UltraData-SFT-Agent-2609 is an agent supervised fine-tuning dataset released by OpenBMB. The associated paper is "UltraData-SFT-Agent-2609" and the dataset is designed for supervised fine-tuning on agent-related tasks, providing data support for model training.

The dataset contains 483,661 Agent interaction trajectories, covering four major directions: general agents, tool use, code agents, and search agents. The data records the complete process from instruction understanding, tool invocation, and environment interaction to result verification, error recovery, and final delivery.

Dataset Composition

The dataset mainly consists of the following four parts:

• General-Agent: 311,006 entries, covering comprehensive tasks such as skill invocation, office work, and file processing;
• Tool-Use: 82,760 entries, covering tasks such as function calling, database operations, and tool usage;
• Code-Agent: 69,895 entries, covering tasks such as software engineering, code writing, testing, and bug fixing;
• Search-Agent: 20,000 entries, covering tasks such as web search, multi-hop question answering, and information retrieval.

Citation

@misc{ultradata_sft_agent_2609,
  title        = {UltraData-SFT-Agent-2609},
  author       = {MiniCPM Team},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/openbmb/UltraData-SFT-Agent-2609}}
}