The Nemotron RL Agentic Conversational Tool-Use Pivot v1 dataset is a dataset focused on reinforcement learning for conversational tool use, released by NVIDIA in 2026. It aims to train large language models in behavioral cloning within complex interaction scenarios by leveraging existing expert tool-use trajectories.

The dataset contains 170,320 training samples with a total size of approximately 3.0 GB, covering tool-calling scenarios across 838 different domains, including EdTech, IT support, and travel, with diverse tasks such as identity authentication, data querying, service booking, and discount acquisition. The data is generated using synthetic methods, leveraging large models such as DeepSeek and Qwen to generate trajectories, combined with GPT-oss and Qwen for label synthesis. The format is structured JSON files compatible with the NeMo Gym framework, suitable for reinforcement learning with verifiable rewards (RLVR) tasks during post-training of large models.

Dataset Composition

The dataset mainly contains the following structure and fields:

• Data format: Structured JSON, compatible with the NVIDIA NeMo Gym framework.
• Number of samples: A total of 170,320 records in the training set.
• Data size: Approximately 3.0 GB.
• Core fields: Each record contains 11 top-level fields, specifically including:
• trajectory_id: Trajectory identifier
• responses_create_params: Response creation parameters
• expected_action: Expected action
• scenario: Scenario description
• num_unique_actions: Number of unique actions
• meta_info: Meta information
• qwen_235b_info: Qwen model-related information
• agent_ref: Agent reference
• pass_rate: Pass rate
• pass_rate_total: Total pass rate
• pass_rate_passed: Pass status