Nemotron-RL-Agentic-Terminal-Pivot-v1 Dataset Overview

[NEMOTRON_RL_AGENTIC_TERMINAL_PIVOT_V1_BADGE] Nemotron-RL-Agentic-Terminal-Pivot-v1 is an agentic terminal reinforcement learning dataset released by NVIDIA in 2026, designed to provide large language models with reinforcement learning training samples tailored for terminal (command-line) interaction scenarios. Specifically developed for the terminus_judge environment within NeMo Gym, this dataset supports post-training reinforcement learning, supervised fine-tuning, and offline behavioral analysis of agents.

The dataset comprises 31,111 training samples derived from 630 ATCB seed tasks and 2,716 distinct source trajectories. Task scenarios span long-term terminal operations across domains such as industrial systems, high-performance computing clusters, healthcare, finance, and media archiving—including data pipeline construction, fault auditing, service diagnostics, and security compliance workflows. Each task features a median of 45 samples; these consist of single-agent decision points extracted from successful final-task completion trajectories. Key components include task prompts, reference next-step actions generated by the Terminus-2 agent using GLM-5.1 as its teacher model, and routing information utilized for reward scoring. On average, task prompts contain approximately 39,900 characters, while reference next-step actions average around 970 characters.

Data Fields

• responses_create_params.input: Prompt containing both the task instruction and the complete history of terminal interactions up to the current decision point.
• expected_answer: Reference action representing the optimal next step taken by the agent at that specific state.
• agent_ref: Routing information pointing to the NeMo Gym terminus_judge resource server, enabling policy model action evaluation and generation of RL rewards.
• metadata: A structured object comprising five sub-fields; internal collection identifiers and infrastructure details have been stripped out.

[JIALIU_0]