Nemotron Agentic Tool Use v1 is an agentic tool-use dataset released by NVIDIA in 2025, containing 335,122 samples with a total size of approximately 5.5 GB, designed to enhance language models' abilities to decompose user goals, plan tasks, decide when to call tools, and understand tool outputs.

The dataset contains synthetic multi-turn dialogue trajectories that simulate interactions among users, agents, and tool execution environments, and are scored and filtered by language models to remove inconsistent or incoherent trajectories. The data is suitable for supervised fine-tuning, data augmentation, and the training and evaluation of agentic models requiring multi-step planning, tool calling, and result reasoning.

Dataset Composition

• Interactive Agent: Contains 19,028 synthetic multi-turn interaction trajectories, simulating complete interaction processes among users, agents, and tool execution environments. Data generation and filtering use models such as Qwen3-235B-A22B-Thinking-2507, Qwen3-32B, GPT-OSS-120B, and Qwen3-235B-A22B-Instruct-2507.

• Tool calling: Contains 316,094 tool-calling instances, generated by collecting tool sets from public datasets and simulating tool usage scenarios. Seed data for the user simulator comes from the Nemotron-Personas-USA dataset, and Qwen3-235B-A22B-Thinking-2507 and Qwen3-235B-A22B-Instruct-2507 are used for dialogue simulation and turn-level judgment.