Overview
Agent-Ark/Toucan-Qwen2.5-7B-Instruct-v0.1 is a 7.6 billion parameter model, fine-tuned from Qwen2.5-7B-Instruct. Its core differentiator is its training on the Toucan-1.5M dataset, the largest fully synthetic tool-agent dataset to date. This dataset, created by Agent-Ark, consists of over 1.5 million trajectories synthesized from 495 real-world Model Context Protocols (MCPs) and 2,000+ tools, focusing on diverse and realistic multi-round, multi-turn, sequential, and parallel tool calls.
Key Capabilities
- Advanced Tool Use: Specifically optimized for complex agentic behaviors involving multiple tools.
- Benchmark Performance: Outperforms larger closed-source models on the BFCL V3 benchmark and extends the Pareto frontier on the MCP-Universe benchmark, demonstrating superior tool-use capabilities.
- Diverse Task Handling: Trained on a supervised fine-tuning (SFT) subset of 119.3K instances, including extensions for irrelevance, diversification, and multi-turn interactions, enhancing robustness.
Training Details
The model was fine-tuned using the Hermes prompt template on a curated subset of Toucan-1.5M. The dataset generation methodology and technical specifics are detailed in the technical report.
Good For
- Developing and deploying agentic LLMs that require sophisticated tool interaction.
- Applications demanding high performance in multi-tool, multi-step reasoning tasks.
- Research into improving LLM capabilities for complex, real-world problem-solving through tool integration.