Nemotron RL Agentic Conversational Tool-Use (Pivot v1)

Nemotron RL Agentic Conversational Tool-Use Pivot v1 is a reinforcement learning dataset released by NVIDIA in 2026 as part of NeMo Gym, its open-source framework for building RL environments that train large language models through Reinforcement Learning from Verifiable Reward (RLVR). It was used in the development of NVIDIA's Nemotron family of models. The dataset takes expert multi-turn tool-use trajectories and recasts each assistant step as a separate behaviour cloning problem, rewarding a policy model for matching the tool call the expert chose. The scenarios span 838 domains such as EdTech, IT support and travel, and exercise tools for authentication, data lookup and servicing actions like booking or amending reservations and applying discounts.

You access it on Hugging Face with no account or application, which is why the access tier is open. It ships as a single train.jsonl file of about 1.75 GB holding 170,320 examples (the card quotes roughly 3.0 GB), with eleven fields per record covering the trajectory, the request parameters, the expected action, scenario metadata and pass-rate statistics. The format is built for NeMo Gym rather than for generic supervised fine-tuning, so you will typically consume it through that framework.

For AI work its primary use is post-training conversational agents to call the right tool at the right turn, and it is best suited to RL and behaviour-cloning pipelines rather than plain retrieval. For RAG it is less a knowledge base than a bank of worked tool-use dialogues you can mine for few-shot examples or to study how agents sequence authentication, lookup and action calls.

The data is fully synthetic, generated with DeepSeek-R1-0528, DeepSeek-V3.2, Qwen3-235B and Qwen3-32B and labelled with GPT-oss-120b and Qwen3-235B-Instruct, so there are no third-party content concerns. It is licensed CC BY 4.0 and the card states explicitly that it is ready for commercial use, requiring only attribution to NVIDIA. It sits alongside our Toucan-1.5M, xLAM and Glaive function-calling training sets, and complements evaluation suites like tau3-bench and BFCL.