilities
- Emotionally Aware Tool Use: Fine-tuned on tool-use dialogues with emotionally expressive users across four customer-service domains (Retail, Telehealth, Telecom, Airline).
- Structured Reasoning: Employs a
<think>block for internal reasoning before generating replies or making tool calls, enhancing its decision-making process. - Improved Performance: Demonstrates significant performance improvements over its base model, Qwen3-4B, on the \u03c4-Emotion-Bench benchmark across various domains and user simulators, particularly in pass@1 and pass@3 metrics.
- High Context Length: Supports a maximum context length of 32,768 tokens, allowing for complex and extended interactions.
Good for
- Customer Service Agents: Ideal for building interactive agents that can effectively respond to and manage user emotions in customer support environments.
- Tool-Using Applications: Suitable for applications requiring robust tool integration and structured reasoning within conversational flows.
- Research on Emotional AI: Useful for researchers exploring the interaction fragility of LLM agents with real-world user emotions, as detailed in the associated paper \u03c4-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions.