Toucan

# 🦤 Toucan-1.5M: Toucan-1.5M is the largest fully synthetic tool-agent dataset to date, designed to advance tool use in agentic LLMs. It comprises over 1.5 million trajectories synthesized from 495 real-world Model Context Protocols (MCPs) spanning 2,000+ tools. By leveraging authentic MCP environments, Toucan-1.5M generates diverse, realistic, and challenging tasks requires using multiple tools, with trajectories involving real tool executions across multi-round, multi-turn, sequential, and parallel tool calls. Models fine-tuned on Toucan-1.5M outperform much larger closed-source counterparts on the BFCL V3 benchmark and extend the Pareto frontier on the MCP-Universe benchmark. - 📄 [Technical Report](https://arxiv.org/abs/2510.01179) - Technical details behind Toucan-1.5M - 💾 [Github Repo](https://github.com/TheAgentArk/Toucan) - Pipeline to produce Toucan-1.5M - 🤗 [HF Dataset](https://huggingface.co/datasets/Agent-Ark/Toucan-1.5M) - Full dataset  ## 🚚 Installation ```bash # Create Env conda create -n toucan python=3.12 -y conda activate toucan # Install Required Packages pip install torch pip install -r requirements.txt # Install Qwen Agent from Source cd Qwen-Agent; pip install -e .; cd ../ ``` ## 📝 Data Synthesis Please refer to [`./datagen`](./datagen/) folder for details. ## 📚 Citation If you find the data or code useful, please cite: ``` @misc{xu2025toucan, title={TOUCAN: Synthesizing 1.5M Tool-Agentic Data from Real-World MCP Environments}, author={Zhangchen Xu and Adriana Meza Soria and Shawn Tan and Anurag Roy and Ashish Sunil Agrawal and Radha Poovendran and Rameswar Panda}, year={2025}, eprint={2510.01179}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2510.01179}, } ```