Zichen1024/CoVe-4B
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm
CoVe-4B is a compact 4 billion parameter interactive tool-use agent developed by Zichen1024, fine-tuned from Qwen3-4B-Instruct-2507. It utilizes the CoVe (Constraint-Verification) post-training framework, trained on a 12K dataset of multi-turn tool-use trajectories. This model excels at complex tool invocation and verification, achieving top performance in its size group and rivaling larger models on tool-use benchmarks.
Loading preview...
CoVe-4B: A Compact Interactive Tool-Use Agent
CoVe-4B is a 4 billion parameter interactive tool-use agent, fine-tuned from Qwen3-4B-Instruct-2507. Developed by Zichen1024, this model leverages the innovative CoVe (Constraint-Verification) post-training framework to enhance its tool-use capabilities.
Key Capabilities & Features
- Constraint-Guided Verification: The CoVe framework uses explicit constraints to guide a User Simulator LLM and deterministically verify the agent's tool invocations, ensuring high reliability.
- Specialized Training Data: Trained on the CoVe-12K dataset, which consists of 12,000 high-quality, multi-turn tool-use trajectories synthesized and verified through deterministic constraint checking.
- Strong Performance: Achieves top performance among models in the \u22648B parameter group on the \u03c4\u00b2-bench benchmark, demonstrating competitive results even against models up to 70B parameters.
- Hermes Tool-Call Format: Utilizes the Hermes tool-call format, facilitating deployment with tools like vLLM.
Ideal Use Cases
- Interactive Agent Development: Suitable for building agents that require precise and verifiable tool interactions.
- Complex Task Automation: Excels in scenarios demanding multi-step tool invocation and robust constraint adherence.
- Research in Tool-Use LLMs: Provides a strong baseline and framework for further research into constraint-guided agent training.